Keywords: wheel impact load detector, wheel-rail contact forces, digital signal processing, deep learning, fault diagnosis, rolling stock maintenance.

Wheel Impact Load Detection (WILD) and Weigh-In-Motion (WIM) systems have been deployed at trackside locations on the Indian Railways network for over two decades, providing reliable detection of severe wheel defects through threshold-based Impact Load Factor (ILF) analysis. While this represents a significant operational capability, the diagnostic reach of conventional WILD systems is fundamentally constrained: they are effectively blind to the broader class of rolling stock health conditions — suspension degradation, axle box misalignment, uneven loading, tread polygonisation — that precede and predict catastrophic failure. This paper presents a CNN-based rail classification platform, designated WILDPulse, deployed as a software augmentation layer above existing WILD hardware, extending the diagnostic pipeline through three interconnected capabilities: (i) physics-informed multi-domain signal analysis in time, frequency, and time-frequency domains, (ii) a ResNet-18-based CNN classification layer mapping STFT spectrogram images of axle events to rolling stock health states, and (iii) a Train Health Index (THI) providing longitudinal, rake-level health records for condition-based maintenance. Results are presented for a real field WILD dataset with fault-modified signals synthesised via physics-informed simulation for five fault types and their combination. Explainability is provided through Grad-CAM attribution maps.

download the full-text of this paper (PDF, 15 pages, 1147 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description