Low-resource scenarios in natural language processing often face the challenge of data scarcity, which is particularly prominent in the field of lighting intelligent customer service. This domain is characterized by dense professional terminologies, sparse specification parameters, and redundant combinations of services and products, making it difficult to obtain high-quality annotated intent data. Existing mainstream data augmentation methods, such as synonym replacement and back-translation, struggle to improve semantic diversity or resolve ambiguous intent boundaries, while generative augmentation methods based on large language models are prone to issues such as intent drift and entity errors, with insufficient domain adaptability. To address these problems, this paper proposes a Guided Constraint Deduplication Data Augmentation method (GCDDA). This method constructs a three-stage collaborative framework consisting of few-shot example guidance, multi-dimensional constraint filtering, and semantic similarity deduplication. Specifically, few-shot example guidance uses a small number of high-quality seed samples as references to enable rapid adaptation of the model to the generation paradigm in the lighting domain, compensating for the lack of large-scale manual annotation. Multi-dimensional constraints impose fine-grained control over the generation process at three levels—intent, entity, and scenario language—effectively suppressing common intent drift and entity violations in generative augmentation. Semantic similarity deduplication employs an LLM-driven duplication scoring mechanism to forcibly eliminate redundant samples that exhibit "same logic, different wording," fundamentally mitigating the deficiencies of traditional methods in semantic diversity and ambiguous intent boundaries. Through the synergy of these three components, the method can generate quality-controllable, low-redundancy intent samples in batches without fine-tuning the generative language model. Experimental results demonstrate that on the self-built Light-Intent dataset, with BERT-base-Chinese as the downstream classification model, the accuracy of the 5-shot baseline is only 40.54%, which increases to 63.34% after GCDDA augmentation, while the F1 score improves from 40.33% to 63.66%. Moreover, the augmented performance (5+15-shot) outperforms the same-scale real data (20-shot) with 60.48% accuracy and 60.88% F1. On three public datasets, ATIS, Kuake-qic, and a_intent, GCDDA consistently outperforms traditional augmentation methods such as back-translation and synonym replacement, as well as generative methods including ZeroGen, AugGPT, and PromptMix under the same sample size conditions. The proposed method requires no fine-tuning of large language models and can be flexibly transferred.

As the volume of scientific literature continues to grow, researchers' information needs are shifting from document-level relevance retrieval toward the acquisition of fine-grained scientific knowledge, including research problems, technical methodologies, experimental settings, and supporting evidence for scientific conclusions. Existing scientific paper representation methods mainly rely on abstracts or fixed-length text chunks, making it difficult to simultaneously preserve semantic completeness and evidence traceability. Moreover, using a single large language model to construct such representations is constrained by limited context windows and insufficient factual consistency, resulting in compromised quality and reusability. To address these challenges, this paper proposes a multi-agent collaborative framework for fine-grained semantic mining of scientific papers. Specifically, it adopts a three-layer architecture consisting of control, state, and operation layers. A deterministic task orchestrator manages stage transitions and feedback-driven rollback. A shared semantic evidence space maintains intermediate objects with explicit types, provenance, and dependencies. Four functional agents perform text enhancement, atomic fact extraction, semantic organization, and verification, progressively transforming multimodal evidence into semantic representations. The verification agent produces structured reports identifying error types and affected objects, enabling the orchestrator to perform local corrections and recomputation along dependency chains. Based on the generated semantic representations, we further develop a semantic representation-based framework for academic paper retrieval and knowledge services to facilitate scientific knowledge discovery. Experiments conducted on a self-constructed arXiv dataset demonstrate that the proposed method achieves the best performance in golden-fact coverage, information amount, and faithfulness, while maintaining a relatively low level of cross-representation semantic redundancy. Furthermore, the proposed semantic representation effectively improves academic paper retrieval and knowledge services by providing researchers with more relevant and trustworthy results. Qualitative cases further show that, under comparable token budgets, semantic-representation-based services improve evidence coverage and the specificity and query alignment of answers and recommendation rationale. These experiments validate the effectiveness of our method in supporting scientific knowledge discovery and offering a new technical paradigm for trustworthy scientific knowledge organization and intelligent knowledge service.

Heat-sensitive moxibustion can enhance therapeutic outcomes by eliciting propagated sensation along meridians (PSM) and facilitating the arrival of qi at the affected site. It is widely used for analgesia, functional regulation, and chronic disease rehabilitation. However, the technique is highly experience-dependent and individualized. Current assessments of practitioner skill rely largely on direct expert observation and subjective judgment. Consequently, evaluation criteria remain poorly defined, scores are susceptible to rater bias, and technical performance is difficult to quantify objectively. Moreover, practitioners must dynamically adjust the moxibustion location, stimulation area, and movement rhythm according to patients’ heat-sensitization responses. Assessment methods based on trajectory-shape matching or static-template matching may therefore incorrectly classify clinically appropriate individualized adaptations as performance errors. To address these limitations, this study proposes an objective scoring method that integrates expert knowledge with multidimensional kinematic features, drawing on acupoint heat-sensitization theory and clinical practice. Three-dimensional motion trajectories of the moxa stick were analyzed using a multidimensional framework encompassing spatial control, movement rhythm, and procedural timing. Technique-specific assessment strategies were developed to accommodate differences in movement patterns among moxibustion techniques. A scoring model based on a modified cumulative distribution function was then constructed. A scale parameter was introduced to control the model’s sensitivity to deviations in individual metrics, allowing feature deviations expressed in different units to be mapped to normalized component and composite scores. The method was validated using an in-house dataset comprising 108 motion-trajectory samples representing four techniques: circling moxibustion, sparrow-pecking moxibustion, stationary suspended moxibustion, and round-trip meridian moxibustion. The mean absolute errors (MAEs) between the model-generated scores and expert reference scores ranged from 0.037 to 0.059, while the corresponding intraclass correlation coefficient (ICC) values, calculated using ICC(3,1), ranged from 0.757 to 0.860, indicating good scoring accuracy and reliability across all four techniques. The proposed method also achieved lower overall scoring errors and more stable performance than several comparison methods based on time-series matching or trajectory similarity. These findings indicate that the method can objectively quantify adherence to moxibustion procedural standards while providing component-level scores with clear practical meaning, thereby offering a reliable and interpretable digital assessment framework for practical training, skill assessment, and performance feedback in heat-sensitive moxibustion.

In multimodal magnetic resonance imaging (MRI) brain tumor segmentation, different MRI modalities make markedly different contributions to tumor subregions. Conventional channel concatenation or simple feature addition cannot adequately characterize modality-specific information, and existing segmentation networks generally lack explicit explanations of modality contributions. These limitations affect the accuracy, robustness, and interpretability of multimodal MRI brain tumor segmentation. To address these issues, a membership-based multimodal MRI brain tumor segmentation network is proposed. The network employs a Private-Shared Encoder to separately extract modality-specific pathological features and cross-modality shared anatomical representations, thereby alleviating multimodal feature entanglement. In the Fuzzy Fusion Layer, a Data-Prior Co-driven Membership Module (DPC-MM) integrates data-driven responses with medical-prior guidance to generate voxel-wise modality memberships, which characterize the dynamic contributions of different MRI modalities across tumor regions. The Fuzzy Feature Aggregation Module (FFAM) then weights the private features of each modality according to their memberships and residually fuses them with shared features, enhancing the representation of tumor spatial heterogeneity and fuzzy boundaries. The network outputs multimodal membership maps and brain tumor segmentation results. Five-fold cross-validation on the BraTS2021 dataset yielded Dice scores of 0.9223±0.0061, 0.8873±0.0064, and 0.8593±0.0110 for whole tumor (WT), tumor core (TC), and enhancing tumor (ET), respectively, with a mean Dice of 0.8897±0.0069. The mean 95th-percentile Hausdorff distance (HD95), average surface distance (ASD), and expected calibration error (ECE) were 5.18±0.53 mm, 3.15±0.54 mm, and 0.1442±0.0073, respectively. Further multi-method comparison experiments were conducted on the BraTS2021 and BraTS2018 HGG datasets. The proposed method achieved Mean Dice values of 0.9037 and 0.8393 and Mean HD95 values of 4.7764 mm and 5.1885 mm, respectively, demonstrating good overall segmentation performance. The results show that the proposed network achieves region-adaptive multimodal feature fusion through voxel-wise modality memberships, while maintaining high segmentation accuracy, boundary quality, and probability calibration. In addition, the generated multimodal membership maps visually reveal differences in modality utilization across spatial locations and tumor subregions, providing a quantitative and visual modality-level representation for model interpretation.

Advances in generative artificial intelligence and localized content editing have extended deepfakes from whole-sequence synthesis to targeted modifications of speech and visual content. Video-level detection can obscure brief anomalies within predominantly authentic content, whereas frame-level detection often produces fragmented predictions. Neither paradigm directly identifies the onset, offset, or duration of manipulated segments. Deepfake temporal localization addresses this limitation by detecting suspicious segments in untrimmed audio, video, or audio-visual streams and returning verifiable temporal intervals. We searched Google Scholar, Web of Science, IEEE Xplore, the ACM Digital Library, CVF Open Access, and arXiv, and included studies on temporal localization methods and the associated datasets, evaluation protocols, and enabling techniques. Taking changes in detection evidence and localization objectives as the organizing thread, we trace the development of the field from local artifact detection and physiological or behavioral cue analysis, through joint audio-visual detection and forged-segment proposal generation, to boundary refinement. Methods are further organized by input modality into audio, visual, and audio-visual approaches. For multimodal methods, we establish a layered framework comprising forgery evidence, cross-modal interaction, and temporal segment decoding. Our analysis shows that current research is largely concerned with the reliability of forgery evidence and the precision of temporal boundaries. Contextual modeling and self-supervised learning can reduce dependence on specific generators and known manipulation types, but excessive feature fusion may suppress short-lived anomalies, while naturally occurring audio-visual asynchrony can trigger false alarms. Multimodal interaction exploits complementary acoustic and visual information, yet remains sensitive to modality quality, cross-modal correspondence, and missing inputs. Among peer-reviewed studies published to date, the best reported AP@0.50 scores on the official LAV-DF and AV-Deepfake1M test sets exceed 99% and 94%, respectively, whereas the corresponding AP@0.95 scores reach only approximately 70% and 51%. Existing methods can therefore identify the approximate extent of manipulated content, but accurate boundary recovery under strict overlap criteria remains difficult. Cross-dataset and cross-lingual evaluations further indicate that performance is affected by generators, languages, editing operations, segment duration, and biases introduced during dataset construction. Strong in-domain results therefore do not necessarily imply stable generalization to real-world media. Future research should strengthen dataset auditing and evaluation under realistic media dissemination conditions, establish more rigorous cross-dataset, cross-generator, and cross-lingual protocols, and incorporate modality-quality assessment, uncertainty estimation, and modality-specific evidence preservation to improve the reliability, interpretability, and practical deployability of temporal localization systems.

Aiming at the problem that the existing multivariate time series prediction methods based on diffusion model have insufficient conditional information representation ability and large computational cost of long sequence modeling, this paper proposes a diffusion model (Wavelet-Conditioned Diffusion Model with Mamba, WCDM-M) that combines wavelet condition enhancement and lightweight Mamba architecture. The wavelet conditional enhancement network is designed in the WCDM-M model. The multi-scale time-frequency analysis of the historical time series is carried out by continuous wavelet transform, and the periodicity, local variation and transient characteristics at different time positions and scales are expressed explicitly. Then, using the convolutional coding structure composed of two layers of InceptionV1 modules, the time-frequency features under different receptive fields are extracted through multi-scale convolution kernels and pooling branches, and the scale dimension is compressed through deep convolution. The conditional vector is generated by linear mapping, which provides a multi-scale time-frequency prior for the diffusion reverse denoising process. In the denoising network, a lightweight time-series modeling backbone based on the selective state space model Mamba is designed. By removing the Conv1D module in the native Mamba, the linear mapping is used to simplify the redundant projection, and the L-Mamba module is constructed to achieve efficient modeling of long-distance timing dependencies. At the same time, the wavelet condition vector and diffusion step are embedded into the joint input multi-layer perceptron to generate feature scaling and offset parameters, and the adaptive feature modulation of different diffusion stages is realized by the condition layer normalization mechanism. Experimental verification is performed on five public time-series datasets of ETTm1, Electricity, Traffic, Weather and Exchange. The results show that in the 96-step prediction task, the MSE of WCDM-M on the five datasets is 0.3212,0.1767,0.4196,0.0071 and 0.0410, respectively, which is better than the optimal results in the corresponding comparison model. Among them, the MSE of the Weather dataset is reduced to 0.0071, indicating that the model can effectively capture high-frequency fluctuation pattern; the MSE of the Exchange data set reaches 0.0410, which verifies the predictive capability of the model for non-stationary time series. The results of ablation experiments show that after removing the wavelet conditional network, the average MSE of all data sets and prediction lengths is increased by 26.4%; after using single-scale wavelet or removing Inception multi-branch coding, the average MSE is increased by 15.1% and 10.5% respectively, which proves that multi-scale time-frequency conditions and multi-receptive field feature extraction play an important role in model performance improvement. In addition, the WCDM-M parameter is 11.2M, which is lower than 18.5M of the Transformer denoising version WCDM-T, and the single-round training time is 89.7ms, which is 26.6% lower than WCDM-T. The WCDM-M model combines the wavelet time-frequency condition enhancement with the lightweight Mamba denoising network, which reduces the computational complexity and improves the prediction ability of the diffusion model for periodic changes, local mutations and non-stationary dynamic patterns. It provides an effective method for the prediction of complex and variable time series with both accuracy and efficiency.

This paper proposes a Prediction Prior and Uncertainty Refinement Network (PURNet) for camouflaged object detection (COD) to address false positives, missed detections, and inaccurate boundary localization in existing methods. The proposed method adopts a Pyramid Vision Transformer (PVT) as the backbone to extract multi-scale features with rich semantic information. Based on these features, a Dual Prior-guided Attention Module (DPAM) is designed. The module uses the prediction result of the current stage as a prediction prior and adaptively enhances features from both spatial and channel perspectives. Specifically, the spatial prior guides the features to focus on potential object regions and suppress irrelevant background responses. The channel prior dynamically generates channel attention parameters according to the current prediction state. It reweights target-related semantic features to enhance the feature representation of camouflaged objects. To further evaluate the reliability of prediction results, an Uncertainty Generation Module (UGM) is introduced. The module performs probabilistic modeling and reparameterization sampling to generate multiple prediction samples. It estimates uncertainty from the variance of these samples. Furthermore, a Group Uncertainty-Guided Residual Refinement (GUGR) Module is proposed. The module adopts different levels of uncertainty-guided strategies to progressively integrate uncertainty information with features and refine unreliable regions. It enables the network to adaptively adjust prediction results according to their reliability. Experiments are conducted on three public COD datasets, namely CAMO-Test, COD10K-Test, and NC4K. Four metrics are used for comprehensive evaluation, including Structure Measure, Enhanced Measure, Weighted F-measure, and Mean Absolute Error. On CAMO-Test, PURNet achieves 0.856, 0.920, 0.804, and 0.050, respectively. On COD10K-Test, the corresponding results are 0.839, 0.914, 0.730, and 0.029, respectively. On NC4K, the corresponding results are 0.870, 0.926, 0.812, and 0.037, respectively. Complexity and inference efficiency experiments show that PURNet requires 10.77 GFLOPs and 25.82 M parameters and achieves 78.6 FPS. Overall, the experimental results demonstrate that PURNet effectively integrates prediction priors and uncertainty information and achieves more accurate object localization and segmentation in complex backgrounds.

