Abstract:With the development of communication technology and cloud computing technology, the front-end obtains data and uploads it to the cloud for analysis, and then transmits the cloud analysis results back to the front-end for display APP design and audio mode are gradually popularized in life. Drawing on its ideas, a JS+Python test signal analysis cloud platform design technology is proposed. Among them, the web-based JS front-end page program is responsible for signal acquisition and analysis result display, while the Python program running on the cloud server is responsible for cloud computing and analysis of uploading data. Run the test experiment platform. After the data and script are uploaded, the server can successfully receive and call the corresponding test technology and control theory algorithm in the script to process the data, and return it to the browser to visualize the results.This minimizes the hardware and software requirements of the front-end equipment, and facilitates the realization of the equipment operation fault monitoring and diagnosis network composed of front-end measurement device+back-end test signal analysis cloud platform.

Abstract:To address the issue of drift and the high cost of calibration in the F-P filtering demodulation method, this paper had developed a F-P filter multi-channel resonance demodulation system using the FPGA and the fiber optic switch. The system utilizes DDS signals to generate specific triangular wave driving voltages and achieves dynamic calibration of wavelength-time by linking multiple-dimensional reference FBGs in series. Three dynamic calibration modes, including linear and quadratic relationships, are set. The FPGA-controlled fiber optic switch enables fast channel switching of modes and step-by-step acquisition of reference wavelength and wavelength signal to be measured. he average deviations in the three modes are 0.028 02 nm, 0.018 14 nm, and 0.010 9 nm, respectively, at a frequency of 100 Hz. The research results indicate that, under a unit voltage of 1 V, the deviation is reduced by over 45% compared to static calibration, improving the demodulation accuracy Improves demodulation accuracy, reduces calibration costs and addressing the issue of misjudgment caused by the proximity of the reference and measured wavelengths.

Abstract:Axial flux permanent magnet synchronous motor has the characteristics of short axial length, compact structure, small volume, high torque density, etc. However, its magnetic field is a three-dimensional magnetic field, and the model is too complex to be solved and calculated directly by the classical analytical method. In order to quickly complete the preliminary estimation of the axial flux permanent magnet synchronous motor, a simplified equivalent magnetic circuit model is proposed to quickly calculate the basic performance of the axial flux permanent magnet synchronous motor. The magnetic circuit model transforms the three-dimensional axial flux permanent magnet synchronous motor model into a multi-layer two-dimensional equivalent model through reasonable equivalence, and then uses the equivalent magnetic circuit method and superposition method to calculate the motor performance. Finally, taking a 10 pole 12 slot double stator single rotor axial flux permanent magnet synchronous motor as an example, the model described in the article was used to calculate its air gap magnetic density, permanent magnet flux linkage, and no-load back electromotive force, and finite element simulations were conducted on motor models with different layers. The calculation results show that the calculation accuracy of the equivalent magnetic circuit model in this paper is similar to that of the three-dimensional finite element method, and the time consumption is very short, which is more conducive to the initial engineering calculation of this kind of motor.

Abstract:Aiming at the problems of long equalization time and large energy loss in the equalization process of series-connected lithium-ion battery packs, a two-level equalization topology is designed in this paper, and a variable universe fuzzy Logic control strategy based on battery state of charge is designed for this topology. The proposed topology adopts an improved Buck-Boost circuit within the battery pack to optimize the equalization path, and a centralized single-inductance equalization circuit is used between the battery packs, which can achieve equalization between any battery packs. The proposed strategy introduces a contraction-expansion factor on the basis of fuzzy logic control to flexibly regulate the input domain, which further improves the equalization speed and energy utilization by precisely adjusting the equalization current. Finally, the equalization system is built for validation, and the results show that the topology of this paper reduces the equalization time by about 12.53% compared to the grouped Buck-Boost topology. Under the same static and charging/discharging conditions, the strategy in this paper not only reduces the equalization time by about 20.98% but also improves the energy utilization by about 7% compared with the FLC algorithm. The feasibility of the equalization scheme in this paper is verified.

