Abstract:Energy management strategy (EMS) is a key technology to improve the economy and durability of fuel cell electric vehicles. To overcome the shortcomings of most existing strategies, a new EMS based on quadratic utility function (QUF) is proposed in this paper. Mathematical models of vehicle and components are established, and the accuracy of the models is verified by bench test. QUF is used to decompose the demand power into the output power of fuel cells and batteries, the strategy considers the historical output state and load change rate of the battery and fuel cell. In order to improve the durability of power sources and the economy of the vehicle, the coefficients of the QUF are determined and optimized by using the Nash equilibrium. Simulation results show that the proposed strategy can effectively improve the durability of fuel cells and batteries, reduce hydrogen consumption, and extend driving range. Compared with the FSM strategy, this strategy can reduce the degradation of the batteries by 21.15%. Compared with the fuzzy control strategy, the proposed strategy can reduce the degradation of fuel cells by 36.52%.

Abstract:Compared with the three-coil relay wireless power transmission structure, the two-coil structure has higher efficiency but short distances, while the three-coil structure is mostly used for medium and long-distance wireless power transmission, the efficiency is low in the short-distance system. For electric vehicles with different chassis distances, resulting in low charging efficiency, a U-shaped coil switching system based on vertical relay is presented here. The system switches at a suitable distance point, that is, the optimal efficiency can be reached at different transmission distance, the problem of low efficiency caused by changes in spacing is solved. This paper has analyzed the three-coil circuit model, its coil characteristics, load characteristics and energy transmission distance characteristics are analyzed and optimized, efficiency influence factors have been deduced, and system simulation analysis are performed. The simulation results show that compared with the single two-coil or three-coil structure model, the designed switching system improves the wireless power transfer efficiency by 58% at both near and far distances compared with the two-coil structure, and compared with ordinary three-coil structure, the system has increased to its highest efficiency of 71%.

Abstract:The semi-airborne transient electromagnetic realizes the detection of underground structures by arranging loop source on the ground and using UAV equipped with receiving coils to get signals. However, the response will be affected by the flight height and reception offset in the actual application. Aiming at this problem, the research on the spatial distribution of semi-airborne transient electromagnetic response is carried out. Firstly, the influence of different factors on the electromagnetic response is analyzed by using the method of control variables with the forward calculation of the uniform earth model and the layered earth model; Secondly, through the forward calculation of the model with abnormal body, the influence of parameters such as the buried depth and resistivity of the abnormal body on the ability of the semi-airborne transient electromagnetic exploration underground target body is analyzed. The results show that the effects of flight height and offset on the semi-airborne transient electromagnetic response are mainly concentrated in the early stage. With the increase of the flight height or the offset, the amplitude of the early vertical induced electromotive force gradually decreases, and the polarity reversal phenomenon occurs in the external response of the loop source. The response curve shows different decay rates in different resistivity formations, which indicates that it has a good identification effect on the formation resistivity change. The response curve can reflect the buried depth and resistivity of underground target body, which proves the ability of semi-airborne transient electromagnetic to distinguish different objects.

Abstract:Aiming at the problem that the traditional graph convolutional network ignores the relationship between spatial and temporal features, a dual-stream network model based on the combination of residual structure and graph convolutional network is proposed. First of all, the network includes two channels of space flow and time flow. The gesture skeleton information is constructed into a space diagram and a time sequence diagram as the input of the two channels. The training speed is greatly improved by separating the time dimension and the space dimension. Then, in order to increase the depth of the network and avoid problems such as the disappearance of gradients, the residual structure is embedded and improved to make more effective use of time features and ensure the diversity of features. Finally, the spatial point set sequence and the time edge set sequence output by the two channels are converted in series and input into the Softmax classifier for classification, and the recognition result is obtained. The newly proposed method is tested on the CSL and DEVISIGN-L gesture datasets, and the results show that the recognition accuracy on the two datasets reaches 96.2% and 69.3%, which proves that the method has a certain degree of advancement.

Abstract:In order to meet the requirements of real-time monitoring output voltage and output current of power supply and distribution control equipment of measurement system. This paper designs a high-precision current measurement circuit, which is based on TI's precision current monitor INA260, which can realize real-time monitoring and accurate measurement of 0-15A continuous current in the 0-36V power supply range. The current measurement value is output through the I2C interface, and the signal measurement circuit is electrically isolated from the digital control circuit. Through a large number of data tests, the error range between the measured output voltage and the theoretical value is 0~10mV, and the error range between the measured output current and the theoretical value is 0~5mA. The experimental results show that the application circuit's current measurement accuracy within 0~15A and voltage measurement accuracy within 0~36V are better than 1‰, realizing the function of real-time monitoring and improving the measuring precision.

