Abstract:Aiming at the problem of large amount of the steering gear test data and unbalanced samples, an anomaly detection model is proposed that uses the Grey Wolf Optimization (GWO) to optimize the Deep Neural Networks (DNN) and combines it with the Logistic Regression Classification (GWO-DNN-LRC). The construction of the model effectively solves the problem that small samples in the steering gear test data are difficult to be accurately classified, and is suitable for the deep feature extraction and multi-fault classification of the steering gear test data. The accuracy of this method reaches 99.261%, which is 4.931%, 0.205%, and 0.087% higher than LRC, DNN, and GWO-DNN, respectively. The precision, recall, and F-score reach 98.417%, 98.062%, and 98.217%. In the comparison of classification accuracy of different categories, the categories of 6 small samples can reach 100%. Experimental results show that this method fully improves the performance of anomaly detection of steering gear, and is an effective application of deep learning technology in steering gear test data.

Abstract:In my country's air traffic control security surveillance, the control of surveillance accuracy is very important. In recent years, aviation unsafe incidents caused by "low, slow and small" aircraft have gradually increased, increasing the difficulty of airspace surveillance. In order to strengthen the orderly management of the airspace and increase the monitoring requirements for "low, slow and small" aircraft, a monitoring method is proposed that uses an external radiation source as the transmitter and only establishes a receiving station. The monitoring of "low, slow and small" aircraft in low-altitude airspace enables the monitoring accuracy to meet the monitoring requirements near the airport's take-off and landing route. By deriving the HDOP, the precision factor is obtained as an index to evaluate whether the monitoring method can achieve the monitoring accuracy, and the simulation results are carried out for different numbers of receiving stations, failures of receiving stations and different heights of receiving stations. The accuracy in all cases is less than the derived HDOP, which proves that this monitoring method can meet the needs of civil aviation for the monitoring accuracy of low-altitude airspace near the airport.

Abstract:Landing exploration of extraterrestrial objects is an important form and method of deep space exploration in China. To extract the pose of the probe rover in the ground exploration simulation training system, an approach of combining laser scanner and target ball is proposed to fit the position and attitude data of the rover, and the ground simulation training system is established. First of all, the laser scanning data is preprocessed to remove the outliers of the point cloud, and then the appropriate parameters are selected for resampling the data. The point cloud is segmented using the improved region growth method based on classification, a set of prespecified number of points are adopted to fit multiple target spheres and a local coordinate system is established. Through the analysis of the experimental data, the fitting accuracy of the target sphere meets the maximum allowable error of 3mm, and the point cloud processing speed is effectively improved, which verifies the accuracy and efficiency of the feature recognition method. Finally, the position and the attitude matrix of the rover is obtained through coordinate transformation to calibrate the camera accuracy.

Abstract:In order to achieve power mutual benefit between AC and DC sub-microgrids in island mode, a power allocation method of coordinated hybrid microgrid system is proposed in this paper. Firstly, the topology of the microgrid structure is analyzed for gathering characteristic quantity which can represent the reference value of power transmission. In addition, the switching condition of the operation stage is designed reasonably according to the characteristic quantity, where the operating mode of the system is divided into the interlinking converter free control section and ILC working control section. Secondly, a power feedback algorithm based on AC side bus voltage and DC side voltage feedback is designed to achieve the power mutual benefit purpose. Finally, a simulation model is built in MATLAB Simulink. The simulation results show that the AC-DC sub-microgrid can bear the power fluctuation of the system according to its own conditions and ensure the overall stable operation of the system by using the power mutual benefit method proposed in this paper. Furthermore, the response time of ILC power change is less than 50ms in the simulation, which indicates that the control strategy has a high dynamic response.

