An Unsupervised Learning Approach for Wayside Train Wheel Flat Detection

Abstract

One of the most common types of wheel damage is flats which can cause high maintenance costs and enhance the probability of failure and damage to the track components. This study aims to compare the performance of four feature extraction methods, namely, auto-regressive (AR), auto-regressive exogenous (ARX), principal component analysis (PCA), and continuous wavelet transform (CWT) capable of automatically distinguishing a defective wheel from a healthy one. The rail acceleration for the passage of freight vehicles is used as a reference measurement to perform this study which comprises four steps: (i) feature extraction from acquired responses using the specific feature extraction methods; (ii) feature normalization based on a latent variable method; (iii) data fusion to enhance the sensitivity to recognize defective wheels; and (iv) damage detection by performing an outlier analysis. The results of this research show that AR and ARX extraction methods are more efficient techniques than CWT and PCA for wheel flat damage detection. Furthermore, in almost every feature, a single sensor on the rail is sufficient to identify a defective wheel. Additionally, AR and ARX methods demonstrated the potential to distinguish a defective wheel on the left and right sides. Lastly, the ARX method demonstrated robustness to detect the wheel flat with accelerometers placed only in the sleepers.

Keywords

• wheel flat detection
• wayside condition monitoring
• train-track interaction
• unsupervised learning