Abstract:Aiming at the strong real-time demand of image processing in the focus detection system of projection lithography machine, a focus detection image real-time processing system based on SOPC is designed with FPGA chip as the carrier. Firstly, the MicroBlaze embedded soft core is used to execute the image processing algorithm, and the system scheme verification and algorithm debugging can be carried out without changing the hardware architecture; Then, using the scheme of software and hardware cooperation, the algorithm is transplanted to FPGA hardware to accelerate the focus detection process. The FPGA chip carries out data interaction between software and hardware through Axi bus, and the off chip is interconnected with the worktable through optical fiber interface to complete the high-speed upload of focus detection results. The experimental results show that the system runs stably, and it takes 104 μs to complete the entire focus detection process, and the software-hardware collaboration is about 1700 times faster than the soft-core implementation.

Abstract:Aiming at a structural stress monitoring system, it is proposed that using the principle of fiber grating sensing to measure structural stress.A F-P tunable filter fiber grating demodulation system based on fiber etalon is designed. Using the stability of the etalon to calibrate the F-P filter to eliminate the interference of the ambient temperature on the F-P filter. The driving source of the F-P filter is designed by using the DDS principle, and the resolution of triangle wave can be less than 0.42mV, ensuring that the final resolution of the demodulation system reaches 1pm. The photoelectric conversion circuit is designed to realize the conversion and amplification of the weak light signal, and finally it is converted into a voltage signal that is easy to collect and display. Finally, the demodulation system designed in this paper can achieve a strain measurement of ±3000με, and the measurement accuracy can reach 2.59%.

Abstract:In view of the problems of poor measurement reliability and low accuracy caused by traditional pull-wire displacement sensors due to collisions, bad weather, etc., a virtual displacement sensor for excavator working devices that establishes the mapping relationship between the cylinder displacement length and the pixel coordinates of the marking point through a neural network is proposed system. Using image processing technology to extract the center pixel coordinates of the hydraulic cylinder marking point, taking the pixel coordinates and the actual cylinder displacement signal as input, and establishing the mapping relationship between the cylinder displacement and marking point center coordinates through the neural network optimized by genetic algorithm, predicting the cylinder displacement and obtaining mining The attitude of the machine working device. Experiments show that the accuracy of the cylinder displacement predicted by this method is as high as 99.5%, and the predicted mean square error of the working device's attitude is 1.1329, which meets the requirements of practical applications and can be used in the actual measurement of the excavator's attitude.

Abstract:When the sampling covariance matrix contains the desired signal, the model mismatch will greatly reduce the performance of MVDR (minimum variance distortion response) adaptive beamforming algorithm. A robust adaptive beamforming algorithm based on spatial smoothing difference algorithm (SSDA) is proposed in this paper. The algorithm first obtains the prior knowledge of the number of sources through the spatial smoothing difference algorithm, then searches the wave crest of the MVDR spatial power spectrum to obtain the expected signal, and finally removes the expected signal covariance matrix from the sampling covariance matrix. The simulation results show that under the condition of high signal to noise ratio input, the output signal to interference plus noise ratio of the improved algorithm has been greatly improved after removing the expected signal, which is 2 ~ 3 times higher than that of MVDR algorithm.

Abstract:Due to the strong nonlinearity of the working process of compression-ignition aero piston engine, only using Model Predictive Control (MPC) algorithm to realize the torque control of compression ignition aero-piston engine would lead to the unsatisfactory accuracy of torque prediction based on state space model. The above problems could be solved by the compound predictive control of engine torque based on Radial Basis Function (RBF) neural-network and MPC. Firstly, the engine operation data obtained by the MAP control method were used as the neural network training sample-set on the self-built engine joint simulation platform. Secondly, the Simulated Annealing (SA) mechanism introduced into particle swarm optimization (PSO) algorithm was to train the RBF neural network torque prediction model. Finally, through joint simulation, the control algorithm was continuously modified to verify the superiority of SA + PSO algorithm in training engine torque prediction model on RBF neural network. Also it could effectively improve the problem that PSO algorithm was easy to fall into local optimization. The effectiveness of torque compound predictive control was verified by the engine bench experiment.

