Abstract:In-situ inspection and maintenance of wind turbine blades play a crucial role in ensuring the safe operation of wind turbines. Currently, the mainstream UAV inspection method requires panoramic stitching of wind turbine blade images to further locate and analyze minor blade defects and assess the overall blade condition. An improved image stitching technology based on NCC algorithm is proposed to solve the problem of difficult stitching caused by single structure and sparse texture of wind turbine blades. Canny edge detection algorithm is used to extract the blade edge and filter the duplication to get the boundary coordinates, NCC algorithm traverses the blade boundary coordinates for searching and matching to optimize the search strategy and speed up computation while increasing the weight of key information to improve the stitching accuracy, and combining image pyramid coarse-fine matching thoughts to further speed up the algorithm. Finally, the spatial corresponding relation is obtained according to the optimal matching position to achieve stitching. The experimental results show that the matching time of the proposed method is about 6% of the original NCC algorithm and 3%~10% of other classical gray matching algorithms, and lower than other improved NCC algorithms. The stitching success rate is 94.74%, which is higher than all comparison methods, and finally, the visible panoramic image of the wind turbine blade is obtained successfully, demonstrating its good stability in panoramic stitching of large-size wind turbine blades visible images.

Abstract:Aiming at the significant size discrepancies and various shapes among different models of workshop tools, a workshop tool detection method based on sparse learnable proposal is proposed. Firstly, sparse representation and learnable proposal mechanism are integrated to improve the robustness of the model and reduce the required parameters in the detection process. Secondly, Swin-Transformer structure is introduced to enhance the global and detail learning ability of the model, which can effectively overcome the shortcomings of traditional convolution neural network in high-level semantic information fusion. Thirdly, an improved multi-scale feature fusion network architecture is used to improve the detection ability of the model for various scale targets according to effective fusion of different scale features. Finally, multi-head attention and dynamic convolution are combined to establish a more precise and detailed connection between different feature layers, thereby furtherly improving the accuracy of target detection. The CIoU loss function is applied to make the regression prediction of the boundary box more comprehensive and accurate by considering the location, scale and shape information. The experimental results show that the average detection accuracy of the proposed method for workshop tool detection reaches 91%, which is at least 2.3% higher than the current mainstream methods. At the same time, the detection speed of a single picture is about 53 ms, which meets the needs of real-time detection and reflects the excellent comprehensive performance.

Abstract:To address the issue of low detection accuracy in remote sensing image target detection methods for small devices with limited resources, an efficient and lightweight method based on the improved YOLOv7-tiny algorithm is proposed. To address the issue of dense distribution of small targets in remote sensing images, a low-span context decoupling detection head module is designed. This module fuses deep and shallow features to perform classification and regression tasks for target detection. It effectively solves the problems of leakage and misdetection of small targets in remote sensing images. Meanwhile, a parallel series attention mechanism is proposed to enhance the network’s ability to extract multi-scale target features for remote sensing image targets. This is achieved by combining the parallel three-branch network with the spatial attention module. Additionally, the model’s generalization ability is improved through the introduction of the Focal-EIoU loss function. The comparison experiments, ablation experiments, deployment experiments and generalization experiments were conducted on the model. Experimental results indicate that the detection accuracy on DIOR-5s and NWPU VHR-10 datasets improved by 2.6% and 1.7%, respectively, compared to the original model. The model size is only 19.1 MB, and the detection speed is 64.1 fps, verifying the algorithm’s effectiveness.

Abstract:To address the issues of detail information loss and blurred salient target contours in existing infrared and visible image fusion models during deep feature extraction, we propose an infrared and visible image fusion method that combines semantic segmentation with cross-modality differential feature compensation (CMDFC). By incorporating a crossmodality differential feature compensation module with a convolutional block attention module (CBAM), complementary features from different modalities are integrated into the original features for deep feature extraction. Additionally, a semantic segmentation network is introduced to perform pixel-level classification on the fused image, constructing a semantic loss to constrain the fusion network, and a decoder is used to reconstruct the fused image. Experimental results on public datasets show that compared to the best metrics of the reference models, the proposed model achieves various degrees of improvement in five selected metrics, with mutual information (MI) and visual information fidelity (VIF) increased of 4.41% and 4.25%, respectively. These results indicate that the proposed model generates clearer fused images with stronger correlation to the source images, effectively mitigating the issue of feature detail loss during the fusion process and enhancing the visual quality and contrast of the generated images.

