Multi-domain feature fusion and sparse regression-driven physical information neural network for remaining useful life prediction

Abstract

Rolling bearings, as critical components of mechanical equipment, are widely used in aircraft engines, marine propulsion systems, and electric motor drive systems. Most bearings operate under harsh conditions involving high loads and temperatures, and prolonged operation can lead to various types of bearing failures, potentially causing safety incidents or equipment damage. Therefore, accurately predicting their remaining useful life (RUL) is of significant importance. Current bearing life prediction methods face the following challenges: Most statistical features contain redundant information and noise components, which cannot provide effective global degradation information for bearing life prediction; Traditional physical information neural network (PINN) utilizes a deep hidden physical model (DeepHPM) to identify nonlinear relationships between RUL and input data. However, when the input data contains health indicator (HI), this module suffers from underfitting. Additionally, traditional PINN only assigns physical loss hyperparameters in the loss function, neglecting the balance between data loss and physical loss. To address these issues, this paper proposes a sparse regression-driven PINN prediction framework. First, principal component analysis is used to fuse multi-domain statistical features to construct a bearing health indicator dataset; Then, sparse regression is introduced in the physical information network to identify partial differential equations (PDE) that align with actual physical laws; Finally, a brand-new loss function weighting strategy was introduced to enhance the model’s ability to learn degradation information in bearing health indicators. Experiments on the XJTU-SY and SMU datasets validated the effectiveness and good generalization performance of the proposed model method.