Abstract:Aiming at the problem that UWB sensor measurement errors are caused by electrical conversion time delay and antenna delay in UWB technology, a UWB combined positioning algorithm based on graph optimization and EKF is proposed in this paper. Firstly, EKF algorithm is used to obtain the initial positioning value, and then the objective function including UWB ranging error is constructed. Then, the graph optimization algorithm was used to solve the base station location that minimized the overall location error of UWB in the whole positioning process. Then, the solved base station location was substituted into EKF algorithm for the second calculation to obtain more accurate location results. Finally, Mean-Shift algorithm was used to perform cluster analysis on the location results. Experimental results show that compared with the EKF algorithm and the least square method, the horizontal positioning accuracy of the proposed algorithm is improved by 27% and 38% on average, and the obtained target trajectory smoothness is much better than the traditional positioning algorithm.

Abstract:Aiming at the problems of low safety, slow thermostat adjustment, poor effect and low intelligence of the heating equipment currently on the market, a heating equipment control system based on fuzzy PID was designed. The system completes the acquisition of temperature and humidity as well as intelligent interactive control functions through sensor arrays and power adjustment modules. At the same time, the system uses the fuzzy PID control algorithm to achieve precise control of the ambient temperature, which can make timely prediction and processing of some behaviors that cause potential safety hazards. After MATLAB system simulation and actual test, the fuzzy PID control system can not only achieve no overshoot basically, but also have high robustness. When compared with the traditional PID, both the oscillation period and adjustment time are shortened by 46% and 25% respectively. In addition, the system can also handle all kinds of potential safety hazards well, which greatly improves the safety during heating and has a good market prospect.

Abstract:The accurate measurement of the rolling attitude of the ammunition is the premise of precise guidance. In the common attitude measurement schemes that combining satellite navigation and magnetometer, large-scale maneuvering of the ammunition can cause the large angle measurement error. In order to solve the above problems, a new combined attitude measurement scheme of magnetometer and satellite navigation based on BP neural network calibration is proposed. For the same type of missile, the control command, ballistic speed, historical attitude angle and other parameters are used to estimate the sideslip angle information of the missile body in real time. Simultaneously, the declination angle of the missile body calculated by the satellite navigation system is calibrated online. Finally, the rolling attitude of the projectile is measured by fusing the measurement information of the magnetometer, and the accurate rolling attitude estimation is provided through the fitting and filtering scheme for the real-time control of the ballistic trajectory. The numerical simulation results show that for the same type of ammunition system, the sideslip angle information error of the projectile that estimated by the trained BP neural network is within 5% of the full scale. The roll attitude angle accuracy obtained by the fitting and filtering after the combination of the satellite positioning and navigation system and the magnetometer is within ±1°, which greatly improves the roll angle measurement accuracy of the ammunition during maneuvering.

Abstract:This paper presents a seamless switching control strategy based on virtual synchronous generator (VSG) control and voltage and current internal model control (IMC) for voltage distortion and impulse current in the switching process betweeen grid-connected and off-grid of the Inverter. Firstly, the control structure and mathematical model of the inverter are established.Secondly, the control structure with voltage and current double-loop based on IMC is designed and the parameters of internal model controller are adjusted. Thirdly, the pre-synchronization control method based on amplitude and phase of the voltage is adopted to enable the inverter to switch smoothly between the two models, and the transient voltage and current of switching is stable, which realizes the smooth transition. Finally, the effectiveness of the proposed control strategy is verified by the Matlab/Simulink simulation.

Abstract:Goal Aiming at the problem that the image of motor vehicle registration certificate is printed incorrectly and the characters are crossed by lines, a set of character recognition algorithm is designed based on Halcon machine vision software. Method Firstly, the image is tilt corrected、enhanced、binarized; Extraction of table lines by specific structure corrosion method, the original image area is divided into two parts: table and character; Based on the vertical projection segmentation method and the character spacing merging algorithm, a single character is segmented, in which the characters crossed by the table line will have obvious fracture. The region detection and filling algorithm specifies the structural elements and parameters to detect and fill the internal and edge fractures of the region to realize fracture splicing. Finally, the spliced characters are trained and recognized based on multi-layer perceptron, and the results are typeset and displayed in the table. Result Experimental results show that the algorithm can preprocess the registration certificate image and recognize characters, and the recognition accuracy is 97.28%. Conclusion Through the above algorithm, the efficiency of extracting information from complex background can be greatly improved.

