Underwater Signal Modeling for Subsurface Classification Using Computational Intelligence.
Abstract
In the thesis a method for underwater layered media (UWLM) modeling is proposed, and a simple nonlinear structure for implementation of this model based on the behaviour of its characteristics and the propagation of the acoustic signal in the media accounting for attenuation effects is designed. The model that responds to the acoustic input is employed to test the artificial intelligence classifiers ability. Neural network models, the basic principles of the back-propagation algorithm, and the Hopfield model of associative memories are reviewed, and they are employed to use min-max amplitude ranges of a reflected signal of UWLM based on attenuation effects, to define the classes of the synthetic data, detect its peak features and estimate parameters of the media. It has been found that there is a correlation between the number of layers in the media and the optimum number of nodes in the hidden layer of the neural networks. The integration of the result of the neural networks that classify and detect underwater layered media acoustic signals based on attenuation effects to prove the correspondence between the peak points and decay values has introduced a powerful tool for UWLM identification. The methods appear to have applications in replacing original system, for parameter estimation and output prediction in system identification by the proposed networks. The results of computerized simulation of the UWLM modeling in conjunction with the proposed neural networks training process are given. Fuzzy sets is an idea that allows representing and manipulating inexact concepts, fuzzy min-max pattern classification method, and the learning and recalling algorithms for fuzzy neural networks implementation is explained in this thesis. A fuzzy neural network that uses peak amplitude ranges to define classes is proposed and evaluated for UWLM pattern recognition. It is demonstrated to be able to classify the layered media data sets, and can distinguish between the peak points for identification purposes. Fuzzifying the extracted data set through a nonlinear extended sigmoid type function causes an associative memory property in the proposed networks. It has been found that the response of the fuzzy relation rule based network to grade of membership of fuzzified data set is much more convincing than the case of working with crisp or nonfuzzified data. Naturally, because of associative memory properties of the networks, substituting mono a connection instead of dual connections may reduce the networks size, computation time and number of features. The model has been tested by a real data set from southern California sediments that were provided by the Naval Research Laboratory. It has responded well to identification of the sediments of the data set. Neural networks have shown great promise in several areas and its synergism with the fuzzy concept appears to have a wide range of applications in data classification for oil and natural gas explorations. The results suggest that the application of fuzzy neural networks in pattern recognition will enable automatic classification of the marine data collections for layered media identification. (Abstract shortened by UMI.).
- Publication:
-
Ph.D. Thesis
- Pub Date:
- 1995
- Bibcode:
- 1995PhDT.......176S
- Keywords:
-
- Engineering: Electronics and Electrical; Artificial Intelligence; Physics: Acoustics