A convolutional neural network based on a capsule network with strong generalization for bearing fault diagnosis

Abstract

Abstract Bearing fault diagnosis is a significant part of rotating machine health monitoring. In the era of big data, various data-driven methods, which are mainly based on deep learning (DL), have been applied in the field of fault diagnosis. As a common deep learning method, convolutional neural networks play an important role in the field of image classification. To take advantage of existing research results of convolutional neural networks that were originally designed for images, short-time Fourier transform is applied to convert signals to two-dimensional graphs. However, the traditional convolutional neural network with max-pooling layers may disregard the positional relationship between features. Inspired by dynamic routing capsule net, a novel capsule network with an inception block and a regression branch is proposed. First, during the preprocessing stage, the one-dimensional signals are transformed into a time-frequency graph. Second, the graph data are fed into the network, and two convolution layers are applied to extract higher information; following that, an inception block is applied to the output feature maps to improve the nonlinearity of the capsule. Then, after dynamic routing, the lengths of the capsules are used to classify the fault category. There are two other branches: one uses the longest capsule to regress the damage size of the capsule, and the other reconstructs the input graph. To verify the generalization ability of the proposed model, comparisons among the proposed model, conventional methods and state-of-the-art convolutional neural network models are carried out.