Abstract:Pruning is an effective approach to reduce weight and computation of convolutional neural network, which provides a solution for the efficient implementation of CNN. However, the irregular distribution of weight in the pruned sparse CNN also makes the workloads among the hardware computing units different, which reduces the computing efficiency of the hardware. In this paper, a fine-grained CNN model pruning method is proposed, which divides the overall weight into several local weight groups according to the architecture of the hardware accelerator. Then each group of local weights is pruned independently respectively, and the sparse CNN obtained is workload-balancing on the accelerator. Furthermore, a sparse CNN accelerator with efficient PE and configurable sparsity is designed and implemented on FPGA. The efficient PE improves the throughput of the multiplier, and the configurability makes it flexible to compute CNN with different sparsity. Experimental results show that the presented pruning algorithm can reduce the weight parameters of CNN by 70% and the accuracy loss is less than 3%. Compared to dense accelerator research, the accelerator proposed in this paper achieves up to 3.65x speedup. The accelerator improves the hardware efficiency by 28~167% compared with other sparse accelerators.

Abstract:The existing defect detection methods for photovoltaic panels mostly use methods such as electroluminescence excitation and area array camera, which have problems such as complex operation and low efficiency. Therefore, research on photoluminescence imaging system for photovoltaic panels based on linear array InGaAs camera is carried out. Firstly, design the hardware and logic framework of the linear InGaAs camera, and use FPGA to drive the linear InGaAs camera to complete data acquisition and image display. By using fixed mode noise removal algorithm and histogram bidirectional equalization algorithm, fixed mode noise in defect images was removed, while improving the contrast and clarity of the images. Finally, by building an overall imaging system and conducting imaging experiments on various defects of different types of photovoltaic panels through photoluminescence imaging, with a detection accuracy of 0.2 mm/pixel. The experimental results show that the system can detect defects such as hidden cracks, black spots, damaged chips, mixed grades, and dirt in monocrystalline silicon and polycrystalline silicon photovoltaic panels.

Abstract:In view of the high-resolution requirements of singlephoton counters for high-speed flight photon time measurement, the traditional time-to-digital converter TDC has the disadvantage of large errors in time measurement. This paper designs a TDC that uses the internal logic delay unit Carry4 of the FPGA to cascade to build a delay chain. The method firstly uses the method of sub chain average to sample the data and avoid the data "bubble". Secondly, the width of each delay cell is calibrated to nearly uniform width by combining the code density test and bin-by-bin calibration to improve the measurement accuracy of the system. Finally, Vivado software was simulated and burned to ZYNQ7000 for board-level testing. The experimental results show that the TDC can achieve time resolution 10.91 ps in the dynamic time range of 3 ns. DNL(DNL) range is ［0.75,1.01］LSB,INL (INL) range for ［1.74, 2.19］ LSB.

Abstract:Accurate estimation of the state of health (SOH) of lithium batteries plays an important role in the health management of battery systems. In order to improve the accuracy of SOH estimation, a SOH estimation method that combines the parameter-optimized multivariate variational modal decomposition (MVMD) and stochastic configuration network (SCN) is proposed. Multiple health factors (HF) are extracted from the lithium battery charging and discharging process as inputs to the SOH model, and adaptive weights and optimal domain fluctuation strategies are introduced in the global stage of the Zebra Optimization Algorithm (ZOA) to improve its global searching ability, to obtain the Improved Global Zebra Optimization Algorithm (IGZOA), which is utilized to search for the optimization of the MVMD and the SCN parameters, and finally, the MVMD and SCN parameters are tested in nine benchmark functions IGZOA performance, the proposed combined method is compared with different methods for lithium battery SOH estimation on NASA and CALCE datasets, and the results show that the average values of root mean square error and absolute error of the proposed method are 0.84% and 0.93%, respectively, and the proposed method has higher prediction accuracy and generalizability.