Weak boundaries, low-contrast lesions, and imaging artifacts often cause localization shifts, missed regions, and over-segmentation in two-dimensional medical image segmentation. Meanwhile, patch-level semantics from a frozen vision foundation model do not fully correspond to pixel-level medical structures. To address these problems, a DINOv3 prior-guided parameter-efficient network, termed HiFi-DinoMECNet, is proposed. The network employs frozen DINOv3-S/16 as the semantic encoder and trains only the spatial adaptation components and decoder, retaining self-supervised semantic priors while reducing the number of parameters updated for the target task. First, patch tokens are extracted from the 3rd, 6th, 9th, and 12th Transformer blocks. Class and register tokens are removed, and four pseudo-multiscale semantic interfaces are constructed through token-to-map transformation and scale alignment, allowing features at different semantic depths to interact progressively with the spatial branch. Second, an encoder-side semantic-spatial fidelity injection module (SSFI) uses shallow convolutional features to generate spatial queries, predict local sampling offsets and aggregation weights, and perform spatially guided local sampling on frozen semantic feature maps. This process strengthens semantic responses around weak boundaries, textures, and local structures. Third, a multi-scale efficient collaborative attention decoder (MECAD) successively performs upsampling calibration, semantic-guided gating, and multi-scale contextual aggregation to suppress shallow noise and spurious boundaries and progressively recover high-resolution segmentation results. Experiments are conducted on four public two-dimensional medical image datasets: ISIC 2017, ISIC 2018, BUSI, and DSB18. All methods are independently trained under fixed data splits using the same five random seeds, and results are reported as mean ± standard deviation. Key hyperparameters are selected on an independent validation set, while the test set is used only after the final model configuration has been fixed. HiFi-DinoMECNet achieves Dice similarity coefficients of 91.04%, 91.08%, 83.18%, and 93.12% on the four datasets, respectively. On the ISIC 2018 test set, it obtains a mean intersection over union of 83.18%±0.12%, a Dice similarity coefficient of 91.08%±0.08%, a BoundaryF1 score of 85.42%±0.11%, and an HD95 of 8.34±0.15 pixels. Ablation experiments show that SSFI and MECAD contribute to encoder-side semantic-spatial adaptation and decoder-side feature reconstruction, respectively, and jointly improve region overlap and boundary quality. At an input resolution of 224×224, the model contains 27.64 M total parameters and 6.28 M trainable parameters, requires 5.12 G FLOPs, and achieves 57.20±1.25 FPS. These results indicate that HiFi-DinoMECNet provides a practical balance between segmentation accuracy and adaptation cost and offers a feasible parameter-efficient transfer strategy for applying frozen vision foundation models to two-dimensional medical image segmentation.

[Objective] To address blurred lesion boundaries, high-frequency detail loss, and insufficient global modeling in medical image segmentation, we propose FreqUNet, a frequency–spatial dual-domain collaborative network for improving the segmentation accuracy and contour continuity of weak-boundary lesions. [Methods] A frequency–spatial dual-branch encoder was constructed based on U-Net, with a Multi-scale Frequency Filtering Module (MFFM) embedded in the first three encoder stages. The amplitude and phase spectra were decomposed using the two-dimensional Fast Fourier Transform. The amplitude spectrum was divided into low-frequency structural and high-frequency detail components, which were enhanced using a squeeze-and-excitation module and parallel 3×3, 5×5, and 7×7 multi-scale convolutions, respectively. A learnable parameter γ, initialized to 0.2, was introduced to adaptively regulate high-frequency energy. At the bottleneck layer, a Frequency-Space Dual Attention (FSDA) module was designed to fuse frequency-domain channel attention and spatial multi-scale atrous convolution features. A joint loss consisting of binary cross-entropy, Dice, and boundary losses was adopted with a weight ratio of 0.6:0.3:0.1. Comparative experiments were conducted under unified settings on the BUSI, ISIC, Kvasir, and CVC-ClinicDB datasets against U-Net, U-Net++, TransUNet, FAT-Net, Swin-UNet, H2Former, FFC, SFUNet, and PFESA. Ablation, generalization, and failure-case analyses were also performed. [Results] FreqUNet achieved an mDice of 0.8758 on BUSI, representing a 10.39 percentage point improvement over U-Net, with an HD95 of 13.34 pixels. The mDice values on ISIC, Kvasir, and CVC-ClinicDB were 0.8990, 0.8889, and 0.8886, respectively. FreqUNet achieved the best mDice and HD95 on BUSI, ISIC, and Kvasir, and generally outperformed the other comparison models except FAT-Net on CVC-ClinicDB. Bonferroni-corrected Wilcoxon signed-rank tests showed that FreqUNet significantly outperformed all comparison models on BUSI (p<0.0056), whereas its differences from FAT-Net on ISIC, Kvasir, and CVC-ClinicDB were not statistically significant. The model contained 35.84M parameters, involved 58.63 G floating-point operations, and achieved an inference speed of 144.81 FPS. Failure-case analysis showed that the model may still produce local false negatives, false positives, or boundary discontinuities in samples with extremely low contrast, strong noise, adhesive lesions, or tiny lesions. Three cross-dataset experiments demonstrated stable generalization performance on ISIC 2018→PH2, CVC-ClinicDB→ETIS-LaribPolypDB, and BUSI→UDIAT. Ablation experiments confirmed the synergy between adaptive amplitude enhancement and dual-domain attention: the complete model improved mDice by 9.66 percentage points over the baseline, and adding boundary loss reduced HD95 from 15.62 to 13.34 pixels. [Conclusion] FreqUNet improves structural representation and boundary delineation for weak-boundary lesions through frequency-domain decomposition, differentiated amplitude enhancement, and dual-domain adaptive fusion. It provides a useful approach for automatic segmentation of breast ultrasound, dermoscopic, and endoscopic images, although further optimization is needed for complex noise and extremely weak-boundary cases.

With the continuous migration of Industrial Internet of Things (IIoT) data to cloud platforms, achieving dynamic and trustworthy data access control while ensuring data confidentiality has become a critical issue in cloud data sharing scenarios. Existing access control schemes can provide fine-grained authorization based on user attributes, but they usually directly transform attribute satisfaction into data acquisition capability, resulting in strong coupling between authorization and data acquisition. When user states change, access policies are updated, or authorization relationships become invalid, previously issued authorization materials may still remain usable, thereby causing invalid authorization risks. In addition, existing schemes mainly rely on static authorization results and have difficulty continuously perceiving changes in user states, policy states, and data states, making it difficult to ensure trustworthy access throughout the data access process. To address these problems, an authorization and data acquisition decoupled cloud data access control scheme is proposed. First, an authorization and data acquisition decoupled access control structure is constructed. Attribute satisfaction is restricted to the qualification for recovering the data encryption seed rather than directly granting data acquisition capability, thereby separating authorization decisions from the actual data recovery process. By controlling the recovery of the data encryption seed, changes in user states and access policies can continuously affect subsequent data acquisition, thus reducing the security risks caused by the long-term validity of historical authorization materials. Second, a two-stage continuous verification mechanism oriented to data acquisition is designed. During the authorization stage, access conditions are verified by combining user identity, attribute policies, and on-chain state information. During the data acquisition stage, the authorization state, policy version, and data integrity are further verified through the generation of partial decryption results and the validation of the data encryption seed, thereby continuously constraining the access process. Furthermore, to address the conflict between anonymous user access and accountability for abnormal behaviors in cloud environments, an anonymous accountability tracing mechanism bound to the access process is constructed. Pseudonymous credentials, request signatures, and on-chain access records are used to hide the real identities of legitimate users while supporting abnormal access tracing and responsible-user revocation. Performance evaluation results show that, compared with similar access control schemes, the proposed scheme reduces ciphertext generation time and data decryption time by 68.9% and 84.3%, respectively. The additional storage overhead introduced by access control remains at the KB level, and the average latency of basic on-chain operations is below 56 ms, demonstrating that the proposed scheme can satisfy the efficiency requirements of cloud data access scenarios. In terms of security, theoretical security analysis is first conducted, followed by ablation experiments based on the Comprehensive Attack Blocking Rate (CABR) to evaluate the effectiveness of the proposed security mechanisms. For multiple attacks, including invalid authorization, policy version anomalies, ciphertext replacement, and binding anomalies, the ablation results show that removing the authorization and data acquisition decoupling mechanism, dynamic on-chain state verification mechanism, and BKD binding mechanism reduces the CABR to 91.70%, 66.15%, and 75.50%, respectively. To further distinguish the security boundaries of individual mechanisms, independent-module experiments show that the schemes containing only the authorization decoupling mechanism, dynamic on-chain state verification mechanism, and BKD binding mechanism achieve CABRs of 33.33%, 66.97%, and 58.51%, respectively. These results indicate that the three mechanisms provide independent protection in access-process constraint, dynamic state maintenance, and key-data binding, respectively, while the complete scheme achieves more comprehensive security protection through their cooperation. In conclusion, the proposed scheme realizes continuously trustworthy access control for cloud data sharing by decoupling authorization from data acquisition and integrating dynamic on-chain state verification with key-data binding.

To address the challenges of small object size, feature degradation, strong background interference, and the difficulty of balancing detection accuracy with lightweight deployment in UAV aerial imagery, this paper proposes BEL-YOLO, a lightweight small-object detection model with coordinated optimization based on YOLO11s. The model is specifically optimized in terms of multi-scale feature fusion, upsampling detail reconstruction, high-resolution prediction, and bounding-box regression. First, a Bidirectional Feature Pyramid Network (BiFPN) is introduced into the neck to enhance the interaction between shallow spatial details and deep semantic information through bidirectional weighted cross-scale feature fusion, thereby improving multi-scale feature representation for small objects. Second, a Gated Shift Channel Efficient Up-Convolution Block (GSC-EUCB) is proposed, which integrates channel rearrangement, shift-channel mixing, and residual gated fusion to strengthen cross-channel interaction and local detail reconstruction while alleviating weak-feature loss during upsampling. Subsequently, Detect_LSCD-P2 is designed based on the LSCD shared-convolution mechanism. The detection scales are reconstructed from P3/P4/P5 to P2/P3/P4, allowing high-resolution features to enhance tiny-object perception while cross-scale parameter sharing reduces redundancy in prediction branches. Finally, a Normalized Wasserstein Inner-Shape Loss (NWIS-Loss) is proposed by jointly modeling the distribution-distance constraint of NWD and the internal-overlap and shape constraints of Inner-ShapeIoU, thereby improving robustness to positional deviations of tiny bounding boxes and enhancing geometric consistency. Experimental results on the VisDrone2019 dataset show that, compared with YOLO11s, BEL-YOLO reduces the number of parameters by 29.6%, while improving mAP50 and mAP50–95 by 8.3 and 5.6 percentage points, respectively, and increasing precision and recall by 6.2 and 7.2 percentage points, respectively, while maintaining an inference speed of 157.41 FPS. Further experiments on the TinyPerson dataset show improvements of 5.3 and 1.84 percentage points in mAP50 and mAP50–95, respectively, demonstrating good generalization to extremely small objects and complex backgrounds. Overall, BEL-YOLO effectively improves small-object detection accuracy while controlling model size and inference overhead, providing an efficient solution for real-time small-object detection and edge deployment on resource-constrained UAV platforms.

To address the problem that symbolic music generation models often fail to adequately capture harmonic relationships, rhythmic organization, and music-theoretical constraints among different instrument tracks in multi-track arrangement tasks, this paper proposes a multi-track music generation model integrating aesthetic computing and hybrid learning, namely HL-MTMG. Existing symbolic music generation methods mainly focus on note sequence modeling or general controllable generation. Although these methods can learn the distribution of music data to a certain extent, they still tend to produce loose harmonic relationships, unstable rhythmic structures, excessive repeated notes, and insufficient pitch control when generating multi-track music involving piano, guitar, and bass. To improve the coordination and artistic expressiveness of generated multi-track music, HL-MTMG combines cross-track collaborative modeling with music-theoretical rule constraints from the perspective of music aesthetic computing, enabling the model to learn both the internal sequential patterns of individual tracks and the interactions among different instrument tracks. Technically, HL-MTMG consists of a hybrid learning module and a reward feedback network. The hybrid learning module introduces both masked self-attention and cross-track attention mechanisms. The masked self-attention mechanism is used to learn temporal dependencies within each single track, while the cross-track attention mechanism calculates the correlations among piano, guitar, and bass tracks, thereby enhancing the model’s ability to represent multi-track collaborative relationships. The reward feedback network is constructed based on the idea of generative adversarial learning. It converts melodic pitch-range constraints and note repetition suppression rules into reward signals, which are combined with the cross-entropy loss during training. In this way, the generator is guided to learn the data distribution while maintaining better melodic regularity and harmonic rationality. Experiments are conducted on a three-track symbolic music subset constructed from the Lakh MIDI dataset. During data preprocessing, pop MIDI files containing piano, guitar, and bass tracks are selected, followed by track mapping, beat alignment, abnormal sample removal, and four-bar segmentation. Finally, 11,840 three-track music segments are obtained. To avoid data leakage, the dataset is divided at the original MIDI-file level rather than at the segment level, and the training, validation, and test sets are constructed at a ratio of 8:1:1. All comparison models and ablation models adopt the same data split, input representation, generation length, and evaluation procedure. To further evaluate the generalization ability of the proposed model, an External-MIDI test set containing 300 external three-track music segments is also constructed to examine its adaptability to non-training data sources. HL-MTMG is compared with three representative baseline models, namely MuseGAN, HRNN, and XMusic, as well as three ablation variants. Objective evaluation results show that HL-MTMG achieves a chord matching degree of 0.61 and a harmonious degree of 0.72 in the main experiment, outperforming the selected comparison models and ablation variants overall. In terms of note-level features, the ratio of three consecutive notes to the total number of notes (NCTN) of HL-MTMG is 0.22, which is lower than that of the other models. Meanwhile, the model achieves 0.47, 0.72, and 0.64 on qualified rhythm (QR), used pitch classes (UPC), and polyphonicity (PP), respectively, indicating better rhythmic organization, pitch richness, and multi-track polyphonic performance. On the External-MIDI test set, HL-MTMG obtains 0.56, 0.67, 0.25, 0.45, 0.68, and 0.60 in chord matching degree, harmonious degree, NCTN, QR, UPC, and PP, respectively, showing that the model can still maintain relatively stable generation quality on external data sources. For subjective evaluation, 10 professional composers and 90 music enthusiasts are invited to rate the generated results. The evaluation by music enthusiasts shows that HL-MTMG obtains scores of 7.87, 8.17, and 7.47 in melody, music form, and audibility, respectively. The professional composers’ evaluation shows that HL-MTMG achieves scores of 6.6, 7.6, 7.8, 7.4, and 9.1 in chord, interval, harmony, pitch, and music style, respectively. The Kendall’s coefficient of concordance test further shows that the Kendall’s W values for rhythm, melody, integrity, music form, audibility, and fluency are 0.74, 0.81, 0.76, 0.79, 0.72, and 0.75, respectively, all reaching a significant level, indicating a high degree of consistency in the subjective ratings. The experimental results demonstrate that the cross-track attention mechanism can enhance the model’s ability to capture collaborative relationships among piano, guitar, and bass, while the music-theoretical reward feedback helps improve melodic pitch control and repeated-note suppression. In the constructed three-track symbolic music generation task, HL-MTMG improves harmonic coordination, rhythmic organization, and several subjective evaluation dimensions of the generated results, providing a feasible approach for introducing aesthetic computing and music-theoretical rules into multi-track music generation. The current experiments are mainly conducted on fixed-length symbolic music segments with three fixed tracks. Future work will further verify the adaptability of the model in more instrument combinations, music styles, and long-structure music generation tasks.