Abstract:Light field imaging technology can capture both the spatial and angular information of light in a scene simultaneously. It is commonly used in various computer vision tasks. A two-stage 6D pose estimation method leveraging light field decoupled feature fusion is proposed. The aim of this method is to overcome the limitations of RGB image pose estimation methods when predicting pose in complex scenes with severe occlusion and truncation, illumination changes, and similarity between objects and backgrounds. Various feature extractors are utilised to decouple the light field macro-pixel image and map it to the feature space. An attention mechanism is then introduced to fuse the spatial, angular and EPI information to provide effective and reliable features for the downstream pose estimation network. Additionally, the back-projection is applied to the keypoints prediction network to minimise information loss during feature transfer. Experiments on the LF-6Dpose light field pose estimation dataset demonstrate that this method achieves 91.37% and 70.12% for the average closest point 3D distance for symmetric objects (ADD-S) and 2D Project metrics, respectively. This represents a 12.5% improvement compared to existing state-of-the-art methods in 3D distance metrics and more effectively solves the problem of estimating the 6D pose of objects in complex scenes.

Abstract:Dynamic magnetic resonance imaging (DMRI) is an imaging technology that acquires images through continuous scanning to capture their changes over time and space. Applying compressed sensing technology to DMRI tends to result in unsatisfactory visual quality of the reconstructed magnetic resonance images. Therefore, to address the deficiencies of compressed sensing in DMRI reconstruction, a reconstruction model based on low-rank and sparse decomposition is proposed by using 1 norm to characterize the sparsity of magnetic resonance image data and utilizing low-rank to describe the intrinsic correlation of dynamic magnetic resonance image sequences. This effectively reduces artifacts in dynamic magnetic resonance imaging. In the modeling phase, the sparse component is modeled using the 1 norm, while the low-rank component is modeled using the nuclear norm. In the model optimization phase, a wavelet framework regularization method is introduced, and the reconstruction model is transformed into a non-smooth convex optimization problem, which is then solved by using a momentum-accelerated proximal gradient method. Finally, experiments are conducted on cardiac cine, cardiac perfusion, and phantom membrane image data to verify the effectiveness of the proposed model. The experimental results show that the average PSNR and the average SSIM of the proposed method reach 33.709 0 dB and 0.966 0 at a sampling ratio of 30%, respectively, which further improves the reconstruction accuracy of the dynamic magnetic resonance image.

Abstract:Aiming at the problems of low timeliness of single centralized data processing in the cloud, low accuracy of bird’s nest detection on overhead lines, high consumption of model’s arithmetic power on edge computing devices, and inaccurate target localization, an algorithm for detecting and localizing bird’s nests on overhead lines based on the collaboration of cloud-edge and end-end is proposed. The algorithm solves the problem of low efficiency of centralized processing in the cloud through the collaboration of cloud, end and edge, and solves the problem of unclear images due to angle and light through the collaboration of cloud-edge data visualization. In order to improve the accuracy of bird’s nest detection on overhead lines, the algorithm is optimized on the basis of YOLOv5x model. First, by replacing the C3 module in the backbone feature extraction network with the C2f module, and adding the SE (squeeze and excitation) attention module in the last layer to improve the model’s ability to detect small targets. Secondly, the activation function is replaced with the Mish function to solve the problem of neurons stopping learning due to the saturation of the training gradient. In order to reduce the model’s consumption of computing power on edge computing devices, the improved model is pruned and fine-tuned to reduce the scale of model parameters. Based on this optimized model, a 3D target localization algorithm is proposed, and the localization results are corrected by combining with the GIS (geographic information system) system, which achieves accurate localization of the detected target. The experimental data show that the mean average accuracy of the improved model reaches 93.25%, which is 3.44% higher than the original YOLOv5x model, and the pruning rate of the optimized model reaches 45%. The detection target is able to locate to the corresponding pole tower after 3D spatial modeling calculation and position correction, which effectively guides the staff to quickly and accurately eliminate hidden dangers.