Abstract:An improved A* algorithm is proposed to address the issues of low search efficiency, path diagonally crossing obstacle vertices, and excessive turns in mobile robot path planning. Firstly, a strategy is introduced to avoid diagonally crossing obstacle vertices during the neighborhood expansion in the A* algorithm. Secondly, an exponential weight is applied to the evaluation function based on obstacle factors to reduce unnecessary search and improve the efficiency and adaptability of the A* algorithm, favoring paths with fewer obstacles. Finally, a three-phase optimization strategy is employed, considering the obstacle safety distance, to minimize redundant nodes and turns in the path. MATLAB simulations are conducted in grid maps of sizes 20×20 m, 40×40 m, and 60×60 m. The results demonstrate that the improved A* algorithm significantly reduces search time by 70.12%, 84.31%, and 91.44%, respectively, and reduces the number of expanded nodes by 53.77%, 71.20%, and 74.30%, respectively. Moreover, the accumulated turning angles in the path are reduced by 70.48%, 76.31%, and 82.18%, respectively. The improved A* algorithm effectively enhances the efficiency of mobile robot path planning, resulting in smoother and safer paths, especially in complex environments.

Abstract:This paper addresses challenges in unmanned surface vessel (USV) trajectory tracking, such as significant errors induced by wind and waves, limited adaptability of adaptive gains, and vibrations in sliding mode control. A novel adaptive Super-Twisting sliding mode control algorithm is proposed to mitigate these issues. The USV mathematical model, derived from its structure, is transformed into a second-order system differential equation using trajectory reference points. An adaptive Super-Twisting sliding mode controller is designed, and adaptive gains ensuring closed-loop system stability are derived through the construction of a Lyapunov function. Simulation experiments, considering wind and wave resistance, compare the proposed method with Super-Twisting sliding mode control and traditional adaptive Super-Twisting sliding mode control. Results demonstrate a reduction in average tracking errors of 0.60 m and 0.27 m, respectively, over a 30-second simulation period, affirming the effectiveness of the proposed method in enhancing system control performance, suppressing vibrations, and reducing trajectory tracking errors.

Abstract:To enhance the real-time and robustness of digital instrument digital display area detection, an improved AdvancedEAST algorithm is proposed to complete the detection of substation digital instrument digital display area quickly and accurately. First, under the framework of the AdvancedEAST model, an ES-MobilenetV3 lightweight backbone network enhanced by the attention mechanism is constructed. By considering the influence of different layers of extracted features on the results, the ECA and multi-dimensional attention mechanism (ECA-SE) is introduced to The Bneck block of MobileNetv3 is improved to highlight key features while maintaining a balance between performance and complexity. A depthwise separable convolution is introduced in the neck network to reduce the computational complexity of the network and improve the detection speed. At the same time, the transfer learning strategy is used to improve the generalization ability of the model under small samples. Finally, the experimental verification was carried out on the constructed substation digital instrument dataset. The results showed that the proposed algorithm reduced the number of parameters of the model by 82% compared to the AdvancedEAST algorithm and increased the detection speed by nearly 2 times while ensuring detection accuracy.

Abstract:Aiming at the problem that the ship motion prediction model does not have high accuracy and the error of prediction results is too large, an extreme learning machine (ELM) prediction model is proposed to optimize the model parameters using the improved African vultures optimization algorithm (IAVOA), and use the model to predict the ship motion conditions. machine (ELM) prediction model, and use the model to predict the ship′s motion conditions. Circle chaotic mapping is introduced in the initialization of the population to increase the diversity of the population; adaptive operators are added to adjust the guiding role of two types of vultures to other vultures to improve the convergence speed and the quality of the algorithm. The IAVOA-optimized ELM model is used to predict the ship model pool test motion data, and the root-mean-square error and the mean absolute error are used to judge the prediction model. Comparing with other existing heuristic algorithms to optimize the ELM model, the proposed IAVOA-ELM has a better prediction accuracy and generalization ability.

Abstract:For the trajectory planning of multi-stage continuum robots, a trajectory planning algorithm based on deep deterministic policy gradient reinforcement learning is proposed. Firstly, based on the piecewise constant curvature hypothesis, the forward velocity kinematic model of joint angular velocity and end pose of the continuum robot is established. Then, the reinforcement learning algorithm is used to take the current pose and target pose of the robot arm as state input, the joint angular velocity of the robot arm as the output action of the agent, and a reasonable reward function is set to guide the robot to move from the initial pose to the target pose. Finally, a simulation system is built in MATLAB, and the simulation results show that the reinforcement learning algorithm successfully performs trajectory planning for the multi-segment continuum robot and controls the end of the continuum robot to move smoothly to the target pose.

