A neural network-based passive sonar detection and classification design with a low false alarm rate

Abstract

The Standard Transient Data Set (STDS) Phase 1 data were used to design detection and classification algorithms. Two separate processing chains were constructed, using neural networks for the short-duration transients and conventional processing for tonals. The design activity emphasized the judicious matching of acoustic digital signal processing (DSP) and neural networks, plus the construction of optimized training sets. The resulting design achieved 92% correct classification of the events in the testing files (204 correct out of 221 total events), with only four false alarms in approximately 35 min of data.<>