A Fault Diagnosis Method for Unbalanced Data Based on a Deep Cost Sensitive Convolutional Neural Network

Abstract

For the research of industrial machinery intelligent fault diagnosis, traditional methods generally require that the samples share the same distribution or are balanced. In fact, the machines switch working conditions frequently during operation, accordingly resulting in changes in data distributions and the data can be unbalanced. To solve the above, combining transfer learning method, an intelligent diagnosis method for imbalanced data based on Deep Cost Sensitive Convolutional Neural Network is proposed. According to this model, the cost sensitive classification loss function adaptively assigns different error classification costs for all class, and it can mitigate data unbalancing. The domain adaptation is constructed by domain adversary and distance measure to automatically learn domain invariant features. The proposed method is verified by transfer experiments under various working conditions with three published bearing vibration data sets, and compared with other method under different imbalanced ratio. The results show that the proposed method is effectiveness and able to classify the imbalanced data.