An Interactive Neural Network System for Acoustic Signal Classification
Abstract
The objectives of this project was to develop an understanding of the effect of neural networks, implemented in interactive systems, on sonar operators and other naval personnel. Specifically, the project called for the development of a prototype system, employing neural networks to test the effect of interactive (man-in-the-loop) operations. ARD developed such a system which is used to classify time domain signals generated from the insonification of an underwater mine-like target. The system converts the time domain signals to frequency domain and frequency over time (spectrograms) and displays the signals at the users' request in all three formats. A time windowing function is also provided to allow the user to closely inspect specific portions of the time domain signal. In addition, a neural network system classifies the signal according to three parameters: shell thickness, interior content and angle of insonification. Results have shown that most users exhibit a large bias towards the use of the neural network analysis because of their highly accurate classification. Future work will concentrate on the integration of neural network tools into existing systems in real-world situations. A better understanding of the human-network interactions will be gained when the ability of the networks to classify real world signals is decreased due to the complex geometries of actual mines and environmental effects on the sonar returns (thermoclines, shallow water, surface returns).
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 28, 1990
- Accession Number
- ADA219278
Entities
People
- Georgianna Meagher
- Michael Philips
- Nelson Steel