In the context of ongoing advancements in educational informatization, constructing precise and efficient curriculum knowledge graphs has become key to promoting personalized education development. As a structured knowledge representation model, curriculum knowledge graphs reveal complex relations between curriculum content and learning objectives to optimize the allocation of educational resources, and tailoring personalized learning paths for learners. This survey presents a discussion around the techniques used to construct curriculum knowledge graphs, starting with an explanation of the basic concepts; intrinsic connections; and significant differences among general, educational, and curriculum knowledge graphs. It then delves into the key technologies used for building curriculum knowledge graphs, covering aspects such as curriculum ontology design, entity extraction, and relation extraction, and provides a detailed analysis and summary of their evolution, key features, and limitations. Furthermore, it explores the application value of curriculum knowledge graphs in scenarios such as learning resource recommendation, learner behavior profile and modeling, and multimodal curriculum knowledge graph construction. Finally, it focuses on the challenges in constructing curriculum knowledge graphs, such as data diversity and heterogeneity, difficulties in quality evaluation, and the lack of cross-curriculum integration, and provides future-oriented insights based on cutting-edge technologies such as deep learning and Large Language Models (LLMs).

Large Language Models (LLMs) have made significant progress in dialogue, reasoning, and knowledge retention. However, they still face challenges in terms of factual accuracy, knowledge updates, and a lack of high-quality domain datasets for handling knowledge-intensive tasks in the electricity sector. This study aims to address these challenges by introducing an improved Retrieval-Augmented Generation (RAG) strategy. This strategy combines hybrid retrieval with a fine-tuned generative model for efficient knowledge capturing and updating. The Metadata-driven RAG framework (Meta-RAG) is proposed for knowledge Question Answering (QA) tasks in the electricity domain. This includes data preparation, model fine-tuning, and reasoning retrieval stages. The data-preparation stage involves document conversion, metadata extraction and enhancement, and document parsing. These processes ensure efficient indexing and structured processing of power regulation documents. The Electricity Question Answering (EleQA) dataset, consisting of 19 560 QA pairs, is constructed specifically for this sector. The model fine-tuning stage uses multi-question generation, chain-of-thought prompting, and supervised instruction fine-tuning to optimize the reasoning abilities in specific tasks. The retrieval reasoning stage employs mixed encoding and re-ranking strategies, combining retrieval and generation modules to improve answer accuracy and relevance. Experiments validate the effectiveness of Meta-RAG. Compared to baseline models such as Self-RAG, Corrective-RAG, Adaptive-RAG, and RA-ISF, Meta-RAG shows higher answer accuracy and retrieval hit rates. Meta-RAG with the Qwen1.5-14B-Chat model achieves an overall accuracy of 0.804 3, surpassing the other methods. Ablation and document recall experiments indicate that document retrieval significantly impacts the framework performance, with a 0.292 8 drop in accuracy when the retrieval capability is lost.

When implementing Federated Learning (FL) in a cell-free network environment, user scheduling and resource allocation strategies are crucial for optimizing system time overhead, improving user reachability, and accelerating FL convergence rate. To address the issue of uneven resource allocation, this study designs an optimization scheme that combines user scheduling, CPU processing frequency, and power allocation. This scheme aims to achieve fair resource allocation by maximizing the minimum user rate in the system, thus enhancing FL performance. The joint optimization problem is decomposed into two subproblems: user scheduling and power allocation. For user scheduling, this study proposes a greedy scheduling algorithm based on k-means clustering to comprehensively evaluate channel conditions and data "value" of users and categorize users into different groups. Subsequently, for the resource occupation situation, a personalized CPU processing frequency allocation plan is developed for users within each group based on their resource occupancy. Finally, by independently executing user scheduling within each group, user selection is performed efficiently and precisely, and the complexity of user selection is effectively reduced via early grouping. For power allocation, this study introduces a Bisection Method-based Power Allocation (BM-PA) algorithm. This algorithm not only considers fairness among users but also prioritizes resource-constrained users to ensure that they can obtain superior resource allocation. The BM-PA algorithm achieves fast convergence of power allocation using a low-complexity iterative optimization process, significantly improving the resource utilization efficiency without deteriorating the system performance. In this study, a reasonable user scheduling strategy serves as the foundation for obtaining optimal solutions for the power allocation subproblem. This study adopts an alternating iteration method that allows independent optimization in each subproblem while considering the solution of the other subproblem. Via multiple rounds of iterative optimization, this interdependent relationship ensures that power resources are reasonably allocated to users who need them the most or are most likely to effectively utilize them, thus enhancing the overall system performance. This study realizes joint optimization solutions that significantly improve overall system performance. Simulation results show that compared with the baseline algorithm, the proposed algorithm exhibits outstanding performance in terms of downlink achievable rates-the average improvement reaches up to 103.34% under optimal conditions. Additionally, the uplink achievable rates improve by up to 102.78%. Furthermore, the proposed algorithm can save 67.44% of the FL task training time on average compared to the baseline algorithm, particularly when the FL learning model accuracy reaches 90%, wherein the time overhead of the proposed algorithm is minimal.

Post-Training Quantization (PTQ) is an efficient model compression method that converts the parameters of high-precision floating-point models into low-bit integer representations without requiring retraining, using only a small amount of unlabeled calibration data. This method significantly reduces storage and computational overhead while maximizing the retention of the original model's inference accuracy; therefore, it is widely recognized and adopted in both academia and industry. This paper systematically summarizes the progress of research on PTQ from four dimensions: quantization steps, method classification, tool ecosystem, and application advancements. First, a clear framework for the quantization process is constructed, covering steps such as dynamic range statistics, quantization parameter calculation, weight and activation quantization, error optimization, and model generation. Second, a complete classification system for quantization methods is proposed, which includes quantization granularity, bit width, calibration methods, and structure-guided quantization. Third, the tool ecosystem supporting the large-scale application of PTQ is analyzed, and its value in hardware adaptation and engineering deployment is discussed. Finally, this paper summarizes the progress in the integration and application of PTQ methods and highlights practical challenges, particularly those related to cross-modal consistency, extremely low-bit semantic collapse, and hardware adaptation. These challenges not only reveal the limitations of current technologies but also provide important directions for future research. This review provides a reference framework for PTQ methods in academia and industry, thereby facilitating the widespread application of artificial intelligence in resource-constrained scenarios.

The rapid development of blockchain technology and the rise of cryptocurrencies not only challenge traditional financial and asset management concepts but also introduce significant privacy and security concerns. Coin mixing, a primary privacy protection mechanism in blockchains, is designed to enhance transaction anonymity and protect user identities. However, this enhanced anonymity also enables the misuse of cryptocurrencies for illicit activities such as money laundering, theft, and fraud. Therefore, a comprehensive review of technologies for the identification, detection, and regulation of coin mixing services is urgently required to support law enforcement and promote the secure and compliant use of cryptocurrencies. This paper first reviews the technical principles and mechanisms of coin mixing services and proposes a general classification framework for mixing tools while discussing the criminal activities associated with them. Subsequently, it focuses on recent advances in five key regulatory approaches: manual rule-based methods, address clustering, entity classification, anomalous transaction detection, and mixing service traceability. It also surveys existing industrial tools and platforms related to coin mixing regulations. Finally, it identifies current challenges in the field and suggests directions for future research.

As multivariate time series data become increasingly prevalent across various industries, anomaly detection methods that can ensure the stable operation and security of systems have become crucial. Owing to the inherent complexity and dynamic nature of multivariate time series data, higher demands are placed on anomaly detection algorithms. To address the inefficiencies of existing anomaly detection methods in processing high-dimensional data with complex variable relations, this study proposes an anomaly detection algorithm for multivariate time series data, based on Graph Neural Networks (GNNs) and a diffusion model, named GRD. By leveraging node embedding and graph structure learning, GRD algorithm proficiently captures the relations between variables and refines features through a Gated Recurrent Unit (GRU) and a Denoising Diffusion Probabilistic Model (DDPM), thereby facilitating precise anomaly detection. Traditional assessment methods often employ a Point-Adjustment (PA) protocol that involves pre-scoring, substantially overestimating an algorithm's capability. To reflect model performance realistically, this work adopts a new evaluation protocol along with new metrics. The GRD algorithm demonstrates F1@k scores of 0.741 4, 0.801 7, and 0.767 1 on three public datasets. These results indicate that GRD algorithm consistently outperforms existing methods, with notable advantages in the processing of high-dimensional data, thereby underscoring its practicality and robustness in real-world anomaly detection applications.

