Identification of high-impact wheels on rail transit rolling stock through acoustic sensing and machine learning

This study investigates the potential for developing and leveraging machine learning algorithms to identify problematic train wheels that generate high impact loads using sound classification analysis. The data used in this research were collected from a heavy rail transit agency in the US. Both the wheel-rail interface load magnitude and sound pressure levels were collected and processed. Each audio sample was transformed into a spectrogram, which provided a visual representation of the sound signal. To classify spectrograms into two distinct categories representative of high impact loads and normal wheel loads, a convolutional neural network (CNN) was developed and trained on the spectrograms as input, which were labeled based on their loading conditions. The performance of the trained model – 72% accuracy – was satisfactory given constraints present and proves the feasibility to predict railway wheel loading condition based on the sound generated during train passage. This potential alternative method to the current complex system of track-mounted strain gauges for monitoring wheel health could yield substantial benefits to rail operators with limited resources or that require unintrusive or portable wheel condition monitoring.