最新刊期
- 摘要：Path planning is a core step in autonomous navigation for mobile robots， aiming to generate a collision-free and cost-minimized trajectory from the starting point to the target point. To enhance the efficiency and global optimality of path planning for mobile robots， an IBA-IDWA algorithm that integrates improved bidirectional A* search and dynamic window approach is proposed. This algorithm introduces a heuristic function based on obstacle distribution in A* search to improve search efficiency， and designs a cost function based on the distance between trajectory endpoints and local goals in the dynamic window approach to enhance global consistency. Additionally， the algorithm adaptively adjusts the robot speed window size based on the density of environmental obstacles by predicting obstacle movement patterns and identifying dangerous areas. Simulation experimental results show that compared to the traditional A* algorithm， the planning efficiency of the improved bidirectional A* algorithm is improved by 14.07%. In dynamic obstacle avoidance tasks， compared to existing fusion algorithms， the obstacle avoidance success rate of IBA-IDWA is increased by more than 5%. The IBA-IDWA algorithm significantly improves path planning efficiency and path quality in both static and dynamic environments， providing an effective solution for autonomous navigation of mobile robots.关键词：mobile robot;path planning;bidirectional A* algorithm;dynamic window method4|1|0更新时间：2026-07-30
- 摘要：Imbalanced data classification remains a significant challenge in machine learning. Conventional methods treat feature selection and oversampling as independent processes. On the one hand， this independence distorts feature importance evaluation and introduces sampling noise； on the other hand， feature assessment typically considers only the discernibility between samples from different classes while neglecting the similarity within the same class， making it difficult to accurately identify discriminative features. To address this issue， this paper proposes a unified processing algorithm for imbalanced data based on the similarity-discernibility matrix （ISDM）. The algorithm introduces the Similarity-Discernibility Matrix （SDM）， which simultaneously captures the discernibility between different classes and the similarity within the same class， enabling multi-perspective quantification of discriminative information for both samples and features， thereby providing a unified measurement basis for feature selection and sample quality evaluation. The algorithm first employs SDM to remove redundant features and construct a low-dimensional feature space， then performs clustering and weighted interpolation on minority class samples based on SDM-computed sample importance， ensuring that feature selection and oversampling collaborate under a consistent information measure， avoiding performance loss caused by information fragmentation. Experimental results on 12 benchmark datasets demonstrate that ISDM achieves F-measure of 85.69% and 84.43% under CART and NN classifiers respectively， and AUC of 91.67% and 92.23% respectively， with average improvements of 2.22% to 9.88% over comparative algorithms， validating the effectiveness of the algorithm.关键词：imbalanced data;similarity-discernibility matrix;feature selection;oversampling;rough set theory4|0|0更新时间：2026-07-30
- 摘要：Real-time and accurate traffic flow prediction holds significant practical application value for urban management and governance. Existing methods have limitations in modeling dynamic spatio-temporal dependencies and feature extraction. To further improve prediction accuracy， this study proposes a Feature-Enhanced Dynamic Graph Convolutional Neural Model with Attention Fusion for Traffic Flow Prediction （AFEDGCN）. The model integrates spatial attention with dynamic graph convolution to enhance the modeling of node spatial correlations. Meanwhile， interactive convolution is utilized to strengthen the features of individual nodes， and the Gated Recurrent Unit （GRU） is incorporated to model synchronous spatio-temporal dependencies. Multi-head temporal attention is employed to capture temporal correlations， while global graph convolution is used to model the global spatial correlations over the entire time period. Finally， spatio-temporal features are adaptively fused to predict future traffic flow conditions. Experiments conducted on multiple real-world traffic datasets show that， compared with the advanced GDGCN and DGCRN models， the proposed model reduces the MAE （Mean Absolute Error） by 1.65% and 4.80% respectively， and the MAPE （Mean Absolute Percentage Error） by 2.36% and 3.93% respectively. These results indicate that the proposed model has better performance compared with the current state-of-the-art models.关键词：traffic prediction;spatial attention;multi-head temporal attention;feature enhancement;dynamic graph convolution3|0|0更新时间：2026-07-30