The Siamese tracking network is a popular target tracking framework that includes three modules: backbone, fusion, and positioning networks. The Transformer is a relatively new and effective implementation method for fusion network modules. The encoder and decoder of the Transformer use a self-attention mechanism to enhance the features of the Convolutional Neural Network (CNN). However, the self-attention mechanism can only enhance features in the spatial dimension without considering feature enhancement in the channel dimension. To enable the self-attention network of the Transformer to enhance features both in the spatial and channel dimensions and provide accurate correlation information for the target localization network, a Transformer tracker based on dual-dimensional feature enhancement is proposed to improve the Transformer fusion network. First, using the third- and fourth-stage features of the backbone network as inputs, channel dimension feature enhancement is performed via CAE-Net in the self-attention module of the Transformer encoder and decoder to enhance the importance of the channel. Subsequently, two-stage feature-weighted fusion and linear transformation are performed via SAE-Net to obtain the self-attention factors Q, K, and V. Finally, spatial dimension feature enhancement is performed via a self-attention operation. Experiments conducted on five widely used public benchmark datasets reveal that the improved Transformer feature fusion module can improve the tracking performance of the tracker with minimal reduction in speed of tracking.

】Multimodal sentiment analysis leverages the complementary information of speech, text, and visual modalities to enhance emotion recognition accuracy and robustness. However, existing approaches still face three major challenges: (1) the lack of unified modeling for multi-scale emotional dynamics across fast and slow temporal rhythms; (2) the difficulty in explicitly characterizing semantic dominance and subordination among modalities; and (3) the limited ability to adaptively regulate modality intensity and information contribution. To address these issues, this paper proposes a multimodal sentiment analysis framework that integrates multi-scale encoding with a polarity-aware fusion mechanism. Specifically, a Multi-Scale Mamba encoder (MS-Mamba) is introduced for visual and audio modalities to jointly capture global and local temporal dependencies; a Polarity-Aware Fusion (PAF) module is designed to explicitly model inter-modal enhancement and suppression through semantic residuals and signed weights; and a Polarity-Driven Gating (PDG) mechanism is developed to adaptively control information flow via a saliency–direction disentanglement strategy. These components collaboratively form a closed-loop structure of “temporal modeling–polarity alignment–global gating.” Experimental results on the CMU-MOSI and CMU-MOSEI datasets demonstrate that the proposed model achieves binary classification accuracies of 86.58% and 86.50%, with F1 scores of 86.59% and 86.26%, respectively—yielding an average improvement of approximately 1.3% over mainstream baselines. The results validate the effectiveness and robustness of the proposed method in semantic alignment, temporal modeling, and adaptive fusion.

In the modern industrial sector, the perception and analysis of text data are essential for promoting intelligent manufacturing and optimizing production processes. However, industrial text data are typically characterized by high specialization, diversity, complexity, and annotation costs, making traditional large-scale annotation methods unsuitable. Existing few-shot Named Entity Recognition (NER) methods often use prototypical networks to classify entities, where the prototype is the average of the features of all samples belonging to the same category. However, these methods are highly sensitive to the support set data and prone to sample-selection bias. To address this, a few-shot NER model based on Distribution Calibration called DC-NER is proposed. A two-stage approach for span detection and entity classification is adopted for the few-shot NER task. In the first stage, an entity extractor is used to select candidate entities. In the second stage, a type discriminator assigns these entities to predefined categories. To address the challenge of capturing the true class distribution with a small number of labeled samples in few-shot NER, which leads to inaccurate class prototype calculations, a method is proposed that leverages useful information from the source domain data to calibrate the class distribution in the target domain. Thus, more samples are generated based on the refined class distribution to construct more accurate prototypes, thereby improving the performance in few-shot NER tasks. Experiments on both the in-domain dataset Few-NERD and the cross-domain dataset Cross-NER reveal that DC-NER significantly outperforms the comparative models in terms of F1 value, validating its effectiveness for few-shot NER.

Existing Knowledge Graph Completion (KGC) methods fail to effectively distinguish semantic differences among paths at different levels and cannot dynamically adjust relation representations by fully leveraging neighborhood context information. Consequently, contextual semantics remain insufficiently understood. To address these issues, this paper proposes a KGC model integrating relational context semantics, named RCSKGC, for enhancing the semantic representation of learned paths and neighborhood information. First, the local and global encoding of multi-hop paths at different levels is performed using a Bidirectional Gated Recurrent Unit (Bi-GRU) and an attention mechanism to extract effective path information. Meanwhile, the fine-grained semantic features of the path information at different levels are further enhanced through relational embedding contrastive learning. Second, a dual-attention mechanism and dynamic weighting strategy are employed to capture the neighborhood-level information of relations and learn their semantics through a "neighborhood—entity—relation" approach. Finally, the two types of relational representations are aggregated and used as inputs for the decoder to perform KGC. The results of experiments on the FB15k-237 dataset demonstrate that, compared to the best results of the baseline methods, RCSKGC achieves improvements of 1.4, 0.8, 1.3, and 2.1 percentage points in the Mean Reciprocal Rank (MRR), Hits@1, Hits@3, and Hits@10, respectively. Further, on the WN18RR dataset, RCSKGC shows improvements of 0.8 and 1.0 percentage points in MRR and Hits@3, respectively. These results validate the effectiveness of the proposed method.

Vision-and-Language Navigation (VLN) tasks aim to guide an agent to move to a target location in a 3D or real-world environment based on language instructions. However, the traditional end-to-end VLN algorithms have certain limitations. When an erroneous action occurs during navigation planning, the agent tends to enter incorrect paths, resulting in an inability to continue following the instructions or exploring unnecessary areas. To address this issue, an agent called Nav-Explore is proposed, which is based on a Large Language Model (LLM) and an exploration module. The proposed agent leverages the reasoning capabilities of the LLM to predict the next action according to the language instructions and current visual information and uses an exploration module to balance exploration and utilization. The exploration module employs an ε-greedy strategy for toggling between the normal navigation and exploration modes. When the random probability is below ε, the agent explores possible future paths to assess the feasibility of the next actions, thus avoiding incorrect decisions. If the probability exceeds ε, it directly uses the output of the LLM for navigation. This modular design enables Nav-Explore to effectively enhance navigation success rates and improve the generalization ability of the agent in unseen environments. The experimental results demonstrate that Nav-Explore achieves superior performance on two outdoor VLN benchmark datasets, Touchdown and Map2seq, significantly increasing navigation success rates. Furthermore, Nav-Explore exhibits strong generalization capabilities, effectively completing navigation tasks in different environments.

Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain, thereby improving the performance of the target domain model. However, traditional UDA methods assume that the category spaces of the source and target domains are identical; hence, they cannot handle unknown categories present in the target domain, which limits their applicability in real-world scenarios. Open-Set Domain Adaptation (OSDA) addresses this issue by introducing the recognition of unknown categories. Nevertheless, effectively mitigating the negative impact of inter-domain discrepancies and class imbalances on the model performance remains a significant challenge. Existing OSDA methods tend to overlook domain-specific features and directly minimize domain discrepancies, which may lead to ambiguous decision boundaries between classes and weaken the generalization ability of the model. To address this problem, this paper proposes an open-set domain adaptation method based on Optimal Transport Distance Regularization (OTR) and neighborhood clustering, called OTRNC. Specifically, the method employs OTR to maximize the distributional distance between high-confidence and low-confidence sample groups, thereby reducing the interference of unknown categories in the domain adaptation process. Subsequently, it leverages dynamic neighbor retrieval and invariant feature learning to reduce intra-class variations in the target domain and enhance the generalization capability of the learned features. Experimental results demonstrate that OTRNC achieves superior performance across multiple benchmark datasets, thereby validating its effectiveness.

