Research on power battery welding defect detection based on deep learning models with transformer architecture

Abstract The welding quality of new energy vehicle battery terminals is crucial for the safety and reliability of the entire vehicle. Addressing the limitations of current weld defect detection methods, this paper proposes a lightweight point cloud classification model based on Transformer, named TDSPnet, which incorporates a dynamic similarity mechanism to enhance the accuracy and efficiency of weld defect detection. The dynamic similarity attention module is designed to strengthen the response to morphology-sensitive weld regions, such as missing solder, grooves, local height discontinuities, and surface-continuity breaks, thereby linking the algorithmic attention weights with physical defect characteristics. The TDSPnet network structure consists of a feature embedding module, a dynamic similarity attention module, and a classification module for segment-level defect judgment in battery weld point clouds. Experimental results show that TDSPnet achieves an accuracy of 0.9555 on the BWS dataset provided by Hexagon and a power battery manufacturer, with a miss rate and false alarm rate of 4.52% and 4.37%, respectively. Compared with PointNet, PCT, and DGCNN networks, TDSPnet demonstrates superior performance in convergence speed, training stability, detection accuracy, and model-inference efficiency. The research findings provide a new solution for defect detection in power battery terminal welds and indicate promising engineering application value.