Adaptive time series representation for out-of-round railway wheels fault diagnosis in wayside monitoring

Through the integration of advanced sensor technologies and machine learning algorithms, artificial intelligence has revolutionized wayside monitoring in the railway sector. Although several algorithms have been proposed for detecting out-of-roundness, i.e., flats, wear treads and polygonization, they generally fall short in isolating the root cause of the wheel's issue in a train passage. In this context, the paper presents a novel approach for wheel out-of-roundness diagnosis with (1) detection of aberrant train behavior; (2) isolation of specific defective wheels; (3) identification of the severity. For this, the methodology automatically segments a strain gauge signal, capturing the complex nature and temporal dependence of vibration patterns. This segmentation allows the extraction of localized accelerometer features in both the time and frequency domain, as well as implicit axle count and labelling of each wheel passage. Moreover, a single-value damage indicator based on anomaly detection algorithms was proposed. To validate the effectiveness of the proposed methodology, experiments on a set 3D numerical train-track dynamic interaction simulations are performed for different wheel profiles, track irregularities, train speeds, sensor placement and noise, associated to other environmental and operational variations. This demonstrates the potential of artificial intelligence for real-time assessment of wheels without interfering with normal service conditions, suggesting the possibility of automated fault diagnosis. © 2023 Elsevier Inc. All rights reserved.