Data Augmentation Based on Varying-parameter Time-frequency Analysis for Imbalanced Fault Diagnosis of Rolling Bearings

The health condition of rolling bearings is crucial to the safety of rotating machines in aerospace engineering. However, the bearing abnormal samples are rare in engineering practice, thus the historical bearing dataset is imbalanced between the normal and abnormal samples, which degrades the generalization of traditional intelligent diagnosis model. This paper proposes a new data augmentation method by using varying-parameter time-frequency analysis to expand the data amounts of bearing abnormal samples for imbalanced bearing fault diagnosis. First, each of the abnormal signals is converted to several time-frequency spectra by the time-frequency analysis with varying time-frequency parameters. Then, the time-frequency spectra of the abnormal samples are combined with those of normal samples using single time-frequency parameter to construct rebalanced bearing dataset. Finally, the rebalanced dataset is used to train a model of convolutional neural networks (CNN) for classification of different bearing health conditions. The proposed data augmentation method is applied to a bearing dataset with the imbalanced ratio of 100:1 between the normal and abnormal samples, and is compared to three existing data augmentation methods, which proves the superiority of the proposed varying-parameter time-frequency analysis method on classification accuracy and clustering performance.