Deep learning-based object detection has significantly improved the detection of medium and large targets. However, when detecting small objects, traditional algorithms often face challenges such as missed detections and false positives owing to the inherent issues of small scale and complex backgrounds. Therefore, this study aims to enhance the accuracy of small object detection by improving the YOLOv8 model. First, the convolutional module in the backbone is replaced with the RFAConv module, which enhances the ability of the model to process complex images. Second, a Mixed Local Channel Attention (MLCA) mechanism is introduced in the neck part, allowing the model to fuse features from different layers more efficiently while maintaining computational efficiency. Third, the Detect head of YOLOv8 is replaced with the Detect_FASFF head to address the inconsistency between different feature scales and improve the ability of the model to detect small objects. Finally, the Complete Intersection over Union (CIoU) loss function is replaced with the Focaler-IoU loss function, enabling the model to focus more on small objects that are difficult to locate precisely. Experimental results show that the improved model increases mAP@0.5 by 4.8 percentage points and mAP@0.5:0.95 by 3.0 percentage points on the FloW-Img dataset, which is sparse in small objects. On the VisDrone2019 dataset which has a high density of small objects, mAP@0.5 increases by 5.9 percentage points and mAP@0.5:0.95 improves by 4.0 percentage points. In addition, generalization comparison experiments are conducted on the low-altitude dataset AU-AIR and the pedestrian-dense detection dataset WiderPerson. The optimized model significantly improves the accuracy of small object detection compared with the original model and expands its applicability.

Chinese medical Named Entity Recognition (NER) aims to identify entities with specific meanings from medical texts, such as diseases, drugs, symptoms, and anatomical parts. This task provides robust support for clinical decision making, medical information integration, and medical record management. However, existing research on Chinese medical NER has not fully addressed the complexity of medical texts, which are characterized by an abundance of specialized terminology, limited embedding diversity, and insufficient utilization of semantic information. To address these issues, this paper proposes a Chinese medical NER model that integrates multigranularity features. The model first employs a BERT pretrained model to generate character embeddings for the text. It then uses both one-dimensional and two-dimensional convolutional neural networks to extract character shape and stroke features, while external lexicons are incorporated to introduce word-level features, enhancing the representation of word and entity boundaries. Additionally, sentence-level features are included to capture global semantic information. A cross-attention mechanism is utilized to iteratively fuse these multigranular features, resulting in embeddings enriched with deep semantic information. Finally, Conditional Random Fields (CRF) are used to output the entity recognition results. Experimental results on the CCKS2017 and CCKS2019 datasets demonstrate that the proposed model achieves F1 values of 92.88% and 87.86%, respectively, outperforming mainstream models in recognition performance.

Cross-domain recommendation systems are widely applied in e-commerce and content platforms. Although the dual-target cross-domain recommendation (DTCDR) proposed in recent years has achieved a breakthrough in simultaneously improving the performance of both domains, it still faces two major challenges: 1) the generated user-item representations lack sufficient correlation and diversity; 2) the semantic noise mixed in the shared preferences leads to negative transfer problems. To address these issues, a dual-target cross-domain recommendation model based on heterogeneous graph and hierarchical preference disentanglement (HGPD-DTCDR) is proposed. Its core innovations include: 1) a heterogeneous graph collaborative learning framework is proposed to integrate user-item interactions, user social networks, and item attribute similarities, constructing a multi-relation heterogeneous graph, and generating high-order semantic representations through a relation graph convolutional network (R-GCN) to enhance the diversity and correlation of the representations; 2) a two-stage decoupling process is designed, first separating domain-specific and shared preferences through a variational graph encoder, and then introducing a semantic filtering network to optimize the quality of shared preferences. Experiments on five real cross-domain datasets show that the performance improvement of this model stems from the synergistic effect of heterogeneous graph modeling and hierarchical decoupling mechanisms. Compared with the best baseline, it achieves average improvements of 3.55%, 7.27%, and 15.57% in hit rate, normalized discounted cumulative gain, and mean reciprocal rank, respectively. In data-sparse scenarios, the performance improvement is even more significant, with an average gain of 10.35%. Ablation studies further verify the effectiveness of each technical component and their synergistic effects.

With the continuous evolution of network environments and the constant emergence of Internet applications, the applicability of machine learning classifiers trained on previous traffic data to new sample spaces gradually weakens. This leads to a decline in the recognition capability of classification models; thus, these models are ineffective in meeting the ever-increasing demands of network services and security. If the classifier is manually updated based on experience, it consumes a significant amount of effort, and it is difficult to ensure the generalization performance of the new classifier. Moreover, the continuous influx of new data poses severe challenges in balancing model training accuracy with the storage of computational resources. To address this issue, an incremental learning strategy employing spatial optimization techniques is proposed to achieve efficient network traffic classification. First, by optimizing the spatial distribution of old and new traffic samples, clusters of different categories (both old and new) are maintained with minimum intervals to avoid distribution conflicts arising from old and new tasks shared feature space. Within the optimized feature space, a small number of old data samples are replayed, and knowledge distillation techniques were utilized to maintain the stability of the original model parameters. Only the extended parts of the model are adjusted to update the classifier at minimal cost. Experimental results on the USTC-TFC2016 dataset demonstrate that, compared with other methods, this approach exhibits higher stability and effectiveness in terms of model accuracy, resource consumption, and performance.

Convolutional Neural Networks (CNNs) are widely used in the field of object detection, earning widespread acclaim in scholarly circles due to their precision and scalability. It has spawned numerous notable models, including those in the Region-based Convolutional Neural Networks (R-CNNs) (such as Fast R-CNN and Faster R-CNN) and You Only Look Once (YOLO) series. After the success of Transformers in the field of natural language processing, researchers began exploring their application in computer vision, leading to the development of visual backbone networks such as Visual Transformer (ViT) and Swin Transformer. In 2020, a Facebook research team unveiled DEtection TRansformer (DETR), an end-to-end object detection algorithm based on Transformers, designed to minimize the need for prior knowledge and postprocessing in object detection tasks. Despite the promise shown by DETR in object detection, it has limitations including low convergence speed, relatively low accuracy, and the ambiguous physical significance of target queries. These issues have spurred a wave of research aimed at refining and enhancing the algorithm. This paper aims to collate, scrutinize, and synthesize the various efforts aimed at improving DETR, assessing their respective merits and demerits. Furthermore, it presents a comprehensive overview of state-of-the-art research and specialized application domains that employ DETR and concludes with a prospective analysis of the future role of DETR in the field of computer vision.

World models are generally believed to understand and represent the external world and predict future states based on current world states and actions. Large models leverage massive training data and vast parameter scales to exhibit outstanding capabilities in learning, understanding, representing, and generating textual knowledge, as exemplified by language large models such as GPT-4 and LLaMA. In recent years, research on world models has attracted significant attention from both industry and academia, leading to significant research and commercial achievements in domains such as autonomous driving, social simulation, embodied intelligence, and video generation. Moreover, researchers have applied the remarkable results of various large models to world models, further enhancing their performance. This paper comprehensively reviews world models built using large models across different domains, covering both language large model- and Vision Large Model (VLM)-based approaches. Several important application areas, including embodied intelligence, smart cities, social simulation, and physical environment simulation, are selected to introduce relevant models. This paper classifies world models based on the modality of the large models used, highlighting the functional differences between world models based on different modalities. Subsequently, important open-source resources and benchmarks for world models are presented to help researchers in related fields understand and utilize world models quickly. Finally, this paper is summarized and future research directions are presented.

Timely and effective disinformation detection is crucial for curbing the spread of disinformation and minimizing social harm. Numerous deep learning methods have been employed for disinformation detection. Summarizing the detection principles and paradigms of existing research is essential for identifying directions for technical optimization. Therefore, this paper comprehensively reviews existing research based on the principles and implementation paths of disinformation detection, and for the first time, summarizes and compares the applications of large language models in this field. First, the relevant concepts of disinformation detection tasks are introduced and the data structures of commonly used disinformation detection datasets are summarized. Then, based on detection principles and implementation methods, the paper presents ways to detect textual and multimodal disinformation through semantic feature representation, auxiliary task design, internal knowledge inference, and fact verification, refining them into ten subcategories and summarizing the potential characteristics of detection methods for each subcategory. Finally, the paper summarizes disinformation detection paradigms based on deep neural networks and large language models, compares the detection performance of representative methods from these paradigms across seven disinformation detection datasets, and highlights the advantages and limitations of large language models in detecting disinformation. It also presents the anticipated opportunities and challenges brought about by large language models in the field of disinformation detection, providing a reference for future research.