Abstract:To address the current issues of low accuracy and misidentification in indirect tire pressure monitoring systems (iTPMS) when detecting underinflation in all four wheels simultaneously, as well as the misidentification caused by engine vibrations, a study was conducted on the wheel speed spectral characteristics generated by road excitation. An indirect tire pressure monitoring algorithm based on wheel speed sensors and onboard hardware is proposed to improve the accuracy of underinflated tire identification. First, the wheel speed signal is preprocessed through signal denoising and gear error filtering to eliminate the issues of gear skipping and multi-tooth interference. The gear ring error is corrected using the recursive least squares (RLS) method. Then, combined with the analysis of tire vibration characteristics, the fast fourier transform (FFT) method is used to obtain the spectral characteristics of the wheel speed. Band-pass filters (BSP) and notch filters (NF) are applied to acquire the wheel speed spectral characteristics within a specified range, eliminating the influence of engine rotation. The tire pressure status is determined by the resulting resonance frequency peaks, where the peak frequency of an underinflated tire is 2~3 Hz lower than that of a normally inflated tire. Based on this characteristic, the tire pressure result is provided. Real vehicle test results indicate that this algorithm can eliminate the impact of engine rotation on wheel speed, ensuring the identification accuracy of single, dual, and triple underinflated wheels, as well as accurately identifying simultaneous underinflation in all four wheels. The condition recognition capability increases by approximately 18%, and the accuracy of identifying conditions where engine speed affects tire resonance improves by about 25%. Compared to traditional indirect tire pressure monitoring, this algorithm can more accurately and promptly inform the driver to avoid the risk of tire blowouts.

Abstract:The time-to-digital converter (TDC) is a device designed to convert the continuous analog value of time interval between signal pulses into discrete digital values. The tapped-delay-line time-to-digital converter (TDL-TDC) is commonly implemented using the internal carry chain resources of field programmable gate array (FPGA) chips. However, the delay time of each delay bin in the TDL-TDC is significantly impacted by variations in operating temperature. Currently, methods of TDC calibration, such as code density calibration, linear compensation, or high-order Taylor function fitting, struggle to accurately model the changing delay times of individual bins within a long delay line under varying temperature conditions. To maintain the required precision of TDC operations, a neural network calibration based on multilayer perceptron (MLP) is proposed. This method utilizes delay time and corresponding temperature data from 128 delay bins in the delay line as training data to construct a four-layer MLP. By feeding back temperature information when working, the network can independently calculate delay time of different bin to determine the time interval between signal pulses. Experimental results confirm the effectiveness of the network calibration in compensating for temperature variations, with the potential for deployment across different FPGA chips. The network achieves an accuracy of 91%, and the resolution of TDC is 34 ps.

Abstract:Addressing the issues of TDOA estimation accuracy being limited by sampling frequency and the large localization errors of a single microphone array, this paper proposes two improved methods and conducts simulation verification. Firstly, to tackle the problem of significant TDOA estimation errors caused by low sampling frequency, a modified cross-correlation algorithm based on cubic spline interpolation is proposed. This method involves performing cubic spline interpolation after the first cross-correlation to increase the sampling frequency, followed by a second cross-correlation to obtain the TDOA. Simulations under 10x interpolation conditions demonstrate that this method reduces the TDOA error from 7.6% to 0.6%, effectively minimizing the estimation error.Secondly, to address the large localization errors associated with a single microphone array, a multi-microphone array data fusion method for sound source localization is proposed. This approach involves fusing the localization results of multiple microphone arrays to obtain the final sound source position. Using a quaternary cross array as an example, simulations were conducted at 10 points in space, comparing far-field localization accuracy between dual-microphone arrays, quad-microphone arrays, and single microphone arrays. The results show that the localization error for a single microphone array exceeds 1 meter when the sound source is distant, whereas the proposed dual-microphone array design achieves localization errors below 0.3 meters, and the quad-microphone array achieves errors below 0.2 meters, significantly enhancing the accuracy of sound source localization.

