Refined Fusion Graph Spectrum Model: A Graph-Informed Multivariate Signal Processing Method for High-Speed Train Bogie Bearing Fault Diagnosis

Abstract

Fault feature of high-speed train bogie bearing is readily masked by strong noise, while multisensor measurements provide complementary information for robust diagnosis. This article proposes a graph-informed multivariate signal processing method, termed the refined fusion graph spectrum model (RFGSM) for weak fault feature extraction. Refined fusion spectrum decomposition integrates super Fourier analysis with adaptive fusion spectrum segmentation to achieve interpretable multivariate signal decomposition and isolate fault frequency band. The fusion graph spectrum model then embeds graph theory into local fusion spectrum and reconstructs high-order graph spectrum components through the graph Fourier transform, enhancing fault spectrum structures and suppressing in-band noise. The enhanced fusion fault feature is synchronously mapped back to all channels, enabling collaborative feature enhancement while preserving cross channel consistency. Fault components are identified using frequency domain signal-to-noise ratio, and their energy distribution is visualized through a fusion diagnostic diagram. Results from simulated and experimental signals show that RFGSM outperforms advanced multivariate signal processing methods in multichannel fault feature extraction and noise robustness.