Abstract:Fugl-Meyer Assessment is one of the most commonly used methods in stroke impairment evaluation. However, Fugl-Meyer assessment needs guidance and grading from professional rehabilitation medical doctors. Therefore, there are challenges in stay-home Fugl-Meyer assessment. In this paper, we present a system that can make Fugl-Meyer assessment from videos taken by common cameras. The proposed system consists of two modules: A motion data capture module for fetching motion data in Euler Angles from videos and a Fugl-Meyer assessment module for grading through motion data from the former module. Experimental tests are conducted on the Human 3. 6 M dataset and demonstrate that our video-based Fugl-Meyer assessment system performs well in accuracy and covers most of the test items in Fugl-Meyer assessment table.

Abstract:Aiming at the low detection efficiency of existing circle detection algorithms in high-resolution marker images, this paper proposes a new type of rapid circle detection algorithm for markers. First, the YOLOv4-tiny algorithm is used to quickly locate the markers in the image to filter out the background interference, and then use the proposed arc completeness sorting method to screen out candidate circles with different completeness from the markers, and finally filter through the distance and direction angle constraints. And fit all the marker circles. This paper conducts experiments based on high-resolution marker images. The experimental results show that this algorithm has extremely fast detection speed. When the detection accuracy is basically unchanged, the detection time of this algorithm is only 11% of the EDcircle algorithm and 5. 0% of the AAMED algorithm. It provides an efficient solution for tasks such as visual positioning and calibration that require rapid positioning of markers.

Abstract:Aiming at the problem that traditional deep learning strategies used in steel plate surface defect images classification rely on abundant labeled samples. This paper proposes an efficient deep active learning method with a lightweight convolutional neural network and a novel uncertainty based active learning strategy. The network adopts a simplified convolutional base to do feature extraction, and replaces the hidden layer in the final densely connected classifier with global pooling layer to mitigate overfitting. To better measure model uncertainty about unlabeled image samples, this method first passes unlabeled images through the model trained by labeled image samples to obtain the probability distribution over classes ( PDC) for every unlabeled sample, then uses the same model to make predictions on the labeled samples to get an average PDC for every class. The KL-divergence value between these two kinds of distributions can be used as a new uncertainty measure to select unlabeled images for annotation. According to the experiments on NEUCLS dataset, the proposed method can reach 97% accuracy with 44% labeled data, which can reduce annotation cost greatly.

Abstract:Coke optical texture analysis is an important way to evaluate the quality of coke, aiming at the problems of fuzzy edge, low contrast and halo artifacts in coke photomicrographs, an automatic detection and extraction method for coke optical texture based on semantic segmentation and fully connected conditional random field is designed. Firstly, a coke optical texture measurement platform is built by using microscope, industrial camera and computer; secondly, the Unet is improved using residual module and attention module and the output weight of the coke optical texture is enhanced to realize automatic detection and segmentation of coke optical texture; finally, the spatial characteristics of coke optical texture are modeled using the fully connected conditional random field to refine the segmentation edges and achieve the accurate extraction of coke optical texture. The experimental results show that the precision, recall, F1-score and accuracy of the proposed method reach 0. 967, 0. 959, 0. 963 and 0. 965, respectively, which are better than other comparative semantic segmentation networks, proving that the method has high segmentation performance and can realize automatic detection and extraction of coke optical texture.

Abstract:Aiming at the problem of waste oil leakage from power plant rainwater outlets in monitoring images, a pollutant leakage detection algorithm based on improved Faster R-CNN is proposed. The improved Faster R-CNN detection algorithm first uses ResNet-50 as the backbone network, and builds a multi-scale feature map pyramid structure ( FPN) on this basis to achieve information fusion between high-level semantics and low-level semantics, and improve detection accuracy; Secondly, the CIoU loss and DIoU-NMS methods are used to improve the accuracy of bounding box regression in Faster R-CNN; Finally, by introducing Focal Loss function, it solves the problem of unbalanced positive and negative samples in the R-CNN training stage caused by redundant anchor generated by RPN network. The experimental results show that the improved algorithm performs well in real samples, and the accuracy rate reaches 90. 2%. Compared with the original Faster R-CNN algorithm, the accuracy rate is improved, and the false positive rate and false negative rate are significantly reduced. It can be effectively used in the actual environment.

