Non-Gaussian noise reduction in system identification

The synergy of signal processing filtering methods and established parameter estimation techniques reduces unwanted effects from noise. Measurement noise, particularly of non-Gaussian type, is known to be problematic in identification and parameter estimation; and a relatively small amount may wreak havoc on linear estimation schemes. Opportunities for improved system identification exist using noise-reducing filtering techniques. The challenge is to merge filtering methods with parameter estimation algorithms appropriately. The application of nonlinear filters as thresholding mechanisms for linear prediction-error identification strategies reduces the impact of noise on parameter estimates. Dynamic thresholding allows a modified algorithm to detect potentially contaminated data before it influences the estimation process. To test this approach, additive nonGaussian measurement noise contributes to outputs from linear systems with well-defined structures. System identification is performed using the resulting noisy input-output signals, and the outcomes are compared against identification made using noise-free signals. Examples demonstrate almost a 25% reduction in residual error. Regardless of filtering method employed, the performance from incorporating nonlinear filters demonstrates improvements in system identification