Deep Learning Fault Diagnosis Method Based on Feature Generative Adversarial Networks for Unbalanced Data

Due to its powerful feature representation capabilities, DNN is quickly being used in the field of fault diagnosis. When the samples used to train DNN are unbalanced, the fault features extracted from only a small number of fault samples via DNN can be masked by large number of normal samples, which may result in high misclassification rate. This paper solves this problem by using feature Generative Adversarial Network (FGAN) and stacked Auto-Encoder (SAE). The main innovation is to design a feature guided generator to improve the quality of the generated samples. The generator designed can generate more qualified new samples in the sense that it uses the fault feature extracted by SAE to guide the training of the generator. Rolling bearings experiments is used to show the effectiveness of our proposed algorithm.