Abstract:At present, most mangoes still need to be sorted and graded through manual identification of volume and quality, resulting in low efficiency and lack of data management. Machine vision is an effective means to improve the efficiency of mango grading, but traditional industrial cameras can only obtain two-dimensional projections. In response to this situation, this paper uses a 3D structured light system to obtain mango shape descriptors combined with three-dimensional depth information. Then 80 correction sets are used as samples, and the fisher judgment method is used for pose detection, and the non-linear support vector machine establishes the volume and mass prediction models in the “flat” and “upright” poses. Error analysis is performed on the prediction set. The results show that after adding depth information, the accuracy of pose detection can be increased to 100%, and the average error of volume quality can be reduced to less than 5%.

Abstract:Aiming at the actual problems of low detection accuracy and low detection speed in the detection of power insulators and insulator defects in a complex environment background, an improved you only Look once v4(YOLOv4) algorithm for power insulator images and existence Method of detecting defective insulators is proposed. By making a dataset of power insulators and insulators with defects, using K-means clustering (K-means) algorithm to cluster the power insulator image samples to obtain different sizes of a priori box parameters; then by improving the balance cross entropy (Balanced Cross Entropy, BCE), it introduces a weight coefficient to increase the contribution of the loss function. Finally, the depth of the network is deepened by adding convolutional layers before and after the spatial pyramid pooling ( SPP) structure. The experimental results show that the single sheet detection time of the improved model is 3. 27 s, and the average detection accuracy of insulator defects is improved by 24. 36% compared with the original YOLOv4 algorithm. At the same time, through the improved YOLOv4 algorithm, the value of mean average precision(mAP) on the test set is 84. 05%, which is 17. 83% higher than the original YOLOv4 algorithm, which fully demonstrates the ability to locate and identify of the defect in power insulator images well.

Abstract:Due to the need of extracting skeleton information and gait parameters for walking quadruped, a skeleton extraction model of quadruped is proposed by using context information enhancement and multi-scale information fusion based on HRnet, and a quantitative analysis method for gait parameters is established. The validity of the model is verified on the test data set of images. Experiments show that, in the key point estimation of quadruped, the model achieves good performance with the mean similarity of 81. 04%, the accuracy of 92. 77%, and the recall rate of 92. 75%. Based on the skeleton extraction model, the frequency of buffalo and alpaca were analyzed and calculated. Compared with the manual statistical results, the maximum relative error was 2. 73%. Through analyzing the angle variations of buffalo’ s and alpaca’ s hip and knee joints during a complete gait cycle respectively, the joint motion logic and gait sequence of walking quadruped can be extracted automatically. Finally, taking the rhinoceros as a sample, it is demonstrated that the proposed method can work adaptively in a range of different shooting angle of images. The results can provide a reference for intelligent perception of quadruped motion information.

Abstract:The quality inspection of metal parts is a labor-intensive work. Generally, fatigue test is used to evaluate the quality of metal parts and the basis for accurate evaluation of the quality of metal parts, which can guide the design of reasonable mechanical structure and effectively reduce safety accidents and economic losses. In order to quickly screen qualified metal parts with strong vibration resistance, an online detection algorithm of vibration fatigue crack based on computer vision is proposed for metal parts with rapid vibration in fatigue test. Firstly, the displacement and motion blur caused by vibration are eliminated based on the front and rear frame alignment method, and then the surface changes of parts are detected by the inter frame difference method. Finally, the crack shape and length parameters are obtained by the segmentation algorithm according to the texture and geometric characteristics of the crack. The experimental results show that compared with the crack detection algorithm under static conditions, the proposed algorithm can automatically obtain the location, length and shape of parts’ cracks without stopping the shaking table, and record the crack propagation process, which can greatly improve the work efficiency of vibration crack detection.