Abstract:With the rapid development of microgrid, its operation and economic optimization control have attracted more and more attention. Traditional economic optimization control is centralized control, which is faced with single point of failure, large resource consumption and other problems. Distributed control can solve the above problems well, and the economic scheduling problem in microgrid can be solved by combining consensus algorithm. The system scheduling process requires the communication network to complete the information interaction, but the communication network is usually affected by the noise. For this problem, a fully distributed economic scheduling strategy based on finite time consensus is proposed to deal with the interference of noise, and in a distributed system of minimum power generation cost and environmental pollution as the goal . By introducing consensus gain function to suppress noise, the completely distributed environment/economic optimal scheduling is realized. The simulation results show that the fully distributed scheduling strategy proposed in this paper can suppress the interference of noise and achieve good optimization effect while ensuring consistent convergence.

Abstract:Aiming at the problem that the traditional speech enhancement network is not ideal for unknown noise enhancement, this paper proposes an improved method from the aspects of spectral enhancement, network structure and feature fusion mechanism. Firstly, in order to extract the deep feature information of the spectrum, VGG19 structure was used to replace the encoder part of UNet structure, and residual network was added to the decoder part to deepen the network depth and prevent the training degradation. Secondly, in order to better combine the feature information in the spectrogram, an adaptive feature fusion mechanism is added to the jump connection part of THE UNet structure to fuse the deep and shallow features. In addition, in order to enhance the speaker information, the histogram equalization algorithm is used to optimize the feature of the spectrogram, and the histogram equalization enhancement spectrogram is obtained. In different noise environments, the proposed method outperforms other enhancement methods in terms of quality and comprehensibility.

Abstract:Aiming at the problem that the search path of A* algorithm is not smooth and does not have dynamic obstacle avoidance in 3D environment, this paper proposes a fusion A* algorithm. Based on the A* algorithm, this algorithm first introduces pitch angle and yaw angle as search constraints, which shortens the time for path planning, and secondly uses variable weight evaluation functions and UAV range, flight height, and threat cost Finally, the smoothed A* algorithm is combined with the artificial potential field method, and the particle swarm algorithm is used to optimize the parameters involved in the A* algorithm and the artificial potential field method. The simulation results show that compared with the traditional A* algorithm, the fusion algorithm saves 5.34% of fuel consumption, improves the search efficiency, shortens the path length, the planned path is smoother, and can achieve real-time dynamic obstacle avoidance.

Abstract:The harmonic detection method based on Complementary Ensemble Empirical Mode Decomposition (CEEMD) is susceptible to the interference of iterations and auxiliary white noise, resulting in the defects of false component and mode mixing. And CEEMD method has the defect of low precision in detecting harmonic signal under noise background. To solve the above problems, a new PE-CEEMD harmonic detection method was proposed based on the combination of Permutation Entropy (PE) algorithm and CEEMD. Firstly, the harmonic signal is decomposed by CEEMD to obtain a series of Intrinsic mode-functions (IMF) with frequency from high to low. The PE algorithm was used to quickly select and eliminate the noise components with large randomness, and then CEEMD decomposition was performed for the remaining signals.  Simulation results show that PE-CEEMD method can overcome the problems of modal aliasing and false components better than CEEMD method, and the detection accuracy of frequency components and amplitude of complex harmonic signals is improved by 4.424% and 9.3%, respectively.

Abstract:Machine olfaction is an emerging bionic technology based on sensor arrays and computer algorithms to simulate biological olfaction. The characterization of odor substances is a field worthy of research in machine olfaction. At present, olfactory perception is in the preliminary research stage, and the general classification theory of odor is not yet mature. . In this paper, starting from the electronic information of material odor, aiming at the relatively balanced fragrance data in the collected data, using machine learning algorithms and parameter adjustments, grid search and other model optimization methods, the material odor classification model based on electric nose data is proposed, and the connection between the information and perception of the material odor electronic nose is established. The experimental results show that the random forest model performs better than other machine learning algorithms in each evaluation index, and the average accuracy of odor classification based on random forest reaches 93.6%.

Abstract:Image super-resolution (SR) reconstruction refers to the generation of a corresponding high Resolution image from a low Resolution image. SR has important application value in monitoring, remote sensing, digital HD, video coding communication and other fields. In this paper, we first review the history of single image SR and summarize the non-learning based SR methods. Among them, interpolation and reconstruction based methods are introduced, and then learning-based methods are introduced, in which SR based on deep learning is analyzed such as SRCNN, ESPCN and SRGAN are summarized in detail, and compared with recursive structure, dense structure and attention mechanism network structure. Then the function of loss function and upsampling method in image SR is analyzed, common data sets and image evaluation indexes are introduced, and visualization results of image SR are displayed. Finally, the progress and deficiency of single image super-resolution technology are summarized.

