A Novel Cross-Scenario Transferable RUL Prediction Network With Multisource Domain Meta Transfer Learning for Wind Turbine Bearings

Abstract

Accurately predicting the remaining useful life (RUL) of bearings is quite significant for ensuring the operation reliability of wind turbines. Due to the limitation of real-world life cycle data, the accuracy of existing wind turbine RUL prediction methods needs to be improved. This article suggests a multisource domain meta transfer learning (MD-MTL)-based cross-scenario transferable RUL prediction method. The MD-MTL uses test rig data on multiple operating conditions to build up the prediction model and migrate the prediction knowledge to real wind turbines. In addition, a novel RUL prediction network based on a convolutional encoder and ProbSparse multihead attention decoder (CE-PSAD) is proposed for the improvement of prediction accuracy. It can extract key degradation features and long-term dependence relationships from raw vibration signals, and complete the precise mapping to RUL. The efficiency of the proposed method is proven based on two test rig datasets and two wind turbine bearing datasets.