Book Description
In this research we propose two novel adaptive signal processing techniques to enhance the quality of a received signal that is transmitted over a wireless communication link. In the first we develop a family of efficient echo cancellation algorithms that require no knowledge of the echo-path. Echoes significantly degrade the quality of service in wireless communications because of extra delay introduced into the transmission path from the speech compression/decompression process. To remove these echoes we must identify the echo-path impulse response that in general is sparse (predominantly zero). This makes it desirable to adapt only those filter coefficients corresponding to the non-zero regions of the impulse response. To accomplish this we exploit the hierarchical structure and temporal localization property of the wavelet decomposition. In this way we are able to adapt a small subset of filter coefficients and then, based on the coefficients that are significantly different from zero, to accurately identify the remaining coefficients that require adaptation as well. This approach was previously applied to the case of the Haar transform. The current work generalizes it to the wavelet decomposition in which any orthogonal or bi-orthogonal wavelet can be used. Due to the ability of longer wavelets to achieve greater input decorrelation, the resulting algorithm is capable of a significant improvement in convergence speed and computational complexity over LMS. The other novel adaptive technique we propose is in the physical layer to reduce the effect of multiple-access interference (MAI) for direct-sequence code-division multiple-access (DS-CDMA) wireless communication systems. A DS-CDMA system allows many users to share the same spectrum and distinguishes different user's data through the use of signature codes. In practice, these signature codes are not orthogonal and therefore data from different users interfere with each other and creates MAI. Past approaches have either used classical estimation techniques where each unknown user symbol to be estimated is treated as a deterministic parameter, or Bayesian techniques that treat each unknown user symbol as a stochastic parameter but restrict it to a finite set of discrete possible values. In this research, we employ Bayesian estimation such that each user symbol is treated as a stochastic parameter with continuous Gaussian distributions centered at the true symbol values of the data constellation. This allows us to develop a gradient-based maximum a posteriori (MAP) estimator that takes the structure of the particular symbol constellation into account in order to dramatically improve symbol estimation accuracy.