Combination of data-driven feature selection methods with domain knowledge for diagnosis of railway vehicles

Abstract

Railway vehicles are generally maintained preventivelywithin certain time periods. Condition based predictivemaintenance strategies have a great economic potential sothat modern trains are equipped with many sensors in orderto perform diagnostics and prognostics of components.Methods for fault detection need appropriate feature subsetsin order to achieve small in-sample and out-sample errors. Inour case the typical feature selection approach using puredata-driven methods is difficult, as the number of possiblefeature sets is very large. On the other hand there exists richdomain knowledge and detailed physical models of themechanical system. The aim is to combine this knowledgewith the often used mathematical methods for featureselection for improving classification of cases when a faultydamper is present. Based on the dynamic equations ofmotion, this paper presents heuristic feature selection via theanalysis of transfer functions. We describe several wellknownmethods of automated feature selection and aworkflow which combines domain knowledge withautomated methods. Results show that it is difficult to definefeatures based only on domain-knowledge, but incombination with data-driven techniques good classificationperformance can be achieved.