A feasible strain-history extraction method using machine learning for the durability evaluation of automotive parts

The feasibility of a machine-learning algorithm is discussed in terms of strain-history prediction and durability evaluation of automotive components. Additionally, an additional benefit using the machine-learning is reported; that is the acquisition of strain histories from sensors other than strain-gages. More specifically, the strain-history prediction performance was compared depending on how to prepare and learn data from each driving event, and the accuracy was analyzed with respect to the dominant direction of wheel-forces. Moreover, detailed information on the strain-history extraction was provided. Consequently, the performance of the network that learned the data generated from whole PG was better than one that learned a single driving data. For better application of machine learning to strain estimation for automotive components, timesteps of 80 were found to be the optimal parameter. And it was suggested that the data ranges for machine-learning training be larger than that in real-application to minimize discrepancies around peaks.