Abstract:With the accelerated transformation of aerospace equipment toward lightweight and intelligent design, the traditional passive maintenance model based on periodic inspections can no longer meet the urgent demands for high reliability and real-time safety assessment. During long-term service, key aircraft structures such as the fuselage and wings are subjected to complex coupled effects, including multiaxial cyclic loads, fatigue accumulation, environmental corrosion, and dynamic impact. These factors can induce the progressive expansion of hidden damage, highlighting the urgent need for online load monitoring technologies to enable accurate characterization and spatial localization of damage evolution mechanisms. FBG sensing technology, with its core advantages of strong electromagnetic immunity, high sensitivity, and long-term stability, has emerged as a promising alternative to overcome the limitations of traditional electrical sensors. It is steadily progressing from laboratory research toward practical engineering applications. However, the deployment of FBG-based systems still faces several technical challenges, such as cross-sensitivity, miniaturization of demodulation units, and precision integration, which hinder their widespread adoption in aircraft structural load monitoring. This study focuses on the research and application of FBG technology in aircraft structural load monitoring. It systematically reviews the background and current development of load monitoring requirements, summarizes representative research progress and typical application cases both domestically and internationally, and discusses key technical issues and future development trends of FBG-based structural load monitoring in aviation.

Abstract:Indoor positioning plays a crucial role in enabling location-based services in sensor networks that cannot be reached by GNSS. In the wireless positioning system, the wireless fingerprint-based positioning method only needs to compare the signal of the device to be located with the known features to determine the location, which is widely used in indoor scenes because of its low complexity. However, due to the fading and multipath effects caused by the complex and changeable indoor environment, which will lead to the fluctuation of indoor signal values, thereby reducing the positioning accuracy, most of the current methods ignore the temporal and spatial information of fingerprint collection, in order to solve these problems, this paper proposes an algorithm model combining deep residual network (ResNet) and indoor fingerprint positioning. In order to solve the problem of random fluctuation of indoor signals, the particle filter can better adapt to the changes of the dynamic environment, and the self-attention mechanism algorithm is used to dynamically adjust the particle weight, so that the algorithm model proposed in this paper can better capture the signal features in the room, so as to improve the positioning accuracy and robustness. Finally, the corresponding experimental verification is carried out, and the experimental results show that the average positioning error of the SA-ResNet indoor positioning algorithm model fluctuates between 0.56 and 0.62 m, which has good stability.

Abstract:Path planning is a key technology for unmanned vehicles to realize autonomous navigation. Whether a safe and smooth travelable path can be quickly planned in a narrow channel determines the efficiency of unmanned vehicles in performing tasks in narrow and complex environments. However, common path planning algorithms usually have the problems of slow convergence speed, long planning time and poor path quality in the narrow channel environment. For this reason, this paper proposes a RRT-Connect algorithm Based on dual-layer guided sampling (DLGS-RRT-Connect) algorithm. First, the guided path is pre-constructed in the narrow channel, and the searching connection strategy is used to guide the random tree to expand along the guided path in the narrow channel, so as to reduce the invalid sampling and improve the exploration efficiency of the algorithm in the narrow channel. Secondly, the algorithm introduces a target bias strategy to reduce the randomness in the sampling process and provide directional guidance for the growth of the random tree, thus further improving the efficiency of path planning. Finally, the simulation results show that compared with the common Goal_bias RRT, Informed-RRT*, and RRT-Connect algorithms, the DLGS-RRT-Connect algorithm proposed in this paper improves the planning success rate in narrow channel environments by 35%, 60%, and 26%, respectively, and reduces the average planning time by 70.62%, 70.62%, 70.65%, and 97.65%, and 63.92%, and the average path length is also reduced by 14.53%, 16.70%, and 18.84%, respectively, which can effectively improve the smoothness and safety of planning paths in narrow environments.

