With the widespread adoption of Transformer models in computer vision, natural language processing, and related domains, inference workloads have grown rapidly, placing increasingly stringent demands on inference throughput. Existing inference frameworks primarily improve local execution efficiency through operator fusion, kernel generation, and graph-level optimization. However, operators in Transformer models exhibit distinct resource preferences with respect to hardware resources such as Tensor Cores, CUDA Cores, and memory bandwidth. Constrained by data dependencies, these operators are typically executed in topological order, making it difficult to achieve sufficient overlap among heterogeneous hardware resources and thereby limiting further improvements in inference throughput. To address this issue, this paper proposes a batch-partitioned pipelined inference method for Transformer models. Transformer computation consists of diverse operators, including matrix multiplication, reduction, element-wise computation, and tensor reshaping. Based on operator resource-occupancy characteristics and offline performance profiling, we classify Transformer operators into three categories: Tensor Core-intensive, CUDA Core-intensive, and memory-bandwidth-intensive operators. To overcome the limited overlap among operators with different resource preferences under conventional execution, the original inference batch is partitioned into three sub-batches with adjustable sizes and scheduled concurrently while preserving the computational-graph dependencies within each sub-batch. By allowing different sub-batches to progress through different computation stages, operators with complementary resource preferences can be overlapped in a pipelined manner. This reduces contention for the same type of hardware resource, mitigates pipeline bubbles, and improves overall hardware utilization. Furthermore, we formulate the With the widespread adoption of Transformer models in computer vision, natural language processing, and related domains, inference workloads have grown rapidly, placing increasingly stringent demands on inference throughput. Existing inference frameworks primarily improve local execution efficiency through operator fusion, kernel generation, and graph-level optimization. However, operators in Transformer models exhibit distinct resource preferences with respect to hardware resources such as Tensor Cores, CUDA Cores, and memory bandwidth. Constrained by data dependencies, these operators are typically executed in topological order, making it difficult to achieve sufficient overlap among heterogeneous hardware resources and thereby limiting further improvements in inference throughput. To address this issue, this paper proposes a batch-partitioned pipelined inference method for Transformer models. Transformer computation consists of diverse operators, including matrix multiplication, reduction, element-wise computation, and tensor reshaping. Based on operator resource-occupancy characteristics and offline performance profiling, we classify Transformer operators into three categories: Tensor Core-intensive, CUDA Core-intensive, and memory-bandwidth-intensive operators. To overcome the limited overlap among operators with different resource preferences under conventional execution, the original inference batch is partitioned into three sub-batches with adjustable sizes and scheduled concurrently while preserving the computational-graph dependencies within each sub-batch. By allowing different sub-batches to progress through different computation stages, operators with complementary resource preferences can be overlapped in a pipelined manner. This reduces contention for the same type of hardware resource, mitigates pipeline bubbles, and improves overall hardware utilization. Furthermore, we formulate the

Abstract: The Common Vulnerability Scoring System (CVSS) serves as the core foundation for vulnerability prioritization, remediation decision-making, and compliance auditing. In real-world Security Operations Centers (SOCs), semi-structured data—such as ports, protocols, services, banners, and version numbers—captured by scanners often arrives prior to the formal Common Vulnerabilities and Exposures (CVE) descriptions. Existing automated scoring methods largely rely on structured vulnerability description text, making them ill-suited for workflows where scanner output precedes vulnerability disclosure. Furthermore, the CVSS base vector comprises eight distinct metrics—Attack Vector, Attack Complexity, Privileges Required, User Interaction, Scope, and the impact on Confidentiality, Integrity, and Availability. A misjudgment in any single field can significantly alter the risk prioritization of a vulnerability. Consequently, a model must not only predict individual labels but also generate a semantically consistent and verifiable complete base vector in a single pass.To address these challenges, we propose VulXbert, an interpretable multi-label prediction model for CVSS v3.1 base vectors. The model employs a dual-channel semantic alignment framework to encode vulnerability description text and host scan outputs independently, leveraging the security-domain pre-trained language model SecureBERT to enhance the representation of vulnerability terminology, service names, port protocols, and semi-structured symbols. During the training phase, the model utilizes temperature-scaled cross-entropy contrastive learning to map inputs from both description and scan channels into a unified embedding space, thereby narrowing the expression gap between public vulnerability disclosures and raw scan logs. During the prediction phase, the model utilizes the shared representation as input to a shared architecture with eight task-specific softmax classification heads, enabling the output of all CVSS v3.1 base vector fields in a single forward pass.To mitigate the impact of class imbalance on minority label performance, we introduce category-weighted cross-entropy, with weights normalized by label frequency in the training set. Furthermore, by incorporating input-gradient saliency analysis, the model provides token-level key evidence for each prediction. This allows security analysts to pinpoint the specific fields supporting judgments on attack vectors, privileges required, or scope, thereby improving the transparency, auditability, and efficiency of human review. Experiments were conducted using approximately 60,000 records from the National Vulnerability Database (NVD) (2016–2024) containing complete CVSS v3.1 base vectors, augmented with real-world scan data. Compared to existing methods such as CVSS-BERT, BERT-multi, Costa et al.’s methods, Kühn et al.’s open-source intelligence-based approach, and AutoCVSS, VulXbert achieved a Macro-F1 of 0.89, Micro-F1 of 0.91, Hamming Loss of 0.054, and Exact Match Accuracy of 68%. Metric-specific analysis demonstrates that VulXbert achieves more stable performance across Attack Vector, Privileges Required, Scope, and the three impact metrics. Ablation studies confirm that the security-domain pre-trained encoder, cross-source semantic alignment, and the joint multi-head prediction structure each contribute significantly to model performance. In summary, by utilizing cross-source semantic alignment, joint multi-label prediction, and evidence-based interpretability, VulXbert provides a scalable solution for automated, interpretable CVSS base vector generation in environments where scan logs arrive ahead of vulnerability descriptions.

In the digital economy era, digital finance plays a crucial role in alleviating financing constraints for SMEs and serving the real economy. However, SMEs generally suffer from irregular financial practices and a lack of collateral, leading to severe information asymmetry in credit risk assessment. Traditional risk control models heavily rely on static quarterly or annual financial statements, resulting in significant data lag and making it difficult to objectively and in real-time depict the true operational dynamics of enterprises. In recent years, high-frequency power load data has gradually become a core alternative risk control data source due to its ability to objectively, continuously, and tamper-resistantly reflect the actual production status of enterprises (such as production line operating rates and equipment utilization rates). Existing credit assessment methods based on power data mainly rely on expert rule engines or purely data-driven deep learning models. However, these methods have significant limitations in practical financial business applications: traditional expert rules are too rigid and struggle to capture hidden anomalies in massive nonlinear time series; while purely data-driven models are prone to falling into a "black box" dilemma when faced with lengthy high-frequency sequences and are highly susceptible to interference from normal fluctuations such as sudden changes in weather conditions and short-term shutdowns (such as shift changes and routine maintenance). This results in a lack of ability for the model to accurately focus on core default risks and an inability to provide a traceable chain of evidence that meets the requirements of financial compliance audits. To address these issues, this paper proposes a joint research approach for electricity load forecasting and credit risk early warning based on multi-source heterogeneous data, aiming to bridge the mapping mechanism from underlying physical power fluctuations to high-level credit defaults. First, in the cross-frequency feature perception stage, this paper utilizes a zero-order hold mechanism to achieve precise alignment between daily macro-meteorological covariates and 15-minute high-frequency power consumption sequences, establishing a dynamic prediction baseline and removing operational dynamic residuals after excluding environmental temperature interference, thereby physically decoupling and extracting real micro-level operational anomalies. Second, in the knowledge representation stage, this paper abandons the simple variable concatenation method and instead uses multi-hot encoding and feature projection techniques to deeply characterize enterprise business semantics (industry category), qualification labels (such as litigation default status), and quantitative capital (such as registered capital and contract capacity) into a dense heterogeneous knowledge matrix, constructing a unique priori risk control profile for each enterprise. Finally, this paper designs a knowledge-guided multi-head cross-attention mechanism, using the enterprise profile knowledge matrix as a prior query filter to perform deep noise reduction and feature purification on high-frequency residual sequences. This mechanism can guide the network to adaptively ignore load drops caused by normal production rhythms and accurately amplify structurally abnormal power outage signals caused by precipitous order shrinkage or broken capital chains. Furthermore, to collaboratively optimize prediction and recognition tasks within the same framework, this paper introduces a multi-task adaptive dynamic loss function based on homoscedastic uncertainty. This mechanism, combined with multi-classification focus loss, effectively addresses the inherently extremely unbalanced nature of the four risk levels. Simultaneously, the model is supplemented with electrical physical capacity upper limit constraints and risk monotonicity logic constraints, significantly improving joint discrimination ability and internal interpretability while ensuring high-precision load forecasting. We conducted large-scale experiments on real business datasets of leading global power companies. Experimental results show that the proposed method outperforms baseline models in both load forecasting and credit rating tasks, achieving state-of-the-art performance.

Visual and acoustic modalities provide important nonverbal cues for multimodal sentiment analysis, but they are also easily affected by environmental noise, individual expression differences, and modality-quality variations. When unreliable nonverbal information is directly introduced into multimodal fusion, it may cause unstable cross-modal interactions and biased sentiment prediction. To address this problem, this paper proposes a visual-acoustic residual compensation method for multimodal sentiment analysis. Instead of redesigning the whole multimodal fusion framework, the proposed method constructs a controlled residual compensation path on top of a language-focused multimodal fusion model, aiming to calibrate the remaining prediction bias of the base multimodal prediction. Specifically, textual, visual, and acoustic modalities are first encoded to generate a base multimodal prediction through a language-focused fusion structure. Then, the language-focused visual and acoustic representations are concatenated and fed into a residual compensation branch composed of fully connected mappings, nonlinear activation, dropout, shortcut connection, and a scalar output layer. This branch produces a visual-acoustic compensation value for output-level incremental calibration. To avoid excessive correction caused by auxiliary modalities, a forward compensation weight, a residual clipping threshold, a compensation amplitude regularization term, and a residual supervision objective are introduced. In this way, the compensation branch is encouraged to learn the remaining bias between the base prediction and the ground-truth sentiment label, rather than directly fitting the complete sentiment target. Considering that the stability of base prediction and residual patterns may vary under different data conditions, two residual-training paths are designed. For CMU-MOSEI, the model adopts a Residual+Stage path, in which the base prediction branch is trained first, the residual compensation branch is then optimized, and the whole model is finally fine-tuned jointly. For CMU-MOSI, the model adopts a Residual-only path, where the base prediction branch and the residual compensation branch are optimized end-to-end. To reduce the confounding factors caused by cross-dataset comparison, controlled down-sampling experiments with different training scales are further conducted on CMU-MOSEI while keeping the validation set, test set, modality features, and evaluation scripts unchanged. Experimental results show that the proposed method outperforms the DLF baseline on all six evaluation metrics on CMU-MOSEI and CMU-MOSI. On CMU-MOSEI, the proposed method achieves 55.26% Acc-7, 57.04% Acc-5, 86.91% Acc-2, and 86.88% F1, improving over DLF by 1.36, 1.34, 1.49, and 1.61 percentage points, respectively. The correlation coefficient increases from 0.764 to 0.780, and the mean absolute error decreases from 0.536 to 0.530. On CMU-MOSI, the proposed method achieves 48.19% Acc-7, 54.50% Acc-5, 86.67% Acc-2, and 87.70% F1, improving over DLF by 1.11, 2.17, 1.61, and 2.66 percentage points, respectively. The correlation coefficient increases from 0.781 to 0.800, and the mean absolute error decreases from 0.731 to 0.721. Ablation results show that different residual-training paths are suitable for different data conditions. On CMU-MOSEI, Residual+Stage achieves the best overall performance, while on CMU-MOSI, Residual-only performs better. The controlled training-scale experiments on CMU-MOSEI further show that Residual-only is more stable when only 10% or 25% of the training data is used, whereas Residual+Stage gradually becomes more effective as the training ratio increases to 50%, 75%, and 100%. This indicates that training sample size affects the stability of base prediction error patterns and the adaptability of residual-training paths. The compensation-value analysis shows that the average absolute compensation values on CMU-MOSEI and CMU-MOSI are only 0.018 and 0.021, respectively, and the proportions of samples close to the compensation boundary are 2.6% and 3.1%. These results indicate that the compensation branch does not largely rewrite the base prediction, but performs controlled calibration within a small range. The computational cost analysis further shows that the residual compensation branch introduces only about 2.03×104 additional parameters, increasing the total parameter size from 111.23M to 111.25M, with only slight increases in training and inference time. Overall, the proposed method provides an effective output-level calibration strategy for controlled use of visual and acoustic modalities, and improves multimodal sentiment prediction with limited additional computational cost.

Braille is a fundamental textual medium through which blind and visually impaired people access information, receive education, and participate in social life. Automatic and accurate Chinese-to-Braille translation provides important technical support for promoting information accessibility and information equity. However, several technical challenges remain in automatic Chinese-to-Braille conversion. Traditional rule- and dictionary-based methods rely heavily on manually designed pinyin mappings, Braille word segmentation and concatenation rules, and punctuation conversion rules. Owing to their limited ability to understand deep contextual semantics, these methods tend to achieve relatively low accuracy in scenarios involving polyphone disambiguation, ambiguous sentence structures, and long-distance semantic dependencies. Although general neural machine translation models provide end-to-end sequence transformation capabilities, they are susceptible to insufficient generalization and sequence alignment errors under low-resource Chinese–Braille parallel corpus conditions. Moreover, most existing neural network models employ a unified feedforward structure, making it difficult to handle differentiated linguistic phenomena, including semantic understanding, Braille word segmentation, and symbol generation, in a specialized manner. Consequently, these models may produce insufficiently fluent translations and fail to maintain consistency with Braille rules when processing complex texts. To address these limitations, this paper proposes a Chinese-to-Braille translation model that integrates a pretrained language model with a Mixture-of-Experts (MoE) network, thereby constructing an end-to-end translation architecture that combines contextual semantic understanding with task-specific feature representation. The proposed model adopts an encoder–decoder architecture. In the encoder, a pretrained Bidirectional Encoder Representations from Transformers (BERT) model is employed as the core semantic feature extraction module. Its bidirectional contextual encoding capability is used to improve the identification of Chinese polyphones, ambiguous structures, and long-distance semantic dependencies. An MoE enhancement layer is introduced after the BERT output layer. A learnable gating network computes the routing weights of different experts and dynamically integrates their outputs, thereby improving the model’s ability to process differentiated linguistic information related to semantic features, Braille word segmentation boundaries, punctuation marks, and special symbols. The decoder consists of six stacked Transformer Decoder layers. Conditioned on the enhanced semantic representations produced by the encoder, the decoder combines target Braille sequence embeddings with positional encoding information and generates the target Braille symbol sequence autoregressively. To support model training and evaluation, a Chinese–Braille parallel corpus was compiled in collaboration with a school for blind students. The corpus covers several common text categories, including news reports, literary works, science education materials, public service documents, and daily communication texts. Data cleaning, duplicate removal, and Braille symbol normalization were performed to improve the consistency and usability of the corpus. The model was evaluated using the Bilingual Evaluation Understudy (BLEU), Character Error Rate (CER), Character Accuracy (CharAcc), and Exact Match (EM). These metrics respectively measure overall n-gram consistency, character-level editing errors, character-level correctness, and complete sequence consistency. Under the same training data, data split, preprocessing procedure, and evaluation settings, the proposed model achieved BLEU, CER, CharAcc, and EM scores of 98.15%, 1.46%, 98.54%, and 63.99%, respectively, on the self-built test set. Compared with the standard Transformer baseline under the same experimental conditions, the proposed model improved BLEU by 11.64 percentage points and reduced CER by 17.24 percentage points. The ablation results showed that introducing the MoE network alone increased BLEU from 86.51% to 92.12% and reduced CER from 18.70% to 10.17%. Introducing the BERT encoder alone increased BLEU to 97.02% and reduced CER to 1.84%. When the BERT encoder and the MoE network were used together, the model achieved the best results among the ablation models evaluated in this study, indicating that the two components provide complementary benefits. A subjective experiment involving visually impaired users was also conducted to assess the practical reading quality of the generated Braille texts. The results showed that the translation method had a significant effect on Braille reading scores. The proposed model received relatively high ratings for reading fluency, content comprehensibility, and semantic accuracy across daily-life, science education, and literary texts, while its rating distributions were relatively concentrated. The combined results of automatic evaluation, ablation experiments, and the user study indicate that BERT-based semantic encoding enhances the contextual representation of Chinese inputs, whereas the MoE mechanism dynamically integrates expert outputs according to input features. Their combination helps reduce character-level Braille errors and improve the exact-match rate of complete sequences. At a medium model scale, the proposed method achieves favorable Chinese-to-Braille translation performance and can provide technical support for Braille publishing, accessible conversion of digital reading materials, intelligent assistive devices, and public Braille information services.

In response to the problems in image tampering detection, such as the insufficient perception ability of tampered regions with small scales and blurred edges, which easily leads to missed detections; the fusion mechanisms of multi - domain features in the spatial and frequency domains often adopt simple splicing or serial attention, failing to fully explore the complementarity and collaborative discriminative ability of cross - domain features; complex models usually have a large number of parameters and high computational costs, making it difficult to achieve a balance between accuracy and efficiency in actual deployment scenarios with limited resources. This paper proposes an image tampering detection network FDU - Net that integrates dual - stream features and an attention aggregation mechanism. Based on the U - shaped encoder - decoder architecture, this method constructs a parallel input mechanism for RGB and JPEG dual - streams. The RGB stream directly processes the original image, aiming to extract spatial - domain apparent features such as color, texture, and structure to capture the differences in visual consistency between the tampered region and the background. The JPEG stream maps the image to the frequency - domain representation through a carefully designed artifact extraction module and uses multi - layer convolution to capture hidden traces in the compression domain, such as block effects, quantization errors, and spectral anomalies left by tampering operations, thus effectively compensating for the lack of robustness of single - spatial - domain features to post - processing operations.In terms of the backbone network, to balance feature expression efficiency and computational economy, this paper uses the lightweight MobileNetV4 as the shared backbone for the dual - streams. Considering the special sensitivity of the tampering detection task to high - frequency edge information, this paper proposes to use two general inverted bottleneck module variants, UIB - 1 and UIB - 2, for differential combined configuration. UIB - 1 alleviates the loss of high - frequency information during the down - sampling process through pre - depth convolution pre - processing, and UIB - 2 enhances the receptive field and channel interaction through post - depth convolution, thus outputting feature maps rich in multi - scale semantics and details at five scales.In the feature fusion stage, this paper designs a dual - stream attention feature fusion module. This module breaks through the limitations of traditional serial or single - dimensional attention mechanisms, models the long - range dependence between RGB and JPEG features in parallel in the channel and spatial dimensions through cross - domain multi - head self - attention, and introduces a linear compression layer to reduce computational complexity, achieving in - depth interaction and adaptive purification of dual - stream features and effectively solving the problem of insufficient multi - domain feature fusion.In addition, to improve the positioning accuracy of tampering boundaries, this paper introduces an accuracy detection head module, which combines the dual mechanisms of channel attention and spatial attention. Channel attention filters discriminative feature channels through global pooling and fully connected layers, and spatial attention fuses the results of average pooling and maximum pooling to strengthen the boundary response. The synergistic effect of the two significantly enhances the model's sensitivity to blurred - edge and small tampered regions.This paper conducts a large number of experiments on multiple public and self - built datasets, such as CASIAV1, CASIAV2, COVERAGE, IMD2020, and the self - made VOC2012TAMPER based on PASCAL VOC2012. The experimental results show that the F1 score of the model reaches 0.8921, and it outperforms existing mainstream algorithms in key indicators such as F1, MCC, and AUC. The proposed method performs excellently in terms of detection accuracy, robustness, and generalization ability, and the model has a relatively low number of parameters, demonstrating good practicality and deployability.

