An Intelligent Multi-Local Model Bearing Fault Diagnosis Method Using Small Sample Fusion

Abstract

It is essential to accurately diagnose bearing faults to avoid property losses or casualties in the industry caused by motor failures. Recently, the methods of fault diagnosis for bearings using deep learning methods have improved the safety of motor operations in a reliable and intelligent way. However, most of the work is mainly suitable for situations where there is sufficient monitoring data of the bearings. In industrial systems, only a small amount of monitoring data can be collected by the bearing sensors due to the harsh monitoring conditions and the short time of the signals of some special motor bearings. To solve the issue above, this paper introduces a transfer learning strategy by focusing on the multi-local model bearing fault based on small sample fusion. The algorithm mainly includes the following steps: (1) constructing a parallel Bi-LSTM sub-network to extract features from bearing vibration and current signals of industrial motor bearings, serially fusing the extracted vibration and current signal features for fault classification, and using them as a source domain fault diagnosis model; (2) measuring the distribution difference between the source domain bearing data and the target bearing data using the maximum mean difference algorithm; (3) based on the distribution differences between the source domain and the target domain, transferring the network parameters of the source domain fault diagnosis model, fine-tuning the network structure of the source domain fault diagnosis model, and obtaining the target domain fault diagnosis model. A performance evaluation reveals that a higher fault diagnosis accuracy under small sample fusion can be maintained by the proposed method compared to other methods. In addition, the early training time of the fault diagnosis model can be reduced, and its generalization ability can be improved to a great extent. Specifically, the fault diagnosis accuracy can be improved to higher than 80% while the training time can be reduced to 15.3% by using the proposed method.