An Improved RBF Network Research And Its Application In Material's Performance Prediction

Satellite Longevity Prediction (SLP) is a very significant problem that directly affects construction and development of aeronautic equipments.Because satellites'key exposed parts material performance degenerating law research is one important aspects of the whole predicton, their performance degenerating law prediction (PDLP) becomes more important. Currently, by means of physics experiments people often use mathematics modelling methods that are usually limited and cost a large quantity of labor, material resources and also time.Moreover, the method can not usually achieve goals.Therefore, we will propose an improved method, with which we will model the PDRP and provide a method for assitantly extrapolating materials'performance degenerating curves by means of computers.Satellites'key exposed parts material PDLP is one of the applications of Time Series Prediction (TSP). Firstly, we discuss some common problems about TSP, briefly illustrate the main process of solving TSP problems by means of an example, and summarize its research actuality. Secondly, we detailedly introduce a typical neural-network method that is RBF network and its common training algorithms, while discussing the advantages and disadvantages.One of widely-used RBF network algorithms is the algorithm that is based on both k-means clustering algorithm which is used in the first training phase and least means square (LMS) algorithm which is employed in the second training phase.The method has advantage of concise algorithm and high precision.It obtains the value of hidden cells'center vectors(when the kernel function is Guassian) by k-means which has high monopolization that means each sample completely vests in one cluster and learns weights from hidden layer to output layer by LMS which has too long computation steps in each iteration and so slow convergence rate.So,we propose an new method based on fuzzy c-means (FCM) clustering algorithm and approximate LMS with adaptive learning rate (ALR) for RBF network,which is able to restrain the monopolization efficiently, to fastern convergence rate and to improve learning precision.For the characteristics of the TPS and background of the application, we propose the experiment method of satellites'key exposed parts (Thermal Control Coasting) PDLP, which mainly includes transform process of time data, determination of network top structure and concrete operatoion process of the experiment. At last, we inllustrate our method is efficient and practical by means of the comparison of experiments'results between the two methods we talked first.In the end, we complete a demo system of SLP. The system includes Longevity prediction function, subsystem integration function, uniform model expression function and graph display function.