Research And Implementation Of Lithium Battery Defect Detection Algorithm Based On Deep Learning

Nowadays,the state actively advocates green and environmentally friendly travel,and people are paying more and more attention to new energy vehicles.As the main power source of new energy vehicles,lithium batteries can effectively protect the life and property of drivers by improving their safety.During the coating operation of lithium battery pole pieces,affected by the coating process and environmental factors,it is easy to produce up to 18 types of defects such as bubbles,joints,pits,and metal leakage,which greatly reduces the safety of lithium batteries.The traditional lithium battery defect detection is mainly through manual visual inspection.This method has low efficiency and high cost,and the shape of some defect types is very similar,which cannot be accurately distinguished by the human eye.Traditional machine vision algorithms need artificially designed features to complete the defect detection of lithium batteries.The algorithm is not very versatile,and there are few detection types,so the detection accuracy needs to be improved.Deep learning has achieved many achievements in the field of image processing.Based on this,this thesis studies the application of deep learning in the classification and detection of coating defects of lithium battery pole pieces to improve the types and accuracy of defect detection.Firstly,the data set containing 18 kinds of coating defects was statistically analyzed,and the data is enhanced by turning and rotating operations for the problem of unbalanced defect data.For some problems with similar types of coating defects,contrast enhancement was used to preprocess data to improve the identification between defects.In addition,the data sets before and after processing are compared in the network for experimental analysis to verify the effectiveness of image processing.Secondly,comparative experiments are carried out on 5 classical convolutional neural network models.According to the experimental results,Alex Net was selected as the basic model,and then the network structure and parameters were optimized.For the problems of too many network parameters,incomplete feature extraction,and unreasonable distribution of network data,the convolution kernel and the number of layers are optimized,and a batch normalization layer is constructed in the network.The experimental results show that the optimized network inference prediction efficiency and detection accuracy are effectively improved.Then,in order to solve the problem of gradient disappearance caused by network deepening and loss of defect details in the process of feature extraction,skip connections are constructed to solve the problem.In the structure,dilated convolution is used for multi-scale feature extraction,and then feature fusion is performed.The experimental results show that the network recognition accuracy can be further improved after adding skip connections to the network.Next,in view of the problem of neuron death during network training,the network parameters cannot be updated,and the Leaky Re LU activation function is introduced into the network to retain useful negative features to prevent neuron death.The experimental results show that the recognition accuracy of the network is further enhanced after the network activation function is improved.Finally,a visual interface is built,which can call the model to monitor the defect classification and detection process in real time.The analysis of multiple sets of comparative experiments shows that the accuracy of the improved network can reach99.34%,and the detection time of a single sample is 51 ms,which meets the accuracy and time requirements of the classification and detection of lithium battery pole piece coating defects in the factory.