Abstract:High magnitude, narrow pulse time domain pulse signal sources and ultra-wide band, low later-time ring time domain pulse antennas are the important components of the pulse ground-penetrating radar. In this paper, the electro-magnetic interference can be decreased by using ducking technology. High amplitude time domain pulse signal sources with 156.2V pulse amplitude, 1. 6ns pulse width are designed. Besides this, a novel UWB time domain pulse antenna with semi-circle arms has been proposed. Compared to the traditional bow-tie antenna with triangular arms, the VSWR of the UWB time domain pulse antenna is less than 3.5 from 0. 5GHz to 7GHz. And the later-time ring decreased from 17.5% to 8.3%. Lastly, an experiment for detecting the buried targets is carried out using the integrated radar system, the experiment demonstrated high resolution ratio especially for shallow targets. Superior performances and huge value of the pulse ground-penetrating radar front-end detecting subsystem designed in this thesis have been further verified.

Abstract:The active noise control system uses the theory of acoustic interference cancellation, and the reference noise is processed by an adaptive algorithm to generate an anti-noise signal. Aiming at the problem that the compressor of gas station has a high sound pressure level in the middle and low frequency bands and is difficult to cancel, this paper first analyzes the compressor noise in time-frequency, and then proposes an active noise control system based on adaptive IIR. The system consists of two It consists of a digital filter with adjustable parameters and a corresponding adaptive algorithm. Finally, use Xilinx FPGA as the core control module of the system to design the hardware of the system. The actual measurement results show that the noise reduction effect of this method can reach 20dBA in the middle and low frequency bands, the system convergence speed is 30us, and the hardware resource consumption is lower.

Abstract:In order to solve the alignment problem between the detection head of the large-size photoelectric detection device and the reaction hole, a small, in the traditional chemiluminescence enzyme immunoassay method. A multi-mode optical fiber combined with the optical fiber-type photon counting probe is used to detect the target signal, and an aflatoxin B1 optical fiber detection system based on photoelectric detection technology is designed. The ZEMAX simulation and test results show that the optical fiber photon collection ratio reaches more than 95.83% and RSD ≤ 0.21% in the reaction hole range compared with the reaction hole center position detection, and the alignment requirements of the optical fiber detection port and the reaction hole are greatly reduced. The experimental results showed that the detection limit of the fiber optic detection system was 0.63 μg/L. The correlation coefficient of the standard curve was 0.9977. The RSD was 1.03~2.38 %, and the sample spiked recovery was 92.62~96.03 % with higher linearity, precision, and reproducibility.

Abstract:In view of the problems that the current traction battery system has not yet established a sound testing and evaluation standard system and the testing technology lags behind the development of the industry, the traction battery system testing and evaluation standard system is discussed and the key testing technologies are analyzed. Based on the current domestic and foreign traction battery testing standards, the traction battery testing and evaluation standard system covering electrical performance parameter testing, safety performance testing, life testing and battery management system testing has been preliminarily constructed. And the methods and ideas for key testing technologies of traction battery such as the testing and evaluation of key parameters under dynamic working conditions, the safety test of local thermal runaway of thermal diffusion, and the functional safety evaluation of battery management system are also provided, laying a solid foundation to achieve traction battery system testing.

Abstract:Simulated annealing algorithm is an effective method for solving unconstrained optimization problems, but it has the disadvantages of poor accuracy, easy to fall into local optimum and slow convergence when solving travel quotient problems. In order to improve the above problems, a parallel simulated annealing algorithm based on Spark platform is proposed in this paper. The cooling function of the simulated annealing algorithm is modified to construct the solution space of the travel quotient problem, the local search capability is enhanced by using the large neighborhood search technique and the 2-opt operator, the global search capability is enhanced by introducing the OX crossover idea, and the crossover cooperative experiment parallel strategy is proposed to be implemented in parallel with the Spark platform. A number of TSPLIB datasets are selected for simulation experiments to test both solution quality and running time, and to compare the experiments with other parallel algorithms of Spark framework.The simulation results show that the solution accuracy of this method is greatly improved, and the solution speed is 3-10 times higher than other algorithms, which can effectively solve the traveling salesman problem.

