258 / 2021-04-15 16:58:35

A transfer learning method for bearing fault diagnosis

Fault diagnosis; Transfer learning; Convolution neural network; Siamese network

Special Sessions > Dynamics analysis and PHM of hydroelectric machinery

In the era of big data, intelligent mechanical fault diagnosis technology based on neural network is an important tool to ensure the healthy operation of equipment. However, in practical engineering, machines are mostly in healthy and faults seldom happen. It’s difficult to collect massive high-quality fault data. In order to overcome the problem of lack of a large number data, the theory of transfer learning is introduced into the intelligent mechanical fault diagnosis technology. The knowledge acquired from the source domain which is similar to but not same as the target domain is used to solve the problem of the target domain. Therefore, a deep transfer diagnosis method based on convolution neural network and siamese network is proposed. In the proposed method, the parameters-shared convolution neural network is first used to extract the features of the source and target domain. And the similarity between source domain features and target domain features is calculated by contrastive loss function. Then, the regularization terms of domain adaptation are introduced into the training process of the deep transfer diagnosis model. The proposed method is validated by a case study on a bearing data sets. The results show that the proposed method is able to use the fault diagnosis knowledge from source domain to identify the health states of the bearing.