Weighted Support Vector Machine Applications In Reliability Prediction

As the developing of modern industry, the reliability of products has become more and more important. A precise reliability prediction of an industrial manufacture can help discover a series of problems that may be found during its lifetime as soon as possible, which may lead to an easier control to the lifecycle. Therefore, the research for the precise reliability prediction is significant in modern system engineering.Traditional functions in reliability prediction are mainly constituted of the models those have already got efficient implementations in the field of nonlinear regression, including the Lifecycle Distribution Model, the Fault Tree Analysis(FTA), the Monte Carlo Model and the Artificial Neural Network(ANN), et al. The ANN, which may be especially mentioned, has become an important field that absorbs global researchers'attentions for its great precision of sampling fit, and thus has got many improved models. However, the ANN still remains some limitations in its own theory. The principle of minimizing the empirical risk, which leads to an"over-fit"that limits the ability of generalization of training machine, is the main one. On the other hand, training for ANN needs a lot of samples while the samples in a real world may usually meet a limit. So, in most situations, there's no satisfying precision for prediction in an ANN without enough samples.Support Vector Machine (SVM), which is mainly used in pattern recognition problems at the beginning, is a fresh machine study method put forward by Vapnik using statistics principles in early 1990s. As the import ofεinsensitive loss function, SVM has already been extended in the regression estimation of nonlinear systems, and has represented its good study ability in regression under the situation of solving small sample. SVM makes minimizing of structure risk as its criteria, and gets both satisfying precision and great extending ability in data fitting. In addition, the solutions of SVM transform to the solutions of quadratic programming problems at last. SVM is thus the only solution and the global optimal solution too.As people's in-depth studying of SVM theory, new SVMs were being built. In the prediction of reliability, data sampled in different time made different effect in the result of prediction. Recent data affected the result more than the earlier data. Based on the very characteristic of reliability prediction, the papers used a new SVM model called Weighted SVM (WSVM) to model and predict the system reliability. Different from the traditional SVMs'giving same punish parameters to all the samples, WSVM gave different punish parameters to the samples of different periods. Thereby, a significantly increase in the proportion of recent samples took in the prediction was made. At the same time, an optimization to relative parameters of WSVM through genetic algorithm made a further improvement to the prediction accuracy of WSVM. The analysis to the mean square deviation showed us the conclusion, that the prediction accuracy of WSVM was better than the ANN and traditional SVM models.