A neural networks approach to multiple passive sonar targets identification by source separation technique

Abstract

A source separation and neural network unsupervised learning procedure has been proposed and applied to the identification of multiple passive sonar targets. Acoustic noises radiated from 101 fishing boats were collected from two widely separated underwater hydrophones. A noise background whitening algorithm was applied to flatten the power density spectra (PDS) [W. A. Struzinski and E. D. Lowe, J. Acoust. Soc. Am. 76, 1738–1742 (1984)]. The system was trained by using the single-target spectrum shapes derived from one of the hydrophones and then used these to identify the sources from the other hydrophone for both single and multiple targets. Multitarget signals were preprocessed by a source separation technique to obtain the individual signals. Results of practical testing indicated that the system could correctly identify 90.1% of the recordings for a single sonar target. Identification rate of the multi-target signals can achieve 84% for 50 different combinations of single-target signals. This paper describes the system configuration, the experiment design, and experiences with the practical applications. [Work supported by CSIST.]