In Printed Circuit Board Assembly (PCBA), defect detection is key to improving production line efficiency. However, after assembly, printed circuit boards are usually inspected manually, leading to labor and time wastage, as well as missed and false detections. To address these issues, this paper proposes an improved lightweight YOLOv8s network that effectively reduces model complexity while enhancing the accuracy of PCBA defect detection. First, owing to the lack of publicly available PCBA-related datasets, a dataset called PCBA-DET is constructed for post-assembly PCBA defect detection. Various data augmentation techniques are applied to PCBA-DET to simulate real-world production scenarios and improve the dataset balance. Second, the last C2f module of the YOLOv8s backbone is replaced with a Re-parameterized Large Kernel convolution Network (RepLKNet) to reduce computational cost and increase the effective receptive field of the model. In addition, in the neck network of YOLOv8s, a P2 small object detection layer and Ghost Convolution are introduced to capture more detailed information and effectively reduce the number of model parameters. On the augmented PCBA-DET dataset, the improved model achieves an increase of 2.6 and 0.1 percentage points in terms of mean Average Precision (mAP)@0.5∶0.95 and mAP@0.5, respectively, compared with the baseline model, whereas the number of parameters is reduced by 36.8%.

In recent years, Large Language Models (LLMs) such as GPT, LLaMA, Qwen, and DeepSeek, have achieved significant breakthroughs in natural language processing, computer vision, multimodal learning, and other fields. However, constrained by factors such as their reasoning mechanisms, parameter scales, and the inherent knowledge contained within their training data, these models often suffer from issues like ″hallucinations″—characterized by inaccurate answers and even factual deviations—when handling complex tasks, addressing questions from professional domains, or generating time-sensitive content. These limitations severely hinder their application in high-reliability scenarios. The ″tool learning″ paradigm is attracting increasing attention as a promising solution to these capability bottlenecks. Its primary objective is to enable LLMs to understand and utilize external tools to complete specific tasks. By invoking external tools, such as databases, search engines, and mathematical tools, LLMs can transcend their parameterized knowledge; enhance their reasoning, decision-making, and execution capabilities; and mitigate hallucination problems. This paper systematically reviews the development context and technical advancements in LLM tool learning, analyzes the expansion of LLM capabilities through tools, summarizes tool invocation mechanisms ranging from in-context learning to fine-tuning training, and discusses key issues including performance optimization and adaptive tool generation. The paper also analyzes evaluation methods for LLM tool invocation, summarizes the current challenges in tool learning, and outlines future research directions.

Object detection and multi-target tracking technologies are becoming increasingly mature. However, when performing aerial multi-target tracking tasks in complex scenarios, issues such as small target size, large size variation, and occlusion lead to unsatisfactory detection and tracking performances. Therefore, this paper proposes an aerial multi-target tracking algorithm based on an improved YOLOv8 and ByteTrack method called YBTrack. First, a detector called Multi-Scale Attention YOLO (MSA-YOLO) is constructed. The original convolution in YOLOv8 is replaced with a space-depth convolution that transforms spatial information into channel dimensions, effectively preserves target details, and reduces missed and false detections caused by information loss during multiscale feature map fusion. Simultaneously, a lightweight accelerated space-channel attention module is designed for neck convolution to reduce computational complexity. This module also acts as a feature refinement module before the detection head, further enhancing its ability to extract target feature information. Next, the ByteTrack tracking model is optimized to improve the tracking performance. A Spatial-Appearance Similarity Matrix (ASM) is designed to enhance the model's ability to distinguish similar targets. In addition, a target correction function is proposed to reduce the error accumulation of the Kalman filter and decrease the target offset and loss rates. Finally, MSA-YOLO and the optimized ByteTrack are combined for multi-target tracking experiments. MSA-YOLO achieves a 9.4 percentage points improvement in mAP@0.5 on the VisDrone2019-DET dataset. The multi-target tracking algorithm improves the Multiple Object Tracking Accuracy (MOTA) by 11.2 and 8.3 percentage points and the Identification F1 Score (IDF1) by 8.9 and 7.4 percentage points on the VisDrone2019-MOT and MOT17 datasets, respectively. Experimental results demonstrate the excellent tracking performance of the proposed method. Furthermore, comparison experiments with other multitarget tracking algorithms confirm the superiority of the proposed algorithm.

In multi-agent game simulations, the performance of Large Language Model (LLM) has been widely studied; however, their decision-making ability to guide multi-agent cooperation in fuzzy task objectives or uncertain environments is often unreliable. To address this issue, a multi-level collaborative decision-making framework based on distributed Bayesian inferences is proposed. This framework integrates three major functional modules: decision making, peer evaluation, and supervision. It utilizes multiple LLM for collaborative decision making and has been experimentally validated in a spatial prisoner's dilemma game. The experimental results show that the framework effectively overcomes the decision-making bottleneck of LLM in fuzzy task environments and successfully promotes the emergence of multi-agent cooperative behavior. Additionally, a quantitative evaluation of the model's decision-making ability in different experimental scenarios reveals that the decision error of the model is not linearly related to the model size. Under fuzzy task instructions, the decision error of the LLaMA3 (70×10⁹) model is 16.6% higher than that of the LLaMA3 (8×10⁹) model and 7.2% higher than that of the LLaMA2 (7×10⁹) model. This indicates that in more complex environments, relying solely on the expansion of the model size does not significantly improve the decision-making performance. By contrast, LLM collaborative decision making has shown significant advantages in improving decision consistency and effectiveness. These results reveal the crucial role of multi-model collaboration in complex decision-making environments and provide important references for the future design of intelligent agent systems for uncertain tasks.

Deep learning has driven the development of artificial intelligence, which is widely used in computer vision. It provides breakthroughs and remarkable results in complex tasks such as image recognition, object detection, object tracking, and face recognition, demonstrating its excellent recognition and prediction capabilities. However, vulnerabilities and loopholes in deep learning models have been gradually exposed. Deep learning techniques, represented by convolutional neural networks, are extremely sensitive to well-designed adversarial examples, which can easily affect the security and privacy of the models. This paper first summarizes the concept of adversarial attacks, reasons for generating adversarial examples, and related terms. It outlines several types of classical adversarial attack strategies in the digital and physical domains and analyzes their advantages and disadvantages. Second, it focuses on computer vision and summarizes the latest research in adversarial attacks during tasks such as object detection, face recognition, object tracking, monocular depth estimation, and optical flow estimation, from both the digital and physical domains, as well as the various datasets commonly used in the study. It also briefly introduces the current stage of adversarial example defense and detection methods, summarizes the advantages and disadvantages of these methods, and describes examples of the applications of adversarial sample defense for various visual tasks. Finally, based on the summary of adversarial attack methods, it explores and analyzes the deficiencies and challenges of existing computer vision adversarial attacks.

Artificial intelligence has made remarkable progress across many fields, encouraging countries to attach great importance to its research and development. However, the rapid development of artificial intelligence has also brought about a series of problems and threats, and overreliance on and blind trust in such models can lead to serious risks. Therefore, interpretable artificial intelligence has become a key element in building trusted and transparent intelligent systems, and its research and development requires immediate attention. This survey comprehensively summarizes the research progress on explainable artificial intelligence at home and abroad comprehensively from multiple dimensions and levels. Based on current research results in the industry, this survey subdivides the key technologies of explainable artificial intelligence into four categories: interpretation model, interpretation method, safety testing, and experimental verification, with the aim of clarifying the technical focus and development direction of each field. Furthermore, the survey explores specific application examples of explainable artificial intelligence across key industry sectors, including but not limited to education, healthcare, finance, autonomous driving, and justice, demonstrating the significant role it plays in enhancing decision-making transparency. Finally, this survey provides an in-depth analysis of the major technical challenges of interpretable artificial intelligence and presents future development trends, in addition to a special investigation and in-depth analysis of the interpretability of large models, which has attracted considerable attention recently.

The traditional path-planning algorithm of a mobile robot usually requires a map to effectively plan the path. By contrast, path planning based on Deep Reinforcement Learning (DRL) does not require maps for navigation, owing to which it has received considerable attention. However, traditional DRL path-planning algorithms often face challenges such as low sample utilization, slow training speed, and insufficient generalization ability. To solve these issues, the Twin Delayed Deep Deterministic (TD3) policy gradient algorithm is improved to enhance its path-planning performance for mobile robots. First, to solve the problem of the TD3 algorithm having limited ability for continuous space exploration, the exploration strategy is improved and the continuous space exploration ability of the algorithm is enhanced using pink noise with time correlation. Second, the n-step method is combined with the Loss-Adjusted Approximate Actor Prioritized (LA3P) experience replay method. The n-step method can capture the long-term reward signal more accurately by expanding the immediate reward in the experience replay buffer to the n-step cumulative discount reward, whereas the LA3P method can improve the sample utilization and performance of the algorithm by efficiently using the n-step experience. Finally, three different environments are built in Gazebo for the experiments and compared with various algorithms. The experimental results show that the improved algorithm has more advantages in terms of training time, average success rate, and average distance, which proves the effectiveness of the improved algorithm.