- 摘要：To address the issues of multi-source information uncertainty and conflicting expert opinions in the risk assessment of electric power material contract performance， a risk assessment and consensus reaching model integrating Hesitant Fuzzy Linguistic Term Sets （HFLTS） and improved Dempster-Shafer Evidence Theory （DSET） is proposed. Firstly， a multi-dimensional risk index system is constructed， utilizing HFLTS to handle the fuzziness of expert evaluations. Secondly， a two-dimensional dynamic reliability measurement method based on hesitancy degree and similarity degree is proposed， incorporating expert reliability into the calculation of Basic Probability Assignment （BPA） to dynamically revise the weights of evidence sources. Thirdly， a dual feedback consensus mechanism featuring "explicit guidance + implicit suggestion" is designed to improve the consistency of group opinions through iterative optimization. Finally， the comprehensive risk assessment results are generated based on the improved DSET fusion rule. Case analysis demonstrates that the model identifies "Conflict with Government Planning" （risk value 0.302） as the primary risk. When dealing with conflicting evidence， the dynamic reliability mechanism automatically reduces the weight of low-quality evidence sources from 17.0% to 5.1%， and the dual feedback mechanism achieves a high-level consensus of 0.902 within only 6 iterations. The proposed model outperforms traditional methods in terms of convergence speed and anti-interference capability， providing quantitative decision-making support for electric power material management.关键词：electric power materials;risk assessment;D-S evidence theory;group decision-making;system simulation4|1|0更新时间：2026-07-30
- 摘要：Adversarial examples can mislead classification models into making incorrect predictions through minor perturbations， making the development of adversarial example detection technology crucial for enhancing model robustness. Existing detection methods predominantly rely on features from the model's final output layer， leaving the hierarchical representational information within the model's internal layers underutilized. This study analyzes the representational differences between adversarial examples and original samples in the hidden layer outputs of pre-trained language models and proposes a hierarchical adversarial example detection framework. By recording and computing the similarity of hidden layer outputs between original and adversarial samples before and after model fine-tuning， we find that the variation patterns of inter-layer similarity differ significantly between adversarial and original samples. Using layer-wise similarity as the core metric， a One-Class SVM model is trained to learn the distribution characteristics of layer similarity in normal samples， thereby effectively identifying adversarial examples. The proposed detection framework was evaluated on four benchmark text classification datasets under three attack methods. Experimental results show that it outperforms the current state-of-the-art algorithms in terms of both F1 and AUC scores， achieving up to an 8.6% improvement over the strongest baseline. Additionally， the framework demonstrates good generalization capabilities across different pre-trained encoder architectures.关键词：adversarial sample detection;defense mechanisms;intermediate layer features;model robustness enhancement;One-Class SVM4|1|0更新时间：2026-07-30
- 摘要：Chest X-ray imaging served as an essential tool for clinical screening and diagnosis of thoracic diseases. Existing approaches often relied solely on local feature extraction， failing to effectively utilize contextual information and anatomical priors， which limited their diagnostic accuracy and robustness. To address these issues， we proposed a dual-guided context refinement module， which integrated anatomy-guided spatial attention and edge-guided channel attention to enhance feature representation from both anatomical and structural perspectives. Meanwhile， to overcome limitations in modeling label co-occurrence and semantic dependencies， we constructed a learnable label co-occurrence knowledge base to capture co-occurrence relationships and semantic similarities among disease labels. Experimental results demonstrated that the proposed method achieved AUC scores of 85.0% and 89.6% on two public CXR datasets， ChestX-ray14 and CheXpert， respectively. These findings indicate that the proposed framework effectively improves multi-label chest disease recognition and exhibits strong generalization capability.关键词：multi-label classification;anatomical prior;label co-occurrence;knowledge base;medical imaging4|1|0更新时间：2026-07-30