In high-load edge environments, bursty task arrivals and limited multidimensional resources at edge nodes, including CPU, I/O, bandwidth, and memory, can cause queue congestion and resource fragmentation, thereby affecting scheduling metrics such as resource utilization, running time, and waiting time. Existing learning-based scheduling methods often struggle to capture higher-order resource competition between tasks and servers, and their continuous outputs cannot be directly converted into discrete placement decisions that satisfy capacity constraints. This paper proposes EdgeHypOp, an edge task scheduling method based on hypergraph modeling and bin-packing decoding. EdgeHypOp jointly optimizes resource utilization, running time, and waiting time. For each scheduling window, it constructs a bipartite hypergraph over task and server nodes with four hyperedge types. Capacity, task-server, latency, and utilization hyperedges represent shared resource budgets, candidate placement relations, task waiting states, and the overall system load, respectively. To limit online scheduling overhead under high-load conditions, EdgeHypOp calculates task urgency from waiting time and execution duration and bounds the candidate batch size per round. A multilayer hypergraph neural network propagates information between nodes and hyperedges to produce task and server representations that encode resource demands, remaining capacities, and higher-order competition relations. Bilinear matching then generates a task-server score matrix. Since this matrix does not directly satisfy discrete capacity constraints, EdgeHypOp applies a Tetris-inspired multi-resource bin-packing decoder. Among feasible candidates, the decoder considers the learned score, the match between task demands and remaining server resources, task waiting priority, and execution duration. It places tasks iteratively, updates server states after each placement, then applies lightweight local optimization to reorder tasks within each server and reduce the blocking of short tasks by long tasks. Experiments are conducted on three real-world datasets, Google Cluster Trace, Alibaba Cluster Trace, and EUA, with FCFS, HRRN, RLPNet, PSNet, HyperJet, and aRL as baselines. All methods use the same task scales, server scales, and test instances, and results are reported as means and standard deviations over 10 runs. Under a high-load setting with 2,000 tasks and 5 servers, EdgeHypOp achieves the highest resource utilization on all three datasets. Compared with the strongest baseline for each metric, it improves resource utilization by 3.93 percentage points and reduces running time and waiting time by 4.18% and 14.03%, respectively, on Google Cluster Trace. On Alibaba Cluster Trace, the changes are an increase of 2.01 percentage points and reductions of 2.25% and 8.16%; on EUA, they are an increase of 2.10 percentage points and reductions of 2.19% and 0.65%. Scale experiments further show that EdgeHypOp maintains more consistent advantages in resource utilization and waiting time as the number of tasks increases or server resources become constrained, while the performance gaps among methods narrow under low-load or resource-abundant conditions. Ablation experiments show that removing bipartite hypergraph modeling or multi-resource bin-packing decoding increases the composite objective value by 7.89% and 4.29%, respectively, indicating that higher-order relation representation and constraint-aware decoding are the main sources of performance gains. Removing local optimization increases the composite objective value by 2.46%, and this component provides greater improvements for CPU-intensive and memory-intensive tasks. Online overhead tests show that when the task scale increases from 500 to 2,000, inference time rises from 704.49 ms to 2,178.42 ms, while peak CPU and GPU memory usage increases only slightly. Model training is performed offline. During online scheduling, EdgeHypOp performs only batch selection, hypergraph construction, forward propagation, scoring, bin-packing decoding, and local optimization, without gradient updates. The results show that EdgeHypOp combines higher-order resource competition representation, multidimensional capacity constraints, and discrete task placement in a unified scheduling process. It is suitable for edge task scheduling scenarios with high task concurrency and pronounced resource competition and fragmentation, while low-load settings, single-resource bottlenecks, and workloads with strong task dependencies require further extensions.

The large-scale deployment of the Industrial Internet of Things (IIoT) exposes industrial networks to increasingly complex security threats, making network intrusion detection a key measure for ensuring the safe operation of IIoT systems. However, significant disparities exist in sample sizes among attack categories in IIoT traffic, resulting in insufficient recognition capability of existing detection models for high-risk minority-class attacks. Meanwhile, methods based on complex architectures such as Transformers incur large parameter counts and high computational overhead, making them difficult to deploy under the resource constraints of edge devices. To this end, a Hierarchical Layer-Adaptive Feature Network (HLAF-Net) for intrusion detection is proposed to achieve high-precision and low-latency detection for class-imbalanced traffic. At the data-processing level, certain continuous features in IIoT traffic follow long-tailed or sparse distributions, whose discriminative information is prone to being dominated by extreme values or diluted by a large number of zero entries during direct normalization. A logarithmic transformation is applied to such features to compress the order-of-magnitude gap between extreme and ordinary values, so that the discriminative distances among samples are preserved after normalization. At the structural-design level, the classification task is decomposed into several levels according to the sample counts of attack categories. Since each level only needs to process the sample subset associated with that level, some features used to distinguish shallow-level categories become constant within this subset and no longer provide discriminative information; a zero-variance criterion is therefore introduced to remove such redundant features level by level, and a level-specific feature subspace is constructed for each level. At the network-architecture level, each level first adopts a gated attention mechanism to adaptively recalibrate the input features, and then employs a lightweight Multi-Layer Perceptron (MLP) in parallel with an Explicit Feature Interaction Branch (EFIB) to extract semantic representations and high-order interaction features, respectively, which are fused and passed through a classification output layer to obtain level-wise predictions. During inference, all levels output their prediction probabilities in parallel, and a cascaded level-wise decision logic determines the final category. HLAF-Net is experimentally validated on two public datasets, Edge-IIoTset and X-IIoTID. Ablation studies show that the logarithmic transformation yields the most significant gain, raising the Macro-F1 score from 77.16% to 91.04%; the hierarchical architecture, supported by high-quality input features, further improves the Macro-F1 score to 95.06%; and the gated attention mechanism enhances the convergence stability of the deeper levels during training. On the Edge-IIoTset dataset, HLAF-Net achieves a detection accuracy of 98.89%, a Weighted-F1 score of 98.91%, and a single-sample inference time of only 0.0979 ms, meeting the real-time requirements of edge deployment while maintaining detection accuracy. Across five independent runs, the standard deviations of accuracy and Weighted-F1 score remain below 0.0002, indicating stable performance that does not depend on a specific random initialization. Compared with models based on the same hierarchical divide-and-conquer paradigm, HLAF-Net improves accuracy and Weighted-F1 score by 2.18% and 2.70%, respectively, verifying the effectiveness of the level-specific feature subspaces and the network branch design. On the X-IIoTID dataset, both accuracy and Weighted-F1 score reach 98.98%, further demonstrating the effectiveness and robustness of the model across different IIoT environments.

Integrated Sensing and Communication (ISAC) has been recognized as one of the key enabling technologies for future wireless networks. Reconfigurable Intelligent Surface (RIS)-assisted multi-Unmanned Aerial Vehicle (UAV) cooperative ISAC systems can simultaneously exploit the advantages of wireless propagation environment reconfiguration and aerial mobility. However, such systems face significant challenges, including strong coupling among optimization variables, high-dimensional decision spaces, and the difficulty of jointly enhancing communication and sensing performance. To address these issues, this paper proposes a joint optimization framework for RIS-assisted multi-UAV cooperative ISAC systems. A three-dimensional discrete-time system model consisting of multiple UAVs, multiple ground base stations, a sensing target, and an RIS is established, where UAV mobility constraints, RIS reflection mechanisms, air-to-ground communication processes, and target sensing processes are jointly characterized. A non-convex optimization problem is formulated to maximize a weighted utility comprising the system sum communication rate and a sensing performance metric based on the log-determinant of a dimensionless Fisher Information Matrix (FIM), with transmit beamforming, UAV trajectories, and RIS phase shifts jointly optimized. The sensing utility is derived from the cumulative FIM contributions of both direct sensing links and RIS-assisted sensing links, which explicitly captures the impacts of UAV spatial deployment, RIS configuration, and target observation geometry on sensing performance. To tackle the non-convex objective function, unit-modulus constraints, and the strong coupling between communication and sensing components, an Alternating Optimization (AO) framework is developed. Specifically, the original problem is decomposed into three subproblems: transmit beamforming optimization, UAV trajectory optimization, and RIS phase-shift optimization. For transmit beamforming optimization, the Weighted Minimum Mean Square Error (WMMSE) method is employed to transform the weighted sum-rate maximization problem into an equivalent iterative minimization problem, while incorporating sensing-aware weighting for resource allocation. For UAV trajectory optimization, the Successive Convex Approximation (SCA) method is adopted to locally convexify the communication rate term, sensing utility term, and safety-distance constraints, and trust-region and smoothing regularization techniques are introduced to improve trajectory feasibility and numerical stability. For RIS phase-shift optimization, the non-convex quadratically constrained quadratic programming problem under unit-modulus constraints is transformed into a Semidefinite Relaxation (SDR) problem, and a grouping strategy combined with Gaussian randomization is employed to reduce the complexity of high-dimensional RIS phase recovery. Simulation results demonstrate that the proposed algorithm exhibits stable convergence behavior under different system parameters and network geometries, typically reaching convergence within 8–15 iterations. Statistical results obtained from 100 independent Monte Carlo trials show that, compared with the non-RIS scheme, the proposed RIS-assisted joint optimization method improves the average system communication rate by 24.83%, indicating that RIS-enabled propagation environment reconfiguration can effectively enhance air-to-ground communication link quality. Compared with the communication-only optimization scheme, incorporating the FIM-based sensing utility increases the average flight altitude of multiple UAVs by 16.58%, demonstrating that the proposed joint objective can guide UAVs to form more favorable spatial observation geometries for target localization while simultaneously satisfying multi-UAV safety constraints. Furthermore, sensitivity analysis with respect to communication-sensing weighting factors and experiments involving different numbers of UAVs reveal a pronounced synergistic effect among RIS-enabled environment reconfiguration, multi-UAV spatial mobility, and FIM-based sensing information modeling. Overall, the proposed method achieves communication rate enhancement, sensing geometry improvement, and comprehensive utility maximization while satisfying UAV safety-flight constraints and RIS unit-modulus constraints, providing an effective solution for joint resource allocation and trajectory design in complex air-ground ISAC networks.

In the task of insulator defect (e.g., breakage and flashover) detection in transmission lines, traditional object detection models suffer from receptive field mismatches, semantic gaps in feature fusion, parameter redundancy in detection heads, and high missed detection rates for minute defects. This is primarily due to the strong background interference in aerial images, the minute size and extreme scale variations of the targets, and the geometric anisotropy of insulators, which exhibit elongated structures and arbitrary spatial orientations. To satisfy the strict constraints of edge computing devices such as unmanned aerial vehicles (UAVs) regarding model storage overhead and real-time inference, this study proposes DDE-YOLO, a lightweight and efficient insulator detection algorithm based on an improved YOLO11n. The proposed algorithm reconstructs the network across three dimensions: feature extraction, feature fusion, and prediction regression. First, a Dynamic Inception Mixer Block (C3K2-DIMB) is innovatively designed in the backbone network. Its internal Dynamic Inception Depthwise Separable Convolution employs parallel square, horizontal strip, and vertical strip convolution kernels to extract multi-directional features. Combined with a channel-level dynamic weighting mechanism, the network adaptively perceives the elongated distribution characteristics and spatial orientations of insulators, effectively filtering background noise and enhancing valid feature representation. Second, addressing the information loss problem during multi-scale feature fusion, a Dynamic Adaptive Double Fusion Network (DAFPN) is constructed. Utilizing a Dynamic Attention Gated Fusion (DAGF) module, this network dynamically filters multi-level semantic information through grouped interactions and a bidirectional gating mechanism. By generating feature weights based on global contexts, it significantly suppresses background interference, bridges the semantic gap between deep semantics and shallow details, and strengthens the capability to capture minute defect textures in complex backgrounds. Third, an Efficient Shared Detail Head (ESDH) is designed, which implements cross-scale sharing of feature extraction and regression weights among multi-scale detection layers and adopts Group Normalization (GN) to improve training stability. To mitigate the degradation of scale sensitivity caused by parameter sharing, a learnable scale calibration factor is introduced at the end of the regression branch. Furthermore, Detail Enhancement Convolution (DEConv) is employed to strengthen high-frequency edge detail extraction via differential convolution branches during the training phase. In the inference phase, structural re-parameterization technology losslessly merges these branches into a single-layer standard convolution, achieving performance improvement with zero additional inference overhead. Comprehensive experiments on the insulator dataset demonstrate that DDE-YOLO delivers exceptional detection performance, achieving a precision of 93.67%, a recall of 89.75%, an mAP50 of 93.86%, and an mAP50:95 of 65.01%. Compared with the baseline YOLO11n, the mAP50 and mAP50:95 increase by 4.74 and 5.73 percentage points, respectively. Meanwhile, the total parameter count is only 2.35 M, representing a decrease of approximately 9.3% from the baseline, with computational complexity controlled at 7.0 GFLOPs. Ablation studies confirm that the C3K2-DIMB module improves localization quality while reducing parameters by 6.6%, the DAFPN significantly increases the AP of broken and flashover defects by 4.29% and 3.47%, respectively, and the ESDH alone reduces the parameter count by 12.4%. Moreover, generalization validation on the public Unmanned Aerial Vehicle Power Line Insulator Dataset (UPID) yields an mAP50 of 87.64%, a 1.62% improvement over the baseline. On the cross-domain NEU-DET steel surface defect dataset, the mAP50 increases by 3.31% to 79.27%, fully proving its outstanding cross-domain generalization capability. The proposed algorithm substantially enhances the detection precision of minute breakages and flashovers while effectively pruning parameter redundancy and reducing computational complexity, exhibiting excellent robustness.

To address the challenges of existing task chain deduction methods in capturing deep semantics and long-range dependencies under complex environments, which lead to insufficient accuracy and stability, this paper proposes a knowledge graph-enhanced large language model for combat task chain deduction method to improve the accuracy and stability of task sequence prediction in complex battlefield scenarios. The method first employs a bidirectional encoder-based entity linking module to accurately link combat elements in task texts with external general knowledge graphs for disambiguation, obtaining semantic descriptions of the elements. A task composition graph is then constructed, and TransE graph representation learning is applied to capture complex relationships among elements, yielding knowledge-augmented vector representations. Subsequently, a self-attention mechanism integrates the task chain context with candidate task sets to generate knowledge-level task state representations, which are concatenated with chain-level temporal representations to form a comprehensive state infused with external knowledge. Finally, task chain deduction is formulated as a reinforcement learning problem, and a multi-level reward mechanism combining chain-level rewards (BLEU) and knowledge-level rewards (cosine similarity) is designed to guide multi-step deduction. Experimental results on four public datasets (NYT, Finance, ICEWS14, and ICEWS15) show that the proposed method achieves one-step prediction accuracies of 76.23%, 74.39%, 72.11%, and 70.09%, respectively, outperforming state-of-the-art baselines by an average of approximately 19 percentage points. For four-step task chain prediction, the accuracies reach 71.76%, 70.39%, 68.97%, and 69.35%, significantly surpassing all baseline models. Furthermore, three ranking metrics—MRR, Hits@3, and Hits@5—are introduced for multi-dimensional evaluation, and the proposed method is compared with recent representative approaches including SPARK, ThinkTank-ME, and the zero-shot reasoning baseline of the DeepSeek-V4-Flash large language model. The proposed method achieves optimal or near-optimal results across all four metrics, with MRR reaching 78.64%, 73.15%, and 72.83% on NYT, ICEWS14, and ICEWS15, respectively. Ablation studies confirm the critical roles of the knowledge fusion mechanism and reward design in performance improvement. This study validates the effectiveness of knowledge graphs in enhancing semantic understanding and deduction performance, providing technical support for intelligent combat command and decision-making systems.

Federated Prototype Learning (FPL) significantly reduces communication overhead but faces critical security risks in open networks, including identity forgery and poisoning attacks. Existing statistical defenses struggle to distinguish malicious perturbations from benign heterogeneous features under non-independent and identically distributed (Non-IID) settings, often wrongfully eliminating honest long-tail clients. To address these vulnerabilities, this paper proposes a lattice trapdoor-based controllable identity tracing scheme for FPL, termed Federated Lattice Hidden Trapdoor (FedLHT). Shifting from probabilistic statistical detection to deterministic mathematical computation, FedLHT establishes a lightweight framework for pre-training anonymity, real-time authentication, and post-attack accountability. It constructs a certificateless signature system via lattice cryptography, utilizing preimage sampling to bind client identities with private keys. Traceable tags and rejection sampling guarantee information-theoretic anonymity. During aggregation, nonlinear group operations are linearized over lattices for lightweight batch verification. For targeted tracing, regulators utilize the lattice trapdoor master basis to solve the bounded distance decoding problem, accurately recovering malicious identities from abnormal noise. FedLHT's security is rigorously reduced to average-case Learning With Errors (LWE) and Short Integer Solution (SIS) problems, ensuring post-quantum security. Furthermore, utilizing the linear additive homomorphism of prototypes, the server seamlessly erases malicious historical contributions via lightweight vector subtraction for rapid model self-recovery. Compared with state-of-the-art defenses, such as ShieldFL (Mitigating Poisoning Attacks in Privacy-Preserving Federated Learning) and FLTrust (Byzantine-robust Federated Learning via Trust Bootstrapping), the proposed mechanism maintains a global model accuracy above 84.65% on the CIFAR-10 dataset. Furthermore, in a severe Non-IID setting with a heterogeneity coefficient of 1.0, it achieves a 98.75% retention rate for honest long-tail data, realizing zero-misjudgment response. For 40 concurrent clients, local computing and server-side authentication overheads are merely 1.22 ms and 6.42 ms, effectively overcoming computational bottlenecks of traditional pairing-based schemes for resource-constrained Internet of Things (IoT) scenarios.

