Imbalanced fault classification of rolling bearing based on an improved oversampling method

Many works of bearing fault diagnosis based on vibration signals have been present. However, most of them work under ideal conditions that the fault data are enough, but it is unrealistic in real applications. In this paper, we proposed a method to address the problem. It is a pipeline method of ensemble feature engineering, data improvement, and classifier design. First, multiple features of the bearing vibration signal are extracted. And then, the fault data are augmented by a novel oversampling method. Finally, the faults with a stacking classifier were identified. The proposed oversampling method can remove noise instances before and after oversampling, to overcome the risk of synthesizing noisy instances. It is not a new oversampling paradigm, but an enhancement module of the existing methods. Relying on it, four common oversampling methods improve the classification performance by 1-2% on publicly available NASA software defect datasets. Empirical results indicate the superiority of the proposed method: When the imbalanced ratio is 10:1, our method can obtain 99.2% and 91.5% accuracy rates on CWRU and Paderborn datasets, respectively.