- 摘要：Infrared small target detection aims to identify small， low-contrast targets from complex backgrounds. Its performance depends on the model's ability to extract weak signal features， accurately characterize boundary regions， and suppress background interference. Existing methods generally suffer from insufficient interaction between shallow structural information and deep semantic features， and poor adaptability of static single-scale fusion strategies， leading to blurred target boundaries and decreased recognition accuracy in complex backgrounds. Based on the MSHNet model， this paper proposes a detection model with a bidirectional feature interaction and dynamic scale-aware fusion mechanism. The model achieves efficient feature interaction between the encoder and decoder through a bidirectional feature interaction module， enhances target region response and boundary discrimination capabilities using lightweight channel attention， and then introduces a dynamic scale-aware fusion module. This module adaptively allocates fusion weights based on the contribution of features at different scales to target recognition， thereby optimizing feature integration and strengthening boundary representation， effectively improving the target boundary recognition capability. Experimental results on the NUDT-SIRST and IRSTD-1K datasets show that the proposed model improves the intersection-union ratio by 1.7% and 2.0% respectively compared to the benchmark algorithm， validating the model's effectiveness.关键词：infrared small target detection;bidirectional feature interaction;dynamic scale-aware fusion;U-Net;deep learning2|0|0更新时间：2026-07-30
- 摘要：Optimizing planting schemes under limited arable land is vital for rural development. This study employs a mixed-integer linear programming and evolutionary strategy model to optimize village planting strategies. By setting minimum planting area constraints and applying discretization， a net income-maximizing model is built. Long-term needs are incorporated through Spearman’s rank correlation to assess crop substitutability and complementarity， while evolutionary strategies enhance adaptability. Experimental results show that this method better responds to market and environmental changes than traditional linear programming， significantly improves revenue， and demonstrates strong robustness and generalizability， providing scientific support for agricultural planning.关键词：optimal planting plan;mixed-integer linear programming model;evolutionary strategy model;Spearman’s rank correlation coefficient;time series analysis4|0|0更新时间：2026-07-30
- 摘要：Multimodal video analysis serves as a crucial approach for understanding complex dynamic scenes and human behavior. However， existing vision-language models often rely on a single， rigid temporal modeling paradigm， which struggles to adaptively handle diverse dynamic patterns in videos （such as static backgrounds， short-range actions， and long-range periodic behaviors）， thereby limiting the depth and generalization capabilities of video analysis. To address this， a Dynamic Multi-Expert Fusion （DMEF） framework is proposed. First， an adaptive expert scheduler is designed to dynamically select and integrate three different granularities of temporal modeling experts based on the content features of the input video： a global static expert maintains static semantic generalization， a short-range dynamic expert focuses on capturing local frame-to-frame changes， and a long-range contextual expert models global temporal dependencies. Second， an intelligent modeling architecture is established that can adaptively adjust according to video content， enhancing the interpretability of video content analysis. In few-shot learning and cross-domain generalization tasks， the DMEF method achieves optimal performance on the HMDB51 and UCF101 datasets； on the SSv2 dataset， it performs best when K=8. In cross-domain generalization tasks， the DMEF method demonstrates excellent recognition accuracy on the HMDB51 dataset.关键词：video action recognition;few-shot learning;cross-domain generalization;vision-language model;temporal modeling;multimodal video analysis3|1|0更新时间：2026-07-30
- 摘要：Named entity recognition is an important task in natural language processing. However， real-world applications often suffer from class imbalance， which makes minority entity types difficult to recognize and consequently affects the overall performance of NER models. Existing methods mainly rely on linear or hyperspherical decision boundary assumptions， which are insufficient to capture the complex feature distributions of minority samples in the representation space， thereby limiting the discriminative capability of models for minority entities. To address this issue， an adaptive boundary optimization-based named entity recognition method （ABO-NER） is proposed. The proposed method constructs flexible decision boundaries using nonlinear discriminative functions and independently models features for different entity classes. Meanwhile， dual constraints of intra-class compactness and inter-class separability are introduced to achieve adaptive optimization of the decision boundaries， thereby improving the recognition performance of minority entities. Experimental results on six public NER datasets （ACE2005，GENIA，MOVIE，OntoNotes5.0，JNLPBA and Few-NERD） demonstrate that our method outperforms all existing baseline models in both Macro-F1 and Micro-F1 scores. Specifically， the Macro-F1 score is increased by 3.93%， 3.60% and 2.94% on the GENIA，JNLPBA and OntoNotes5.0 datasets under extremely imbalanced settings， respectively， which verifies the effectiveness and robustness of our method in class-imbalanced scenarios.关键词：named entity recognition;class imbalance;adaptive boundary optimization;decision boundary;nonlinear discriminant function;minority class recognition4|1|0更新时间：2026-07-30