As core carriers of human civilization and Red (revolutionary) culture, historical documents harbor densely packed spatiotemporal clues and intricately interwoven entity associations within their texts. However, conventional knowledge-organization methods in the digital humanities have long been constrained by bottlenecks such as the fragmentation of multi-source historical materials, the heterogeneity of narrative styles, and referential ambiguity among persons, places, times, and events. When general-purpose large language models (LLMs) are applied directly to historical text processing, the absence of domain-specific prior knowledge readily leads to inaccurate identification of entity boundaries. At the same time, the pervasive phenomena of homonymous entities (identical names denoting distinct entities) and synonymous entities (distinct names denoting an identical entity) cause traditional entity-alignment algorithms to confront a cost crisis in which computational complexity grows quadratically with scale during the comparison of massive numbers of graph nodes. Moreover, conventional retrieval-augmented generation (RAG) approaches overlook the implicit spatiotemporal topological constraints among texts and therefore frequently produce logical distortions when addressing narrative-style question answering, such as cross-period trajectory reconstruction and complex spatiotemporal transitions. In view of these challenges, this paper proposes a spatiotemporal narrative enhancement framework for historical documents that integrates large language models with knowledge graphs, aiming to achieve efficient structured extraction, high-precision association fusion, and deep spatiotemporal reasoning-based question answering over large-scale historical texts. The proposed framework adopts a hierarchical, collaborative design comprising three progressive core modules: knowledge extraction, entity alignment, and spatiotemporal narrative intelligent question answering. In the knowledge-extraction stage, in light of the dense proper nouns and idiosyncratic grammar characteristic of historical texts, a history-constrained prompt-chain framework is designed that decomposes the single extraction task into three cascaded stages—step-by-step reasoning prompting, domain prior-knowledge injection, and structured-output constraint correction. By introducing military terminology, historical place names, and a Ganzhi (sexagenary-cycle) chronology reference table into the context stream, the framework achieves precise, high-volume conversion of unstructured text into graph-structured data. In the entity-alignment stage, to eliminate the computational bottleneck in cross-source, multi-graph merging, a two-stage entity-alignment algorithm is proposed. The first stage introduces an index-optimized filtering mechanism that constructs a four-dimensional hybrid index covering textual aliases, life cycles, geographical-hierarchy mappings, and social topological relations; by means of a bucketing strategy, it converges the search space of nodes awaiting alignment to single digits, thereby compressing the overall time complexity of the algorithm from quadratic to linear. The second stage performs LLM-enhanced semantic discrimination, fusing the static and dynamic attribute features of candidate entities with multi-hop association networks to compute a composite similarity, and establishes a tiered, human-machine collaborative closed-loop decision process via a confidence-evaluation mechanism. In the question-answering stage, a spatiotemporal narrative graph RAG module is designed; through instruction fine-tuning or semantic parsing, natural-language queries are mapped into structured constraint conditions with explicit spatiotemporal boundaries, and a graph-database query language is dynamically generated to retrieve multi-hop associated spatiotemporal subgraphs from the global knowledge graph. The temporal evolution features and spatial topological structure are thereby conveyed in full to the generative model, driving it to produce spatiotemporal narrative text with rigorous logical chains and high interpretability. To validate the robustness and effectiveness of the proposed framework in practical settings, multi-dimensional empirical evaluations were conducted on the self-constructed Red historical-document dataset N4A-Hist, the publicly available historical dataset CHisIEC, and the multilingual general-purpose dataset DBP15K. Experimental results show that, on the knowledge-extraction task for unstructured text, the prompt-chain framework attains F1 scores of 94.8% and 92.1% on N4A-Hist and CHisIEC, respectively—improvements of 16.3% and 14.0% over the baseline of direct extraction by the base model—while ablation experiments confirm the decisive role of stepwise constraints in mitigating format hallucination. In cross-source entity-alignment testing, the two-stage algorithm achieves a Hits@1 of 91.3% and an MRR of 0.962 on N4A-Hist, with a total online runtime of only 0.8 hours; compared with the state-of-the-art LLM-based alignment methods ChatEA and LLM4EA and the traditional method EasyEA, it reduces runtime cost by more than 75% while maintaining comparable or higher accuracy. On the general-purpose datasetDBP15K, it likewise sustains a Hits@1 of 98.3% with a rapid convergence time of 4.2 hours, fully corroborating the superiority of the four-dimensional index in curbing the expansion of computational scale. In a question-answering test comprising 530 complex reasoning instructions, the spatiotemporal narrative graph RAG module raises answer accuracy on spatial-displacement questions and cross-period comprehensive-reasoning questions by substantial margins of 47% and 54%, respectively, relative to traditional vector-database retrieval. In summary, this study demonstrates that organically integrating the deep semantic representation capabilities of large language models with the rigorous topological structural knowledge of knowledge graphs constitutes an effective means of overcoming the bottlenecks of high ambiguity and strong spatiotemporal correlation in digital-humanities research on historical documents. While safeguarding the efficiency of automated processing workflows, the proposed framework mitigates the uncertainty risks of practical deployment through confidence filtering and multi-dimensional constraint mechanisms, thereby providing a highly feasible and academically valuable system-level solution for the digital reconstruction, deep semantic association, and intelligent knowledge services of massive, heterogeneous historical-document resources.

Small target detection in low-altitude UAV imagery is challenged by small object scales, dense spatial distributions, strong background interference, and complex imaging conditions. Compared with general objects, small targets occupy only a limited number of pixels, making their edges, textures, and shape cues prone to being overwhelmed by background responses during feature extraction. Existing detection methods usually obtain high-level semantic features through stacked convolutions and progressive downsampling. However, this process rapidly reduces the spatial resolution of feature maps, causing the local structures and pixel-level details of small targets to be weakened or even lost. Although feature pyramids and attention mechanisms can alleviate scale variation and background interference to some extent, most methods focus mainly on multi-scale semantic fusion and provide limited direct constraints on the fine-grained spatial information loss caused by downsampling. Meanwhile, insufficient collaborative modeling among channel features makes it difficult to integrate small-target cues scattered across different channels into stable and discriminative representations. To address these problems, this paper proposes RFINet, a small target detection network based on recalibrated fine-grained downsampling and inter-group collaborative modeling. First, a recalibrated fine-grained downsampling (RFD) module is constructed, which replaces conventional downsampling with spatial slicing and channel reorganization. By mapping local spatial information into the channel dimension, RFD preserves fine-grained structures of small targets while reducing the feature map size. To suppress redundant responses and background noise introduced by channel expansion, RFD further introduces a cross-channel adaptive recalibration mechanism that reallocates weights according to aggregated channel responses, thereby enhancing target-related channels and suppressing irrelevant background information. Compared with stride convolution or pooling, RFD does not simply discard spatial sampling points; instead, it transfers information that would otherwise be compressed into the channel domain through structured rearrangement, providing more complete local context for subsequent feature extraction. Second, an inter-group collaborative modeling (ICM) module is designed to divide input features into multiple channel subspaces, extract locally salient information within each group, and promote intra-group enhancement and inter-group interaction through weight redistribution, residual connection, and channel concatenation. This design preserves the fine-grained local modeling ability of grouped features while alleviating information isolation caused by fixed grouping, enabling the network to integrate small-target cues distributed across different channels more effectively. Finally, object classification and bounding box regression are performed by the detection head to achieve small target detection in complex low-altitude scenes. To verify the effectiveness of the proposed modules, experiments are conducted from multiple perspectives, including different input resolutions, ablation studies, model complexity, detection accuracy, and cross-dataset generalization, so as to evaluate the overall performance of RFINet under dense small-target distributions, significant scale variation, and strong background interference. Experimental results on the VisDrone2019 dataset show that RFD effectively preserves fine-grained spatial information of small targets and suppresses redundant background noise, while ICM reduces redundant background responses and strengthens the model’s focus on key feature regions. The combination of the two modules brings stable performance gains under different input resolutions. RFINet achieves mAP@0.5 scores of 40.8% and 47.5% with input sizes of 640×640 and 960×960, respectively, outperforming the compared mainstream two-stage detectors, one-stage detectors, and lightweight detection models. The results demonstrate that the proposed method enhances the representation of small targets in dense, occluded, and cluttered scenes while maintaining high detection accuracy, providing a more targeted feature modeling scheme for low-altitude UAV visual perception tasks.

Existing point cloud completion networks mainly rely on the spatial coordinates and local neighborhood features of the single-modal point cloud for shape reasoning. However, for some point clouds, there are usually problems such as sparse point distribution and insufficient global contour information. These results in the appearance of missing details and discontinuous topological structures in the completed structure. Previous cross-modal point cloud completion methods strictly rely on paired input multimodal data. The collection and alignment of multimodal data also increase the cost of dataset construction. This limits its direct application in real engineering projects. To address these issues, this paper proposes a network named Dual-State Guided Cross-Modal Point Completion Network (DSG-Net). It aims to make up for the inherent limitations of single-modal features. DSG-Net only uses the partial point cloud input as the original data. During the data preprocessing stage, it generates the corresponding depth map through the preset pinhole camera model for projection and rendering to assist in the completion of the single-modal point cloud. Thus, it overcomes the dependence of traditional cross-modal methods on the strict pairing of multimodal data. The designed Cross-Modal Fusion Generation (CMFG) module utilizes convolution to map the extracted point cloud features and depth map features to the same feature space for concatenation and fusion. Then, the multi-head self-attention mechanism models long-range dependencies in the fused features. And the modeled complementary features are split point by point to generate the initial coarse point cloud. Therefinement stage introduces the Dual-State Guided Refinement (DSGR) module. It comprises two parallel processing channels: structural analysis and similarity comparison. The structural analysis channel measures the spatial correlation between the generated point cloud and the input partial point cloud. This guides the network to focus on structural differences in the missing regions and enhances the continuity of the overall contour and local structures. The similarity comparison channel focuses on mining potential similarity patterns among different local regions. It guides topological refinement in the missing regions, thereby improving the shape consistency and detail completeness of the completed point cloud. The State Flow Block (SFBlock) is integrated within the DSGR module. It enables dynamic feature-state updating and effective information propagation during the iterative refinement process. This module adaptively controls the retention ratios of historical accumulated states and current input features via a dual-gating unit. This mechanism dynamically adjusts local details and positional deviations during the refinement process. It also prevents the recovered shape information from being weakened or redundantly propagated across multiple iterations. To verify the effectiveness of DSG-Net, experiments are conducted on two public datasets, PCN and ShapeNet55, and a self-built Terracotta Warrior dataset. The experimental results show that DSG-Net achieves an average CD-L1 of 6.72 on the public PCN dataset. This corresponds to a relative reduction of 19.81% compared with PoinTr’s average CD-L1 of 8.38. Meanwhile, DSG-Net outperforms the latest network, 3DMambaComplete, in terms of CD-L1 across all eight categories. These results further demonstrate that DSG-Net can effectively improve the structural completion accuracy of missing regions. On the ShapeNet55 dataset, DSG-Net achieves an average CD-L2 of 0.89 across three tasks with different difficulty levels. Meanwhile, among the five common categories, DSG-Net obtains the best completion performance in three of them. On the self-built Terracotta Warriors dataset, DSG-Net achieves an average CD-L1 of 5.23. This corresponds to a relative reduction of 10.45% compared with PoinTr’s average CD-L1 of 5.84. This indicates that DSG-Net has strong generalization ability on non-public data. In summary, DSG-Net achieves favorable completion performance on both public and self-built datasets. This indicates that it can not only effectively restore the shape structure of complex objects, but also exhibit strong cross-dataset adaptability. Therefore, it can provide valuable technical reference for future cultural heritage digital modeling and incomplete artifact restoration.

Aiming at the problems that electric load series are highly non-stationary, exhibit significant multi-scale fluctuations, and that periodic information is difficult to model stably and is prone to attenuation and phase shift in multi-step forecasting, as well as the complex coupling relationships between exogenous variables such as temperature, humidity, electricity price and load, this paper proposes a short-term load forecasting method based on time-frequency fusion and periodic residual coupling (TFF-PRC). The proposed method aims to improve the joint representation capability of the model for long-term trends, local fluctuations and periodic patterns of load series, while enhancing the effective utilization of exogenous information in multivariate scenarios, thereby improving forecasting accuracy and stability under different prediction horizons. Existing deep learning methods mainly rely on time-domain modeling or global dependency modeling, and often fail to fully exploit low-frequency trends, high-frequency disturbances and explicit periodic information. Meanwhile, in multivariate modeling, premature mixing of exogenous variables may introduce redundant noise, whereas complete channel independence may weaken effective interactions. Therefore, this paper constructs the forecasting framework from three aspects: time-frequency collaborative modeling, multi-period residual compensation and load-centered channel interaction.First, in the time domain, moving average decomposition is adopted to extract the trend and seasonal components. In the frequency domain, multi-level discrete wavelet transform (DWT) is employed to obtain the low-frequency approximation component and multi-scale high-frequency detail components. Through a directional fusion mechanism, the low-frequency information is used to enhance the trend component, while the high-frequency information is used to enhance the seasonal component. In this way, frequency-domain information participates in the construction of time-domain components in a residual manner, thereby strengthening the model’s ability to characterize long-term trends and local disturbances. Subsequently, the enhanced trend and seasonal components are separately fed into a channel-independent temporal patch feature extraction network. Convolution is used to capture local patterns, and multi-layer LSTM is employed to model long- and short-term dependencies, so as to obtain stable and clean high-dimensional representations.To address periodic attenuation and phase shift in multi-step forecasting, the proposed method identifies the dominant periods of load series based on fast Fourier transform (FFT), extracts historical load segments that are phase-aligned with the prediction interval, and generates periodic compensation features through sample-level adaptive weighting. These features are then incorporated into the backbone forecasting results in a residual form. Furthermore, the load-centered channel interaction module takes load as the query center and selectively aggregates information from exogenous variables such as temperature, humidity and electricity price through the query-key-value mechanism. This design preserves the multivariate modeling capability while effectively suppressing noise interference among channels.Experiments are conducted on two public electric load datasets, namely New South Wales (NSW), Australia, and Spain. The input window is set to 48 time steps, and the forecasting horizons are set to 1, 2, 24 and 48 time steps. Compared with mainstream baseline models, TFF-PRC reduces the average MAE by 11.03%~20.11% and the average RMSE by 11.41%~18.53% on the NSW dataset. On the Spain dataset, the average MAE is reduced by 3.99%~5.23%, and the average RMSE is reduced by 4.05%~4.10%, indicating that the proposed model has good generalization capability. Ablation experiments show that after removing the TFF, PRC or LCA module, MAE increases by 7.54%, 10.63% and 7.00%, respectively, while RMSE increases by 7.03%, 8.18% and 8.10%, respectively, verifying the contribution of each module to forecasting performance.In summary, TFF-PRC effectively alleviates non-stationary fluctuations, periodic attenuation and multivariate noise interference through time-frequency directional fusion, multi-period residual coupling and load-centered channel interaction. The proposed method achieves high accuracy and stability across different datasets and forecasting horizons.

Existing vehicle-infrastructure cooperative perception task scheduling methods face two major challenges: insufficient characterization of computational resource hard constraints, and the difficulty of meeting scheduling decision efficiency in dynamic environments. Furthermore, during the task scheduling process, bidirectional data transmission exists between Connected Autonomous Vehicles (CAVs) and Roadside Units (RSUs), where RSUs send perception data to CAVs, and CAVs offload tasks to RSUs and receive the results. In this process, the RSU-side computational delay in the task completion delay model is inversely proportional to the allocated computational resources, which introduces fractional nonlinear terms into the delay model and further increases the mathematical complexity of the problem. To address the above issues, this paper considers the scenario where both RSUs and CAVs are subject to computational resource upper bound constraints, and constructs a vehicle-infrastructure cooperative perception task scheduling model aimed at minimizing the total system delay. At the communication level, the model adopts Orthogonal Frequency Division Multiple Access (OFDMA) as the multiple access scheme, where the RSU allocates orthogonal subcarriers to each vehicle to ensure equal communication bandwidth between CAVs and the RSU. Meanwhile, considering the Doppler shift and Inter-Carrier Interference (ICI) caused by high-speed vehicle mobility, a quantitative relationship between vehicle speed and data transmission rate is established. At the computational resource allocation level, the model explicitly characterizes four types of constraints: the RSU total computational resource upper bound, the per-task computational resource upper bound, the RSU maximum number of parallel tasks, and the CAV onboard computational resource upper bound. From the RSU perspective, the core decision is to select which CAVs' tasks to admit and to allocate computational resources to them, thereby coupling task offloading decisions with computational resource allocation. On this basis, this paper takes minimizing the sum of task completion delays of all CAVs in the system as the objective function, takes the binary task offloading decision variable, the continuous RSU computational resource allocation variable, and the continuous CAV local computational resource allocation variable as the core optimization variables, and introduces the task completion delay as an auxiliary variable, to construct a mixed-integer nonlinear programming model with fractional nonlinear constraints. To solve the difficulty caused by nonlinear constraints, this paper proposes a customized solution algorithm based on the outer approximation method. The algorithm exploits the convex separable property of the delay function, performs a first-order Taylor expansion on the delay expression of RSU task processing to generate outer approximation cuts, and transforms the original nonlinear model into a mixed-integer linear programming problem. To address the shortcomings of the traditional outer approximation method, namely its sensitivity to initial points and insufficient convergence stability, this paper proposes an initial point selection strategy: taking the maximum computational resource allowable to a single task as the initial point, to improve the solution quality and efficiency of the model. To evaluate model performance, experiments compare the proposed model with two baseline schemes: Baseline 1: all CAVs process tasks locally without using RSU resources; Baseline 2: CAVs prioritize offloading tasks to the RSU, with tasks exceeding the RSU's capacity processed locally by CAVs, and the RSU allocates equal computational resources to each offloaded task. Experimental results demonstrate that, compared with Baseline 2, the proposed model reduces the total task completion delay by 15.6% and achieves a maximum reduction of 32.4% in the task completion delay for a single CAV. This paper also compares the proposed algorithm with the Genetic Algorithm and the Adaptive Large Neighborhood Search Algorithm. Experimental results demonstrate that, in terms of solution efficiency, the proposed initial point selection strategy enables the algorithm to converge in 2.55 rounds on average, with the solution time controlled between 0.91 and 2.86 seconds. Compared with the Genetic Algorithm, the proposed algorithm achieves an average saving of 56.2% in solution time. These results indicate that the proposed vehicle-infrastructure cooperative perception task scheduling model and method under computational resource constraints can effectively reduce the total system delay and improve the efficiency of scheduling decisions.

To address significant heterogeneity among modality representations, redundant noise introduced by non-textual modalities such as vision and audio, and the difficulty of dynamically modeling sample-specific modality contributions in multimodal sentiment analysis, this paper proposes a Text-Guided and Dual-Gated Adaptive Graph Fusion model, namely TGAGF. In video-based sentiment analysis, textual, visual, and acoustic modalities provide sentiment cues from semantic content, facial expressions, and acoustic variations, respectively. However, these three modalities differ significantly in representation formats, information density, and noise distributions. The textual modality usually has strong semantic representation capability and can more directly reflect the speaker’s opinions and sentiment orientation. Visual and acoustic modalities can provide complementary nonverbal information, such as facial expressions, head poses, intonation, and speaking rate, but they are also susceptible to individual differences, environmental interference, and acquisition quality. If multimodal features are directly fused, weakly relevant information and redundant noise from non-textual modalities may be introduced into the joint representation and further propagated or amplified during cross-modal interaction, thereby weakening the model’s discriminative ability for sentiment polarity, fine-grained sentiment levels, and continuous sentiment intensity. To this end, TGAGF takes the textual modality as semantic guidance to enhance the representation capability of effective sentiment cues in visual and acoustic modalities during cross-modal fusion, while suppressing the influence of redundant noise on sentiment prediction. First, the model performs feature extraction and dimensional alignment on textual, visual, and acoustic modalities, mapping features from different modalities into a unified semantic space to provide structurally consistent input representations for subsequent fusion. Then, a shared-specific dual-branch disentanglement network is introduced to separate cross-modal consistent information from modality-specific information. Specifically, the shared branch extracts sentiment-related consistent semantic information across different modalities, while the modality-specific branch preserves complementary discriminative information within each modality. Meanwhile, reconstruction consistency, modality-specific consistency, metric constraint, and orthogonality constraint are jointly employed to constrain the disentangled representations from the perspectives of semantic recovery, modality-specific stability, cross-modal consistency, and subspace independence, thereby reducing semantic entanglement between shared and specific features. Second, a text-guided dual-stage gating mechanism is adopted to suppress redundant noise in visual and acoustic modalities before graph-based relational reasoning and to reweight modality representations after graph propagation. The first-stage gate pre-selects the modality information entering the graph reasoning module, reducing the diffusion of invalid information during relational propagation. The second-stage gate readjusts the contribution of each modality according to the cross-modal interaction results after graph propagation, enabling the model to dynamically allocate modality weights according to the sentiment expression characteristics of different samples. On this basis, textual, visual, and acoustic modalities are regarded as graph nodes, and a learnable adjacency matrix is used to model inter-modal dependencies. In this way, semantic correlations among different modalities are explicitly captured at the modality level, and complementary information across modalities is enhanced through graph message propagation. Finally, a cascaded multi-task prediction strategy is employed to jointly model sentiment polarity, fine-grained sentiment levels, and continuous sentiment intensity, allowing the model to simultaneously perform sentiment direction discrimination, sentiment level classification, and continuous sentiment intensity estimation. Experimental results on the CMU-MOSI and CMU-MOSEI datasets show that the proposed model achieves competitive performance on multiple evaluation metrics. Specifically, the Acc7, Acc5, Acc2, and F1 scores on CMU-MOSI are 48.4%, 56.1%, 85.3%, and 85.3%, respectively; the Acc2, F1, and MAE on CMU-MOSEI are 85.2%, 85.1%, and 0.534, respectively. Compared with various classical methods and recent representative methods, TGAGF demonstrates strong competitiveness in both sentiment classification and sentiment intensity prediction tasks. The experimental results verify the effectiveness of the proposed method in alleviating modality heterogeneity, suppressing redundant noise interference, dynamically modeling modality contributions, and improving the stability of multimodal sentiment recognition.