Abstract:Arrhythmias is a common cardiovascular disease, which seriously affects the quality of life and safety of patients. The automatic classification of arrhythmia utilizing electrocardiogram (ECG) is of great significance for timely diagnosis and prevention. An arrhythmia classification method using multi-feature fusion is proposed. Firstly, the short time Fourier transform (STFT) features and wavelet transform (WT) features are respectively extracted from denoised ECG. Then, its deep STFT features is obtained by the branch aggregated residual network (BCAR-NET) with STFT features as input, and its deep WT features is obtained by the 1D-CNN with WT features as input. Moreover, the LSTM is used to extract deep ECG features. Finally, a fully connected network is used to concatenate and fuse the three deep features, then arrhythmia classification is realized. The proposed arrhythmia classification method is evaluated on the MIT-BIH arrhythmia dataset. The accuracy of the proposed method is 98.66%, and the macro-average F1 score is 94.22%, which is better than traditional arrhythmia classification methods. The experimental results show that the constructed multi-feature fusion network improves the classification performance of arrhythmia by effectively exploiting the complementarity between deep STFT features, WT features, and ECG features.

Abstract:Aiming at the abuse of gaseous substances such as nitrous oxide (N2O), the detection method based on the combination of quantum cascade laser and wavelength modulation technology can detect the concentration of nitrous oxide exhumed by users, and realize the rapid on-site identification of users. Aiming at the measurement error caused by laser light intensity jitter, a digital phase-locked amplifier for trace N2O gas detection is proposed. The laser light stress modulator is integrated in the phase-locked amplifier. When the mercury cadmium telluride detector receives the absorbed light intensity signal, the first and second harmonics are demodulated in FPGA at the same time, and the second harmonics are normalized to avoid the influence of light intensity jitter on the measurement. The performance of the detection system is evaluated by using the phase-locked amplifier. The mass flow controller is used to set different concentrations of N2O gas. When the laser modulation frequency is 5 kHz, the linear fitting degree between the normalized second harmonic peak value and the N2O gas concentration is 0.992 1. The detection system detection limit is evaluated using Allan variance. When the system integration time is 0.1 s, the detection limit of N2O gas is 17.7×10-9; when the integration time is 6.3 s, the detection limit of N2O gas reached its lowest value of 8.9×10-9. The phase-locked amplifier has the advantages of high integration, fast detection speed and high signal-to-noise ratio, and is suitable for the field rapid detection of trace N2O gas.

Abstract:In the field of NDT, infrared thermal imaging testing has an important place. In the use of infrared thermal imaging testing to detect defects in the process, by the processing technology and the influence of the surrounding environment, the surface roughness of the equipment to be detected are different, the emissivity of the surface of the material is not uniform, resulting in a large error in temperature acquisition. To address this problem, a global detection method for infrared thermography with correctable emissivity is explored in conjunction with the fundamental law of thermal radiation. The intention is to correct the surface emissivity of the material and to reduce the risk of safety issues such as misdetection and omission of defects that may occur during the inspection process. The simulation and experimental results show that: on the one hand, when using infrared thermography detection, the surface roughness is different and it will seriously interfere with the accuracy of the detection, the larger the roughness, the higher the detected radiation temperature; on the other hand, the use of the separation of temperature method can be achieved by correcting the surface of the object, which results from the roughness of the different emissivity that brought about by the different impact of the different simulation, experimental results show that the corrective accuracy of up to 70% or so, and can reach a maximum of more than 75%. Based on this corrective detection method, it can effectively achieve global detection based on infrared thermal imaging technology, reduce the impact of the emissivity on the detection accuracy of infrared thermal camera, improve the detection accuracy, and greatly improve the efficiency of defect detection.

