Abstract:Spacecraft exhibit intricate structural designs and operate in highly dynamic environments, where anomaly detection plays a pivotal role in monitoring spacecraft status and ensuring mission success. Traditional anomaly detection methods face challenges of low accuracy and poor adaptability when applied to non-stationary and noise-contaminated telemetry data. This paper proposes an intelligent anomaly detection framework integrating representation learning with dynamic threshold optimization to enhance detection accuracy and reliability. The methodology comprises three key components. First, a stacked autoencoder performs nonlinear dimensionality reduction on high-dimensional time-series data to extract low-dimensional intrinsic features while suppressing noise interference. Second, a Neural Circuit Policies model with bio-inspired sparse connectivity and adaptive time constant mechanisms is employed for temporal pattern modeling and prediction. Finally, a multi-objective optimization algorithm dynamically adjusts anomaly thresholds to balance precision and recall, effectively addressing the performance degradation of fixed thresholds under abrupt data distribution shifts. In the drift anomaly detection experiments on the simulation dataset, the proposed method achieved F1-score improvements of 65.1%, 50.5% and 8.8% compared to LSTM, Transformer and TFT methods, respectively. For the real-world satellite dataset, our approach demonstrated superior performance with F1-score enhancements of 53.0%, 51.0% and 41.0% over the same three baseline methods. Experimental results on two spacecraft datasets demonstrate that the proposed method significantly improves prediction accuracy compared to baseline methods. The proposed framework establishes a novel technical approach for on-orbit autonomous health management of spacecraft, with substantial implications for enhancing the safety and reliability of deep space exploration missions.

Abstract:Short-wave infrared(SWIR) gas imaging technology is affected by factors such as detector fabrication processes and readout noise, which result in non-uniformity and large clusters of blind pixels in infrared images. These issues degrade the imaging quality, leading to errors in gas contour extraction. To overcome the aforementioned limitations, a CO2 imaging system based on a quantum dot short-wave infrared focal plane detector was developed. The system, integrated with a supercontinuum laser and monochromator, generates short-wave infrared laser wavelengths at 1 200 nm and 1 578 nm. This configuration facilitates gas absorption spectral imaging using tunable diode laser absorption spectroscopy. Correlated double sampling technology was employed to remove reset and low-frequency noise from the detector. A differential two-point correction algorithm was used to reduce image non-uniformity to 4.91%. Additionally, a spatial compensation-based correlated pixel compensation algorithm was implemented to effectively eliminate both clustered and isolated blind pixels, achieving a blind pixel rate of 0.006%. Finally, background subtraction was applied to extract the gas plume contours. Experimental demonstrates that the system is capable of imaging and detecting 2% CO2 gas plumes at different flow rates within 1 s.

Abstract:Aiming at the problems of insufficient utilization of data spatial-temporal characteristics and poor generalization ability of e traditional network traffic anomaly detection methods, a traffic anomaly detection method based on multi-head attention mechanism and spatial-temporal feature fusion is proposed. The convolutional neural network(CNN) is utilized for the extraction of the spatial local features presented within the traffic data. The multi-head attention mechanism is introduced to achieve multi-angle adaptive reweighting of key features through parallel computation of multiple attention heads, thus improving the sensitivity of the model to abnormal traffic. The re-weighted features are then input into the bidirectional long short-term memory network(BiLSTM) to mine the long-distance temporal dependencies in the traffic data. Finally, Softmax is used to classify and identify the traffic data. Experiments are carried out on the publicly available dataset NSL-KDD and CIC-IDS-2017 with a detection accuracy of 85.40% and 99.41%, respectively, which verifies the effectiveness of the method in the task of network traffic anomaly detection.