Aiming at the challenges of real-time generation of high-precision radio maps (RMs) in low-altitude multi-UAV networks, including the rapidly changing environment, limited computing capability of individual UAVs, and edge computing overload caused by concurrent requests from multiple UAVs, this paper proposes an air-ground collaborative RM shared generation mechanism and a joint optimization method. First, by exploiting the RM generation capability of conditional diffusion models in complex propagation environments, scenario information such as building distributions and radiation-source locations is encoded into conditional prompts. At the ground edge servers, UAV requests are clustered according to the similarity among their scenario conditions. For similar requests within the same cluster, a representative medoid sample is selected as the shared condition to perform the prefix denoising process of the diffusion model, after which each UAV completes the subsequent personalized denoising process based on its own scenario condition. In this way, intermediate computations among similar requests can be reused, thereby reducing redundant inference overhead. Second, by comprehensively considering the air-ground communication latency between UAVs and base stations, the shared denoising computation latency at edge servers, and the impact of differences among request conditions on RM generation quality, a mixed-integer nonlinear programming model is established to coordinate the trade-off between RM generation quality and end-to-end system latency through the joint optimization of the UAV-base station association strategy and the intra-server clustering structure. To address the strong coupling among the decision variables and the high computational complexity of the optimization problem, an Association-Clustering Collaborative Optimization (ACCO) algorithm is proposed. In the outer layer, a genetic algorithm incorporating an elitism-preserving strategy is employed to globally optimize the association strategy. In the inner layer, the optimal number of clusters and medoid samples for each server are adaptively determined based on the silhouette coefficient and hierarchical clustering. Through iterative optimization between the outer and inner layers, the service association strategy and the shared denoising structure are jointly optimized. Simulation results show that, compared with simulated annealing, GAI-IoV, random association, local computing, and STACKING, the proposed method reduces the average system computation latency by approximately 13.7%, 9.5%, 18.4%, 42.2%, and 9.9%, respectively, while maintaining high and stable RM generation quality. Under different numbers of UAVs and edge servers, as well as different server computing capabilities, the proposed method demonstrates favorable latency performance, generation stability, and network scalability. Furthermore, a physical test platform is constructed to evaluate the proposed shared denoising mechanism in real-device and wireless communication environments. The test results verify the feasibility of the proposed air-ground collaborative RM generation method, which can provide a technical reference for real-time electromagnetic situational awareness in highly dynamic low-altitude scenarios and the coordination of large-scale UAV networks.

Cognitive diagnosis estimates students’ mastery of knowledge concepts from their response records, providing a basis for personalized instruction. Existing methods primarily focus on static relationships among students, exercises, and knowledge concepts, with limited joint modeling of temporal context and latent affective factors. Consequently, short-term response fluctuations may be incorporated into estimates of cognitive states. Affective information is typically used as an additional predictive feature, without a dynamic modulation mechanism tailored to individual interactions, and some datasets lack affective labels. To address these issues, we propose the Affective-Temporal Neural Tensor Ring Cognitive Diagnosis model (AT-NTRCD) and develop a transfer variant, T-AT-NTRCD, for settings without affective labels. AT-NTRCD maps students, exercises, and time into a low-rank tensor ring space and uses Q-matrix constraints to maintain semantic consistency between exercise representations and knowledge concepts. A gated recurrent unit encodes historical time embedding sequences and aggregates information from adjacent time slices to generate temporal context representations. Cognitive challenge representations are then constructed from the discrepancy between students’ knowledge mastery and exercise difficulty at the concept level. These representations are combined with students’ affective tendencies and temporal context to infer multidimensional affective states for individual interactions. The model parameterizes these affective states as affective modulation core slices, which form a closed tensor ring interaction structure together with the student, exercise, and temporal cores. An encoder–decoder neural network then extracts nonlinear interaction features to accurately predict student response outcomes. To improve the reliability of affective representations and extend the model’s applicability, training strategies are designed for different affective label conditions. For datasets with affective pseudo-labels, auxiliary supervision constrains affective representation learning. For datasets without affective labels, T-AT-NTRCD reuses and freezes the source-domain affective generation network while learning student affective tendency mappings, cognitive challenge projections, and affective basis matrices compatible with the target tensor ring rank in the target domain. This enables dynamic affective modulation without affective supervision in the target domain. Experiments on multiple real-world educational datasets show that, compared with the best-performing baselines, AT-NTRCD improves AUC and ACC by 0.44 and 0.40 percentage points, respectively, and achieves a relative RMSE reduction of 0.55% on ASSIST12. On ASSIST17, the corresponding improvements in AUC and ACC are 1.90 and 1.23 percentage points, with a relative RMSE reduction of 11.87%. On Junyi, which lacks affective labels, T-AT-NTRCD improves AUC and ACC by 3.53 and 2.39 percentage points, respectively, and achieves a relative RMSE reduction of 19.69%. Ablation studies further confirm the effectiveness of temporal context encoding, affective modulation, Q-matrix constraints, and neural tensor ring interactions. In summary, the proposed models improve response prediction through low-rank high-order interactions and interaction-level affective modulation, and extend affective modeling to settings without explicit affective labels.

Graph anomaly detection has important application value in industrial scenarios such as microservice call chain tracing and communication network monitoring. Real-world systems often contain multiple types of entities and interaction relationships, exhibiting significant heterogeneous characteristics. However, existing heterogeneous graph anomaly detection methods usually adopt feature concatenation or simple average pooling when aggregating cross-relation information, implicitly assuming that different relation types have equal importance for anomaly detection. This assumption often does not hold in real data. When anomalies are only reflected in a few key relations, regular relations tend to dilute important anomaly signals. In addition, existing methods are mostly limited to a single pooling mechanism when generating graph-level representations, making it difficult to simultaneously handle both global weak anomalies and local strong mutations. To address these issues, this paper proposes HRA-GNN, a new heterogeneous graph anomaly detection framework based on relation deviation-modulated attention. The method defines relations as type triplets and maintains independent message transformation parameters for each relation. A relation deviation-modulated dynamic attention mechanism is introduced to characterize the deviation strength of each relation channel relative to normal prototypes by establishing normal relation prototypes, achieving anomaly-sensitive relation weighted fusion and effectively suppressing the interference of regular relations. An adaptive gated hybrid readout module is designed to balance local anomalies and global anomalies. Heterogeneous graph data augmentation and self-supervised learning mechanisms are introduced to provide additional structural discriminative signals for unsupervised training. The framework does not require predefined meta-paths and can adaptively capture abnormal patterns in multi-type interactions. Through the synergy of deviation modulation and gated readout, the model maintains stable discriminative capability when facing complex anomaly patterns where local mutations and global distribution shifts coexist, and exhibits good generalization across different data scales and relation types. Experimental results demonstrate that HRA-GNN achieves an AUROC of 0.84 and an AP of 0.78 on the TraceLog dataset, attains 1.0 for both AUROC and AP on the FlowGraph dataset, and yields an AUROC of 0.87 and an AP of 0.85 on the ADFA-LD dataset, verifying the effectiveness of the proposed method.

To address the severe challenges in Unmanned Aerial Vehicle (UAV) aerial imagery, including drastic target scale variations, feature degradation caused by low pixel coverage of small targets, and intense background noise interference, a Multi-attention Enhanced Recursive Feedback Detector (MERF-DETR) is proposed, utilizing the real-time end-to-end detector RT-DETR-L as the baseline framework. First, a Multi-Scale Triplet Attention (MSTA) module is designed at the end of the backbone network, using a partial channel split strategy to perform parallel depthwise separable convolutions with varying kernel sizes exclusively on key feature channels, establishing multi-scale local spatial representations while controlling parameter redundancy, and combining cross-dimensional triplet attention to enhance feature interaction between spatial and channel dimensions for improved multi-scale target representation. Second, a Neck Recursive Feedback Feature Pyramid Network (NRF-FPN) is constructed with local recursive feedback paths, utilizing deep semantic features for top-down semantic calibration of shallow high-resolution features to alleviate small target feature degradation during progressive downsampling. Finally, to combat background noise such as complex textures and shadows, a Coordinate guided Spatial frequency complementary Feature Enhancement (CSFE) module is introduced, leveraging coordinate attention mechanisms to focus on spatial directional information, depthwise convolutions to compensate for local edge textures, and 2D real Fast Fourier Transform-based frequency-domain filtering to model global context for complementary feature enhancement. Experimental results demonstrate that MERF-DETR achieves mAP@50 scores of 51.3% and 47.4% on the VisDrone2019 and UAVDT datasets, representing improvements of 9.1 and 5.8 percentage points over Deformable-DETR, respectively. Ablation studies indicate that individually incorporating the MSTA, NRF-FPN, and CSFE modules improves mAP@50 over the baseline by 1.5, 3.6, and 1.1 percentage points, respectively. Meanwhile, the CSFE module effectively mitigates the computational redundancy introduced by NRF-FPN, and the synergy of all three modules achieves optimal overall performance, significantly enhancing small target detection accuracy in drone aerial scenarios.

To address character confusion, discontinuous cross-sentence emotional progression, and unstable transitions between narration and dialogue in long-text multi-role audiobook generation, this paper proposes an agent-based decision-planning method for generating emotional multi-role audiobooks. The proposed method employs a scene chain as a discourse-level intermediate representation and integrates long-form text parsing, character modeling, emotion control, and rhythm planning into a structured front-end decision process without modifying the architecture of the underlying acoustic model. First, a script annotation agent identifies scene boundaries according to temporal changes, spatial transitions, variations in the character set, and shifts in event topics. It then performs narration/dialogue recognition, speaker annotation, and character-set extraction within each scene, thereby constructing a scene chain containing scene types, scene texts, character identities, and discourse-style information. This representation provides unified and traceable contextual information for subsequent modules. Reference voices are then assigned to characters according to attributes such as gender, age, identity, and personality, and the same reference voice is maintained for each character throughout the generation process to reduce cross-scene timbre drift. Second, an emotion-planning module consisting of an emotion vectorization agent and an emotion constraint agent is constructed. The emotion vectorization agent maps textual semantics into an eight-dimensional emotion vector comprising happiness, anger, sadness, fear, disgust, depression, surprise, and neutrality. The emotion constraint agent further refines the initial vector by incorporating character traits, scene-level emotional tone, and historical emotional states. Cross-turn emotional discontinuities are detected using cosine similarity. Plot-triggered emotional changes are preserved, whereas unsupported emotional jumps are suppressed. Third, a contextual rhythm-planning mechanism is designed. It establishes intra-segment pauses and discourse-style transition pauses according to scene type, and dynamically adjusts inter-segment pauses by considering character alternation, narration length, and changes between adjacent emotion vectors. This mechanism improves rhythmic continuity in long narration, consecutive utterances by the same character, character switching, and transitions between narration and dialogue. The experiments use 40 Chinese novel excerpts covering eight representative scene types, with five excerpts for each type. Each excerpt contains three to four speakers, including one narrator and two to three story characters. Reference voices from 24 speakers are selected from publicly available Chinese speech datasets. To ensure a fair comparison, all methods use the same texts, reference recordings, and predefined character-to-voice mappings. Experimental results show that, when the emotional-transition threshold is set to 0.65, the proposed method achieves subjective scores of 4.09, 4.21, 3.92, and 3.96 for naturalness, emotion accuracy, character consistency, and rhythm comfort, respectively. Compared with the second-best baseline, emotion accuracy and rhythm comfort improve by 24.2% and 29.8%, respectively, while the blind-listening preference rate reaches 72.1%. In the objective evaluation, the proposed method obtains a word error rate of 7.55%, a speaker similarity score of 0.89, a speaker assignment error rate of 0.13%, and an intra-speaker timbre consistency score of 0.027 under cross-scene and cross-emotion conditions. These results indicate that the method improves character identity stability while preserving speech intelligibility and reference-speaker characteristics. Comparative and ablation experiments further demonstrate that the scene-chain representation, multi-level emotion constraints, and contextual rhythm planning make complementary contributions to emotional continuity, character consistency, and rhythm naturalness. In particular, removing the character and scene constraints causes the most evident degradation in character consistency, whereas removing the rhythm-control module leads to the largest reduction in rhythm comfort. The results demonstrate that structured scene modeling and staged agent-based planning can effectively coordinate character, emotion, and rhythm control in long-form speech generation, providing an interpretable and extensible front-end control framework for discourse-level multi-role audiobook synthesis without requiring modifications to the acoustic model.

Graph neural networks (GNNs) have been widely applied to various complex tasks due to their capability of effectively processing non-Euclidean data. However, the strong dependence of GNNs on graph structures during data modeling makes them highly sensitive to structural variations. Even slight structural manipulations can lead to severe deviations in model predictions, thereby significantly reducing model stability and reliability. Existing adversarial defense methods mainly focus on two directions. The first direction processes and corrects perturbed graph data before model training in order to restore the statistical characteristics of the graph structure as much as possible. The second direction enhances and improves the information aggregation mechanism during the training of GNNs to strengthen the robustness of learned representations. However, these methods emphasize superficial structural repair or enhancement of the aggregation process, while overlooking the fundamental mechanism by which adversarial attacks impair GNN performance. In essence, adversarial attacks disrupt the original distribution of node representations, weaken inter-class separability, and increase intra-class confusion, which consequently limits the generalization ability of defense methods across different attack scenarios. To address this limitation, this work systematically analyzes the underlying reasons why adversarial attacks damage node representations. Based on the fundamental mechanism that adversarial perturbations destroy the discriminative structure of the representation space, we propose a dual-view decision boundary purification method, termed DV-DBP. This method explicitly models and repairs the discriminative structure in the node representation space, thereby achieving systematic suppression and mitigation of the impact of adversarial perturbations. Specifically, DV-DBP performs collaborative purification of perturbations from two complementary perspectives, namely inter-node and node-edge relationships. On the one hand, it characterizes the disordered state of node representations to identify and alleviate neighborhood confusion caused by anomalous nodes. On the other hand, it constrains abnormal edges and explicitly models potential low-quality connections between nodes, thereby mitigating the adverse effects of perturbations and spurious associations during representation propagation. On this basis, an inter-node relationships graph and a high-confidence graph are further constructed to establish a consistency-driven collaborative modeling mechanism across the structural and representation levels, enabling the model to simultaneously perceive node representation quality and structural reliability during aggregation. Through joint optimization of the dual perspectives, the method purifies the impact of perturbations while effectively improving inter-class decision boundary separability and intra-class compactness, thereby significantly enhancing robustness and generalization under various attack scenarios. Experimental results on multiple real-world graph datasets demonstrate that DV-DBP achieves superior defense performance under two representative adversarial attack scenarios, with performance improvements of up to 3.88% and 2.41% compared with state-of-the-art methods. In generalization experiments, DV-DBP maintains stable advantages under stronger adversarial attack settings, with performance gains of up to 2.33% and 3.38%, which verifies its robustness in complex attack environments. In addition, ablation studies show that the construction of the inter-node relationships graph provides effective perturbation correction capabilities, while the high-confidence graph helps to finely identify and correct potential abnormal connections. The synergy of these two components enables the proposed dual-view modeling strategy to form a systematic defense mechanism against various attack types. Furthermore, experiments on low-quality graph data containing structural noise show that DV-DBP can still effectively correct node representations, demonstrating strong robustness and good generalization ability under scenarios with large-scale perturbations and noise. In summary, this work conducts a systematic analysis of graph adversarial defense from the perspective of representation space, and verifies the necessity and effectiveness of explicitly modeling and optimizing representation space structure through a dual-view decision purification mechanism.

