Neural Networks, Adaptive Signal Processing

Artificial neural network attracts more and more attention in recent years. It has been successfully applied to many fields, such as signal processing, automatic control and pattern recognition. In this paper, researches on neural networks for adaptive signal processing, including adaptive filtering, adaptive equalization, identification and the adaptive predistortion of the RF amplifier, are presented.The main originality in this paper can be summarized as follows:1. A complex-valued Hebbian -type learning algorithmIn this dissertation, a complex-valued Hebbian-type learning algorithm is proposed. This algorithm is a generalization of the real-valued anti-Hebbian algorithm to complex-valued case. The performance of the proposed algorithm is demonstrated with application in complex adaptive IIR filtering. Computer simulation results show that it has many good properties.2. A novel method for adaptive equalizationA novel method for adaptive communication channel equalization based on a multi-layer feed-forward neural network training algorithm was proposed in this dissertation. It trains a complex neural network whose inputs, outputs, weights and active functions are all complex-valued. The training of the neural network is 'based on the combination of supervised and unsupervised learning process, while the update of the weights based on the TLS (total least square) criterion. Computer simulation results demonstrate that the proposed equalizer has powerful properties both in linear and nonlinear channels.3. A Neural Networks Training Algorithm with Application in System IdentificationA fast new algorithm for the training of multilayer neural networks is presented in this dissertation. This algorithm is based on the exponentially weighted local least squares (EWLLS) object function and the Euclidean direction set (EDS) method. During the course of training, the local desired outputs have been estimated, the multilayer neural networks can be decomposed into a set of adaptive linear elements (Adaline), and the Adalines are trained by EDS method.The performance of the algorithm is demonstrated with application in nonlinearsystem identification.4N Using RNN for Adaptive Predistortion Linearization of RF AmplifiersA novel adaptive predistorter for linearizing a RF power amplifier in a mobile transmitter is studied. Unlike most other predistorters reported in the literatures, this predistorter was constructed as a complex-valued recurrent neural network (RNN). And the weights of the RNN were adjusted by using complex RTRL(Real Time Recurrent Learning) algorithm. Thus the AM/AM and AM/PM responses of the proposed predistorter are simultaneously implemented. The proposed scheme is shown to attain superior performance in comparison with most other well-known predistortion structures. The performance of the proposed predistorter is demonstrated by computer simulations.Limited to the length of this dissertation, we present here only a part of my research works. For other aspect of my works, please refer to the papers and research works listed in the appendix of this dissertation.