Abstract:The probe volume of medical endoscope is limited, the use environment is special, and the lighting conditions are limited, so the dark part processing of image is particularly important. This design proposes a fast dark part enhancement algorithm for color image. Firstly, the collected RGB888 format video stream images are channel separated, and then the mean filter of specific window size is carried out for each channel through convolution to extract the regional features. Finally, the average value of each corresponding point of the three channels after filtering is taken and substitute it into the Logistic function of specific parameters. The result is the gain of the pixel point at that position of the frame image, which is applied to the original image. The design takes the Zynq series ARM+FPGA SoC platform launched by Xilnx as the carrier, uses Vivado HLS to develop the AXI-Stream interface video stream processing algorithm, and generates IP to run on the FPGA. After experiments, this algorithm only needs 1.6 milliseconds to process a frame of 400*400@30fps on the Zynq7020 platform, which ensures the real-time output of the video stream. At the same time, the endoscope is inserted into the oral cavity to observe the video stream images before and after the algorithm processing.The algorithm enhances the brightness of the dark part while ensuring the regional contrast, and improves the dark part quality of medical endoscope video stream.

Abstract:To address the problems of the Rapidly-exploring Random Trees (RRT) algorithm in mobile robot path planning including time cost, long and tortuous path, and poor smoothness, an improved RRT algorithm of new sampling node generation and path node selection is proposed. The algorithm combines the probability of sampling points and the gravitation of the target point, and dynamically changes the random step size, thus accelerating the expansion to the target point and reducing the planning time. Finally, to complete the path optimization, the path nodes obtained after the forward optimization and the secondary selection are processed using the B-spline function, which comprehensively improves the path in terms of length and smoothness. The proposed algorithm has been compared with traditional RRT algorithm and P-RRT algorithm in the simulation experiments, whose results show that the proposed algorithm has improved the path length, planning time and the number of path nodes to a certain amount, and effectively the path smoothness.

Abstract:Aiming at the problem that the traditional Census algorithm is too dependent on the center pixel, which is susceptible to noise, and the AD-Census algorithm can not make full use of the advantages of different algorithms, this paper proposes an improved Census transformation and adaptive weight stereo matching algorithm. Firstly, the mean value of the Census transform window and the pixel information of the center point and neighborhood in four directions are used to automatically classify the close pixels into one class, which improves the robustness of the Census transform against noise. Secondly, the SAD algorithm and Sobel edge detection are introduced, and the weight of SAD and Census transform is determined according to the gradient information, which improves the adaptability of the algorithm in different regions. Finally, the final disparity map is obtained by the cost aggregation method of the cross-domain and subsequent optimization. The parallax maps of different images are verified on the Middlebury platform, and the average error of the proposed algorithm is 9.33%, which is 3.39% lower than the AD-Census algorithm. Compared with other algorithms, the algorithm has better matching accuracy in the parallax discontinuous region and repeated texture region, and better robustness against noise and light.

Abstract:In this paper, the existing commonly used lithium-ion battery models are analyzed, and a second-order RC network equivalent circuit model which is convenient for engineering application is established. The corresponding battery models are built in MATLAB, and the measured data are used to identify the battery model parameters offline, and the accuracy of the model is verified. Considering that the values of model parameters are not constant in the process of battery charging and discharging, but constantly change due to factors such as charging and discharging rate and battery SOC, in order to improve the accuracy of the model, the recursive least square method with forgetting factors is adopted to identify the model parameters online, and the optimal range of forgetting factors is determined by comparing the influences of different forgetting factors through simulation analysis. The experimental results show that as the forgetting factor decreases from 1, the accuracy of the model will first increase and then decrease. The appropriate forgetting factor range of this model is about 0.90~0.95, and the best value should be around 0.94. At this time, the average voltage error of the model is only 0.00043V, which proves the correctness and high accuracy of the identification method in this paper.

