Compound artificial neural network based classifiers with applications to high resolution signal recognition

A novel class of compound artificial neural network based classifiers are presented. The proposed architectures contain different types of neurons and higher-order components, and may satisfy the Stone-Weierstrass theorem directly. They automatically map the input into a linearly separable feature space, then give the class code of the input through a set of Adalines. Intensive computer simulations with underwater signals have been conducted. It is shown that the proposed approach converges more rapidly, and achieves much higher recognition accuracy and robustness with smaller nets as compared with the conventional neural models, even for such signals distorted by noise and other factors as can hardly be discriminated using conventional techniques. The architecture simplification and high performance encourage practical engineering applications in wide fields