SA-DCWGAN-GP Based Sample Generation Model and Bearing Fault Diagnosis

For the problem of low diagnostic accuracy due to less fault data in real working conditions, this paper is based on deep convolutional neural network architecture, replacing the traditional cross-entropy loss function with Wasserstein distance and combining the gradient penalization GP term, which finally effectively mitigates the gradient disappearance and pattern collapse problem of the traditional Gan, and on the basis of which, adding the self-attention mechanism in the discriminator makes the discriminator model better converged. Model converges better. In the process of fault diagnosis, the wavelet transform is firstly used to change the one-dimensional time series signal into the time-frequency domain signal; then the SA-DCWGAN-GP model is used to expand the data samples, and the samples generated by it are evaluated for their similarity; finally, the balanced dataset is inputted into each classification model. Experiments show that the method learns the distribution of the original data, and the balanced dataset achieves a significant improvement in the fault diagnosis accuracy on each classifier; the method proposed in the paper has better results in sample generation than other models.