Abstract:Aiming at the problems that traditional human action recognition algorithms cannot effectively suppress spatial background information, the lack of information interaction between networks, and the inability to model global temporal correlation, a human action recognition algorithm of feature fusion Bi-LSTM based on segmentation attention is proposed. First, 30 frames of images are sampled at a certain sampling rate, extract the depth features of the images by split-attention network, and introduce a feature fusion mechanism to enhance the information interaction between different convolutional layers. Then input the depth features into the Bi-LSTM network to model the long-term information of human actions, and finally use the Softmax classifier to classify the recognition results. Compared with the traditional two-stream convolutional network, the accuracy of this algorithm on the UCF101 and HMDB51 datasets is increased by 6. 6% and 10. 2%, respectively, which effectively improves the recognition accuracy.

Abstract:In an unconstrained environment, face detection is challenging due to differences in light, occlusion, and expressions. The accuracy of the multi-task cascaded convolutional neural network ( MTCNN) face detector is reduced in low-light environments. To improve the accuracy of face detection in low-light environment, a MSRCR-based optical frequency segmented filtering enhancement algorithm ( 3CGF-MSRCR) is proposed. This paper uses MTCNN for face detection, and uses a RGB three-channel decomposition guided filtering (GF) method to improve the multi-scale retinex with color restoration. Firstly, face images are enhanced by MSRCR and decomposed into RGB three channels to obtain the image weights of RGB. Then the GF method is used to filter each channel separately and update the weights of the RGB images. Finally, we reconstruct the face image. Training and testing are conducted on the actual lowlight scene face dataset: Dark Face and the public standard face dataset CelebA. Meanwhile, the running time of the proposed algorithm is compared with other enhancement algorithm. The results show that 3CGF-MSRCR can effectively suppress the high-frequency noise of MSRCR, retain the brightness enhancement effect, and improve the accuracy. Meanwhile, 3CGF-MSRCR has a faster running speed.

Abstract:Optical fiber connectors have attracted much attention due to the essential role in optical transmission systems, but the impurities attached on the fiber surface will generate noise on the recovered morphology. Moreover, the existing detection methods cannot accurately locate the noise. It needs to be processed for multiple overall noise reduction, and the image detail retention ability obtained by this method is inadequate. To this end, we proposed an improved Gaussian mean region denoising technology based on GA-BP neural network. Firstly, the interference data is processed by dimensionality reduction. Secondly, select the dimensionality reduction data as the training data, and use the neural network to locate the noise. Finally, the improved Gaussian mean filter is used to filter the noise position of the three-dimensional image. Furthermore, the results show that the noise pixel obtained by the neural network discrimination method is 2. 45%, which is higher than the threshold discrimination method. And the noise difference obtained by the improved Gaussian mean filtering method is 474. 7, and the PSNR value is 32. 56. Compared with the mean and median filtering methods, the image detail retention ability is higher, and the restored image noise bulge is significantly reduced. Therefore, it is more suitable for automatic detection based on the principle of white light interference.

Abstract:In the automatic cleaning process of embedded aeronautical ground lights in airport, in order to obtain the center of lamps emitting port and the relative position of the manipulator to achieve precise positioning of the manipulator, a positioning system of lamps cleaning for manipulator arm based on monocular vision is designed. Firstly, the kinematics model of the manipulator was established by D-H method. Then, according to the strong backlight characteristics of lamps and environmental interference at night, an improved Otsu algorithm was proposed to segment the image by optimizing the threshold criterion and reducing the search range of threshold, and then the center of mass method was used to extract the center position of the lamps emitting port. Finally, under night conditions, the experiment was analyzed and the least square method was used to compensate the positioning error. The experimental results show that the designed manipulator arm positioning system of lamps cleaning has high speed and high precision. Compared with the traditional Otsu algorithm and the improved random Hough transform algorithm, the positioning accuracy was improved by 72. 5% and 55. 5% respectively, and the average positioning accuracy reaches 8. 7 mm, which meets the requirements of lamp cleaning.

