An Intelligent Bearing Fault Diagnosis Method with Transfer Learning from Artificial Damage to Real Damage

Abstract

Traditional intelligent fault diagnosis methods for bearings have achieved much success when the training data and testing data are collected from the same type of damage generation. In actual industrial environments, it is difficult to collect enough real damaged samples, but artificial damaged samples are easy to obtain. In this paper, a novel transfer learning method with convolutional neural network (CNN) from artificial damage to real damage for bearing fault diagnosis is proposed. Firstly, the CNN model is constructed and pre-trained by huge and easily obtained artificial damaged samples. Then, parameters of some layers from the pre-trained CNN model are frozen to save the same feature between artificial damaged samples and real damaged samples. Finally, a small number of real damaged samples are used to fine-tune the model. Experiments on an open bearing fault database showed that the proposed method can not only improve the diagnosis accuracy in the real damage dataset, but also save training time and samples. The superiority of the proposed method is also proved by comparing with other classical methods and visualizing the learned features of the networks.