A batch-adapted cost-sensitive contrastive feature learning network for industrial diagnosis with extremely imbalanced data

Data-driven intelligent fault diagnosis models provide a great convenience for industrial diagnosis applications. However, in real industrial scenarios, most machines are in normal condition, and the collected amount of normal monitoring data is much greater than that of fault data. The extremely imbalanced monitoring data makes the existing models unable to output satisfactory diagnosis results. In this paper, we propose a batch-adapted cost-sensitive supervised contrastive feature learning network to address this issue. In the proposed network, we use supervised contrastive learning based on data augmentation to extract features from fault data of minority classes in the imbalanced set. Besides, the channel attention mechanism is applied to further improve the sensitivity of the extracted fault features. Finally, we design a batch-level cost-sensitive loss function, which is adjustable adaptively according to every training batch data, to ensure that the proposed network can accurately identify fault samples in the minority classes. We performed diagnostic experiments with different imbalance ratios under two experimental cases and added significant noise to validate the performance of the model. The results indicate the excellence as well as the potential of our proposed network for fault diagnosis in extremely imbalanced data scenarios and demonstrate considerable robustness.