Abstract:blood pressure is an important physiological index of human body. It can judge the cardiovascular function and heart condition of the body. Many diseases are closely related to blood pressure. Therefore, the correct determination of blood pressure is of great significance for the diagnosis and treatment of cardiovascular diseases. We proposed a noninvasive blood pressure measurement method based on BiLSTM network. Firstly, taking the BiLSTM network and the traditional LSTM as the experimental model, and comparing the output evaluation index coefficients, it is found that the BiLSTM network has a better effect on blood pressure measurement. Because the attention mechanism can assign weight coefficients from rows according to the importance of features, it is introduced into the BiLSTM network with good measurement effect for experiments. According to the results, it is found that compared with the original BiLSTM model, the MSE value and Mae value of the introduced attention mechanism model are greatly reduced by 18.29% and 21.27% respectively, and the R-square value is increased by 0.17%.

Abstract:In view of the difficulty of feature extraction and the low accuracy of fault type recognition in rolling bearing fault diagnosis, a fault diagnosis method based on Improved Complete Ensemble Empirical Mode Decomposition with adaptive noise (ICEEMDAN) and Multi-scale Permutation Entropy (MPE) combined with Aquila Optimizer (AO) to optimize the regularization parameters and kernel parameters of Least Squares Support Vector Machine (LSSVM) is proposed. Firstly, the original vibration signal of the bearing is decomposed by ICEEMDAN. Secondly, according to the double principles of correlation coefficient and variance contribution rate, the eigenmode component (IMF) that meets the standard is selected, and the MPE of the corresponding component is calculated to comprehensively obtain the fault characteristic information; Finally, the multi-dimensional feature vector is formed, and the bearing fault diagnosis is realized by using AO-LSSVM identification model. At the same time, several groups of comparative experiments are carried out. The results show the superiority of the proposed method in rolling bearing fault diagnosis, and the recognition accuracy can reach 98.95%.

Abstract:To solve the problem of low resolution of infrared target images, lack of texture details, and low detection accuracy caused by complex background interference, an infrared target detection algorithm based on improved YOLOX is proposed. First, an effective spatial channel mixed attention module is introduced into the feature extraction backbone network CSP-Darknet53 to reduce the accuracy loss of the network due to long-distance transmission; secondly, in order to further improve the detection accuracy of infrared targets, based on the original enhanced feature extraction network PANet, an improved path feature fusion method is proposed; finally, in order to solve the problem of low recognition rate of small objects in infrared targets, a deconvolution operation is performed at the YOLOX output detection-head to expand the output feature map. Experiments are carried out on the FLIR infrared public data set. The experimental results show that the mean Average Precision (mAP) of the proposed algorithm recognition reaches 91.00%, which is 5.04% percentage points higher than that of the benchmark YOLOX network, it is effective to improve the detection accuracy of infrared targets.

Abstract:Aiming at the problems of complex multi-target image detection scenes and redundant target position data with different length, width and height, the neural network algorithm can effectively improve the accuracy and stability of parallel detection of different types of targets. A multi-target detection method based on the improved YOLOv5 network is proposed. First, according to the spatial scale of different objects, the feature fusion method of the model is improved, and a multi-scale feature detection layer is added to reduce the error of multi-target detection. At the same time, Adaptive Feature Adjustment module is added to reduce the false detection rate and missed detection rate of the network; then K-means++ algorithm is used to estimate the candidate frame to obtain better frame parameters; finally, Efficient IOU Loss is used in the loss function for optimization. Experiments show that the mean average precision of the improved method reaches 76.48%, which is 3.2% higher than the classic YOLOv5 network, and the average detection accuracy of small-sized objects increases by 6.3%. The improved method network continues the lightweight and high-efficiency of the YOLOv5 network, obtains better detection accuracy for multi-scale target detection and can achieve more accurate real-time multi-target detection.

