Research on Gear Surface Damage Identification Based on the ResNet Network

Abstract

Abstract Gear damage will directly affect the normal operation of the whole transmission system. Aiming at the problem of gear surface damage recognition, a gear surface damage recognition model based on the improved ResNet-34 network is proposed. Firstly, the most suitable network model is selected by training the official describable textures dataset. Secondly, channel pruning, decomposition of the large convolutional kernel, application of global average pooling and transfer learning operations are performed on the primitive network to reduce the computational and parametric quantities and accelerate the training. Finally, experiments are conducted to compare the training effects of the improved network by introducing different pruning methods and scaling factors to the dataset. The experiments show that for the gear surface image dataset, the Slim Pruning (20% Pruned) method results in a 20.78% reduction in computation, a 21.71% compression in the number of parameters, and a 7.38% reduction in the average time per training epoch at a relative error rate of only 1.17%, thus showing that channel pruning can effectively compress the model within the accuracy error and reduce the training cost.