I. Introduction
With the advent of Industry 4.0, traditional industries are transforming to achieve automation, digitization, and intelligence, which makes the demand for scientific and efficient health management of mechanical equipment more urgent. The signal features fed into the conventional machine learning (ML) models, e.g., multilayer perceptron (MLP) and support vector machine (SVM), are usually concentrated in the time domain, the frequency domain, and the time–frequency domain. These models are only suitable for samples with small data sizes and simple patterns. The recent deep learning models represented by convolutional neural networks (CNNs) [1], [2], autoencoder (AE) [3], and generative adversarial networks (GANs) [4] have been widely applied in the field of fault diagnosis [5]. The feature extraction still focuses on the data values but lacks exploration of their relationship and structure hidden in the data.