Abstract:Aiming at the problem of defect detection of circular screen printing pattern, an improved algorithm model based on yolov5 is used to detect the defect of printing pattern. In this experiment, the network structure of yolov5 model is changed according to the actual situation. Firstly, the backbone of yolov5 network is optimized and improved, and the attention mechanism module is introduced to extract the features of channel attention and spatial attention of input pictures respectively. Secondly, aiming at the small target of printing defects, the detection layer structure of the network is modified. The experimental results show that the accuracy of the improved yolov5 detection algorithm is improved by 14.4%, and the detection speed is also improved, reaching 43.1fps, which meets the requirements of real-time detection.

Abstract:For the problems of losing the model feature information and causing easily the point cloud surface holes in single simplification algorithms, a streamlined algorithm for curvature classification optimization based on dichotomous k-means clustering is proposed. First, the least squares method was used to fit the neighborhood surface, calculate the curvature value, and divide the significant and non-significant feature regions based on the curvature value，Second, dichotic k-means clustering was used to divide non-significant feature regions, select the subfeature points with feature importance retained according to the curvature threshold of subclusters, and finally the datasets and subfeature points were merged to obtain simplified results. The simplification algorithm is compared with the space surrounding box algorithm and the curvature reduction algorithm by the simulation experiments in terms of speed and information entropy. The results show that the proposed algorithm outperforms the other two algorithms in streamlining quality and has a certain application value in point cloud data reconstruction.

Abstract:In order to explore the causes of gait movement disorders in children with cerebral palsy from the level of neural control, this paper designed the surface EMG signal of 8 muscles of each lower limbs in 17 subjects during the gait process,and used the Kendall model to map the surface EMG signals,to the corresponding spinal motor neuron location. At the same time, in order to describe the changes of the activation position of the cord activation centor segments, this paper proposes the CoA curve of the central point,and calculated the number of extreme of the CoA curve. The results showed that the activation of spinal cord segments in both healthy children with cerebral palsy, it is found that there is no obvious difference in the spinal cord segments and uneven CoA curve in the gait movement of the two. The range of extreme points for healthy children is 8~34, cerebral palsy range from 50~144. The results show that the research results in this paper will help to better understand the neural control mechanism and the pathology of gait movement disorders in children with cerebral palsy, and further extend Kendall model method to the field of abnormal gait movement assessment in children with cerebral palsy.

Abstract:Raindrop image enhancement methods based on deep learning generally have some problems, such as highly dependent on aligned sample datasets, and blurred background detail after removing raindrops. In this regard, this paper proposes a weakly supervised raindrop image enhancement method guided by dual attention. This method designs and constructs a weakly supervised raindrop image enhancement network. It only needs images from the raindrop image domain and the clean image domain for training, which can effectively reduce the dependence on the aligned sample datasets. At the same time, dual attention is introduced into the generation network to guide the feature extraction and multi-branch masks generation. After the masks are fused with the input raindrop image, a clean image with a clear background is obtained, and the input raindrop image is enhanced. The experimental results show that the PSNR is 27.0711 dB and the SSIM is 0.8996 of the proposed method respectively on the Raindrop. The background details and color information of the image are better preserved than the previous methods, which prove the feasibility and effectiveness of the proposed method.

Abstract:The depth of field for imaging equipment is limited, the problem of out-of-focus of part of the acquisition image.An effective multi-focus image fusion algorithm is proposed, to further improve the contrast and sharpness of fused image. Firstly,the source image is decomposed into approximate subbands and detailed subbands by NSCT; Secondly,the FMO and ISML were used to combine the approximate subband coefficients and the detailed subband coefficients respectively;finally, the fused image is obtained by inverse NSCT. Experiments were conducted using a gray scale multi-focus image dataset, and comparative analysis with commonly used multi-focus image fusion algorithm shows that, the proposed fusion algorithmhas superior performance in terms of visual inspection and 7 commonly used objective evaluation indicators.

