Railway device diagnosis using sparse Independent Component Analysis

This paper presents a study on the potential interest of sparse Independent Component Analysis (ICA) for the diagnosis of a complex railway infrastructure device. This complex system is composed of several spatially related subsystems, i.e. a defective subsystem not only modifies its own inspection data but also those of other subsystems. In this context, the ICA model is used to extract from inspection data indicators of each subsystem state. We assume here that inspection data are observed variables generated by a linear mixture of independent and nongaussian latent variables linked to the defects. Furthermore, physical knowledge on the inspection system provides prior information on the mixing structure. We investigate then the ability of sparse ICA to recover this structure and to provide meaningful defect indicators. We also show that introducing sparsity in the mixing process slightly improves the results.