- 摘要：Aiming at the key problems of answer leakage， uncontrollable strategies， and insufficient guidance effectiveness in the current Large Language Model （LLM） technology in teaching， a guided mathematics teaching system， CoT_MathTutor， based on the thinking chain and LLM， was designed with a focus on the scenario of mathematical intelligent tutoring. Firstly， by combining the big language model with the thought chain and constraining the generation process of LLM through a finite state machine mechanism， a controllable progressive heuristic teaching has been achieved； Secondly， the StepCoT method is proposed to break down each problem-solving step into a fine-grained structured structure， transforming macro problem-solving ideas into a logically rigorous and semantically clear multi-step chain structure， in order to promote the cultivation of students' autonomous reasoning ability. Experiments have shown that in primary school mathematics application problem tutoring tasks， the problem solving success rate of the CoT_MathTutor system based on GPT4 has increased by 18%， and the teaching score has increased by 0.18， demonstrating strong dynamic adaptability and strategic flexibility. Compared with the open generation GPT-4， the improvement is significant. The modular design of the system not only ensures the controllability of teaching strategies， but also provides a flexible framework for optimizing and expanding subsequent teaching methods， providing important references for the application of LLM in the field of education and promoting the development of intelligent tutoring systems towards more effective and controllable directions.关键词：LLM;CoT;math teaching;heuristic teaching;state machine;intelligent education4|2|0更新时间：2026-07-30
- 摘要：With the widespread adoption of Internet technology， the southern Xinjiang region faces severe challenges from the dissemination of multimodal sensitive information. This paper innovatively proposes a "cultural semantics-technical features" dual-driven analytical framework. By constructing a multimodal dataset， developing a cross-modal feature fusion model， and incorporating dialect-robust representation learning， among other key technologies， a hierarchical intelligent recognition system is designed. Experimental results demonstrate that the system achieves a sensitive content recognition accuracy of ≥92% and a misjudgment rate of ≤10%， while reducing the detection time to the millisecond level. Specifically， in standard corpus scenarios， the accuracy reaches 93.2%； in dialect scenarios， the false detection rate is reduced to 8.7%； and the misjudgment rate for cultural symbols is controlled within 3.2%， representing a 5-fold efficiency improvement over traditional methods. This study provides a three-in-one digital paradigm of technology empowerment， ethical constraints and culture respect for border area governance.关键词：Uyghur language;sensitive information;multimodal dataset;feature fusion model;robustness5|1|0更新时间：2026-07-30
- 摘要：Large language models can be effectively adapted to specific tasks through fine-tuning， while federated learning enables in-situ computation using local device data， further enhancing privacy-preserving fine-tuning. Existing methods typically employ low-rank adaptation techniques for efficient federated fine-tuning but rely on traditional federated learning aggregation strategies， failing to fully exploit the potential of large language models and inadequately addressing the challenges posed by heterogeneous model configurations. To this end， a novel approach called FedPDS is proposed， combining the difference of low-rank matrix products with adaptive singular value decomposition （SVD）. First， the difference of low-rank matrix products is utilized to introduce additional model parameters， expanding the representation space and effectively overcoming the representational limitations of low-rank matrices in downstream tasks. Second， an adaptive SVD mechanism is deployed on the server side to dynamically reallocate parameters in response to the computational constraints of heterogeneous clients. Experiments demonstrate that FedPDS significantly outperforms existing low-rank adaptation methods in both homogeneous and heterogeneous federated settings， achieving faster convergence and superior performance.关键词：federated learning;large language model;low-rank adaptation;model heterogeneity;SVD3|0|0更新时间：2026-07-30