Abstract:In order to solve the problems of single detection category, low detection accuracy and difficult detection of complex objects, a cascaded bag detection method integrating convolution and transformer is studied. CT-CBDet First, a deformable conformer is designed as a backbone network for feature extraction, which uses deformable convolution and spatial pyramid pooling modules to achieve geometric feature transformation and multi-scale feature fusion on the basis of the fusion of transformer and convolutional double network. feature modeling ability; then, a region proposal network with adaptive positive and negative sample selection based on anchor statistical features is proposed to balance the fairness of positive and negative selection of object samples at different scales and enhance the training stability of the model; finally, the cascade detection component of the model is trained end-to-end using multi-stage loss. The results show that the method improves the mAP by 5.8% and the small-scale object detection accuracy by 10.9% compared to the baseline method Cascade RCNN. It can be seen that CT-CBDet can effectively perform the bag detection task in complex scenes.

Abstract:Printed circuit boards are the core and most basic components of many electronic products, and their defect detection has the characteristics of high complexity and small defect targets. An improved YOLOv4-tiny printed circuit board defect detection method is proposed to meet the detection speed. On the premise of improving the detection accuracy. First, a spatial pyramid pooling module is added to the backbone network to reduce network parameters and improve network prediction speed while using the local and global features of the image to fuse multiple receptive fields; secondly, add a convolutional block attention module in the FPN part to further Enhance the effect of feature fusion at different stages and improve the accuracy of target detection for small target defects. Finally, Adam optimizer is used to improve the convergence speed and accuracy of the regression process, and the cosine annealing decay and label smoothing strategies are used to optimize the network loss function. In order to suppress the overfitting problem during network training. By using the improved algorithm to conduct comparative experiments on the printed circuit board defect data set, it shows that the weight file of the algorithm in this paper is only 22.85M, the average detection accuracy is improved by 13.38% compared with the original algorithm, and the detection speed reaches 149.03FPS (on GeForce RTX3060). ), with better effectiveness and feasibility.

Abstract:Aiming at the problems of large amount of network computation and poor detection effect in the current traffic sign detection method, a lightweight traffic sign detection method with embedded coordinate attention mechanism is proposed. First, the coordinate attention mechanism CA module is embedded in the residual block of MobileNetv2 to retain the coordinate information in the channel attention; Secondly, the improved MobileNetv2 is used to lighten the YOLOv4 backbone network, and the depthwise separable convolution block is used in PANet to reduce the amount of computation; Then, ASFF adaptive feature fusion is used to improve the PANet structure to balance the inconsistency of different feature layers. Finally, attention is added to the feature fusion module to increase the weight of the target information; and the K-Means++ algorithm generates new a priori box cluster centers. Experiments show that the weight file is reduced by 60% from 136M to 54.5M, the network volume is reduced by 80%, and the accuracy reaches 96.84%, lose only 0.46% accuracy compared to YOLOv4 network.

Abstract:Aiming at the problems of poor boundary recognition effect and low segmentation accuracy of rail transit wheel set tread damage image, an improved DeeplabV3+ algorithm is proposed to recognize and segment the damage area. Firstly, the lightweight network MobileNetV2 is used as the backbone feature extraction network to speed up the speed of semantic segmentation; Then, the expansion convolution in Atrous Spatial Pyramid Pooling module and the ordinary convolution after feature fusion are replaced by Deep Separable Convolution, so as to reduce the amount of parameters and reduce the complexity of the model; Finally, ECA mechanism is added to the shallow and deep feature layers of the backbone network output to strengthen the network feature learning ability and make the model more sensitive to the damaged area, so as to improve the segmentation accuracy of the model. The experimental results show that the size of the improved DeeplabV3+ model is reduced to 5%, the mPA value is 90.70%, and the mIou value is 84.33%. While the model is lighter, the segmentation effect of tread damage image is ensured.

Abstract:In order to tackle the problem that the original point cloud simplification was easy to lose key features and complex latent surface information, this passage proposed a weighted local optimal projection (WLOP) point cloud simplification algorithm based on FPFH. Firstly, this passage used Fast Point Feature Histogram (FPFH) to find and extract feature points in the original model. Then, the original dense point cloud was reduced by the WLOP algorithm to generate point cloud which had no noise, no outliers, and was evenly distributed. Finally, a point cloud fusion method was used to combine the feature points with the simplified model and remove redundant points. This passage carried out comparative experiments between algorithm with minimum rectangular bounding box algorithm, farthest point sampling algorithm and weighted local optimal projection. The experimental conclusion indicates that the algorithm in this paper is better than other algorithms in terms of distribution uniformity and feature retention when the reduction rate is 30%. In addition, the visual analysis results show that the algorithm in this paper not only guarantee the integrity of the simplified model, but also better preserve the key features of the original point cloud. The results of information entropy analysis show that the simplified point cloud contains richer information and expresses more accurate feature. The algorithm can provide important application value for point cloud reconstruction.

