Identification of aerodynamic coefficients with a neural network.

摘要

The components of a framework for the procurement, identification, and employment of aerodynamic coefficients are developed. The basic structure follows the estimation-before-modeling (EBM) technique. In the EBM methodology, state estimation and model determination are broken into two independent steps. An extended Kalman-Bucy filter and a modified Bryson-Frazier smoother are used to estimate state and force histories from a measurement vector. This data is used for maintenance of the aerodynamic mapping. The model satisfies the accuracy, smoothness, and differentiability requirements demanded by nonlinear control laws.; A-priori information drawn from the entire input-space is employed to establish a baseline model. Dynamic-system measurements are processed to provide the accurate state and force histories required for on-line updates of the identification model. An extended-Kalman Bucy filter provides state estimates and in combination with a random-walk model accurate force histories. A modified Bryson-Frazier smoother refines these estimates based on future measurements.; The identification scheme employs a neural network to provide models of aerodynamic coefficients during dynamic-system operation. These models are valid over the entire input-output space. Prior to flight, a-priori data is incorporated into a base neural network using a new design and training algorithm. This algorithm functions in the face of an eight-dimension input vector. During flight, the parameters of the base neural are fixed, and a second set of activation functions are available for learning the surface created by the difference between the base neural network and the current dynamic-system information. The new neural network is demonstrated on a longitudinal-motion aircraft model, with static and dynamic training data, and its training speed, accuracy, and parsimony abilities versus existing neural networks are established.; The identification framework is used to identify the three longitudinal-motion coefficients of a twin-jet, transport aircraft. A localized feature is introduced into the lift-coefficient surface and performance of the model. The network learns new information from the dynamic-training data patterns, without loss of information in regions distant from the dynamic maneuvers. Approximation performance is evaluated with respect to both training and generalization data sets.