HYBRID NEURAL NETWORK ALGORITHMS FOR THE RECOGNITION AND CLASSIFICATION OF WEAK TEMPORAL SIGNALS

Neural networks as well as classical transformation techniques have been successful in the field of static pattern recognition. However, the problem of reliably detecting and recognizing weak, temporally varying signals remains a challenge. New techniques are needed to detect and recognize conventional signal types within a dense, dynamically, changing, interference environment. Given their successful application to the field of static pattern recognition, neural networks offer a promising approach to solving this difficult problem. The approach for detecting and recognizing weak, temporal Signals of Interest (SOIs) and Signals Not of Interest (NSOIs) consists of three distinct parts: algorithm development, simulation environment development, and algorithm validation. These three parts are designed to achieve the stated objectives by providing careful analyses of existing neural network architectures as well as the development of hybrid architectures and by thoroughly validating these algorithms in a controlled, reproducible, simulated environment. This research is expected to produce a set of algorithms which can be applied to separating SOIs and SNOIs, as well as a set of heuristics for selecting which algorithms to apply given characteristics of the type of signal being analyzed. This project deals with the three major parts of research which address algorithm development, signal simulation environment, and algorithm validation.