Abstract:Since the existing color constancy algorithms do not perform well for color cast removal in the case of non-uniform illumination and complex scenes, this paper proposes a deep learning algorithm that comprehensively uses the skin color model and perceptual loss to remove color casts. The algorithm integrates the skin color model and perceptual loss, so that can recognize and focus on skin color information in the calculation process, and pay more attention to the understanding of image semantics, rather than simple calculation between pixels. At the same time, the skin color model is combined with the attention mechanism, which highlights the role of the skin color area. The experimental results show that the color constancy calculation method proposed in this paper can accurately eliminate the color cast of images in single-illumination and multi-illumination scenes at the semantic level. Compared with other algorithms, this algorithm can achieve better results.

Abstract:The model uses the BERT to extract the semantic feature representation of the short text, inputs the semantic features into the Bi-GRU and extracts the semantic information with contextual timing features. The model feeds the features into the Maxpooling layer to filter the optimal features and classify them to get the category of the short text. A correction algorithm is added to mitigate the performance degradation for the momentum bias generated by the Adam algorithm in the fitting. The Adam algorithm is improved by comparing the corrected momentum values at two consecutive time steps and selecting the maximum value of momentum in the two time steps to substitute into the gradient calculation. The improved Adam algorithm adds an adaptive adjustment factor to the learning rate and uses the gradient value of the previous iteration to achieve adaptive adjustment of the learning rate and improve the classification accuracy. Experiments show that the classification accuracy of DTSCF-Net is 94.86%, which is 2.07% and 1.71% higher than that of the benchmark model BERT and BERT-Bi-GRU respectively in the same experimental environment. The results demonstrate that the proposed method in this paper has certain performance improvement.

Abstract:Accurate segmentation of open-loop saw chain images under traction motion is the key to automatic detection of saw chain defects. In order to achieve accurate segmentation of parts in saw chain images, this paper proposes a saw chain image segmentation algorithm that combines assembly features and regression analysis. Firstly, by analyzing the assembly features of the saw chain, the Hough circle detection algorithm is used to initially obtain the position information of the rivets in the saw chain image; then the outlier elimination method based on the least squares method is established, and the missed rivets are judged by the position of the adjacent rivets, so as to solve the problem of false detection and missed detection in the Hough circle detection process; then perform affine transformation on the pixel coordinates of the adjacent rivet area to realize the segmentation of the blade, connecting piece and transmission piece in the saw chain image; finally, an experimental platform is built, and the algorithm is verified by collecting images with a dual-position camera. The experimental results show that the saw chain segmentation algorithm can accurately and quickly segment normal and defective saw chain images, and the saw chain segmentation accuracy rate reaches 94.4%, which has good reference significance and practical value for the automatic detection of similar products.

Abstract:To improve the localization accuracy of sensor network nodes and reduce the computational workload, a novel algorithm based on low-rank approximation was proposed. Given distance measurements obtained between sensors in the neighborhood, the proposed algorithm first fulfilled the Euclidean distance matrix (EDM) completion. Then, sensors’ positions were obtained by rigid transformation using anchors’ positions. To achieve accurate range information, the EDM completion stage exploited the low-rank essence of the Gram matrix of sensors’ coordinate matrix, resulting in a semidefinite programming (SDP) problem. Furthermore, some regularization term was introduced in our localization model to avoid degenerate solutions in the EDM completion stage. In practice, solving a large-scale SDP problem is still a challenging task. To improve the scalability of the proposed algorithm, an alternating direction method of multipliers (ADMM) was further developed. Compared with traditional algorithms (including multidimensional scaling method and other Euclidean distance-filling algorithms), this algorithm reduces the root mean square error by 28.2%~46.6% and the reconstruction error by 18.4%~64.5% in the case of large noise through simulation experiments, and the computation time is only 7% of that of SDP algorithm.

