Deep learning-based fast recognition of commutator surface defects

With low accuracy and poor efficiency, the manual and traditional detection of commutator surface defects cannot meet efficiency and timeliness requirements. Thus, a method is proposed to detect and identify commutator surface defects. YOLOv3 target detection is applied to commutator surface defect detection and recognition, and the model size and parameter number are reduced without significantly reducing detection accuracy. First, this paper proposes a separable residual module based on deep separable convolutions and residual networks. A network with shallower layers and fewer channels is designed for quick detection and recognition of commutator surface defects. Finally, the algorithm herein is evaluated on SD_data. The experimental results indicate that while maintaining good accuracy, this method has a smaller model size, fewer parameters and faster running time than the YOLOv3 network. Moreover, accuracy is improved to a certain extent compared to the current commutator surface defect detection methods, so it can be better applied to real-time detection and recognition of commutator surface defects.