A new method is developed for determining the linear autoregressive moving average (ARMA) model parameters for system identification from the time domain impulse response of a circular elastic cylinder of finite length, struck axially (to induce longitudinal vibrations). An artificial neural network is employed to find the coefficients of the digital rational transfer function that approximates the time domain impulse response. Neural networks provide an efficient solution where most algorithms like the Steiglitz-McBride method become computationally time-consuming to obtain good parameter estimates of ARMA models of high order. Deconvolution methods are known to lead to ill-posed problems. Computer simulations using real and simulated data show that accurate linear ARMA model parameters can be obtained with this algorithm. Comparison of the spectral representation of the elastic waves propagating in the cylinder obtained by the new algorithm and that obtained experimentally show that the new algorithm yields, in a reliable way, to understandable transfer function relations between the exciting force and the acceleration measured at the opposite end of the cylinder.
展开▼
