Neural network based optimum target detection in non-Gaussian noise environments

The architecture, principle of operation, and training algorithms for two classifiers, namely the Gram-Charlier Neural Network (GCNN) and Generalized Probabilistic Neural Network (GPNN), for detection of signals in Gaussian symmetric and non-Gaussian non-symmetric noises (e.g. Weibull and Lognormal) are presented. The GCNN classifier is based on applying the Gram-Charlier series approximation of probability density functions. The GPNN is based on Gram-Charlier series and Parzen's windowing technique for approximation of density functions. The classifiers are implemented in neural network type parallel architectures. Training of these networks using a modified Kohonen training algorithm is presented.; Performance of these classifiers have been evaluated in terms of probability of detection and compared to those of two other neural network based classifiers namely Backpropagation and Bayesian classifiers for both scalar and vector samples from sinusoidal signals in Gaussian and Weibull noises and radar signals in Gaussian, Weibull and Lognormal noises.; It is observed that the GCNN and GPNN attain high probability of detections of 0.9 or higher in all cases having reasonable signal-to-noise ratios, whereas the Backpropagation and Bayesian classifiers fails to attain probability of detection of 0.9 for many cases and for any value of signal-to-noise ratio. Also, the GCNN and GPNN attained probability of detections of 0.9 at considerably less signal-to-noise ratio than the other classifiers. The performance of the Backpropagation classifier is formed to vary case to case.; This dissertation also presents two parametric detectors that are optimum for detection in Weibull noise. Performance of these detectors are evaluated analytically and by simulation, and compared to that of a matched filter detector. It is observed that the parametric detector performed better than the Gaussian matched filter in Weibull noise. Whereas the Matched filter outperformed the new detector in Gaussian noise.; These classifiers will be useful in many practical situations (e.g. radar, acoustic underwater monotoring, and communication systems) in which the noises are non-symmetric and non-Gaussian.