Abstract:The data of three-dimensional point cloud obtained by laser sensor scanning is inevitably mixed with noises and outliers, resulting in a decrease in the fitting accuracy of the point cloud plane. In order to solve this problem, this paper proposes a method that combines M-estimate Sample Consensus (MSAC) algorithm and principal component analysis (PCA) method to fit the point cloud plane. Firstly, via this method, the MSAC algorithm is used to remove the abnormal points for the point cloud data, and the ideal point cloud plane is obtained. Then, the PCA method is used to fit the retained point cloud data to obtain more accurate point cloud plane parameters.Using the battery tray as the test object, a 3D line laser profile sensor is used to scan the test object and transmit the point cloud data to a computer for processing.Through experiments with the simulation data and battery tray point cloud data, it is found that compared with the method of random sample consensus (RANSAC) combined with PCA and least square median (LMedS) combined with PCA,the proposed method can significantly reduce the influence of outliers on point cloud plane fitting and obtain more accurate plane fitting parameters when the time consumption is approaching.When plane fitting is performed on the two parts of the battery tray point cloud after filtering, it can be found that the standard deviation of the proposed method is reduced by 28.6% and 22.5%, 24.0% and 29.0%, respectively, compared with the other two methods.Thus, this method has strong plane fitting accuracy and practicability.

Abstract:This study delves into the application of a deep Q-Network (DQN) algorithm, which integrates strategies of Navigational Priority (NP) and Prioritized Experience Replay (PER), for intelligent path planning in maritime environments. Unlike conventional path planning algorithms, our optimized model autonomously explores and learns the patterns of the maritime environment without relying on manually constructed global maritime information. We have developed a maritime simulation environment based on the Gym framework to simulate and validate our improved DQN model. This model incorporates the mechanisms of Navigational Priority and Prioritized Experience Replay, enhancing the algorithm′s learning efficiency for critical decisions by adjusting the frequency of experience sample utilization during the learning process. Additionally, the introduction of a novel reward function has further strengthened the model′s adaptability and stability in addressing path planning issues. Simulation experiments demonstrate that our model significantly outperforms baseline methods in avoiding obstacles and finding optimal routes, showcasing notable generalizability and exceptional stability.

Abstract:Objective: To solve the problem of motion blur in the monitoring image of the pantograph carbon slide caused by the fast running speed of high-speed railway, an image deblurring method based on improved multi-stage progressive network is proposed. Methods: First, a hybrid dilated convolution is introduced as a feature extraction network, which can increase the local receptive field without changing the calculation and resolution of the feature map, and then obtain high-quality image texture and detail information. Secondly, the pixel attention mechanism was introduced to adaptively select the weight value of each pixel to enhance the deblurring quality of the model. Thirdly, a hybrid loss function was introduced to improve the robustness of the model to different types of fuzziness. Finally, a synthetic data set of 1 600 pairs of pantograph carbon slide monitoring images was made for the model to train and test. The experimental results show that the peak signal-to-noise ratio (PSNR) reaches 38.82 dB and the structural similarity (SSIM) reaches 0.972 3. Compared with the other seven classical methods, the proposed network can better restore the edge contour and texture detail information of the image. The robustness of the model is effectively improved.

Abstract:Unsupervised stereo matching algorithms have important applications in areas such as autonomous driving, however, unsupervised stereo matching algorithms have low disparity accuracy in the region of object continuity, edges and other detail information, this paper proposes a new unsupervised stereo matching algorithm to improve the accuracy of detail information region by combining spatial attention mechanism and parallax fusion. Specifically, the spatial feature pyramid network is designed by introducing a spatial attention mechanism and residual structure into the feature pyramid network, to suppress the loss of edge and small target details in the process of feature extraction. Further, a disparity fusion module is constructed to improve the accuracy of the continuous detail information region, where the original disparity generated by the semi-global block matching algorithm and the initial disparity generated by disparity regression are fused with confidence disparity. Moreover, For the network loss function, the original disparity supervised loss and confidence masking loss are integrated to retain more detailed information at image edges and continuous regions. Finally, the experimental results show that the mis-matching rate of the proposed algorithm in the non-occluded region and all regions in the KITTI 2015 test set are 6.24% and 5.89%, respectively, which greatly improves the effect and accuracy of the detailed information region compared with other classical algorithms.