Abstract:To address the issues of local optima and slow convergence speed encountered by traditional genetic algorithms, several improvements have been made to the genetic algorithm in this paper. Firstly, the conventional 8-direction search is extended to a 24-neighborhood, 16-direction search to enhance the global search ability. Secondly, a Piecewise and Tent (PT) chaotic mapping fusion strategy is introduced, where the sequence generated by the Piecewise chaotic mapping is used as the parameter for the Tent chaotic mapping to improve population diversity. Furthermore, the Levy flight strategy is integrated to avoid local stagnation, and a new strategy for handling out-of-bounds particles is proposed to prevent the initialization population from exceeding boundaries. A novel pairing exchange and differential perturbation mechanism is then designed to prevent the loss of good individuals, which may lead to the algorithm getting stuck in local optima. Lastly, a new pressure level splitting selection mechanism and an adaptive crossover and mutation probability adjustment operator are proposed. Coefficients are adjusted to resolve the issue of excessive selection pressure, and a nonlinear exponential function is used to adjust the crossover selection probability to avoid early divergence. Additionally, complementary adjustments to mutation probabilities are introduced to expand the search space and reduce convergence oscillations. Experimental results show that the proposed method significantly improves path planning performance compared to traditional genetic algorithms and other improved algorithms, with path lengths reduced by 5.13% and 2.06%, respectively. The superiority and practicality of the method in robot path planning are validated.

Abstract:A method is proposed for multi-robot path planning in complex environments, employing an improved rapidly-exploring random tree (IRRT) algorithm and predicted-improved artificial potential field (P-IAPF) algorithm to achieve obstacle avoidance in multi-robot systems. Firstly, in view of the shortcomings of slow convergence speed and random search range of RRT algorithm, the target-biased strategy is used to guide the generation of random sampling points, simultaneously, the improved artificial potential field method is integrated into the bidirectional random search tree to rapidly identify the global path. Secondly, in response to the problem of traditional APF algorithm being prone to getting stuck in local minima and having low path planning efficiency, a predicted APF algorithm with multiple virtual keypoints is proposed, the Douglas Peucker (DP) algorithm is used to find the sequence of sub keypoints in the planned global path, and the multi robot system switches keypoints to escape from local minima, thereby enhancing both the efficiency and smoothness of multi-robot path planning. Ultimately, to confirm the effectiveness of the proposed algorithm, simulation experiments in complex environments with U-shaped and long rectangular obstacles are carried out, and it has the advantages of high path planning efficiency and avoiding multi-robot collision.

Abstract:Aiming to address the issue of suboptimal positioning accuracy in non-ranging node localization algorithms for wireless sensor networks, particularly in the context of multi-hop distance and average hop distance estimation methods that are limited in their capacity to rectify errors, resulting in the propagation of computational errors and consequent reduction in positioning accuracy, an improved black-winged kite algorithm-3D distance cetor-hop (IBKA-3DDV-Hop) localization algorithm is proposed. First, to reduce the hop quantization error, the number of hops between nodes is refined by using the multi-communication radius, and then the hop distance correction factor is introduced to compensate for the error of hop distance. Secondly, the optimal latin hypercube mechanism (OLHS) is employed to optimize the population initialization in the improved black-winged kite algorithm. This approach overcomes the limitations of random initialization and generates a reverse population through the Elite Reverse Learning strategy, which further enhances the quality of the initial population. In conclusion, the Levy flight strategy is integrated into the migration behavior of BKA. This integration serves to optimize and enhance the algorithm’s global search capability, thereby preventing the algorithm from attaining a local optimum. The simulation results demonstrate that, in comparison with the conventional 3DDV-Hop algorithm, multi-communication radius algorithm, GOOSE-3DDV-Hop algorithm, and WOA-3DDV-Hop algorithm, the proposed IBKA-3DDV-Hop localization algorithm reduces the normalized localization error by approximately 22%, 17%, 11%, and 6%, respectively. This improvement effectively enhances the accuracy of the non-ranging node localization algorithm.

Abstract:To address the limitations of traditional trajectory planning algorithms in confined spaces, such as low planning success rates, poor adaptability, and deviations from human driving habits, this paper proposes an adaptive two-stage trajectory planning algorithm integrated with virtual obstacle decision-making. The first stage combines dynamic programming and quadratic programming to achieve path planning and velocity optimization for autonomous vehicles. Subsequently, an adaptive aggregation sampling strategy is introduced to resolve navigation challenges in narrow environments. Finally, a random forest-based virtual obstacle decision model is developed to enhance decision-making rationality under diverse vehicle interaction scenarios. The results on the simulation platform Carla show that, compared with the traditional method, the path length and path curvature of the proposed method are reduced by 2.4% and 85.6% respectively, and the success rate of planning, safety and stability are improved by about 20%, 20.6% and 44.9% respectively in the static multi-obstacle avoidance in narrow areas. In the dynamic multi-obstacle avoidance in narrow areas, the path length and path curvature are reduced by 8.3% and 76.4%, respectively, and the planning success rate, safety and stability are improved by about 36%, 78.2% and 45.3%, respectively. Finally, the method was deployed to the actual unmanned vehicle, and the obstacles were set up in the narrow and long corridor scene for testing, which verified the effectiveness of the method.

