Multidomain variance-learnable prototypical network for few-shot diagnosis of novel faults

A multidomain variance-learnable prototypical network (MVPN) is proposed to learn transferable knowledge from a large-scale dataset containing sufficient samples of multiple faults for few-shot diagnosis of novel faults (i.e., disjoint with fault types in the large-scale dataset). Signal characterizations in time, frequency, and time-frequency domains are first constructed to make full use of information contained in severely limited labeled data. Mahalanobis distance is proposed as a criterion for improving classification performance by considering different spreads between classes in the embedding space. The spread variance of each class is learned by constructing an additional deep learning network in the original prototypical network. Multidomain signals are used to learn the prototype representations and spread variances separately, and are finally fused for classification. With the proposed MVPN, deeper variance-learnable embedding learning from wider domain characterizations improves the ability of few-shot fault diagnosis. Experiments are conducted to evaluate the performance of MVPN using datasets collected from a benchmark bearing and a Delta 3-D printer. Results indicate that the proposed MVPN performs competitively compared to state-of-the-art few-shot learning algorithms.