- 摘要：Weld quality is directly related to the structural integrity and operational safety of industrial equipment. Efficient and accurate detection of weld surface defects is essential for intelligent manufacturing and quality control. However， traditional weld defect detection methods often suffer from limited detection accuracy and poor generalization ability in complex interference environments. To address these challenges， this paper proposes a lightweight improved model， termed YOLOv12-DSD， based on YOLOv12 for efficient and accurate weld surface defect detection. First， a dynamic convolution mechanism is introduced into the feature extraction stage， and a C3k2_D module is constructed to enhance the model's adaptability to complex texture features and multi-scale defect targets. Second， an SMPDIoU loss function is designed for the bounding box regression stage to improve the modeling capability of geometric structural differences and localization errors. Finally， a DyHead module is incorporated into the detection head， leveraging attention mechanisms across scale， spatial， and task dimensions to strengthen feature representation and discrimination for multiple defect categories. Ablation and comparative experiments are conducted on a weld defect dataset containing 1，414 images. The results demonstrate that the proposed YOLOv12-DSD model contains only 2.13 M parameters while achieving a precision of 92.41%， a recall of 74.80%， and an mAP@0.5 of 87.66%， outperforming existing mainstream object detection algorithms in overall performance. The experimental results indicate that the proposed model improves detection performance while maintaining a lightweight architecture， demonstrating strong potential for industrial deployment and practical applications.关键词：weld defect detection;dynamic convolution;lightweight model;YOLOv124|1|0更新时间：2026-07-30
-
Technical Research and Implementation of A PCIe Endpoint Mode Cryptographic Card in System-on-Chip AI导读
摘要：Cryptography serves as the foundation of information security， ensuring data integrity， confidentiality， and non-repudiation. PCIe cryptographic cards， a key class of security hardware， are widely deployed but often hindered by complex designs， insufficient anomaly detection， and cumbersome firmware updates. To overcome these challenges， this paper presents a PCIe cryptographic card based on a System-on-Chip（SoC） architecture. By integrating a Dual-Mode（DM） PCIe controller and cryptographic algorithm engines into the SoC， the hardware design is significantly streamlined. A flexible software stack is also implemented， comprising boot firmware， a Linux kernel endpoint（EP）framework， algorithm engines， and service daemons. The system supports access control， key management， cryptographic operations， enhanced status monitoring， exception handling，firmware hot-upgrade， and comprehensive logging. Experimental results on the RK3588 platform demonstrate that the DMA throughput achieves 72% of the theoretical maximum with 2 MB data packets.关键词：system-on-chip （SoC）;cryptographic card;PCIe endpoint （EP）;direct memory access （DMA）;state monitor3|1|0更新时间：2026-07-30 - 摘要：To address the issue of degraded localization accuracy or even positioning failure in autonomous vehicles under visually degraded environments， this paper proposes an IMU-centric Global Navigation Satellite System（GNSS）-visual-LiDAR-IMU multi-sensor fusion SLAM system （ICAE-SLAM）. The overall system framework is designed with three subsystems： the Visual-IMU system （VIS）， the LiDAR-IMU system （LIS）， and the GNSS-IMU system （GIS）. First， the IMU mechanization process is designed to provide precise prior attitude information to both VIS and LIS， where the IMU-centric assistance in the initialization of other subsystems. Subsequently， observations from IMU， visual， LiDAR， and GNSS are transformed into error factors for constructing a multi-sensor fusion factor graph model. Finally， to tackle the high computational load in the back-end optimization process， an adaptive elimination method is adopted within the sliding window to convert the factor graph into a Bayesian network. Tested on the KITTI and M2DGR datasets， the proposed method achieves average trajectory errors of 2.379 m and 2.699 m， respectively， with an average CPU utilization rate of 31.9%. Compared to VINS-Mono， LIO-SAM， LVI-SAM algorithms， the localization accuracy is improved by 77.26%， 19.78% and 45.75% on average， respectively.关键词：autonomous driving;multi-sensor fusion;simultaneous localization and mapping （SLAM）;factor graph optimization;variables elimination5|1|0更新时间：2026-07-30