Abstract:This paper designs and implements a high-speed image acquisition and storage system based on FPGA, aiming to address the performance bottlenecks in high-speed image data acquisition and storage. The system receives high-speed image data through the Camera Link interface and utilizes an FPGA-implemented SATA protocol controller to efficiently and stably store the data on SATA hard drives. The image acquisition module employs Xilinx FPGA’s built-in LVDS transceivers and specific primitives(such as IDELAYE3 and ISERDESE3) to directly process the Camera Link protocol, replacing traditional dedicated chips. A lightweight file system is designed within the MicroBlaze for configuring the SATA controller and controlling data flow. By integrating an ethernet module, the system can export data to a host computer in real-time. A C# host software was developed, providing a visualized file system interface for managing and monitoring disk and file operations.Experimental results show that the SATA controller achieves write speeds of up to 504.8 MB/s and read speeds of up to 542.0 MB/s. At a 400 MB/s data acquisition rate, the system demonstrates excellent performance and reliability, making it suitable for high-performance image acquisition and storage applications.

Abstract:With the background of campus intelligence construction, a speed guidance strategy for campus unmanned vehicles based on vehicle road coordination environment is proposed to address the traffic congestion problem caused by unmanned transport vehicles at three-way intersections on campus, and adjust the operating status of campus unmanned vehicles at three-way intersections. Firstly, establish the driving rules that unmanned transport vehicles must follow for safe operation on campus, and design the layout and indication rules of traffic lights at three-way intersections. Secondly, define and assume the vehicle dynamics model, determine the range of the speed induction area, and design speed induction control strategies for campus unmanned vehicles under different signal light conditions. Adopting the S-shaped acceleration and deceleration algorithm to adjust the speed of campus unmanned vehicles and improve the smoothness of the speed curve of campus unmanned vehicles induced by vehicle speed. Finally, design a VISSIM simulation analysis platform and establish a campus unmanned vehicle speed induction model. Using the designed VISSIM simulation analysis platform, build a road network model of the campus three-way intersection, and conduct experimental analysis based on the coverage of unmanned vehicles on campus.The real vehicle experiment results show that after using the speed induction strategy for scheduling, the travel time, number of queued vehicles, and delay time of campus unmanned vehicles passing through three-way intersections are reduced by 19.3%, 47.5%, and 24.3%, respectively. The proposed campus unmanned vehicle speed guidance strategy based on vehicle road collaborative environment effectively improves the transportation efficiency of campus unmanned vehicles, which can effectively avoid campus safety accidents caused by traffic congestion, achieve the goal of intelligent scheduling, and meet the needs of campus smart construction.

Abstract:Addressing the prevalent issues of low resource utilization, incompatibility with multi-antennas, and inadequate estimation accuracy in gonal frequency division multiplexing(OTFS) pilot designs, a novel embedded pilot sequence design scheme is proposed，based on which the high accuracy channel estimation algorithm is addressed. Specifically, the proposed pilot sequence is consisted of multiple ZC sequences deployed across the Doppler domain and cascaded in the delay domain, and each ZC sequence is derived from a common root sequence through cyclic shifts of varying lengths. At the receiver, multiple channel estimations are conducted using local sequences and each corresponding received signal sequence, followed by averaging for enhanced channel state information accuracy. Based on these estimates and local sequences, interference from the pilot signal to the data signal can be cancelled. Due to the excellent orthogonality of ZC sequences, this scheme can adapt to multi-antennas systems by employing different root sequences, and this scheme markedly improves pilot accumulated signal-to-noise ratio and correspondingly the channel estimation accuracy. Simulation results show that the proposed scheme has about 6 dB signal-to-noise ratio gain for the same channel estimation accuracy, and bit-to-error rate is better, as compared to the traditional scheme, which proves that the proposed scheme has the advantages of promising performance and being very valuable to promote practicality.