Abstract:Aiming at the non-stationary capacity degradation trend caused by the capacity recovery during the use of lithium batteries, which makes the prediction accuracy of the model vulnerable to interference, a long short-term memory network (LSTM) prediction method of lithium battery remaining useful life based on variational mode decomposition (VMD) and bayesian optimization (Bo) is proposed. Firstly, the capacity data of lithium battery is decomposed by variational modal decomposition, and a finite number of modal components are obtained; Then the decomposed components are denoised and reconstructed; Finally, the Bayesian optimized long and short-term memory neural network algorithm is used to predict the service life of the processed data, and the final prediction result of remaining useful life (RUL) of lithium battery is obtained. Through the experiment on the lithium-ion battery data set of CALCE center, the proposed VMD-BO-LSTM lithium battery combination prediction model has high prediction accuracy and stability, and the average value of the root mean square error of the battery used in the experiment is less than 7%, and is better than other prediction models.

Abstract:In order to improve the detection accuracy of P300 EEG signals in non-invasive brain-computer interface (BCI) system, this paper proposes a CNN-LSTM combined network model based on convolutional neural network (CNN) and long short-term memory (LSTM) network. The convolutional network adopts a hierarchical structure, and designs a one-dimensional convolution kernel that matches different feature dimensions; long short-term memory network (LSTM) is used to explore the interdependence of data time series, learning Correlation of global features for object classification. The test results show that the model proposed in this paper has a detection accuracy of 91.28% for the single-trial P300 signal induced by the experiment. Compared with the EEGNet network and the support vector machine(SVM) algorithm, the accuracy is increased by 2.18% and 8.31%, respectively. It also achieves the optimal performance under the evaluation indicators of Precision, Recall, F1 score and AUC value, and has strong generalization performance.

Abstract:Whether the quality of parts is qualified or not affects the service life of the entire assembly. How to quickly and accurately detect whether the quality of parts is qualified has become one of the research hotpots. Machine vision defect detection is increasingly used, but due to the existence of the texture background of the parts after milling, It often leads to insufficient precision in the detection of surface defects of parts. This paper proposes a new image surface defect extraction method that combines the gray-level co-occurrence matrix and the fuzzy C-means clustering algorithm. The improved gray-level co-occurrence matrix is used to increase the contrast between the defect and the milling background, and then the defect and the milling background are used to increase the contrast. For the feature of large gray scale difference between milling backgrounds, fuzzy C-means clustering method is used to segment the image. The algorithm can effectively distinguish processing defects and processing textures, and quickly and accurately extract part defect features. Through the experiment of defect extraction, and compared with the traditional segmentation algorithm, it can be concluded that the algorithm can quickly extract the surface defects of milling parts, and has good adaptability to extract multiple types of defects. Keywords: Milled parts; machine vision; Surface defects; feature extraction; intelligent algorithm.

Abstract:In order to effectively improve the node coverage of wireless sensor networks, a network coverage optimization algorithm based on hybrid strategy sparrow search algorithm is proposed. Firstly, the Tent chaotic mapping is used to improve the initialization sparrow population and increase the diversity of the population; Reverse learning strategy is used to generate inverse solutions to expand the search range and improve the global search capability; Then the inertia factor is added to select Levy strategy and update the sparrow position to improve the local search ability of the algorithm; Finally the optimal sparrow position is perturbed by random walk strategy to further improve the local search capability. The simulation results show that HSSSA algorithm resulted in a more uniform distribution of nodes and a significant improvement in coverage rate.

Abstract:In order to realize the accurate identification of production processes, a process identification model based on machine learning was proposed. Time convolution network, long and short term memory network and support vector machine were selected to build the process identification model, and the model was tested and verified with the production energy consumption data of a titanium metal refining enterprise. Firstly, the historical power and process data were preprocessed, and then the model training and testing data set was constructed according to the production characteristics. Finally, the model was trained and tested based on the data set. The results shows that the recognition model based on time convolution network has a high accuracy of process identification, and the accuracy of process identification for test sets reaches 96.94%.