Hybriding data-driven and model-based approaches for fault diagnosis of rail vehicle suspensions

This paper introduces a hybrid method to identify and isolatefailures on rail vehicle suspensions, composed of dampersand springs using some spectral information. The systemunder investigated was introduced at a competition named as2017 Data Challenge organized by prognostics and healthmanagement (PHM) society. With limited information, bothdata-driven approach and physics based approach areintroduced. First, a data-driven approach was introducedwhich computes root mean square error (RMSE) betweentraining data set and validation data set at all the sensors.Since the failure on one component (i.e., spring or damper)has impact on adjacent sensors, the failure can be detected byidentifying the maximum RMSE values among all thesensors. Second, an ensemble method, integrating physicalmodel based method and Pearson correlation coefficient(PCC) based method, was developed for the experimentsunder unknown track condition. In physical model basedmethod, the models for each suspension were designed as aform of transfer function, explaining the relation betweencomposing sensors. In PCC based method, correlation valueswhich are independent with track conditions were calculatedto detect and identify the failure. The proposed method led tothe third prize in 2017 Data Challenge.