抄録
In this paper, a new partial-normalized least mean square (partial-NLMS) algorithm for updating the coefficients of an adaptive FIR digital filter (ADF) is proposed. The new adaptive algorithm is modified to overcome the NLMS method's problem, about unstability when very small input enters into the filter, and the LMS method's gradient noise amplification problem by using the preset value pσ2x to decide which method is to be used. This value is determined beforehand by the knowledge of the statistical properties (probability density function and variance) of the input signal. In the new partial-NLMS (P-NLMS) algorithm, the measured input signal's square norm is compared with the preset value pσ2x. If the square norm is larger than pσ2x the coefficients of the ADF are renewed by the NLMS algorithm. Renewal of the ADF's coefficients is done with the LMS algorithm when the square norm is smaller than pσ2x.
The simulation results and theoretical analyses show the effectiveness of the proposed P-NLMS, and a good agreement between both results. The saturated residual error of the P-NLMS ADF under the noisy circumstances is smaller than that of the ordinary NLMS ADF and also have good stability. Even in bad condition, corrupted input, the proposed method still shows good performance.