Abstract:Due to the large neural network structure, vast number of parameters, complex environment, and poor performance of hardware equipment in actual engineering inspection sites, the real-time detection rate of defects is slow and the accuracy is low. This study combines the lightweight concept of Depthwise Separable Convolution from MobileNet with the ECA attention mechanism module, as well as the feature extraction model of the U-Net network, to propose a photovoltaic panel defect detection method based on an improved U-Net network model. At the same time, according to the characteristics of photovoltaic cell defects, suitable activation functions were selected and the cross-entropy loss function was improved. Experimental results show that the improved U-Net algorithm not only reduced the number of parameters by 36% compared to the original algorithm, but also achieved a detection accuracy of 97.05% for defects such as cracks and black spots, demonstrating better performance in segmenting surface defects of photovoltaic cells than traditional networks.

Abstract:Aiming at the small target scale, complex background, and serious leakage and misdetection of aerial images, an aerial small target detection algorithm based on location awareness and cross-layer feature fusion, DC-YOLOv8s, is proposed.DC-YOLOv8s adds a new small target detection layer, which enhances the sensitivity to the small target scale and improves the detection accuracy. In order to reduce the loss of feature information, a cross-layer feature fusion module is designed to fully fuse the small target shallow semantic information and deep semantic information to enrich the feature representation. Improve the C2f structure, combined with variability convolution to design a sensory field attention module based on position-aware incorporation of residuals, adapting to the changes in the shape of aerial small targets, quickly extracting sensory field features, and reducing the rate of leakage detection and false detection. Finally, the dynamic detection head based on the attention mechanism is used to improve the localization performance of small targets in complex scenes in terms of scale perception, spatial perception, and task perception. The experiments show that on VisDrone2019 dataset, DC-YOLOv8s improves 7.2%, 7.5%, and 9.1% on P, R, and mAP respectively compared to YOLOv8s, which significantly improves the performance of small target detection, and the FPS is 71 frames, which meets the realtime requirement. Experimental verification of model generalizability is carried out on VOC2007+2012, and the effect is better than other classical algorithms.

Abstract:To address the issue of uneven distribution and insufficient correctly matched feature points in weakly-textured aircraft skin, an improved LoFTR algorithm is proposed for stitching aircraft skin images. Based on the posture of the camera, cylindrical back-projection is utilized to correct skin image curvature. By determining the feature extraction area through the overlapping regions between images, the generation of falsely matched point pairs is reduced. The LoFTR algorithm is employed for feature extraction, and the RANSAC algorithm is applied for feature point sorting. Adhering to the idea of image partitioning, grid division is used on overlapping areas for further sorting of feature points, ensuring a more even distribution, thereby yielding a more accurate transformation matrix for image registration. Experiments conducted on aircraft skin images collected via our self-developed unmanned vehicles confirmed the efficacy of this improved method. A feature matching rate comparison experiment with SIFT, SURF, ORB, BRISK, and AKAZE showed match rates of 4.84%, 0.47%, 2.9%, 0.86%, and 5.08%, respectively, while the proposed algorithm achieved a feature match rate of 55.21%. The average SSIM increased by 44.38% to 88.46%. The proposed method is effective for stitching tasks of aircraft skin images, and it eliminates the issue of missed stitches due to weak textures.

Abstract:Traditional backbone convolutional networks for feature extraction in few-shot image classification suffer from the problems of single context information, limited receptive field, and lack of global edge feature similarity measurement. In this paper, we propose a few-shot image classification algorithm based on prototype embedding graph network. First, the feature obtained by multiplying the weight values generated by CBAM with the features of different scales obtained by ASPP at different sampling rates is used as the node embedding feature of the graph network. Then, the prototype nodes are constructed in the measurement module using the prototype network method, transforming the similarity calculation between paired nodes into the sum of the similarity between a single node and the prototype node, and using the obtained similarity as the edge feature input to the graph neural network. Finally, the label information is propagated from labeled samples to unlabeled samples through the dual graph structure after multiple update iterations. In the classification task using ResNet-12 as the backbone convolutional network, our algorithm achieves classification accuracies of 71.47%, 75.41%, 86.21%, and 79.84% on the miniImageNet, tieredImageNet, CUB-200-2011, and CIFAR-FS datasets, respectively, for the 5way-1shot task. In the backbone convolutional network using Conv-4, our proposed algorithm outperforms existing graph network methods in both 5way-1shot and 5way-5shot tasks.