Abstract:Deep learning methods can improve the speed and accuracy of AOI. However, due to the variable nature of industrial production scenarios, the models have to be updated continuously to ensure performance, which increases time and labor consumption. In order to improve the model iteration efficiency of deep learning methods in actual practice of AOI, an automatic training system for defect detection models of PCBA electronic components is developed in this paper. The system can automatically train the defect detection models required for four common types of electronic components(Chip, IC, SOT, and Plug-in). The automatic training process is divided into three parts: automatic data enhancement, parameter tuning, and deployment. Experimental results show that the models automatically trained by the system outperform the manual training models. Compared with manual training, the training time is shortened by 36% to 42%, and the overall accuracy is increased by 1.3% to 4.1%. At present, the proposed system has completed testing and the automatically trained model can meet the requirements of actual practice of AOI. It effectively improves the speed of model iteration, reduces labor costs and demonstrates its good application prospects.

Abstract:Due to the limitation of the satellite platform, the bandwidth and power resources of the satellite transponder are limited, and the improvement of resource utilization is particularly important. Aiming at the problem of joint power and bandwidth allocation in multi-beam satellite communication system, this paper proposes a joint bandwidth and power allocation scheme based on improved particle swarm optimization algorithm. The scheme aims to minimize the total second-order service rejection of the system. The optimal solution is obtained by introducing an improved particle swarm optimization algorithm with inertia weight, local suppression mechanism and penalty function. Finally, a balance between the total capacity of the system and the fairness of the user′s access capacity is achieved.The simulation results show that compared with the traditional algorithm, the method based on the improved particle swarm optimization algorithm proposed in this paper can significantly improve the effect of joint bandwidth and power allocation of multi-beam satellite communication system.

Abstract:In view of D2D communication reuse cellular user spectrum can improve the system throughput and frequency utilization, a joint resource allocation scheme is designed to maximize the system throughput under system user communication quality, firstly designing a mode selection scheme based on throughput optimization, and then optimizing D2D users by hybrid grey Wolf optimization algorithm under the premise of known channel allocation vectors. The simulation results show that the proposed scheme can effectively improve the total system throughput by 20% and reduce the interference suffered by cellular users by 90% compared with other schemes, and can also improve the convergence effect and operation speed.

Abstract:In recent years, the response speed of temperature sensors applied in extreme temperature measurement situations in aerospace has reached the microsecond scale, however, how to realize dynamic calibration of them at the microsecond scale is still a challenge to be solved. The shock tube method and the laser method are two potential technological approaches to the problem, both of which can provide a microsecond temperature step response but also have different characteristics. The shock tube method is suitable for high-temperature and high-pressure environments, but the rising process and amplitude of its step signal are unstable. The laser method is characterized by non-contact and high accuracy, but a unified evaluation system has not yet been established to directly compare the dynamic response characteristics of different temperature sensors. Finally, three key research directions are summarized: improving existing calibration methods, exploring new calibration methods, and constructing new evaluation theories to facilitate the development of microsecond temperature sensor dynamic calibration research and related technology applications.

Abstract:Aiming at the problem that the bandwidth of large-range hall current sensors cannot be tested with high currents of different frequencies due to the limitation of detection equipment, a bandwidth detection circuit is designed and the relationship between the bandwidth of large-range hall current sensors and the test current is explored. A large range pulse test current of different frequencies can be obtained by chopping the DC high current through the bandwidth detection circuit, which flows through the hall current sensor and the constantan wire array. According to the -3 dB principle, the distortion degree of the output signal amplitude of the hall current sensor relative to the condon wire array is compared, and the bandwidth of the hall current sensor is determined. The bandwidth of the hall current sensor was tested by using 30～100 A test current of different sizes, and the results showed that the test current was different, the bandwidth of the large-range hall current sensor was different, and the bandwidth gradually decreased with the increase of the test current.