Abstract:Aiming at the problems of deep learning-based gesture recognition model with large parameters, slow training speed and high equipment requirements, which increase the cost, a gesture recognition and detection algorithm based on lightweight convolutional neural network is proposed. First, use the Ghost module to design a lightweight backbone feature extraction network to reduce the amount of parameters and calculations of the network; improve the feature fusion network by introducing a weighted two-way feature pyramid network to improve the network detection accuracy; finally use the CIoU loss function as the bounding box regression loss function And add Mosaic data enhancement technology to speed up model convergence and improve the robustness of the network. Experimental results show that the size of the improved model is only 17.9M, which is 92.4% smaller than the original YOLOv3 model, and the average accuracy is increased by 0.6%. Therefore, the new detection method can not only reduce the amount of model parameters, but also ensure the accuracy and efficiency of the model, providing a theoretical reference for gesture recognition and detection.

Abstract:Aiming at the problems of low mapping accuracy and map ghosting and drift in outdoor large scene environment, a simultaneous localization and mapping system integrating filtering and graph optimization theory is proposed. The system consists of three parts: point cloud data preprocessing, filtering based tight coupling inertial odometer and back-end pose map optimization. Firstly, the point cloud data preprocessing uses the random sampling consistency algorithm to segment the ground, extracts the ground model parameters, and constructs the ground constraint factors in the back-end optimization. Then, the front-end tightly coupled inertial odometer adopts iterative error state kalman filter, takes the laser odometer as the observed value and the result of IMU pre-integration as the predicted value, and constructs a joint function to filter and fuse to obtain a more accurate laser inertial odometer. Finally, combined with the graph optimization theory, the closed-loop factor, ground constraint factor and odometer factor matched between frame and graph are introduced as constraints to construct the factor graph and optimize the map pose. The closed-loop factor adopts the improved closed-loop detection algorithm of scanned text for position recognition, which can reduce the environmental false recognition rate. The algorithm proposed in this paper completes scene mapping in multiple scenes such as outdoor plant buildings, parking lots and indoor workshops. The cumulative deviation in the three directions of distance, level and elevation is controlled by about 10 cm, which can effectively solve the problem of map ghosting and drift, and has high robustness and high precision.

Abstract:In view of the visual limitations of the current camera attitude estimation methods. In order to realize the relative attitude estimation of wide-FOV( Fields of view camera), this paper used fisheye lens as visual sensor for attitude estimation. While fisheye imaging has the advantages of a wide-FOV, it is accompanied by serious nonlinear distortions, which leads to the problem of different distortion diffusion at different azimuths and distances. Therefore, this paper proposed a method to directly use the non-linear characteristics of the fisheye image to measure the relative pose of the camera. First, established the fisheye dataset kitti_FE; Secondly, used convolutional neural network for feature extraction and then combined with Long short-term memory network for bidirectional loop training to achieve the end-to-end output of the relative posture of the camera; Finally, the method of transfer learning was used to estimate the pose of the fisheye camera in the actual scene. Experiments show that the proposed method is 32%、29% and 25% higher than the camera pose estimation accuracy under the existing frameworks of CNN 、DeepVO and CNN-LSTM-VO-cons, respectively, and the proposed method is more stable under high-speed motion.

Abstract:Scanning ion conductivity microscopy (SICM) is capable of achieving micron-scale and nanometer-scale morphology measurements, which has attracted the attention of scholars for research. A bilateral filtering algorithm based on wavelet hierarchical thresholding is proposed to solve the problem that SICM morphology images are vulnerable to contamination, which can affect subsequent applications. For the multi-feature fusion noise of SICM morphological image, pseudo-median filtering is applied to partly process the strong speckle noise in the image, and wavelet threshold denoising and bilateral filtering are organically combined to remove the high-frequency and low-frequency noise of the image. Finally, a morphological image with good denoising effect is obtained. The algorithm is verified several times through the simulation experiment and the real test experiment, and its peak signal-to-noise ratio improvement is greater than 9.8% when comparing the three denoising algorithms of median filtering, bilateral filtering and wavelet denoising. The experimental results show that this algorithm has more advantages in the denoising of SICM morphological images.

