贝叶斯证据框架下最小二乘支持向量机的软件老化检测方法

A Software Aging Detection Method Based on Bayesian Framework Using Least Squares Support Vector Machine

针对当前软件老化的检测、分析和软件再生的不确定性问题,提出了一种基于贝叶斯证据框架的最小二乘支持向量机(LS-SVM)的软件老化检测方法,即:使用最小二乘支持向量机分类器进行数据分类,以此解决数据采集时出现的小样本、高纬度,非线性、局部最小值等问题;通过贝叶斯证据框架来优化LS-SVM的超参数,从而提高分类器的学习精度和泛化能力。实验结果表明,在状态清晰区间,软件老化的概率均在0.7至0.9之间,而高维模型检测出的软件老化的概率为0或1。如果从概率粒度层来描述软件老化,则软件再生的时间点选取效率更高,根据概率值的变化可进一步解析软件老化的不确定性。实验结果及分析显示,概率粒度所描述的软件健康状态更符合软件老化的客观状况。

A software aging detection method based on Bayesian evidence framework using least squares support vector machine （LSSVM） is proposed to solve the uncertainties in software aging detection and analyzing, and software regeneration. The least squares support vector machine classifier is used to classify the data, so that the problems such as small sample of the data collected, the highlatitude, nonlinearity, and local minimum can be solved. Then, the Bayesian evidence framework is employed to get the optimal LSSVM hyperparameters, and hence, the learning accuracy and generalization ability of the classifier are improved. Experimental results in the state clear interval show that the probabilities of software aging given by the proposed method are all between 07 and 09, while the probabilities given by the highdimensional model are 0 or 1. When the software aging is described by the probability granularity, the regeneration time can be more effectively selected; moreover, the uncertainty of the software aging can be further analyzed from the changes of the probabilities. Experimental and analyzing results show that the software health status described by the probability granularity is better in line with the practical status of software aging.