The Prediction Of Small Sample Of Events Based On Support Vector Machine

Firstly, the research status of fault prediction and health management (PHM) andseveral typical prediction methods of the small sample event are summarized, and thesupport vector machine (SVM) is also discussed in detail. In many complex industrialsystems, due to the high risk of system and other components failures, it is eager that areliable scheme of the fault detection and prognosis should be developed to guarantee asafe operation in spite of the existence of the uncertainties, which motives the PHM.Moreover, the fault prediction has attracted considerable attention in response to itsincreasing importance in the PHM. The basic methods of the fault prediction can beclassified into three kinds including model-based one, knowledge-based one anddata-based one. With the increasing complexity of the practical systems, model-basedapproach has been becoming restrictive. In contract, data-base approach becomes popularbecause of the convenience of the data access. Unfortunately, in many practical systems,the obtained sample may be limited. So it is desired to establish the theories to guide ourlearning and prediction under the case of small sample. Nicely, the SVM addressed in thethesis is just an effective method to solve the small sample problem.Then the simulations were performed to check the effectiveness and deficiency ofSVM. Here, we focus on the regression ideal of the SVM. First of all, we try to be familiarwith the effects which results from the method of pretreatments and the selection ofparameters. Then the optimal parameters and the scheme of the pretreatment are derived,which provide us convenience to proceed with one-step and two-step prediction. Thesimulation results demonstrate the effectiveness of the SVM method for predicting futuredata. It is worth mentioning that the example used in our simulation comes from a trackingcontrol problem of the city bus, which was declared by the International Federation ofAutomatic Control (IFAC) in1990.Finally, for the sake of improving the effectiveness the SVM, we combine the fuzzyinformation granulation (FIG) and the SVM to reduce the error between the predictionvalue and the real value. As we know, it may be very difficult to predict the precise value in practice because of the existence of the external disturbances and uncertainties. Thus, itwill also make sense to predict future trends instead of accurate values. So thecombination of the SVM and the FIG is utilized to predict the future trends and the FIG isused for the pretreatment of SVM. For comparing with previous results, the samesimulation example is adopted, which shows that the combined method can predict thefuture trend better.