Neural network based signal prediction and parameter estimation for underwater layered media systems identification

A method for underwater layered media (UWLM) modeling accounting for attenuation effects is proposed. A simple nonlinear structure and an algorithm for implementing and simulating the model based on the behavior of the media is designed. The assumption of suitable parameters in the modeling permits...

Full description

Saved in:
Bibliographic Details
Main Authors: Setayeshi, El-Hawary
Format: Conference Proceeding
Language:English
Subjects:
Backpropagation
Feedforward neural networks
Multilayer perceptrons
Neural network applications
Parameter estimation
Prediction methods
Sonar signal analysis
Underwater acoustic propagation
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:A method for underwater layered media (UWLM) modeling accounting for attenuation effects is proposed. A simple nonlinear structure and an algorithm for implementing and simulating the model based on the behavior of the media is designed. The assumption of suitable parameters in the modeling permits us to design a more accurate model. A model that responds to the acoustical input is employed to test the artificial neural network's ability in signal prediction and parameter estimation. Experimental results from the implemented programs are presented. A feed-forward multi-layered neural network trained by the back-propagation algorithm using the min-max amplitude ranges of an output signal of the UWLM is designed and evaluated. This is based on a computer simulation of synthetic data, to estimate the parameters of the media including attenuation factors, reflection coefficients, travel times and decay parameters. Based on experiments estimating the parameters of the media and predicting its output signal, the network produces the same computational results as those of the original assumed media structure. The results of computer simulation of the underwater layered media modeling in conjunction with the proposed neural network training process are given. The results suggest that the proposed network can supplement, or replace conventional techniques for parameter estimation and output prediction in system identification. The method presented also offers advantages in speed and efficiency over existing estimates techniques.< >
DOI:10.1109/CCECE.1994.405867