In view of missed and false detection phenomena caused by numerous small target instances and occlusions among targets in drone images, this paper proposes a lightweight small target detection algorithm for Unmanned Aerial Vehicle (UAV) images based on an improved YOLOv8. The Triple Feature Encoder (TFE) and Scale Sequence Feature Fusion (SSFF) modules are introduced in the neck to enhance the ability of the network to extract features at different scales. Furthermore, a Small Object Detection Head (SMOH) is designed and fused with the improved neck feature extraction network, and an additional detection head is also introduced to reduce the loss of small target features and enhance the recognition ability of the network for small targets. Additionally, considering the defects of Complete Intersection over Union (CIoU), a regression loss function, Wise-Inner-MPDIoU, is proposed by combining Wise-IoU, Inner-IoU, and Minimum Point Distance based IoU (MPDIoU). Finally, to realize the lightweight application requirements of the algorithm in mobile and embedded systems, amplitude-based layer-adaptive sparse pruning is performed to further reduce the model size while ensuring model accuracy. Experimental results demonstrate that, compared to the original YOLOv8s model, the improved model proposed in this paper improves mAP@0.5 by 6.8 percentage points, while reducing the number of parameters, amount of computation, and model size by 76.4%, 17.1%, and 73.5%, respectively. The proposed model is lightweight, improves detection accuracy, and has strong practical significance.

Biomedical imaging is crucial in the diagnosis and treatment of various diseases. The application of deep learning methods to medical image analysis can enhance the readability of medical images and provide more reliable support for clinical decision-making. However, traditional medical image processing methods are limited in effectively capturing spatial features and complex structural information in 3D images, especially when handling complex 3D medical images generated by different imaging modalities. This often challenges the accuracy and generalization ability of the model. To address this challenge, an MTM3D model is proposed for medical image classification tasks. This model combines the excellent performance of the Mamba model in complex sequential tasks with the external memory storage function of the improved Token Turning Machine (TTM) network. By introducing a cyclic chain storage structure, MTM3D enables effective interaction of features from different spatial structures within memory units, thus enhancing its ability to capture complex spatial relationships. Furthermore, the incorporation of Mamba further strengthens the interaction between the memory and processing units, strengthening the generalization capability of the model and enhancing its performance across different medical imaging datasets. Experimental results on the MedMNIST v2 dataset demonstrate that MTM3D exhibits outstanding capabilities in understanding medical images. Compared with the current best medical image analysis networks, MTM3D improves the average Accuracy (ACC) by 3.97% and the average Area Under the Curve (AUC) by 2.00%, fully showcasing its tremendous potential in interpreting medical images and assisting healthcare professionals in diagnosis and treatment planning.

Single-modal images are easily affected by light, weather, and other environmental conditions in all-weather ship detection. This leads to a low ship detection accuracy and high leakage rate. To address these issues, this paper proposes a ship detection method, VIF-RTDETR, which fuses visible light and infrared image information. The method fully utilizes the rich details and color information of visible images and the stable performance of infrared images in low-light environments, and constructs a four-channel input model. The complementary fusion of varied modal information is realized by designing the fusion module VIF such that it makes more reasonable use of the information from the two modalities (visible light and infrared) in the detection network. The channel attention in the backbone feature extraction network is combined to further optimize the feature extraction capability by dynamically assigning different weights to the channels, thereby enhancing the feature expression capability of the channels. To further enhance the detection performance of small targets in ship detection, a weighted bounding box loss function is designed so that the model can effectively focus on the feature expression of targets of different sizes and improve the detection accuracy under different target sizes. The experimental results show that in the visible and infrared datasets for the ships, the detection precision AP_0.5∶0.95, AP_0.5 of the model reaches 78.3% and 98.5%, respectively, reflecting improvements by 4.7 and 9.2 percentage points relative to AP_0.5∶0.95 the single-modal visible and infrared models. Further, the recall rate AR_0.5∶0.95 reaches 85.2%, reflecting improvements by 3.1 and 7.3 percentage points relative to the single-modal visible and infrared models, respectively. Thus, the findings contribute to significantly improving the precision of ship detection and reducing the leakage rate.

To address the issue of unmanned vehicles being unable to efficiently reach target points in muddy and rugged terrains such as tidal flats, an improved algorithm based on the A* algorithm, Tidal-A* (TA*), is proposed to plan optimal paths for unmanned vehicles. Considering the characteristics of tidal flat environments, the quality of the generated paths is jointly evaluated using Soil Moisture Content (SMC) along the path, path height fluctuation, and path length. To address the difficulty of directly obtaining environmental information, a drone equipped with a hyperspectral sensor and LiDAR is used to scan the target area. A dimensionality reduction method combining spectral preprocessing and the Pearson correlation coefficient is proposed to train the SMC inversion model. In response to the limitations of the traditional A* algorithm, which only searches for paths based on path length, a cost function that integrates multiple constraints is proposed based on the design of the cost functions for three individual constraints. To address the issue that the traditional A* algorithm cannot change the path according to requirements, a coefficient combination is designed to control the proportion of each constraint in the cost function while solving the problem of inconsistent orders of magnitude between different constraints. To address the potential issue that the traditional A* algorithm may overlook better solutions, the calculation range of the heuristic function is improved, allowing the algorithm to trade off-path redundancy for the optimization of other constraints. The simulation results show that when using this algorithm to train the model, the determination coefficient R² is 0.784, and the Ratio of Standard Deviation (RPD) is 2.151, which are 38% and 33.8% higher, respectively, than those of the direct inversion methods. Compared to those generated by traditional algorithms, the length, SMC value, and height fluctuation of the paths generated by the TA* algorithm are reduced by 3.4%, 5.1%, and 18.7%, respectively.

This study proposes a Common and Differential Cross-Attention Module-Bird's-Eye View (CDCAM-BEV) algorithm that combines 4D millimeter-wave radar and vision fusion to improve target detection accuracy for pedestrian and vehicle target recognition and localization in autonomous driving scenarios. First, a radar cylinder network is designed to encode the 4D radar point cloud into a pseudo image and convert the monocular image into a Bird's-Eye View (BEV) feature through Orthogonal Feature Transformation (OFT). Second, based on the cross-attention mechanism, a Common Information Extraction Module (CICAM) and a Differential Information Extraction Module (DICAM) are used to fully explore the common and differential information between radar and images. Finally, a BEV feature fusion module is designed based on CICAM and DICAM to achieve feature-level fusion of image and radar information in the BEV space. Experiments are conducted on the VOD dataset, and the CDCAM-BEV algorithm is compared with five other 3D object detection algorithms. The experimental results show that CDCAM-BEV achieves better detection performance in multiple modes. In the 3D mode, the average detection accuracy of CDCAM-BEV is 3.65 percentage points higher than that of the second ranked Part-A²; in the BEV mode, it is 5.04 percentage points higher than that of the second ranked PointPillars; in the Average Directional Similarity (AOS) mode, it is 2.62 percentage points higher than that of the second ranked Part-A². These results show that CDCAM-BEV exhibits excellent performance in all modes, effectively fusing images and 4D radar point cloud features, which significantly improves the accuracy and reliability of object detection.

The multi-view subspace clustering algorithm, a type of multi-view clustering algorithm, emphasizes discovering potential subspaces in multi-view data and clustering based on these subspaces. The Multi-view Subspace Clustering algorithm with Grouping Effects (MvSCGE) is also a type of multi-view subspace clustering algorithm. The basic principle is to learn the subspace representation of each view through smooth regularization, while ensuring cross view consistency, and ultimately learning a consistent clustering index matrix. The clustering results are obtained after processing. However, this algorithm only considers the local structure of a single view and has certain limitations. To further explore the diversity between views, this paper proposes a Diversity-induced Multi-view Subspace Clustering algorithm with Grouping Effect (DiMvSCGE). It preserves the local structure of each view while using the Hilbert-Schmidt Independence Criterion (HSIC) to measure the diversity between views, and iteratively uses alternating direction minimization on this basis. Based on the clustering index matrix obtained after the iterations, k-means clustering is performed to obtain the final result. Comparative experiments with several advanced algorithms on four public datasets show that this algorithm offers advantages, such as low parameter sensitivity and fast convergence speed, and demonstrates good performance on different datasets.

