A neural network approach for data inversion, application to the characterization of marine sediments, and sonar targets
Abstract
The paper presents an innovative method for inversion based on the use of neural networks. The inversion procedure will be split in three major steps: the understanding and enhancing of the physical phenomena involved, the learning of the physics from extracted parameters, and the inversion itself. This separation allows a very convenient distribution of computation load: the first two steps (which are the most demanding in terms of computation time) can be done in a prior phase, while the third one (which becomes very simple) can be implemented on-line (real time). Another major advantage of the proposed method (with respect to conventional approaches such as conjugate gradient) is that it requires only poor a priori information thanks to a judicious choice of the characteristic physical phenomena and the selection of relevant signal processing tools for parameter extraction. Two examples of the proposed procedure are described: the inversion of geoacoustic parameters of seabed and the characterization of sonar targets. The inversion is applied to both simulated and experimental data sets. In both cases, the training will be achieved on simulated data and then applied to the experimental.
- Publication:
-
Acoustical Society of America Journal
- Pub Date:
- November 2002
- DOI:
- 10.1121/1.4779296
- Bibcode:
- 2002ASAJ..112.2307Z
- Keywords:
-
- Neural Nets;
- Underwater Acoustics;
- Acoustics