Remaining Useful Life Prediction for Bearings Across Domains via a Subdomain Adaptation Network Driven by Spectral Clustering

Accurate <i>remaining useful life</i> (RUL) prediction of bearings is essential, as bearing failures compromise operational safety. However, distribution discrepancies caused by varying working conditions often degrade prediction performance. <i>Domain adaptation</i> (DA) has been widely used to mitigate this issue, but most DA methods align feature distributions on a global scale, overlooking fine-grained discrepancies within the same domain. <i>Subdomain adaptation</i> (SDA) offers a promising alternative by aligning feature distributions at a subdomain level. Despite its potential, existing SDA methods often use fixed subdomain boundaries, overlook the unequal importance of subdomains, and lack clustering mechanisms for similar features. These limitations hinder further improvements in RUL prediction accuracy. To address these issues, this paper proposes a novel model, <i>subdomain adaptation network driven by spectral clustering</i> (SC-SAN), which dynamically adjusts subdomain boundaries, assigns higher weights to key features, and clusters similar features during model training. The effectiveness of SC-SAN is validated through ablation, comparison and generalization experiments on the XJTU-SY and PRONOSTIA datasets. Experimental results show that SC-SAN achieves an average MAE of 0.1009 and RMSE of 0.1231 across two datasets, representing reductions of 19.86% and 23.41%, respectively, compared to existing state-of-the-art methods.