Retrieval-augmented generation introduces external knowledge and effectively alleviates the knowledge gaps and hallucination problems of large language models. However, existing methods still face two challenges in complex question-answering scenarios. First, relevant evidence is often scattered across multiple text chunks. Existing retrieval methods have difficulty explicitly modeling the associations among cross-text events. This problem leads to insufficient evidence integration. Second, existing candidate text ranking methods lack sufficient explicit constraints on event-mainline consistency. They tend to rank redundant texts or noisy evidence at the top. Such evidence is superficially relevant but deviates from the target event. This problem affects the accuracy of final answer generation.To address these problems, this paper proposes ED-RAG, an Event-Driven Retrieval-Augmented Generation framework based on a knowledge graph. ED-RAG uses events as the core semantic units for cross-text fact organization and evidence aggregation. It structures event information, entity information, and source relations in unstructured texts. It also explicitly links these elements. This design improves evidence retrieval, integration, and ranking in complex question-answering scenarios.ED-RAG first takes raw text chunks as input. It uses a large language model to jointly extract events and entities. It identifies event names, time, locations, participants, and related entities. It then constructs a lightweight event knowledge graph with text nodes, event nodes, and entity nodes. The graph contains two types of relations: event-source text relations and event-related entity relations. These relations explicitly record event sources and entity constraints. They organize factual information from scattered texts around event semantics.ED-RAG further introduces an event fusion mechanism to handle different natural language expressions of the same event. The mechanism first uses an event vector index to retrieve semantically similar candidate events. It then uses a similarity threshold and shared-entity constraints to perform preliminary filtering. Next, it uses a large language model to determine event identity. The mechanism merges nodes that refer to the same event. It keeps events with key attribute conflicts or insufficient evidence as independent nodes. This strategy reduces the risk of incorrect event merging.In the question-answering stage, ED-RAG extracts entities from the question. It normalizes the entities and evaluates their importance. It selects core entities and maps them to graph nodes. It then retrieves candidate evidence along the “entity-event-text” path. This process enables the retrieved evidence to cover key entities in the question. Event semantics also constrain the retrieved evidence. This design improves the match between the evidence and the question intent. When graph-path retrieval fails, the system uses the event vector index to retrieve the events that are semantically closest to the question. It then traces these events back to their source texts and uses the texts as fallback evidence.Finally, ED-RAG integrates multiple signals, including entity coverage, question-text semantic relevance, and question-event semantic consistency. It uses these signals to rerank candidate evidence. It then inputs high-quality text chunks into a large language model to generate the answer.This paper conducts experiments on four datasets: QA(1-document), QA(2-document), QA(3-document), and UHGEval. The experimental results show that ED-RAG achieves F1QE scores of 63.05 and 57.08 on the cross-text question-answering datasets QA(2-document) and QA(3-document), respectively. Compared with the best baseline method, ED-RAG improves F1QE by 7.34 and 4.25 points on the two datasets, respectively. The experimental results show that ED-RAG has stronger advantages in cross-document evidence localization and integration. Events serve as intermediate semantic units for cross-text evidence organization. This design enhances the aggregation of scattered evidence and strengthens event-mainline constraints. The ranking strategy integrates entity, text-semantic, and event-consistency signals. It effectively suppresses noise interference and provides more reliable context for large language models. Therefore, ED-RAG improves the accuracy of answer generation in complex question-answering scenarios.

To address the problems of complex rain streak distributions, confusion between rain streaks and background textures, and redundant parameters in deraining networks, this paper constructs a lightweight adversarial deraining network based on feature mapping. Existing deep learning-based deraining methods have achieved good performance on synthetic rainy images. However, in heavy rain, oblique rain streaks, local occlusion, and regions where rain streaks overlap with background edges, they may still suffer from residual rain streaks, blurred edges, and insufficient detail recovery. Meanwhile, some deraining models rely on complex feature extraction structures, leading to large parameter numbers and high computational costs, which makes them less suitable for resource-constrained scenarios such as vehicle-mounted terminals, roadside perception devices, and edge computing platforms. Therefore, this paper improves an adversarial deraining network from two aspects: input feature representation enhancement and model lightweighting.In terms of network structure, an adversarial generation framework is adopted as the basic model. The generator maps rainy images to derained images, while the discriminator distinguishes generated images from real rain-free images. During training, reconstruction loss and adversarial loss are jointly used to balance pixel-level consistency and visual naturalness. Considering that rain streaks are usually thin, directional, unevenly distributed, and rich in high-frequency components, Fourier feature mapping is introduced at the input stage. Two-dimensional spatial coordinates are expanded into multi-frequency positional encodings and concatenated with shallow image features to form enhanced input features. This provides the generator with richer spatial-frequency information, improves its ability to perceive rain locations, edge structures, and local texture regions, and reduces restoration errors caused by the confusion between rain streaks and background details.For model lightweighting, a frequency-domain channel pruning strategy is designed. The intermediate features of the generator are first transformed into the frequency domain by Fourier transform, and the frequency-domain energy distribution of each channel is then calculated. Channel importance is evaluated according to the response to high-frequency details and structural information. Channels that contribute more to rain modeling and detail recovery are retained, while low-contribution redundant channels are pruned to reduce the number of parameters. This pruning process is mainly performed after training, so it does not introduce additional computational burden during inference.In the experiments, Rain14000 is used as the training set, and Rain100H, Rain100L, Rain12, and the real rainy dataset SPA-Data are used for testing. Quantitative results show that the proposed method achieves PSNR/SSIM values of 30.96 dB/0.900, 37.62 dB/0.975, and 35.41 dB/0.959 on Rain100H, Rain100L, and Rain12, respectively. Compared with the Base model, the proposed method improves PSNR by 0.76 dB, 0.82 dB, and 0.61 dB on the three datasets, respectively, indicating stable restoration performance under different rain conditions.Ablation experiments show that Fourier feature mapping improves PSNR and SSIM on Rain100H and SPA-Data, demonstrating its ability to enhance the modeling of rain spatial distribution and high-frequency details. When only frequency-domain channel pruning is applied, the number of parameters is reduced, while the restoration metrics decrease slightly. When Fourier feature mapping and frequency-domain pruning are combined, the model achieves 30.96 dB/0.900 on Rain100H and 29.42 dB/0.925 on SPA-Data, with the parameter number reduced to 0.69M. This indicates that the two components are complementary in feature enhancement and model compression. Qualitative results show that the proposed method can reduce residual rain streaks and better preserve human edges, umbrella regions, and complex background textures. Overall, the experimental results demonstrate that the proposed method improves deraining quality while controlling model size, providing a lightweight solution for image restoration in complex rainy scenes.

To address the fragmentation of cooperative-communication research in "classical-quantum" heterogeneous fusion systems—where the module-coupling mechanism remains unclear and a unified analytical framework is lacking—this paper reviews the technical status, key bottlenecks, and trends of heterogeneous cooperative communication from an architectural perspective. First, taking the timing-coupling depth of feedback loops as the essential dimension, we classify cooperative modes into three types: open-loop feedforward, variational iteration, and real-time closed-loop. Moreover, cooperative-architecture differences are comparatively analyzed across four physical systems, namely superconducting, photonic, trapped-ion, and neutral-atom. Second, a "classical-quantum" bidirectional data-flow interaction model is constructed. In this model, the control flow carrying timing synchronization and gate-trigger pulses is strictly decoupled from the feedback flow carrying measurement results, and the microsecond-level real-time closed-loop demand of quantum error correction serves as a benchmark to quantify the interaction constraints. Then, according to communication distance and deployment form, the technical system is organized into three categories: near-field high-frequency measurement and control, distributed cooperation, and quantum cloud communication. Specifically, the performance boundaries of three near-field control architectures—FPGA room-temperature control, cryogenic CMOS, and single-flux-quantum—are compared. In addition, the evolution of distributed cooperation from offline optimization toward dynamic cooperation is revealed, and two cloud-communication routes are outlined: "near-quantum computing" runtime optimization and heterogeneous hardware abstraction. Finally, the bottlenecks are dissected from four layers—physical-layer timing, system-layer bandwidth, computing-layer coupling, and ecosystem-layer standards. Here, a performance-constraint model is used to clarify the coupling between bandwidth and decoding computing power, thereby yielding optimization paths and evolution tiers. The study shows that a 3-to-6-order-of-magnitude time-scale gap exists between quantum gate operations and classical communication, and that the decoding latency is about 57 times the error-correction cycle period. For ten-thousand-qubit superconducting systems, moreover, interconnect challenges emerge: more than 30,000 cables cross the temperature zones, and the heat load exceeds the 4 K cooling capacity by an order of magnitude. In addition, when frequency-division multiplexing reads out N qubits, the signal-to-noise ratio degrades as the square root of N. Furthermore, as the code distance grows, bandwidth and decoding computing power fall into a coupled imbalance; consequently, the syndrome uplink bandwidth of a million-qubit system will reach the Tbps level. Meanwhile, proprietary protocols aggravate standard fragmentation, and measurement-and-control costs grow superlinearly. We therefore conclude that the essence of cooperative communication lies in the timing-coupling depth of feedback loops, and that the relevant bottlenecks must be broken through jointly via timing optimization, cross-layer scheduling, technology integration, and protocol unification. Accordingly, a phased strategy is proposed: frequency-division multiplexing for the near term, on-chip routing for the mid term, and optical interconnection for the long term. Ultimately, end-to-end cross-layer optimization will become the key to advancing toward million-qubit fault-tolerant computation.

The propagation of online rumors has evolved from single-modal text to composite forms incorporating multimodal information such as images and videos. Leveraging stronger visual impact and emotional appeal, multimodal content significantly accelerates the dissemination speed and expands the coverage of false information, posing severe challenges to cyberspace governance. This paper proposes a detection model based on feature enhancement and dual-path decision fusion. The overall architecture consists of three core modules: lightweight visual feature enhancement, dual-path collaborative discrimination, and adaptive decision fusion. First, a Lightweight Parallel Multi-scale Feature Enhancer (LPMFE) is designed, which employs three parallel branches—global self-attention, spatial depthwise separable convolution, and temporal 1D convolution—to capture features from different dimensions. A learnable weight vector dynamically fuses these multi-scale outputs, while a gated residual mechanism and feedforward network refine local image features while preserving the integrity of the backbone representation. Second, a dual-path collaborative discrimination mechanism is constructed. The generative path first aligns the visual feature sequence into the word embedding space of a large language model (LLM) via a cross-attention based image projector. The projected visual tokens are then inserted into the text sequence and fed into the LLM for autoregressive inference. The hidden state of the last token is extracted and mapped through a linear classification head to logits over the vocabulary, which are finally converted into a generative prior probability via a Softmax function, leveraging the model’s parametric knowledge for macroscopic logical reasoning. The discriminative path explicitly verifies image–text semantic consistency by sharing the deep hidden states from the same LLM and aggregating them via masking to obtain a global text vector; meanwhile, the enhanced visual features are pooled. Both modalities are projected into a shared 256-dimensional semantic space. A supervised contrastive loss is introduced to optimize the discriminability of modal features. The absolute difference and element-wise product of the projected features are computed to capture fine-grained semantic conflicts. The original projected features and the two interaction features are concatenated along the channel dimension and fed into a multilayer perceptron to output a consistency discrimination score. Finally, at the decision fusion level, a learnable weight parameter is introduced to adaptively fuse the dual-path output probabilities. A two-threshold mechanism is set to form a three-interval decision logic, categorizing samples into high-confidence rumors, high-confidence non-rumors, and ambiguous cases. A boundary penalty term is added to the optimization objective to prevent threshold collapse and ensure stable convergence of the fusion strategy. Experiments are conducted on the Weibo, Fakeddit, and FineFake datasets. On Weibo, the method achieves an accuracy of 93.23%, with precision/recall for non-rumors at 95.72%/90.50% and for rumors at 90.99%/95.95%. On Fakeddit, the accuracy reaches 91.56%, with F1-scores of 92.98% for non-rumor and 89.42% for rumor, a difference of 3.56 percentage points. On FineFake, the accuracy is 81.21%, with non-rumor/rumor F1-scores of 83.42%/78.31%. Ablation studies demonstrate that removing LPMFE decreases FineFake accuracy to 79.20% and rumor recall by 5.32 percentage points; replacing the cross-attention projector with a linear mapping reduces Fakeddit rumor F1 by 6.01 points; and removing the discriminative branch lowers Weibo accuracy to 86.59% and rumor recall by 16.27 points. This study demonstrates that incorporating a visual feature enhancement mechanism to suppress information redundancy and complex scene interference, combined with an LLM-based parametric commonsense reasoning and explicit cross-modal consistency verification via a generative-discriminative dual-path collaborative design, overcomes the limitations of single-paradigm detection models, effectively mitigates LLM hallucinations and unimodal semantic bias, and substantially improves the overall accuracy of multimodal rumor detection.

To address the key challenges faced by existing graph neural network (GNN) based multi‑criteria recommendation algorithms under data‑sparse conditions—namely the long‑tail effect, the difficulty in synergistically integrating graph attention networks (GAT) and graph convolutional networks (GCN) , and the ambiguous discrimination between positive and negative samples caused by the difficulty of contrastive learning anchor selection in multi‑criteria rating scenarios—this paper proposes a multi‑criteria recommendation algorithm that fuses noisy graph neural networks with contrastive learning, termed NGCLR. The research objective is to systematically mitigate node embedding degradation, enhance cross‑criteria representation consistency, and optimize contrastive learning effectiveness by introducing dynamic noise augmentation, a progressive fusion mechanism, and a neighbor‑similarity‑based anchor selection strategy, thereby significantly improving the rating prediction accuracy and generalization capability of multi‑criteria recommender systems in sparse interaction environments. On the technical implementation side, the model first constructs a user item bipartite graph for each evaluation criterion independently and normalizes the adjacency relationships using a layer wise trained bipartite graph neural network matrix, so as to preserve the independent rating distribution characteristics of each criterion. To tackle the degradation phenomenon where long tail items' embeddings converge towards head items due to sparse neighborhoods, a Gaussian noise enhanced LightGCN module is designed: during each layer of graph convolution propagation, dynamic Gaussian noise with an initial intensity of 0.2 and an exponential decay rate of 0.5 per training epoch is injected into the current layer's node embeddings. This perturbation does not alter the graph topology but acts solely on the feature space, thereby breaking feature homogenization and encouraging a more uniform embedding distribution for long tail items. To fuse the complementary advantages of graph attention networks in capturing fine grained local interactions and graph convolutional networks in perceiving global structural patterns, a progressive fusion mechanism is proposed. At each propagation layer, the multi head attention outputs from GAT and the neighborhood aggregation outputs from LightGCN are concatenated, and a learnable parameter adaptively controls the contribution of each layer, achieving hierarchical fusion from low level local details to high level global semantics. To overcome the difficulty of reliably selecting contrastive learning anchors caused by the inconsistency between rating criteria and samples in multi criteria data, an anchor selection strategy based on average neighbor cosine similarity is designed: within each criterion view, the average similarity between each node's embedding and those of all its neighbors is computed, and the node with the highest similarity is selected as the anchor for that view, serving as a stable contrastive benchmark. On this basis, a dual contrastive learning task is constructed—the local contrastive loss maximizes the consistency of positive sample embeddings of the same node across different criteria while pushing away negative samples, and the global contrastive loss enhances cross criterion alignment among global embeddings (node means) of each criterion. These two losses are jointly optimized, and the total loss function combines the local loss, global loss, and an L₂ regularization term, balancing node level discriminability and distribution level consistency. To validate the model's effectiveness, experiments are conducted on three publicly available real‑world datasets: Yahoo!Movies, BeerAdvocate, and TripAdvisor. Yahoo!Movies contains 6,078 users, 976 movies, and 758,405 four‑criterion ratings; BeerAdvocate comprises 8,831 users, 2,698 beers, and 3,880,359 four‑criterion ratings; TripAdvisor includes 15,000 users, 1,325 hotels, and 20,830 seven‑criterion ratings, with a high sparsity level. Using mean absolute error and root mean square error as evaluation metrics, the proposed NGCLR is compared against more than ten baseline methods, including BMF, MSVD, MLR, SVR, CIC, DMCF, CFM, DNN‑MF, MCAE‑FADNN, and the best‑performing baseline D‑MGAC. Experimental results show that NGCLR achieves MAE and RMSE of 0.6028 and 0.8211 on Yahoo!Movies, corresponding to relative reductions of 2.94% and 3.79% over D‑MGAC; on BeerAdvocate, the values are 0.3931 and 0.5380, with relative reductions of 4.93% and 2.73%; and on TripAdvisor, they are 0.7338 and 0.9724, with relative reductions of 2.04% and 1.96%. Ablation studies demonstrate that removing the noise augmentation module, the progressive fusion mechanism, or the contrastive learning module leads to varying degrees of increase in both MAE and RMSE. In particular, the variant adopting a random anchor selection strategy yields MAE values that are 2.86%, 2.21%, and 0.69% higher than the full model on the three datasets, respectively, confirming the independent contributions and synergistic gains of each core component. Hyperparameter analysis indicates that the optimal performance is achieved with an embedding dimension of 256, 8 attention heads, and equal weights of 0.5 for both local and global contrastive losses; the dynamic decay noise strategy outperforms fixed‑noise and standard‑adaptive‑noise strategies, and the neighbor‑similarity‑based anchor selection surpasses both global‑average‑similarity and PageRank‑based strategies. Head‑vs‑tail group experiments show that the model achieves significantly greater MAE improvements on tail item groups than on head groups, and noise enhancement visualization further confirms that the embedding distribution of long‑tail items shifts from clustering towards dispersion. In summary, NGCLR effectively addresses the three major challenges in multi‑criteria recommendation—long‑tail degradation, insufficient fusion of graph modules, and anchor bias in contrastive learning—achieving consistently superior prediction accuracy across multiple real datasets, and thus provides an efficient and robust solution for multi‑criteria recommender systems.

