Abstract
Using simulated data, this study investigated the development and comparative performance of support vector machine (SVM) and feedforward neural network (FNN) models in identifying faults in trains’ wheelset axle bearings. The simulated dataset consisted of 20 trains, each with 50 wagons and 4 wheelsets per wagon, for a total of 20040 entries. The parameters included the sound pressure level (SPL), frequency, and harmonic distortion. The classification accuracy across the "Normal," "Alert," and "Critical" conditions was evaluated. The FNN demonstrated superior performance across all the metrics, achieving near-perfect accuracy, precision, recall, and F1-scores, with ROC curves indicating perfect sensitivity and specificity (AUC = 1.00) for all the classes. In contrast, the SVM, while effective in detecting critical conditions, exhibited lower accuracy in identifying normal and alert conditions, as shown by AUC values of 0.76 and 0.86, respectively, highlighting a trade-off between computational resources and accuracy. This variability in the SVM performance suggested potential limitations in handling less distinctive classes, which could impact the safety and reliability of axle bearing monitoring systems. Consequently, the FNN is identified as the superior model for axle bearing fault detection, offering robust and reliable classification across all conditions, making it the preferred choice for deployment in safety-critical environments such as train maintenance systems.