查看文章

In modern industry, obtaining expert labels for bearing fault data is often difficult, which makes traditional detection methods (e.g., supervised learning) less effective and reduces the detection accuracy. Therefore, this paper proposes an unsupervised bearing fault detection based on Time-Frequency transform and generative adversarial networks (TF-FDGAN), which mainly consists of WGAN and izi _f encoder and uses the time-frequency map as input. The WGAN uses normal data for unsupervised learning to learn its distribution. Then, the izi _f structure is used to train the encoder to map the input image to space, and the loss function of the izi _f structure is used as an anomaly score to evaluate whether the input data is in normal working condition. The experimental results show that the method maintains high accuracy and stability in different datasets, and the detection accuracy is greatly improved compared with …