| Summary: | 碩士 === 國立臺灣科技大學 === 電機工程系 === 87 === This thesis focuses on the development of subspace adaptive filters such that the computation cost of the traditional adaptive filters can be reduced and overall performance can be improved as well.
We begin with the discussion of a traditional adaptive filters such that a general background can be phased in and some terminologies can be formally defined. After that, a time domain eigensubspace adaptive filter was developed to improve the performance of traditional LMS algorithm. Some stability problems of the LMS method stemming from the input autocorrelation matrix with large eigenvalue spread can be relieved by this method.
To improve the performance of the adaptive filters even further, the transform domain adaptive filters were introduced for this purpose. We focus on the wavelet domain adaptation for their finer resolution in the frequency domain which enables us to reduce the data dimension. After that, a signal subspace method called ULV decomposition is included. This subspace method is numerically stable and can be updated without too much effort, when a new data sample is presented. Combining the wavelet subband characteristics and the ULV algorithm, we develop a new adaptive algorithm called wavelet-ULV RLS algorithm, which approximates the ULV-RLS and plain RLS algorithm well, and is computationally more efficient. The computational cost reduction of this new algorithm has been analyzed and shown in detail.
To show the feasibility of this new algorithm, it has been implemented in an active noise control system. The result is very encouraging.
|
|---|
