Research On Subband Adaptive Filter And Its Implementation

Adaptive filter has been used in many applications such as communication, sonar, radar, audio processing and image processing, which is one of the most popular research subjects in the field of adaptive signal processing. The performance of adaptive filter depends on its adaptive algorithm and structure. Because of the special structure of subband adaptive filters(SAF), they can partition and then decimate the input signal via a filter bank to reduce the correlation of the fullband input signal, to increase the convergence rate and to decrease the speed of signal processing. Therefore, subband adaptive filters are used widely in the field of signal processing. In order to meet the requirement of project, this dissertation is devoted to a study of subband filter banks and adaptive filter algorithm of SAF. The main works of this dissertation are outlined as follows:1. The structure of uniform subband filter banks and the design of prototype filter are researched. Considering the prototype filter of cosine-modulated filter banks generally leads to severe amplitude distortion of the recovered signals by using iterative design methods, a new cost function is proposed to make filter banks have perfect reconstruction condition and transition band have square root cosine roll-off. Meanwhile, the cut-off frequency of passband is adjusted to ensure that the 3dB passband cut-off frequency of the designed prototype filter approximates to that of the ideal filter. Then the cost function minimum point and the best prototype filter are obtained by iterative approach. The result of the simulation shows that the proposed prototype filter performs better than those designed by existing iterative approach.2. In the applications of system identification, two improved variable parameter SAFs based on multiband structureare proposed. One is a variable step-size matrix design, which is proposed to solve the conflict between fast convergence rate and low misalignment for general affine projection normalized SAF(AP-NSAF). The variable step-size matrix consists of the optimal step size of each subband, which is obtained by minimizing the upper bound of mean-square deviation(MSD) in iteration. In addition, the variable step-size matrix AP-SSAF(VSSM-AP-NSAF) is similar to the AP-NSAF in terms of computational complexity. Experimental results demonstrate that the proposed VSSM-AP-NSAF achieves better convergence performance in system identification, when compared with the variable parameter NSAF. The other is variable regularization parameter affine projection sign SAF(VRP-AP-SSAF), which is presented to get a trade-off between fast convergence rate and low steady-state misalignment by minimizing the MSD of the system weight vector. Besides, the update of regularization parameter is realized by the normalized stochastic gradient descent of MSD. Simulation results show that the proposed VRP-AP-SSAF not only maintains the robustness against impulsive noise and double-talk, but also has a performance enhancement in convergence rate compared with variable parameter SSAF.3. In the applications of channel equalization, two variable parameter subband adaptive equalization(SAE) based on multiband structure are proposed. First, multiband structure is used to solve the alias error in traditional SAE. Then, two new variable parameter SAEs are given to solve the conflict between fast convergence rate and low misalignment of SAE based on multiband structure. One is a kind of variable step-size matrix sign SAE(SSAE). In this method, the energy of subband posteriori error vector is treated as the cost function. By solving optimal step length for each subband to minimize the cost function after each iteration calculation, the best variable step size matrix is obtained. The other is a new variable regularization parameter(VRP) for the normalized SAE(NSAE), which is obtained by using the stochastic gradient descent to minimize the energy of subband error vector. Both of the two proposed algorithms have good convergence performance in terms of faster convergence speed and lower steady state misadjustment.