Abstract:Gait recognition refers to the technology of identity verification by identifying the walking posture of pedestrians. Different from the physiological characteristics such as fingerprints and palmprint that need close contact, gait, as a behavioral feature, has the characteristics of high non-invasive, low camouflage and long-distance recognition. Therefore, gait recognition has broad application prospects in various fields. This paper proposes a gait feature recognition method based on capsule network, and introduces spatial attention mechanism in the capsule network to improve the weight of effective gait features in the capsule, and updates the input image through the design of feedback weight matrix to improve the performance. The designed gait recognition model based on capsule network has been tested on casia-b dataset. The average recognition rate is 93%, 85% and 67% respectively under three different walking conditions: normal walking, walking with bags and walking with coats. At the same time, a multi view gait recognition experiment is carried out on the ou-mvlp data set, and the average recognition rate reaches 85%.

Abstract:In order to achieve a good balance between infrared and visible image information, this paper proposes a deep multi-classification method of infrared and visible image fusion based on generative adversarial network technology. In this method, the idea of principal and auxiliary is introduced into the gradient and intensity information extraction of generator, and the depth and shallow network information extraction ability of generator convolution layer is improved. In the discriminator, multiple classifiers are used to estimate the distribution of visible and infrared regions simultaneously. After continuous face-off learning, the fusion results have remarkable contrast and rich texture details. The obtained information entropy and Shannon entropy value are 6.86, mutual information value is 13.72, standard deviation value is 34.82 and structural similarity value is 0.71.The experimental results show that the proposed method achieves better performance of infrared and visible image fusion in subjective and objective evaluation.

Abstract:Aiming at the problems of low spectrum utilization, large equipment volume, poor reliability and complex design in missile borne telemetry communication system, this paper designs a zero intermediate frequency telemetry transmitter with high spectrum utilization CPFSK(Continuous Phase Frequency Shift Keying) modulation. Firstly, according to the CPFSK modulation mode, the framed 2Mbps telemetry data signal is subjected to orthogonal modulation with an index of 0.715. The modulated data is processed by interpolation filtering and upper edge frequency within the AD9364 chip, and then transmitted by the antenna through the amplifier circuit. Finally, the spectrum, transmission distance and bit error rate of the radio frequency modulation signal are tested by the test system. It is found that the output signal has the advantages of narrow bandwidth, concentrated energy of the main lobe spectrum and faster attenuation of the sidelobe, and the transmission distance can reach more than 20Km. The results show that the designed missile-borne telemetry transmitter has the characteristics of small size, low complexity, high reliability and high spectrum utilization. It not only meets the performance index of missile-borne telemetry system, but also can be applied to future aerospace telemetry and remote control system.

Abstract:Blood pressure is an important indicator of people's health. With the increasing distribution of hypertension, continuous blood pressure monitoring becomes more and more important. This paper presents a method for continuous noninvasive measurement of blood pressure based on long-term recursive convolution network. Firstly, the pulse wave signal collected by optical capacitance product method is normalized, threshold processing and feature extraction, and then the blood pressure is calculated from the pulse wave by long-term recursive convolution network. The experimental results show that when the pulse wave signal of optical capacitance product is directly input, the average absolute error and mean square error of the method are increased by 56.00% and 73.25% respectively. When the characteristic parameters are used as input, the average absolute error and mean square error of the experiment are increased by 59.55% and 87.41% compared with the direct input of optical capacitance product pulse wave signal. Compared with the direct input, the experimental effect of characteristic parameter input is better, and the accurate measurement of blood pressure is realized.

Abstract:Rail is the infrastructure of railway traffic, and magnetic leakage technology is commonly used to detect the surface defects of ferromagnetic materials such as rail. For the problem of the leakage magnetic detection signal, a correlation-based filtering algorithm is proposed. The detection data is segmented according to the damage size, inspection speed and sampling speed, and then the amplitude of the lift-off interference is determined by comparing the correlation of the sampling data of the adjacent magnetic sensitivity sensor x direction and the relative size of the sampling data of each magnetic sensitivity sensor x and z direction. The experimental system is constructed to detect the artificial and natural defects of the rail surface, and the experimental results show that the proposed algorithm can effectively inhibit the lift-off method.,the SNR gain is above 1.7.