Trusted multimodal classification, which integrates heterogeneous data streams and quantifies predictive uncertainty, is of great significance in high risk domains such as medical diagnosis and autonomous driving. However, when handling multimodal conflicts, most existing methods still suffer from three core issues: mechanical noise processing leads to overconfidence when evidence is scarce; the unified fusion paradigm causes modality specific deep semantics to be lost due to “feature assimilation”; and uncertainty estimates are not utilized as a dynamic feedback signal to guide fusion decisions in a closed loop manner. To address these challenges, this paper proposes a Joint Adversarial like Learning and Dynamic Fusion (JALDF) method for trustworthy multimodal classification. The method aims to systematically improve the accuracy, noise robustness, and trustworthiness of multimodal classification through the synergistic optimization of feature alignment, semantic preservation, and adaptive fusion. In terms of technical implementation, JALDF consists of three core innovative modules. First, the cross modal class similarity module applies sigmoid sparsification and L₁ regularization to each modality’s features to select highly discriminative subsets, then extracts samples by class and uses the Wasserstein distance to measure the distribution discrepancy of the same class across different modalities. By minimizing this distance, the network is forced to learn a shared discriminative representation across modalities. This module leverages optimal transport theory to precisely characterize the minimal transport cost between distributions. Compared with symmetric measures such as KL divergence, the Wasserstein distance provides stable gradient guidance even when there is a large distribution shift in the feature space, thus ensuring effective cross modal alignment. Second, to prevent excessive modal convergence caused by the above constraint, the semantic adversarial module treats each modality as an independent domain, assigns modality origin labels to samples, and trains a domain classifier. This classifier forms an adversarial game with the similarity module: the similarity module strives to confuse modal differences, while the domain classifier tries to distinguish the source of each modality. As a result, while keeping inter class distributions close, the model precisely preserves the unique deep information of each modality. This adversarial design is introduced for the first time in the field of trustworthy multimodal classification and effectively resolves the conflict between feature collapse and information loss.Finally, the confidence supervised self attention fusion module maps each modality’s features into queries, keys, and values, computes inter modal dependency weights via multi head self attention, and uses the True Class Probability (TCP) as a trustworthy supervision signal to train a confidence estimation branch, enabling the model to dynamically evaluate the evidence quality of each modality during testing without requiring labels. The fusion loss combines classification cross entropy and confidence mean square error, driving the attention weights to be adjusted in real time according to current confidence, thereby enhancing high quality modalities and suppressing low quality or noisy ones. This closed loop supervision mechanism allows uncertainty measures to directly participate in the generation of fusion weights, breaking the limitations of traditional static fusion. Experiments on five public multimodal datasets (Cub, Caltech101, Scene15, Animal, and PIE) demonstrate that JALDF achieves leading accuracy results across all datasets. Specifically, it attains 81.40% on Scene15, outperforming the suboptimal RMVC by 8.35 percentage points, and reaches 95.14% on PIE, significantly surpassing existing methods. Additional experiments on hyperparameter analysis, conflict sensitivity, uncertainty estimation, ablation, and visualization further validate the effectiveness and superiority of JALDF from multiple perspectives. In summary, through the organic synergy of adversarial like learning and dynamic fusion, JALDF effectively solves the three major problems of overconfidence, feature assimilation, and rigid fusion in multimodal classification, offering a highly accurate and robust technical solution for trustworthy multimodal classification.

Cognitive diagnosis ( CD ) aims to accurately identify students ' mastery of different knowledge points based on their response records in online teaching platforms. It is an important basis for personalized learning recommendation, learning path planning and teaching intervention. However, the real education data generally have obvious long tail distribution characteristics : a small number of head students have rich answer interaction, which can provide sufficient supervision signals for the model ; however, a large number of tail students only have a small number of answering records, and their cognitive state is difficult to be stably estimated, which is prone to representational aliasing, mean regression and even representational collapse. Most of the existing cognitive diagnosis models rely on student ID embedding and limited interaction records, and the relatively complete cognitive models in head students are not fully utilized, which is difficult to effectively alleviate the cold start and sparse representation of tail students. At the same time, some models do not fully distinguish the cognitive semantic differences behind the correct and wrong answers, resulting in insufficient mining of discriminative information in sparse interaction. In view of the above problems, this paper proposes a student cognitive diagnosis framework CFECD based on conditional manifold enhancement and positive and negative two-view contrastive learning. In the design of the method, CFECD first introduces the semantic enhancement feature initialization module, which integrates the semantics of the exercise text, knowledge point information and ID embedding to improve the generalization ability of node representation in sparse scenes. Then, LightGCN is used to encode the high-order interaction between students, exercises and knowledge points to obtain structured cognitive representations. On this basis, the conditional manifold enhancement module is constructed to store the high-quality cognitive representation of the head students through the global memory library, and the reverse projection, latent space interpolation and forward generation of the tail students in the continuous probability space are carried out by using the reversible mapping ability of the manifold to realize the distribution completion and feature calibration of the sparse representation of the tail students. In order to further improve the reliability of the enhanced features, the model combines the original graph features, manifold generation features, peer reference features and context interaction features to form a multi-source collaborative tail student representation. At the same time, aiming at the different cognitive meanings reflected by the correct and wrong behaviors of the answers, this paper constructs a positive and negative dual-view contrastive learning mechanism, decouples the interaction graph into a correct-response mastery view and an incorrect-response deficiency view, and enhances the separability of different cognitive states in the hidden space by comparing the constraints. Finally, the model designs a diagnostic prediction function with scaling factor to moderately enlarge the marginal difference between students ' ability and the difficulty of exercises, and combines the adaptive loss weight to generate the network dynamic balance prediction loss, sorting loss, consistency loss and contrast learning loss to improve the stability of the joint optimization process. The experimental results on the three public education datasets of ASSIST2009, Junyi and XES3G5M show that CFECD is superior to various baseline models in terms of prediction accuracy and ranking ability, especially for tail students. The ablation experiment and parameter sensitivity analysis further show that the conditional manifold enhancement can effectively complement the tail student representation, the positive and negative two-view contrast learning can improve the semantic discrimination of the hidden space, and the moderate manifold constraint, network depth, contrast learning weight and scaling factor help to achieve a balance between representation enhancement and main task prediction. In summary, CFECD can generate more stable and interpretable portraits of students ' knowledge mastery in the long-tail online teaching scene, especially for the early diagnosis of low-active students, newly registered students and students with insufficient interactive records. The model output can further serve the identification of weak knowledge points, personalized exercise recommendation, learning path planning, teacher teaching early warning and class learning situation analysis. The experimental results verify the effectiveness, robustness and practical teaching application potential of the proposed method in the cognitive diagnosis task of long tail students.

Low-light image enhancement aims to recover clear, natural, and detail-preserving images from scenes with insufficient illumination, low contrast, noise degradation, and color distortion. Existing methods can improve image brightness to some extent, but they often suffer from noise amplification, unstable color restoration, texture loss, and over-enhancement in complex dark scenes. This is mainly because noise and textures may show similar high-frequency responses in the spatial domain, while luminance and chrominance information are highly coupled in the RGB space. To address these problems, this paper proposes SFSNet, a spatial-frequency synergistic network based on color space decomposition. The input image is first transformed into the HVI color space, where the intensity and chrominance components are separated to reduce the mutual interference between brightness recovery and color correction. Before spatial encoding, a frequency denoising module is introduced to suppress noise-related frequency disturbances and provide cleaner features for subsequent restoration. Then, a dual-branch backbone is constructed to process intensity and chrominance features separately, while a multi-scale convolution module is used to capture local textures, directional structures, and contextual information. To improve the consistency between brightness and color restoration, a cross-branch feature collaboration mechanism is designed to promote information interaction between the two branches. In addition, a deep interaction module and an uncertainty-based adaptive weighting strategy are employed to enhance deep feature representation and improve training stability. Experiments are conducted on LOL-v1, LOL-v2-real, and LOL-v2-syn, which provide paired normal-light reference images, and further evaluated on real unpaired datasets including DICM, LIME, MEF, and VV to verify the generalization ability of the proposed method. Compared with representative low-light enhancement methods, the proposed method can better improve dark-region brightness, suppress noise amplification, maintain natural color appearance, and preserve local structural details. Ablation studies further demonstrate the effectiveness of frequency denoising, multi-scale spatial modeling, cross-branch feature collaboration, and deep interaction supervision.

Vision-Language Models (VLMs) have demonstrated exceptional capabilities in cross-modal understanding by aligning Large Language Models (LLMs) with visual encoders. However, cross-modal hallucination remains a critical bottleneck that limits their deployment in accuracy-sensitive scenarios. The fundamental cause lies in the fact that the underlying LLM base, pre-trained on massive text corpora, establishes strong language priors that often override authentic visual features during autoregressive decoding. A detailed analysis of the internal attention mechanism during decoding reveals significant hierarchical disparities in visual feature evolution: shallow layers exhibit highly diffuse attention containing substantial redundant noise, which easily triggers attention drift, while deep layers suffer from severe attention decay due to language dominance, effectively severing the link between text generation and visual facts. Existing decoding-time intervention methods typically adopt global, coarse-grained strategies, failing to systematically address the contradiction between shallow-layer diffusion and deep-layer decay. Consequently, this paper proposes a Multi-layer Cumulative Visual Prior (MCVP) method, which reconstructs feature distributions at the decoding stage through three synergistic mechanisms. First, a sparsified focusing strategy is introduced to dynamically evaluate shallow-layer responses and filter background noise to extract high-confidence features, fundamentally preventing attention drift. Second, a layer-wise recursive accumulation mechanism is designed to construct a temporally robust global visual prior by applying weighted aggregation to preceding shallow-layer sparse features via a sliding window. Finally, an adaptive soft injection strategy is proposed to calculate the real-time visual dependency of the current generation step, dynamically incorporating the robust prior into deep-layer cross-attention modules to rebalance modal weights and enforce strict alignment with visual facts. Extensive empirical evaluations on LLaVA-1.5, MiniGPT4, and DeepseekVL, based on CHAIR and POPE benchmarks, demonstrate that MCVP significantly outperforms baseline sampling and state-of-the-art intervention methods. In the CHAIR benchmark, MCVP achieves a relative improvement of 18.8% in hallucination suppression performance; specifically, it reduces the sentence-level hallucination rate of LLaVA-1.5 from 48.4% to 38.2% while maintaining a high recall of 77.7%, successfully overcoming the tendency of traditional exploratory strategies to over-fit language probabilities and achieving an optimal balance between precision and coverage. In POPE testing, MCVP achieves a state-of-the-art F1 score of 88.45% on MiniGPT4 and effectively corrects the over-prediction issues caused by shallow-layer diffusion in DeepseekVL. Comprehensive ablation and visualization analyses confirm that these mechanisms effectively prevent deep-layer feature decay and refocus attention onto target regions. In summary, MCVP accurately captures the evolutionary patterns of hierarchical attention and provides an efficient, training-free approach to suppressing cross-modal hallucinations without compromising general reasoning capabilities, establishing a new theoretical and practical paradigm for white-box hierarchical intervention in large model optimization.

Brain tumor MRI images usually exhibit large variations in tumor scale, strong morphological heterogeneity, and blurred lesion boundaries that are often intertwined with normal brain tissues. Existing detection models still suffer from insufficient local detail extraction, global semantic modeling, and cross-scale feature fusion, which may lead to inaccurate boundary localization, increased missed and false detections, and high deployment costs. To address these issues, this paper proposes GDC-DETR, a high-accuracy and lightweight brain tumor detection model based on RTDETR-r18. The proposed model aims to improve the detection accuracy of complex lesions while reducing the number of parameters and computational cost, making it more suitable for clinical computer-aided diagnosis and lightweight deployment. GDC-DETR is improved from three aspects: backbone enhancement, feature fusion, and efficient feature interaction. First, a gated enhancement module with contextual awareness, named GLCA-CG, is designed. It extracts edge and texture details and long-range contextual information through local and global dual branches, respectively, and adaptively selects discriminative features using a convolutional gating mechanism. This enhances the model’s ability to distinguish small lesions, weak boundaries, and complex backgrounds. Second, a hierarchical local-global attention fusion module, named HLGA-FM, is constructed. By employing parallel patch-aware attention at local and global scales, this module achieves semantic alignment and adaptive fusion between high-level and low-level features, alleviating information dilution, channel redundancy, and semantic deviation caused by traditional feature concatenation, thereby improving cross-scale lesion localization. Finally, a low-resolution lightweight attention-based intra-scale feature interaction module, named LR-AIFI, is proposed. High-resolution features are first mapped into a low-resolution space through anti-aliasing downsampling for attention interaction, and then spatial details are compensated through upsampling and depthwise convolution refinement. As a result, the computational complexity of the main attention term is reduced to approximately 1/16 of the original strategy, achieving a better balance between global modeling capability and computational efficiency. Experiments are conducted on a public brain tumor MRI dataset containing three types of tumors: glioma, meningioma, and pituitary tumor. Under unified training parameters, evaluation metrics, and experimental environments, GDC-DETR is compared with RTDETR series models, YOLO series models, and several brain tumor detection methods. The experimental results show that GDC-DETR achieves an mAP50 of 92.9%, which is 2.7 percentage points higher than the RTDETR-r18 baseline. The precision and recall reach 91.4% and 91.1%, respectively. Meanwhile, the number of parameters is reduced to 15.6 M, representing a 21.6% decrease compared with the baseline, and the computational cost is 47.8 GFLOPs, maintaining a relatively low computational overhead among the mainstream comparison models. Visualization results indicate that, for challenging samples with blurred boundaries, irregular shapes, and high interference between lesions and surrounding tissues, GDC-DETR can more accurately focus on the tumor core region, reduce invalid responses in background areas, and obtain more stable bounding-box localization results. Ablation experiments demonstrate the effectiveness of each proposed module. Introducing GLCA-CG alone improves mAP50 by 1.2 percentage points and significantly reduces the parameter scale. Combining HLGA-FM with GLCA-CG further improves mAP50 to 92.4%. LR-AIFI reduces the computational burden while maintaining stable detection accuracy. The best overall performance is achieved when the three modules are jointly integrated, indicating that detail enhancement, multi-scale semantic fusion, and low-resolution efficient interaction are complementary to each other. Generalization experiments further show that the proposed model achieves mAP50 values of 94.9% and 92.9% on two external MRI datasets, respectively. In the zero-annotation transfer scenario, the mAP50 reaches 73.5%, and after fine-tuning with only 10% and 20% of target-domain annotations, it increases to 86.5% and 88.6%, respectively, demonstrating favorable cross-domain adaptability and few-shot transfer capability. In summary, GDC-DETR effectively alleviates the trade-off between high accuracy and low computational cost in brain tumor MRI detection, improves the sensitivity and localization ability for complex lesions, and reduces the dependence on hardware resources, providing a lightweight reference solution for clinical auxiliary brain tumor detection.

With the rapid development of Large Language Model (LLM), LLM has shown promising application prospects in text classification tasks. Existing text classification methods based on prompt engineering heavily rely on the generative capability of LLM, and often suffer from insufficient generation ability and poor reasoning stability in lightweight LLM scenarios, making it difficult to achieve satisfactory performance. In addition, different hidden layers of LLM exhibit significant differences in representation capability, and intermediate layer representations often contain richer semantic information, even outperforming the last layer outputs in certain tasks. To address these issues, a multi-level semantic fusion for text classification based on LLM is proposed. By extracting and fusing text embedding representations from different hidden layers of lightweight LLM, the proposed method achieves efficient and stable classification performance without fine-tuning LLM. Firstly, prompt engineering is introduced into the construction process of text embedding representations. By designing prompt templates suitable for text classification tasks, original sentences are inserted into specific contextual semantics to guide the model in adjusting the attention distribution when generating hidden layer representations, thereby enhancing the discriminative capability of text embedding representations. Secondly, curvature is employed to characterize the geometric smoothness of hidden layer representations. Hidden layer representations with lower curvature correspond to smoother semantic trajectories and are more suitable as text embedding representations for text classification tasks. By calculating the average curvature of each hidden layer, several hidden layer representations with lower curvature are selected as feature sources for subsequent fusion. Thirdly, intra-layer position-weighted pooling and inter-layer multi-scale convolution fusion are used to obtain more expressive sentence embeddings. The intra-layer position-weighted pooling assigns different weights according to the relative positions of tokens in the input sequence, preserving important semantic information at the tail while also considering contextual information at the beginning of the sequence, thereby generating more complete and stable text embedding representations. The inter-layer multi-scale convolution fusion models relationships among different hidden layers through convolution operations to achieve the fusion of multi-level semantic information. Finally, the classifier is trained using a joint loss function, simultaneously improving classification accuracy, the learning capability for hard samples, and the discriminative of the representation space, thereby optimizing model parameters from multiple perspectives to obtain the final prediction results. Experimental results demonstrate that the proposed method achieves accuracies of 96.11%, 93.48%, 89.92%, 74.15%, and 96.33% on the SST-2, MR, 20NG, Ohsumed, and R52 datasets, respectively, outperforming existing mainstream baseline methods. In addition, Macro-F1 is introduced as a supplementary evaluation metric on the relatively imbalanced Ohsumed and R52 datasets, where the proposed method achieves 62.93% and 73.12%, respectively, demonstrating its effectiveness under complex data distribution conditions. Furthermore, ablation studies, different intra-layer pooling methods analysis, different prompt templates analysis, and different numbers of layers analysis are conducted to verify the rationality and effectiveness of each module design. Sensitivity analysis of loss weights is performed to evaluate the influence of different loss weight variations on classification performance. Different layers performance analysis is performed to explore the relationship between different hidden layer representations and classification performance. In conclusion, the proposed method can effectively extract and fuse LLM embedding representations for text classification, significantly reducing computational overhead while improving the representation quality and discriminative capability, demonstrating strong generalizability and practical applicability.

To meet the demand for sharing massive and continuous time-series data in Internet of Things (IoT) scenarios, traditional certificateless signcryption (CLSC) schemes based on bilinear pairings or elliptic curve cryptography face security limitations under the threat of quantum computing. Moreover, repeated signcryption for multiple data segments incurs considerable computational and communication overhead. To address these challenges, this paper proposes a lightweight post-quantum certificateless hybrid signcryption scheme for time-series data sharing. First, a bidirectional hash-chain-driven symmetric key generation and management mechanism is designed to enable rapid key derivation for multiple consecutive data segments while providing forward and backward security. In addition, a lightweight hybrid signcryption framework with conditional privacy protection is constructed over module lattices. By further introducing an offline/online paradigm that shifts part of the computation to the precomputation stage, the proposed scheme achieves post-quantum security while significantly reducing both the computational burden on end devices and the transmission overhead over communication links. Security analysis in the random oracle model demonstrates that the scheme satisfies IND-CCA confidentiality and EUF-CMA unforgeability. Prototype implementation and evaluation are further conducted on resource-constrained devices. Comparative experimental results show that, compared with representative schemes, the proposed scheme offers significant advantages, reducing the overall end-side latency and transmission cost to about 40% of those of comparable post-quantum schemes on average. In long-duration and continuous time-series data scenarios, these advantages become even more pronounced. The proposed scheme strikes a balance among post-quantum security, lightweight design, and scalability, making it well suited for time-series data sharing applications in next-generation intelligent IoT systems.