Abstract:This paper proposes a power cabin temperature inversion method based on POD-RBF surrogate model and feature point KNN correction, aiming to address the challenges of high computational complexity and poor adaptability in numerical simulation methods for calculating the temperature distribution of power cabin under different working conditions. The proposed method utilizes simulated temperature field data obtained from various operating conditions to construct a temperature inversion model for the power cabin using the proper orthogonal decomposition (POD) and radial basis function (RBF) approach. This surrogate model effectively avoids redundant calculations, enabling rapid approximation of the simulated model. Meanwhile, the K-Nearest Neighbor (KNN) algorithm is employed to introduce the feature temperature points into the inversion model for correcting the temperature inversion error and improving the accuracy and adaptability of the in-version. Taking an actual power cabin as an example, the temperature inversion of the power cabin under the specified working condition is conducted. The results show that the proposed method can achieve real-time temperature inversion of the power cabin under the condition of cable current flow and known temperature of feature temperature points, with the maximum relative error between the inversion temperature and the simulation calculation temperature of 0.96%, meeting the engineering application standard.

Abstract:In proton exchange membrane fuel cell (PEMFC) life prediction, the unknown degree of influence of the characteristics in the fuel cell on its life makes the problem of predicting the remaining life of the fuel cell relatively complex. In order to more accurately predict the remaining service life of the fuel cell. In this paper, the original stack voltage was de-noised by wavelet analysis to filter the noisy data. pearson correlation coefficient (PCC) was used to reduce the dimension of influencing factors, extract key influencing factors, and simplify the model structure. Then, the improved sparrow search algorithm (ISSA) is used to optimize the BP neural network, find the optimal weights and thresholds of the network, and establish the ISSA-BP model. Finally, the processed data is input into the ISSA-BP model to predict the remaining life of PEMFC.The experimental results show that the average absolute error percentage, average absolute error, and root mean square error of PCC-ISSA-BP are 0.125%, 0.003 97, and 0.005 68, respectively, which are better than other models and can more effectively predict the remaining life of fuel cells.

Abstract:Power flow calculations are the basis for the operation and control of power systems. In order to solve the problems of uncertainty of voltage fluctuation at the point of load caused by the increasing penetration rate of renewable energy in the distribution network, and the inaccuracy of power flow calculation caused by the insufficient power flow data collection capacity of traditional power system. In this paper, a data-driven power flow analysis model is proposed, and a power flow calculation method based on back propagation neural network combined with genetic algorithm and adaptive moment estimation optimization algorithm is constructed to analyze the power flow calculation method of distribution network under randomness. Firstly, the initial power flow information, topological structure characteristics and power factor indicators are introduced to construct the training set, and the mapping relationship between node voltage and power is fully explored through the training of the regression model. Secondly, the GA-ADAM algorithm is used to optimize the initial value and weight parameters of the model. Finally, based on the IEEE-33 bus distribution network model, the maximum error is 3.93×10-3， average absolute error is 1.46×10-3， and root mean square error is 1.81×10-3 of the model power flow calculation in this article, the optimized BPNN power flow calculation voltage error value is reduced by 37.66%. The simulation results of actual examples show that compared with other methods, the model constructed in this paper has smaller error indicators and higher accuracy, which improves the efficiency and accuracy of power flow calculation.

Abstract:Aiming at the problem of large fluctuation of winding tension when the winding system is working, an inverse nonsingular fast terminal sliding mode tension control method based on neural network interval observer is proposed. The mathematical model of the winding system is constructed, and the neural network is used to approximate the random response caused by the change of parameters such as the radius and inertia of the winding system. The interval state observer is designed to estimate the upper and lower bounds of the system speed and the winding tension. According to the estimated state value, the backstepping nonsingular terminal sliding mode controller is constructed to make the tension tracking error converge to zero quickly in a finite time, which effectively enhances the robust performance of the system. The simulation results show that the designed control method makes the tension on the coil reach a given value and remain constant after 1.6 s. Compared with the conventional sliding mode controller and the sliding mode controller in the published literature, the adjustment time is reduced by 57% and 33% respectively, which proves the effectiveness and reliability of the proposed control method and meets the requirements of the winding process of the winding equipment.