Abstract:In order to meet the demand of high efficiency and accuracy of weld appearance quality defect detection in wind turbine tower cylinder industry, the 3D machine vision technology was used to develop the weld appearance quality defect detection system based on wind turbine tower cylinder. First, the point cloud data was preprocessed by point cloud filtering, point cloud segmentation and point cloud simplification to ensure the accuracy of defect evaluation in the later stage. Secondly, the contour characteristics of 3D data were obtained by slice processing and breakpoint fitting. Thirdly, the improved recursive rough extraction algorithm was used to extract the feature points, and the defect evaluation was carried out to obtain the detection results of weld appearance defects. Finally, according to the evaluation process and standard of weld defects in the system, a typical weld sample is selected to test the weld width, weld dislocation and weld straightness. The weld detection accuracy can reach 0. 001 mm, and the speed is 3 times of the current manual detection speed. The detection results show that the system has the characteristics of high accuracy, high speed and high precision, which can replace manual detection, and has a good application prospect.

Abstract:In order to achieve rapid and accurate recognition and location of insulators for the railway catenary, an insulator recognition and location approach was proposed through deep integrating the YOLOv4 target detection algorithm with the ORB feature matching algorithm under the complicated background. To begin with, the transfer learning strategy was adopted to train the YOLOv4 detection network with an objective of addressing the overfitting problem resulted from few insulator data sets. Then, image multi-scale features were extracted using Gaussian pyramid. By doing so, the original ORB algorithm is equipped with scale invariance; Finally, the insulator recognition frame was marked on the image acquired by a binocular camera by integrating the above two algorithms. On this basis, the three-dimensional coordinates of the insulator relative to the camera can be restored with the parallax principle upon extracting feature points in the left and right image recognition frames for matching. Experimental results demonstrate that the proposed approach featuring high precision and real-time performance can accurately locate the three-dimensional coordinates of the insulator through effectively preventing complex background interference. The maximum positioning error within 4 meters is 2. 1%, and the detection speed is 35 fps.

Abstract:The appearance-based gait recognition methods are easily affected by the carrying objects, clothing and other occlusion factors. In order to solve this problem, Dynamic Gait Image is proposed. Dynamic Gait Image divide gait image into dynamic part and static part, which is more conductive to extract dynamic information less affected by occlusion factors. This paper proposes Bi-Route gait recognition network, which can minimize the influence of occlusion factors by increasing the proportion of dynamic features and reducing the proportion of static features. The global silhouettes features and frame level silhouettes features of the gait sequences were extracted by 2D-convolutional neural network with the input of dynamic gait image. Then 3D-convolutional neural network extracts dynamic features from frame level silhouettes features. The accuracy of the proposed method evaluated on CASIA-B dataset is 92. 9%, 87. 2% and 65. 6% in NM, BG and CL conditions. The result shows that the proposed method can reduce the impact of occlusion factors.

Abstract:In this paper, aiming at the highlight phenomenon on the surface of optical device metal wire in the process of visual detection, a method for detecting and removing the highlight area of optical device metal wire image is proposed. First, according to the feature analysis of the collected metal wire image, the highlight area is segmented by the OTSU algorithm. Then, the bwareaopen function in MATLAB is called to delete small-area objects, that is, the highlight closed area, and the highlight area image is extracted through the image difference operation. Finally, the pixel value of the highlight area is updated by the method of the average value of the eight-neighborhood pixels. This method not only removes the highlight area, but also ensures the balanced distribution of the pixel value of this area in the image. Through the analysis of simulation experiment results, the effectiveness and feasibility of the method for detecting and removing the highlights of the metal wire image are verified. Meanwhile, it provides a strong guarantee for the subsequent quality inspection of the metal wire.

