Contrastive Generative Replay Method of Remaining Useful Life Prediction for Rolling Bearings

Recent research has shown that deep learning algorithms used in prognostics of remaining useful life (RUL) have the potential to greatly benefit industrial automation. However, these algorithms are typically trained on a fixed situation that rarely occurs in engineering applications. When confronted with different working conditions, the model’s performance may unpredictably decrease. To address this issue, continual learning methods allow models to incorporate new knowledge after deployment. Despite being a relatively new field, current research on continual learning for RUL prediction relies on regularization-based methods, which do not perform well in noisy and diverse problems such as RUL prediction in bearings. Moreover, traditional memory-based methods for bearings RUL prediction are also ineffective. As a solution, this article proposes a contrastive generative replay (GR) method for continuous bearings RUL prediction. The proposed method utilizes contrastive learning when training GR models, allowing the model to better learn new and old data features and improve the effectiveness of continual learning. Additionally, the method is verified using the FEMTO-ST dataset and compared to state-of-the-art algorithms. Experimental results demonstrate the effectiveness of the proposed method in the task of continuous RUL prediction for bearings.