Abstract:Dynamic path planning is a critical factor in ensuring the safe flight of unmanned aerial vehicles (UAVs) in complex interference environments. To address the issues of high iteration counts, slow convergence, and dynamic obstacle avoidance in dynamic path planning, this paper proposes a UAV dynamic path planning method based on obstacle motion prediction and improved artificial potential fields (APF). First, for dynamic obstacles, a target detection algorithm based on laser radar and a motion prediction algorithm based on Kalman filtering are designed to estimate dynamic obstacle information. A velocity direction similarity detection method is introduced for local position evasion decisions. Secondly, for static obstacles, a simulated annealing algorithm is introduced to perturb the current state, coupled with a neighborhood optimization function based on target points for dynamic path planning. Simulation results show that the proposed algorithm reduces dynamic obstacle avoidance time by 69% when dealing with static obstacles and reduces obstacle avoidance distance by 19.7% and task duration by 23.6% when dealing with dynamic obstacles, thereby enhancing the safety and efficiency of UAV mission execution.

Abstract:The dimensionality reduction performance of Local Linear Embedding algorithm LLE is closely related to the manifold structure mined. However, the manifold structure mined by LLE is singular and sensitive to the selection of neighborhood parameters, making it difficult to extract a comprehensive local structure of the manifold, which limits its dimensionality reduction performance.Therefore, this article proposes an adaptive local linear embedding algorithm based on multiple information fusion MIF-ALLE. MIF-ALLE firstly uses tangent space approximation criterion to adaptively select neighborhood parameters to obtain more accurate local neighborhood; Then, the angle information of Tangent space contained in the local neighborhood is fused with the local linear information to mine a more comprehensive local structure of the manifold and reduce the deviation of local low dimensional embedding; Finally, the experimental verification is carried out on the bearing data set published and the bearing data set extracted from the laboratory. The experimental results show that MIF-ALLE can mine more comprehensive manifold structures, extract more significant features, and achieve bearing fault diagnosis accuracy of up to 100%.

Abstract:In fingerprint indoor positioning, constructing a highquality fingerprint database is a prerequisite for achieving highprecision positioning. Collecting enough signal samples at each reference point during the fingerprint dataset establishment stage usually consumes a lot of manpower and time costs, to solve this problem, this paper proposes a fingerprint database augmentation method based on an improved conditional deep convolutional generative adversarial network. The network model uses the reference point index as conditional information to generate corresponding samples for each reference point. It uses the least squares loss function instead of the cross-entropy loss function to avoid the problem of gradient disappearance that often occurs during training. Experimental results demonstrate that this method can effectively increase the sample size of each reference point, improve the training effect of the convolutional neural network and the positioning accuracy in small sample cases. The root mean square error is reduced to 0.44 meters, and the proportion of positioning errors within 1 meter is 86.98%, while that within 2 meters is 92.72%.

Abstract:In order to further strengthen the monitoring of university electric bicycle charging process and facilitate electric bicycle users and battery researchers to obtain charging data, an electric bicycle charging process monitoring system based on the IoT was designed. The system mainly relies on Zigbee module for local networking. STM32 issues control commands to each IM1281B module through the network to collect and aggregate charging data of multiple electric bicycles. The summarized data is uploaded to the cloud server through the WiFi module, and the WeChat mini program obtains real-time information of the cloud server to realize user interaction. Open data sites store long-term charging data and provide battery researchers with a wealth of data. The experimental results show that the networking range of ZigBee module is 150 m, which meets the networking distance requirements of electric bicycle charging points in universities. The charging data collected is calibrated by standard electrician digital clamp meter VC866A, and the relative error is less than 3%. The system successfully realizes the collection, transmission, storage and display of charging data of multiple electric bicycles.

