EVOLUTIONARY METHODS FOR TRAINING NEURAL NETWORKS FOR UNDERWATER PATTERN CLASSIFICATION

The ability to accurately detect and classify underwater objects of interest in an automated system is of great importance. It provides human operators with enhanced capability and can greatly reduce information overload. As is often the case, human classification of signals of interest is often impractical due to situational or operational constraints. Two benefits of automated classifier systems are immediately apparent. The first is the tireless nature of computers to perform tedious tasks such as data analysis. The second and equally important benefit of machine pattern recognition is the ability to perform classification in areas where the presence of humans may not be desirable or possible. Highly accurate classification and fast parallel processing speeds can be obtained by using neural network classifiers for pattern recognition problems. Problems associated with conventional training methods can be either alleviated or bypassed by using Evolutionary Programming as a training mechanism. Combining these efficient classifiers with effective signal pre-processing techniques provides the basis for robust automated detection of undersea objects of interest. This paper considers combined spectral and parametric techniques on CTFM active sonar data as a preprocessor for neural network classifiers to detect and discriminate between mine like objects (spheres) and naturally occurring objects.