ST-DGAT: A Fault-Characteristic Frequency-Guided Dynamic Graph Attention Transformer for Bearing Remaining Useful Life Prediction

Abstract

Accurate remaining useful life (RUL) prediction of rolling bearings remains challenging because fault-related spectral components evolve across frequency and time. This article proposes a spatio-temporal dynamic graph attention transformer (ST-DGAT) that embeds bearing kinematics into frequency-graph learning. A logarithmic continuous wavelet transform (CWT) representation combines fault-characteristic anchors with complementary background frequencies. Frequency positional embedding (FPE) preserves node identity, while adaptive dual-channel fusion integrates complementary horizontal and vertical information. A learnable degradation prototype (LDP) produces sample-dependent continuous gates, and mechanism-consistent edge residuals bias graph interactions toward kinematic relations while retaining communication among all node pairs. Global bottleneck injection (GBI) supplies graph-level context, and a Transformer models temporal evolution. The method is validated on two real run-to-failure datasets. Comprehensive ablation and local-bandwidth sensitivity studies support the design, while comparisons with recent approaches demonstrate competitive overall prediction accuracy.