A Neural Expert Approach to Self Designing Flight Control Systems.
Final rept. 15 Jul 93-14 Jan 94,
CHARLES RIVER ANALYTICS INC CAMBRIDGE MA
Pagination or Media Count:
Based on the simulations performed in this phase I study, we show that Hopfield and RBF feedfoward network architectures may have a great potential in the control of nonlinear systems. In particular, Hopfield implementation of Lagrange multiplier method is suitable for real-time adaptive optimal control. Similarly, RBF feedforward neural network architectures are suitable for learning inverse dynamics and inverse trim in aircraft FCS applications. In addition, RBF feedfoward are easier to train than backpropagation sigmoid networks since RBF formulation results in linear parameters. The initial simulations we performed show very promising results as exemplified by the small control errors in closed-loop Simulations using the nonlinear A-18 longitudinal dynamics. Further studies are needed to test the applicability of the techniques to real world problems and to study the robustness, stability and general reliability of the proposed neural techniques. Neural networks by themselves cannot be the panacea to all the nonlinear control problems. An effort has to be made to incorporate all the available knowledge about the dynamics system to achieve good performance.
- Flight Control and Instrumentation