In high-concurrency access scenarios with massive small files, frequent context switches caused by metadata operations become a major performance bottleneck for user-space file systems. Existing caching and prefetching mechanisms find it difficult to strike an effective balance among memory overhead, adaptability to random access patterns, and performance. To reduce the per-component lookup overhead during multi-level path resolution in the FUSE (Filesystem in Userspace) framework, this paper proposes a path-aware batch metadata prefetching scheme named Lookupplus. The core design philosophy of Lookupplus is to change the request interaction granularity between the kernel and the user-space daemon. In native FUSE, path resolution follows a serial per-component processing mode: whenever the directory entry cache of a component expires, the FUSE kernel module issues an independent lookup request, and each expired component incurs two context switches. As directory depth increases, the cumulative overhead of multiple round trips significantly amplifies metadata access latency. Lookupplus observes that when the cache of an intermediate component expires, the remaining subsequent components on that path will also require revalidation within a very short time. Based on this observation, when the first expired component triggers a lookup request, Lookupplus captures the complete remaining path from the current process control block, attaches it to the current lookup request, and transmits it to the user-space daemon. The daemon parses each component in the remaining path and returns their metadata entries, allowing the kernel to update the attributes and validity periods of all child directory entries in a batch, thereby enabling subsequent components to be verified through the fast path. The user-space file system does not need to be aware of the batch prefetching logic or maintain additional cache consistency states. The additional memory overhead comes only from temporary path copying and batch response entries, and is released immediately after request processing completes. The prototype system of Lookupplus is implemented on the Linux 5.2.2 kernel and evaluated against state-of-the-art schemes including ExtFUSE and RFUSE. Single-threaded metadata performance tests show that at a directory depth of 10, the average latency of create operations in Lookupplus drops to 32% of that in native FUSE. During continuous runs across depths from 1 to 10, the total number of lookup requests falls from 65 in native FUSE to 20 in Lookupplus. Thirty-two-thread micro-benchmarks cover directory widths ranging from 10 to 1,200, with 80% of the file set pre-created. Results show that Lookupplus achieves less than 5% performance improvement in concurrent access to a shared file set. Meanwhile, ExtFUSE performs approximately 20% lower than native FUSE. RFUSE shows significantly lower throughput under narrow directory configurations due to queue pressure, but gradually recovers to a high throughput level as directory width increases. In typical real-world macro-benchmarks, under the file server workload, Lookupplus improves throughput by 7.4% over native FUSE; under the web server workload, Lookupplus achieves the highest throughput, with a 9.7% improvement over native FUSE and a 12.9% improvement over ExtFUSE; under the mail server workload, due to the flat directory structure and significant synchronization overhead, Lookupplus achieves only a marginal improvement. Memory overhead evaluation shows that the Slab increment of Lookupplus under various workloads remains close to that of native FUSE, with the vast majority being kernel-reclaimable memory. In summary, Lookupplus effectively reduces the context-switch overhead of FUSE in deep-directory scenarios through a path-level batch metadata prefetching primitive with minimal architectural changes. It does not rely on complex cache consistency protocols or prediction models, nor does it introduce new kernel data structures. The extra memory overhead remains controllable and reclaimable. The benefits are most pronounced in deep-directory, cold-start, and low-reuse scenarios, making Lookupplus a lightweight and feasible optimization path for building high-performance user-space file systems in metadata-intensive environments.

Real-time, high-precision perception and 3D reconstruction of road potholes are core technologies for ensuring driving safety and active suspension control in autonomous driving. Existing road pothole perception methods primarily rely on a single visual modality or static cross-modal fusion strategies. Single visual methods are susceptible to drastic illumination changes, texture-less road surfaces, and shadow occlusions, making it difficult to stably acquire robust depth information. Meanwhile, 3D point cloud-based methods are limited by the beam distribution of LiDAR sensors, resulting in extremely sparse point clouds in long-distance or high-speed driving scenarios, which leads to insufficient expression of the microscopic geometric features at pothole boundaries. Existing cross-modal fusion schemes mostly adopt global static concatenation, which fails to fully consider the non-linear spatial registration errors between heterogeneous data. This easily leads to feature contamination or the drowning of valid geometric information when sensors suffer from spatio-temporal misalignment, thereby limiting the 3D reconstruction accuracy of pothole boundaries. To address these problems, this paper proposes a centroid-guided road pothole perception and 3D reconstruction method based on the cross-modal fusion of LiDAR and depth camera, termed DeepRoad. The proposed method constructs an end-to-end deep neural network, which mainly comprises three core technical stages. In the first stage, multi-source heterogeneous feature rough alignment and multi-scale extraction are performed. The system concurrently inputs 2D images captured by the depth camera and spatial point clouds generated by the LiDAR. A nearest-neighbor adaptive rough matching module is utilized to establish a preliminary correlation between the pixel coordinate system and the point cloud projection matrix across spatio-temporal scales, and the underlying representations of both modalities are captured via a multi-scale feature extraction network. In the second stage, centroid-guided non-linear feature fine fusion is executed. To handle feature drift caused by spatial registration errors, a centroid guidance mechanism is innovatively introduced. A lightweight prediction network is designed to localize the geometric centroid of candidate pothole regions. Taking this centroid as a spatial anchor, a 3D asymmetric attention network is constructed to dynamically adjust the interaction weights of cross-modal features, achieving adaptive aggregation and deep fusion of high-frequency image textures and sparse point cloud microscopic geometry in high-value local regions. In the third stage, Bird's-Eye-View (BEV) space mapping and edge semantic consistency constraints are applied. To resolve the boundary blurring and topological distortion issues caused by downsampling operations, DeepRoad projects the fused cross-modal high-dimensional features uniformly into the BEV space to preserve the global topological structure, and designs a novel edge semantic consistency regularization loss function. By explicitly enforcing spatial continuity and semantic opposition between the physical boundaries of potholes and the surrounding normal road surface, this loss function effectively suppresses reconstruction artifacts and substantially improves the fitting precision of pothole edges. To verify the effectiveness and robustness of the DeepRoad method, extensive comparative and ablation experiments were conducted on the synthetic dataset SYN-UDTIRI and the real-world road dataset RSRD. In the road pothole segmentation comparative experiments on the SYN-UDTIRI dataset, DeepRoad achieves the optimal performance across all key metrics: its maximum F1-score (MaxF) reaches 97.58%, Average Precision (AP) reaches 94.04%, Precision (Pre) is 97.53%, and Recall (Rec) reaches 97.86%, comprehensively outperforming mainstream cross-modal fusion methods such as RoadFormer and SNE-RoadSeg+, demonstrating a significant recall advantage particularly in depicting boundaries and fine details. Structural module ablation experiments indicate that progressively introducing modules such as centroid guidance into the baseline cross-modal fusion network validates the effectiveness and complementarity of each module. In the candidate region construction strategy ablation, the model reflecting excellent structural independence and stability. Furthermore, in the generalization and robustness tests on the real road dataset RSRD, DeepRoad maintains stable geometric estimation capabilities under sensor noise and sparse point cloud interference, achieving an absolute error of 0.479 cm and a root mean square error (RMSE) of 0.568 cm, with the proportion of large errors (>0.5 cm) restricted to only 35.1%. Compared with the state-of-the-art RoadBEV model, all metrics of the proposed method are superior, extreme deviations are significantly reduced, and no error amplification occurs. In conclusion, the DeepRoad method proposed in this paper successfully overcomes the technical bottlenecks of blurred 3D perception boundaries and difficult alignment of heterogeneous features for road potholes in complex scenarios through the centroid-guidance mechanism. It forms a systematic end-to-end perception and reconstruction solution, providing high-precision data support for intelligent obstacle avoidance planning in autonomous driving, and possessing important theoretical innovation and engineering application value. Future research will focus on network model lightweight pruning and edge-side real-time deployment, as well as exploring the deep integration of DeepRoad with multi-modal Vision-Language Models (VLMs) to drive the intelligent decision-making and reliable evolution of autonomous driving technology under all-weather and all-scene conditions.

To address the problems of underwater images suffering from color shift, contrast attenuation, and loss of high-frequency details due to selective light absorption and scattering by suspended particles, as well as the insufficient boundary recognition capability and limited segmentation accuracy of existing general-purpose semantic segmentation models for high-value rare objects, this study aims to improve the fine-grained segmentation performance for blurred underwater targets from two aspects: feature representation and loss supervision, with particular emphasis on enhancing the reconstruction and discrimination of edge contours of rare categories.Technically, this work adopts the UWSegFormer model, which has demonstrated excellent performance in underwater scenes, as the baseline architecture. While retaining its underwater image quality attention module and multi-scale aggregation attention module, we propose targeted improvements from two perspectives: frequency-domain enhancement and edge supervision. On one hand, we design a Frequency-Spatial Multi-scale Aggregation Attention (FS‑MAA) module. This module introduces Two-Dimensional Real-valued Fast Fourier Transform (2D‑RFFT) during multi-scale feature fusion in the decoder, explicitly mapping spatial-domain features to the frequency domain. It leverages the decoupling property of the frequency domain to separate low‑frequency components, which represent global semantics, from high‑frequency components, which characterize textural details. A fixed empirical gain coefficient is applied to the high‑frequency components to compensate for high‑frequency loss caused by water scattering. The enhanced frequency-domain features are then restored to the spatial domain via inverse 2D Fourier transform and finally aggregated with the original spatial features through residual connection, thereby injecting high‑frequency detail information while preserving global contextual semantics. On the other hand, we construct an Online Hazy Edge Mining‑based Edge Learning Loss (OHEM‑ELL). During training, this loss computes edge response maps for both predicted masks and ground‑truth labels using Scharr operators, obtaining the prediction error for each boundary pixel. The errors are then sorted in descending order, and the top K% boundary pixels with the largest losses are selected as hard examples; only these pixels are used for gradient backpropagation and parameter updates. This mechanism dynamically focuses the network’s learning attention on the most difficult‑to‑recognize blurred edge regions with negligible additional computational cost, effectively alleviating the dominance of easily classified pixels (such as background water) in the loss. The overall loss function is a linear weighted combination of the standard semantic segmentation loss and the proposed OHEM‑ELL, enabling joint supervision of pixel‑wise classification and edge optimization.Experiments are conducted on the SUIM underwater semantic segmentation benchmark dataset and the DUT‑USEG real‑world underwater image dataset, using MiT‑B0 as the encoder with ImageNet‑1K pretrained weights. On the six major categories of the SUIM dataset, the proposed method achieves a mean Intersection over Union (mIoU) of 86.92%, which is 1.86 percentage points higher than the baseline UWSegFormer (85.06%). Ablation studies demonstrate that the combination of both modules yields the best performance, confirming their synergistic effect in improving segmentation accuracy. On the DUT‑USEG dataset, our method improves mIoU from the baseline of 68.12% to 68.75%, verifying its good cross‑dataset generalization ability. The model has 16.81M parameters and a computational complexity of 6.48 GFLOPs, almost identical to the baseline, indicating that the proposed improvements introduce no significant extra computational burden. Visualization results further show that the FS‑MAA module effectively recovers lost textural details in turbid underwater environments, while the OHEM‑ELL loss produces more coherent and sharper object boundaries; the model integrating both modules achieves the best visual quality. Based on the above results, this study draws the following conclusions: frequency‑domain feature decoupling with selective high‑frequency enhancement can effectively compensate for detail attenuation caused by scattering during underwater imaging, enabling the network to obtain more complete structural information at the feature level; and the online hard edge mining strategy can guide the network to concentrate its limited parameters and gradient updates on the most discriminatively challenging boundary regions, significantly improving the contour delineation capability for rare objects.

To address the difficulty of bounding end-to-end latency in high-load mixed-traffic scenarios, the degradation of transmission determinism caused by priority inversion and head-of-line blocking, and the large model size and low solving efficiency of joint scheduling over dual-homed redundant topologies, a deterministic transmission architecture and scheduling method based on Time-Sensitive Networking (TSN) is proposed for train-control services. In terms of architecture, an eight-car dual-homed ring network is constructed according to the hierarchical structure of the train backbone and consist networks. A/B-plane isolation and end-to-end cross-layer mapping are employed to support differentiated transmission and redundant-path switching for services with different priorities. In terms of scheduling, a two-stage optimization model combining spatial routing and temporal gating is established based on the time-aware shaping mechanism. The first stage determines transmission paths over the two planes, whereas the second generates conflict-free transmission phases within each scheduling period. To address the rapid growth of disjunctive constraints with the number of flows in the full-scale model, binary routing reduction, a forward-list heuristic, and accelerated binary search are applied sequentially. These strategies discretize routing decisions into plane selections, fix the transmission order of frames competing for the same port, and perform a bounded search for the end-to-end latency upper bound. Under the monotonic-feasibility condition, the binary search terminates within a finite number of steps. For cases in which the demand of critical flows exceeds the available network capacity, a two-stage overload admission mechanism is designed. A priority-based bandwidth-feasibility precheck is first performed, after which the reduced model is solved for the flow set that passes the precheck. If the flow set is infeasible, the lowest-priority flow is removed and the model is solved again until a feasible solution is obtained or a rejection result is produced. Simulations are conducted on an eight-car dual-homed ring topology with background loads ranging from 20% to 95% and with scenarios containing 10 and 20 highest-priority flows. The proposed method is compared with Strict Priority (SP) scheduling, routing-only optimization, and the Online Scalable Scheduling and Routing (OSSR) algorithm. The results show that the proposed method limits the maximum end-to-end latency of critical control flows to less than 0.6ms. For the evaluated static flow sets, it reduces the maximum end-to-end latency by 33.6%–51.9%, 38.6%–48.2%, and 37.5%–43.6% compared with SP scheduling, routing-only optimization, and OSSR, respectively. With 20 highest-priority flows, the three-stage reduction decreases the number of disjunctive constraints from 738 to 0, limits peak memory consumption to approximately 4% of that required by the full-scale model, and achieves a cumulative solving time of approximately 0.5s and a speedup of approximately 1.56 fold. The binary search converges after six feasibility checks. Under overload conditions, the method admits critical flows according to capacity constraints and service priorities while rejecting requests that exceed the available capacity, thereby providing deterministic transmission guarantees for safety-critical train-control services.