Abstract:This paper presents an enhanced A* algorithm aimed at resolving issues of diagonal traversal of obstacles, excessive numbers of turning points, and non-smooth paths produced by the conventional A* algorithm. Initially, by excluding all forced neighboring nodes, the search neighborhood is optimized to prevent the generated path from diagonally crossing obstacles, thereby enhancing path safety and dependability. Subsequently, a safe distance is established, and crucial turning points are extracted from the generated path after optimizing the neighborhood, reducing path redundancies and simplifying the path structure. Lastly, the algorithm employs Bezier curves to interpolate essential turning points, determining the quantity and position of control points for each segment based on the positions of adjacent necessary turning points and the connecting line slopes to achieve segmental smoothing. Simulation experiments demonstrate that compared to the original A* algorithm, the improved A* algorithm exhibits an average 33.68% enhancement in path safety and a corresponding average reduction of 37.00% in the number of turning points. Additionally, the robot′s turning angle and path curvature are continuous, ensuring path smoothness. The paths generated by the refined A* algorithm are smooth, with few turning points, and maintain a safe distance from obstacles, which can be applied to indoor path planning for mobile robots.

Abstract:In response to the limitations of existing deep learning-based object detection methods when faced with real-world unmanned aerial vehicle (UAV) targets, such as poor robustness, low accuracy, and high model complexity, a YOLO-based object detection method called OD-YOLO is proposed. This algorithm addresses the characteristics of UAV targets being small, slow, and low. Several improvements have been implemented. Firstly, to tackle the issue of learning information loss and insufficient emphasis on target information during the downsampling process, a spatial-to-depth convolution is introduced to ensure the preservation of learning information while highlighting the features of UAV targets. Secondly, to further enhance the accuracy of object detection and improve its generalization across different backgrounds, a full-dimensional dynamic convolution is be used. This enhances the accuracy of object detection and improves its generalization capabilities across various backgrounds. Lastly, the backbone network of the model is modified to enhance the semantic features of UAV targets and reduce the size of the skeleton, resulting in a reduced parameter count and improved computational efficiency of the model, while maintaining effective representation capabilities for UAV targets. Experimental simulations were conducted to compare OD-YOLO with current state-of-the-art object detection algorithms. The results demonstrate significant improvements in accuracy and lightweight performance for OD-YOLO. The mAP and Recall distributions increased by 3.4% and 5.1%, respectively, compared to the original model.

Abstract:The monitoring and analysis of massive data from radio spectrum monitoring are essential components of radio regulation work. To address this, the paper proposes a lossless compression method based on wavelet-like transform for radio spectrum monitoring data. This method first converts the one-dimensional spectrum signal into a two-dimensional matrix based on temporal correlation. Once transformed into a two-dimensional matrix, there is redundancy in both the horizontal and vertical directions. The algorithm employs a convolutional neural network to replace the prediction and update modules in traditional wavelet transform, and introduces an adaptive compression block to handle features of different dimensions, thereby obtaining a more compact representation of spectrum data. Furthermore, the paper designs a context-based deep entropy model, which utilizes the wavelet-like transform′s different subband coefficients to obtain entropy coding parameters, estimating cumulative probabilities to achieve spectrum data compression. Experimental results indicate that the proposed algorithm achieves additional performance improvements compared to existing traditional lossless compression methods for spectrum data, such as Deflate. Moreover, when compared with typical two-dimensional image lossless compression methods like JPEG2000, PNG, and JPEG-LS, the proposed method achieves over 20% better compression effectiveness.

Abstract:In actual service environments, 40% of the total failures that occur in aircraft are caused by sealing leaks, and the quality of their sealing performance directly affects the functionality, performance, and reliability of the product. However, the lack of wear monitoring of sealing rings on aircraft makes it difficult to evaluate the health status of sealing rings. To address this issue, this paper proposes a method for predicting the wear of active O-sealing based on physical model and statistical analysis. Firstly, a mechanism analysis and parameter measurement are conducted on the wear of the sealing ring. Secondly, Abaqus finite element software is used to simulate and study the motion process of the sealing ring, obtaining the contact stress of the sealing ring. Then, statistical analysis is conducted on the cumulative stroke of each servo flight to obtain the probability statistical curve of the stroke. Combined with the Holm Archard wear model, the probability distribution curve of the wear volume is obtained. Finally, based on the relationship between time and travel, a mathematical model between wear volume and time is established. Multiple samples are used for validation, and the results show that the probability of the actual wear volume within the predicted density function’s 3σ range is 95.83%, proving that the model in this paper has a high probability of predicting the volume wear of the O-sealing for active seals.

