Combination of Data-driven Feature Selection Methods with Domain Knowledge for Diagnosis of Railway Vehicles

Abstract

Railway vehicles are generally maintained preventively within certain time periods. Condition based predictive maintenance strategies have a great economic potential so that modern trains are equipped with many sensors in order to perform diagnostics and prognostics of components.Methods for fault detection need appropriate feature subsets in order to achieve small in-sample and out-sample errors. In our case the typical feature selection approach using pure data-driven methods is difficult, as the number of possible feature sets is very large. On the other hand there exists rich domain knowledge and detailed physical models of the mechanical system. The aim is to combine this knowledge with the often used mathematical methods for feature selection for improving classification of cases when a faulty damper is present. Based on the dynamic equations of motion, this paper presents heuristic feature selection via the analysis of transfer functions. We describe several wellknown methods of automated feature selection and a workflow which combines domain knowledge with automated methods. Results show that it is difficult to definefeatures based only on domain-knowledge, but in combination with data-driven techniques good classification performance can be achieved.