Abstract:The particularity of the underwater optical environment causes fog, color cast, and low contrast in normally captured underwater images. Existing restoration methods suffer from local misestimation in the depth map and inaccuracy in the transmission map. In order to better improve the quality of underwater images, an improved dark channel underwater image restoration algorithm that can balance attenuation and accurately estimate global ambient light is proposed. Construct a depth map through gradient maps and channel differences to obtain global ambient light; obtain a more accurate transmission map by balancing the attenuated dark channel; correct color casts through a white balance algorithm to further improve image quality. In order to verify the effectiveness of the algorithm, experimental evaluations were conducted on the UIEBD data set and RUIE data set. The UIQM index was 1560 and 1668 respectively, and the UCIQE index was 0470 and 0364 respectively. Experimental results show that the improvements proposed in this article can effectively improve the restoration effect of underwater images.

Abstract:The elbow is a crucial component of a pipeline and can be subject to fluid scour erosion and other defects that threaten its safe operation. A highly effective method for detecting pipeline defects is magnetic flux leakage (MFL) detection, and accurately quantifying these defects is of significant importance. In order to enhance comprehension of defect patterns in small-diameter pipe elbows and improve the measurement accuracy of defects, this paper proposed a novel image quantification method for MFL defects in small-diameter pipe elbows using an improved Canny operator. The image features of metal loss defects at different locations of the elbow were analyzed by establishing a simulation model for MFL detection in small-diameter pipe elbows. The defect image quantization model was constructed by using morphological filtering and OTSU optimized Canny operator, combined with image processing methods. This model corrected the depth quantification of defects from various positions on the elbow. The experimental results clearly showed that there are differences in the images of defects at different positions on the elbow. The accuracy of the quantification model in measuring defect length and width is precise, with an error rate of less than 2 mm. However, quantifying the depth of defects revealed a more significant error rate, with a precision of 86.34% post-correction. Nonetheless, this level of accuracy satisfies the necessary standard for detecting metal loss defects. The suggested approach allows for batch processing of defect images and therefore holds considerable importance in detecting MFL defects in pipelines.

Abstract:Aiming at the problem of poor reconstruction of Multi-View Stereo Networks in challenging regions such as weak textures or non-Lambertian surfaces, this paper first proposes a multi-scale feature extraction module based on three parallel dilated convolution and attention mechanism, which enables the network to capture the dependencies between features while increasing the sensory field to obtain global context information, thus enhancing the multi-view stereo network′s ability to characterize features in challenging regions for robust feature matching. Secondly, an attention mechanism is introduced in the 3D CNN part of the cost volume regularization so that the network pays attention to the important regions in the cost volume for smoothing. Additionally, a neural rendering network is built, which utilizes the rendering reference loss to accurately resolve the geometric appearance information expressed by the radiance field and introduces the depth consistency loss to maintain the geometric consistency between the multi-view stereo network and the neural rendering network, which effectively mitigates the detrimental effect of the noisy cost volume on the multi-view stereo network. The algorithm is tested in the indoor DTU dataset, achieving completeness and overall metrics of 0.289 and 0.326, respectively. Compared to the benchmark method CasMVSNet, there is an improvement of 24.9% and 8.2% in the two metrics, demonstrating high-quality reconstruction even in challenging regions. In the outdoor Tanks and Temples intermediate dataset, the average F-score for point cloud reconstruction is 60.31, showing a 9.9% improvement over the UCS-Net method. This reflects the algorithm′s strong generalization capability.

