Archard model guided feature engineering improved support vector regression for rail wear analysis

Jinlong Wang,Yi Su,N. Alagu Subramaniam,John Hock Lye Pang

DOI: https://doi.org/10.1016/j.engfailanal.2022.106248

IF: 4

2022-07-01

Engineering Failure Analysis

Abstract:A new method combining the split Hopkinson pressure bar (SHPB) technique with theback-propagation (BP) neural network program is proposed. By this method, the treated strainwave signals become smooth with less noises induced by the transverse inertia. Moreover, thedynamic rate-dependent constitutive behaviour of materials can be identified, without any pre-assumption of a constitutive model. It is found that by taking the experimentally measured data ofstrain, strain rate and time as 'input' and the corresponding data of stress as 'output' of the BP neuralnetwork, the dynamic constitutive behaviour with internal damage or phase transformation evolute-ton is easy to be identified, where the time could represent either the internal damage evolution orphase transformation process accompanied with the deformation process. It is emphasized thatthe data learnt by the BP neural network must include both loading and unloading processes, if thewhole loading and unloading response are to be correctly predicted. The comparisons between thepredictions and experimental results are in good agreement for both polyamide (PA) polymer (as anexample of nonlinear viscoelastic materials) and Ti–Ni alloy (as an example of superelastic materialswith stress-induced phase transformation)A new method combining the split Hopkinson pressure bar (SHPB) technique with theback-propagation (BP) neural network program is proposed. By this method, the treated strainwave signals become smooth with less noises induced by the transverse inertia. Moreover, thedynamic rate-dependent constitutive behaviour of materials can be identified, without any pre-assumption of a constitutive model. It is found that by taking the experimentally measured data ofstrain, strain rate and time as 'input' and the corresponding data of stress as 'output' of the BP neuralnetwork, the dynamic constitutive rail with internal damage or phase transformation evolu-tion is easy to be identified, where the time could represent either the internal damage evolution orphase transformation process accompanied with the deformation process. It is emphasized thatthe data learnt by the BP neural network must include both loading and unloading processes, if thewhole loading and unloading response is to be correctly predicted. The comparisons between thepredictions and experimental results are in good agreement for both polyamide (PA) polymer (as anexample of nonlinear viscoelastic materials) and Ti–Ni alloy (as an example of superelastic materialswith stress-induced phase transformation)This paper applied Archard wear law in feature engineering for the improvement of Support Vector Regression (SVR) performance and realized the rail steel wear behavior prediction. The actual complex rail wear multidimensional degradation information was obtained from field maintenance records over a decade and the hidden data outliers raised the modelling challenges. We applied pre-process technologies including feature importance analysis, physical model guided feature generation and outlier detection to build up the SVR based robust nonlinear regression analysis framework. Individual railway parameters' effects on the wear process were investigated and revealed through model interpretation-based post analysis. This work provides a practical approach to deploying machine learning algorithms for rail service maintenance data analysis and treatment of data outliers.