Abstract:The surface crack detection and localization for titanium bar polishing was identified as a fundamental step in the manufacturing of titanium profiles. To address the issues of low detection accuracy, poor generalization ability, and low computational efficiency of traditional target detection models for narrow cracks, an improved YOLOv8s model named DEBM-YOLO was proposed. The ELA attention mechanism was added to capture long-range spatial dependencies of cracks. The DCNv3 convolution module was adopted to enhance the receptive field and representation ability of the backbone network. A bidirectional weighted feature pyramid structure replaced the original feature pyramid structure in YOLOv8 to improve multi-scale feature fusion. Finally, MPDIoU was used instead of CIoU to boost generalization performance and convergence speed. Experiments on a dataset captured in real environments showed that the improved DEBM-YOLO model reduced the number of parameters by 4.5%, increased precision by 1.9%, raised mAP@0.5 by 1.4%, mAP@0.5:0.95 by 1.9%, and recall by 4.9%. The model now achieves both enhanced detection accuracy and lightweight design.

Abstract:In order to mitigate the security risks and economic losses caused by electricity theft, and to efficiently and accurately identify theft users while discriminating their theft patterns, propose an electricity theft detection and type discrimination method that integrates a convolutional neural network(CNN) with a bidirectional long short-term memory(BiLSTM) network. Initially, an open energy data initiative(OEDI) dataset comprising 16 types of electricity users is employed. The dataset is modified according to six distinct theft patterns and subjected to Min-Max normalization to eliminate the influence of differing feature scales. Subsequently, the model extracts multi-scale local features via convolutional layers and further expands the receptive field using dilated convolution, thereby effectively capturing subtle variations amid environmental interference. Thereafter, BiLSTM is utilized to model the sequential data in both forward and backward directions, comprehensively capturing long-range dependencies and contextual information. To enhance the model′s robustness and generalization capability, dropout and dynamic learning rate adjustment mechanisms are incorporated. Finally, experiments are conducted under binary, six-class, and seven-class classification tasks with varying training set ratios. The experimental results demonstrate that the proposed method significantly outperforms traditional approaches in terms of accuracy, AUC, and F1-score, thereby validating its effectiveness in electricity theft detection and type discrimination under complex scenarios.

Abstract:Aiming at the challenges of high efficiency and accuracy of insulator defect detection for embedded devices in resource-constrained and foggy complex environments, this paper proposes a new lightweight detection model, RNSC-YOLOv7-tiny, and achieves important innovative results and practical application value. Firstly, the RepNCSPELAN module is designed by lightweighting the ELAN module in the backbone network, which effectively reduces the number of parameters and computational complexity of the model, while maintaining a significant improvement in detection accuracy. Secondly, the incorporation of the Spatial Group Enhancement module enables the model to focus on the target region overlapping with the background, thereby significantly suppressing the interference of irrelevant information and improving the accuracy of insulator defect localisation and identification. Furthermore, the incorporation of the NWD loss function addresses the issue of gradient vanishing due to deviation points in the detection process, thereby enhancing the overall detection accuracy. Furthermore, the incorporation of the CARAFE upsampling operator enables the model to achieve accurate detection and localisation in low-resolution images and complex foggy environments. The experimental results demonstrate that the RNSC-YOLOv7-tiny model exhibits rapid and highly accurate performance in insulator defect detection, with a detection accuracy of 94.8%. The model comprises 4298150 parameters and 10.5 floating-point operations, yet occupies only 8.69 MB of memory. In comparison to the original YOLOv7-tiny model, the newly proposed model exhibits notable enhancements in several pivotal metrics. The accuracy has been augmented by 3.4%, the number of parameters has been diminished by 28.5%, the number of floating-point operations has been reduced by 19.2%, and the model size has been reduced by 3.01 MB. These outcomes substantiate the algorithm′s high applicability in embedded device environments and its efficacy in practical applications.