To address target feature capture difficulties from extreme organ scale variations, noise interference from shallow-deep semantic gaps, boundary geometric information loss during fixed-grid upsampling, and boundary ambiguity from adjacent organ adhesion in existing medical image segmentation, this paper proposes a Cross-Scale Geometric Calibration Network. First, utilizing PVTv2 as the backbone encoder, the network employs a self-attention mechanism to effectively extract image contextual information and establish global dependencies of anatomical structures. Building on this, a Multi-Scale Collaborative Attention Block is designed. By integrating a Multi-Scale Feature Regulation Module, Shareable Multi-Semantic Spatial Attention, and Progressive Channel-wise Self-Attention, this block constructs multiple receptive fields to flexibly perceive organ size fluctuations from patient differences. Meanwhile, a channel-spatial dual attention mechanism filters key channel semantics and focuses on salient spatial regions, effectively suppressing background noise, collaboratively enhancing multi-granularity features, and improving model sensitivity to complex organ structures. Second, a Dual-path Adaptive Fusion Module is designed. Unlike traditional simple concatenation of high- and low-dimensional features, it uses learnable weights to evaluate feature importance across dimensions, dynamically adjusting the combination ratio of high-resolution encoder spatial details and low-resolution decoder semantics to alleviate the shallow-deep semantic gap. Then, a Dynamic Upsample Block is introduced. Unlike the non-geometrically adaptive stretching of linear interpolation upsampling, this block learns grouped independent spatial offset fields to perform adaptive geometric calibration on upsampling coordinates based on a fixed initial geometric grid. This mechanism enables dynamic model deformation matching real anatomical topology, improving aliasing effects from proportional enlargement during upsampling, and achieving sub-pixel reconstruction for irregular boundaries. Finally, a Foreground Orthogonal Loss is incorporated during supervised training. By penalizing spatial overlaps of predicted probabilities among foreground categories, it effectively disentangles adhered boundaries of adjacent organs, further improving overall segmentation accuracy. The proposed method is compared with advanced networks like ECA-GC-Unet, DCAU-Net, and MCT-Net on three public datasets. Notably, the standard version CSGC-Net-B has only 28.44M of parameters, far fewer than the 91.30M of the second-best model in the main experiments. Regarding accuracy, on the abdominal multi-organ dataset Synapse, CSGC-Net-B outperforms the second-best model by increasing DSC by 0.33% and decreasing HD95 by 1.14, achieving optimal accuracy for the gallbladder, pancreas, and stomach. On the cardiac MRI dataset ACDC, its DSC improves by 0.3%, securing optimal results in the myocardial region. On the intestinal polyp endoscopic dataset CVC-ColonDB, it improves DSC by 1.21%, mIoU by 0.98%, and significantly decreases HD95 by 3.04 compared to the second-best model. Moreover, the lightweight version CSGC-Net-S is reduced to just 4.33M in parameters, yet still achieves DSCs of 82.18%, 91.74%, and 91.93% on the Synapse, ACDC, and CVC-ColonDB datasets, respectively. Overall results demonstrate that compared with existing mainstream networks, CSGC-Net improves segmentation accuracy while maintaining a lightweight architecture, exhibiting excellent generalization.

Aiming at the strong nonlinearity and non-stationarity of microgrid load series under high penetration of distributed energy resources, as well as core drawbacks of existing forecasting models including accuracy degradation in abrupt load regions and distorted extreme value prediction caused by the smoothing effect of global modeling, this paper proposes a multi-scale feature short-term load forecasting model with an error-aware mechanism (MAG-TL). Existing standalone deep learning models cannot simultaneously extract multi-scale features covering long-term periodic patterns and local transient fluctuations; hybrid forecasting models generally rely on manually tuned decomposition parameters, adopt simplistic feature fusion, and lack error correction mechanisms for high-volatility segments, and notably, Transformer-based models with outstanding long-term sequence modeling capability exhibit a smoothing effect due to weighted feature aggregation under drastic load fluctuations, resulting in notable peak and valley prediction errors and failing to support refined microgrid dispatch. The model establishes a four-stage collaborative forecasting pipeline of multi-scale feature decomposition, global dependency modeling, targeted error correction and global residual compensation: first, an improved whale optimization algorithm (IWOA) with adaptive inertia weight is employed to minimize envelope entropy, balancing global exploration and local exploitation to adaptively optimize variational mode decomposition (VMD) parameters, so that the original non-stationary load sequence is decoupled into physically interpretable intrinsic mode functions (IMFs), and a multi-scale input feature set is constructed with selected meteorological correlation features; second, parallel 1D convolutional neural networks extract deep features from each IMF, and an adaptive attention module dynamically assigns modal feature weights for adaptive fusion, enhancing error-sensitive features while preserving global baseline information; third, the fused features are fed into a Transformer encoder to capture long-range temporal dependencies, where a lightweight error-aware projection layer quantifies per-step prediction risk via hidden feature spatial offset, and a Top-K relative ranking mechanism dynamically screens high-error anchor points to autonomously locate load mutation regions without ground-truth labels; finally, an error-aware enhanced LSTM decoder integrates anchor features, local error scores and global error background via residual fusion, generates multiplicative gains for targeted amplitude correction in anchor regions with a dynamic threshold to avoid over-compensation in stationary periods, and a lightweight fully connected residual network is introduced to output global additive compensation, forming a hierarchical correction mechanism of multiplicative mutation amplitude correction and additive systematic bias compensation. Verification is conducted on 18-month field data of a high-renewable-penetration microgrid in Guiyang, with the dataset split into training, validation and test sets at a 7:2:1 ratio, and multi-dimensional performance comparisons are carried out against 3 standalone models and 4 hybrid models. The results show that on the Guiyang dataset, the proposed model achieves a root mean square error (RMSE) of 10.71 MW, a mean absolute error (MAE) of 4.78 MW, a mean absolute percentage error (MAPE) of 0.77%, and a coefficient of determination (R⊃2;) of 0.992; compared with the best-performing hybrid baseline model, its RMSE is reduced by 12.23 MW, and the maximum RMSE reduction reaches 34.81 MW compared with standalone forecasting models. Ablation experiments verify the independent contribution of each core module, and the error-aware mechanism alone reduces RMSE by 5.71 MW; the Pearson correlation coefficient between error-aware scores and actual prediction residuals reaches 0.8071, confirming the high precision of the risk quantification mechanism. The single-sample inference latency of the model is 2.34 ms, which can fully meet the real-time requirement of microgrid online closed-loop dispatch. The proposed model effectively alleviates the inherent smoothing bottleneck of global modeling, improves forecasting accuracy with controllable computational overhead, and can accurately capture the transient variation characteristics of load fluctuation errors, providing reliable data support and technical reference for precise microgrid dispatch under high renewable penetration scenarios.

To address the insufficient distinction between origin and destination roles, inadequate utilization of regional-pair interactions, and weak structural constraints on predicted matrices in inter-regional origin-destination (OD) traffic flow forecasting, this paper proposes BiHOD-Net, an OD traffic flow prediction method based on an origin-destination role bipartite graph. Since the same region reflects traffic generation when acting as an origin and traffic attraction when acting as a destination, using a unified node representation may mix these two different traffic semantics. Moreover, optimizing only the prediction error of individual OD edges may cause local errors to accumulate along the row and column dimensions of the OD matrix. To address these issues, BiHOD-Net first extracts the outflow and inflow distributions of each region toward different destinations and from different origins from historical OD matrices, respectively, and constructs origin-role and destination-role sequences. A patch-based time series Transformer (Patch Time Series Transformer, PatchTST) is then employed to independently encode the two role sequences, capturing temporal variations within local patches and dependencies across different patches. These temporal features are further combined with regional static attributes to generate origin-role and destination-role representations with distinct traffic semantics. Subsequently, the two types of role nodes are placed on opposite sides of a bipartite graph to construct an origin-destination role bipartite graph according to historical OD connections. Origin-to-destination propagation is normalized according to historical outflow relationships, whereas destination-to-origin propagation is normalized according to historical inflow relationships. Through bidirectional message propagation, the model captures directional and asymmetric interactions between the traffic generation and attraction sides. At the edge-level prediction stage, the propagated origin-role and destination-role representations are integrated with their absolute difference, element-wise product, and edge historical features derived from recent flow, mean flow, maximum flow, standard deviation, and activity ratio, forming an explicit regional-pair representation that characterizes role differences, matching relationships, and the historical movement patterns of individual OD edges. On this basis, a gated linear unit (Gated Linear Unit, GLU) is employed to adaptively filter different types of information and predict future edge-level traffic flows. Meanwhile, an OD matrix marginal consistency constraint is introduced in addition to the edge-level traffic flow prediction loss. By using the row and column marginals of the ground-truth OD matrix as supervision signals, this constraint reduces deviations in regional total outflow and inflow caused by accumulated local edge-level prediction errors, thereby balancing the prediction accuracy of individual regional pairs and the overall structure of the OD matrix. Experiments are conducted on four taxi OD traffic flow datasets from Beijing, Chengdu, Nanjing, and NYC-Taxi, with prediction intervals of 20, 40, and 60 min. BiHOD-Net is compared under a unified protocol with historical average, support vector regression, deep temporal models, spatio-temporal graph neural networks, and multiple OD demand and OD matrix forecasting methods. Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) are adopted as evaluation metrics. The results show that, among the 36 evaluation results formed by four datasets and three prediction intervals, BiHOD-Net achieves 28 best and 6 second-best results, with 94.4% of the results ranking among the top two. In particular, it achieves the best performance for all nine evaluation results on both the Beijing and NYC-Taxi datasets. For the 60 min prediction task, the MAE, RMSE, and MAPE of BiHOD-Net are 1.0389, 1.5087, and 47.6338%, respectively, on the Beijing dataset, and 0.8499, 0.8822, and 42.4931%, respectively, on the NYC-Taxi dataset. Ablation experiments show that, for the 60 min prediction task on the Beijing dataset, removing edge historical features, origin-destination role separation, O-D bipartite graph propagation, and the matrix marginal consistency constraint increases the MAE from 1.0389 for the complete model to 1.3728, 1.1084, 1.1048, and 1.0990, respectively. Additional key ablation experiments on the Chengdu, Nanjing, and NYC-Taxi datasets further demonstrate that these core designs provide stable improvements under different regional partition schemes and OD network structures. Repeated experiments using three random seeds, 42, 123, and 2024, also show that the prediction results remain generally stable across all datasets. These results demonstrate that BiHOD-Net can effectively characterize directional inter-regional OD traffic demand through origin-destination role separation, bidirectional bipartite graph propagation, explicit regional-pair interaction, and OD matrix marginal consistency constraints, providing favorable overall prediction performance, stability, and cross-dataset adaptability for dynamically sparse OD traffic flow forecasting.

As the number of training jobs in distributed deep neural network clusters continues to grow, communication overhead has become a significant factor affecting training completion time, making network sharing among multiple concurrent training jobs increasingly critical. Existing studies have mainly focused on distributed training communication optimization and network resource scheduling by employing task-level scheduling or single-job communication optimization to reduce communication overhead. However, task-level scheduling cannot differentiate the priorities of communication flows within the same training job, while single-job communication optimization fails to adequately address the competition among communication flows from multiple concurrent training jobs. Consequently, these approaches are unable to effectively mitigate the straggler effect caused by delayed critical communication flows during synchronous communication, leading to inefficient utilization of network resources. To address these challenges, a multi-job communication flow scheduling method based on flow-granularity priority assignment and path selection joint optimization is proposed. First, the network resource competition problem among multiple Ring-AllReduce training jobs is formulated as a mathematical optimization model. The model takes the priority assignment and path selection of all communication flows as input and estimates the throughput of each flow under the assumption of equal bandwidth sharing among flows within the same priority queue. The communication completion time of each training job is then estimated, and the overall communication-phase duration of all training jobs is minimized as the optimization objective. Since jointly optimizing the priority and routing configurations of all communication flows leads to exponential computational complexity, a heuristic solution combining local optimal scheduling and a runtime feedback mechanism is developed. By integrating binary search and backtracking techniques, the proposed algorithm significantly reduces computational complexity while maintaining high scheduling accuracy. During execution, the runtime feedback mechanism dynamically adjusts flow priorities according to detected straggler flows, thereby further alleviating the straggler effect in synchronous communication. To evaluate the effectiveness of the proposed method, a distributed training simulation platform based on container technology and virtual switches is developed to emulate both computing nodes and network switches. A communication flow scheduling system consisting of a centralized scheduler, node agents, and runtime monitoring modules is implemented. The centralized scheduler computes the priority assignment and routing decisions for communication flows and dynamically updates node traffic control policies and switch forwarding rules through remote procedure calls (RPC) and the OpenFlow protocol, respectively. Meanwhile, the runtime monitoring module continuously detects straggler communication flows during training and reports the monitoring results to the centralized scheduler, enabling adaptive optimization of scheduling decisions. Simulation experiments compare the proposed method with the priority assignment strategies of Crux and Gsched using communication-phase duration and cluster computational resource utilization as evaluation metrics. Experimental results demonstrate that, compared with Crux, the proposed method reduces communication-phase duration by 18.4% and improves cluster computational resource utilization by 10.4%. Compared with Gsched, it reduces communication-phase duration by 13.5% and improves computational resource utilization by 6.3%. In addition, ablation studies, approximation ratio analysis, and straggler-threshold sensitivity analysis further verify the effectiveness of the proposed flow priority assignment, path selection, and runtime feedback mechanisms. The results also indicate that a straggler threshold of ρ = 20% achieves the relatively good overall scheduling performance.