In scenarios related to human-agent collaboration, the explicit/legible representation and accurate recognition of agent intent serve as vital means to enhance collaborative safety and efficiency, and constitute the foundation for establishing and improving human trust in agents. The core lies in enhancing the legibility of agent behaviors to facilitate effective human understanding. Existing research primarily conveys intent by introducing differentiated behaviors that deviate from the optimal path. While effective in certain contexts, such schemes inherently sacrifice partial execution efficiency. In highly confusing scenarios, they face two critical challenges: 1) In environments characterized by overlapping and dense routes, sacrificing efficiency yields diminishing returns in intent conveyance; 2) Overlapping routes obscure the informational value of the agents' original behavioral sequences, making it difficult for observers to discern their true intentions. Addressing these issues, this study investigates key factors influencing behavioral legibility in complex human collaboration, focusing on action posture, velocity, and distance. Without altering the agents' original action sequences, this paper proposes a multi-dimensional fusion representation mechanism through the multi-dimensional modeling and integration of key state information, including posture, velocity, and distance. Specifically, we design a posture mapping function aligned with human reasoning, utilize Gaussian kernel functions to transform and map velocity and distance representations, and integrate these features via fusion parameters to enhance behavioral legibility. Furthermore, corresponding to this mechanism, we propose an efficient intent recognition scheme based on Bayesian update. By fully leveraging the fused state information, the scheme achieves rapid and accurate identification of agent intent. Consequently, the proposed method—which integrates the mechanism and the scheme—not only enhances the explicit expressive capability of overall behaviors regarding their intent but also improves task execution efficiency by reducing the number of differentiated behaviors required. The method is comprehensively validated across three experimental scenarios: unconstrained, complex obstacle-laden, and highly confusing environments. Experimental results demonstrate that, compared to baseline methods, our approach incorporates additional dimensions of state information, improving the accuracy of agent intent recognition by an average of 44.47%. Moreover, by conveying sufficient intent information through shorter behavioral sequences, the method significantly boosts task execution efficiency, reducing agent execution costs by an average of 42.86%. Ablation studies verify the positive contributions of posture, velocity, and distance to legibility enhancement, with posture identified as the most significant factor. Sensitivity analysis determines the optimal value ranges for color and velocity factor parameters, ensuring stable performance. Additionally, experiments on continuous control tasks confirm the method's applicability in complex engines and real-world engineering. On the algorithmic side, complexity analysis reveals that the inference time increases by only 2.89 milliseconds compared to the base method, satisfying real-time interaction requirements. This study provides a novel pathway for reducing redundant behaviors while enhancing legibility, balancing belief transmission and execution efficiency without disrupting original action sequences. The method can be directly embedded into various baseline approaches to achieve varying degrees of improvement in recognition accuracy. Future wor

Efficient cross-domain resource allocation is the foundation for achieving optimized resource distribution and ensuring service quality in the Space-Air-Ground Integrated Network (SAGIN). However, constrained by the inherent multi-layer highly complex topology of SAGIN and a massive number of differentiated network slicing service requests, traditional flat architectures and static mathematical optimization models struggle to achieve efficient and dynamic cross-domain resource allocation in the face of dynamic and complex problems. Therefore, this paper proposes a hierarchical cross-domain resource allocation method based on deep reinforcement learning (DLCDS), aiming to maximize system utility while ensuring the efficient utilization of underlying physical resources. Firstly, this paper designs a hierarchical collaborative cross-domain framework, establishing a dual-layer collaborative mechanism between the global domain layer and the local domain layer. Secondly, at the global domain layer, a service level agreement (SLA) decomposition algorithm based on the Twin Delayed Deep Deterministic Policy Gradient is proposed. This algorithm leverages deterministic policies within continuous action spaces to achieve the planning and decomposition of SLA constraints across subdomains. At the local domain level, a resource allocation algorithm based on Proximal Policy Optimization is designed. This algorithm constructs an SLA mapping model to achieve the conversion from service objectives to resource requirements, and employs the exploratory advantage of stochastic policies to enhance resource allocation efficiency within the domain. Finally, experimental results demonstrate that, compared to other benchmark algorithms, this method achieves an average 12% improvement in resource utilization and a 19% increase in system utility.

To address the problems of low detection accuracy and high missed detection rate caused by the small size,weak texture features,and severe background interference of unmanned aerial vehicle (UAV) targets in low-altitude complex environments,a lightweight UAV object detection method named EF-YOLO (Efficient Fusion YOLO) based on YOLOv11 is proposed. Considering the limitations of existing models in complex scenarios,including insufficient deep semantic representation, multi-scale feature fusion mismatch, and serious background interference, the YOLOv11 network architecture is systematically optimized. First, a C3K2-MSFE module is introduced into the deep layers of the backbone network. By combining grouped convolution with multi-scale convolution kernels, the module enhances the discriminative ability of weak deep target features while maintaining the stability of the main semantic flow through a linear branch, thereby improving the feature representation capability of small UAV targets in complex backgrounds. Second, an improved multi-scale feature fusion structure named PSG-FPN is proposed. A Soft Nearest Neighbor Interpolation mechanism is adopted during the upsampling stage to alleviate semantic coverage problems in cross-scale feature fusion through adaptive soft coefficients. Meanwhile,GSConvE is employed to optimize the downsampling process,and an additional P2 detection layer is introduced to enhance the perception ability for small-scale UAV targets, thereby improving multi-scale feature alignment. Furthermore, to reduce parameter redundancy and computational overhead in traditional detection heads, a lightweight shared-feature detection head named SFE-Head is designed. By adopting a shared convolutional stem and grouped convolution structure, classification and regression branches can share features,which significantly reduces model parameters and computational complexity while maintaining detection accuracy. Finally,the original C2PSA module in YOLOv11 is improved by introducing a frequency-domain enhanced feedforward structure. Through frequency-domain feature modeling, redundant texture noise in complex backgrounds can be effectively suppressed, thereby improving the response capability and anti-interference performance for weak UAV targets. Experiments are conducted on the DUT Anti-UAV and Det-Fly datasets. The results show that, on the DUT Anti-UAV dataset, EF-YOLO achieves Precision, Recall, mAP50, and mAP50-95 values of 96.76%, 88.16%, 92.59%, and 62.41%, respectively, representing improvements of 4.01%, 3.44%, 3.04%, and 6.18% compared with the original YOLOv11. Meanwhile, the number of model parameters is reduced from 2.58M to 2.27M, and the model size decreases from 5.3 MB to 4.8 MB, while maintaining favorable real-time detection performance under relatively low computational complexity. On the Det-Fly dataset, EF-YOLO achieves mAP50 and mAP50-95 values of 96.4% and 68.06%, respectively, which are 2.39% and 2.16% higher than those of YOLOv11, demonstrating excellent generalization ability and stability under different low-altitude complex scenarios. In addition, ablation experiments verify that the proposed modules provide significant collaborative benefits. Specifically, the C3K2-MSFE module effectively enhances weak deep feature representation, PSG-FPN improves cross-scale semantic alignment, SFE-Head reduces model complexity while maintaining detection accuracy, and the frequency-domain enhancement structure further suppresses complex texture noise. Compared with conventional object detection models, EF-YOLO exhibits more stable detection performance in complex backgrounds, long-distance target scenarios, and low-contrast environments, effectively reducing missed detections and false detections while preserving model lightweight characteristics. Overall, EF-YOLO achieves a favorable balance among detection accuracy, lightweight design, and real-time performance, showing strong engineering deployment potential and providing reliable technical support for low-altitude airspace supervision, edge intelligent monitoring, and UAV security applications.

This paper studies operator-level reconstruction for the structural mismatch between YOLOv11 and Ascend NPUs. The SPPF, C3K2, and C2PSA modules are optimized without changing network semantics or model scale. Three Ascend C operators are designed: the SPPF operator uses an on-chip data loop and halo cache to reduce redundant global-memory traffic in multi-stage pooling; the C3K2 operator integrates multi-core task assignment and multi-queue asynchronous pipelining to reduce fine-grained kernel launch overhead; and the C2PSA operator reconstructs attention communication through a parallel reduction-broadcast primitive. On an Ascend 910B NPU, the complete reconstruction reduces the COCO training time per epoch by 27.6%. The results show that matching Ascend on-chip memory, asynchronous queues, and multi-core synchronization mechanisms improves the training execution efficiency of key YOLOv11 modules and keeps inference performance stable.

Reinforcement Learning from Human Feedback (RLHF) is a key technology for achieving professional value alignment in large legal language models. However, existing reviews lack systematic analysis of the specific adaptation mechanisms and challenges of legal scenarios. This paper aims to fill this gap, systematically construct a panoramic view of the research and application of RLHF in legal large language models, and provide a reference for method selection in engineering practice. This article takes the six-dimensional framework of "technical foundation - algorithm evolution - task migration - evaluation system - application challenges - future path" as the main analytical thread, systematically reviewing 80 core papers from top conferences and journals such as NeurIPS, ICML, ICLR, and ACL from 2017 to 2025. Analyze the complete evolution logic of RLHF in the legal field from algorithm adaptation to evaluation system. Analysis shows that the core predicament of legal RLHF lies in the inaccurate reward modeling caused by the conflict of multi-dimensional rewards and the sparsity of expert preferences. The PPO algorithm has insufficient stability in the generation of long text laws, while new algorithms such as DPO and GRPO simplify the process but face the bottleneck of preference data quality. Further research reveals that different legal tasks vary significantly in terms of risk levels, reasoning complexity, and feedback acquisition costs. Therefore, there is no unified optimal solution for the RLHF algorithm. For structured tasks such as the generation of legal documents and summaries of regulations, DPO has good applicability due to its advantages such as a simple training process and relatively high stability. For tasks such as legal reasoning and case analysis that require multi-path argumentation, GRPO uses intra-group relative ordering to construct learning signals and shows great potential in handling complex legal arguments. In high-risk scenarios such as judicial auxiliary decision-making and sentencing recommendations, it is necessary to combine expert feedback, rule constraints, and knowledge retrieval mechanisms to achieve reliable alignment. The evaluation system is evolving from automated metrics to process interpretability and bias auditing. RLHF provides a key evolutionary path for legal large language models from usability to credibility. However, its implementation relies on endogenous rewards and hybrid feedback mechanisms to reduce annotation costs, and the development of reward traceability and process evaluation to meet judicial transparency requirements. Meanwhile, with the development of new methods such as direct preference optimization and group relative strategy optimization, legal RLHF is gradually evolving from the traditional reward model-driven mode to lightweight preference optimization and relative ranking optimization, providing a new technical path for reducing training complexity and enhancing the adaptability to legal scenarios. Existing research indicates that relying solely on a single reward model is insufficient to fully depict the complex trade-off relationships within the legal value system. In the future, it is necessary to further integrate rule knowledge, expert experience, and process feedback information to construct a more stable and interpretable legal alignment mechanism. This paper not only constructs a six-dimensional analysis framework for legal RLHF research, but also further proposes a legal task-oriented RLHF adaptability analysis system, systematically comparing the applicable boundaries and failure mechanisms of algorithms such as PPO, DPO, and GRPO in different tasks such as legal document generation, legal reasoning, and compliance review. In addition, this paper conducts a horizontal comparison of different RLHF methods by combining the risk levels of legal tasks and the characteristics of algorithms, providing a reference basis for algorithm selection and engineering deployment in legal intelligent systems. Research shows that the core contradiction of legal RLHF is not merely the insufficiency of algorithm performance, but rather a structural contradiction among multi-dimensional reward conflicts in legal scenarios, inconsistent expert preferences, and the difficulty in quantifying judicial value. This paper, for the first time, comprehensively reveals the entire process logic of legal RLHF from algorithm adaptation to engineering implementation from four aspects: technological evolution, reward modeling, evaluation system, and judicial governance, providing a systematic research framework and practical guidance for the trustworthy construction of legal large language models.

Globally escalating marine ecological pollution necessitates efficient and precise automated intelligent monitoring and cleaning technologies for underwater litter leveraging autonomous underwater vehicles (AUVs) or underwater robots, which has become an urgent academic consensus in the field of marine environmental protection. To address the inherent physical degradation of underwater visible light imaging—such as color casts, blurring, and low contrast—as well as the bottlenecks of cross-domain heterogeneity caused by varying optical properties of water bodies across different sea areas and the severe scarcity of high-quality fine-grained annotated samples, this study aims to overcome the feature shift and negative transfer phenomena induced by deploying models in unknown water environments. To this end, it proposes an underwater litter object detection framework that merges pixel-level image enhancement with a two-stage robust cross-domain adaptive learning scheme. This framework adopts a design strategy of "step-by-step construction and dynamic alignment," decoupling the pipeline construction into two mutually collaborative training stages. The first stage aims to construct an interference-resistant baseline perception architecture, establishing a solid feature foundation of deep representations for the entire network. To maintain high computational efficiency while circumventing the limitations of complex deep generative networks—which easily introduce spurious edges and grid-like artifacts under small-sample conditions—a low-cost pixel-level physical alignment pipeline is implemented at the data loading layer. Specifically, the input blue-green (BGR) raw images are first transformed into the LAB color space, where the L channel is extracted and subjected to adaptive pixel-level linear stretching to maximize the restoration of edge contrast and fine details. Subsequently, a Channel Stabilization Module compensates for color casts in the A and B channels to robustly suppress underwater-specific cyan-blue artifacts, before inversely mapping the calibrated feature maps back to the standard red-green-blue (RGB) space. On this basis, 30,000 iterations of supervised perceptual training are executed on the source dataset (TrashCan Instance), enabling the RetinaNet detector to acquire high-quality low-level general geometric textures and high-level robust semantic features, thereby establishing a solid perception baseline for multi-source scenarios. The workflow then naturally proceeds to the second stage, which focuses on domain alignment and precision enhancement for cross-domain deployment. To eliminate deep feature shift and response distortion triggered by unknown new water domains, a lightweight Feature Calibration Module (FCM) is symmetrically injected after the multi-scale output layers of the pre-trained backbone network. A differential initial assignment based on the Kaiming normal distribution is directly applied to the internal parameters of the FCM at the very beginning of the fine-tuning process to explicitly induce adaptive dynamic calibration. During the subsequent 10,000 iterations of cross-domain adaptive fine-tuning, to avoid catastrophic forgetting under low-data regimes, this stage completely freezes the feature extractor parameters of the backbone network, serving as a locked low-level feature stabilizer. Concurrently, a differential asymmetrical dual-learning-rate mechanism is deployed, assigning high gradient sensitivity (initial learning rate 0.01) to the newly injected FCM while allocating a smaller step size (initial learning rate 0.002) to the top-level Feature Pyramid Network (FPN) and detection head. Furthermore, an Exponential Moving Average (EMA) technique with a fixed decay rate (λ = 0.9999) is introduced throughout the entire fine-tuning process. It automatically switches to the smoothed EMA weights during the single complete evaluation performed every 1,000 iterations, leveraging a smooth weight loss plane to shield the model from training noise brought by extremely small samples and ensuring a lossless migration of feature space alignment. To guarantee the scientific rigor of the experimental design and its practical deployment reference value, this study introduces completely independent third-party cross-domain underwater datasets for blind generalization validation. Quantitative results demonstrate that while maintaining a lightweight and compact architecture with only 55.86 million parameters, the proposed method achieves a real-time inference speed of up to 44.61 FPS on a single NVIDIA RTX 4080 SUPER GPU, showcasing immense potential for edge-side practical deployment. In in-domain closed-loop testing, the framework significantly improves the bounding box Average Precision (AP) by 5.33 percentage points compared to the traditional RetinaNet baseline model, with a notable jump of 8.97 percentage points in the highly challenging task of detecting small targets characterized by blurred edges and low visibility underwater. In the most rigorous cross-water and cross-dataset generalization blind tests, the proposed method demonstrates remarkable performance gains: on the completely unseen independent J_Litter dataset, the AP_50 of the baseline model is as low as 0.31% due to multi-source interference, whereas the proposed method catapults the AP_50 to 28.66% through dynamic feature calibration; on the DeepPlastic deep underwater dataset, the method successfully elevates the baseline AP_50 from 3.30% to a stable 12.81%. These outcomes robustly validate the scientific validity and synergetic efficiency of the "physical foundation construction" and "dynamic feature calibration" within the two-stage architecture, proving its capability to effectively overcome negative transfer in cross-water detection and highlighting its strong practical generalization potential for real-world engineering deployment in automated marine litter removal.

As a core research task within the fields of artificial intelligence and multimodal understanding, Visual Question Answering (VQA) requires models to jointly comprehend the visual content of images and natural language questions in order to generate accurate answers. Addressing the limitations of existing VQA methods—which typically lack large-scale pre-training and suffer from interference by weakly correlated background elements and redundant regions during multimodal feature modeling—this paper proposes a multi-module collaborative VQA method based on multimodal filtering and spatial-channel attention. Specifically, the proposed method first introduces a Multimodal Filtering Module. This module extracts question feature vectors from specific positions to serve as a global semantic representation; it then performs a Hadamard product interaction with aligned visual features and utilizes a perceptron combined with an activation function to generate normalized filtering gate weights. By element-wise scaling the image features prior to deep-level interaction, this process explicitly suppresses redundant background information, thereby providing a "cleaner" visual input for subsequent deep interactions. Subsequently, a Spatial-Channel Self-Attention Unit is introduced to replace the traditional self-attention architecture. Within a six-layer stacked encoder-decoder framework, this unit reshapes features into 2D maps along the spatial dimension; by combining multi-scale convolutional kernels with global pooling, it generates a spatial attention distribution that effectively captures spatial dependencies between different image regions. Simultaneously, along the channel dimension, it employs pooling and fully connected networks to dynamically adjust the semantic importance of individual channels. Finally, a gating mechanism is utilized to perform a weighted fusion of the spatial-channel-modulated local features with the global context captured by multi-head self-attention, thereby effectively enhancing the discriminative power of the visual features. Building upon this foundation, a Dynamic Modality Fusion Module is constructed to move beyond static fusion mechanisms. After multiple layers of interaction, the model generates attention maps for both visual and textual features, concatenates them, and feeds them into a fully connected classification network (incorporating dropout) to adaptively learn 2D dynamic fusion weight factors. This process ultimately achieves a dynamic weighted integration of multimodal features, with the overall model parameters optimized using a binary cross-entropy loss function. Extensive quantitative experiments conducted on the VQA-v2 and GQA datasets demonstrate that the proposed method achieves significant performance improvements, even without relying on massive-scale cross-modal pre-training data. Specifically, on the VQA-v2 test-dev split, the proposed method achieved an overall accuracy of 71.85%, representing a significant improvement of 1.22% and 2.33% over comparable baseline models—MCAN and BAN—respectively. Notably, the accuracy rates for binary "Yes/No" questions and "Other" question types reached as high as 88.03% and 62.39%, respectively. On the GQA test set—which emphasizes compositional reasoning and relational understanding—the model achieved an overall accuracy of 57.92%, with binary "Yes/No" questions attaining an accuracy of 76.68%. Furthermore, rigorous ablation studies quantitatively validated the efficacy of each core module; the introduction of the multi-modal filtering, spatial-channel attention, and dynamic modality fusion modules yielded absolute performance gains of 0.63%, 0.69%, and 0.52%, respectively. Additionally, an analysis of the impact of model depth confirmed that feature interactions are most comprehensive—and overall model performance reaches its peak—when both the encoder and decoder consist of six stacked layers. In summary, by employing fine-grained feature modeling and dynamic interactions, the proposed method successfully mitigates issues such as severe noise interference, isolated spatial-channel processing, and the poor adaptability of static fusion mechanisms, thereby offering an improved solution for visual-language multi-modal understanding and deep feature representation within specific vertical domains characterized by limited computational resources.

