ARMA model order estimation based on the SVD of the data matrix

Abstract

In many fields of digital signal processing such as acoustic echo path estimation, modeling of the unknown systems has been required. Since the selection of the model order is very important in such modeling, various methods of selection have already been proposed. However, the order estimation of the ARMA model is very difficult due to the noise of the observed data, even in the case of known input data. Since, in the presence of additive noise in the output data, traditional methods fail to select a reasonable order, we propose a new algorithm to overcome this problem. Our proposed algorithm for ARMA model order selection is based on singular value decomposition (SVD) and new criteria of threshold values for the smallest singular value of the noisy data matrix. In the simulation, we show that our proposed algorithm is very effective for obtaining an appropriate rank of the noisy data matrix. We also show the good performance of our algorithm for practical estimation of the room acoustic transfer function. © 1994, Acoustical Society of Japan. All rights reserved.