- 摘要：Case retrieval is a key technology for improving the consistency of judicial rulings and the accuracy of legal application， and is increasingly important in the smart judicial system. Existing methods often focus on text semantic similarity and surface matching， making it difficult to model the judicial logic and legal requirements behind the case. Therefore， a case retrieval method based on large language model element fact extraction and inference guidance is proposed. Firstly， utilizing the big language model and syntactic analysis to identify the core behavioral facts in the judgment document， and constructing legal element facts； Secondly， introducing a legal knowledge graph for semantic enhancement and reasoning guidance enhances the consistency modeling ability of judicial logic. The experimental results indicate that the proposed method is effective in Precision@5 /10 （49.09%， 48.93%）， MAP （50.38%）， and NDCG@10 /On indicators such as 20/30 （79.27%， 84.16%， 88.94%）， it outperforms the baseline model and has a significant advantage in case matching， providing theoretical support and technical path for building case retrieval with the ability to perceive judicial logic.关键词：intelligent judiciary;case retrieval;large language models;legal elements;legal knowledge graphs3|0|0更新时间：2026-07-30
- 摘要：Joint Query Encryption Multi Mapping Technology （cEMM） supports multi label joint queries while ensuring the confidentiality of outsourced data. In order to address the data privacy risk caused by the leakage of query result patterns， a joint query encryption multi mapping scheme with result pattern hiding characteristics was designed. By optimizing the query token and server storage structure， joint queries can be completed without leaking the query result pattern. Security analysis shows that this scheme can effectively hide the query result pattern， and the communication overhead is reduced by 69.75% compared to existing representative schemes.关键词：symmetric searchable encryption;encrypted multi-map;conjunctive search;access pattern;communication efficiency4|1|0更新时间：2026-07-30
- 摘要：Foggy environments reduce the clarity of road images and seriously affect road surface distress detection. To address the problems of insufficient detail recovery and relatively high model complexity in existing dehazing methods for road distress detection， a multi-scale all-in-one dehazing network is proposed. First， a multi-scale feature extraction and weighted fusion module is introduced into the network. Features are extracted through three branches at different scales， enhanced by channel and spatial attention mechanisms， and then aggregated by a fixed-weight fusion strategy， thereby strengthening the contextual representation of both near and far scenes. Second， a hybrid loss function composed of Mean Absolute Error （MAE）， Mean Squared Error （MSE）， and Structural Similarity （SSIM） is designed. Then， a parameter optimization strategy based on an improved RIME algorithm is employed to optimize key parameters such as the learning rate. Finally， experiments are conducted on a synthetic foggy road defect dataset. The results show that the proposed network has only 0.137M parameters and runs at 115 f/s， indicating a high degree of lightweight design， while outperforming other methods in terms of Peak Signal-to-Noise Ratio （PSNR）， Structural Similarity （SSIM）， and related metrics. In addition， it restores richer texture details in critical distress regions， providing a solid foundation for subsequent foggy road surface distress detection.关键词：image dehazing;road damage detection;hybrid loss;YOLOv11;FRIME optimization algorithm4|2|0更新时间：2026-07-30
- 摘要：Speech emotion recognition is an important branch of natural language processing that aims to automatically recognize and classify emotional information in speech through computer technology. In the specific language environment of Tibetan， due to the lack of relevant research and the widespread use of multi feature fusion strategies in existing studies， there are problems such as cumbersome feature extraction， high feature dimensionality， and high computational overhead. To this end， a lightweight network model ST CapsNet is proposed by using 45 dimensional Mel frequency cepstral coefficients （MFCC） to represent audio emotions. Firstly， the spatiotemporal feature extractor and capsule network are used as the core to capture the temporal and spatial information of MFCC through the spatiotemporal feature extractor； Secondly， input the information into the capsule network for deep analysis. The experiments on the self built Tibetan speech emotion corpus TSEC-4528 and the publicly available corpora EMO-DB and RAVDESS （Speech&Song） showed that the proposed model achieved unweighted accuracies of 78.59%， 96.34%， 87.60%， and 97.54%， respectively， demonstrating the effectiveness and cross lingual adaptability of the method for Tibetan speech emotion recognition.关键词：Tibetan;speech emotion recognition;Mel-frequency cepstral coefficients;capsule net4|2|0更新时间：2026-07-30
- Technical support is provided by Beijing Founder electronics co., LTD 鄂ICP备18021494号-4
鄂公网安备42010602004949号 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0