Generative face-swapping technology poses a severe threat to personal privacy. Existing proactive defense methods suffer from poor generalization in black-box scenarios and insufficient robustness under image processing distortions. To address these issues, this paper proposes a Robust Black-box Proactive Defense (RBPD) framework. The framework aims to inject imperceptible adversarial perturbations into the source image to disrupt the identity feature extraction process of unknown face-swapping models, thereby achieving reliable identity protection.The framework employs a two-stage generation mechanism. In the first stage, a Semantic-Aware Encoder (SAE) and a Texture-Guided Decoder (TGD) are utilized to generate initial semantic perturbations. The SAE uses facial semantic masks for guidance and incorporates a Convolutional Block Attention Module to accurately focus on identity-critical regions. The TGD fuses shallow texture features through skip connections, constraining the perturbations to adapt to the local texture distribution and gradient intensity of the source image. This ensures effective attack performance while significantly reducing visual artifacts and improving visual quality.In the second stage, a Dual-Stream Fusion Encoder (DFE) and a Multi-Scale Aggregation Decoder (MAD) are introduced. The DFE extracts deep features from both the source image and initial perturbations and performs nonlinear fusion, deeply embedding adversarial information into the image’s semantic feature space. The MAD employs three parallel dilated convolutions to capture multi-scale contextual information and integrates a Squeeze-and-Excitation module for adaptive channel recalibration, thereby enhancing the robustness of the perturbations against image distortions. Additionally, a Meta-learning Adaptive Attack (MAA) strategy is designed, which integrates gradient feedback from four heterogeneous identity feature extractors (ArcFace, FaceNet, MagFace, and AdaFace), dynamically adjusts optimization weights, and achieves precise breakthroughs against strongly robust extractors, thereby improving the cross-model generalization of the perturbations on unknown black-box models.Evaluations on the CelebA-HQ and RaFD datasets, targeting three mainstream face-swapping models (SimSwap, E4S, and DiffSwap) as well as commercial facial recognition APIs from Baidu and Tencent, yield the following results: the protected images achieve average Top-1 and Top-5 identity matching rates of 0.311 and 0.396, respectively, representing decreases of 63.28% and 54.79% compared to unprotected images, and further reductions of 16.17% and 19.02% compared to the best baseline methods. On the unseen RaFD dataset, the method maintains stable performance with average Top-1 and Top-5 matching rates of 0.349 and 0.394. In black-box face-swapping tests, the cosine similarity between swapped images and source images mostly drops below 0.3, with the lowest reaching 0.185, achieving reliable identity mismatch. When facing common social network distortions such as JPEG compression, Gaussian blur, noise, and resizing, the Distortion Defense Volatility (DDV) averages only 3.4%, significantly outperforming the baseline methods Saliency (55.5%), DF-RAP (27.1%), NullSwap (19.70%), and ID-Eraser (13.90%). In terms of visual quality, the protected images attain a PSNR of 37.38 dB, SSIM of 0.976, and LPIPS of 0.0065, demonstrating good perceptual naturalness. In commercial API tests, the face matching pass rate on Baidu API decreases from 89.92% to 0.40%, and on Tencent API from 90.37% to 4.25%. The total processing time for a single image is only 8.23 ms, indicating strong practical deployment potential.The proposed RBPD framework effectively addresses the deficiencies of existing methods in black-box generalization and distortion robustness. Through the synergistic design of semantic-texture guidance and multi-scale deep fusion, it achieves a strong balance among attack effectiveness, visual quality, and robustness, providing an efficient and practical proactive defense solution for personal privacy protection in complex social network environments. This work holds significant theoretical importance and application value.

Question answering models aim to enable machines to understand natural language questions and automatically reason over text or knowledge bases to generate accurate answers. Existing retrieval-augmented generation (RAG) based QA methods improve generation quality by incorporating external knowledge, yet they still face two challenges: first, insufficient semantic understanding makes it difficult to capture the relationship between the question and the retrieved knowledge; second, the retrieval stage inevitably introduces noisy information with low relevance to the question, which is then amplified during the generation process, easily leading to answer deviation or hallucination. To alleviate these issues, this paper proposes a Semantic Retrieval-Augmented Generation Question Answering model based on a two-stage filtering mechanism, named BiS-RQA. By constructing a progressive framework of explicit and implicit filtering, the model improves answer accuracy in QA scenarios while maintaining efficient inference. The model adopts a two-stage progressive filtering architecture: the first stage filters out obviously irrelevant knowledge entities, and the second stage performs noise suppression on the retained candidate knowledge, allowing the large model to focus on the most valuable pieces of information. The first stage is explicit filtering, which introduces a Semantic-guided Personalized PageRank algorithm (SPPR). In implementation, for an input natural language question, a pre-trained language model is used to extract the question embedding vector; entities and relations in the knowledge graph also obtain their respective embedding vectors through the same model. First, the similarity between the question vector and each entity vector is computed and used as a personalization vector to guide the initial preferences in a random walk with restart. Then, the similarity between the question vector and each relation vector is computed, and this similarity is employed to adjust the transition matrix from the current node to its neighbors during the walk, i.e., the jump probabilities are modified according to the question-relation similarity. On this basis, the algorithm iterates until convergence, yielding stationary distribution scores for all nodes. Finally, the nodes are sorted in descending order of scores, the top-K entities and their associated relations are selected, and a corresponding subgraph is constructed. This process explicitly filters out irrelevant nodes and edges from the original knowledge graph, thereby effectively improving the relevance of the retrieved information. The second stage is implicit filtering, which proposes a Multi-Attention Information Fusion method (MAIF). The MAIF method utilizes self-attention and cross-attention mechanisms to weight entities and relations, respectively. First, entities and relations are pooled to serve as the subgraph embedding; subsequently, the entities, relations, and subgraph pass through a self-attention module to capture internal dependencies, and then cross-attention is further performed to allow them to interact with each other to obtain the correlation degree between information at different granularities, thus achieving implicit filtering and weighting of the candidate information. Finally, the weights output by the cross-attention mechanism serve as soft prompts, while the entities and relations obtained from the first stage serve as hard prompts, and together with the question they are fed into the large language model for answer generation. On the dataset WebQSP, BiS-RQA achieves an accuracy of 77.21%, a relative improvement of 4.6% over the best baseline model G-Retriever (73.79%); on the dataset CWQ, BiS-RQA achieves an F1 score of 35.38%, a relative improvement of 8.1% over the second-best baseline (32.70%). In terms of memory consumption, BiS-RQA uses 31.5 GB of memory, an increase of 1.5% compared to the baseline G-Retriever; in terms of inference time, BiS-RQA requires 0.9427 seconds, an increase of 56% compared to GRAG. Experimental results on multiple public QA datasets validate the effectiveness of the proposed method in semantic retrieval and answer generation, and demonstrate strong competitive advantages.

While punctuation restoration of standardized text has made significant progress, real-time punctuation restoration within Automatic Speech Recognition (ASR) systems remains a challenging task. The core challenge lies in the colloquial nature of ASR transcripts and the severe imbalance of punctuation categories. Moreover, despite strong semantic understanding capabilities, large language models (LLMs) suffer from high inference latency and deployment costs. To alleviate these problems, this paper proposes a real-time punctuation restoration method based on progressive iterative optimization. First, multiple public Chinese corpora are integrated to construct a large-scale training dataset covering diverse text domains. In addition, a punctuation-aware data weighting strategy is introduced to mitigate the learning bias caused by imbalanced punctuation distributions. By assigning different weights to training samples, the proposed strategy improves the alignment between the training corpus and real-world ASR outputs. Then, by introducing dynamic masking to restrict the attention range, only the context information within a limited window near the current position is retained, ensuring the ability to obtain key semantic information and effectively reducing long-distance noise interference, thereby achieving a balance between accuracy and latency. Finally, a progressive iterative optimization mechanism is introduced. The sliding-window average loss is used to estimate sample difficulty, and a dynamic weighting strategy is designed accordingly. For difficult samples that continuously generate high prediction losses, a multi-sentence concatenation method is used to generate new samples with more complex context dependencies, and their training weights are increased, enabling the model to gradually focus on complex semantic boundaries, achieving the collaborative evolution of training data and model capabilities, and thereby continuously improving the model's ability to capture complex semantic structures. This paper conducts systematic experiments on the real punctuation restoration test set and compares it with representative models such as CT-transformer, Qwen2.5-7B, Llama3.1-8B, Gemma2-9B, and DeepSeek-V3. The experimental results show that the F1 value of this method is 4.92 percentage points higher than that of the traditional real-time punctuation restoration baseline CT-transformer model. Moreover, when compared with large language models with parameter scales much larger than this model, this method still achieves better restoration performance. In terms of inference efficiency, the average inference time per sentence of the model is only 26 ms, which is nearly 200 times faster than DeepSeek-V3, meeting the requirements of low latency response in real-time speech processing scenarios. Ablation experiments further verify the effectiveness of the controllable delay mechanism, data weighting strategy, and progressive iterative optimization module in improving model performance. Among them, the iterative optimization process gradually increases the F1 value from 38.78% to 43.70%. Compared with general large models relying on large-scale parameters, the data and model collaborative optimization scheme designed based on task characteristics can achieve better punctuation restoration effects while maintaining low resource consumption and low-latency inference, demonstrating strong practical value for real-time ASR post-processing applications.

To address the issues of severe background interference, large variations in defect target scales, inconspicuous features of small-scale defects, and progressive loss of edge information in deep networks during insulator defect detection in UAV inspection scenarios, and given the favorable balance between real-time performance and parameter count of YOLOv11n, this paper proposes an improved insulator defect detection method based on YOLOv11n to enhance detection accuracy and robustness under complex background conditions. This method performs collaborative optimization of the network structure from three aspects, namely feature enhancement, feature fusion, and feature extraction, thereby improving the model's perception and recognition capabilities for insulator defect targets in complex scenes while maintaining lightweight characteristics.The proposed model incorporates three core optimization modules. First, to alleviate complex background interference and enhance fine-grained defect representation capability, a Dynamic Dual-Domain Feature Enhancement Module (DDFEM) is designed in the feature extraction stage of the backbone network, establishing a collaborative working mechanism between a global semantic branch and a local detail branch. The global branch establishes long-range spatial dependencies through bidirectional global pooling and cross-dimensional matrix interaction, achieving global semantic modeling with low computational complexity. The local branch extracts fine-grained texture features such as flashover traces and damage using multi-branch depthwise separable convolutions, combined with a dynamic attention fusion mechanism for adaptive recalibration of local features, thereby achieving complementary enhancement of global semantic information and local detail information. Second, to address the progressive weakening of edge detail information during deep semantic feature enhancement, a Sobel-Edge-Guided Weighted Fusion Module (SEGWF) is designed. This module explicitly extracts edge information from shallow features using the Sobel operator and dynamically fuses edge structural information with deep semantic features through a channel-wise weighted fusion mechanism, thereby enhancing the model's perception of key structural features such as damaged edges and contours of insulators, and improving small target detection performance under complex backgrounds. Finally, to enhance multi-scale target feature extraction capability and address the fixed receptive field of traditional convolutions that struggles to adapt to defects of different scales, a Receptive-Field Attention Convolution (RFAConv) mechanism is introduced into the backbone network to replace traditional convolutional structures. This mechanism adaptively adjusts the receptive field response range according to the feature distribution of different spatial regions, thereby improving the model's feature extraction capability for insulator defects at different scales and enhancing multi-scale target detection performance.To verify the effectiveness of the proposed method, experimental studies are conducted on a self-constructed insulator defect dataset, with comparisons against mainstream object detection models. Experimental results show that the proposed method achieves 92.2% mAP@0.5, representing a 4.9 percentage point improvement over the original YOLOv11n.Precision increases from 89.7% to 93.0% with a 3.3 percentage point improvement, while Recall increases from 79.6% to 88.3% with an 8.7 percentage point improvement, demonstrating that the proposed method effectively reduces the miss detection rate and improves defect detection accuracy in complex scenes. Furthermore, to validate the generalization capability of the model, tests are conducted on the public insulator defect dataset IDID. Experimental results show that the proposed method achieves a Precision of 89.6%, a Recall of 79.8% and an mAP@0.5 of 88.9%. It obtains the optimal mAP@0.5 among all comparative models, with an improvement of 1.9 percentage points over the original YOLOv11n, which verifies its strong cross-dataset generalization ability and stable detection performance. Meanwhile, the proposed model only has 3.12M parameters and a computational cost of 7.6 GFLOPs, maintaining low computational complexity while boosting detection accuracy. In conclusion, the improved YOLOv11n algorithm proposed in this paper can effectively complete the task of insulator defect detection under complex backgrounds.

Multi-view subspace clustering aims to learn latent shared structures by exploiting the consistency, complementarity, and discrepancy information among multiple feature sources. How to simultaneously preserve global structural relationships, local geometric characteristics, and representation stability during feature fusion remains a critical issue in this field. Existing methods are mainly built upon low-rank representation, sparse constraints, or graph learning mechanisms. Although they can exploit cross-view shared information to some extent, they still lack sufficient collaborative modeling of global structure preservation, local smoothness constraints, and robust representation learning, making them prone to local structure distortion, unstable representations, and degraded clustering performance under noise contamination, outlier perturbations, and complex data distributions. To address these issues, this paper proposes a Structured Local Smooth Multi-View Subspace Clustering (SLS-MVSC) method. Under a unified self-representation learning framework, the proposed method jointly incorporates low-rank constraints, Total Variation (TV) regularization, graph regularization, and robust error modeling to achieve collaborative optimization of global information learning, local relationship preservation, and stable representation learning. Specifically, low-rank constraints are employed to learn shared representations across multiple views, thereby capturing the latent consistent global subspace structure among different views and enhancing the modeling capability for cross-view common information. Considering that noise interference may cause severe fluctuations in the self-representation matrix, TV regularization is introduced into the self-representation learning process to preserve representation continuity and boundary structures by restricting abrupt local variations, thereby improving the smoothness and stability of learned representations. Furthermore, graph regularization is adopted to maintain neighborhood relationship consistency among samples. By explicitly modeling local manifold structures, the proposed method strengthens the characterization of data geometric relationships and enables the learned representations to better conform to the intrinsic distribution characteristics of the original data. Meanwhile, a robust norm is introduced into the reconstruction error term to enhance the model’s adaptability to outliers and complex noise, thereby improving representation quality and clustering robustness from multiple perspectives. To optimize the proposed model, an iterative algorithm based on the Alternating Direction Method of Multipliers (ADMM) is developed. By introducing auxiliary variables, the complex objective function is decomposed into several independently solvable subproblems, and the update rules for all variables together with the overall optimization procedure are derived, ensuring training stability and computational efficiency. Extensive experiments are conducted on six publicly available datasets to evaluate the effectiveness of the proposed method and compare it with several representative multi-view clustering approaches. Experimental results demonstrate that the proposed method achieves superior performance in terms of Clustering Accuracy (ACC), Normalized Mutual Information (NMI), Purity (PUR), Adjusted Rand Index (AR), and F-score, attaining optimal or near-optimal results on multiple datasets and exhibiting strong clustering capability and cross-dataset adaptability. Further ablation studies verify the positive contribution of each module to performance improvement. Among them, low-rank constraints and graph regularization play important roles in global structure learning and local geometric relationship preservation, respectively, while TV regularization is particularly effective in enhancing representation smoothness, alleviating noise-induced representation fluctuations, and improving model stability. Robustness experiments further demonstrate that the proposed method maintains stable performance under different noise intensities, indicating its effectiveness in mitigating the adverse influence of noise on feature representation learning and clustering results. In summary, the proposed SLS-MVSC method effectively improves clustering performance, representation stability, and noise robustness, providing an effective solution for complex multi-view data clustering tasks.

Existing salient object detection (SOD) methods generally follow the principle of passive visual stimulation. These methods rely on low-level features such as color, texture, and contrast to determine salient regions. They treat the object with the strongest visual features in the scene as the focus of user attention. However, they neglect the decisive role of active user demands in determining salient targets. In real scenarios such as human-computer interaction and robot inspection, target objects are often submerged in complex backgrounds. In these cases, user attention has a strong intent orientation. Recently, the user-demand-driven salient object detection (UserSOD) task is introduced. It shifts the perception paradigm from passive visual response to active intent matching. This task also provides corresponding benchmark datasets and baseline models. Existing methods lack deep fusion and dynamic calibration between visual features and user demand semantics. Consequently, models only capture shallow semantic associations of keywords. Moreover, the downsampling process in hierarchical backbone networks easily loses spatial details. Therefore, models fail to accurately locate targets that match abstract demands in complex backgrounds. The detection performance remains heavily limited. To address these issues, this paper proposes a depth calibration and attribute-aware network (DCA-Net) within the UserSOD task framework. This network adopts the Swin-Transformer model as the visual backbone. It combines a pre-trained Contrastive Language-Image Pre-training (CLIP) text encoder as the text branch. The network designs a cascaded semantic recalibration encoder (CSRE) as its core. The CSRE contains four independent semantic recalibration modules (SRM). Each SRM utilizes a cross-modal cross-attention mechanism to achieve step-by-step semantic alignment between visual features and user intent at the source of feature extraction. Meanwhile, each SRM introduces a gated feature flow mechanism (ScGate) based on a multilayer perceptron structure. The ScGate uses full-connection transformation and ReLU-Tanh activation to generate adaptive weights. These weights dynamically adjust the intensity of semantic flows in each channel. This mechanism accurately locks the target region in the early stage of the feature flow. Meanwhile, the network constructs a convolutional cross-scale interaction module (CCIM) to achieve cross-level feature compensation. The CCIM aggregates global features using a pyramid pooling aggregation operator and a 1×1 convolution. It captures multi-scale context information through three parallel 3×3 depthwise separable convolution branches with different dilation rates. Then, the module applies weights by combining serial channel and spatial attention modules. Finally, it uses bilinear interpolation upsampling to restore the feature size. This module significantly enhances the perception ability of the model for spatial details of multi-scale targets.Furthermore, the network introduces a fine-grained attribute-aware decoder (AGD) to provide fine-grained constraints. The AGD adopts a top-down path to step-by-step upsample and merge high-level features. It utilizes the CLIP text encoder to initialize four types of preset attribute prompts, including category, color, appearance, and functional requirements. This initialization provides semantic priors and avoids the random convergence of attribute queries. Next, the AGD extracts explicit semantic constraint vectors of the four attribute dimensions from user instructions via multi-head cross-attention. This operation decouples abstract intent into four independent fine-grained attribute constraints. It avoids semantic confusion caused by multi-attribute demands. Then, the decoder feeds visual features into four parallel attribute-specific feature branches to perceive the visual patterns of corresponding attributes. These features then fuse with the constraint vectors. Finally, global text features dynamically predict the attribute existence probability weights. The decoder adaptively adjusts the contribution intensity of each branch based on the attribute of the four categories. This process ensures the precise correspondence between fine-grained constraints and user intent. It guides the model to generate detection masks with clear boundaries and consistent semantics.Experimental results on the UserSOD benchmark dataset demonstrate the effectiveness of the model. Compared with the state-of-the-art methods, DCA-Net improves the S-measure, F-measure, and E-measure by 4.5%, 6.8%, and 3.8%, respectively. It also reduces the mean absolute error (MAE) to 0.028. These results effectively validate the superiority and strong robustness of the proposed architecture in complex background interference, multi-scale target capture, and abstract intent alignment.

