A fault diagnosis method for multi-class imbalanced data based on compressed sensing and adaptive oversampling techniques

Abstract

In the context of fault detection for rotating machinery, such as bearings, the high noise levels, large data volumes, and imbalanced data distributions in the collected signals have consistently impacted the accuracy of fault detection. To address these challenges, this paper proposes a multi-class imbalanced data fault diagnosis method that integrates compressed sensing and adaptive oversampling techniques. First, an improved compressed sensing technique is employed to compress and reconstruct the original signals, thereby removing noise and redundancy while extracting significant features to construct a feature data set. Second, to mitigate the data imbalance issue, an adaptive synthetic minority oversampling method based on natural neighborhoods is introduced. Natural neighborhoods are constructed for all classes in the multi-class imbalanced data set, and adaptive oversampling is performed on minority class samples to create a balanced data set. This approach alleviates the problem of insufficient fault sample information during the fault detection process. Experiments were conducted on six imbalanced data sets derived from two public data sets and compared with seven state-of-the-art methods. The experimental results demonstrate that the proposed method significantly enhances the accuracy of fault detection, providing a more robust foundation for the reliable operation of motor-driven rotating mechanical equipment.