Abstract:To address issues such as low detection accuracy and the occurrence of missing or misidentifying bovine faces due to their small size, we propose an enhanced model called Mask R-CNN+MResNet. Firstly, we introduce a MResNet network based on the ResNet101 architecture, which enhances the detection accuracy of the model by improving upon ResNet101. Secondly, we adjust the anchor frame size of the model′s RPN network to enhance its capability in detecting small targets. Experimental results demonstrate that compared to the original network model, MResNet achieves a 12.6% improvement in bovine face detection accuracy. Furthermore, the improved model exhibits a 2.4% increase in average accuracy for detecting small targets, compared to the original model. These results indicate that this model effectively detects small target cow faces and holds practical application value.

Abstract:The tenon tooth serves as a vital component of aero-engine blades, where its profile dimension precision plays a pivotal role in establishing the firmness of blade installation and ensuring the safety of engine operations. However, the conventional manual projection continues to be employed as the prevailing measurement approach in current industrial production, exhibiting drawbacks including low efficiency and inconsistent accuracy. Drawing upon machine vision technology, this paper puts forth a novel measurement method and designs a corresponding system aimed at measuring the dimension of aero-engine blade tenon tooth profiles. Employ an industrial camera to acquire realtime images of tenon tooth profiles. The obtained images are subsequently processed using machine vision software based on HALCON, which includes sequential operations such as image pre-processing, edge extraction, geometric feature extraction, and geometric dimension transformation. Through these operations, precise dimensional information of the tenon tooth profile is obtained. Additionally, a visual graphical software with interactive information is designed using Visual Studio 2019 to facilitate the analysis and visualization of the obtained data. The experimental results demonstrate that the developed system achieves an average accuracy error of 0.023 6 mm and 0.270 1° for straight line and angle measurements, respectively, and the time consumption for a single measurement does not exceed 2 s. Compared to the conventional manual projection measurement method, the developed system significantly enhances the accuracy and speed of measurements while reducing manual labor costs. Furthermore, it exhibits the promising potential for user-friendly on-site deployment, offering enhanced convenience in practical applications.

Abstract:Aiming at the problems of complex smoke occurrence scene and low accuracy in smoke, an improved YOLOv5s smoke and flame detection method is proposed. Firstly, in order to solve the problem that the Neck feature fusion of smoke and flame is not accurate and the effect is poor, a new channel attention mechanism, Scoring module, is proposed to score features of each channel. Features with high scores are selected for feature fusion and features with low scores are filtered to avoid introducing too many redundant features. On the premise of not increasing too much computation burden, the module can enhance feature fusion ability and detection accuracy. Then, in order to improve the prediction ability of the Head layer, α-EIOU is used to replace GIOU as the prediction box regression loss to improve the prediction accuracy of the prediction box. Finally, the improved Mosaic data enhancement method is used to solve the problem of small data set and single data form, expand the sample data, and improve the generalization ability of the model. As a result, the mean average precision of the modified YOLOv5s model is improved by 4.7%, while the detection speed reaches 212 frames per second. Meanwhile, it performs well in the comparison experiment with other improved YOLOv5s. It achieves good detection effect in the image with complex environment, and can meet the task of smoke and flame detection in complex environment.

Abstract:In order to solve the problem of low object detection accuracy in complex driving scenarios which makes it difficult to meet the needs of autonomous driving, an efficient network model MEPNet based on YOLOP is proposed. MEPNet can simultaneously handle three tasks: vehicle detection, drivable area segmentation, and lane detection. First, YOLOv7 is used as the main structure to balance the accuracy and real-time performance. Second, the FRFB module is designed to enlarge receptive fields and enhance the feature extraction capability of the network. The proposed small object detection layer added to the head of the detection network effectively alleviates interference caused by vehicle occlusion and overlap. Finally, CARAFE is used as the upsampling operator to accurately locate object contours while preserving semantic information in images. Experimental results show that the algorithm achieves a inference speed of 42.5 fps, and compared with the baseline YOLOP, it improves the mAP50 and Recall of vehicle detection by 6.8% and 6.3%, the accuracy and IoU of lane detection by 6% and 1%, and the mIoU of drivable area segmentation reaches 92.5%, which significantly improves performance. Furthermore, MEPNet-s has been further designed to accomplish four-task object detection, while simultaneously meeting the accuracy and real-time requirements of autonomous driving.

