New unbiased adaptive IIR filter: it’s robust and variable step-size versions and application to active noise control

Abstract

This paper proposes a robust variable step-size adaptive IIR filter realized by a new bias-free structure (BFS). Unlike equation error (EQE) method that uses a desired signal contaminated with observation noise, the BFS employs a filter driven by the output of the plant estimate and this achieves a bias-free estimate of the denominator of the system function. In addition, the adaptation is made robust to the observation noise by the Griffiths’ LMS adaptation, which uses the cross-correlation estimate between the input and the desired signal for its adaptation gradient computation. A robust variable step-size adaptation is also realized by the Griffiths’ gradient. The proposed structure is referred to as BFSGV and has good modeling capability with improved convergence rate and reduced misadjustment. For system identification, the proposed BFSGV algorithm gives a 3 dB improvement in the performance index over EQE method. The proposed BFSGV has been applied to active noise control (ANC). The BSFGV structure is used for secondary path (SP) estimation, and for the main path (MP), BFS structure with step-size varied according to Okello’s method (BSFV) is used. The new ANC system for narrowband noise field is found to be having 4 times faster convergence rate and an additional noise reduction of 15dB over that FIR for MP and the EQE for SP. Further, the use of the proposed ANC IIR algorithm achieves computational savings compared to that of FIR. For the broadband noise field, the proposed method that uses BSFV for MP and BSFGV for SP provides 18 times faster convergence rate and 2.5 dB reduction in ANC error compare to that of the ANC using FIR for MP and the EQE for SP estimation.