Smart railways: AI-based track-side monitoring for wheel flat identification

Abstract

The wheel flat detection in trains using Artificial Intelligence (AI) has emerged as a critical advancement in railway maintenance and safety practices. AI systems can effectively identify geometric deformation in wheel rotation patterns, indicative of potential wheel flat damage, resorting to wayside monitoring systems and machine learning algorithms. This study aims to propose an unsupervised learning algorithm to identify and localize railway wheel flats, which considers three stages: (i) wheel flat detection to distinguish a healthy wheel from a damaged one using outlier analysis, achieving 100 percent accuracy; (ii) localizing the damage to pinpoint the location of the defective wheel through the Hidden Markov Model (HMM); (iii) classification of wheel damage based on its severity using k-means clustering technique. The unsupervised learning algorithm is validated with artificial data attained from a virtual wayside monitoring system related to freight train passages with healthy wheels and defective wheels with single and multiple defects. The proposed methodology demonstrated efficiency and robustness for wheel flat detection, localization, and damage severity classification regardless of the number of defective wheels and their position.