Traditional Retrieval-Augmented Generation (RAG) methods predominantly focus on pure-text scenarios. In these scenarios, their retrieval and generation mechanisms encounter difficulties in effectively modeling common visual elements, spatial layouts, and structural semantics within multimodal documents. This drawback restricts their performance in tasks related to text-image hybridization, long documents, and cross-document reasoning. To tackle this issue, Multimodal Retrieval Augmented Generation (MRAG), by integrating text, image, and layout structure modeling, and incorporating multimodal evidence retrieval and scheduling during the generation process, has already developed into a core technical paradigm for Question & Answer (Q & A) and reasoning in visually-rich documents. This paper conducts a systematic review of research progress in MRAG applications for document Q & A tasks. Firstly, based on the practical requirements for multimodal document understanding, we analyze the key challenges in MRAG implementation, including multimodal alignment, long-context modeling, evidence traceability, and system robustness. Secondly, from the perspective of how MRAG systems support the generation process, we compare representative methods across four dimensions: embedding paradigms, document retrieval scope, layout-aware mechanisms, and multimodal retrieval strategies. We focus on how design choices influence generation stability, reasoning accuracy, and system complexity. Thirdly, we summarize the characteristics and limitations of existing multimodal document Q & A datasets and evaluation frameworks, and analyze the current constraints in evidence granularity and reasoning explainability. Finally, we point out that MRAG is evolving from static similarity-matching retrieval mechanisms to dynamic evidence planning paradigms centered on generation and reasoning needs, and should continuously enhance the reliability and explainability of complex document Q & A systems through collaborative multimodal modeling with multi-granularity approaches.

To address the privacy leakage risks associated with collecting and annotating sensor data for wearable device-based fitness action recognition as well as the limitations of traditional centralized model training in adapting to new users, this study proposes a fitness action recognition method based on a wearable sports knee sleeve and personalized Federated Learning (FL). The method accomplishes two tasks, namely fitness action type recognition and user performance level recognition, to help users understand their fitness performance and improve exercise effectiveness. First, each user is treated as an independent task, and a global embedding network is meta-trained in a federated manner to learn user-agnostic shared representations that can be effectively generalized for any user. Subsequently, through an adaptation procedure, the local classification network is fine-tuned in two stages on top of the global embedding network, enabling each user to obtain a personalized model. Finally, extensive experiments are conducted on a real-world fitness dataset to validate the performance of the proposed method. The experimental results demonstrate that the proposed method achieves 100% accuracy in fitness action type recognition and 95.94% accuracy in user performance level recognition, significantly outperforming existing state-of-the-art methods. The proposed system can protect user privacy while exhibiting a strong generalization ability for new users.

With the development of social networks, people are increasingly expressing their emotions through multimodal data, such as audio, text, and video. Traditional sentiment analysis methods struggle to process emotional expressions in short videos effectively, and existing multimodal sentiment analysis techniques face issues such as low accuracy and insufficient interaction between modes. To address these problems, this study proposes a Multimodal Sentiment Analysis method based on Dense Co-Attention (DCA-MSA). First, it utilizes the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model, OpenFace 2.0 model, and COVAREP tool to extract features from text, video, and audio, respectively. It then employs a Bidirectional Long Short-Term Memory (BiLSTM) network to model the temporal correlations within different features separately. Finally, it integrates different features through a dense co-attention mechanism. The experimental results show that the model proposed in this paper is competitive in multimodal sentiment analysis tasks compared to some baseline models: on the CMU-MOSEI dataset, the highest increase in binary classification accuracy is 3.7 percentage points, and the highest increase in F1 value is 3.1 percentage points; on the CH-SIMS dataset, the highest increase in binary classification accuracy is 4.1 percentage points, the highest increase in three-classification accuracy is 2.8 percentage points, and the highest increase in F1 value is 3.9 percentage points.

Knowledge Graphs (KGs) are structured semantic knowledge representations with entities as nodes and relations as edges. They accurately depict various things in the real world and their complex associations and have become a core supporting technology across multiple domains, including artificial intelligence, natural language processing, recommendation systems, and intelligent question-answering. Hence, they form an important foundation for machines to understand semantics and achieve cognitive intelligence. This paper describes the basic concepts and system architecture of a KG, clarifies the knowledge representation unit with the ″entity—relation—attribute″ triple as the core, and analyzes the applicable scenarios and technical characteristics of both top-down and bottom-up construction approaches. The technical evolution of three core links in the KG construction process, namely, information extraction, knowledge fusion, and knowledge reasoning, are analyzed. Subsequently, the technical development context is systematically discussed, and the advantages and limitations of different methods are discussed. Further, through an in-depth analysis of the differences in technical route selection between DBpedia and Baidu as two typical KGs, the theoretical method is combined with an actual KG construction scenario. Finally, the challenges faced by current KG construction in terms of data quality, semantic consistency, and dynamic evolution are summarized, and future research directions are examined. This review provides comprehensive guidance for both theoretical research and practical applications in KG construction, thereby advancing technological development in this field.

As core data for maritime traffic, ship trajectory data can be used for trajectory prediction, early warning, and other tasks with pronounced temporal characteristics. However, owing to factors such as harsh marine environments and poor communication reliability, missing ship trajectory data is a common problem. Learning from time series containing missing data can affect the accuracy of time series analysis significantly. The current mainstream solution is to approximate the imputation of missing data, mainly based on convolutional models, to reshape the time series along a timeline to capture its local features of the time series. However, the ability to capture the global features of long time series is limited. The Transformer enhances the ability of a model to capture the global features of a time series by capturing the relationships between various time points in the time series through its core self-attention mechanism. However, because its attention is calculated through matrix multiplication, it ignores the temporal nature of the time series, and the obtained global feature weights do not have a time span dependency. Therefore, to address the issue of capturing global features in long time series, this study proposes the GANet, a variant network based on the self-attention mechanism. GANet first obtains the basic global feature weight matrix from the time series points through the self-attention mechanism and then uses gated recurrent units to forget and update this global feature weight matrix on the timeline, thereby obtaining a global feature weight matrix with time span dependency, which is then used for data reconstruction to impute the missing data. GANet combines the self-attention mechanism and gating mechanism to capture global features while considering the impact of the time span on different time points, making the captured global feature time span dependent. Experimental results show that compared with existing models, such as Autoformer and Informer, GANet achieves better interpolation performance on Trajectory, ETT, and Electricity datasets.

To address the issue of interference factors such as seasonal changes, climate, and illumination that affect high-resolution remote sensing images of the same geographical space but different temporal phases, a remote sensing image building Change Detection (CD) method based on a multi-temporal ChangeFormer is proposed. This method uses multiple remote sensing images from different temporal phases and fuses the latest temporal remote sensing image with multiple prechange remote sensing images at different scales for feature difference extraction. Additionally, it focuses on both the comprehensive semantic features of the images and the details of the semantic information between the images. This approach helps reduce false detections caused by changes in factors such as season and illumination. Additionally, the method fuses the feature differences of multiple prechange remote sensing images from different temporal phases and introduces a regularization term as a loss function. This eliminates interference from nonbuilding changes and illumination shadows in the nonchanging areas of buildings, thereby enhancing the generalization ability of the model. A three-temporal remote sensing image dataset covering changes from agricultural land to construction land is constructed. The experimental results show that, compared to the current optimal BIT method, the multi-temporal ChangeFormer method improves the F1 value, Intersection over Union (IoU), precision, and recall by 9.04%, 9.87%, 15.27%, and 3.4%, respectively, thus significantly enhancing detection accuracy. Furthermore, it outperforms classical CD methods in terms of detailed information processing.

Multisensor data fusion methods can enhance the accuracy of bearing fault diagnoses. However, in the context of motor-bearing fault diagnosis, most existing multisensor data fusion methods suffer from issues such as a single type of sensor data and difficulty in fully exploiting the redundancy and complementarity among different modal data. To address this issue, a bearing fault diagnosis method based on a Multiple-Constraint Modal-Invariant Graph Convolutional Fusion Network (MCMI-GCFN) is proposed. First, features are extracted from the original current and vibration signals using a Convolutional Auto-Encoder (CAE) and Squeeze-and-Excitation block (SE block). Second, a source domain classifier and domain discriminator are introduced to capture the modal invariance among different modal data based on domain adversarial training, thereby fully exploiting the redundancy and complementarity among multimodal data. Finally, the spatial aggregation characteristics of the Graph Convolutional Neural Network (GCN) are utilized to capture the dependency relationships between features at similar time steps in the current and vibration modalities, enabling precise fusion of their contextual semantic information. This method is validated using a publicly available bearing damage current and vibration dataset obtained from the University of Paderborn, Germany. The experimental results demonstrate that the MCMI-GCFN method achieves a bearing fault diagnosis accuracy of 99.6%, which is 9 to 11.4 percentage points higher than that achieved by nonfusion methods, thereby verifying the effectiveness of the proposed model.

