A parameter based transfer learning fault diagnosis method under different working conditions

In industrial applications, equipment is often worked under multiple working conditions, making it prone to various failures. Because of the lack of enough training data, the use of data-driven fault diagnosis methods is often restricted. In this paper, to address such a problem, a parameter based transfer learning(TL) method for few-shot fault diagnosis under different working conditions is proposed. In the methodology, Marginal Distribution Adaptation (MDA) is first used to decrease the divergence via minimizing the maximum mean distance between two marginal distributions of target and source domain. Then, the fault diagnosis model is built by training the from new target domain dataset and new source domain dataset. Finally, part of new target domain dataset are used to test the diagnosis accuracy of the model. Experimental results on bearing benchmark data sets from the University of Ottawa validate the proposed method. Compared with other effective fault diagnosis approaches reported in the literature, the proposed method can obtain higher diagnosis recognition rates and stronger robustness.