Information Fusion-Based Meta-Learning for Few-Shot Fault Diagnosis Under Different Working Conditions

Abstract With the development of deep learning and information technologies, intelligent fault diagnosis has been further developed, which achieves satisfactory identification of mechanical faults. However, the lack of labeled samples and complex working conditions can hinder the improvement of diagnostics models. In this article, a novel method called Information Fusion-based Meta-Learning (IFML) is explored for fault diagnosis with few-shot problems under different working conditions. Firstly, an information fusion and embedding module is applied to perform both data- and feature-level fusion of multi-source. The embedding module only contains one input layer and multiple convolutions, residual and batch normalization (BN) layers, which has the advantage of low computational cost and high generalization. Then the prototypical module is proposed to reduce the influence of domain-shift caused by different working conditions using the fusion representation, which can improve the performance of fault diagnosis. The approach is verified on artificial and real faults under 4 different working conditions from the KAt-DataCenter at Paderborn University. For the 3-way 1-shot classification on Task T1, the average testing accuracy of the proposed method is 97.14%. For the K-shot classification on different tasks, the proposed method achieves the highest average testing accuracy of 94.21%. The results show the proposed method outperforms other typical meta-learning methods in terms of testing accuracy and generalization capability.