A randomized integral error criterion for parametric identification of dynamic models of mechanical systems

This paper proposes a new approach to the identification of reduced order models for complex mechanical vibration systems. Parametric identification is commonly conducted by the regression of time-series data, but when this includes significant unmodelled modes, the error process has a high variance and autocorrelation. In such cases, optimization using least-squares methods can lead to excessive parameter bias. The proposed method takes advantage of the inherent boundedness of mechanical vibrations to design a new regression set with dramatically reduced error variance. The principle is first demonstrated using a simple two-mass simulation model, and from this a practicable approach is derived. Extensive investigation of the new randomized integral error criterion method is then conducted using the example of identification of a quarter-car suspension system. Simulation results are contrasted with those from comparable direct least-squares identifications. Several forms of the identification equations and error sources are used, and in all cases the new method has clear advantages, both in accuracy and consistency of the resulting identification model.