A Dual-Strategy Deep Learning Framework for Automated Defect Detection: Enabling Efficient Echelon Utilization of Spent Lithium-Ion Batteries

Echelon utilization, as the preferred method for reusing spent power batteries, relies on accurate detection of appearance defects to ensure safe and reliable operation.  However, traditional methods rely heavily on manual detection, making it difficult to meet the demands of industrialization. This work proposes a dual mode deep learning framework for automated defect detection in spent power batteries, aiming to satisfy both real time inspection and high precision detection requirements. In this framework, a YOLOv5 model enhanced by transfer learning is employed for rapid real time detection, while a Mask R-CNN model improved with SoftNMS is developed for high precision detection in overlapping target scenarios. Compared with conventional NMS, SoftNMS reduces missed detections by continuously decaying the confidence scores of overlapping candidate boxes rather than directly suppressing them. To support model training and performance evaluation, a specialized dataset of appearance defects in spent batteries is constructed. Experimental results show that the improved YOLOv5 achieves a precision of 96.7%, while the improved Mask R-CNN reaches a precision of 97.44%. In addition, a PyQt5 based visual inspection system is developed to verify the practical applicability of the proposed framework in automated battery sorting. The results demonstrate that the proposed method provides an effective technical solution for intelligent defect detection in the echelon utilization of spent power batteries.