Research on Gear Surface Damage Recognition Based on Small Sample Deep Learning

Abstract

Gear surface damage is an important factor affecting gear transmission. It is extremely important to improve the efficiency and accuracy of gear surface damage identification. ResNet recognition model of gear surface damage is established based on Pytorch architecture, dataset is expanded by means of data enhancement, model training is optimized by means of transfer learning, and four ResNet structures are compared. The results show that the dataset composed of 640 images after the enhancement of 64 original image is not enough to meet the needs of model training for a large amount of data; using transfer learning can improve the speed and accuracy of model training, and meet the requirements of gear surface damage identification; the ResNet-101 model is the optimal structure in this framework. This research has important scientific significance and engineering value for the recognition of gear surface damage.

Keywords

• Convolutional neural network;Gear surface damage;Deep learning;Transfer learning