Abstract:Aiming at the problems of small, diverse and inconspicuous features of existing printed circuit board defects, a defect detection algorithm based on ARS-YOLOv9s is proposed, which is optimised on the basis of YOLOv9s network architecture. To address the problem of information loss in multi-scale feature fusion in the original algorithm, AFPN is integrated into the image feature fusion so as to enrich the semantic information; by introducing the iRMB attention mechanism in the backbone network, the attention to the tiny defects in the shallow features is improved; to address the problem of the small target defects, the large target detection layer is deleted and a new tiny target detection layer is added, which lightens the model and improves the detection accuracy; the original model loss function is replaced by the loss function, and the original model loss function is replaced by the loss function, and the original model loss function is replaced by the loss function. The original model loss function is replaced by Shape-IoU to improve the impact of positive and negative sample imbalance on the model and accelerate model convergence. The experimental results show that the mAP of this paper′s algorithm is 98%, and that of mAP@0.5:0.95 is 68.2%, which is 2.8% and 9.3% higher than that of the original YOLOv9s, respectively, and the mAP of defects of each category is significantly improved, which proves the effectiveness of this paper′s algorithm.

Abstract:Glass insulators are critical components in transmission lines, and their defects can significantly impact the reliability of power systems. This paper proposes an improved defect detection algorithm for lightweight glass insulators based on enhanced RT-DETR, addressing issues related to low contrast and multi-scale defects. First, we introduce the lightweight backbone network RE-FasterNet, which enhances feature extraction efficiency and improves the detection of small targets and low-contrast defects through innovative partial duplication and an efficient multi-scale attention mechanism. Second, during the feature fusion stage, a partially repeated cross-stage feature fusion module is proposed to further enhance the detection capability for multi-scale defects. Finally, an attention scale sequence fusion framework is integrated into the small target detection head, significantly improving the network′s spatial feature extraction ability for small defects. Experimental results demonstrate that the proposed algorithm increases the mean average precision by 2.8%, reduces the model size by 23.6%, and decreases computational requirements by 13.1% compared to the benchmark model. In the domain of automatic defect detection for glass insulators, this approach exhibits strong practicality and broad applicability.

Abstract:To address the challenges of complex vamp component shapes, low nesting efficiency, and insufficient material utilization in the two-dimensional nesting for safety shoe production, this paper proposes an intelligent nesting algorithm based on an improved No-fit polygon algorithm. First, a tangential vector-based arc contact detection strategy is introduced to effectively solve the overlap detection problem of vamp components with arcs. Then, an incomplete no-fit polygon algorithm is employed to generate the feasible nesting area, reducing nesting time. Finally, a vamp component shape feature-based nesting strategy is developed to enhance material utilization. Experimental results show that the proposed nesting algorithm achieves a maximum material utilization rate of 91.27% and an average material utilization rate of 79.10%, representing an 8.36% improvement over manual nesting. The nesting time for a vamp component ranges from 1.21 to 1.63 s, reducing time by 68.2% compared to manual nesting. The proposed algorithm effectively solves the online nesting problem for complex and irregularly shaped vamp components.

Abstract:To deal with the challenges faced by target detection algorithms due to the small scale, weak features, and high background interference in images of a drone dataset, a multi-scale target detection algorithm for unmanned aerial vehicle images in complex scenarios is proposed. This algorithm enhances the overall accuracy, reduces false negatives and positives, through the incorporation of modules such as DConv, AIFI, and Dyhead. These components address the limitations of the original network in handling multi-scale targets. Furthermore, the use of the DIoU loss function improves the model′s convergence capability. The effectiveness of this approach is demonstrated through its application in detecting multi-scale targets on the VisDrone-DET2019 dataset. Compared to the original network, there is a 3.7% increase in precision, a 1.2% increase in recall rate, and a 2.3% improvement in average accuracy. Moreover, extensive experiments demonstrated that the proposed algorithm exhibits strong robustness and excellent overall performance, suggesting significant industrial application potential.

