Neural network technology for aeronautical system optimization
Thursday September 18, 2003, 14:00-14:50
Neural networks provide a general framework to solve problems of classification, pattern recognition and non linear regression. The parsimonious approximation property of neural architectures is attractive to implement neural solution of various data processing problems. This explains that neural networks are often used in space-oriented research, mostly for earth observation and more generally in space imagery and multi-sensor fusion. The ability of neural networks to perform non-linear control is another source of space applications for various non-linear control problems where neural networks are generally used as correctors of a basic linear solution or approximators of rule-based control systems (neuro-fuzzy control). Neural networks are less common in areonautical applications for several
reasons:
Aeronautic technology is based on precise physical models with
customerized computation code.
The validation of aeronautical systems is based on very restrictive
rules that are not generally compatible with statistical validation.
However, recently, new ways of using neural-based computing systems arose to support man decision or to help aircraft design. Aeronautical system optimization becomes too complex to allow straightforward runs of physical-model based computation. So the approximation properties of neural networks are used to build reduced models of aeronautical systems in order to allow further optimization of the system. Then, neural network technology is a component among other statistical tools as surface response model, design of experiments, evolutionary optimization. So a new methodology becomes an active field of research in aeronautical engineering. Several application examples will illustrate this approach. They are selected from international engineering litterature or from current research in partnership with aeronautic industry.