Optimal transport strategy-based meta-attention network for fault diagnosis of rotating machinery with zero sample

Deep learning-based methods are widely applied to fault diagnostics, which depend on adequate samples. However, fault samples are often limited and even not available in industrial applications. To solve this problem, an optimal transport strategy-based meta-attention network (OTS-MAN) is proposed for the fault diagnosis by exploiting the fault knowledge learned from few source domains to diagnose the target domain with zero sample. Firstly, a new meta-attention network is built to mine discriminative features of each class from the source domain. Then, an optimal transport strategy is designed to align the feature distribution of each category between known fault in the source domain and unknown fault in the target domain. Finally, the similarity scores are obtained to assess the health status of the target domain. The proposed OTS-MAN is trained only with known source domain data and can diagnose unknown faults without previous access to target domain data. The validity of the proposed method is implemented through using two cases. The results indicate that the OTS-MAN has a better fault diagnosis accuracy than existing methods, and its noise immunity is also improved.