Physiological time series classification plays a critical role in various healthcare tasks, including sleep monitoring, electrocardiogram diagnosis, and epilepsy detection. However, in practical applications, severe class imbalance makes it challenging to learn effective representations for minority classes. Although these minority samples are scarce, they often contain highly informative signals, and accurate identification of their patterns is crucial for timely disease intervention, precise health assessment, and optimized decision-making in clinical and other application scenarios. Due to the limited number of minority samples and the predominance of background or normal waveforms, models struggle to learn discriminative features from entire sequences. Within these minority samples, a small number of critical local segments exist, whose unique structures or dynamic patterns carry essential class-discriminative information. Recognizing these segments is vital for effective modeling of minority classes. Contrastive learning offers strong generalization and feature distribution awareness and has been widely used to address class imbalance. Yet, most existing approaches rely on global sample representations or originate from image-based methods, making it difficult to capture sparse yet discriminative local waveform segments in minority samples. Moreover, current methods often underutilize frequency-domain information, and data augmentation typically lacks class differentiation, ignoring distributional differences between minority and majority classes. To address these challenges, this study proposes KWave-CL, a key waveform-aware contrastive learning method for class-imbalanced physiological time series classification. Hard-to-learn key segments in minority samples typically exhibit large reconstruction errors. KWave-CL employs a variational autoencoder jointly modeling time- and frequency-domain information to reconstruct minority waveforms. By computing reconstruction errors in both domains, the method identifies key and non-key waveform segments, providing critical guidance for subsequent contrastive learning. To fully exploit local discriminative information in minority samples, a key waveform-aware contrastive loss is designed to pull key segments closer while pushing key and non-key segments apart in the representation space, enhancing the discriminability of minority class embeddings. To increase minority diversity while maintaining the stability of majority features, class-differentiated data augmentation is applied, imposing stronger perturbations on minority samples and weaker perturbations on majority samples, thereby mitigating class distribution shifts in the augmented space. The method adopts a joint optimization strategy, integrating self-supervised contrastive loss, key waveform-aware contrastive loss, and time-frequency variational autoencoder reconstruction loss, enabling cooperative learning of global and local features. KWave-CL is also highly flexible and can be embedded into various time series contrastive learning frameworks. Experiments on three publicly available physiological datasets show that KWave-CL outperforms multiple baseline methods for class-imbalanced learning under two representative contrastive learning frameworks. On the PhysioNet 2017 dataset using an instance-level framework, the overall F1 score improves by up to 6.69%, while the minority class F1 score increases by up to 11.67%. Ablation studies further demonstrate that the key waveform-aware contrastive loss, time-frequency variational autoencoder, and class-differentiated data augmentation all play crucial roles in enhancing minority class performance. These results indicate that KWave-CL effectively mitigates class imbalance and provides reliable decision support for healthcare applications.

To address the problems of extremely small target scale, weak texture information, low signal-to-noise ratio, and severe background interference in infrared UAV aerial images, while satisfying the lightweight and real-time deployment requirements of UAV platforms, this paper proposes a lightweight infrared small object detection algorithm named ACFF-YOLOv12n based on YOLOv12n. Existing infrared small object detection methods generally suffer from insufficient fine-grained feature modeling, inefficient multi-scale feature fusion, and poor adaptability of detection heads to low-contrast targets, resulting in feature dilution, missed detections, and false alarms during deep feature propagation. To improve the perception capability of weak infrared targets in complex scenes, the proposed method performs collaborative optimization from three aspects: backbone feature extraction, feature fusion strategy, and detection head design. First, an A2C2f-ACmix++ feature enhancement module is constructed in the backbone network. By integrating global self-attention and local enhanced convolution through a dual-path collaborative mechanism, together with a Lightweight Channel Attention (LCA) module, the network can adaptively strengthen weak target channel responses and enhance the extraction capability of edge and texture information under limited computational cost. Second, a Semantic-Gated Dynamic Fusion Module (SGDFM) is designed in the neck network. High-level semantic features are used as dominant guidance, while shallow detail features are treated as auxiliary information. Through dynamically generated gating weights, adaptive cross-level feature fusion is achieved, effectively alleviating the feature dilution problem of infrared small targets during deep propagation and improving robustness under complex backgrounds and scale variations. Finally, a CondConv-based Lightweight Detection Head (CLD-DET) is proposed for infrared small object detection. The detection head introduces Conditionally Parameterized Convolution (CondConv) to dynamically adjust convolution kernels according to input features and incorporates a Small Object Enhancement (SOE) module to strengthen weak edges and low-contrast spatial details, thereby improving the localization and classification capability for infrared weak targets. Experiments are conducted on the HIT-UAV infrared dataset, and additional generalization evaluations are performed on the SIRST and NUDT-SIRST datasets. Experimental results demonstrate that ACFF-YOLOv12n achieves an mAP@0.5 of 87.1% on the HIT-UAV dataset, which is 5.7 percentage points higher than YOLOv12n, while the mAP@0.5:0.95 reaches 56.8%, improving by 2.8 percentage points. Meanwhile, the number of parameters is reduced to 2.13M and the computational complexity is reduced to 5.0 GFLOPs, achieving an effective balance between detection accuracy and lightweight deployment. On the SIRST dataset, the proposed method achieves an mAP@0.5 of 78.6%, outperforming YOLOv12n by 5.6 percentage points. On the NUDT-SIRST dataset, the mAP@0.5 reaches 76.7%, improving by 3.4 percentage points over the baseline model, demonstrating strong cross-scene generalization and robustness. Furthermore, visualization and heatmap analyses show that the proposed model can generate more accurate and balanced feature attention regions in complex backgrounds, long-distance weak target scenes, and dense small-target scenarios, effectively reducing missed detections and false alarms. Overall, ACFF-YOLOv12n achieves a superior balance among detection accuracy, computational complexity, and real-time performance, making it highly suitable for infrared UAV small object detection tasks and practical applications such as UAV inspection, border surveillance, and complex environment monitoring.

Simultaneous localization and mapping (SLAM) in dynamic scenes is highly susceptible to moving objects, occlusions, and illumination variations, which often lead to pose drift and reconstruction artifacts. Existing methods remain limited in identifying unreliable dynamic observations and constraining map updates. They frequently depend on category priors, fixed thresholds, or scene-specific distributions, and may incorporate moving objects into the map as spurious static structures. To address these issues, a Multi-Source Consistency Residual-Enhanced Gaussian SLAM method, termed MCRGS-SLAM, is proposed for dynamic scenes. It models dynamic interference from the perspective of observation reliability in a continuous and interpretable manner, and embeds reliability constraints into both front-end pose estimation and back-end Gaussian mapping to improve the localization accuracy and static map reconstruction quality of monocular SLAM. The method is built on the physical constraint that static regions should satisfy multi-view consistency. It constructs four complementary consistency residuals, namely appearance, geometric, motion, and structural residuals, thereby transforming dynamic observation identification into a quantifiable physical measurement. Specifically, the appearance residual characterizes cross-view brightness and texture consistency, the geometric residual measures the stability of depth projection relationships, the motion residual detects independent displacement that cannot be explained by camera ego-motion, and the structural residual describes variations in local edges and geometric patterns. Unlike strategies that directly remove dynamic regions using binary masks or fixed thresholds, these residuals are modeled as continuous physical constraints to represent the degree to which pixel observations deviate from the static-scene assumption. On this basis, a semantic–geometric dual-stream reliability inference network, named Multi-source Consistency Residual Network (MCR-Net), is designed. The semantic stream extracts high-level semantic features to provide category-level dynamic priors, while the geometric stream encodes residual evidence to represent multi-view consistency. The two streams are fused through an attention mechanism to generate a pixel-wise reliability map. This map is introduced into both the front-end and back-end of the SLAM system as soft weights. In front-end pose optimization, appearance, motion, geometric, and structural constraints are weighted by reliability to reduce the interference of dynamic outliers in camera pose estimation. In back-end 3D Gaussian Splatting (3DGS) mapping, the reliability map guides the initialization, update, and removal of Gaussian primitives, thereby adaptively suppressing map contamination caused by dynamic observations. In this way, dynamic observation handling is transformed from discrete removal into reliability-based continuous weighting, which preserves boundary regions and weakly reliable static observations while reducing the cumulative influence of dynamic outliers on pose estimation and Gaussian map updates. In addition, MCRGS-SLAM establishes a self-supervised closed-loop optimization mechanism based on reprojection errors and rendering consistency, enabling the network to adapt online to dynamic variations in unknown scenes. Experiments on dynamic-scene datasets, including Bonn and TUM, show that MCRGS-SLAM achieves competitive performance in both localization accuracy and reconstruction quality. In localization evaluation, MCRGS-SLAM obtains an average ATE RMSE of 2.35 cm on the Bonn Dynamic dataset, outperforming several representative methods. These results indicate that reliability-weighted optimization effectively reduces the impact of dynamic observations on pose estimation. In the reconstruction and rendering quality evaluation, the proposed method achieves average PSNR, SSIM, and LPIPS values of 17.99 dB, 0.730, and 0.272 on the TUM dynamic dataset, respectively. Compared with Dy3DGS-SLAM, a representative dynamic 3DGS-SLAM method, it improves PSNR by 0.14 dB and SSIM by approximately 1.1%, while reducing LPIPS by approximately 4.6%, effectively mitigating reconstruction artifacts caused by dynamic objects. Results on real-world complex scene sequences further demonstrate that MCRGS-SLAM maintains stable performance under scene distribution shifts and unstructured dynamic interference, indicating good cross-scene applicability.

The anonymity inherent in blockchain provides a concealed environment for transactions, and Bitcoin address type detection, which aims to categorize addresses with distinct behavioral patterns, is of great significance for blockchain de-anonymization analysis. Existing methods face three major challenges: (1) the Bitcoin transaction network is large and complex, leading to low training efficiency when directly using graph neural networks; (2) deep graph neural networks are prone to over-smoothing, causing node representations to converge and discriminative ability to decline; (3) most methods rely solely on static graph structures or simple statistical features, ignoring temporal transaction patterns and failing to capture dynamic address behaviors. To address these issues, this paper proposes F-BAC (Filter-enhanced Bitcoin Address Classifier), a Bitcoin address type detection method based on a temporal graph feature network. The method consists of four core steps. First, transaction modeling: each transaction of a labeled address is constructed as an independent homogeneous graph, where nodes are addresses involved in the transaction and edges represent fund flows. To handle the large variance in transaction scales, a fine-grained address aggregation compression strategy is proposed, which divides nodes into four tiers based on their total occurrence frequency in all transaction records and aggregates them accordingly, transforming the original heterogeneous graph into a small-scale homogeneous graph with regular structure. This significantly simplifies topological complexity and lays an effective foundation for subsequent graph representation learning. Second, address feature extraction and augmentation: local features are computed for each address, and statistical features are computed for each aggregated address group. Then, a graph feature network (GFN) is applied to propagate and augment the features through three layers while preserving the original node information at each layer, effectively mitigating the over-smoothing problem. Third, temporal feature optimization: the time-ordered sequence of transaction graph representations of the same address is fed into a filter-enhanced MLP (F-MLP), transformed to the frequency domain via discrete Fourier transform, where a trainable linear filter suppresses obfuscation behaviors and noise, and finally reconstructed to an enhanced temporal signal via inverse Fourier transform. This design significantly improves the model's ability to capture periodic transaction patterns. Fourth, address classification: the filtered and optimized temporal graph representation sequence is input into a classifier to output the address type label, trained under the supervision of the cross-entropy loss function. This paper also constructs a dataset incorporating temporal features, providing an important data foundation for subsequent research. Systematic experiments are conducted on our self-constructed dataset and three public datasets. Ablation studies show that removing the address aggregation module, the GFN module, or the temporal filtering module degrades model performance to varying degrees, validating the necessity of the three strategies. On our self-constructed dataset, F-BAC achieves 96.7% precision, 95.9% recall, 96.3% F1-score, and a macro-averaged F1 of 91.1%. On the BAC dataset, F-BAC achieves an F1-score of 98.4%; on the BATC dataset, 98.0%; and on the Google/CvdxBp dataset, 94.6%. Comparative experiments show that F-BAC outperforms existing methods including BAClassifier, Multi-hop GAT, Balanced-BiEGCN, and MDST-GNN. Furthermore, F-BAC reduces data collection volume by 89.3% compared to the BABD method, and the GFN module saves approximately 30%–45% of training time compared to GCN, significantly reducing training costs. In summary, the proposed F-BAC method effectively alleviates the problems of large graph scale, over-smoothing, missing temporal patterns, and obfuscation interference in Bitcoin address type detection through fine-grained address aggregation, graph feature network enhancement, and temporal frequency-domain filtering. Experiments on multiple datasets validate that F-BAC has good generalization performance for this task.

A Multi-distribution and Multi-scale Adaptive Prior Transformer for Anomaly Detection (MMAPT-AD) is proposed to address the limitations of existing multivariate time series anomaly detection methods in modeling complex temporal dependencies, multi-scale dynamic patterns, and prior information utilization. The model employs a data embedding module to obtain unified representations of multivariate time series from both temporal and variable dimensions, enhancing the representation of variable coupling relationships and dynamic temporal patterns. A multi-distribution and multi-scale prior generation mechanism is further introduced by jointly incorporating Gaussian, Laplace, and Cauchy distributions under different temporal scales to model latent correlation structures within the sequence. Specifically, the Gaussian distribution characterizes stationary variation patterns, the Laplace distribution captures local abrupt changes, and the Cauchy distribution improves the adaptability to long-tail anomalies and complex fluctuation patterns. To describe heterogeneous temporal dependencies across different scales, prior correlation matrices are generated at multiple temporal granularities for joint modeling of local dependencies and global temporal structures. Learnable fusion weights are adopted to integrate heterogeneous prior information from different distributions and scales, improving the representation capability for complex statistical characteristics and dynamic patterns. Based on the generated priors, a prior-guided anomaly attention mechanism is designed by incorporating multi-scale prior information into the attention weight computation process. The proposed mechanism introduces structural prior constraints while learning temporal correlation features and guides the model to focus on anomaly-related temporal segments and variable channels, thereby improving the detection capability for local anomalies, sparse anomalies, and complex structural anomalies. To enhance anomaly discrimination capability, a joint optimization objective combining reconstruction error, attention-prior discrepancy constraints, and multi-scale inconsistency constraints is constructed. In addition, a comprehensive anomaly scoring strategy integrating reconstruction error, attention distribution discrepancy, and multi-scale structural deviation is developed for time-step-level anomaly detection. Experiments on five public datasets, including SMD, MSL, SWaT, SMAP, and PSM, show that MMAPT-AD achieves F1 scores of 93.40%, 94.99%, 96.06%, 96.67%, and 98.06%, respectively. On the SMD dataset, the proposed method improves the F1 score by 7.18 percentage points and 1.07 percentage points compared with InterFusion and Anomaly Transformer, respectively. On the MSL dataset, MMAPT-AD achieves an F1 score of 94.99%, outperforming TransDe by 0.63 percentage points. On the SMAP dataset, the Recall reaches 99.35%, indicating strong anomaly coverage capability. Ablation studies demonstrate that multi-distribution prior modeling, multi-scale structural constraints, and the joint optimization strategy all contribute to performance improvement. Robustness experiments further verify the stability and generalization capability of the model under different input perturbation conditions. Experimental results demonstrate that MMAPT-AD effectively integrates multi-distribution statistical characteristics and multi-scale temporal dependencies, exhibiting strong anomaly detection capability and structural adaptability in complex dynamic scenarios.

To address the problems of decreased accuracy in causal structure learning and unreliable causal direction determination among observed variables under latent-variable interference, a latent variable causal discovery method based on non-Gaussian information is proposed. In real-world data, there often exist latent variables that cannot be directly observed or completely recorded. These latent variables may simultaneously affect multiple observed variables, thereby inducing statistical correlations among observed variables due to common causes. Causal discovery methods that do not consider the existence of latent variables tend to misinterpret such latent-variable-induced correlations as causal relationships, resulting in spurious causal edges, incorrect causal directions, and incomplete structure recovery. To improve the accuracy of causal structure recovery in scenarios with latent variables, the proposed method utilizes the non-Gaussian information contained in data under a linear non-Gaussian acyclic causal model with latent variables, and identifies both direct causal relationships among observed variables and undirected causal relationships affected by latent variables. Specifically, in the first stage, the method takes a complete undirected graph over observed variables as the initial structure, and gradually removes statistically independent variable pairs through conditional independence tests based on regression residuals, thereby obtaining an initial skeleton structure among observed variables. This stage can preferentially determine some reliable direct causal relationships among observed variables without explicitly modeling latent variables, providing a structural basis for subsequent latent variable detection and orientation of remaining edges. In the second stage, for variable relationships that remain unoriented and may be affected by latent variables, higher-order cumulants are introduced to characterize the higher-order statistical information in non-Gaussian distributions. Fourth-order joint cumulants are used to estimate the effect strength of latent variables on observed variables, and whether multiple observed variables are affected by a common latent variable is determined by testing whether the shared single latent component condition is satisfied. On this basis, the remaining unoriented edges among observed variables affected by latent variables are further oriented, and a causal structure containing both observed variables and latent variables is finally output. Theoretical analysis shows that, when the variables are generated by a linear non-Gaussian acyclic causal model, the noise variables are mutually independent, and the relevant identifiability conditions hold, the proposed method can identify causal structures with latent variables by using independence constraints and higher-order cumulant information. To verify the effectiveness of the method, the proposed method is compared with several representative methods under five simulated causal graphs and three sample-size settings. Experimental results show that the proposed method maintains relatively high precision in most scenarios, with precision at least 30% higher than that of existing methods. Meanwhile, the proposed method achieves the best results in both recall and F1-score, and its F1-score is greater than or equal to 75% under different sample sizes. In particular, in simulated scenarios with strong latent-variable effects and complex structures, the proposed method can effectively reduce the omission of true causal edges and improve the overall structure recovery ability. Ablation experiments show that both the local structure identification stage without latent-variable interference and the stage of latent variable detection and unoriented causal edge identification play important roles in final causal structure recovery. In the real-world financial data experiment, return data of multiple constituent stocks in the Hong Kong stock market are used to further verify the applicability of the proposed method in practical scenarios. Experimental results show that the proposed method can identify latent variable structures with clear economic meanings while maintaining reasonable sparsity of the causal graph, and can characterize a multi-level causal transmission relationship from market-wide macro factors to industry sectors and then to individual stocks. Compared with several representative methods, the proposed method avoids the problem of overly dense causal graphs and compensates for the limitations of some methods that lack global driving factors or do not explicitly characterize latent variable structures. Overall, the results of simulation experiments, ablation experiments, and real-world data experiments demonstrate that the proposed method can effectively utilize higher-order statistical information and conditional independence constraints, improve the accuracy, completeness, and interpretability of causal structure recovery in scenarios with latent variables, and provide an effective approach for causal analysis and intelligent decision-making in complex data environments.

