Gear Fault Detection Method Based on the Improved YOLOv5

Abstract

Gears are used as transmission elements in a wide range of industries, so detecting faults in them is important. Current deep learning-based fault detection is difficult to apply to industrial embedded devices due to the complexity of the model and the huge computational effort. To address this problem, we propose a lightweight gear fault detection model, LG-YOLOv5. To obtain a lightweight network, the introduction of ShuffleNetV2 and GSConv. Then, to ensure excellent detection performance, we integrate a multi-span hybrid spatial pyramid pooling model, attention mechanism modules and cross-scale feature pyramids to improve the detection performance. Finally, to evaluate the gear fault detection capability of the LG-YOLOv5 on the Rockchip RK3568 embedded platform. Image acquisition to create a gear fault dataset. Experimental results show that the LG-YOLOv5 model has a volume of S.SM, which is only 61.5% of the YOLOv5 model, a computational cost of 13.6% of the YOLOv5, a 45% increase in detection speed and a 1.5% increase in accuracy, and is able to accurately identify gear faults such as wear, bulging and missing tooth.