StRUL: stacking-based RUL prediction for bearings with unknown working conditions and failure modes

Abstract Bearings are critical components in equipment, and it is important to predict their remaining useful life (RUL) so that early intervention can be applied before their failure. A challenge in predicting the RUL is that the target bearing’s working condition and/or failure mode is usually unknown. Although various methods have been proposed to address this challenge, few of them consider two practical issues. (1) When various datasets are available and used to train a single deep learning model, training datasets related to bearings of quite different statuses can affect prediction accuracy. (2) Transfer learning can be used to alleviate issue (1), but this requires data collection from the target bearing and fine-tuning, and the additional time required by this process may delay RUL prediction and intervention. To address these issues in industrial practice, we propose stacking-based RUL (StRUL) prediction for bearings of unknown working conditions and failure modes. StRUL is based on the stacking of transformers with novel designs: a modified amplitude spectrum comparison approach, a similarity-based attention mechanism, and a distribution-based attention mechanism. First, StRUL pre-trains each transformer with a specific dataset so that each transformer can generate an encoding for the input data from the target bearing. Second, it applies a modified amplitude spectrum comparison approach to calculate the similarity value between the input data and each transformer’s training dataset. StRUL then uses a similarity-based attention mechanism to prioritize transformer encodings with relatively large similarity values in prediction. Third, it includes an additional transformer trained with all the training datasets and uses a distribution-based attention mechanism to determine how much the additional transformer contributes to the predicted RUL when the distribution of the similarity values is nearly uniform. Case studies performed using the XJTU-SY and PRONOSTIA data, each containing more than 10 datasets, demonstrate that StRUL can efficiently use all the training datasets without additional data collection or fine-tuning to achieve high prediction accuracy and speed useful for practical deployment.