Investigation of Multi-Dimensional Interpolation Methodologies for Vehicle Maneuvering and Design
Abstract
Recurrent and feedforward neural networks were studied and compared for their ability to adequately capture dynamics from a physics based computational model of submarine maneuvering dynamics. The comparison was done using one trajectory of a rise maneuver. Very good approximation was achieved and recurrent neural networks. The results with modular feedforward networks were not as good, but were reasonable. When the networks were tested for an unseen dive maneuver, the response followed the shape of the desired response but there were large overshoots. This preliminary study demonstrated that it is possible to learn dynamics from computational models, that recurrent neural networks are much better suited for this task, and that additional data are required for good generalization. However, an approach for very limited training data set design was developed for further investigation. The applications of this approach are numerous, including online use of detailed models, rapid iterative design update, improved control of a submarine and eventually its noise characteristics, condition based maintenance in the presence of aging and degrading parts, and potentially, damage assessment and mitigation in hostile situations.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 24, 2000
- Accession Number
- ADA372704
Entities
People
- Amulya K. Garga
- Farhad Davoudzadah
- Howard J. Gibeling
Organizations
- Pennsylvania State University