Abstract:Aiming at the problems of low definition, poor contrast, and blurred electrode contours in lithium battery X-ray images, an X-ray image enhancement algorithm for lithium batteries based on improved multi-scale Retinex was proposed. First, in the traditional multi-scale Retinex algorithm,bilateral filtering is used to estimate the luminance component, while image dynamic range compression is performed globally adaptively based on the mean logarithmic luminance value. Then use the improved MSR algorithm to extract the reflection component of the image, use the sobel operator to obtain the longitudinal gradient of the reflection component, and then fuse the gradient information with the reflection component to enhance the image details, and then use the CLAHE algorithm to enhance the contrast of the fused image, and finally use Bilateral filtering is used to remove noise to obtain the final enhanced image. The experiments are carried out on the self-constructed data set. The experimental results show that the proposed method can significantly improve the clarity and contrast of X-ray images of lithium batteries, and the edge contours of the cathode lines in the images are significantly enhanced. It is significantly better than the traditional multi-scale Retinex algorithm in highlighting the edge details of lithium battery X-ray images and enhancing image contrast.

Abstract:Aiming at the problem that gait recognition is vulnerable to environmental interference, this paper focuses on the gait feature extraction method, an improved attitude estimation algorithm is proposed based on the anti learning network framework to extract gait features. The improved residual network is used to obtain the gait features from low level to high level. With the deepening of the number of network layers, the residual network is adjusted accordingly to highlight the focus on the local detail feature information; A timing encoder is designed, which not only improves the generalization of gait features to environmental changes, but also reduces the impact of environment on feature extraction. Finally, a large number of experiments are carried out based on CASIA data set under three different experimental modes, the recognition accuracy is more than 83%, which finally proves that the feature extraction method proposed in this paper shows good flexibility in complex environment.

Abstract:Aiming at the problem of 3D irregular point cloud format and uneven density, a fusion of multi attention mechanism and pointrcnn network is proposed for 3D point cloud target detection. This experiment mainly improves the pointrcnn two-stage network respectively. Firstly, the channel attention and spatial attention mechanism are serially input to the distribution of each network layer in the first stage by adjusting and normalizing in batch to further quickly identify three-dimensional features; Secondly, the cross position attention mechanism is introduced into the second stage network to avoid the position deviation of the cross path, so as to further refine the three-dimensional target position for feature extraction. The experimental results on Kitti data set show that compared with pointrcnn detection network, the improved network improves the average mean accuracy (map) of car and pedestrian tests by 1.2% and 1.9% respectively. Therefore, the improved method not only solves the problems of irregular point cloud format and uneven density, but also ensures the detection accuracy.

Abstract:With the rapid and comprehensive development of China's economy and the improvement of people's living standards, China’s demand for transportation is increasing. In order to meet the higher requirements of locomotive operation safety, a visual algorithm of the advanced driver assistance system of trains based on single-stage instance segmentation in railway scene is proposed. The algorithm model is optimized for detecting the characteristics of multiple overlaps of objects in railway scenes. The accuracy of the model is improved by improving the Backbone network and multi-scale fusion method. The model is accelerated by TensorRT semi-precision acceleration and CUDA code refactoring. The performance evaluation and comparative test of this method and other methods are carried out. Finally, this method achieves 71.2MAP and 108ms on the embedded platform Xavier. High-precision detection of the surrounding environment of the train under vehicle deployment is realized.

Abstract:To realize online detection of solder ball quality and coplanarity in Ball Grid Array chip visual inspection, a detection method combining point cloud and image data is proposed. First, the color camera has been calibrated, the chips color image and point cloud data were acquired, and the unit normal vector of the main plane in the workpiece is achieved by Random sample consensus. The point cloud is rotated to fulfill the tilt of the point cloud alignment. Then, the rotating point cloud is projected and imaged by the color camera’s internal and external parameters, to produce the point cloud grayscale image, which is matched with the color image in point mode, and the translation amount and deflection angle between the point cloud data and the color image are obtained to complete the point cloud. In this way, the point cloud data is registered with the color image data. The solder ball is extracted from the registered data to determine its size, breakage, deformation, bridge, and other faults, as well as to measure its height and complete the solder ball’s coplanarity measurement. To perform comparison trials, this method was compared with the ICP algorithm. Two measurement results show that the proposed method can save 62.8% of the time, which also has consistent detection results. The proposed method for 3D point cloud has a wide application prospect in advanced manufacturing.

