Research on bearing vibration signal generation method based on filtering WGAN_GP with small samples

In the practical application of bearing fault diagnosis, the data imbalance problems caused by the lack of available fault data lead to inaccurate diagnosis. The high cost and difficulty of obtaining fault samples has become an obstacle to the development of intelligent diagnosis technology. Aiming at the problem of data imbalance caused by small samples, this paper proposes a data generation method called FEF_WGAN_GP based on Wasserstein generative adversarial networks with gradient penalty (WGAN_GP) and feature Euclidean distance filtering (FEF) theory. Firstly, WGAN_GP is used to obtain signals with similar distribution to the small sample data, which can alleviate the imbalance of the dataset. Then, the FEF method is used to filter the generated data in order to obtain a higher quality of the samples. In the test validation part, not only the used dataset is evaluated to obtain a more reasonable dataset, but also the generated signals are evaluated from multiple perspectives. In addition, this paper evaluates the effects of the number, length and signal-to-noise ratio of the parent data on the quality of the generated signals, as well as the effect of the setting of the threshold of the data filtering method on the accuracy of the classifier. The experimental results indicate that this method performs well in processing unbalance fault data. It has better stability and diagnostic accuracy than the current stable method.