• Volume/Page
  • Keyword
  • DOI
  • Citation
  • Advanced
   
 
 
 

Journal of the Acoustical Society of America

SECTION LISTING

BROWSE VOLUMES

Year Range: 
Search Issue | RSS Feeds RSS
Previous Issue Next Issue

Nov 1990

Volume 88, Issue S1, pp. S1-S200

Return to All Sections
back to top
RSS Feeds
back to top Session 9UW: Underwater Acoustics: Neural Networks: Their Applications to Underwater Acoustics
Invited Papers
FREE

Application of neural networks to signal processing and control problems (A)

B. Widrow

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S199-S199 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
Neural networks can be used to make nonlinear transversal filters for signal processing. In some cases, signal estimation in noise can be enhanced by nonlinear filtering. Examples will be presented. Applications of neural networks in nonlinear control systems will also be presented. Self‐learning neural control systems for continuous bang‐bang control (the “broom‐balancer”) and for transient analog control (The “truck backer‐upper”) will be explained and demonstrated.
FREE

Detection and classification of nonspeech complex sounds (A)

J. C. Solinsky

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S199-S199 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
Complex sounds generated by nonspeech mechanisms contain acoustic energy represented in both temporal and spectral domains. Conventional energy detection and spectral classification approaches have inferior performance relative to non‐Gaussian and nonlinear approaches when the sounds are from unknown sources. A two‐stage approach will be presented that incorporates polyspectral and neural‐network technologies in a sonar application that performs in a manner similar to human expert operators.
FREE

What does a matched‐field processor try to do? (A)

W. A. Kuperman

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S199-S199 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
Matched‐field processing (MFP) presents a fertile area for applying, developing, and understanding new neural network architectures. Orders of magnitude in computational efficiency are still required to fully take advantage of MFP. Essentially, MFP is a signal processing technique that uses received acoustic data across an array to locate the position of radiating sources in, at best, an imperfectly known ocean environment. This ocean domain is huge. The concepts of MFP are reviewed emphasizing those aspects of the problem that would make it an attractive candidate for applying neural networks. The importance and subsequent treatment of the environment, which is characterized by oceanography, bottom geophysics, and ambient noise, are probably the most central issues from which the benefits of neural networks might emerge. Also relevant is the connection between MFP and remote sensing that further magnifies the impact of new computational structures.
FREE

Artificial neural networks for field reconstruction and source localization (A)

S. Chin, M. Beran, J. Howard, and B. Steinberg

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S199-S200 (1990); (2 pages)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
Neural networks have been used to successfully solve the inverse problem of determining the source location based on field or coherence measurements. The conditions and types of neural networks that are effective in solving this problem are discussed. Neural network topology will be considered, as well as the source‐array geometry and its effect on performance. Simulation results will be provided, which demonstrate under what conditions interpolation is possible. Finally, a discussion of other source localization algorithms will be given. These methods will be contrasted to the neural network methodology.
Contributed Papers
FREE

Application of artificial neural networks to the classification of underwater ambient noise signatures (A)

John N. Kriebel

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S200-S200 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
More than 2500 ambient noise samples were collected in Behm Canal, Alaska, between July 1989 and January 1990. The data were obtained using a system deployed from a MINIMET buoy, and consist of one‐third octave analyses covering the 50 Hz through 63 kHz bands. Signatures obtained on an hourly schedule, were transmitted via the ARGOS satellite and entered into a LOTUS 1‐2‐3 database. Attempts to categorize the data by conventional means, e.g., manual sorting of plots of the signatures, or on a purely statistical basis, were unsatisfactory: thus the application of artificial neural networks was investigated. This approach yielded reasonable classifications and allowed atypical signatures to be identified and deleted from the database prior to performing statistical analysis. The groupings assigned by the network parallel those which might have been made by an analyst in that obviously different signatures are assigned to different classes, but in some cases the network makes more subtle distinctions than an analyst might. One network separated the signatures into 22 major groups (those containing at least 26 signatures) that include 90% of the data.
FREE

Classification of underwater acoustic transients by artificial neural networks (A)

Ronald L. Greene and Robert L. Field

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S200-S200 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
Artificial neural networks have been trained using the backpropagation algorithm to classify a variety of model transient source signals. The networks were then tested on signals propagated to 25 different receiver sites by the time‐domain parabolic equation model. Despite the interference effects from surface and bottom reflections, the classification accuracy is about 90% in the noise‐free case, virtually identical to that of a nearest‐neighbor classifier on the same problem. Classification in the presence of noise is considerably reduced; however, the redundancy provided by the multiple receivers in most cases allows the network to correctly classify all signals from sources on which it was trained. In addition, it shows a robustness in the presence of unknown signals not shown by the nearest‐neighbor classifier. [Work supported by the Naval Oceanographic and Atmospheric Research Laboratory through the U. S. Navy/ASEE Summer Faculty Research Program.]
FREE

Matched‐field processing using a neural network with preprocessing (A)

John M. Ozard, Pierre Zakarauskas, and Peter W. Ko

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S200-S200 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
Model‐based signal processing for source localization usually requires a comparison of the measured and replica acoustic fields at all possible source positions. It is anticipated that this problem may be solved more efficiently as a pattern recognition problem. In order to make such a solution applicable to arbitrary array shapes, the covariance matrix of the simulated data was preprocessed and represented as the excitations of the eigenvectors. The orthogonal bases for the total possible signal space was employed. Localization of a source in depth was performed using a linear perceptron with the excitations as input. The perceptron was trained with the excitations for a distribution of source depths. Precision of localization of the resulting processor and its robustness in the presence of noise were measured and compared to the performance of a minimum variance matched‐field processor. Extensions of the processor to estimate range and azimuth are being investigated by using a multilayer neural network with backpropagation training.
FREE

Neural nets for seafloor classification: A computer simulation study (A)

D. Alexandrou

J. Acoust. Soc. Am. Volume 88, Issue S1, pp. S200-S200 (1990); (1 page)

Online Publication Date: 14 Aug 2005

Full Text: | Download PDF

Show Abstract
The point scattering model of reverberation offers estimates of a large number of reverberation statistics. However, fitting the model to data is not immediately meaningful owing to the model's weak physical connection. The approach taken here utilizes parametrizations of the reverberation probability density function (pdf) to create acoustic “feature vectors” representative of different seafloor types. The problem of seafloor classification is then tantamount to a problem of pattern recognition. Computer simulations are used to create a number of different scatterer distributions representing different seafloor regimes. Several flexible pdf families, including the most general pdf supported by the point scattering model and the generalized Gaussian pdf, are used to parametrize synthetic reverberation data. A neural net classifier is trained to recognize the prevailing seafloor type based on its acoustic signature and is tested for its ability to discriminate among different scatterer distributions. A number of neural net algorithms are evaluated in this setting, including the linear perceptron and a backpropagation of error algorithm. [Work supported by ONR.]
Return to All Sections
Close

Home Publications Meetings Contact ASA Site Map Back to Top

Copyright © Acoustical Society of America

close