Abstract:To solve the problem of low identification accuracy of two-phase flow pattern, a flow pattern identification algorithm based on two-channel hybrid network fusion support vector machine was proposed. Multi-scale feature extraction of rich feature layer information was carried out by multi-scale convolution check of capacitance vector. Squeeze-and-excitation networks was used to focus on the important feature tensor of convolutional kernel channel and adjust the importance proportion of each channel. In addition, multi-scale attention mechanism was introduced to learn the depth feature of capacitance vector. The multi-scale convolutional channel with SENet and multi-attention channel were fused to form a two-channel identification model. Finally, the features of capacitance vector effectively captured by the two-channel model were sent to support vector machine for training and testing. Simulation results show that compared with BP neural network, SVM and 1DCNN algorithms, the average identification rate of the proposed algorithm in flow pattern identification is significantly improved, reaching 98.6%.

Abstract:For the problem of high complexity of high-voltage cable faults and high cost of real-time monitoring, a combined diagnosis method based on ensemble empirical modal decomposition (EEMD) and tennies whisker search algorithm optimized convolutional neural network (CNN) is proposed. The cable sheath current data are decomposed into several eigenmodal components (IMF) by EEMD, and the component with the highest correlation with the original signal is selected by combining the correlation coefficients as the input of the CNN network. In order to improve the classification accuracy of the network model, the hyperparameters of the CNN diagnostic model are optimized using the aspen whisker algorithm (BAS). Taking the high-voltage cable current data of a coal mine in Huainan as an example, the experimental results show that EEMD effectively decomposes the current signal, and the designed BAS-CNN network has the highest classification accuracy with 96.95% monitoring accuracy compared with 2 groups of networks with artificially determined CNN hyperparameters.

Abstract:In view of the problems that the existing fault diagnosis methods cannot identify the long-time dependence relationship and have insufficient precision when processing the observed data in the chemical production process with high dimensions and obvious dynamic characteristics, the long-term memory model is improved in this paper, and a classification model based on depth learning and attention mechanism is proposed, Then the model classification effect is verified and the model in this paper is compared with the model before improvement. Finally, the sample data is drawn and the distribution of feature vectors in two-dimensional space is output at each level of the model by t-sne algorithm. The experimental results show that the improved in-depth learning model can achieve a recall rate of 92.71% and an accuracy rate of 93.05% for fault classification, which are improved by 16.84% and 13.66% respectively compared with the model before the improvement. The learning effect on data characteristics is better, and it is more suitable for chemical data.

Abstract:The health management system is an important support resource for radar, and the rapid assessment of the radar health status is an important part of the health management system. Radar is a complex electronic information system. It consists of a large number of equipment. In the past, the health of radar was calculated by the way that people assign weights. The accuracy of the method is low, and it is difficult to update. This paper proposes a radar health assessment model based on logistic regression, which can be used trained by practice data samples. The accuracy of the model in test samples can reach 83.23%. The model’s average training time is 126ms, which means on-line training and update can be implemented. Through the expert online feedback mechanism, the accuracy of the model in test samples can be promoted to 85.37%. In the daily use of radar, the accuracy of health assessment can be improved continuously by on-line feedback mechanism, in order to ensure the comprehensive support capability of radar life.

Abstract:In the process of electromagnetic parameters testing with waveguide transmission line, the measurement reference plane can not be at both ends of the material being measured, resulting in a large measurement deviation.In this paper, the expression of permeability and permittivity of the measured dielectric material is obtained by constructing a test simulation model using transmission reflection measurement method and electromagnetic parameter inversion algorithm.On this basis, in order to solve the measurement reference plane problem, TRL(Through-reflect-line) calibration algorithm is being proposed,as a result，TRL calibration algorithm reduce the deviation of scattering parameters and to 12% and 16% in theory. In order to solve the embedding deviation problem, the de-embedding algorithm is being proposed,and the de-embedding algorithm can reduce the deviation of scattering parameters and to 18% and 6% respectively.The simulation results show that the two optimization algorithms can effectively reduce the measurement deviation and improve the measurement accuracy of electromagnetic parameters.