Abstract:Pavement crack is the most common defect of road. With the development of deep learning technology, more and more methods are used to extract the crack information from pavement images. Aiming at the problems of low accuracy and lack of real-time due to incomplete extraction of crack features by existing deep learning pavement crack detection methods, a road crack detection method combining attention mechanism and GhostUNet is proposed. This method is composed of encoder and decoder. The conventional convolution in U-Net is improved to Ghost convolution and the number of model parameters is reduced. In coding and decoding, in order to improve the ability to extract crack features, ECA attention mechanism and residual connection are introduced. ECA attention module can filter irrelevant feature information, and residual connection can be used to avoid network degradation. To evaluate the effectiveness of this method in fracture detection, two publicly available fracture data sets were used, and ablation and comparison experiments were conducted. The experimental results of F1_score, P and R increased by 14.48%, 14.35% and 14.45%, respectively, compared with U-Net. The number of parameters in this model decreased by 14.2 MB compared with U-Net. Compared with similar models, this model has higher segmentation accuracy and fewer parameters.

Abstract:Aiming at the problem that the object detector relies too much on the classification effect of convolutional network and cannot use motion information when detecting falls, this paper designs a fall detection model based on YOLOv5s and improved centroid tracking. To solve the problem of resource consumption, the MobileNetV3 network and Slim Neck module are used to lightweight YOLOv5s, and the SE module in the MobileNetV3 network is replaced with the more efficient ECA module, which reduces the network complexity while maintaining high accuracy. Hash sensing algorithm is introduced to improve centroid tracking, increase the basis of target association, and improve the accuracy of fall detection. The experimental results show that the size of the improved YOLOv5s model is reduced by 52.2%, the computational capacity is reduced by 51.8%, and the accuracy is as high as 90.3%. The accuracy of fall detection model with improved centroid tracking was increased by 4.3%. The results show the effectiveness and superiority of the proposed model.

Abstract:To address the problem of artifacts and information residuals in the existing multi-focus image fusion algorithms, an algorithm is proposed to maximize the retention of information and clarity of each region based on the focusing characteristics of the image. Firstly, the focus region decision map is obtained by region detection, which is then used for initial fusion and boundary extraction to obtain the boundary region decision map; secondly, the ACS Network is used to learn the fusion rules of multi-focus images and generate the network fusion image; finally, the initial fusion image and the network fusion image are weighted and summed according to the boundary region decision map to obtain the final fusion image. The experimental results demonstrate that the algorithm outperforms other comparable algorithms in both the focus region and the boundary region, and the evaluation indexes are improved by more than 4.8% and 1.5%, respectively; Meanwhile, the subjective effect is more in line with HVS. The experiments have proved that the algorithm achieves good results in retaining the detailed information of the source image and avoiding visual artifacts in various regions.

Abstract:In response to the current problems of dimensional measurement of workpieces at the stage of accurate extraction of key measurement points of the workpieces, this paper proposes a workpiece target detection method based on the fusion of polarization features and intensity information. The polarization features are introduced on the basis of the workpiece intensity image, and a dual-stream network model with differentiated and efficient interaction between intensity information and polarization features is established to achieve a more efficient fusion of polarization features and intensity information. To validate the algorithm′s effectiveness, we have established a dataset for detecting the saliency of workpiece targets in polarization images. On this dataset, the proposed algorithm outperforms comparison algorithms in terms of Precision, max-F, S-measure, and visual results, underscoring its exceptional performance in workpiece target detection and its outstanding results in workpiece target detection.