The Research On Avionics Prognostics And Health Management Based On Walevet Neural Network

With the development of aviation technology, the complexity of avionics system is more and more high, the frequency of the failure and the resulting loss of it is more and more. Prognostics and Health Management(PHM) is a new technology to realize the maintenance and protection of aircraft, has become an important part of the design and use of the new generation avionics system.PHM is a technology based on the maintenance of visual about comprehensive condition monitoring, fault diagnosis, fault predition and health management. It is in the abnormal symptoms, to forecast the development trend of the fault, determine the system's remaining life and the future of the health status at a certain time, and select the appropriate time to take the maintenance strategy, prevent the system to complete failure, achieve security, reliability, reduce the use and security costs of the target.For the avionics PHM system of three modules about condition monitoring, fault diagnosis and fault prediciotn, this paper give a research on the multi-wavelet threshold denoising, wavelet packet feature vector extraction, wavelet neural network fault classifier and predictor etc. a number of key technologies. Specific research contents are as follows:1. The multi-wavelet threshold denoising. The step of the GHM multi-wavelet threshold denoising is studied, and the performance of four typical threshold estimators is analyzed. For the disadvantages of hard threshold and soft threshold function for multi-wavelet threshold denoising, an improved threshold function is proposed. Through simulation analysis, the improved threshold function is better than hard threshold and soft threshold function.The input SNR is 12 d B, the out SNR is 20.5867 d B when using the improved threshold function, denoising signal to noise ratio promote 71.56%, minimum mean square error is 0.3742.2. The wavelet packet feature vector extraction. The energy of the wavelet packet is used as the fault feature, which can decompose the weak early fault to the fault space easily detected, and provied training samples for neural network learning. The process of extracting feature vector of wavelet packet and optimal wavelet packet is simulated, the optimal wavelet packet to extract feature vector can largely improve the problem of ‘dimension disaster'.3. The health assessment. By comparing each layer of wavelet packet decomposition coefficient of residual error with preset threshold, determine the current system is healthy or abnormal.4. The neural network fault classifier and performance prediction. The traditional wavelet neural network learning algorithm and the improved wavelet neural network learning algorithm are studied. The improved algorithm adds damping term and adaptively changes the learning rate, so that it can improve the speed of convergence of the traditional steepest descent method, which is easy to fall into a very small point. And through the simulation experiment, verified the effectiveness of the algorithm. Fault diagnosis and fault forecast instance is given to validate the rationality and feasibility of wavelet neural network as fault classifier to fault diagnosis and as nonlinear function fitting to fault prediction.