MLE and RBF for AOA estimation in a multipath environment.

摘要

The problem of estimation of angle-of-arrival (AOA) in multipath environments is addressed in this thesis. In particular, two new estimation techniques are developed. The first technique is based on the maximum likelihood estimation (MLE). This algorithm is unique in that a highly deterministic multipath signal model is used when formulating the likelihood function, which is then maximised with respect to the AOA. This model makes use of the geometrical information and a priori knowledge of a number of physical parameters. By using the deterministic multipath signal model with the MLE estimator, one is essentially making more information available to the estimation process. The net result is that the estimator's performance can be greatly enhanced. The Cramer-Rao bounds that apply specifically to this model have been derived to provide a performance measure for the mean-squared errors (MSE) in the estimated AOAs.;Although the MLE method is optimum in a statistical sense, the computational load of the nonlinear optimisation procedure inherently required by the MLE method is too heavy for real-time processing. Accordingly, we propose a novel approach to the AOA estimation problem, which is based on the use of an associative memory. The functionality of an associative memory is identical to that of the inverse mapping network. This provides a more comprehensive explanation for the rationale of exploiting the inverse mapping concept in the AOA estimation problem. In particular, the AOA problem is considered as a mapping from the space of AOA to the space of the sensor output. A nonlinear associative memory is used to form the inverse mapping from the space of sensor output to the spae of AOA and this memory is realised using the generalised radial basis function (RBF) neural network. The RBF network is much more efficient in terms of computation than the MLE algorithm.;Simulations are carried out to understand the efficiency of the RBF neural network approach. The learning and estimation performance is inversely proportional to the number of learning samples and the number of hidden units. At relatively low SNR, the estimation performance of the RBF network becomes insensitive to both the number of learning samples and the number of hidden units. The estimation performance of both the MLE technique and the RBF network is also evaluated as functions of the number of snapshots and SNR. The performance of the MLE algorithm is consistent with the Cramer-Rao bound. The MLE method is more efficient in terms of estimation than a RBF network, provided that the search resolution used in the MLE method is sufficiently high. For equivalent computational complexity, the RBF network gives much better performance than the MLE method. (Abstract shortened by UMI.).