Abstract:In order to detect the impact energy generated by spacecraft micro-impact, a femtosecond fiber grating impact detection system for spacecraft micro-impact was designed and the impact energy was detected. The strain sensing model and edge filtering demodulation model of femtosecond fiber grating are established. The finite element analysis software is used to simulate the impact energy distribution of micro-collisions. An impact detection system was built to test the impact energy sensing characteristics of femtosecond fiber grating, and the energy transfer characteristics were analyzed by using the time-domain envelope extraction method. Experimental results show that when the balls with the mass of 10, 20, 40, 60 and 80 g fall at the same position, the collision signals detected by the grating system show a linear increasing trend, the maximum impulse voltage retrieved by edge filtering algorithm is 30. 8 mV. When the same mass ball falls at different positions, the collision signals detected by the grating system show a linear downward trend, the detected impulse voltage can be as low as 16. 5 mV. The system can detect the impact energy generated by the micro-collision, which has a certain reference significance to the problem of spacecraft micro-collision detection.

Abstract:Water content plays a crucial role in the growth and metabolism of standing trees. Real-time and accurate measurement of water content is of key guiding significance for standing tree cultivation and forest management. A wood moisture content diagnosis system based on wireless acoustic emission sensor network (WASN) was designed and implemented for the nondestructive testing of living wood. Firstly, the acoustic emission signals of the trunk epidermis were sampled at high speed by the WASN node, and then the characteristic parameters were calculated and transmitted to the gateway wirelessly. After that, the optimal feature combination was selected by the MRMR criterion, and the water content identification model was established by the support vector machine (SSA-SVM) optimized by the sparrow algorithm. Finally, on-line real-time long-term monitoring and diagnosis can be carried out. The system has been tested on four species of met sequoia, poplar, pine and beech respectively, and the results show that the lowest diagnostic accuracy is 95. 5%. The design of WASN was fully capable of long-term observation of tree transpiration.

Abstract:To address the problem that the performance of UHF RFID tags in the dense environment of vaccine transportation is affected by the mutual coupling effect and liquid environment, the computational expressions of mutual impedance and power transmission coefficient of tags under the action of proximity liquid interference and tag mutual coupling effect are derived from the perspective of impedance change of single tags in liquid environment based on inductive coupling model and using two-port network analysis method. Using the control variable method to design simulation experiments and actual measurements, the simulation measurement results show that under the interaction of multiple environments, the liquid environment mainly affects the real part of the label, and the mutual coupling effect of the label mainly affects the imaginary part of the label impedance; the research results have certain guiding significance for the label design and the application of the label in the cold chain transportation of vaccines.

Abstract:Water content is an important indicator to measure the quality of crude oil. In the process of production, storage and transportation of crude oil, the interface of oil-water mixture is constantly changing. Thus, high-precision sensors are used to detect it in the whole process. In this article, an intrusive planar capacitance sensor was designed based on the capacitive edge effect. Its main structure consists of a substrate and a planar electrode array. The 8-electrode array sensor model was established by using finite element software. The electric field distribution of different electrodes was studied, and the detection sensitivity and imaging accuracy of the planar capacitive sensor were analyzed. In addition, the influence of the width, length and adjacent distance of the electrodes on the sensitive field distribution of the sensor was studied. Through the image reconstruction of the dielectric distribution, the designed planar capacitance array sensor can detect the height of the three interfaces, and the size parameters are optimized to improve the imaging accuracy of the sensor. The feasibility and effectiveness of the method of using planar capacitor array to detect oil-water interface have proved by experiments.

Abstract:Coiled tubing operation technology is widely used in drilling and repair operations, and the core of the technology is coiled tubing. The quality of the bias weld directly determines the fatigue life. Whether the microstructure of the bias weld is consistent with the base material determines whether the fatigue resistance of the weld reaches the level of the base material. This paper uses multifrequency eddy current testing technology to detect the reactance of the CT90 and CT100 coiled tubing steel strips in the bias weld area and the base material area to distinguish the microstructure and grain size of the base material area and the weld area, and judge whether the microstructure of the bias weld reaches the level of the base material. According to the signal characteristic cloud map, the weld area and the base material area are identified. The results show that the grain size of the coiled tubing bias weld area is significantly larger than that of the base material. It can be seen that the weld quality of CT90 is better than CT100. According to the microstructure difference between the weld area and the base material area, the quality of the bias is evaluated, which has guiding significance for the quality control and improvement of the coiled tubing production process.

