A Method for Imbalanced Fault Diagnosis Based on Self-attention Generative Adversarial Network

Abstract

In the real industrial scenario, the rotating machinery generally works in a normal state, and the sensor can only collect fault signals when the mechanical equipment fails in a few cases. This leads to the problem of data imbalance in fault diagnosis, i.e., the number of normal samples far exceeds the number of fault samples. In this case, the performance of the data-driven fault diagnosis classifier will significantly decrease, leading to misdiagnosis and missed diagnosis. To solve the above problem, we propose a novel method called WT-SAGAN-CNN based on wavelet transform and self-attention generative adversarial networks, which can generate high-quality fault samples to achieve data balance. This method uses continuous wavelet transform (CWT) to convert the one-dimensional signals into images. Then uses the self-attention generative adversarial networks (SAGAN) to generate fault images and stops training until the generative adversarial networks reach the Nash equilibrium. The convolutional neural networks are used as the fault diagnosis classifier, mix the generated images into the imbalanced dataset, and input them into the classifier for training. Experiment shows that the proposed model can generate samples similar to real samples, and as the generated samples continue to be added to the imbalanced dataset, the accuracy of the fault diagnosis classifier has also been significantly improved.