Abstract:A hybrid meta-heuristic scheduling algorithm is proposed for the effective synergy between lithium battery processing sequencing process and robot multi-station handling scheduling. The algorithm minimises the cycle time as the optimisation objective function, constructs a meta-heuristic mixed-integer linear programming model for multi-station robot handling, and initially optimises the performance of the model by introducing effective constraints on the production cycle time. In order to solve the problem of the sharp increase in computing time caused by the increase in the number of lithium battery types, a hybrid meta-heuristic algorithm fused with genetic algorithm and tabu search is designed, which balances the depth of the search with the efficiency of the computation, and achieves the method of obtaining the approximate optimal solution in a short time. Simulation and application experiments show that compared with the traditional mixed integer linear programming scheduling algorithm, the proposed algorithm can improve the time efficiency of the solution by up to 57.92%, which effectively improves the scheduling efficiency of the multi-station robot handling.

Abstract:Aiming at the problem that the detection accuracy of PCB surface defects is insufficient and it is difficult to balance the detection accuracy and real-time performance of the model, which cannot meet the stable operation requirements of modern electronic manufacturing systems, an improved PCB surface defect detection method named HDH.YOLO is proposed. This method replaces the backbone network of the original YOLO11 with an optimized HGNetV2 to achieve model lightweighting. It also improves the HGBlock by referring to the idea of Dynamicconv, and replaces the last four HGBlocks in the HGNetV2 network with the improved Dynamic_HGBlock. This approach introduces more network parameters without significantly increasing the computational load, thereby enhancing the network′s ability to learn generalized features and improving detection accuracy. In addition, a DSM attention mechanism layer is added at the end of the backbone network to enhance the model′s feature extraction capability by amplifying the spatial and frequency domain responses of key regions. Comparative experiments and ablation studies were conducted on the PKU-Market PCB and DeepPCB datasets. The results show that, compared with the baseline YOLO11n model, the proposed HDH-YOLO model reduces the number of parameters by 6.20% and the computational load by 12.70%, while increasing the mAP50 and mAP50.95 by 2.6% and 2.3%, respectively. HDH.YOLO thus achieves a better balance between lightweighting and detection accuracy and demonstrates high reliability and practicality in modern electronic manufacturing systems.

Abstract:Timely and accurate detection of road potholes in unstructured environments is crucial for ensuring traffic safety. Current detection algorithms face challenges related to missed detections and insufficient accuracy, particularly in complex scenarios. To enhance detection performance, an improved approach based on the YOLOv7 model is proposed. This method incorporates several enhancements: first, an enhanced hierarchical multi-scale fusion module is introduced to optimize feature extraction capabilities; second, the integration of an efficient channel attention mechanism enhances the model′s focus on critical target regions; finally, depthwise separable convolutions are employed to reduce computational complexity while maintaining high detection efficiency. The improved model was validated on the self-made dataset, compared with the existing YOLOv7x, YOLOv7-d6, YOLOv5x and YOLOv5m models, and the improved model was transferred to the public dataset. The evaluation metrics used include precision, recall (R), mean average precision, parameter count, and frames per second. The experimental results show that the improved model improves the precision, recall and average accuracy by 5.47%, 4.42% and 6.65%, respectively, and maintains a high efficiency in the detection speed. Compared with commonly used object detection models, the performance is excellent; after the transfer learning of public datasets, the precision, recall, and average accuracy are worth further improving. This improvement significantly enhances the detection performance and robustness of the model, not only strengthening the ability to ensure traffic safety, but also providing reliable technical support for autonomous driving.

Abstract:In order to solve the problems of low detection accuracy, high computational complexity and high false detection and missed detection rate of the current road defect detection model in complex background, this paper is improved based on the YOLOv8 model. Firstly, the EMA attention mechanism is integrated into the feature extraction network(Backbone) of the model to improve the feature representation ability of the model, while retaining important information and reducing the computational cost .Secondly, the lightweight feature fusion network structure SlimNeck and the weighted feature fusion mechanism Weighted Fusion were combined to form a new neck network structure SWNeck, which effectively reduced the number of model parameters and computational complexity, improved the feature fusion efficiency, and reduced the feature redundancy of noise. Finally, the Slide Loss weight function is introduced to give greater weight to the samples that are difficult to classify correctly, improve the learning ability of the model for difficult sample data in road defects, and further enhance the detection performance of the model.The experimental results show that the improved road defect detection model improved mAP by 2.7% compared to the original YOLOv8n model, and the amount of parameters and computational complexity of the model were reduced by 7% and 10%, respectively.

