This thesis addresses the crucial issue of detecting track conditions, including
vertical, lateral, and combined track conditions, which profoundly affect railway
safety and passenger comfort, by utilizing vehicle dynamic responses. Traditional
track maintenance methodologies, primarily manual inspections and using
specialized monitoring vehicles, are constrained by their inspection frequency.
Employing models that correlate vehicle dynamic responses with track conditions
could significantly enhance fault detection, particularly when these models are
integrated into commercial vehicles for near-real-time monitoring.
In this study, we investigate the application of deep learning for classifying track
conditions using vehicle dynamic responses. Our approach employs a variant of
convolutional neural networks (ConvNets), specifically GoogleNet, to detect track
conditions. Additionally, we use the Gramian Angular Summation Field (GASF) to
transform carbody acceleration from a temporal representation to a twodimensional
image.
To address the issue of dataset imbalance that inherently occurs in real-world data,
we employ several techniques, including cross-validation and image augmentation.
The image augmentation approach can be applied to time-series data, such as
vehicle dynamic responses, by leveraging the advantages of the GASF
transformation.