An unsupervised fault diagnosis method for rolling bearing using STFT and generative neural networks

In recent years, the technique of machine learning or deep learning has been employed in intelligent fault diagnosis methods to achieve much success using massive labeled data. However, it is generally difficult or expensive to label the monitoring data in practical engineering due to its complex working conditions. Therefore, an unsupervised fault diagnosis method is proposed in this paper for rolling bearings, which incorporates short-time Fourier transform (STFT) as well as categorical generative adversarial networks (CatGAN). The proposed method first adopts STFT to transform raw 1-D vibration signals into 2-D time-frequency maps to serve as the input of CatGAN. Then, it obtains a CatGAN model via an adversarial training process to generate fake samples with a similar distribution to the maps extracted by STFT and cluster the input samples into certain categories. Furthermore, the performance of the proposed ST-CatGAN method is verified using a classic rotating machinery dataset, and the experimental results demonstrate its high diagnosis accuracy and strong robustness against the motor load changes.