Application of system identification in analysis of automobile crash.

Abstract

Occupant protection during an automobile crash is an important design consideration. Extensive tests and analysis are typically performed during the early phases of design to improve desired crashworthiness properties in the automobile. This thesis presents methods to develop physically meaningful analytical models directly from crash test measurements, and then demonstrates use of the data based analytical models in predicting crash performance, as well as in analyzing the crash test data. A model structure is selected based on the understanding of the crash event. The model is made up of time varying lumped parameters--mass, stiffness and damping--representing the rigid body response of the system and a transfer function model representing vibration response of the system during crash. The parameters are estimated by minimizing quadratic criterion of one step ahead prediction error using a Gauss-Newton algorithm. The time varying nature of the parameters is addressed using a recursive parameter estimation approach. The crash of an automobile is a highly complex event. The structural characteristics are time varying, the measured data is corrupted with noise, the event of interest is transient and estimation of parameters by minimizing multi-criteria in multi dimension space is not easy due to the possibility of a number of local minimas. Because of these problems estimation of structural parameters from the crash data is a challenging task. In this thesis these challenges are addressed by using physical insights such as initial values, known parameters, variance of noise and understanding of the known characteristics of the components (e.g., knowledge of the differences in loading/unloading and tension/compression). In addition the estimation algorithms are also modified to address the problems in selection of the forgetting factor because of noisy output and changing parameters. A data based analytical model for side impact is developed to demonstrate the use of this approach. The model parameters are estimated directly from the test data using Kalman filter and physical insights of the event. Further the selection of design changes based on the model predictions is also demonstrated. The use of the data based analytical models in design indicates that these models are cost and time effective tools. It is also shown that they are very useful in understanding contribution of various components in the crash event and in predicting the crash performance for various design alternatives. These capabilities may be useful in shortening the design cycle and reducing the number of physical tests.