Data Augmentation-Based Prediction of System Level Performance under Model and Parameter Uncertainties: Role of Designable Generative Adversarial Networks (DGAN)

Abstract

Owing to uncertainty factors present in the system, computer-aided engineering (CAE) models suffer from limitations in terms of accuracy of test model representation. This paper proposes a new predictive model, termed designable generative adversarial network (DGAN), which applies the Inverse generator neural network to GAN, one of the methods employed for data augmentation. Statistical model-based validation and calibration technology, employed for improving the accuracy of a predictive model, is used to compare the prediction accuracy of the DGAN. Statistical model-based technology can construct a predictive model through calibration between actual test data and CAE data by considering uncertainty factors. However, the achievable improvement in prediction accuracy is limited, depending on the degree of approximation of the CAE model. DGAN can construct a predictive model through machine learning using only actual test data, improve the prediction accuracy of an actual test model, and present design variables that affect the response data, which is the output of the predictive model. The performance of the proposed prediction model was evaluated and verified, as a case study, through a numerical example and system level vehicle crash test model including parameter uncertainties.