Research on surface defect detection and fault diagnosis of mechanical gear based on R-CNN

Abstract

Gears are the basic units in modern power systems. The detection and diagnosis of surface defects and faults in gears are conducive to improving product quality, ensuring the safety of mechanical equipment, and reducing maintenance costs. However, the accuracy of manual and traditional automated target detection algorithms is not satisfactory. Therefore, this research uses the R-CNN algorithm for gear detection, improves its non-maximum suppression algorithm and multi-task loss function, and obtains the improved Faster R-CNN algorithm. The test was carried out on the built data set. The actual measurement shows that the recall rate of the improved Faster R-CNN is up to 0.951 and the lowest is 0.816. Its AP value is as low as 0.677 and as high as 0.858, and the mAP value is 0.843. Horizontal comparison, the comparison results show that the mAP of Faster R-CNN is 0.80 1, second only to R-CNN among the tested algorithms, and 8.83% higher than the original Faster R-CNN. Under the condition of [email protected]:0.95, among all the tested algorithms, its AR index is the highest at 54.3, and the detection speed is 18 FPS/s. Although the detection speed has decreased, the detection and recognition accuracy has been significantly improved, which provides feasibility for the R-CNN series of algorithms new optimization directions. The research provides a better automatic detection method for product quality inspection in gear manufacturing industry.