关键词: 气缸套, 缺陷检测, 珩磨网纹, 网络优化

Abstract: Honing texture defects of the cylinder liners were the important quality indicator of an engine. Widely used manual quality inspection, however, owned problems of low detection accuracy, low efficiency, and manual uncertainty, as well as manual errors, and related automation detection research mightn not identify different types of defects. Therefore, a cylinder liner defect detection method was proposed based on an improved YOLOv8n model, which might accurately identify seven types of honing mesh defects. Firstly, the seven types of honing defects were defined based on geometric feature parameters.Then, the SC-C3 module was designed based on lightweight convolutional SCConv to reduce computational parameters of the model. Simultaneously, the channel prior convolutional attention(CPCA) mechanism was introduced to enhance the feature extraction ability of network. Finally, the Wise-IoU loss function was used to lower negative impacts of low-quality samples. The results show that the proposed detection method may effectively identify and distinguish composite defects under complex mesh backgrounds. The detection models mAP@0.5(mean of average precision of IoU is as 0.5) reaches 96.7%, and frame per second(FPS) approaches 476 frames/s. Moreover, the proposed model improves recognition accuracy by 2% and reduces computational load of 0.7 GFLOPs(Giga floating-point operations per second) compared to those of the YOLOv8n model. The paper provides an automated high-speed high-precision solution for detecting surface defects on cylinder liners.

Key words: cylinder liner, defect detection, honing texture, network optimization