Abstract:Due to the complex and changeable working environment, the motion parameters of the cross-domain vehicle are greatly affected by transient impacts during the flight process, which leads to the reduction of navigation accuracy and difficulty in stable control. Aiming at the problem of transient impacts that make it difficult to obtain accurate motion parameters of the cross-domain vehicle, an inertial navigation device based on the combination of large and small-range accelerometers is developed, and a switching strategy of large and small-range accelerometers is proposed, which utilizes the Butterworth filter to filter the collected data, and realizes the detection of the impact point through the accumulation of preset thresholds and control chart algorithms and an autonomous variable sliding window, and adopts a segmented linear interpolation method to ensure the data synchronization. Segmented linear interpolation method ensures the synchronization of the data. Finally, the switching strategy proposed in this paper is experimentally verified by simulation and drop table test, and the test results show that the estimation error of the segmented linear interpolation method is no more than 0.6 g, and it switches to the measured value of the large-range accelerometer after the small-range accelerometer measurement value is saturated to satisfy the requirement of real-time switching, which further proves that this inertial navigation This further proves that the inertial navigation device and the small-and large-range switching strategy can be applied to the cross-domain vehicle for cross-domain experiments.

Abstract:A joint DOA and polarization parameters estimation algorithm based on modulus-constrained reduced-dimensional and root multiple signal classification (MUSIC) is proposed for coprime polarization sensitive arrays to address the high computational complexity and ambiguity pairing errors. Firstly, by reconstructing the three-dimensional spectral function, the DOA and polarization parameters are decoupled to achieve dimensionality reduction in the three-dimensional MUSIC method. Then, the DOA is solved using polynomial roots, and the beamforming method is used to solve the problem of ambiguity angle mismatch in the coprime array. Finally, the cost function is constructed using the modulus boundedness of the polarization vector to derive the closed form solution of the polarization parameters. The numerical simulation results have verified the effectiveness of the algorithm. The simulation results show that the parameter estimation performance of the proposed algorithm is better than that of the estimating signal parameter via rotational invariance techniques (ESPRIT), and is basically equivalent to the one-dimensional total spectral peak search MUSIC (1D-TSS-MUSIC) algorithm. However, our algorithm significantly reduces computational complexity and can still obtain reliable parameter estimation in multi source scenarios.

Abstract:In order to solve the problem that the actual muscle force is represented by the extremity force and the degree of muscle fatigue is not taken into account in muscle force measurement, this paper studies an upper limb muscle force prediction method based on surface EMG and muscle fatigue. The musculoskeletal model of upper limb was established by AnyBody software, and the muscle force of a single muscle was obtained by simulation of the end force of upper limb. The time of isometric muscle contraction was used to characterize the degree of muscle fatigue. Ten healthy male subjects were subjected to the upper limb isometric contraction experiment, and six eigenvalues of integrated electromyography, root mean square, median frequency, average power frequency, wavelet coefficient and frequency were extracted during the experiment. After analyzing muscle force, eigenvalue and muscle fatigue degree, it is found that the three are highly correlated. The Sparrow search algorithm (SSA) was used to optimize the weights and thresholds of BP neural network, and the upper limb muscle strength prediction model was constructed and trained. The test results show that the error of this method is less than 12%, and it can predict the muscle strength accurately.

Abstract:Currently, the brain-machine interface based on steady-state visual evoked potential (SSVEP) has received wide attention in human-computer collaboration, and the existing deep learning classification methods oriented to the phase and frequency information of SSVEP signals still have problems such as poor classification of SSVEP signals due to insufficient utilization of the information. And a variety of classification algorithms have appeared for solving the above problems. In this paper, a deep neural network model for SSVEP signal classification is proposed based on the idea of migration learning, which takes the complex vectors after the fast Fourier transform as inputs, and convolves the real and imaginary part vectors of each lead to learn the corresponding phase frequency characteristics. The model is divided into two parts: the first part uses the statistical commonality among all subjects to obtain the global phase-frequency feature module for phase and frequency information; the second part uses the trained global phase-frequency feature module to initialize the local phase-frequency feature module, and fine-tunes the training parameters through further reinforcement learning of the local phase-frequency feature module in order to reduce the individual differences between each subject. Tested on the public dataset BETA, the average accuracy and average information transfer rate are 89.98% and 71.80 bit/min, respectively, when the time window length is 1.5 s. The experimental results show that the classification algorithm model in this paper achieves a relatively good classification effect compared with other methods, and the designed global and local phase-frequency feature modules are able to improve the effect of individual differences on the classification results. The designed global and local phase-frequency feature module can improve the influence of individual differences on the classification results, which provides a brand new idea for the in-depth mining and utilization of phase and frequency information in SSVEP signals.