Improved convergence in the sequential regression algorithm for the adaptive identification of IIR systems

Conventional gradient algorithms are known to have slow convergence. The work presented here shows an improvement in the convergence speed, using a hybrid FIR-IIR adaptive filter for IIR system identification. The IIR identification problem is tackled first via an adaptive FIR filter with the recursive least-squares (RLS) algorithm, extracting most of the unknown system's features. Through the application of the balanced model truncation (BMT) technique the FIR approximation is mapped to a lower order IIR structure which initializes an adaptive IIR filter to perform further adaptive iterations. The sequential regression (SER) algorithm is deployed for the adaptive IIR filter. Our method shows speed and accuracy improvements for the unknown system identification, when compared to the use of the SER algorithm alone with an IIR filter initialized with zeros coefficients. Robustness in the presence of high levels of additive white noise is another advantage of our approach.