Data-based modeling of vehicle collisions by nonlinear autoregressive model and feedforward neural network

Abstract

Vehicle crash test is the most direct and common method to assess vehicle crashworthiness. Visual inspection and obtained measurements, such as car acceleration, are used, e.g. to examine impact severity of an occupant or to assess overall car safety. However, those experiments are complex, time-consuming, and expensive. We propose a method to reproduce car kinematics during a collision using nonlinear autoregressive (NAR) model which parameters are estimated by the use of feedforward neural network. NAR model presented in this study is derived from the more general one – nonlinear autoregressive with moving average (NARMA). Suitability of autoregressive systems for data-based modeling was confirmed by application of neural networks with a NAR model to experimental data – measurements of vehicle acceleration during a crash test. This model allows us to predict the kinematic responses (acceleration, velocity, and displacement) of a given car during a collision. The major advantage of this approach is that those plots can be obtained without additional teaching of a network.