Abstract:The advanced driving assistance system domain controller is responsible for processing and analyzing data from various sensors. However, as the number of in-vehicle cameras continues to grow, various stages within the Domain Controller, including deserialization, serialization, and image processing, may encounter frame loss and pixel anomalies, which can adversely affect the results of image processing. To accurately evaluate the GMSL2 video output functionality of the ADC, a dual-channel GMSL2 video capture and video quality comparison system has been researched. The system involves a hardware card design that initially deserializes the GMSL2 video signal into MIPI CSI-2 signals. Subsequently, a bridging IC separates the MIPI signals into LVDS and CMOS signals recognizable by the FPGA. The XLINX XC7K325T-FFG900 main control chip is then utilized for FPGA logic design, enabling the parsing of MIPI signals, conversion of YUV422 to RGB888 video format, DDR3 buffering, and PCIe 2.0×8 bus transmission. Finally, by integrating image feature extraction, digital tube threading recognition algorithms, and the RGB weighted euclidean color difference formula, the system achieves detection of frame loss and color differences in the video. The experimental results indicate that this system can collect dual-channel YUV422 8 bit, 4K, 30 fps video data from the GMSL2 interface in real time, and conduct a quantitative analysis to determine whether the video output from the intelligent driving domain controller has issues with frame loss and color differences, thereby distinguishing between qualified and unqualified device under test. This has increased the reliability of the test results for the video input and output functions of the advanced driving assistance system domain controller.

Abstract:In recent years, multi-view 3D reconstruction technology based on deep learning has become one of the research hotspots in the field of machine vision and is applied in many fields. However, the 3D reconstruction technology still has problems such as edge missing, serious network memory consumption and low reconstruction accuracy. In this paper, based on the existing problems of 3D reconstruction technology, EGF-MVSNet network is proposed based on deep learning. First, a feature extraction network incorporating the ECA attention layer is used to improve the network′s attention to the channel feature information; then, an improved combination of GRU modules is used to obtain the GC regularization network for regularization and to reduce the computation of the network; finally, the SmoothL1 loss function and Adam optimizer are used to improve the convergence accuracy at the later stage of model training and to optimize the model′s losses and parameters. Through testing and validation on the DTU public dataset, the EGF-MVSNet network proposed in this paper improves the completeness by 22.1% and the overall model score by 11.5% compared to the MVSNet network, which confirms that the EGF-MVSNet network can significantly improve the quality of the reconstruction results and reduce the network′s consumption of memory.

Abstract:In the realm of cross-user gesture recognition research, addressing the challenges of negative transfer and subpar model generalization observed in single-source domain transfer learning, this study presents a multi-source domain transfer learning strategy centered around the fusion of EMG and inertial features. The pivotal innovation lies in amalgamating data resources originating from diverse source domains, and subsequently employing techniques for domain-specific feature alignment and domain classifier alignment. The primary objective of this approach is to bolster model performance in gesture recognition across different users, thus significantly enhancing the accuracy of cross-user gesture recognition systems. Initially, the long short-term memory (LSTM) network model is introduced to extract time series features, encompassing metrics such as average absolute value, variance, and peak value derived from EMG and inertia data. Subsequently, domain-specific feature alignment and domain classifier alignment procedures are executed, facilitating feature extraction within the target domain utilizing data from multiple source domains. Lastly, the fusion of three loss functions—classification loss, domain-specific feature difference loss, and domain classifier difference loss—is undertaken to collectively optimize the overall loss. The experimental results demonstrate that the proposed method exhibits an improvement in average recognition rate compared to various traditional methods, such as single-source domain and source domain combination approaches. On the NinaPro DB5 dataset, the average gesture recognition accuracy for the target users exceeds 80%.

