Automatic clustering-based approach for train wheels condition monitoring

The main goal of this paper is to present an unsupervised methodology to identify railway wheel flats. This automatic damage identification algorithm is based on the acceleration evaluated on the rails for the passage of traffic loads and deals with the application of a two-step procedure. The first step aims to build a confidence boundary using baseline responses evaluated from the rail, while the second step involves the damages' classification based on different severity levels. The proposed procedure is based on a machine learning methodology and involves the following steps: (i) data acquisition from sensors, (ii) feature extraction from acquired responses using an AR model, (iii) feature normalization using principal component analysis, (iv) data fusion, and (v) unsupervised feature classification by implementing outlier and cluster analyses. To evaluate whether the number of sensors used to detect and classify wheel flat can be optimized, the influence of sensors' number is performed.