RSE-YOLO: error-proofing method for solenoid valve assembly process integrating multi-scale and attention mechanism

Abstract This study proposes an improved RSE-YOLO (Res2Net and Squeeze-and-Excitation Networks) model to address the challenges of error-proofing sequence detection of parts during manual assembly of solenoid valves, especially in detecting small parts and overcoming the problems caused by hand occlusion. The proposed model introduces the Res2Net module into the YOLO backbone network to enhance the representation ability of multi-scale features and integrates the Squeeze-and-Excitation (SE) attention mechanism into the C3 module to improve the attention of key feature channels. In addition, the model uses an improved ELU activation function to speed up convergence. The experimental results demonstrate that the RSE-YOLO model achieves a mean average precision (mAP0.5) of 97.1% and a detection speed of 84.3 frames per second (FPS). Compared to the original YOLOv5s model, the accuracy is improved by 2%, and the detection speed is increased by 2.31%. Additionally, a dedicated platform was constructed, and experimental validations were conducted. The findings confirm that the proposed method exhibits excellent real-time performance and high reliability when applied to the assembly lines for small solenoid valve components.