【摘要】 在电动汽车退役动力电池的回收再利用场景中，实现自动化拆解对提高拆解效率、减少人力成本及降低安全风险具有重要意义。针对拆解过程中复杂的工作环境，传统单通道视觉算法在目标检测上表现出问题，因此，提出了一种基于深度学习Yolov5s算法的改进型退役动力电池包零件及连接形式目标检测模型——FDE-Yolo。该模型通过三重策略优化Yolov5s算法：一是优化主干特征提取网络，将CSPDarknet53替换为FasterNet以减少运算规模；二是设计Dilated ReparamNCSPELAN4模块，替代C3模块以增强特征提取融合能力；三是引入ECA(efficient channel attention)注意力机制，提升对关键特征信息的关注度。实验结果表明，FDE-Yolo模型在数据集增广处理后，mAP(mean average precision)值较原模型提升4.0%,达到84.8%,模型参数量由79.6 M减少至31.6 M,运行速度提升4倍，完成了轻量化改进，显示出更高的稳定性、精确度和运行速度，适用于退役动力电池包的拆解场景。更多还原

【Abstract】 In the recycling scenario of EoL electric batteries, automated disassembly is of great significance to improve disassembly efficiency, reduce manpower costs and lower safety risks. For addressing the limitations of traditional single-channel vision algorithms in target detection within complex disassembly environments, this paper proposes an improved target detection model for components and connection configurations of EoL battery packs based on the deep learning Yolov5s algorithm-FDE-Yolo. The model optimises the Yolov5s algorithm through a triple strategy: firstly, it optimises the backbone feature extraction network by replacing the CSPDarknet53 with FasterNet to reduce the operation scale; secondly, the Dilated ReparamNCSPELAN4 module is designed to replace the C3 module to enhance the feature extraction fusion capability; and thirdly, the ECA(efficient channel attention) attention mechanism is introduced to enhance the attention to the key feature information. The experimental results show that the FDE-Yolo model achieves the mAP(mean average precision) of 84.8%, a 4% improvement compared with the original model after the dataset augmentation and generalisation process, increases the operation speed by 4 times, and reduces the number of model parameters significantly from 79.6M to 31.6M. These improvements yield a lightweight model with higher stability, accuracy and operation speed, making it suitable for the disassembly scenarios of EoL power battery packs.更多还原