Abstract:To address the issues of complex mapping processes, difficult parameter tuning, and poor generalization in traditional simultaneous localization and mapping (SLAM) algorithms, this paper proposes a wheeled robot experience map construction method based on spatial multi-scale continuity feature extraction. First, PSA and ASPP modules are integrated into the ResNet18 architecture. PSA groups intermediate features and calculates attention weights across channels to capture multi-scale information, thereby enhancing feature representation. ASPP incorporates dilated convolutions with varying dilation rates and global average pooling to aggregate global contextual information, further strengthening the representation of spatial multi-scale continuity features. Then, the improved ResNet-PSA-ASPP model is trained on datasets collected in both the donkey_sim simulator and real-world robot racetrack scenarios. Finally, model performance is evaluated in both simulated and real-world environments using the donkey_sim simulator and the robot operating system （ROS）. Experimental results show that the proposed model reduces steering angle prediction errors by 38.47%, 44.34%, and 35.51%, respectively, and significantly outperforms classical networks such as ResNet18, ResNet50, and VGGNet in feature extraction capability, computational efficiency, and mapping accuracy.

Abstract:The grassland wind power station is characterized by high wind force levels, undulating terrain, and uneven ground surfaces. When ground inspection robots perform patrol tasks under different wind force conditions, it is challenging to balance path indicators with safety, thus posing higher demands on path planning methods. This paper proposed an A* path planning algorithm enhanced with terrain factors（A* algorithm (TF-A*）. Initially, this paper designed a gradient factor and a step factor based on the undulating terrain with high and low bumps and the terrain with pits and steep slopes of the grassland wind power station, and applied these factors to optimize the cost function and the heuristic function. Building upon these enhancements, the paper successfully developed an A* path planning algorithm enhanced with terrain factors. Subsequently, this paper meticulously tailored the parameters for both the slope factor and the step factor to accommodate varying wind force conditions. By taking into account the significant wind force levels prevalent in the grassland wind power station, which has substantially improved the safety and stability of the inspection robots. Following that, a series of experimental evaluations were meticulously executed, including short-distance and long-distance simulation tests, as well as real-world field experiments, to assess the efficacy of TF-A* path planning algorithm. The experimental outcomes have revealed that the TF-A* path planning method has significantly surpassed the traditional A* algorithm, with a notable 44.55% and 34.82% increase in path length metrics, and a substantial 22.06% and 2316% reduction in search time metrics across two distinct weather conditions. Specifically, under conditions of low wind force, the method strategically prioritizes distance metrics, where-as under high wind force, it adeptly integrates both distance metrics and operational safety into its considerations. It provides a novel approach for robot inspection and path planning in unstructured and uneven terrains of grassland wind power station.

Abstract:High-quality global path planning is one of the key technologies enabling autonomous navigation of unmanned surface vehicles (USVs). To address the global path planning problem for USVs in complex obstacle environments, this paper proposes a global path planning method based on the multi-strategy enhanced slime mould algorithm (ME-SMA). To overcome SMA’s limitations such as uneven initial population distribution, slow convergence speed, and proneness to local optima, ME-SMA employs several enhancements: it optimizes population initialization using improved Logistic chaotic mapping to enhance global exploration; incorporates crossover, mutation, and selection strategies from genetic algorithms to improve local exploitation efficiency; and introduces the golden sine strategy to dynamically adjust the search direction, thereby avoiding premature convergence. To validate the effectiveness of ME-SMA, we tested it on nine types of benchmark functions. The results show that ME-SMA achieves superior convergence accuracy and stability compared to the original SMA and other comparative algorithms. Simulation experiments in identical complex obstacle environments further demonstrate that ME-SMA significantly improves convergence speed, task completion time, and navigation distance. Compared to the other experimental algorithms, ME-SMA achieves an average reduction of 1.8% in path length and an average improvement of 28.22% in stability, highlighting its high efficiency and practical engineering value for USV global path planning applications.

