Research On Wavelet Neural Network And Its Application In Flight Control System

Wavelet neural network is a novel network combined with wavelet analysis and artificial neural network. Because the wavelet neural network inherits the self-learning ability of neural network and time-frequency localization of wavelet analysis, it can tolerant more fault and approach function more closely. The applications of wavelet neural network in approaching nonlinear function or signal, sorting signal, system identification, dynamic model building and predicting nonstationary time series were researched widely. The research and application of wavelet neural network are solidly based on the theories of wavelet analysis and neural network. But the theoretical basis of wavelet neural network is not perfect and complete. There are many difficult problems such as constructing wavelet network and initializing its parameters etc. In this dissertation, how to construct, train and optimize wavelet neural network, and its applications in flight control system, were researched.The base of fault tolerant control and control law reconfiguration of flight control system is fault diagnosis. Diagnosing fault of flight control system exactly in time is crucial for the aircraft flying safety.The main contents and achievements of this dissertation are summarized as follows:1. The constructing algorithms of wavelet network based on wavelet frame and based on multiscale analysis were studied. The structure and algorithm of multi-wavelet network were researched in detail, and then a novel adaptive multi-wavelet network was designed and its training algorithm was given.2. A new method of initializing wavelet network was presented. Three difficult problems that how to select wavelet functions, to decide the number of hidden units and to initialize weights of these units were researched. The wavelet function, which follows the rule that this function must be orthogonal, smooth and symmetrical and must have a tight support, can be selected experiential and experimentally by fact and the signal's properties. After analyzing time-frequency characters of wavelet function and signal, the number of hidden units is decided. The formula of initial weights was also proposed.