Intravoxel Incoherent Motion (IVIM) is a diffusion-weighted magnetic resonance imaging (DWI) analysis technique that enables noninvasive assessment of the tissue diffusion coefficient (D) and perfusion-related parameters (F, D*) without contrast agents, which are clinically important for the diagnosis and treatment of diseases such as liver cancer and brain tumors. Existing studies mainly focus on capturing spatially similar tissue features from DWI signals to assist IVIM parameter estimation and thus improve robustness to noise. However, they often ignore the inherent limitations of the DWI signals themselves: when the signals are severely contaminated by noise, the physiological characteristics contained in the signals are damaged, making it difficult to accurately capture the similarity relationships among features. Inspired by the principle of Deep Image Prior (DIP), this study identifies a phenomenon of Differences in Signal Quality (DSQ) in DWI signals during training, and further proposes a strategy for improving the quality of the training set during the training process based on this phenomenon. Specifically, the DSQ phenomenon indicates that, as training progresses, the overall quality of some fitted signals gradually becomes better than that of the original training signals. Therefore, the fitted signals are used as guidance to optimize the original training signals. Meanwhile, to further enhance the effect of data optimization, an iterative strategy is introduced to regulate the data optimization process, progressively improving the quality of the training signals and reducing the influence of noise on the fitting process. In addition, similarity among features is also widely present in both local neighborhoods and global space. Existing convolution-based methods are limited by the size of the receptive field and thus have difficulty in effectively capturing global information, whereas self-attention-based methods can perceive global correlations among voxels but often sacrifice local high-frequency details. Considering that features belonging to the same tissue usually contain similar physiological information and exhibit higher similarity, a tissue-prior-guided similar feature extraction scheme is proposed in this study. This scheme improves the quality of the fitted signals, thereby enhancing the effectiveness of the iterative optimization strategy and further improving the accuracy of IVIM parameter estimation. Based on the above improvements, a Data Iterative Neural Network (DI-NN) is proposed. To evaluate the performance of the model, the proposed DI-NN is compared with the traditional, voxel-based deep learning, and spatial-correlation-based deep learning methods. Root Mean Square Error (RMSE), Structural Similarity Index Measure (SSIM), and the Mann-Whitney U test are used as evaluation metrics. RMSE results show that across experiments with SNRs ranging from 8 to 50, DI-NN consistently achieves lower RMSE values and more accurate parameter estimation. For example, at SNR 15, the RMSEs of IVIM parameters F, D, and D* are reduced by 37.51%, 18.36%, and 30.92%, respectively, compared with the suboptimal method. SSIM results indicate that DI-NN preserves structural features better than other methods. The Mann-Whitney U test further demonstrates that DI-NN has a stronger capability to distinguish lesions.

Unsupervised anomaly detection identifies samples deviating from dominant distributions or local structural patterns without label information, yet it faces inherent challenges including scarce anomalies, feature overlap between normal and abnormal samples, and unstable anomaly patterns. Existing methods constructing anomaly scores from distance, density or fuzzy similarity still have notable limitations in complex scenarios: uniform attribute weighting weakens the effect of discriminative features, instance-level similarity computation is susceptible to isolated noise and local fluctuations, and there lacks effective integration between global fuzzy similarity information and local aggregation structures, making it hard to capture both global deviation and local sparsity of samples. To address these issues, this paper proposes FEGBAD, an unsupervised anomaly detection algorithm integrating fuzzy entropy and granular-ball density. It organizes original samples into granular-ball structures to build a connection between sample-level and local-region information, derives attribute weights via fuzzy entropy to measure the contribution of different attributes to anomaly identification, and aggregates sample-level fuzzy probabilities into granular-ball density to characterize the similarity between local regions and the dominant distribution, finally fusing attribute weights and granular-ball density to generate comprehensive anomaly scores. Comparative experiments on 18 public benchmark datasets against 8 representative methods demonstrate that FEGBAD achieves 10 best or tied-best AUC results and 11 best or tied-best AP results, with an average AUC of 0.9149 and an average AP of 0.6423, outperforming all comparative algorithms. Ablation studies verify that both core modules contribute positively to detection performance, with the granular-ball density module exerting a more significant effect on improving anomaly detection accuracy. Runtime analysis shows FEGBAD has an average running time of 59.889 seconds, which is longer than some lightweight baselines but significantly shorter than other complex fuzzy rough set algorithms. Overall, FEGBAD achieves a favorable trade-off between detection performance and computational cost by suppressing interference from low-information attributes and enhancing the characterization of local distribution structures; it is applicable to static tabular data scenarios, while its adaptability to strongly nonlinear structures, streaming data and graph-structured data still needs further improvement.

Aiming at the core bottlenecks in intelligent generation of Chinese calligraphy characters under few-shot scenarios, including the lack of detailed stroke features, distorted character structures, insufficient style expression and unstable readability of generated contents, this study proposes a refined controllable generation method for multi-style Chinese calligraphy (MSCDiff: A Controllable Generation Method for Multi-Style Chinese Calligraphy). The core research purpose is to break through the limitations of existing generation models in modeling traditional calligraphic artistic features under few-shot conditions, realize the automatic generation of multi-style calligraphy characters with both artistic expression and standard glyph structure, and provide an efficient and feasible technical solution for the digital inheritance, intelligent creation and stylized text rendering of Chinese calligraphy. In terms of implementation methods and key technologies, the proposed method takes the conditional diffusion model as the basic architecture and abandons the single conditional constraint mode. A multi-dimensional collaborative conditioning injection mechanism is constructed, which takes the target character content semantics, reference calligraphy style features and high-frequency stroke detail features as joint generation constraints, so as to enhance the model’s capability of style transfer, structural constraint and detail depiction for calligraphy characters from the bottom layer. To accurately capture fine-grained stylistic features such as flying white, stroke transitions and edge sharpness in calligraphy, a high-frequency feature extraction module is specially designed. Through high-frequency prior construction, learnable feature encoding and detail feature enhancement strategies, it strengthens the model’s representation and restoration of key artistic features such as stroke edges, flying white textures and local stroke transitions, solving the problems of vague fine artistic details and flattened style expression in traditional models. To ensure the rigor and stability of Chinese character structures and overcome the defects of structural dislocation and component distortion under few-shot conditions, a standard font library structure mapping mechanism and a dedicated content encoder are introduced to accurately extract the skeleton and layout features of target characters. On this basis, a multi-level feature fusion mechanism is built to hierarchically couple and adaptively fuse content structure features, global style features and high-frequency stroke features, and inject the integrated constraint information into the denoising generation process of the diffusion model in an orderly manner, realizing the collaborative regulation of structural, stylistic and detail constraints. To further improve the semantic accuracy and readability of generated results and avoid normative defects such as wrong characters, missing strokes and structural deformations, a character error correction module is innovatively introduced. A semantic supervision feedback mechanism is constructed based on the Connectionist Temporal Classification (CTC) loss, which takes the semantic compliance of characters as a reverse optimization signal to guide the iterative optimization of the generation network, significantly improving the content accuracy and practical reliability of generated characters from the semantic level. To support model training and comprehensive verification, this study independently constructs a high-quality dedicated Chinese calligraphy dataset Calli, which covers the representative writing styles of 11 calligraphers and contains more than 70,000 high-precision and standardized single-character images with standardized annotations and comprehensive style coverage, providing a high-quality experimental basis for few-shot multi-style calligraphy generation tasks. Multiple comparative and ablation experiments are carried out on the self-built Calli dataset and the public IAM handwriting dataset, with comprehensive evaluation from the dimensions of subjective visual effect, objective quantitative indicators and generalization performance. The experimental results show that the proposed MSCDiff method can stably generate Chinese calligraphy characters with complete stroke details, rigorous glyph structures, distinct calligraphic styles and excellent readability. Under the few-shot setting, the model achieves superior comprehensive performance in multiple aspects, including style fidelity, structural accuracy, detail restoration, and semantic consistency. It outperforms existing mainstream generation methods on most evaluation metrics, effectively alleviating the issues of missing details, structural misalignment, and insufficient style expression in calligraphy generation under few-shot scenarios, demonstrating excellent generation effectiveness, generalization ability, and practical value.This method not only achieves high-precision controllable generation of multi-style calligraphy characters, but also provides a referable technical path for the fields of traditional art digitization, intelligent content generation and few-shot visual modeling.

To address the degradation of fine-grained forgery traces, high-frequency information, and boundary artifacts caused by platform transcoding, H.264 compression, JPEG compression, and other forms of degradation during network transmission, which may further lead to local discriminative attention drift, global representation shift, and insufficient cross-quality generalization in deepfake detection models, a Compression-Paired Hyperspherical Consistency (CPHC) learning framework is proposed. Instead of simply treating samples with different compression qualities as ordinary data augmentation and mixing them during training, CPHC organizes the Raw, C23, and C40 versions of the same face frame into source-aligned multi-quality views and explicitly exploits the cross-quality correspondences along the compression degradation chain to learn stable forgery-discriminative representations. At the local spatial-response level, a Local Compressed Attention Consistency (LCAC) module is constructed to characterize discriminative regions under different compression views based on Class Activation Maps (CAMs). Pairwise CAM alignment between adjacent quality views and a shared CAM reference across the three views are employed to guide low-quality views to preserve stable local discriminative evidence, thereby suppressing interference from background noise, compression block artifacts, and irrelevant regions. At the global representation level, a Global Hyperspherical Consistency Constraint (GHCC) is designed to normalize image features onto a unit hypersphere. The von Mises–Fisher (vMF) distribution is further employed to constrain the consistency of class centers and feature distributions for the same real/fake label across different compression qualities, thereby alleviating global class-structure shifts caused by compression degradation. Experimental results show that CPHC achieves an average ACC of 87.02%±0.34% and an average AUC of 91.30%±0.32% on the C40 strongly compressed test set of FaceForensics++, outperforming the baseline by 4.50 and 5.69 percentage points, respectively, and achieving a higher average AUC than seven mainstream methods. Further evaluations under unseen JPEG75, JPEG50, and JPEG25 compression conditions, as well as on the cross-dataset CDF-v2, DFDC, and DeeperForensics benchmarks, demonstrate favorable generalization performance. In the cross-domain evaluation of diffusion-generated forgeries, with FF++ as the source domain and DiFF as the target domain, CPHC achieves an average AUC of 83.13% across four subsets. Ablation studies further show that LCAC and GHCC play complementary roles in stabilizing local discriminative regions and maintaining global class-structure consistency, respectively, and their joint use further improves detection performance under strong compression. These results demonstrate that CPHC effectively mitigates local response drift and global representation shift caused by compression degradation while maintaining stable detection and generalization performance under strong compression, unseen compression, and cross-domain scenarios.

Distributed photovoltaic (PV) stations exhibit decentralized data storage and significant non-independent and identically distributed (Non-IID) characteristics, causing conventional global federated learning methods to suffer from gradient conflicts during cross-site training. Meanwhile, newly joined stations face severe cold-start problems due to scarce historical data. To address these challenges, this paper proposes an Adaptive Granular-Ball Federated Learning (GB-FL) method that organically integrates Granular Computing (GrC) theory with a personalized federated learning framework, enabling adaptive grouping of heterogeneous sites and intra-granular collaborative modeling. First, a 45-dimensional site behavior characterization vector is constructed based on the statistical features of PV power curves to capture differences in power generation patterns across sites. Specifically, this includes 5-dimensional basic statistical features (daily average energy, peak power, peak time index, sunrise index, and sunset indices), 4-dimensional dynamic characteristics (ramp-up rate, ramp-down rate, intraday fluctuation intensity, and clipping ratio), as well as 24-dimensional daily average curve profile and 12-dimensional intraday fluctuation profile features. All features are processed via z-score standardization for subsequent granular-ball partitioning. On this basis, a three-stage adaptive granular-ball construction algorithm comprising "heterogeneity-driven splitting–proximity-based merging–micro-ball fusion" is designed. Specifically, during the heterogeneity-driven splitting stage, intra-granular average Euclidean distance is employed to measure heterogeneity, and a weighted scoring mechanism combined with a "worst-first" greedy strategy is introduced for recursive bisection. In the proximity-based merging stage, a dynamic threshold is defined by the median of inter-granular center distances to merge over-fragmented granular balls. In the micro-ball fusion stage, residual single-site granular balls are incorporated into neighboring granular balls satisfying heterogeneity constraints. The synergy of these three stages achieves adaptive partitioning without pre-specified cluster numbers, dividing 32 sites into 5 granular-ball structures with intra-granular homogeneity and inter-granular mutual exclusivity. Building upon the granular-ball partitioning results, each granular ball independently executes personalized federated learning. The model adopts a two-layer gated recurrent unit (GRU) with a hidden dimension of 64 as a shared temporal encoder to extract common dynamic patterns, while retaining a personalized linear prediction head for each site to fit local characteristics. Local training employs the smooth Huber loss function, and the server performs weighted aggregation of shared encoder parameters based on sample sizes. Model parameters across different granular balls remain mutually isolated, thereby effectively alleviating gradient conflicts induced by heterogeneous data. To address the cold-start problem for new stations, a prior initialization strategy based on the nearest granular ball is further proposed. Specifically, a new station is first assigned to the most similar granular ball according to its behavioral characterization, inherits the shared encoder parameters of that granular ball, and achieves rapid adaptation through few-shot fine-tuning. Experiments are conducted on data from 32 actual PV stations in Shandong Province (each containing 35,040 15-min sampling points) and the La Trobe UNISOLAR public dataset (32 stations, approximately 76,000 fifteen-minute sampling points). Data are partitioned into training, validation, and test sets in a 40%/20%/40% ratio, and experiments are repeated with 5 fixed random seeds to ensure result stability. The results demonstrate that, on the Shandong dataset, GB-FL achieves a 3.8% reduction in root mean square error (RMSE) compared with Personalized Federated Learning (PFL) (15.896 kW vs. 16.517 kW) and a 7.4% reduction in mean absolute error (MAE); a 6.0% reduction in RMSE compared with FedProx; and 2.2% and 4.6% reductions in RMSE and MAE, respectively, compared with ClusterFL. In the cold-start scenario, compared with local independent training (Local), RMSE is reduced by 7.2% (8.791 kW vs. 9.475 kW) and MAE by 11.7%. On the UNISOLAR dataset, GB-FL outperforms FeDaL in both MAE (2.161 kW) and the coefficient of determination (R2) (0.796), while its MAE standard deviation (0.049) is significantly lower than that of FeDaL (0.100), validating the method's cross-dataset generalization capability and robustness. Ablation experiments further indicate that removing the granular-ball grouping mechanism leads to a 3.91% increase in RMSE, demonstrating that the adaptive granularization strategy is the key factor underlying the performance improvement.

To address the difficulty of coordinating local sequence reconstruction with global distribution fitting in generative multivariate time-series anomaly detection, as well as the tendency of such models to reconstruct anomalous signals with high fidelity, thereby narrowing the reconstruction-error gap between normal and anomalous samples, this paper proposes a Progressive Latent Diffusion-Driven Anomaly Detection framework for multivariate time series, termed PLDAD. The framework adopts a three-stage progressive training paradigm consisting of Masked Function Autoencoder pretraining, Latent Prior Learning, and end-to-end Joint Fine-Tuning, which progressively accomplishes basic feature representation, normal-prior modeling, and global collaborative optimization. First, during the Masked Function Autoencoder (MFAE) stage, element-wise independent masking is applied to the input sequence, and a dual-branch encoding structure is employed to extract inter-variable correlations and multiscale temporal dependencies, respectively. The resulting spatiotemporal features are subsequently fused through a self-attention mechanism, and the complete sequence is reconstructed using a cross-attention-based continuous-time decoder, thereby constructing a smooth and compact latent feature space. Second, during the Latent Prior Learning (LPL) stage, the pretrained encoder is frozen, and the complete unmasked sequence is mapped into initial latent features. A forward noising process and a reverse denoising process are then established in the latent space. Unlike conventional diffusion models that take Gaussian noise as the prediction target, the proposed Latent Feature Diffusion Model (LFDM) directly predicts generated latent features, guiding perturbed representations toward the normal feature manifold under the constraint of the normal-data prior. This reduces the likelihood that anomalous information is propagated to the decoder and reconstructed with high fidelity. Finally, during the Joint Fine-Tuning (JFT) stage, the encoder, LFDM, and decoder are unfrozen and optimized end to end using a joint objective composed of reconstruction loss and diffusion loss, enabling local sequence recovery and global prior constraints to be coordinated. During inference, a point-wise anomaly score is calculated from the squared error between the actual observations and their reconstructed values. Experiments are conducted on six public benchmark datasets, namely ASD, MSL, PSM, SMAP, SMD, and SWaT, comprising more than 3.4 million observations, and the proposed method is compared with 11 representative approaches. Model performance is comprehensively evaluated using the F1-score and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). The results show that PLDAD achieves the best F1-score and AUC-ROC results on all six datasets, with an average F1-score of 94.31%, demonstrating stable detection performance across different data distributions and anomaly patterns. Its F1-scores on ASD, MSL, PSM, SMAP, SMD, and SWaT reach 95.12%, 94.01%, 96.51%, 88.12%, 95.67%, and 96.40%, respectively, while its AUC-ROC values on ASD, PSM, and SMD reach 88.13%, 80.12%, and 83.27%, respectively, indicating favorable overall performance in both anomaly-segment detection and the global discrimination of anomaly scores. Statistical significance analysis shows that PLDAD obtains the highest average rank of 1.0000, further confirming its stable advantages over the baseline methods. Ablation experiments demonstrate that progressive training, latent prior learning, LFDM, joint fine-tuning, and the feature purification mechanism all contribute to improved detection performance. Latent-feature visualization further reveals a clearer distribution boundary between normal and anomalous samples in the complete model. Time-efficiency experiments show that PLDAD maintains acceptable computational overhead while preserving high detection accuracy. In summary, PLDAD can alleviate the optimization conflict between local sequence reconstruction and global distribution fitting, reduce the tendency to reconstruct anomalous signals with high fidelity, and achieve favorable detection accuracy, stability, and application potential in complex multivariate time-series scenarios.

