Identification of AR parameters at a very low SNR using estimated spectral distribution in DCT domain

A simple and efficient method for system identification even at a very low signal-to-noise ratio (SNR) is presented. At an SNR as low as -7.5 dB, noise dominates the spectrum and system poles are almost lost in the profound noise. In the proposed method, an enhanced spectrum is estimated in the discrete cosine transform (DCT) domain using the least squares curve-fitting technique. The system modes that were previously indistinguishable become prominent in the enhanced spectrum. The system order is then overestimated using least squares higher order Yule-Walker (LSHOYW) equations to obtain better accuracy. The poles having higher strength in the autocorrelation domain are then identified as system poles.