Discrete Wavelet Transform-Based Detection Transformer for Battery Weld Defect Detection

There are three main challenges in applying target inspection methods to the detection of weld defects between the top cover and casing of a battery: 1) weld defects are difficult to visualize and label; 2) the limited amount of sample data constrains the efficacy of the deep learning model; and 3) the depth sequence information at the weld seam of the battery case is rich in high-frequency features, which helps judge defects. However, existing feature extraction methods fail to consider the high-frequency features of the battery weld. To address these issues, we propose solutions to the problems separately: First, the Keyence Laser Scanner is used to capture the weld long depth sequence, and the optional feature interval visualization (OFIV) method is used to visualize the defects and further annotate the data. Second, the random slicing augmentation (RSA) method is proposed to augment the dataset. Third, the real-time discrete wavelet transform-based detection transformer (RT-DWTR) is employed to capture high-frequency features with the objective of adapting deep sequence data and enhancing detection performance. The experimental results indicate that the RSA method significantly enhances the performance of the detection network, while the RT-DWTR shows improvements in mAP50 by 0.6%, mAP50:90 by 3.4%, mAPs by 6.1%, and F1 score by 1.6% compared to the benchmark network while maintaining a similar number of model parameters.