Abstract:Aiming at the problem of insufficient positioning accuracy of traditional DV _ Hop algorithm in wireless sensor networks, an improved DV _ Hop positioning algorithm based on ranging correction and badger optimization is proposed. Firstly, the minimum hops between nodes are refined by multi-communication radius; secondly, the minimum mean square deviation criterion and correction factor are used to reduce the jump error; finally, the improved badger algorithm with excellent global optimization performance is introduced to replace the least square method to calculate the unknown node coordinates, which further reduces the calculation error. The network simulation results show that under different conditions, the positioning error of the optimization algorithm decreases by 16.62 % and 3.92 % on average compared with the traditional DV_Hop algorithm and the improved algorithm (PDDV_Hop), which can effectively improve the positioning accuracy. The positioning accuracy of the optimization algorithm is less affected by the number of anchor nodes, which can reduce the deployment cost of anchor nodes on the premise of ensuring the positioning accuracy.

Abstract:Optical fiber geometric parameters are an important index that affects optical fiber performance. The gray-scale method is a common method for measuring the geometric parameters of optical fibers. When measuring, it is necessary to illuminate the optical fiber to distinguish the core and the cladding. Since the light is not completely concentrated in the core propagation (part of the light propagates in the cladding), it is difficult to distinguish the interface between the core and the cladding. In order to accurately find the edge of the fiber core, this paper uses two polynomials to fit the light intensity gray distribution of the core and cladding area respectively, and obtains the gray value corresponding to the intersection of the two polynomials as the boundary point between the core and the cladding. So as to get the edge data of the core. The segmented cubic Hermite interpolation is used to correct the measured data, reducing the influence of error points on the fitting. By measuring two sets of fiber end-face images with different imaging qualities, the diameter and out-of-roundness of the fiber core were measured with a standard instrument to be 10.068µm, 0.616%, and 10.397µm, 0.766%. The measured value of the method in this paper is 9.999µm, 0.716% and 10.020µm, 0.857%. Experiments show that the method in this paper has better accuracy and stability for the measurement of optical fiber geometric parameters, and theoretically, the method in this paper has more physical meaning and rationality in the measurement principle than the commonly used gray-scale method.

Abstract:In order to explore the connectivity characteristics of the brain network of patients with depression and its feasibility as an online feedback indicator. First, the brain network is constructed using the imaginary part of coherency (IC) that is not sensitive to the volume conductor effect, this can effectively and conveniently avoid the influence of false connections. Then, the IC value with significant difference is extracted as a feature set, and a combination of Couple entropy (CE) and Relief filtering feature selection method is proposed to optimize the feature set, and the relationship information between features and classes, features and features are combined to improve the quality of feature sets. At the same time, according to the self-referencing brain network module integration feature set, online feedback indicators are constructed. Finally, K-nearest Neighbor (KNN) and support vector machine(SVM) classifiers are used for comparative analysis. The results found that the feature set extracted by the CE-Relief feature selection method in each frequency band is the smallest, and the classification accuracy is higher than 90%; the IC value of the Alpha frequency band has the best effect in identifying depression, and the classification accuracy can reach 100%; the classification ability of the average IC value of the prefrontal area of the self-reference brain network has advantages and stability in each frequency band, and the classification accuracy is higher than 80%.

Abstract:The identification of individual radiation sources is an important technology in the field of electronic countermeasures. By identifying different subtle features between devices, the purpose of distinguishing illegal devices from legal devices is achieved. Aiming at the problem of subtle differences in fingerprint features between individual radiation sources and fewer features extracted under noise interference, this paper proposes a method of identifying individual radiation sources based on a deep residual shrinkage network. This method first splices the feature data of the I/Q map, uses data enhancement technology to expand the sample, and then constructs a deep residual shrinkage network recognition model. Finally the constructed model is trained for individual ADS-B radiation source recognition and the recognition effect is evaluated. The simulation results show that the deep residual shrinkage network constructed in this paper uses the advantage of eliminating data noise, and the overall recognition accuracy of the 20 types of ADS-B radiation source individuals after data enhancement has reached 98.2% when the SNR is as low as 0 dB.Compared with the Resnet network with the same number of layers, its performance is improved by 1.3%, and it is significantly better than other existing methods.