Cross-domain life prediction of extra-large-scale bearings based on transfer learning and dual-model fusion

The application of deep learning in constructing data-driven remaining useful life prediction models through historical bearing degradation data has demonstrated significant potential. However, accurate prognostics for extra-large-scale bearings remain constrained by limited operational lifespan data. While extensive degradation datasets exist for standard-sized bearings, inherent mechanistic disparities between different bearing scales create cross-domain transfer challenges. To address this limitation, this study proposes an innovative dual-model fusion framework that synergizes small-bearing full life-cycle data with mechanical principles for extra-large-scale bearing remaining useful life prediction. Our methodology comprises three core innovations: Development of an attention mechanism-enhanced bidirectional gated recurrent unit network integrated with transfer learning; Construction of a physics-informed degradation model based on ISO281 standards; and a novel threshold continuous triggering algorithm for precise degradation phase segmentation. The framework implements a progressive model updating strategy through coordinated utilization of cross-scale bearing data at different degradation stages, establishing an adaptive “data + mechanism” dual-model fusion prognostic system. Experimental validation confirms significant enhancement in prediction accuracy through iterative updating, ultimately achieving reliable RUL estimation for extra-large-scale bearings.