To address large layout variations, confusion between similar fields, unstable image quality, weak representation of two-dimensional spatial relations after one-dimensional serialization of image patches, and the lack of region constraints in text-only outputs, this paper proposes a logistics certificate information extraction method that combines decoupled two-dimensional positional encoding, coordinate-supervised extraction, and multi-task instruction mixing. Before the downsampled two-dimensional visual features are projected into the language model embedding space, separate learnable positional embeddings are introduced for the row and column directions. The row and column position parameters are maintained independently and concatenated along the channel dimension to form a two-dimensional positional prior. This design supplements row-column relations, local adjacency, and layout structure with low parameter overhead, reduces the loss of spatial topology caused by image patch flattening, and improves the model’s ability to distinguish elongated text regions, dense fields, and adjacent areas. During information extraction, the conventional text-only question-answering task is reformulated as the joint generation of field text and region coordinates. The model outputs field content and normalized bounding boxes in a structured sequence, thereby establishing an explicit correspondence between field semantics and image regions and improving the spatial traceability and verifiability of the extraction results. To reduce the possible optimization conflict between text generation and coordinate prediction caused by direct coordinate supervision, three instruction tasks are further constructed, including joint extraction, visual localization, and text-only extraction. These tasks are randomly mixed at preset ratios, allowing the model to learn spatial localization while maintaining stable text recognition and instruction understanding. A logistics certificate dataset containing 4,000 images was constructed, covering driving licenses, qualification certificates, road transport permits, and vehicle licenses. The dataset includes multiple layouts and image acquisition conditions, with characteristics such as dense fields, elongated text regions, adjacent similar fields, and seal occlusion. Text content and corresponding bounding boxes were annotated for key fields and converted into instruction samples for text extraction, visual localization, and joint text-coordinate extraction. Public datasets, including DocVQA, XFUND-ZH, LSVT, and LLaVA-Instruct-ZH, were further used for mixed training to improve the model’s adaptability to general document structures, Chinese layouts, and complex visual interference. InternVL2.5-2B was used as the main backbone, while InternVL3-2B and InternVL3-8B were used for extended evaluation. Average Normalized Levenshtein Similarity (ANLS), Exact Match (EM), Mean Intersection over Union (Mean IoU), and Grounded Accuracy were adopted to evaluate text recognition, region localization, and text-coordinate consistency. The experimental results show that the proposed method clearly improves field localization while maintaining high text recognition performance. Compared with the InternVL2.5-2B baseline, the macro-average ANLS and Mean IoU over the four logistics certificate categories increase by 3.11 and 11.23 percentage points, respectively. On the road transport permit dataset, Mean IoU increases from 13.42% to 48.39%, while Grounded Accuracy increases from 17.32% to 62.09%. On the driving license dataset, ANLS and Grounded Accuracy reach 98.00% and 48.97%, respectively. Results from three independent runs on InternVL3-8B show macro-average improvements of 0.31, 1.00, 1.28, and 1.46 percentage points in ANLS, EM, Mean IoU, and Grounded Accuracy, respectively. The standard deviations remain close to those of the baseline, indicating stable gains at a larger model scale. Ablation experiments show that direct coordinate supervision alone makes it difficult to balance text generation and spatial localization. Adding multi-task instruction mixing or decoupled two-dimensional positional encoding clearly improves both Mean IoU and Grounded Accuracy, while their combination achieves the best overall performance. Evaluation on XFUND-ZH and DocVQA further shows that the proposed method improves spatial modeling for structured documents without weakening general document question-answering ability. These results confirm the effectiveness of jointly modeling explicit two-dimensional spatial priors and coordinate supervision for complex logistics certificate information extraction and provide technical support for traceable information extraction and structured document understanding in real-world applications.

This paper proposes a security defense method based on the coordination of statistical and geometric domains to cut off the confusion space association in order to address the problem of severe privacy leakage risks in the target oriented adversarial samples generated by image editing technologies such as generative adversarial networks, while achieving pixel level manipulation at the semantic level under high-dimensional chaotic spatial disturbances. The research aims to solve the core bottleneck of existing single encryption or scrambling methods that are difficult to completely separate the two types of orthogonal spatial correlations that attackers rely on to reverse restore sensitive information, namely the local grayscale statistical correlation between pixel values and the global geometric layout correlation between pixel coordinates. This is to construct a full link privacy protection mechanism from encryption to transmission. In terms of technical implementation, this method consists of three layers of collaborative modules. The first layer is the statistical domain association pruning module: using the hash value of adversarial sensitive images to drive logistic mapping to generate content related dynamic S-boxes, and dynamically constructing a 16 × 16 nonlinear replacement table based on the image's own characteristics; Using the Chen hyper chaotic system to generate four sets of chaotic sequences, combined with the coordinate scrambling matrix generated by the two-dimensional Logistic chaotic system, position scrambling and multi-stage bidirectional diffusion are performed on image pixels to homogenize high-dimensional perturbation energy to the global grayscale distribution, thereby cutting off the correlation between neighboring grayscale distribution trajectories and local structural statistical patterns. The second layer is the geometric domain correlation cutting module: the encrypted image in the statistical domain is subjected to two-dimensional integer wavelet decomposition to obtain a sparse coefficient matrix, and a compressed sensing measurement matrix driven by Iterative Chaotic Map (ICM) is introduced for dimension compression; Introducing the Non Periodic Affine Nonlinear (FAN) transformation, a generalized affine scrambling mechanism that includes an infinite number of transformations, to globally remap compressed images, breaking down the spatial topology of pixels, supplemented by a closed bidirectional diffusion link driven by ICM chaotic mapping, allowing any single bit change to diffuse uniformly to the entire matrix, thereby eliminating global layout correlations within the geometric domain. The third layer is the covert transmission module: the encrypted matrix processed in the first two layers is segmented and embedded into the red, green, and blue (RGB) channels of the ordinary carrier image, as well as all 8 bits of the alpha transparent channel. The embedding position map is dynamically generated through a chaotic sequence, so that the ciphertext is distributed discretely and uniformly in space, achieving covert transmission of encrypted data without visual loss. The experimental results show that in low, medium, and high disturbance intensity scenarios, the attack success rates of our method are 0.782%, 3.469%, and 5.887%, respectively, which are 79.9%, 60.4%, and 61.6% lower than the optimal comparison method; The grayscale histogram of the encrypted image shows an absolutely flat distribution, with an information entropy of 7.988. The correlation coefficients between adjacent pixels in the horizontal, vertical, and diagonal directions are all below 0.004, indicating that the statistical independence between pixels has reached an ideal level; The Number of Pixels Change Rate (NPCR) and Unified Average Changing Intensity (UACI) are 99.61% and 33.28%, respectively, which are highly consistent with the theoretical ideal values and demonstrate strong resistance to differential cryptanalysis; In the key avalanche effect test, a 1-bit key flip caused a 99.58% change in ciphertext pixels, approaching the theoretical extreme value, verifying extremely high key sensitivity. In terms of concealment, the Peak Signal-to-Noise Ratio (PSNR) of the carrier image after steganography reached 49.2 dB, and the Structural Similarity Index (SSIM) was 0.9978, making it difficult for the human eye to distinguish between the original carrier and the steganographic carrier; The ablation experiment further quantified the independent contributions of each component: after completely removing the statistical or geometric domain cuts, the attack success rates soared to 15.342% and 18.674%, respectively; After removing dynamic S-boxes, multi-stage diffusion, FAN transformation, and bidirectional diffusion separately, the success rates of attacks increased to 9.113%, 11.456%, 12.237%, and 8.641%, respectively, verifying the irreplaceability of the modules and sub components designed in this paper in the overall defense system. In summary, this article theoretically elucidates the orthogonal independence of the correlation between the statistical domain and the geometric domain. Based on this, a targeted collaborative disconnection mechanism can achieve full link privacy protection for adversarial sensitive images from encryption to transmission while maintaining computational efficiency, providing an effective technical solution for image security defense in open network environments.

Business process remaining time prediction is a core task in predictive process monitoring. Its accuracy provides critical decision support for process optimization, resource scheduling, anomaly early warning, and Quality of Service (QoS) guarantee. However, existing methods generally suffer from two major limitations. First, deep learning models based on Long Short-Term Memory (LSTM) or Transformer, while capable of learning complex temporal dependencies, often rely on recurrent structures or self-attention mechanisms. These models involve large parameter scales and high computational complexity, leading to significant inference latency for long sequences and hindering real-time deployment in computational resource-constrained environments. Second, the utilization of temporal information remains coarse-grained in current research. Most methods rely on single temporal attributes such as timestamps or event intervals, ignoring the importance variations of multi-dimensional temporal semantics—including duration, cumulative execution time, and waiting time—across different process stages. This failure to fully exploit key driving factors limits overall prediction performance. To address these issues, this paper proposes a lightweight business process remaining time prediction method based on adaptive time-weighting, termed LTW-Mixer. First, a multi-dimensional temporal feature system is systematically constructed, covering duration, cumulative execution time, waiting time, resource response time, and periodic information (e.g., day of the week and hour of the day). Building on this, an adaptive time-weighting mechanism is designed to dynamically adjust the contribution of each temporal feature based on the process context. This mechanism enhances key temporal information while suppressing redundant features, thereby improving the model's ability to capture critical time-driven factors. Finally, to balance prediction accuracy and model efficiency, a lightweight backbone based on MLP-Mixer is developed. Sequence mixing and channel mixing are employed to replace the computationally intensive self-attention mechanism, achieving efficient interaction across both temporal and feature dimensions. Additionally, a gated fusion mechanism is introduced to dynamically integrate activity and temporal semantics, allowing the model to adaptively regulate the contribution ratios of heterogeneous features and effectively capture long-range dependencies while maintaining low computational overhead. Experimental results on six public real-world business process datasets demonstrate that LTW-Mixer achieves the lowest Mean Absolute Error (MAE) across all datasets. Specifically, for the BPIC2015_2 and BPIC2015_5 datasets, the MAE is reduced by approximately 26.7% and 39.5%, respectively, compared to the state-of-the-art baseline models. Ablation studies confirm that the multi-dimensional temporal features, adaptive weighting mechanism, gated fusion, and MLP-Mixer backbone all contribute significantly to the model's performance. Furthermore, feature weight analysis and sensitivity experiments show that the model can dynamically prioritize key factors such as waiting time and resource response time across different business scenarios, demonstrating strong interpretability. In terms of lightweight performance, LTW-Mixer requires only 0.19 M parameters and 0.71 MB of storage, with a peak GPU memory usage of 57.15 MB and a per-sample inference time of only 0.023 ms. These results outperform all comparison models, proving that LTW-Mixer significantly reduces storage and computational costs while ensuring high accuracy, thus exhibiting excellent real-time responsiveness and potential for industrial deployment.

Hierarchical Classification (HC) exploits label hierarchies to model multi-granularity semantic dependencies and has important applications in remote-sensing and natural-image analysis. However, real-world degradations, such as noise, occlusion, blur, and low resolution, weaken fine-grained discriminative cues, making a fixed top-down prediction paradigm that always terminates at a leaf node prone to over-refinement, cross-level error propagation, and structural inconsistency. Existing methods primarily emphasize parent–child subsumption consistency, while rarely modeling mutual exclusion among categories at the same level explicitly, thereby limiting the global consistency and robustness of hierarchical predictions. To address the dynamic variation in recognizable granularity under image degradation, the separation between hierarchical-depth selection and category-path prediction, and the inability of local node scores to guarantee globally valid paths, this work further formulates hierarchical classification of degraded images as a granularity-adaptive level–path joint structured inference task. The objective is to avoid forcing every sample to a leaf node, align the predicted depth with the currently recognizable granularity, and simultaneously satisfy parent–child consistency and same-level mutual-exclusion constraints. To this end, we propose VS-HESS, a structurally consistent hierarchical classification framework for degraded scenarios. First, VS-HESS introduces an exclusion-enhanced level–category hybrid prediction paradigm that jointly predicts the deepest level at which a sample can be reliably recognized and its corresponding category path within a unified framework. Second, it incorporates a Hierarchical Exclusion Constraint Graph (HECG) and a state-space matrix to explicitly model same-level category exclusion in addition to parent–child subsumption constraints, and performs consistency-aware inference through structured normalization and marginalization. Third, a hierarchy-semantic-guided visual–semantic cross-attention module uses label-semantic embeddings to guide visual-token aggregation and learns category-specific, cross-granularity discriminative representations. Specifically, a Vision Transformer first performs global visual encoding of a degraded image. Hierarchical label embeddings are then used as queries to aggregate visual tokens in a category-conditioned manner, producing semantic-aware representations and node compatibility scores for categories at different levels. A hierarchical-depth prediction head estimates the deepest semantic level supported by the current visual evidence. On this basis, the HECG jointly represents parent–child subsumption edges and same-level exclusion edges, while the set of structurally valid paths is encoded into a state-space matrix. Consequently, the model no longer makes decisions solely from mutually independent local node scores, but instead performs structured normalization, marginal-probability computation, and final level–path joint selection within the valid path space. The model is optimized end to end using hierarchical-depth supervision, path supervision, and structural-consistency supervision, thereby unifying granularity-adaptive decision making, local category discrimination, and global structured inference in a single training process. On the HRSC-Deg and CUB-Deg degraded hierarchical datasets, VS-HESS achieves the best numerical results across all reported metrics. On HRSC-Deg, the inverse symmetric difference metric, hierarchical precision, hierarchical recall, overall level accuracy, coarse-level accuracy, and fine-level accuracy reach (72.89%), (91.53%), (91.48%), (76.05%), (99.11%), and (90.42%), respectively. On CUB-Deg, the corresponding four overall metrics and the order-, family-, and species-level accuracies reach (65.72%), (90.80%), (92.94%), (73.71%), (98.96%), (92.11%), and (81.12%), respectively. These results demonstrate that visual–semantic evidence encoding, dynamic level selection, and exclusion-enhanced valid-path inference can work synergistically to improve hierarchical path matching and medium- to fine-grained recognition under image degradation. By selecting an appropriate prediction granularity according to the semantic information retained in a degraded image and making structurally consistent decisions within a path space that satisfies both parent–child subsumption and same-level exclusion relations, VS-HESS provides an effective modeling framework for structurally consistent and granularity-adaptive hierarchical image classification across application domains such as remote-sensing ship recognition and natural-image analysis.

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.