Search Results

An advanced flight control system, which has been demonstrated to compensate for unanticipated failures in military aircraft, is proposed for use in general aviation. The method uses inverse control to decouple the flight controls and to modify the handling qualities of the aircraft, while employing artificial neural networks in order to compensate for any modeling error. These errors can stem from any differences between the model and the actual aircraft. Therefore, they can include in-flight hardware failures, rendering the system fault tolerant and reducing the necessity for multiple levels of redundancy. The proposed system is verified in simulations for longitudinal flight and is shown to be able to track pilot-commanded velocity and flight path angle. Also, one example is presented for in-flight changes of the configurations (flap deployment) where the controller is shown to adapt rapidly to these changes without a need for compensation by the pilot.

The architecture and training of artificial neural networks are briefly described. Five applications of these networks to design and analysis problems are presented; three in aerodynamics and two in flight dynamics. The aerodynamics cases are those of a harmonically oscillating airfoil, a pitching delta wing, and airfoil design. The flight dynamic examples involve control of a super maneuver and a decoupled control case. It is demonstrated that highly nonlinear aerodynamic cases can be generalized with sufficient accuracy for design purposes. It is shown that although neural networks generalize well on the aerodynamic problems, they appear lacking comparable robustness in modeling dynamic systems. It is also shown that generalization appears to become weak outside of the training domain.

In nonlinear cases, the SDG method requires multidimensional search procedures. However, in linear cases only one-dimensional search procedures are required to identify the critical gust load conditions. In this study the application of the backpropagation ANN method as a multi-dimensional modeling tool has been proposed to model or identify the global and local extrema of one-dimensional gust load responses. The maximum and minimum response values of ramp-step input gust profiles were considered to investigate the ANN modeling capability and effectiveness. The actual SDG analysis for nonlinear cases was hypothesized to be performed over a large and sparse domain, therefore the ANN could be trained to quickly identify the region of the domain containing the global extrema. The SDG analysis, then, could be concentrated on a smaller region thereby reducing computation time.