Multi Classification of Industrial Imbalanced Data Based on LSGAN

This article proposes the STFT-LSGAN-ResNet equipment fault diagnosis method for the multi classification problem of industrial big data in rotating machinery. This method uses short time Fourier transform (STFT) to extract feature maps from one-dimensional time-series vibration signal data of rotating machinery. Then, the LSGAN is trained using the feature maps of the fault class data, and the LSGAN ultimately generates as realistic a simulated fault class feature map as possible, in order to balance the original image dataset. Finally, this article uses the balanced image dataset to train a residual network model (ResNet) for multi classification. Based on the classification results, equipment fault diagnosis is made and the fault type is classified. The experiment used the ResN et model for image multi class classification, and the experimental results showed that the LSGAN model can effectively improve the problem of neglecting fault categories caused by data imbalance. The use of the LSGAN model has a significant improvement compared to direct classification.