Robust Adaptive Estimation Methods For Complex Environments

The traditional adaptive estimation methods are derived based on the assumptions that the noises obey the Gaussian distribution and the systems are accurately modeled,and they show good estimation performance and robustness in the assumed environments.However,these assumptions are not true in many practical environments,and the performance of the traditional adaptive estimation methods will deteriorate in those environments.In order to solve the estimation problems in the complex environments such as sparse systems,super-Gaussian noises,sub-Gaussian noises and input noises,and further improve the estimation performance of the robust adaptive estimation methods,this dissertation propose some new robust adaptive estimation methods by adopting different improvement strategies and analyze its stability and convergence characteristics in detail.The specific research works of this dissertation include:(1)In order to improve the estimation performance of the proportional adaptive estimation methods in the convergence phase,we propose a new proportional affine projection algorithm.By designing a proportional gain proportional to the deviation between the current estimate and the previous estimate,the estimation method can exhibit rapid convergence throughout the convergence phase.In addition,compared with the traditional proportional adaptive estimation methods,the new proportional affine projection algorithm simplifies the estimation process and further reduces the computational complexity of the algorithm.The performance analysis of the adaptive estimation method plays an important role in guiding the design of the adaptive estimator.However,the performance of the affine projection-based proportional adaptive estimation methods has not been analyzed.Here,we perform the steady-state performance analysis through the energy conservation theory,and derive a general expression of the steady-state excess mean square error.In addition,we perform the tracking performance analysis based on the random walk model,and give the optimal step size parameters that can obtain the minimum steady-state excess mean square error.(2)In the echo channel identification applications,the systems are sparse,the system inputs are speech signals(colored signals),and the background noises are super-Gaussian noises.In order to solve the echo channel identification problems in the presence of the impulsive noises,we propose a new proportional M-estimate affine projection algorithm.This method uses M-estimator to improve the anti-impulsive noise interference capability,and uses the efficient recursive method to reduce the computational complexity.In order to further improve the estimation performance of the adaptive estimation method in estimating the echo channel,we propose a new M-estimate affine projection normalized subband adaptive filtering algorithm and analyze its stability.Furthermore,a new proportional M-estimate affine projection normalized subband adaptive filtering algorithm is proposed to accelerate the convergence speed.(3)Acoustic echo cancellation is an effective strategy for controlling the acoustic echo generated by the hands-free audio terminal.During the double-talk process,the near-end speech signal,as the super-Gaussian noise,will affect the effect of adaptive echo cancellation.In this environment,the performance of the traditional set-membership normalized subband adaptive filtering algorithm deteriorates severely or even does not converge.In order to enhance the anti-interference ability of the set-membership normalized subband adaptive filtering algorithm under the super-Gaussian noise environments,we propose a new robust set-membership normalized subband adaptive filtering algorithm by designing a new robust set-membership error bound,and analyze its steady-state performance.In order to further improve the estimation performance of the adaptive estimation method in estimating sparse systems,we propose some sparsity-aware adaptive estimation methods based on the robust set-membership normalized subband adaptive filtering algorithm,and analyze their steady-state performance.(4)The affine projection-like algorithms have better estimation performance and lower computational complexity than the affine projection algorithm.However,the affine projection-like algorithms does not consider the interference from the filter input,and they yield biased estimates when the input signal is noisy.For the two kinds of affine projection-like algorithms,in order to eliminate the estimation bias,we propose new affine projection-like algorithm based on the unbiasedness criterion.The normalized subband adaptive filtering algorithm exhibits good estimation performance and low computational complexity when the input signal is a colored signal.Since the normalized subband adaptive filtering algorithm does not consider the input noise,the normalized subband adaptive filtering algorithm yields biased estimates in the environments with input noises.In order to eliminate the estimation bias,we propose a new normalized subband adaptive filtering algorithm based on the unbiasedness criterion.In addition,we propose a new input noise estimation method that does not require the input-output noise variance ratio in advance.(5)In the sub-Gaussian noise environments,the normalized least mean fourth algorithm exhibits better estimation performance than the traditional normalized least mean square algorithm.However,the normalized least mean fourth algorithm does not consider the interference of the input noise.The estimated result will have a certain deviation from the true value under the input noise interference.In order to deal with the composite environment with the sub-Gaussian noise and the input noise,we propose a new normalized least mean fourth algorithm based on the unbiasedness criterion.In the super-Gaussian noise environment,the robust set-member normalized least mean square algorithm has the performance of anti-impulsive noise interference and has low computational complexity.However,under the interference of input noise,the estimation result of the robust set-membership normalized least mean square algorithm will have a certain deviation from the true value.In order to deal with the complex environment with the super-Gaussian noise and the input noise,we propose a new robust set-membership normalized least mean square algorithm by minimizing a new cost function.