DGNet: An Adaptive Lightweight Defect Detection Model for New Energy Vehicle Battery Current Collector

• Published in: IEEE sensors journal 2023-12, Vol.23 (23), p.29815-29830
• Main Authors: Lei, Yuan, Yanrong, Chen, Hai, Tang, Ren, Gao, Wenhuan, Wu
• Format: Article
• Language: English
• Subjects: Accuracy; Adaptation models; Adaptive; Batteries; battery current collector (BCC); Computational modeling; Computer networks; Deep learning; defect detection; Feature extraction; Feature maps; Floating point arithmetic; Industrial applications; Jetson Nano; lightweight; Object detection; Passenger safety; Real time; Redundancy; Surface defects; Welding
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3324441

As an essential component of the new energy vehicle battery, current collectors affect the performance of battery and are crucial to the safety of passengers. The significant differences in shape and scale among defect types make it challenging for the model detection of current collector defects. In order to reduce application costs and conduct real-time detection with limited computing resources, we propose an end-to-end adaptive and lightweight defect detection model for the battery current collector (BCC), DGNet. First, we designed an adaptive lightweight backbone network (DOConv and Shufflenet V2 (DOS) module) to adaptively extract useful features adaptively along all four dimensions of kernel space while maintaining low-computational complexity. Second, we designed a lightweight feature fusion network [GSConv and FPN (GS_FPN)], which reduces parameter redundancy and fully utilizes the semantic information of the feature maps of backbone network while ensuring detection accuracy. Experimental results show that DGNet achieves a mean average precision at intersection over union (IoU) threshold 0.5 ( \text {mAP}_{{50}} ) of 91.8% on the self-made BCC surface defect database, with a model size of 4.0M, only 3.7 giga floating-point operations per second (GFLOPs), and frames/s (FPS) of 181.8. To further demonstrate the capabilities of DGNet, we test it on the publicly Northeastern University (NEU) surface defect database, and the results showed that the DGNet exhibited good generalization. Compared with current advanced lightweight network models, it achieves higher detection accuracy and lower computational overhead, reaching the state-of-the-art (SOTA) level. Finally, we deployed DGNet on the embedded platform NVIDIA Jetson Nano for real-time detection, achieving a detection time of 0.074 s per image, meeting the accuracy and real-time detection requirements for BCC defect detection tasks in practical industrial applications.