Estimation of ARMA Model Order Using Artificial Neural Networks

This article presents a novel algorithm for solving the problem of autoregressive moving average (ARMA) model order estimation. The proposed algorithm is based on modeling and designing the artificial neural network (ANN) architecture for a special matrix constructed from the minimum eigenvalue (MEV) criterion. The MEV criterion is based on a covariance matrix derived from the observed output data only. The input signal is unobservable. The proposed ANN-based algorithm is developed by training the MEV dataset using the backpropagation (BP) learning algorithm to select the appropriate ARMA model order. The algorithm uses one-class-one-network (OCON) topology. Hence, the subneural network was developed for each ordered pair. Then, these subnetworks were assembled in one simulator. The ANN-based algorithm was tested on several simulated examples to estimate the ARMA model orders. MEV efficiency was compared with the proposed ANN approach at various signal-to-noise ratios to demonstrate substantial improvements.