Abstract: In order to increase prediction precision of SVR (support vector regression) for catastrophic failures, an improved AdaBoost algorithm was proposed. It could obtain the weights of abnormal data in training sample set by AdaBoost algorithm and a weighted SVR was used to enhance the training of abnormal data, which could improve the prediction precision for catastrophic failure. The samples with small weight were discarded by using an adaptive weight trimming method to improve the training speed. The method was used to predict time series of an engine wear element and the one-step relative prediction error was 0.025. The experiment results demonstrate that the method can improve the speed of fault prediction effectively under desired accuracy.

Key words: methodology of system engineering, support vector regression, AdaBoost algorithm, catastrophic failure, fault prediction