Blind Source Separation for Detection and Classification of Rail Surface Defects Mohamed BENTOUMI 1,2 , Gérard BLOCH 1 , Patrice AKNIN 2 , Gilles MILLERIOUX 1 1 : Centre de Recherche en Automatique de Nancy (CRAN, UMR CNRS 7039) ESSTIN, Rue Jean Lamour, 54519 Vandoeuvre-les-Nancy Cedex, France bentoumi@esstin.uhp-nancy.fr, gerard.bloch@esstin.uhp-nancy.fr 2 : Institut National de Recherche sur les Transports et leur Sécurité (INRETS) 2 avenue du Général Malleret-Joinville, 94114 Arcueil Cedex, France aknin@inrets.fr Abstract. The non-destructive evaluation of the rail is crucial to provide a high safety level in railway transportation. To detect and classify on line rail surface defects (splitted rail, shelling…), a specific double-coils double-frequencies eddy current sensor is used, which gives 8 real differential signals. Independent Component Analysis (ICA) for Blind Source Separation (BSS) is applied to dissociate the different classes of defects from the sensory signals. In an initial context comprising more sources than mixtures, typical defect signatures are analyzed to determine relevant signals for separation and estimate the corresponding separation matrices. A hierarchical separation procedure is then proposed and applied to real signals. 1. Introduction The splitted rail detection is a crucial task for the railway managers to provide a high level of security. The current device providing this function, the track circuit, will disappear in an automatic control context. It is therefore necessary to develop a new system to control the rail in-operating conditions. With that purpose, a specific eddy current (EC) sensor has been developed [1]. This sensor has been designed and optimized according to the following specifications: positioning at 40 mm height, vertical and horizontal displacements of the sensor due to the bogie dynamics, 100 km/h maximum speed of the train, strong acceleration levels (up to 10g), electromagnetic disturbances caused mainly by the traction currents that circulate in the rails. An EC sensor is sensitive to all modification of the geometry and/or electromagnetic characteristics of the target. Obviously, transverse splits of the rail (Figure 1a) are detected, but minor defects as well (shelling (Figure 1b), welded joints or corrugation for example). Detecting and monitoring such defects allow to set up a predictive maintenance policy. Figure 1: a) splitted rail (left), b) shelling (right).