LEARNING NEURAL NETWORKS FOR DETECTION AND CLASSIFICATION OF SYNCHRONOUS RECURRENT TRANSIENT SIGNALS

This paper proposes a neural network solution to the classical signal processing problem of detection of a synchronous recurrent transient s ignal in noise. If a signal exists, it is assumed to be one of M known signals which may sometimes occur (probabilistically) in successive i ntervals. Several neural network configurations are applied to this pr oblem and compared with each other and with the optimum adaptive seque ntial detector. A novel efficient neural network detector is proposed using an XOR-Tree configuration with learning. Tests with synthetic an d real noise, show the excellent performance of this approach as compa red to the optimum adaptive detector and to other neural network techn iques. With real (non-white) noise obtained from sonar data, the XOR-T ree network widely outperforms the likelihood ratio detector. We also discuss the learning time complexity of the XOR-Tree network and compa re it to that of standard three layer network architectures. (C) 1998 Elsevier Science B.V. All rights reserved.