A subspace method for detection and classification of rail defects

A non destructive evaluation system dedicated to rail inspection using a non-contact eddy current sensor embedded in a subway train is presented. An original processing approach borrowed from the MUSIC algorithm is proposed for rail surface defects detection and classification. This approach, based on the eigen decomposition of the signal covariance matrix, produces signal and noise subspaces. The projections of the typical defect signatures on the noise subspace and a multiplicative fusion of the elementary detectors are then performed. Compared to the approaches previously used in the same context, this approach yields to better results, particularly for shelling isolation. The proposed method has been tested successfully on the labeled defect data set of a subway line.