Road waterlogging is one of the most direct risk forms affecting traffic operation and emergency dispatch during urban rainstorm flooding. Timely recognition of waterlogging levels is therefore important for road control, rescue scheduling, and public travel guidance. However, waterlogging level recognition from real road surveillance videos remains challenging. First, heavy rainfall, nighttime low illumination, vehicle occlusion, and water-surface reflection cause large intra-class variations. Second, high-risk categories such as heavy and severe waterlogging occur rarely, resulting in a long-tailed data distribution. Third, video clips from the same camera, road segment, or rainfall event may be highly similar, which can bias the evaluation of generalization if not handled properly. To address these problems, this paper proposes a few-shot road waterlogging level recognition method based on multimodal label constraints. The proposed method first extends the I-JEPA-style joint-embedding predictive architecture to video clips. Cross-frame spatiotemporal blocks in consecutive frames are used as context and prediction targets, and waterlogging-related target-block features are predicted in the representation space. This avoids redundant pixel-level reconstruction and encourages the encoder to focus on semantic cues related to waterlogging levels, such as road water layers, vehicle wheel submergence, reflective textures, and water-surface disturbances. Second, the CLIP text encoder is used to extract semantic features from waterlogging-level descriptions. Textual constraints are introduced during support prototype construction so that visual features are aligned with category semantics such as slight, mild, moderate, heavy, and severe waterlogging in a shared space, thereby alleviating unstable decision boundaries between adjacent water-depth levels under few-shot conditions. Third, dynamic time warping is introduced in the matching stage between query videos and support prototypes. By searching the optimal temporal alignment path through a frame-level cumulative distance matrix, the model can adapt to differences in video length, vehicle-water interaction moments, and key-frame locations, improving matching stability across different shooting rhythms. Two waterlogging level datasets are constructed from real road surveillance videos in Nanjing and Xiamen. Before splitting the data into training, validation, and testing sets, near-duplicate clips from the same source are removed, and camera IDs, road locations, and rainfall-event times are cross-checked to reduce the risk of leakage caused by highly similar clips appearing across different sets. For comparison, ResNet50, VGG-16, ViT-B, SwinT-B, HCL, OTAM, CPEA, CLIP-FSAR, Qwen2-VL, and Qwen2.5-VL are selected as baselines, and the year and publication venue of each method are added to the tables. In addition, AVF-MAE++ and TEAM, two officially published and open-source methods after 2025, are included for method-level comparison to clarify the difference between the proposed method and recent open-source video representation learning and few-shot video recognition approaches. Experimental results show that the overall accuracy of the proposed method reaches 96.4% and 95.0% on the Nanjing and Xiamen datasets, respectively, outperforming CNN, Transformer, few-shot video recognition, and multimodal large-model baselines. Considering the limited number of minority-class test samples, P/R=100% under a single split is not used as independent evidence of stability. Instead, five independent stratified re-splitting experiments are added so that heavy and severe waterlogging samples are covered multiple times in different test subsets. The standard deviation of overall accuracy is below 0.8% over five runs, indicating that the proposed method has good robustness for long-tailed few-shot categories and cross-city scenarios.

With the continuous progress in drone technology, the detection of small targets in drone aerial photography has garnered increasing scholarly attention. To tackle the challenges of low detection accuracy and high missed detection rates caused by small target scales, dense distributions, and complex backgrounds in drone aerial images, we propose PK-YOLO, a drone aerial small target detection model based on an improved YOLOv11n. Using YOLOv11n as the baseline model, we introduce targeted enhancements in three key areas: feature extraction, feature fusion, and loss function, with the objective of improving the detection accuracy and overall robustness of small targets in drone aerial scenarios. In the feature extraction phase, recognizing that YOLOv11n's feature pyramid structure, spanning layers P3 to P5, exhibits limited responsiveness to extremely small targets (less than 32×32 pixels), leading to missed and false detections, we incorporate an additional small target detection layer, namely the P2 detection layer. This layer leverages higher-resolution feature maps to preserve richer spatial position and edge detail information, facilitating a gradual fusion from high-resolution shallow features to semantically rich deep features, thereby significantly enhancing the network's capacity for extracting features from small targets. Furthermore, we design a C3k2_KSFA module that integrates the Kernel Selective Fusion Attention (KSFA) mechanism to replace the original C3k2 module in the backbone network. The KSFA mechanism draws inspiration from dynamic convolution and selective kernel networks, extracting multi-scale features by employing convolution kernels with various dilation rates and sizes in parallel. It then adaptively assigns optimal kernel weights to different spatial locations using a spatial-spectral selection mechanism, enabling the model to dynamically adjust its receptive field based on the actual size of the input target, thus improving feature representation and discrimination for targets of varying scales. During the feature fusion stage, traditional fixed upsampling methods struggle to accommodate the dramatic variations in target scales in drone images, often resulting in the loss of small target features and susceptibility to complex background interference. To address this, we replace the fixed upsampling module in the neck network with DySample dynamic upsampling. Utilizing a point sampling strategy, DySample adaptively adjusts sampling point positions according to the feature map content, effectively suppressing background interference while enlarging the feature map and reducing the loss rate of small target features. In terms of loss function optimization, acknowledging that the CIoU loss function employed by YOLOv11n is prone to interference from low-quality prediction boxes, leading to subpar localization accuracy, we introduce the Inner-IoU mechanism to enhance Wise-IoU v3, proposing the Inner-WIoU loss function. This function improves the fine-grained description of target location and shape while retaining the ability to focus on challenging samples through a dynamic non-monotonic frequency modulation mechanism and auxiliary bounding boxes, thereby effectively enhancing the localization accuracy of small targets. To validate the model's effectiveness, we conduct comparative experiments on three publicly available datasets: VisDrone2019, TinyPerson, and RSOD. The experimental results on the VisDrone2019 dataset reveal that, compared to the baseline model, the PKD-YOLO model achieves improvements of 5.0% and 3.3% in mAP@0.5 and mAP@0.5:0.95, respectively, along with a 5.0% increase in precision. On the more demanding TinyPerson dataset, these three metrics improve by 5.7%, 1.9%, and 7.4%, respectively, fully demonstrating the model's superior detection capabilities for small targets in complex environments. On the RSOD dataset, mAP@0.5 and mAP@0.5:0.95 reach 96.1% and 68.8%, respectively, representing improvements of 2.7% and 3.6% over the baseline model, indicating that the model maintains high localization accuracy while enhancing its ability to discriminate target detection confidence. Combining the experimental results from these three datasets, PKD-YOLO exhibits exceptional small target detection performance across various drone aerial scenarios, confirming the effectiveness and generalizability of our proposed improvement methods.

To address bottlenecks limiting LLMs in high-knowledge-density, logically-constrained domains like healthcare—namely the lack of self-correction in homogeneous multi-agent systems due to cognitive uniformity, and diminishing returns of computational investment—this study proposes a multi-agent validation fusion architecture based on heterogeneous model interaction. The architecture builds a differentiated hierarchical collaborative framework via prompt engineering, disrupting logical blind spots of single-model lineages. A high-performance large-scale model serves as the core decision-making agent, parsing medical records and generating preliminary diagnostic logic. Concurrently, heterogeneous-source models—with distinct output stylistics, conservative logical biases, and divergent training distributions—form an independent verification layer. This layer does not generate answers but conducts multi-dimensional audits through a “verification–issue enumeration–response” paradigm, scrutinizing inference pathways, factual consistency, and logical soundness. The theoretical foundation leverages cognitive divergences among heterogeneous models to enable productive logical conflict and rigorous cross-verification, enhancing output diversity, robustness, and medical factual rigor. In diabetes care evaluations, the system improves accuracy by 10% on multiple-choice questions, 8% on fill-in-the-blank tasks, and 22% on complex reasoning problems versus a single-agent baseline. Compared to Colacare, it shows superior logical consistency and inference stability. Notably, it surpasses Diabetica-7B (a deeply fine-tuned specialized model) without costly annotated data or fine-tuning. Experimental records show average inference cost of 61,881 tokens/query and latency of 71.74 seconds/query. Relative to specialized models, this system avoids prohibitive training costs and long data annotation cycles, while reducing architecturalcomplexity via a modular plug-in design. We conclude that for medical decision-making demanding high knowledge integration, introducing inter-model heterogeneity—not merely scaling homogeneous capacity—is critical for reliability. Optimizing heterogeneous-agent verification remediates logical vulnerabilities via inherent robustness without compromising large-model inference bounds. These findings offer a low-cost technical framework for medical AI and support the paradigm shift from “general intelligence” to “reliable professional intelligence.” Reduction achieved by removing phrases like “fundamental bottlenecks constraining the application,” “attributable to underlying,” “characterizing computational investment relative to performance gains,” “elaborately engineered,” “endowed with massive parameter counts and advanced generalized reasoning capabilities,” “innovatively,” “crucially,” “systematically,” “instantiate productive,” “concurrently improving model adaptability and stability across diverse clinical scenarios,” “specialized experimental evaluations targeting,” “through comprehensive comparative analysis with the state-of-the-art policy model,” “more prominently,” “professionally specialized,” “subjected to deep domain-specific fine-tuning,” “regarding the trade-off between engineering feasibility and resource consumption,” “protracted,” “overall,” “our research conclusions establish that,” “the successful empirical validation of this model demonstrates that,” “these findings furnish not only...but also,” and some redundant adjectives.

Anchor-based bipartite graphs can approximate sample relationships with low computational cost, and have been widely used for graph construction in large-scale multi-view clustering. However, most existing methods generate anchors by random sampling, K-Means centers, or static dictionaries. The anchor positions are usually fixed before graph learning. For data with non-spherical clusters, ring-shaped clusters, elongated clusters, and large density differences among clusters, static anchors are difficult to approach local density peaks and curved cluster boundaries. This may lead to anchor mismatch. The sample-anchor bipartite graph constructed in this way may contain redundant or unreliable connections, which affects cross-view structure fusion and reduces the stability of clustering results. To address these problems, a Multi-view Bipartite Graph Clustering method guided by Mean-Shift (MBGC-MS) is proposed. The aim of the method is to maintain the high efficiency of anchor graphs, improve the adaptive representation ability of anchors for complex data distributions, and integrate graph learning and clustering assignment into a unified optimization process. First, an anchor set is initialized in each view. Mean Shift is then used to iteratively update the anchors along the kernel density gradient, so that the anchors are changed from Euclidean center representatives to local density modes. After that, the local bandwidth is estimated according to the distance from each sample to its neighboring anchors. Only the neighboring anchor relationships of each sample are retained to construct a sparse and density-aware sample-anchor bipartite graph. Second, the bipartite graph of each view is decomposed into a shared consistency graph and view-specific complementary graphs. The consistency graph is used to describe the clustering structure jointly supported by different views, and a nuclear norm constraint is imposed to enhance its low-rank property and structural compactness. The complementary graph is used to preserve the discriminative local information that deviates from the consensus in each single view, and a sparsity constraint is imposed to suppress noise and redundant connections. Third, view weights are adaptively learned according to the reconstruction error, so as to avoid the information interference caused by equal-weight fusion of all views. On this basis, the consistency graph and complementary graphs are fused to construct a unified sample-anchor bipartite graph and an augmented Laplacian matrix. A spectral trace constraint is further imposed to make the unified graph tend to form a given number of connected components. The proposed model adopts a block coordinate descent strategy to alternately update the clustering indicator matrix, consistency graph, complementary graphs, and view weights. Each subproblem can be solved by eigen-decomposition, soft-thresholding operator, singular value thresholding, or quadratic programming. In this way, one-step optimization of graph structure learning and clustering assignment is achieved. Experiments are conducted on four multi-view datasets, including Handwritten, BBC-sport, MSRC_v1, and Caltech101-7. ACC, NMI, and F-score are used as evaluation metrics. The results show that the proposed method obtains the best performance on 8 out of 12 metrics. On Handwritten, the highest NMI is obtained, reaching 0.949. On BBC-sport, the ACC and F-score reach 0.971 and 0.933, respectively. On MSRC_v1, the ACC and F-score reach 0.966 and 0.931, respectively. On Caltech101-7, the ACC, NMI, and F-score reach 0.711, 0.521, and 0.484, respectively, and all three metrics are the best. Ablation experiments further show that Mean Shift anchor updating, consistency-complementarity decomposition, and Laplacian spectral trace constraint all contribute to performance improvement. On Caltech101-7, the ACC, NMI, and F-score of the complete model are higher than those of the model using only the Mean Shift module, whose corresponding values are 0.534, 0.362, and 0.348. The comparison of different anchor strategies shows that the three metrics of Mean Shift anchors are 0.711, 0.521, and 0.484, which are clearly higher than those of K-Means anchors, namely 0.576, 0.395, and 0.456, and also higher than those of random sampling anchors, namely 0.308, 0.260, and 0.320. The convergence experiments show that the objective function usually decreases rapidly in the first 5–10 iterations and becomes stable within 20 iterations. Complexity analysis and running time experiments show that, when the number of anchors and iterations are fixed, the running time of the algorithm grows approximately linearly with the sample size. In summary, the proposed method can calibrate anchor positions by using density information and enhance the representation ability of bipartite graphs for complex cluster structures. It also considers multi-view consistency, complementarity, and one-step clustering optimization. Therefore, it shows good clustering performance, stability, and scalability on non-spherical and uneven-density data.

This paper proposes a distributed task allocation method based on overlapping coalition formation game, to address the collaborative task allocation problem of heterogeneous multi-functional cleaning robot clusters in photovoltaic (PV) power stations under multiple constraints. First, this paper constructs a multi-constraint cooperative task optimization model of heterogeneous robot clusters for PV operation and maintenance (O&M). It systematically analyzes the resource adaptation mechanism between heterogeneous robots and diverse O&M tasks. It designs a matching mechanism between PV module pollution levels and robot cleaning capabilities. It establishes a comprehensive utility function integrating comprehensive resource matching degree, recovered power generation loss, total O&M cost, maximum completion time and redundancy penalty mechanism. The objective function of this task allocation problem is constructed to maximize the total coalition utility. Combined with engineering practice, it introduces constraint conditions including task priority, task uniqueness and safety distance. It completes the refined mathematical modeling of robot cluster O&M tasks. It significantly improves the resource matching degree and O&M scheduling adaptability between robots and tasks. Second, this paper constructs an overlapping coalition formation game framework based on the bilateral mutual benefit criterion. This framework converts the global combinatorial optimization problem into a distributed coalition partition problem. Relying on the defined preference relation and exchange operation, it supports robots to realize cross-coalition dynamic cooperation and independently form an overlapping coalition structure. This paper further proposes a bilateral mutual benefit exchange criterion to describe the mapping relationship between individual selfish profit-seeking and global system optimization. The criterion significantly improves the flexibility of resource scheduling. It effectively coordinates individual rationality and overall system utility. This game is proven to be a potential game. When robots change the resource allocation structure to maximize their own utility, the difference of the objective function is consistent with the difference of the utility function. And there exists at least one Nash equilibrium, namely a stable coalition structure, in the game framework. Finally, this paper designs an overlapping coalition formation algorithm integrating preference gravity guidance and disturbance mechanism. The algorithm introduces a preference gravity mechanism to provide directional guidance for robots to select coalitions. It uses the marginal revenue evaluation mechanism and the proposed double tabu list to eliminate inefficient coalitions and avoid invalid searches. Combined with the co-designed dynamic resource adjustment strategy and random disturbance mechanism, it effectively gets rid of local optimum. This algorithm is proven to converge to a T-stable state within a finite number of iterations. Its complexity is analyzed to be within a controllable range. When solving the task allocation mathematical model composed of the task optimization model and the game framework, the algorithm can meet the task allocation requirements of heterogeneous robot clusters in complex and diverse task cooperation scenarios. It can quickly converge to a high-quality Nash equilibrium solution. Simulation experimental results show that the proposed method has remarkable real-time performance and stability on the premise of ensuring the quality of the global solution. Ablation experiments verify that each improvement strategy can comprehensively improve the comprehensive performance of the algorithm. In experiments verifying algorithm solution quality and O&M indicators, compared with the baseline algorithm, the proposed algorithm increases the recovered power generation loss by 6.63%, reduces the total O&M cost by 2.55%, and shortens the average operation time by 13.29%. In performance comparison experiments under scale changes, all indicators of the proposed algorithm are optimal under different numbers of robots and tasks. In real-time and stability experiments, the average running time of the proposed algorithm does not exceed 6.09s. Its standard deviation and Wilcoxon rank-sum test results have statistical significance. It can formulate economical and efficient task allocation schemes. It provides effective technical support for the refined O&M of PV power stations.

During the feature extraction process of point cloud retrieval, mainstream methods typically form global shape descriptors through hierarchical aggregation of local neighborhood features. Their perception of object structural information mainly relies on indirect inference from the spatial distribution of surface points, lacking direct exploitation of explicit structural priors such as point cloud skeletons. To address the problem of insufficient structural information utilization in existing point cloud retrieval methods when extracting global shape features, this paper proposes a skeleton‑prior enhanced structure‑aware point cloud retrieval network. The method first introduces point cloud skeletons as structural priors, enhances the structural expressiveness of point cloud features through a dual mechanism of explicit fusion and implicit guidance, and designs an adaptive feature aggregation module to aggregate multi‑scale features, thereby forming the final global descriptor. Specifically, the proposed method consists of two modules. First, a dual‑branch feature fusion module. This module extracts skeleton point clouds from the input point cloud, and then employs two independent PointNet++ branches to extract multi‑scale local features from the original point cloud and the skeleton, respectively. At each scale, using skeleton features as keys and values and point cloud features as queries, a multi‑scale cross‑attention mechanism is adopted to weightedly integrate the skeleton structural information into the point cloud features. Meanwhile, a contrastive learning task is constructed: the combination of a cropped point cloud and the complete skeleton serves as the anchor, the combination of the complete point cloud and the skeleton serves as the positive sample, and samples from other categories within the batch serve as negative samples. The contrastive loss implicitly guides the model to learn structural consistency, forming an“explicit fusion + implicit guidance” dual structural enhancement mechanism. Second, a multi‑scale local adaptive aggregation descriptor (MVLAAD) module. This module consists of the Vector Local Adaptive Aggregation Descriptor (VLAAD) and multi‑scale aggregation enhancement. Based on a lightweight Transformer decoder, VLAAD takes the local feature sequence of the input point cloud as keys and values, and the initial generic cluster centers as queries. Through iterative updates via multi‑layer cross‑attention, it dynamically generates personalized cluster centers that adapt to each input sample. A momentum update strategy further combines the generic centers with the personalized centers, balancing adaptability and stability. Subsequently, using the updated cluster centers, soft assignment weights are computed for the refined multi‑scale features and the residuals are aggregated to produce global descriptors at three scales. Finally, a gating mechanism enhances the features and outputs the compact global descriptor. In addition, during training a dynamic weight adjustment strategy is adopted, combining classification loss, triplet loss, and contrastive loss. The training emphasizes contrastive learning in the early stage and shifts focus to triplet loss in the later stage, thereby reinforcing both structure perception and discriminative learning. Experimental results show that the proposed method achieves 82.6% mAP on the ModelNet40 dataset, outperforming the state‑of‑the‑art method CF3D by 1.3%, and attains 84.6% mAP on ShapeNet, exceeding existing methods. Ablation studies verify the effectiveness of each module: the baseline PointNet++ achieves 62.0% mAP, while the full model increases the performance to 82.6%. A lightweight version reduces the number of parameters to 20.34M and further improves mAP to 84.7%. Robustness experiments demonstrate that the method performs robustly under moderate sparsity, low‑level noise, and mild occlusion, but its performance degrades significantly under extreme degradation conditions. In summary, the proposed skeleton‑prior enhanced structure‑aware point cloud retrieval network addresses the deficiency of structural information utilization in existing methods through explicit skeleton prior fusion and implicit contrastive learning guidance. The designed MVLAAD module dynamically generates personalized cluster centers, thereby improving the discriminability of the global descriptor.