Degradation prediction for mechanical components based on transfer learning and a multistage statistical model

Abstract

Rolling bearings serve as vital elements in mechanical components; their degradation significantly influences equipment reliability and safety. Previous studies primarily focus on single-stage degradation methods, limiting their applicability for predicting degradation trends across varying conditions and hindering generalization. To address this challenge, we propose a multistage statistical model and a multistage transfer learning based on bidirectional gated recurrent unit with an attention mechanism (MSTL-BGRUA) to predict the multistage degradation trend of bearings under cross-working conditions. Given the nonlinear properties of bearing degradation, we present an online multistage division method to identify degradation characteristics. The degradation features of bearing’s vibration signal are extracted and classified into multiple degradation stages based on a mutation identification policy via the statistical model. Subsequently, a BGRUA model is proposed to balance forward and backward information and assign different weights, achieving accurate prediction for each identified degradation stage. Based on the presented BGRUA, we incorporate an MSTL method for degradation prediction to maintian accuracy under crossing working conditions. Finally, a prognosis framework using both the statistical model and the MSTL-BRGUA model is presented to predict bearing status under different degradation stages. A validation of the presented method is conducted using a vibration dataset, demonstrating its superior predictive performance compared to other advanced degradation prediction methods.