Abstract:Aiming at the problems of long path length and numerous turning points of jump point search (JPS) algorithm and path tortuosity with low pathfinding efficiency of path finding caused by artificial potential field (APF) falling into U-shaped trap in complex U-shaped obstacle environment. this paper proposes a mobile robot path planning algorithm that integrates an improved JPS algorithm and APF algorithm (JPS*-APF). Firstly, an angle deviation function is introduced into traditional JPS algorithm, and redundant nodes are removed to reduce search distance and turning frequency. Secondly, turning points from improved JPS algorithm are used as sub-goals, guiding APF algorithm in segments to escape U-shaped trap. Adaptive generation of repulsive forces for corner obstacles or dynamic sub-goals enhances path smoothness. Then, symmetric virtual obstacles are added to the target area to resolve target inaccessibility, while external repulsive forces and re-planning strategies are fused to escape local optima and improve pathfinding efficiency. Finally, relative velocity repulsive forces are introduced to ensure safety during dynamic obstacle avoidance. Numerical simulations in different U/L-shaped obstacle environments demonstrate that JPS*-APF algorithm reduces pathfinding time by an average of 51.5% and path length by 7.3% compared to IA*-APF algorithm. Moreover, JPS*-APF algorithm generates smoother paths, effectively escapes U-shaped traps, and enhances mobile robot’s working efficiency. The feasibility of JPS*-APF algorithm is also validated through real-world experimental tests.

Abstract:To address the issues of poor search accuracy, slow convergence speed and easy fallback to local optima in solving unmanned aerial vehicle (UAV) path coordination planning problems using the hyperbolic sine-cosine optimization （SCHO） algorithm, a segment-guided and dynamical partitioned improved hyperbolic sine-cosine optimization （SDSCHO） algorithm is proposed. A three-dimensional geographic model of UAV flight and threat conditions is established, and a path coordination planning cost model is constructed that integrates path length, obstacle threat, flight altitude and turning angle. And SCHO algorithm is comprehensively improved by introducing chaos Circle mapping for population initialization, nonlinear oscillation conversion factor, segment-guided and reverse escape optimization, and dynamic boundary partitioning-assisted position update strategy. The improved algorithm SDSCHO is used to solve the UAV path coordination planning problem. On multiple benchmark functions with different characteristics, the optimizing tests are carried out with seven similar algorithms. The results prove that SDSCHO performs better in optimization accuracy and convergence performance. Finally, by building a three-dimensional mountain model with different obstacles, SDSCHO is applied to solve UAV single-path and multi-path coordination planning scenarios, which can further confirm the superiority of our algorithm in handling actual optimization problems.

Abstract:Railway catenary system is the core equipment to supply power to electric traction vehicles, and its state directly affects the safety of train operation. In order to solve the problem that the key small components (split pin, tube sleeve and nut of positioner bracket) in the railway catenary are difficult to accurately locate due to their small size and complex environment, a small target detection method based on improved YOLOv5s is proposed. Firstly, the C3 module of the feature extraction network is combined with the linear deformable convolution （LDConv） to design a new C3_LD module. The proposed module employs deformable convolution kernels to dynamically adjust receptive fields, which effectively captures geometric deformation characteristics of small targets. This design not only enhances feature extraction capability but also reduces parameter. Secondly, the SPPFCSPC_group structure is designed to replace the original SPPF structure, and the multi-scale feature expression ability of the network is improved by combining group convolution with multi-scale spatial pyramid. Finally, the original loss function is replaced by spatial intersection over union (SIoU), which enhances bounding box regression accuracy through spatial constraints between predicted and ground-truth boxes. The results of ablation experiments and comparison experiments show that the improved algorithm in this paper achieves 93.2% mean average precision (mAP) and 93.1% recall rate in the detection task of railway catenary mesh small components, which are 1.9% and 3.6% higher than those of the original algorithm, which effectively alleviates the missed detection and false detection problems of railway catenary small components.

