Machine learning from computer simulations with applications in rail vehicle dynamics. Issue 5 (3rd May 2016)

• Record Type: Journal Article
• Authors: Taheri, Mehdi; Ahmadian, Mehdi
• Is Part Of: Vehicle system dynamics. Volume 54:Issue 5(2016:May)
• Page Start: 653
• Page End: 666
• Publication Date: 2016-05-03
• Subjects: Stochastic modelling -- meta-models -- surrogate models -- three-piece truck -- global optimisation -- Latin hypercube sampling plan; Motor vehicles -- Dynamics -- Periodicals; Electronic journals; 629.231
• DOI: 10.1080/00423114.2016.1150497
• Languages: English
• ISSNs: 0042-3114

Abstract

ABSTRACT: The application of stochastic modelling for learning the behaviour of a multibody dynamics (MBD) models is investigated. Post-processing data from a simulation run are used to train the stochastic model that estimates the relationship between model inputs (suspension relative displacement and velocity) and the output (sum of suspension forces). The stochastic model can be used to reduce the computational burden of the MBD model by replacing a computationally expensive subsystem in the model (suspension subsystem). With minor changes, the stochastic modelling technique is able to learn the behaviour of a physical system and integrate its behaviour within MBD models. The technique is highly advantageous for MBD models where real-time simulations are necessary, or with models that have a large number of repeated substructures, e.g. modelling a train with a large number of railcars. The fact that the training data are acquired prior to the development of the stochastic model discards the conventional sampling plan strategies like Latin Hypercube sampling plans where simulations are performed using the inputs dictated by the sampling plan. Since the sampling plan greatly influences the overall accuracy and efficiency of the stochastic predictions, a sampling plan suitable for the process is developed where the most space-filling subset of the acquired data with number of sample points that best describes the dynamic behaviour of the system under study is selected as the training data. Results indicated that the stochastic modelling technique is effective in improving the computational efficiency of the MBD model without compromising the accuracy of the predictions, although the improvements in the computational efficiency of the technique could not be quantified due to the inefficiencies associated with transferring the data between multiple software packages (SIMPACK, SIMULINK).