Feedforward neural network system for the detection and characterization of sonar signals with characteristic spectrogram textures

• US 5,502,688 A
• Filed: 11/23/1994
• Issued: 03/26/1996
• Est. Priority Date: 11/23/1994
• Status: Expired due to Term

• Current Assignee: General Dynamics Advanced Technology Systems, Inc. (General Dynamics Corporation)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Russo, Anthony P., Recchione, Michael C.
• Primary Examiner(s): Pihulic, Daniel T.
• Application Number:
  US08/344,450
• Time in Patent Office: 489 Days
• Field of Search: 367/131, 367/135, 367/901, 367/7, 367/11, 367/136, 395/21, 395/22, 395/25, 364/923.4, 364/423, 342/196
• US Class Current: 367/131
• CPC Class Codes: H04B 11/00 Transmission systems employing sonic, ultrasonic or infrasonic waves
• CPC Class Codes: Y10S 367/901 Noise or unwanted signal reduction in nonseismic receiving system

Abstract

The present invention provides a method and system for characterizing the sounds of ocean captured by passive sonar listening devices. The present invention accomplishes this by first generating a spectrogram from the received sonar signal. The spectrogram is characterized in terms of textural features and signal processing parameters. The textural features and signal processing parameters are fed into a neural network ensemble that has been trained to favor specific features and/or parameters. The trained neural network ensemble classifies the signal as either Type-I or clutter.

Claims

1. A method for characterizing a sonar signal captured by sonar listening devices, the method comprising the steps of:

(1) characterizing a spectrogram generated from a sonar signal in terms of textural features and signal processing parameters;

(2) inputting to a plurality of trained neural networks said textural features and said signal processing parameters of said spectrogram, wherein each of said plurality of trained neural networks is trained to favor a different subset of said textural features and said signal processing parameters; and

(3) classifying components of said sonar signal as characteristic texture or clutter using said trained neural network.

2. The method of claim 1, wherein said spectrogram is generated using sliding-window Fast Fourier Transform analysis.

3. The method of claim 1, further comprising the steps of:

(1) filtering said spectrogram to remove Jezmonster and seismic profiler energy to produce a filtered spectrogram;

(2) generating a normalized spectrogram from said filtered spectrogram using a noise spectrum equalization procedure optimized for wideband signals;

(3) generating an integrated spectrogram by reducing the frequency and time resolutions of said normalized spectrogram by integration;

(4) partitioning said integrated spectrogram into overlapping, two-dimensional regions; and

(5) selecting regions whose median powers are above a detection threshold to produce selected regions;

(6) clustering said selected regions into signal objects using a blob-coloring algorithm; and

(7) deriving said textural features and said signal processing parameters from measurements of said signal objects, said filtered spectrogram, and said normalized spectrogram.

4. The method of claim 1, wherein said textural features include co-occurrence analysis results, striation measurements, and aspect determination results of said sonar signal.

5. The method of claim 1, wherein said signal processing parameters include frequency measurements, bandwidth, signal-to-noise ratio, and amplitude of the signal component of said sonar signal.

6. The method of claim 1, wherein said textural features and said signal processing parameters used by said plurality of trained neural networks are selected by an Optimal Brain Damage method.

7. The method of claim 1, wherein each of said plurality of trained neural networks is trained to favor a subset of said textural features and said signal processing parameters by use of higher learning rates for the neurons connected to said subset of said textural features and said signal processing parameters.

8. The method of claim 1, wherein said trained neural network is a sparsely connected, feedforward network, trained with the back-propagation algorithm.

9. An apparatus for characterizing the sounds of ocean captured by sonar listening devices, comprising:

means for generating a spectrogram from a sonar signal;

means for characterizing said spectrogram in terms of textural features and said signal processing parameters; and

a plurality of neural networks trained to classify components of said sonar signal as characteristic texture or clutter based on said textural features and said signal processing parameters, wherein each of said plurality of neural networks has been trained to favor a subset of said textural features and said signal processing parameters.

10. An apparatus according to claim 9, further comprising a preprocessor, said preprocessor comprising:

filter means for removing Jezmonster and seismic profiler energy to generate a filtered spectrogram; and

normalization means for normalizing said filtered spectrogram using a noise spectrum equalization algorithm optimized for wideband signals to generate a normalized spectrogram.

11. The apparatus according to claim 10, further comprising detection means for selecting regions whose median powers are above a detection threshold for inputting to said means for characterizing.