Abstract:In recent years, transformer-based methods have achieved remarkable progress in the field of 3D human pose estimation. However, current approaches are still confronted with two major challenges. First, the computational inefficiency arises from the quadratic complexity of global self-attention when processing large-scale joint affinity matrices in dynamic video sequences. This issue significantly hampers the real-time performance of the models. Second, the suboptimal spatiotemporal feature fusion restricts the model’s ability to capture fine-grained motion patterns and structural dependencies between joints, leading to less accurate pose estimation results. To tackle these limitations, this paper proposes a novel architecture named the dual-stage spatio-temporal convolutional transformer (DSTCFormer). The key innovation of DSTCFormer lies in its decoupling of spatiotemporal feature learning into parallel spatial and temporal pathways. Specifically, the convolutional multi-scale attention (CMSA) module is introduced to hierarchically aggregate local and global correlations through convolution-enhanced multi-head attention. In the spatial pathway, convolutional position embeddings are utilized to encode skeletal topology, enabling the model to focus on intra-frame joint relationships. Meanwhile, the temporal pathway captures inter-frame motion coherence via axial-specific self-attention. Moreover, a cross-stage fusion mechanism is designed to integrate multi-scale spatiotemporal features through depthwise separable convolutions and feature transformation layers, which ensures efficient computation and robust feature representation. Extensive experiments conducted on the Human3.6M and MPI-INF-3DHP datasets demonstrate the superiority of DSTCFormer. Under Protocol 1 (P1), DSTCFormer achieves a state-of-the-art Mean Per Joint Position Error (MPJPE) of 40.1mm on Human3.6M with 243 input frames, outperforming PoseFormer (44.3 mm), MixSTE (40.9 mm), and STCFormer (40.5 mm). On the MPI-INF-3DHP dataset, it attains a percentage of correct keypoints at 150 mm (PCK@150 mm) of 99.1% and an area under curve (AUC) of 85.2%, surpassing existing methods by 0.4% and 1.3%, respectively. In summary, the proposed method not only advances the theoretical frameworks for spatiotemporal modeling but also offers practical implications for real-time applications, paving the way for more efficient and accurate 3D human pose estimation in various scenarios.

Abstract:Data-driven soft-sensing modeling plays a critical role in process industries, yet faces challenges from heterogeneous noise contamination and the coexistence of linear and nonlinear correlations in industrial datasets. These issues significantly compromise model prediction accuracy. To address this, we propose a fused correlation-based collaborative shared noise algorithm for robust soft-sensing modeling. The algorithm integrates Pearson correlation coefficients (linear relationships) and Spearman rank correlation coefficients (nonlinear relationships) to compute data credibility, thereby optimizing noise allocation under mixed correlation conditions. A convolutional neural network (CNN) is subsequently employed to construct the soft-sensing model. Experiments on a debutanizer column dataset demonstrate the superiority of the proposed method. The FC-CSNA outperforms baseline denoising techniques, including wavelet transform, denoising autoencoders, and the original collaborative shared noise algorithm, in noise suppression. The hybrid model achieves state-of-the-art prediction performance, with an R2score of 0.971 6 and mean squared error (MSE) of 0.001 1, validating its effectiveness in handling industrial data complexity.

Abstract:To address the flaws of existing low-light image enhancement methods based on Retinex theory, such as complex training procedures, difficulties in acquiring the ground truths of illumination and reflection components during training, and issues affecting image quality by amplifying dark-region noise and losing structural details when enhancing images under extremely poor lighting conditions, this paper proposes an end-to-end two-stage image enhancement network that combines Retinex theory with diffusion models. In the first stage, guided by Retinex theory, the focus is on improving the brightness of low-light images. A convolutional neural network (CNN) is adopted to estimate the three-channel illumination ratio map, which is then dot-multiplied with the low-light image to obtain the initial enhanced result. Pure Retinex methods barely consider the degradations hidden in dark areas during brightness enhancement. After initially brightening the low-light image, the second stage focuses on denoising and restoring the image using the excellent denoising capability of diffusion models. A brightness-aware diffusion model is proposed, which takes the luminance map in the HSI color space as a condition to fully leverage the advantages of diffusion models in repairing degradations from the initial enhancement. A color correction module is also introduced to mitigate potential global degradation during the inverse process of the diffusion model, yielding the final enhanced image. Experimental results show that compared with 10 other state-of-the-art algorithms on low-light datasets, the proposed method achieves a peak signal-to-noise ratio (PSNR) of 27.517 and a structural similarity index (SSIM) of 0.874 (a near-optimal value), along with an image perception similarity of 0.087-all outperforming the compared methods. The proposed method can well adapt to the distributions of unknown noise and illumination, achieving excellent performance in brightness enhancement, noise removal, and detail preservation, and generating more natural and high-quality enhanced images.

