Research On Surface Defect Detection Method Of Lithium Battery Electrode Based On Deep Learning

In recent years,a large number of new energy products have flooded into the market and daily life.As a new energy power lithium battery,the demand and quality requirements in the market are also increasing.The pole piece is a part used to conduct lithium battery power.During the manufacturing process,it is necessary to complete the process of coating,rolling and other processes,and some defects will occur.These defects will affect the chemical performance of the lithium battery and bring about potential safety hazards.Research on an efficient,accurate and multi-scenario lithium battery pole piece defect detection system can detect defect information in time,provide information feedback for process improvement,and reduce the occurrence of safety accidents and economic losses.Traditional digital image processing technology has the problems of complex feature extraction and weak robustness when faced with multiple types of complex pole piece surface defects.The production lines of different enterprises have different detection requirements for the surface defects of lithium battery pole pieces,mainly including: positioning-level detection for classifying and locating defects and pixel segmentation-level detection for classifying defects and extracting specific contours.Deep learning has the characteristics of automatic feature extraction and good robustness,and is widely used in industrial defect detection.This thesis uses deep learning technology to design two network algorithms to meet the needs of two different detection scenarios.The main work and research results are as follows:（1）Aiming at two different detection requirements of pole piece surface defects,a deep learning-based pole piece surface defect detection method and process are proposed,and a defect dataset with high-quality label information is constructed for deep learning detection research.In the process of constructing the defect data set,data enhancement,data labeling,and data set division have been completed.（2）For the detection at the localization level,the Yolo V3 target detection network is used as the basic network to realize the classification and localization detection of the surface defects of the pole piece.In view of the problems of inaccurate positioning and missed detection in the initial Yolo V3 detection results,the idea of weight migration is firstly used to transfer the weight parameters trained by the Yolo V3 network on the MS COCO dataset to the network in this thesis,and then use the CIo U loss to improve Predict the positioning accuracy of the bounding box,and use Mosaic data enhancement in image preprocessing to randomly combine and increase the number of defective samples to improve the detection ability of the network.The experimental results show that after combining these three methods,the Yolo V3 network detection effect is obtained.With a great improvement,it can efficiently and accurately locate and detect the surface defects of the pole piece.（3）For the detection of pixel segmentation level,Unet semantic segmentation network with simple structure and clear design idea is used to study the segmentation detection of pole piece surface defects.Aiming at the problem of wrong segmentation and the insufficiency of Unet network structure,three stages of experiments are carried out successively to improve the segmentation accuracy and detection effect.First,a classification network with weight transfer is used to replace the coding structure in Unet network.Secondly,a simplified version of SFPN feature fusion is proposed and added into Unet jump connection.Finally,label smoothing is used to optimize the loss function and improve the generalization ability of the network.After the optimization experiment,some noise profiles and missegmented small defect profiles were removed.The experimental results show that the overall optimized Unet network has a good segmentation effect on pole piece surface defects,and can meet the segmentation detection requirements of pole piece surface defects.（4）On the basis of the detection algorithm research,the corresponding detection software system is further designed,and the detection modules of the two algorithms are also designed in the software for the two detection scenarios.The software test results show that the software system is simple to operate,and on the basis of completing the defect location,classification and segmentation,it also increases the visualization of the detection effect,realizes humancomputer interaction,and can save the detection data,which has certain feasibility.