Abstract:Indoor positioning technology, especially the Received Signal Strength Index(RSSI)-based fingerprinting positioning method, has received extensive attention due to its low cost, wide device support, easy deployment, and low computational overhead. In order to enhance the mapping relationship between RSSI and the actual physical distance and improve the ranging accuracy, this paper proposes an RSSI ranging algorithm based on Improved Grey Wolf Optimization(IGWO) algorithm and Back Propagation Neural Network(BPNN). Compared with Genetic Algorithm(GA), Particle Swarm Optimization(PSO) and classical Grey Wolf Optimization algorithm(GWO), the improved GWO algorithm has significant advantages in positioning accuracy and global search ability. Through experiments, the root mean square error(RMSE) of IGWO algorithm is reduced by 21.3%, 15.7%and 14.6%respectively compared with GWO algorithm, GA algorithm and PSO algorithm. IGWO algorithm shows better positioning performance, and is superior to the traditional methods in accuracy and performance.

Abstract:To address the challenges of limited fabric defect categories, significant scale variations, and low detection accuracy in existing models, this study introduces DHR-DETR, a fabric defect detection method based on RT-DETR. Firstly, a Multi-Path Coordinate Attention(MPCA) module is innovatively designed and deeply integrated with the Deformable Convolution Module(DCNv2) to construct a Dynamic Deformable Convolution Module. This integration effectively accommodates the diverse and complex shapes of fabric defects. Secondly, a High-Level Screening Feature Pyramid Network(HS-FPN) is employed to replace the Cross-Scale Feature Fusion Module(CCFM), enabling efficient multi-level feature fusion while significantly reducing model complexity. Finally, a lightweight yet feature-enhancing RetBlockC3 module is developed and incorporated into the HS-FPN network. This module enhances the model′s capability to capture local information and further improves its lightweight design.Experimental evaluations demonstrate that the proposed DHR-DETR method achieves mAP@0.5 scores of 50.9% and 97.5% on public and custom fabric datasets, respectively, reflecting improvements of 2.9% and 0.6% compared to the baseline model. Additionally, the parameter count is reduced to just 17.9 M, with a 37% decrease in computational complexity. These results indicate substantial improvements in detection performance and deployment efficiency, showcasing the potential of DHR-DETR for practical applications in industrial fabric inspection tasks.

Abstract:To address the accuracy and robustness issues of small floating object detection on water surfaces under complex scenarios such as wave disturbances, changes in lighting, and partial occlusion by floating debris, the MEC-YOLOv11n algorithm is proposed. The MEC-YOLOv11n algorithm consists of three parts: Backbone, Neck and Head. To increase the recognition area of the target receptive field, we designed the MSWTC structure and improved the C3k2 structure in the Neck part, which significantly enhances the extraction ability of small floating objects on water surfaces, thus strengthens the model′s ability to capture details in complex backgrounds; next, we proposed a EUCB up-sampling method, replacing the traditional up-sampling module in v11, which enhances the clarity of image edges during up-sampling, making the object contours more accurate in high-resolution feature maps, especially when dealing with complex backgrounds and small target detection tasks, which significantly improves the model′s ability to capture details; finally, we designed an attention module CCA specifically for recognizing edge features before the Head, further optimizing the model′s performance in edge information extraction. Experimental results show that after optimization, the precision P of the model has increased by 3.3%, the recall R has increased by 2.4%, the mAP50 has increased by 2.5%, and the mAP50.95 has increased by 1.5%.