Axle Sensor Fusion for Online Continual Wheel Fault Detection in Wayside Railway Monitoring

TL;DR

This work presents a semantic-aware, label-efficient continual learning framework for railway fault diagnostics that detects minor imperfections due to flats and polygonization, while adapting to evolving operational conditions, such as changes in train type, speed, load, and track profiles, captured using a single accelerometer and strain gauge in wayside monitoring.

Abstract

Reliable and cost-effective maintenance is essential for railway safety, particularly at the wheel-rail interface, which is prone to wear and failure. Predictive maintenance frameworks increasingly leverage sensor-generated time-series data, yet traditional methods require manual feature engineering, and deep learning models often degrade in online settings with evolving operational patterns. This work presents a semantic-aware, label-efficient continual learning framework for railway fault diagnostics. Accelerometer signals are encoded via a Variational AutoEncoder into latent representations capturing the normal operational structure in a fully unsupervised manner. Importantly, semantic metadata, including axle counts, wheel indexes, and strain-based deformations, is extracted via AI-driven peak detection on fiber Bragg grating sensors (resistant to electromagnetic interference) and fused with the VAE embeddings, enhancing anomaly detection under unknown operational conditions. A lightweight gradient boosting supervised classifier stabilizes anomaly scoring with minimal labels, while a replay-based continual learning strategy enables adaptation to evolving domains without catastrophic forgetting. Experiments show the model detects minor imperfections due to flats and polygonization, while adapting to evolving operational conditions, such as changes in train type, speed, load, and track profiles, captured using a single accelerometer and strain gauge in wayside monitoring.

Figures (17)

• Figure 1: (1) semantic extraction of temporal wheel indices (X), strain-based deformation (Y), and wheel count (Z) via peak detection; (2) VAE-based reconstruction of signal embeddings (S); and (3) fault detection through data fusion of S, X, Y, and Z, and continual loss-based experience replay.
• Figure 2: Sequential training causes rapid forgetting of the first task, while interleaved training allows the model to retain both, indicating that forgetting arises from the training strategy rather than model capacity.
• Figure 3: Overview of related work: Offline/Batch processing (light blue), Online processing (blue), Offline Continual Learning (purple), Online Continual Learning (dark blue)
• Figure 4: Related work: Traditional circuit detector (dark blue), Machine vision (purple), Vibration/optical fiber (light blue)
• Figure 5: Semantics extraction of temporal wheel indices (X), strain-based deformation (Y), and wheel count (Z) via peak detection, which are subsequently fused with VAE-based reconstruction of signal embeddings (S), under an ablation of five different strategies: S-WC = S + Z, S-WI = S + X, I-WI = X, I-WD = Y, and S-WD = S + Y. Further details in Table \ref{['tab:ad_summary']}.