The Transformer model, owing to its excellent global information capture and powerful representation abilities, has been widely used in medical image segmentation and has achieved remarkable results. However, when serializing images, these methods divide them into fixed-sized blocks and extract only global features of a single scale, which, to some extent, fragments the semantic features of the images, ultimately leading to poor segmentation accuracy. To address this issue, this paper proposes a Multi-scale self-attention Transformer architecture (MultiFormer). First, continuous convolution and downsampling modules are used to process images. Then, the original 1×1 projection module is replaced with a multi-scale convolution projection module. Finally, deformable convolution is introduced into the feature maps generated by the self-attention module. Compared to the traditional Transformer image serialization process, this continuous convolution effectively enlarges the receptive field while generating features of the same resolution, retains the spatial correlation of 2-Dimensional (2D) images, and avoids the loss of semantic information caused by fixed-position and fixed-size image blocks. Meanwhile, the multi-scale convolution projection module captures contextual information in the images using four convolution kernels of different sizes and achieves multi-scale feature fusion through channel concatenation, reflecting attention between local regions at different scales, rather than being limited to a single scale, which enables the model to aggregate semantic information at different scales and further alleviates the problem of semantic fragmentation. Deformable convolution introduces an additional convolution layer to learn and generate an offset field, allowing the convolution kernel to flexibly adjust its shape to adapt to morphologically diverse lesions or organs in the images and enhancing the ability of the model to process complex medical images. This module is inserted into three network structures—SETR, TransUNet, and TransFuse—and experiments are conducted on the ACDC heart and ISIC 2018 skin lesion datasets. The results show that the Dice coefficient improves by 3.63, 1.06, and 2.30 percentage points and by 1.22, 2.31, and 3.01 percentage points, respectively. MultiFormer is a plug-and-play tool and can be easily integrated into various downstream medical image analysis tasks.

This paper proposes an improved YOLOv8-based model named CAFR-YOLO to address the issues of insufficient cross-level feature interaction and limited feature representation capability in multi-scale object detection under complex scenes. First, a novel cross-scale feature reorganization pipeline was designed, constructing the Channel Attention-guided Feature Reorganization (CAFR) module. By using a specific layer as the fusion backbone and incorporating scale alignment, attention-weighted fusion, and feature subset splicing strategies, it alleviates insufficient cross-level interaction in traditional feature pyramid structures. Secondly, at the local level, the method introduces the C2f_DCNv3 module into the backbone network, significantly enhancing the model's geometric adaptability by exploiting the dynamic sampling characteristics of deformable convolution. From a global perspective, the C2f_SAConv module is constructed by combining Switchable Atrous Convolution (SAC) with the C2f module, optimizing multi-scale semantic feature fusion through dynamic atrous rate adjustment. These two approaches enhance the model's robustness to complex scenes. Finally, SPDConv replaces traditional convolution structures, strengthening feature representation through spatial-channel reorganization while reducing computational complexity. Experimental results demonstrate that CAFR-YOLO achieves 86.3% mAP@0.5 and 67.2% mAP@0.5:0.95 on the PASCAL VOC dataset with comparable computational costs to the original model. On the MS COCO dataset, it improves mAP@0.5 and mAP@0.5:0.95 by 3.5% and 3.9%, respectively. Compared to existing state-of-the-art methods, CAFR-YOLO exhibits significant advantages across multiple metrics. The proposed CAFR-YOLO model substantially enhances multi-scale object detection accuracy and robustness while maintaining computational efficiency, providing a novel solution for real-time object detection tasks.

With the popularization of the Internet and the diversification of applications, fine-grained classification of massive network traffic has become key to optimizing quality of service and analyzing user behavior patterns. This paper presents an overview of Machine Learning (ML)-based and pretrained model-based network traffic analysis methods to promote further research and development in this field through multidimensional comparison and analysis. First, the complete traffic classification pipeline is deconstructed, covering data acquisition, preprocessing, and feature extraction, and the practical value of data balancing techniques is examined. The data format, scale, and scene suitability of mainstream public datasets are introduced, compared, and analyzed from multiple perspectives, highlighting their data distribution, feature redundancy, and timeliness problems. Second, it summarizes the limitations of traditional algorithms in handling high-dimensional data and meeting real-time requirements, and outlines the trend of applying pretrained models in traffic analytics, through a focused comparative analysis of experimental results. This review includes breakthroughs in Transformer-based pretrained models, their fusion with Deep Learning (DL) models, and advances in lightweight pretrained models for traffic classification. Finally, by considering dynamic research trends, the opportunities and challenges in future applications of pretrained models are discussed, and their limitations in terms of computational cost and privacy protection are analyzed.

In recommender systems, sequential recommendations aim to predict the future interests of a user based on historical interaction sequences. However, existing deep learning models typically focus on capturing users' long-term behavior patterns while neglecting the fine-grained modeling of temporal information, which limits improvements in recommendation performance. To address this issue, this study proposes a method that combines the global and local self-attention mechanisms. This method processes user interaction sequences in chunks and introduces a weight decay mechanism that assigns differentiated weights to different sequence blocks based on temporal distance, thereby capturing changes in users' short- and long-term interests more accurately. However, this approach introduces additional parameters, thereby increasing the model complexity. While Stochastic Shared Embedding (SSE) technology can reduce the parameter count and mitigate overfitting, its random embedding approach may introduce noisy data, affecting the accuracy of the recommendations. To solve this problem, this study proposes a strategy that combines a Generative Adversarial Network (GAN) with SSE. By using GAN to generate high-quality interaction data that align with user interest distributions and combining this with the random replacement mechanism in SSE for data augmentation, the data are enhanced by randomly selecting generated data, thus reducing the risk of noise introduction while retaining SSE's advantage in reducing overfitting. Experiments are conducted on three public datasets: MovieLens-1M, Amazon Beauty, and Yahoo Music. The results show that the proposed method performs well in terms of Normalized Discounted Cumulative Gain (NDCG), Hit Rate (HR), and Mean Reciprocal Rank (MRR).

Real-world data often follow a long-tail distribution. Federated learning methods that assume a balanced global data distribution struggle to classify tail-class data within long-tail data accurately. Researchers typically focus on retraining a balanced classifier for the global model, to mitigate the impact of long-tail data. However, this approach does not consider the feature extractor of the balanced model or how the model's feature extractor can be enabled to learn high-quality image features, leading to the poor performance of the global model. To enable the model to learn high-quality image features without bias during the feature learning stage, this study proposes a federated learning method combining rotational self-supervision and Contrastive Language-Image Pre-training (CLIP) guidance. This method uses rotational self-supervision to guide the training of local client models, thereby reducing the impact of long-tail data on the client models and enabling the model to learn high-quality image features. Simultaneously, CLIP is utilized to guide both the normal training of the model and the rotated images, transferring rich knowledge from CLIP to the client model and further enhancing the performance of the feature extractor. In experiments on the CIFAR-10 and CIFAR-100 datasets under different long-tail distributions, the proposed approach improves the global model's classification accuracy by 2.35 to 4.72 percentage points, respectively, compared with other federated learning methods.

To address issues such as a single point of failure, risk of tampering, opaque verification, and the proliferation of false information in centralized news architectures, a blockchain-based decentralized news-retrieval and aggregation architecture is proposed for trusted storage, verifiable retrieval, and transparent governance of multi-source news data. This architecture integrates blockchain, smart contracts, and distributed storage to form an integrated trust system of "chain-contract-storage." The consensus mechanism ensures the credibility of the data sources, whereas smart contracts automatically enforce governance rules for traceability and credibility. A Multi-Valued modulo Function (MVF) robust task allocation algorithm and a Key/Transaction Merkle-Patricia Tree (KMPT/TMPT) dual-layer verifiable indexing mechanism are proposed to optimize task scheduling and index verification, thereby enhancing retrieval and verification efficiency. The system incorporates Merkle Hash Tree (MHT) integrity verification and a multi-source reputation weighting mechanism to achieve adaptive adjustment of the source reputation, thereby enhancing retrieval accuracy and system robustness. The system is deployed in a private cloud environment on an OpenStack and experimentally validated using 106 532 news data collected in 2024. Experimental results show that compared to traditional solutions, the credibility verification accuracy of this architecture is improved by 15.6% (P < 0.01), the tamper resistance detection success rate reaches 99.6%, and the false news suppression rate is 92.4%. Deep integration of the retrieval and verification processes enhances the comprehensive effectiveness of trusted retrieval by 22.3% (P < 0.05). This study verifies the feasibility and engineering effectiveness of blockchain in trusted retrieval and data governance, providing theoretical support and an engineering reference for building a highly transparent and traceable news ecosystem.

