Intelligent Signal Processing with Applications to Radar and Communications
Date
2018-03-13
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The research of intelligent signal processing presented in the thesis is the result of a few main projects: multiple neural network models for chaotic signal prediction and the application to radar signal processing; Genetic algorithm (GA) for optimal multiple neural network training and its application for chaotic underlying dynamics modelling; Autoregressive (AR) system identification using chaos driving signal and non-linear prediction, and the corresponding estimation Mean Square Error (MSE) performance analysis based on Cramer Rao Bound (CRB); The above proposed multiple model (MM) predictor, GA algorithm, and system identification with chaos driving signal are further applied for real-life radar signal processing and channel equalization in chaos communication. Sea clutter refers to the radar electromagnetic wave backscatter from a sea surface. Sea clutter modelling is an important capability with many applications for remote sensing and radar signal processing. Due to a recent discovery that sea clutter is chaotic rather than purely random, computational intelligence techniques such as neural networks have been applied to develop new model for sea clutter. In this thesis, we propose using the multiple neural network model approach to construct a predictive model for sea clutter modelling. The motivation comes from the observation that the sea usually has some unpredictable motions which result in impulsive events such as sea spikes. Although a single nonlinear model could describe the Bragg scattering reasonably as shown in the literature, it is usually incapable of capturing sea spikes motions. Therefore target detection performance might be degraded when such a clutter model is employed. By using a multiple radial basis function (RBF) neural network predictors, we found that a sea clutter signal with different underlying dynamics from sea spikes to normal motions can be modeled accurately, and therefore provides a promising model for clutter suppression in radar detection.
We also discover one main drawback of the above multiple model approach is the training process, where a gradient search is usually employed to obtain the model parameters. Since the parameter set of the experts and gate is usually large, the search cannot always locate the optimal solution. The situation gets worse when all the models are highly nonlinear. The main reason for using this suboptimal training method is its computational efficiency, but the trade-off is a degraded performance. In order to achieve an optimal MM approach for dynamic reconstruction, we propose applying a genetic algorithm to determine the multiple model parameters. This process takes a hybrid approach which uses a GA to search for the optimum values of the following parameters in a MM: RBF net centres, variances, and uses the least square method to determine the RBF weights. Computer simulations are used to illustrate the effectiveness of the proposed method in reconstructing piecewise chaos dynamic. In addition, sea clutter data collected at the Atlantic Ocean and the power pool price data from Alberta, Canada are employed to evaluate the usefulness of the GA-MM approach. Our research further applies nonlinear multiple model prediction to the blind identification of autoregressive (AR) system. Contrary to the conventional statistical techniques with white Gaussian noise, diving signals generated from low-dimensional nonlinear deterministic systems rather than from stochastic processes are applied in this study. In earlier studies, the minimum nonlinear prediction error (MNPE) method, is shown to achieve superior system identification performance using chaotic driven signal over the lease square identification with white Gaussian driven signal. Based on the short-term predictability of chaos, the received signal can be passed through an inverse filter and then the system parameters are estimated by minimizing the nonlinear prediction error of the inverse filter output using the least square technique. However the effectiveness of existing MNPE method is limited to chaos driven signal with its dynamic function known as a priori, which might not always be available in many applications such as radar signal processing and chaos communications. In this thesis, we extend the MNPE approach by using a RBF neural network to identify the systems when the exact dynamic of the chaotic driven signal is not known as a priori. In addition, we propose an improved least square (ILS) method based on the concept of minimizing orthogonal Euclidean distance between the noisy input space and desired prediction outputs to reduce the system estimation bias caused by additive measurement noise. Although the interest of system identification using chaos driving signal is surging, very few theoretical works have been reported on explaining the advantage of chaos identification. Our research investigates the theoretical performance analysis for the AR system identification using chaos driving signal. Based on the Cramer Rao Bound (CRB) analysis for the mean square error (MSE) performance, it is proved that when chaos is used to drive an AR system, identification using only the output signal can be as good as that based on both input and output signals. That is, blind identification performance is equivalent to non-blind identification with chaos driving signal. To further verify this conclusion, a deterministic Maximum Likelihood (ML) is also developed to blindly identify an AR system driven by chaos. Combined with the global search technical Genetic Algorithm (GA), the proposed GA-ML method is found to achieve the optimal identification performance imposed by the asymptotic CRB. The theoretical MSE performance of the proposed GA-ML method is derived and the result is validated using computer simulations. This method is applied to blind equalization of a spread spectrum (SS) communication system where chaos is used as the modulation signal. Compared to conventional AR system identification methods based on white Gaussian driving signal, the chaos approach is shown to have superior performance. The improvement is shown to be the result of the positive Lyapunov exponent of the chaotic signal.
Description
Keywords
chaos, Artificial Intelligence, neural network, genetic algorithm, nonlinear signal processing, system identification, radar signal processing, communications
Citation
Xie, N. (2018). Intelligent Signal Processing with Applications to Radar and Communications (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/31738