Flight Control System Using Dyadic Network

This paper surveys the application of neural networks in flight control system, and analyzes the difficulties in application. Then the paper proposes a brand new neural network, termed dyadic network, and a brand new fuzzy model, termed dyadic fuzzy system. Based on dyadic wavelet and Multiresolution Approximation theory, dyadic networks and dyadic fuzzy systems represent the approximation of the target function. A dyadic network is a linearly parameterized network. It is similar to a RBF network. Its basis functions are products of decomposition scale function of a biorthonormal wavelet. But it is different from the RBF network in approximation theory. Dyadic networks are easy to construct, only the units is needed to be determined, and computation burden is low, even for a network with an enormous number of neurons. A dyadic fuzzy system is a multiresolution model. All initial fuzzy regions of an input variable, called primary regions, have same lengths. The intervals between the centers of two adjacent primary regions are equal. Membership functions are box spline funcitons. Analyzing by discrete wavelet transform, we may combine two adjacent regions. A fuzzy region of the final model comprises 2 j (j∈Z+) primary regions. The goal is to establish a fuzzy system with fewest rules and given precision. The modeling algorithm is high efficient, insensitive to system dimension, and able to reduce noise. This paper focuses on a generic neural flight control, which can be applied to a wide range of vehicle classes. This direct adaptive tracking controller integrates feedback linearization theory with both pre-trained and on-line learning neural networks. Pre-trained neural networks are used to provide estimation of aerodynamic stability and control characteristics required for model inversion. On-line learning neural networks are used to generate command augmentation signals to compensate for errors in the estimates and from the model inversion. The neural flight control system uses reference models to specify desired handling qualities for different kinds of aircrafts. Dyadic networks and dyadic fuzzy system are used as on-line learning and pre-trained neural networks in simulations respectively, and results are presented to illustrate and compare the performance.