Fast Signal Recognition and Detection using ART1 Neural Networks and Nonlinear Preprocessing Units based on Time Delay Embeddings

A new method for fast adaptive signal recognition and detection using neural networks is proposed. The method is essentially based on converting samplesfromthe signals to be detected or classified into a binary "character-like" matrix which can be then used to train fast adaptive neural networks. While this preprocessing method may be applied to any neural architecture designed for character classification tasks, we have used to test the performances on modified ART1 networks. These networks were chosen due to their fast learning capabilities making them very attractive for on-line signal classification tasks. The preprocessing method was much inspired from the embeddology theory which gives appropriate tools for nonlinear systems identification, based only on observing a time-sequence generated by the underlying nonlinear system. Experimental results proved that efficient and fast decisions can be done for signals coming from sources which can be modeled as nonlinear dynamic systems.