Ensemble Data Augmentation for Imbalanced Fault Diagnosis

Data imbalance is a prevalent issue in industrial fault diagnosis, and generative adversarial networks (GANs) have gained significant attention and application in recent years as a solution. GANs address data imbalance by augmenting minority-category data, leveraging their strong generation capability. However, the training process of GANs often suffers from instability, resulting in a skewed distribution of the generated data. To address this challenge, this paper proposes a novel GAN-based ensemble data augmentation framework that draws inspiration from ensemble learning principles. The framework aims to enhance data augmentation through the cooperative and competitive utilization of generated data. In the cooperative aspect, we introduce multi-source data augmentation (MSDA) to mitigate the problem of skewed generated data by combining multiple independently trained GANs as generative sources. To further improve the diversity and efficiency of training multiple GANs, we incorporate cross training and parallel training techniques into the cooperative MSDA (referred to as ct-MSDA). These techniques provide significant support for the practical application of our proposed method. Additionally, in the competitive aspect, we design a data filtering method for df-ct-MSDA. This method evaluates the generated data and selects those with higher scores for participation in the training process. Finally, the effectiveness of our ensemble data augmentation framework is demonstrated through exemplary industrial cases exhibiting typical data imbalance issues. Our proposed method showcases outstanding performance in these scenarios, further validating its superiority.