Abstract:In the fault types of permanent magnet synchronous motors (PMSM), inter-turn short circuit (ITSC) faults are relatively common, making the accurate extraction of fault features particularly significant. However, during fault feature extraction, modal mixing often occurs. In order to accurately extract the fault features of vibration signals in permanent magnet synchronous motor (PMSM) when inter-turn short circuit (ITSC) occurs, proposes an adaptive nonlinear signal processing method based on particle swarm optimized variational mode decomposition (PSO-VMD). Firstly, particle swarm optimization (PSO) is used to find the optimal number of decomposition layers and quadratic penalty factor for variational modal decomposition (VMD) to obtain the optimal decomposition model. Secondly, the optimal decomposition model is used to decompose the motor vibration signals to obtain a series of intrinsic mode functions (IMF). After that, the variance contribution rate (VCR) of each IMF is calculated, and the cumulative variance contribution rate (C-VCR) is further calculated to filter out the IMF that contain fault signature information. Finally, the filtered IMF are analyzed by applying the Hilbert transform (HT), and the three-dimensional time-frequency diagrams are used to output the time, the instantaneous frequency and the amplitude to complete the fault feature extraction. In order to verify the validity and accuracy of the proposed method, an experimental platform of the ITSC in PMSM was built, and the proposed method was used to process the measured signals. The experimental results show that the proposed PSO-VMD method effectively improves the phenomenon of modal mixing, can more accurately extract fault features, and has better engineering applicability.

Abstract:In autonomous aerial refueling, the circular structure of the drogue refueling port is often used to assist target positioning. Still, the complex background interference and oiled plug obscuring significantly reduce the accuracy of circular feature extraction. To address the background interference problem, an adaptive mean filter is designed to obtain the center of mass of refueling ports to obtain the accurate set of edge points in a smaller range using the imaging operation. To address the oiled plug obscuring problem, an outlier elimination algorithm based on convex hull detection is proposed to enhance the anti-interference performance of feature extraction. An iterative reweighted least squares based on geometric distance is proposed to optimize elliptic targets. On the simulation platform, the influence of K value on the fitting accuracy and efficiency of the iterative reweighted least squares algorithm is emphatically analyzed. At the same time, the accuracy and anti-occlusion performance of the fitting algorithm are tested. The average error of the algorithm is less than 0.5% when there is no occlusion and less than 2% when the occlusion rate is 50%. Finally, the feature extraction experiment of the actual drogue is carried out. Compared with other classical algorithms, the accuracy is improved by 49.3%, the average extraction error is 0.79%, the average processing time is 13.9 ms, and the extraction error is controlled within 2% under the special case that the drogue is obscured. Experimental results show that the positioning feature extraction method of drogue meets the requirements of accuracy, rapidity and robustness of image processing for autonomous aerial refueling, and can improve the success rate of autonomous aerial refueling docking and reduce the probability of accidents.

Abstract:In order to solve the problems of high resonance frequency and low output power of traditional piezoelectric energy harvesters, a T-type magnetically coupled piezoelectric energy harvester was proposed. By introducing the magnetically coupled interaction between the two sides, the piezoelectric beam is induced to generate bending-torsional vibration, which increases the deformation of the piezoelectric beam, and further enhances the output voltage response of the harvester. Firstly, the working principle and structural design of the piezoelectric energy harvester were analyzed, and its vibration characteristics were simulated and analyzed, and finally, the experimental prototype was made for experimental testing, and the effectiveness of the piezoelectric energy harvester was verified by simulation analysis and experimental tests. The results show that the optimal load resistance maximizes the output power. The energy harvesting frequency can be controlled by adjusting the weight of the mass at the free end of the device, and the output power improves with the increase of excitation acceleration. Introducing magnetic coupling, the influence of the bending-torsional coupling vibration caused by magnetic coupling on the output performance of the harvester was analyzed through comparative experiments. When the excitation acceleration is 0.2 g, Due to the bending-torsion complex vibration, the working frequency and maximum output power of the magnetic coupling device Ⅲ are 5.5 Hz and 3.71 mW, respectively. Compared to the energy harvesting device I, without magnetically coupled, the output power is increased by about 47%. Therefore, the T-type magnetically coupled piezoelectric energy harvester has a high output efficiency in the low-frequency environment.

