Intelligent classification framework for gear surface damage and gear type using CNN transfer learning

Abstract

In the mechanical transmission system gear has a very important role, and Gear surface damage is an important factor affecting gear transmission. Regular inspection of each gear and its surface damage is of great significance for ensuring the stable operation of a whole mechanical system. To improve the identification efficiency and accuracy for gears and their surface damage, In this paper, we propose a gear and its surface damage recognition method based on PyTorch deep learning library, expand the dataset using data augmentation and data extension techniques, compare the performance of gear and its surface damage recognition under three typical models, namely AlexNet, VGG16, and ResNet-101, and optimize each training model through migration learning and hyperparameter comparison. The results show that the speed and accuracy of model training can be improved by applying transfer learning. By extending the dataset using data augmentation techniques, the robustness of the network is improved considerably. When the batch size is 6 and the initial learning rate is set to 0.01, the model training effect is the best, and ResNet has higher recognition accuracy and stability than AlexNet and VGG16, which is more suitable for the classification of gears and their surface damage. In order to find the optimal ResNet model, four ResNet models with different number of layers, ResNet-34, ResNet-50, ResNet-101 and ResNet-152, were compared, among which the ResNet-101 model showed the optimal performance in gear surface damage recognition.