Abstract:Coronary artery disease (CAD), as one of typical heart diseases, threatens people’s lives and health. However, due to its complex influencing factors and its subtle initial symptoms, many patients miss the optimal treatment window for recovery. To enable early diereses prevention so as to get most appropriate treatment, many machine learning methods have been widely applied in this field, among which deep learning has been acknowledged as one of the cutting-edge techniques for CAD diagnosis. This paper develops a tailored BP neural network, which integrates BP with an attention-residual-mechanism for CAD detection. In order to find the key factors that contribute to CAD prediction, we fist investigate a feature selection strategy based on data visualization and using several statistical methods on the commonly used cleveland heart disease dataset. Then, the attention-residual-mechanism informed BP network is conducted for CAD detection. The amended BP network alleviates the gradient vanishing problem by using a residual structure and captures deep dependencies between features through a multi-head attention mechanism, which can be used for dynamic allocation of feature weights. Extensive experiments demonstrate the better performance of our method than existing machine learning algorithms. It can achieve an accuracy of 97.1% on Cleveland Heart Disease dataset, which verifies the effectiveness of our method in CAD diagnosis.

Abstract:The dual-motor drive pitch system is a strong coupling nonlinear time-varying system. The parameters of the two servo motors will also change during operation, resulting in inaccurate mechanism model of the system and affecting the synchronous control accuracy of the two motors. In this paper, a data-driven model based on improved sparrow search algorithm to optimize hybrid kernel extreme learning machine (CGSSA-HKELM) and a dual-motor model predictive synchronous control system based on quantum genetic algorithm (QGA) to solve the objective function are proposed. Firstly, the kernel extreme learning machine regression principle is used to establish a unified prediction model for two motors, which improves the accuracy, generalization ability and learning speed of the prediction model. Secondly, aiming at the problem that the kernel extreme learning machine is sensitive to parameter settings, the improved sparrow search algorithm is used to optimize its model parameters and conduct offline training to obtain a prediction model with adaptive ability. In the constructed model predictive synchronous control system, quantum genetic algorithm is introduced to optimize the objective function, so as to avoid falling into local optimum and obtain the optimal control of two motors. Finally, in order to prove the effectiveness of the scheme, simulation and experimental verification are carried out. The results show that the torque error of the two motors is reduced by 45% and the torque ripple is reduced by 40% compared with the cross-coupled sliding mode control strategy. The simulation and experimental results effectively prove the rationality and effectiveness of the dual-motor model predictive synchronous control scheme designed in this paper.

Abstract:Aircraft honeycomb sandwich composite panel has the characteristics of complex structure, large size and many types of defects, and the conventional ultrasonic nondestructive testing technology has the problems of low efficiency and poor accuracy. In the study, nonlinear ultrasonic Lamb wave imaging technology for honeycomb sandwich composites is proposed, which combines the advantages of large range propagation of Lamb waves and strong sensitivity of nonlinear response of defects to improve the non-destructive testing ability of honeycomb sandwich composites. Firstly, the propagation characteristics of Lamb waves in the panel are analyzed, and the probe arrangement method is designed and the defect factor is defined based on the anisotropic characteristics of wave propagation. Secondly, the imaging algorithm is designed and optimized to improve the nonlinear Lamb wave imaging effect of the defects in the honeycomb sandwich composite panel. Finally, the ability of nonlinear Lamb wave detection image to display the defects in the panel is analyzed by referring to the air-coupled C-scan and metallographic observation. The results indicate that the nonlinear ultrasonic Lamb wave image has poor resolution compared with the air-coupled ultrasonic C-scan image, which cannot show the honeycomb details. Nevertheless, the debonding and weak bonding defect areas of honeycomb sandwich composite panel can be displayed by the nonlinear ultrasonic Lamb wave testing, and it also possesses the advantages of high efficiency and same-side transceiver probe arrangement which improves the adaptability to the detection environment. Thus, it can be used for the non-destructive detection of bonding defects of honeycomb sandwich composite panels in service.

