Flexible Hyperdisk-Based Support Tensor Machine for Railway Turnout Fault Diagnosis Using Dual-Channel Sound Signals

Abstract

Timely fault diagnosis of railway turnouts is crucial for enhancing maintenance efficiency. Most existing methods rely primarily on distance measures or machine-learning models driven by electrical signals. However, these methods often depend on standardized curves or are limited by poor feature quality and small datasets, which restrict their generalizability and performance. Moreover, electrical signals alone provide limited measurement and physical information, which reduces diagnostic accuracy. To overcome these limitations, a contactless fault diagnosis method for railway turnouts is developed using a support tensor machine combined with dual-channel sound signals. In addition, a novel tensor-based classifier, called the flexible hyperdisk-based support tensor machine (FHD-STM), is proposed to mitigate boundary underestimation and computational burden in existing support tensor learning. The proposed method comprises three primary stages. First, sound monitoring data are collected and processed. Next, fused dual-channel feature tensors are constructed using the continuous wavelet transform and data stacking to capture the spatial characteristics of the sound signal and extract effective features. Finally, an FHD-STM model is built to perform fault diagnosis using these feature tensors. The experimental results indicate superior performance in terms of precision, recall, and F1-score compared to existing approaches, which underscores its utility for reliable railway turnout fault diagnosis.