Real-time and online industrial system for detecting defects of interface surface in automobile brake pipe

A defect detection algorithm based on dual-level cascade and multi-feature scale fusion Faster Region based Convolutional Neural Network (Faster-RCNN) is proposed to overcome the challenges of uneven distribution, small size, and diverse morphology of interface surface defects on the automotive brake pipe. Basing on Faster-RCNN, the proposal of multi-scale feature fusion combines feature information from different scales together to ensure that easily overlooked small targets can be noticed. Region of Interest (RoI) alignment pooling is proposed to reduce the deviation between the original feature regions and the mapped feature regions. To reduce the impact of uneven sample distribution on algorithm performance, a two-stage cascade structure is constructed to enable that different samples use different Intersection over Union (IoU) thresholds. The design of the loss function was optimized to balance the learning ability of the algorithm between easy and difficult samples. A “Defect” dataset is constructed and it validates the effectiveness of the improved algorithm. The improved algorithm achieves Precision (P) of 90.3%, Recall (R) of 84.5%, and mean Average Precision (mAP) of 91.1%. It has been demonstrated that the improved Faster-RCNN defect detection algorithm exhibits high accuracy and robustness. According to the production requirements of the actual industrial production, an integrated online defect detection system has been established, which combines testing for rotation integrity, groove completeness, and interface surface defects. The experimental results indicate that the system achieves an accuracy of 99.4%, a false positive rate of 1.2%, a false negative rate of 0.5%, and a detection time of 0.78 s. The integrated online defect detection system reduces the reliance on manual inspection, enhancing production efficiency and ensuring consistent quality, thereby meeting stringent production standards.