Elastic Target Modeling for Physics-Based Automatic Classification
Automated detection and identification of abandoned munitions through sonar interrogation requires that system algorithms be robust to a wide variety of environmental and target conditions. Empirical tests are a crucial part of the verification of robustness, but even more basic is a physically-justified design of the algorithm and feature selection. In a recent SEED effort SERDP MR-2324, the project team developed a finite element (FE) approach to aid in such an incorporation of the target physics. In this project, the researchers will expand on this approach by developing a wider class of physics isolation techniques and using the identified features to guide classifier design and predict performance in a variety of operation conditions and target states. The researchers will validate the modeling by manufacturing targets that will be interrogated in separately funded sea trials.
Automated acoustic interrogation of seabeds for identification of abandoned munitions will enable the application of mature sensor technology to the vast data throughput necessary in practical application. However, unmanned systems in particular require high confidence in system robustness, and this has not been achieved to date given the limited physics justification even for algorithms which currently perform well in limited tests. FE modeling of munitions has allowed high-fidelity prediction of scattered acoustic returns from sources too complex to analyze using empirical models. The researchers have been focused on developing specific FE techniques to predict responses and to isolate the effects of individual physical mechanisms contributing to the response. Thus, the researchers can illuminate the effects of specific interior structures on the received signal. Once complex target responses are broken down into individual components based on the underlying physics, it is much more straightforward to understand the environmental effects on such components than it would have been on the complete return. The researchers will investigate the effects of burial state (both amount and material) on these individual components and use the results to predict the complete response and allow a physically justifiable assessment of the robustness of features which are being used or proposed for automated munitions identification.
The researchers will utilize a variety of techniques to isolate particular components of a predicted return as opposed to simulating the complex total return. The coupling-region approaches introduced in MR-2324 will be combined with additional constraints derived from research in basic physics, including wave-direction limitations. The researchers will improve existing generic feature representations, and as necessary adapt existing classification architectures, to optimize them for the specific features isolated during modeling. Having identified the sources of these features, the researchers will predict their robustness to target burial state and source/receiver location and validate the results by manufacturing targets and subjecting them to separately funded sea trials.
The result of this effort will be a physically-justified feature set that will be incorporated into a merged operations/classification package, which could be operated from a tow body. More generally, the research will provide guidance on the limits of a particular system's robustness and provide researchers with a path for improved future algorithm performance. (Anticipated Project Completion - 2019)