Few-shot Learning for Rolling Bearing Fault Diagnosis Via Siamese Two-dimensional Convolutional Neural Network

摘要

Data-driven approaches such as deep learning have made great achievements in the field of fault diagnosis. However, most deep learning models acquire a large number of annotated samples to support their training. And for bearing fault detection, it is often difficult or even impossible to collect enough signals for each fault type. In this paper, we demonstrate how a few-shot learning method can be applied in fault diagnosis to lower the amounts of data required to make meaningful predictions. Specifically, we first use a Siamese two-dimensional convolutional neural network to extract the feature vectors of the input fault signal pairs. Second, the absolute difference (L1 distance) between the feature vectors is computed and then input to a fully connected layer with a sigmoid activation function to assess whether the input signal pairs belong to the same class. The Case Western Reserve University bearing data set is used to test the performance of the proposed method. Experimental results show that the proposed few-shot learning approach can obtain good accuracy when training samples are limited.