Abstract:Five approximate multiplier design methods are proposed to address the issues of incomplete models, high on-chip resource consumption, and limited performance of Field Programmable Gate Array (FPGA) in accelerating convolutional neural networks, image processing algorithms, and other approximate computing fields. Based on an 8-bit×8-bit unsigned carry chain approximation multiplier, two LUT based 8-bit×8-bit unsigned approximation multipliers are proposed for different real-world scenarios with a lookup table (LUT) to optimizing the critical path simplification structure by compressed recursive invocation methodology and sub-product recombination computation strategy. This method can save up to 60% of area, about 60.76% of power consumption, and about 25.4% of critical path delay (CPD) compared to similar multipliers within an acceptable range of accuracy. At the same time, in order to meet the needs of more complex scenarios, two 16 bit×16 bit unsigned approximate multipliers with LUT are proposed by doubling the number of multiplies digits. Compared with similar multipliers, the method can save up to about 41.2% of the area, about 77% of the power consumption, about 35.4% of the CPD, which can compensate for the loss caused by the decrease in accuracy. In addition, based on the signed number calculation module, proposed a 16 bit × 16 bit signed approximate multiplier with LUT is proposed to replace Xilinx’s (now ADM) Multiplier IP core, which is deployed in the convolutional neural network convolutional layer with handwritten number recognition function and tested using handwritten number images in the MNIST dataset. It saves about 32.48% of area, about 41.21% of power consumption, and about 24.28% of CPD, at the cost of a 3.4% decrease in accuracy. It is shown that these multipliers can effectively meet the requirements of FPGA accelerated convolutional neural networks and achieve the optimal balance between accuracy and resource overhead.

Abstract:In order to improve the efficiency of multi-class container damage detection in complex yard environment, a lightweight container damage detection method based on improved RETINEX is proposed. The method mainly consists of two parts, image preprocessing and lightweight target detection: In image preprocessing stage, the luminance channel component is introduced and optimized, and when applying the multi-scale Retinex processing method to it, a bilateral filter is used instead of the traditional Gaussian filter to retain the edge details of the original object; the value domain conversion function is improved to reduce the loss of image data; and a color protection is obtained through the calculation of the color balancing strategy factor, which is multiplied with the pixel points of each channel of the original RGB image to get the enhanced image. In target detection stage, MobileNetv3, a lightweight network with improved attention mechanism, is introduced into the YOLOv5 backbone network to construct a text-based target detection network, so as to validate the port container images. The experimental results show that the method helps the target detection network to extract richer feature information in complex port environments such as low illumination, and the average detection accuracy of multiple container damage types is improved by 1.4% to 95.1%, and the model volume is only 20.5 MB, which is a significant advantage compared with multiple mainstream detection algorithms, and it can satisfy the actual detection needs of port containers, proving the effectiveness of the method in this paper.

Abstract:A gearbox is a kind of mechanical transmission device. Aiming at the problem of poor state recognition effect caused by the nonlinearity and instability of the complex fault signal of the gearbox, a gearbox complex fault diagnosis method based on multi-scale Weibull dispersion entropy graph neural network (WB-MDEGNN) is proposed. Firstly, the Weibull distribution (WB) is used to linearize and stabilize the vibration signal to obtain more acute gearbox state information. Then, the Multi-scale dispersion entropy (MDE) is used to extract the quantization features of the given sequence. And construct the node feature matrix. Secondly, use the k-nearest neighbor (KNN) algorithm to extract the correlation of node features and construct the edge index matrix. Combine the node feature matrix with the edge index matrix to construct the feature map. Finally, the feature maps are input into the graph neural networks (GNN) model for classification and recognition. The results show that by collecting gearbox data in five states through piezoelectric acceleration sensors and using the WB-MDEGNN model proposed in this paper for complex fault classification and identification of the collected data, the accuracy rate can be increased by 6.07%~11.69% compared with other existing gearbox fault diagnosis methods. Meanwhile, the accuracy and generalization of the model proposed in this paper are tested by adding Gaussian white noise with different signal-to-noise ratios to the original data and public datasets. The complex fault diagnosis performance of the proposed method, the accuracy difference fluctuation range is between 0.97% and 3.38%, and the generalization test can reach 95%. Therefore, this method has better superiority in dealing with the problem of poor state recognition effect caused by signal nonlinearity and instability, providing a new method for the complex fault diagnosis of gearboxes.

