Aligned sparse non-negative matrix factorization for vehicle-track features decoupling

doi:10.1016/j.ymssp.2026.113907

Aligned sparse non-negative matrix factorization for vehicle-track features decoupling

2026-01-29

0

PRE

AI

DOI：10.1016/j.ymssp.2026.113907

Original

Original request for help

Save

Abstract

En 中文

Vibration signals collected from in-service urban rail vehicles exhibit strong coupling between vehicle dynamics and track geometry excitations, often compounded by environmental noise. This poses a significant challenge for the accurate decoupling of sources and the estimation of track geometric parameters, particularly curve superelevation, from vehicle acceleration data. To address this, we propose an Aligned Sparse Non-negative Matrix Factorization (ASNMF) framework to decouple of vehicle-track features: A Kurtosis-Spectral Peak (KSP) criterion is first applied to construct a Hankel matrix that enhances the representation of non-stationary features; A multi-objective optimization is then formulated by integrating a Gini-based sparsity constraint and a Maximum Mean Discrepancy (MMD) alignment term to ensure consistent component extraction; The resulting multiplicative updating algorithm yields physically interpretable decompositions. Validation using both simulated and real-world vibration data demonstrates that ASNMF effectively separates vehicle and track-induced responses under strong coupling and noise. Compared with existing matrix factorization and blind source separation methods, ASNMF achieves higher signal fidelity and more accurate track-related feature estimation, offering a robust and novel solution for decoupling and interpreting coupled vehicle–track dynamic responses under non-stationary operating conditions.

Aligned sparse non-negative matrix factorization for vehicle-track features decoupling

Abstract

Journal

Organization