Abstract:Aiming at the problems of low convergence accuracy and local optimality in sparrow search algorithm (SSA), an improved sparrow search algorithm (ISSA) is proposed and applied to the diagnosis of inter-turn short-circuit fault in PMSM. Firstly, the PMSM inter-turn short-circuit simulation model is built to simulate the fault of different short-circuit turns ratio. Secondly, the fault is analyzed, and three fault recognition features are extracted. Then, the experiment platform is used to test the fault of different short-circuit turns ratio. Then, the sparrow search algorithm (SSA) is introduced and optimized by using Tent chaotic mapping, adaptive sine-cosine strategy and Levy flight strategy to generate an improved sparrow search algorithm (ISSA). Meanwhile, ISSA algorithm is compared with SSA algorithm, particle swarm optimization algorithm (PSO) and grey wolf optimization (GWO) on the test function. It is proved that it has advantages in optimization ability and stability. Then, the random forest (RF) algorithm is introduced, and the fault diagnosis model of ISSA-RF is built. Finally, four algorithms are used to optimize the basic parameters of RF and achieve fault classification. The results show that the proposed improved method can detect the inter-turn short-circuit fault and its severity, and the accuracy of ISSA-RF model reaches 98.5%, which verifies the effectiveness and reliability of the algorithm.

Abstract:The short-circuit currents distributed from distribution networks with small current grounding mode are relatively little after single-phase grounding faults. However, due to the aforementioned faults the long-term presence of arcing phenomena can increase the potential fire risk. In order to minimize fire threat, the existing methods for identifying fault types are based on whether the arcs occur or not. Whereas, the impact of arcs’ occurrence frequency is not taken into account. Aimed at the problem, on the basis of the actual cases of single-phase grounding faults in a certain distribution network the correlation between zero-sequence current characteristics and corresponding arcing phenomenon is firstly analyzed. And a novel classification method for single-phase grounding faults is proposed involving the arcs’ occurrence frequency. The waveforms characteristics of zero-sequence current are further extracted under the distinct fault types, such as “flat shoulder distortion”, “transient change” and so on. The aforementioned characteristics are mathematically described using the energy proportion of zero-sequence current components with different frequency bands, harmonic centroid and the arcing cycle number. Used these mathematical features as inputs, a fault-type identification model based on long short-term memory (LSTM) networks was developed. At last, the proposed model is tested with a dataset of 223 typical fault cases collected from a certain power company. It is verified that the accuracy rate of proposed model is 96.4%. The distinct fault types can be identified effectively. It is significant for reducing the fire risk and saving on the costs associated with the maintenance of distribution networks.

Abstract:The high temperature of series arc fault is one of the main causes of electrical fire. Aiming at the problem that there is no effective protection method for the series arc fault in the load circuit of industrial frequency converter, a new method of fault detection and line selection for the series arc fault was proposed. First, the series arc fault experiments in different lines were carried out for the load circuit of three-phase frequency converter commonly used in industrial field. Second, the improved variational mode decomposition based on the principle of energy convergence was used to adaptively decompose the A-phase current signal at the front end of the frequency converter into multiple modal components. After multiplying the single modal component by the energy coefficient, the feature enhancement signals of multiple current signals were reconstructed, and the feature matrix was established. Third, the feature matrix was divided into blocks, and the kernel principal component analysis was used to reduce the dimension of each block matrix, and the matrix composed of the reduced dimension signal was reduced twice to construct the fault feature vector. Finally, the support vector machine optimized by the pelican optimization algorithm was used to detect the series arc fault and select the fault line. The results show that the proposed method can realize the fault detection and line selection of the series arc fault in six lines of the whole circuit of the frequency converter only by analyzing the A-phase current signal at the front end of the frequency converter, and the accuracy of fault detection and line selection is more than 98%.