Class-modular multi-layer perceptron networks for supporting passive sonar signal classification

The automatic classification of passive sonar signals in real time is an extremely useful tool for supporting sonar operators decision making during submarine missions. An appealing architecture for the development of accurate, flexible, modular and real-time classification systems is the class-modular multi-layer perceptron (CM-MLP). However, for this task, several design parameters have to be tuned and a more extensive guideline about how to conduct this process is missing in the literature. This work aims at discussing the main design phases related to the development of CM-MLP systems for passive sonar signal classification. For each phase, several approaches are discussed and their cost-effectiveness is experimentally evaluated using statistical significance tests. This analysis is based on real signals produced by 8 vessel classes and acquired by an omnidirectional hydrophone in a shallow water environment, during 263 experimental runs of 34 ships. Results show that some design parameters significantly affect the accuracy of the classification system, especially the complexity of each module. An accuracy of 84.4% is achieved by using tuned design parameters.