Abstract:Low-cycle fatigue cracks of titanium alloy aerospace components are easily underestimated by conventional linear ultrasonic inspection technology, which is a potential safety hazard for aircraft service. In the research, the modulation nonlinear ultrasonic detection and quantitative technology of fatigue cracks in titanium alloys were carried out. A combination of the linear and the nonlinear ultrasonic quantitative method was proposed for the fatigue cracks in titanium alloys. Firstly, the fatigue crack is quantitatively detected by the metallographic method and the ultrasonic phased array technology. Secondly, a modulation nonlinear ultrasonic detection system and a sensor arrangement method are designed for strong anti-interference ability. Finally, a modulation non-linear ultrasonic quantitative method for fatigue cracks is proposed based on the analysis of the sound field distribution, and the quantitative detection results are compared with ultrasonic phased array detection technology. The research results indicate that the length of the macro-crack measured by the ultrasonic phased array method is 12 mm, while the micro-crack measured by the nonlinear ultrasonic method extends to a distance of 20 mm from the end of the narrow groove. Thus, the non-linear modulation detection technology has more significant advantages in detecting micro-cracks and closed cracks.

Abstract:In the vehicle panoramic system, how to accurately detect the position and direction of parking slot is still a problem to be solved. To solve this problem, we designed a two-way parallel multi-scale mark point detection network. The two-way network is used to detect the position and angle of mark point respectively. Multi-scale features are extracted from panoramic images, and a branch network of one high and one low resolution is maintained in parallel. The two branches are fused with each other. The high-resolution features express the location of mark point in the form of Gaussian heatmap. A new method for calculating the direction of parking slots is proposed, which uses the direction of two mark points and the relative position of two mark points to calculate the direction of parking slot. In order to verify the feasibility of the proposed method, the designed network was trained using the training set of the public dataset PS2. 0, and the parking slot detection precision tested on the public dataset PS2. 0 and the self-collected dataset PSS is 99. 4% and 95. 27%, the recall is 99. 88% and 80. 89%, the average error of the mark point position on PS2. 0 is 0. 84 pixel, and the error of the parking direction is 0. 71 degree. The experimental results show that compared with the existing methods, the parking slot detection network proposed reduces the errors in the location of the mark point and the direction of the parking slot, and has a strong generalization ability on the PSS dataset.

Abstract:To address the problem of too many model parameters in predistorters, this paper proposes a predistorter model simplification method for broadband power amplifier (PA) based on the theory of compressed sensing (CS). On the basis of the sparsity adaptive matching pursuit (SAMP) algorithm, the relevant support selection SAMP algorithm based on frequency domain notch (RSS-FNSAMP) is proposed. The PA behavior model can be simplified and used in the proposed segmented dual-feedback DPD system, and the out-ofband distortion masked by the in-band residual can be compensated, which not only enhanced the stability of the system, but also reduced its complexity and improved the linearization effect of DPD. To verify the method, a 35 dBm class F power amplifier is driven by 20 MHz LTE signal. The experimental results show that the normalized mean squared error (NMSE) is improved by 3~ 5 dB compared with ILA-SAMP, ILA-DOMP and segmented dual feedback-DOMP, and the adjacent channel power ratio ( ACPR) is improved by 25 dBc, which shows that the proposed method can improve the linearity of power amplifier while reducing the number of model parameters.

