A variational transformer for predicting turbopump bearing condition under diverse degradation processes

Accurate condition prediction is necessary to ensure the reliability of the turbopump components. Meanwhile, with the ever-increasing complexity of the turbopump system, the corresponding degradation processes of the turbopump bearings are also increasingly diverse. Consequently, the investigation into the prediction of the turbopump bearing condition is of great significance. However, current research mostly reported on the remaining useful life prediction and neglected the predictive analysis based on the object's health condition. To address the problem, this paper proposed a combinational framework for the turbopump bearing condition monitoring and prediction. Firstly, a multi-branch residual network is designed to construct the health indicators (HIs), which are intended to indicate the health condition of the objects. Then, a Transformer model-based predictor is proposed to predict the constructed HIs accurately. By implanting the variational mechanism in the network, the predictor can achieve high accuracy under diverse degradation processes. To demonstrate the effectiveness of the proposed approach, a public whole-lifetime bearing dataset and a turbopump bearing dataset are utilized in the contrast experiments. Compared with some existing approaches, the proposed framework can obtain more reliable HIs and achieve higher prediction accuracy under diverse degradation processes.