Abstract:Grounding grid, as an important equipment to ensure the safety of power system, the research on its corrosion state detection is of great significance. Electrical impedance tomography is one of the important methods for grounding grid corrosion imaging. Due to its pathological nature when solving the inverse problem, the reconstruction effect has a large deviation. In order to improve its imaging quality and accuracy, this paper proposes a total variation- conditional generative adversarial network (TV-CGAN) algorithm to detect its corrosion state. First, the grounding grid forward problem model is established to solve the boundary voltage, and then the total variation (TV) regularization algorithm is used to solve the inverse problem to obtain a preliminary grounding grid conductivity distribution image. Then, the conditional generative adversarial network algorithm is used to perform secondary imaging on the image obtained by the TV method. The generator is a U-Net structure that introduces the convolutional block attention module. The discriminator is a PatchGAN convolutional structure. This method was applied to the detection of grounding grid corrosion status. The reconstructed image structure similarity result was 0.907 8, the peak signal-to-noise ratio was 16.935 6, the corrosion position judgment accuracy was 96.35%, and the corrosion degree judgment error was 8.61%. The results show that this method effectively improves the ill-conditioned problem in solving the inverse problem, improves the quality of grounding grid corrosion imaging, and improves the accuracy of grounding grid corrosion detection.

Abstract:To more accurately and efficiently detect series arc faults under different loads and address the difficulty of setting a unified threshold for different loads when using a single feature, a multi-feature fault arc recognition method based on t-distributed stochastic neighbor embedding (t-SNE) and simulated annealing particle swarm optimization algorithm-optimized BP neural network (SAPSO-BP) is proposed. Firstly, considering the rich high-frequency components in arc fault currents, the traditional coefficient of variation (CV) feature is improved by extracting the CV of current frequency. The improved CV achieves an average recognition accuracy of 96% across different loads. Secondly, wavelet packet detail components and energy entropy, which are time-frequency domain features, are further extracted and fused with CV for the identification of arc faults. During the fusion process, various nonlinear dimensionality reduction algorithms are used, and clustering visualization comparisons are carried out. It is found that reducing the multidimensional features to a three-dimensional space using t-SNE dimensionality reduction provides the highest distinction for fault arcs. Finally, the reduced features are input into SAPSO-BP for training, and ablation experiments are designed to verify the recognition performance and robustness of the proposed method. The results show that the recognition performance of the fusion algorithm tSNE-SAPSO-BP is improved by 3.2%, 16.8%, 27.66%, and 33.5% respectively compared with the recognition accuracy of single features on different loads. t-SNE dimensionality reduction and clustering effectively deal with the nonlinear correlations between features, providing key feature information for the identification of fault arcs by the fusion machine learning method.

Abstract:Aiming at the issue of arc fault detection and localization in photovoltaic arrays, particularly in large-scale photovoltaic systems, this study constructs equivalent circuit models of arc faults under diverse operational scenarios to analyze voltage and current characteristic differences induced by arc faults of distinct locations and varying types. Based on the typical fault characteristics observed in Photovoltaic string-side operations, this study establishes a detection criterion integrating characteristic frequency band energy ratios and differences between the mean values. Furthermore, leveraging the distinct fault signatures of busbar-side faults, a detection criterion combining voltage mean values and differences from the means is established. Finally, by synergistically utilizing multivariate information from both string currents and busbar voltages, an IoT-based strategy for direct current arc fault detection and localization in photovoltaic systems is proposed. The method proposed in this paper not only achieves fault detection but also accurately identifies fault types and determines the fault segments. This method exhibits good anti-interference capabilities, effectively distinguishing between non-fault conditions such as shadow occlusion. Experimental results demonstrate that this method significantly outperforms traditional single-feature detection methods. It represents a novel approach to implementing arc fault detection through Internet of Things technologies.