Abstract:In order to reduce the dependence of neural network on a large number of complete data in the process of mechanical equipment fault prediction and health management (PHM). To solve the problem of rolling bearing fault diagnosis with scarce data, a multi-source domain transfer learning method is proposed. The model uses one-dimensional convolution neural network, takes the original vibration signal as the input of the model, uses two different source domains data to pre-train the model, and uses the target domain data to fine-tune the pre-training model to improve the recognition accuracy of the target domain. Using the measured data of the machinery fault simulator of spectra quest and the bearing data sets of Case Western Reserve University, in the case of few fault samples in the target domain, the classification accuracy, training speed, result stability and multi-source domain effectiveness of the model are verified respectively, and the results of migration diagnosis were compared with those of CNN, TCA JDA and SVM. The results show that the model can achieve higher classification accuracy when the fault data is scarce. In the case of the number of samples in the three target domains, the classification accuracy of the multi-source domain migration method reaches 97. 71%, 96. 28% and 94. 18%, respectively. The model has fast convergence speed and good stability.

Abstract:Planar capacitive sensor works based on edge effect. The sensitive field exhibits obvious soft field characteristics, as a result, its detection ability is easily affected by external factors. In order to ameliorate the sensitive field distribution characteristics and improve the sensor’ s testing accuracy, this paper proposes a new type of 12-electrode coplanar array capacitive sensor and analyzes its characteristics. First, a numerical simulation method is used to set a three-dimensional sensitive field model, and the sensitivity value three-dimensional imaging method is selected to study the sub sensitivity and the overall sensitivity distribution characteristics. Then, the evaluation index P is used to quantify the uniformity of the sensitive field distribution of the new 12-electrode sensor and the rectangular array sensor. Three different object model experiments are implemented combined with the Tikhonov regularization algorithm for image restoration. The final result indicates that the P value of the new sensor is reduced by 3. 6%, and the average correlation coefficient of the reconstructed image is increased by 7. 2%. It is comprehensively verified that the new 12-electrode coplanar array sensor effectively improves the uniformity of the sensitive field distribution, enhancing the detection capability of the sensor.

Abstract:A SOC estimation method for a lithium battery that can track the abrupt state is proposed and applied to the SOC equalization of the multi-lithium battery pack. The strong tracking filter is introduced into the particle filter algorithm, and the current sampling results are incorporated into the prediction error update to get a new correction term, and then the correction term is used to correct the particle set of the particle filter, so as to quickly push the particles to the high likelihood region and restrain the particle degradation. The introduction of the fading factor can adjust the error covariance matrix in real-time so that the particle filter algorithm has both the strong robustness of the strong tracking filter and the tracking ability of the mutation state, which can effectively overcome the uncertainty of the model and further improve the estimation accuracy of the SOC. The proposed method was applied to much cell-active balance. The equilibrium strategy was designed based on the consistency criterion of battery SOC, and the adjacent cells with larger capacity gap were equilibrium first, then energy real-time bidirectional transmission is controlled, so the overall balancing speed was improved. Experimental results show that the average estimation error of the improved algorithm is within 0. 13%, and the standard deviation is 0. 12%. The improved algorithm is about 64%, 85%, and 75% higher than that of the traditional particle filter algorithm, the extended Kalman algorithm and the strong tracking algorithm, and the stability is further enhanced. The application in multi-battery active equalization can effectively reduce the inconsistency of battery pack capacity in the process of charge and discharge, and control the dispersion of the battery pack within 1%, which is beneficial to improve the utilization rate and service life of battery capacity.

Abstract:Battery state of charge (SOC) estimation is helpful to alleviate the mileage anxiety in the process of driving. Aiming at the problem of particle degradation in the estimation of SOC by particle filter, this paper proposes to apply the Gaussian particle swarm optimization particle filter ( GPSO-PF). Compared to estimation of SOC by particle filter, GPSO-PF combines particle swarm optimization algorithm and particle filter to estimate SOC. GPSO-PF solve the problem of particle dilution and improve the estimation accuracy of SOC by continuously optimizing the position of particles in the iteration. As SOC estimation is easily affected by temperature, an equivalent circuit model based on temperature is established and applied to the proposed SOC estimation algorithm. Two LiFePO4 batteries of the same type are selected and the GPSO-PF algorithm is used to estimate the SOC value under different working conditions. The maximum estimation error of SOC is less than 0. 72%. By comparison, GPSO-PF algorithm combined with equivalent circuit model based on temperature can effectively improve the estimation accuracy of SOC.