As a mainstream tool for container orchestration, Kubernetes can support automatic deployment, service discovery, and load balancing. It is known for its high availability and performance. However, scheduling strategies such as the best adaptation algorithm or the minimum negative cut method ignore the heterogeneity and energy differences of nodes. In addition, Kubernetes tools only consider CPU and memory resources and set a unified weight mechanism in advance, which can easily lead to problems such as load imbalance, performance degradation, and the inability to satisfy refined scheduling. To address these existing problems, this study proposes a heterogeneous task scheduling algorithm based on multi-dimensional resources, namely A-KCSS. The A-KCSS algorithm is based on the heterogeneous computing resources of a cluster. It adds disk Input/Output (I/O), network I/O load, and GPU resources as evaluation indicators for filtering and screening, and it more comprehensively considers the heterogeneity of nodes. This study also introduces a weight calculation model based on multi-dimensional resource factors. Based on the resource requirements of the task to be scheduled, the weight value of each dimension of the resource factor of the task to be scheduled is calculated, and the score of each node is calculated based on the real-time resource utilization of the cluster. Nodes are prioritized according to the score, and the node with the highest priority is selected for scheduling. The performance of the A-KCSS algorithm is experimentally verified on the Kubernetes cluster. Compared with the default scheduling algorithm and the KCSS algorithm, the average response time is reduced by 10% and 4%, the throughput is increased by 30% and 15%, the availability is improved by 40% and 30%, and the load balancing performance is increased by 23% and 18%, respectively, thereby improving the overall cluster performance.

Blood pressure is a key indicator of cardiovascular health. To monitor daily blood pressure at home, researchers continue to use non-end-to-end measurement methods, such as combining multiple physiological signals, in addition to traditional blood pressure meters. These methods present some disadvantages: collecting multiple physiological signals is difficult and expensive, and maintaining the time synchronization of signal collection is challenging. Currently, the end-to-end method, which uses face videos to predict blood pressure, has widened the application areas to an extent; however, in most methods, selecting the areas of interest and limited prediction accuracy remain unresolved issues. To solve these problems, this paper proposes a remote blood pressure prediction method based on a multi-scale attention structure in visible light, called EBP-Net. First, each face video candidate window is classified, the effective skin region is extracted by regression, the remote Photoplethysmography (rPPG) signal is extracted from the continuous effective face region by optical flow-based technology, and the complete rPPG signal is filtered by wavelet transform. Robust rPPG signals are extracted using methods such as detrending. Second, EBP-Net introduces a new Efficient Multi-scale Attention (EMA) module and Multi-Scale Fusion (MSF) module, which can enhance the features of depth vision representation without reducing the channel dimension and significantly improve the model's ability to understand and predict physiological signals by capturing and hierarchically expressing multi-scale features. In experiments, Systolic Blood Pressure (SBP) is categorized as grade C according to the British Hypertension Society (BHS) on both datasets, while Diastolic Blood Pressure (DBP) is categorized as grade B. The Mean Absolute Error (MAE) values are 6.82 and 5.17 mmHg for systolic and DBP, respectively, which are lower than those in recent comparable studies. Compared with other models, this method has better generalization ability, a lower error rate, and provides effective methods and suggestions for the detection blood pressure level from face videos.

Cyberattacks are becoming increasingly diverse, and traditional intrusion detection methods exhibit limitations in capturing the spatio-temporal characteristics of complex network traffic. Most traditional methods rely primarily on static feature analysis, making it difficult to adapt to the ever-changing intrusion behaviors in dynamic network environments. When analyzing network traffic, existing deep learning methods often overlook the topological structure between network nodes and the temporal dynamics of traffic. To address these issues, this paper proposes a novel intrusion detection method based on a dynamic spatio-temporal Graph Neural Network (GNN), named DSTG-IDS. By segmenting network traffic through time windows, data packets within each time period are modeled as nodes in a graph, and connections are established based on the relationship between the source and destination IP, thereby constructing a sequence of dynamic graphs over time. To better capture the temporal characteristics of traffic, graph data are encoded with a temporal position to enhance the temporal information expression ability of the nodes within different time periods. In terms of model design, first, Graph Convolutional Network (GCN) are utilized to extract the spatial features of network traffic, and Graph Attention Network (GAT) are incorporated to enhance the focus on key node information. Second, Bidirectional Gated Recurrent Unit (Bidirectional GRU) are employed to model the temporal sequence of traffic, effectively capturing the dynamic characteristics of data changes over time. Finally, a multi-head attention mechanism is utilized to fuse spatio-temporal features and perform classification. Experiments on three widely used datasets—BoT-IoT, ToN-IoT, and NF-CSE-CIC-IDS2018—demonstrate that DSTG-IDS achieves accuracies of 99.69%, 98.61%, and 93.26%, respectively. Compared with other intrusion detection methods, DSTG-IDS exhibits significant advantages in terms of accuracy, recall, False Alarm Rate (FAR), and F1 value.

This study proposes a skin melanoma segmentation algorithm, YOLOv8-Skin, designed to address the issue of imprecise results in existing algorithms caused by diverse shapes and blurred edges. YOLOv8-Skin combines multiscale feature extraction and enhanced edge segmentation based on YOLOv8. First, the backbone network CSPDarkNet53 of YOLOv8 is replaced with U-Net v2, which is more suitable for medical image segmentation. This change introduces rich semantic information into low-level features and refines high-level features, enabling precise delineation of lesion boundaries and effective extraction of small structures in melanoma images. Second, a Deformable-Large Kernel Attention (D-LKA) mechanism is introduced into the neck's C2f, enhancing the model's ability to capture irregular image structures through deformable convolutions and improving multilevel feature fusion using large kernel convolutions. Finally, a Diverse Branch Block (DBB) is incorporated into the head, forming a new segmentation head that enhances the representation capability of single convolutions by combining diverse branches of different scales and complexities. This enriches the feature space and improves feature extraction. Experiments conducted on the ISIC2017, ISIC2018, and PH² datasets verify the algorithm's effectiveness. On the ISIC2017 dataset, the Dice coefficient, Specificity, Sensitivity, and Accuracy reach 88.86%, 91.34%, 97.24%, and 96.29%, respectively. On the ISIC2018 dataset, they reach 91.64%, 95.42%, 96.69%, and 95.83%, respectively. On the PH² dataset, they reach 95.92%, 95.43%, 97.02%, and 96.13%, respectively. The algorithm demonstrates stronger segmentation performance and is better suited for melanoma segmentation tasks compared to existing methods.

To support real-time data transmission and maintain stable connectivity for diverse mobile users, an air-ground cooperative Ultra-Dense Network (UDN) is adopted to enhance network service capabilities. In this novel network architecture, airborne base stations can better support user mobility, while terrestrial UDNs enable more efficient spectrum reuse. However, differentiated channel conditions and coverage disparities introduce new challenges in interference and spectrum allocation. Therefore, this paper focuses on a novel air-ground cooperative UDN and addresses the challenges of interference coordination and spectrum efficiency balancing in mobile scenarios within this network model. A Graph Neural Network (GNN)-based Deep Reinforcement Learning (DRL) algorithm, Graph-DRL, is proposed to dynamically schedule spectrum resources and improve the real-time performance of the algorithm. First, a GNN is utilized to model the user movement correlations among neighboring base station cells. By sharing the demand information of the local cell, cooperation is enhanced to determine the bandwidth reuse factor. Second, a value-based learning network is designed, with the average throughput of the network as the optimization objective. Considering factors such as user mobility, user density, and spectrum reuse degree, a trained Q value network is employed for resource allocation. Simulation results demonstrate that, compared to benchmark algorithms, the proposed algorithm significantly improves the transmission rates for highly mobile users and cell-edge users.