【作者】 许玉格； 莫华森； 杨舒乔；

【Author】 Yuge Xu;Huasen Mo;Shuqiao Yang;School of Automation Science and engineering,South China University of Technology;

【机构】 华南理工大学自动化科学与工程学院；

【摘要】 污水处理故障诊断数据的典型不平衡特性,严重影响了故障诊断的效果,尤其是导致故障类别的正确率偏低.针对此问题,采用基于Stacking的集成算法,以两层的叠加式框架结构建立污水处理故障诊断模型.在此基础上提出了两种改进框架,其一选择支持向量机、相关向量机和加权极限学习机作为元分类器,构造一种多基分类器多元分类的集成结构,多元分类加权投票得出最终诊断结果,降低了单一元分类器可能出现的偏差;其二以不平衡分类性能指标G-mean值为基础,定义基分类器输出权值计算公式,通过权值对基分类器的输出结果进行校正后再融合,使得融合结果更加稳定.在KEEL数据集上验证了模型效果后,建立污水处理故障诊断模型,实验结果表明基于Stacking集成算法的污水处理故障诊断模型性能优于其他对比算法,两种改进框架在性能上比传统Stacking集成算法有所提高,可有效提高G-mean值和整体分类正确率,特别是提高了故障类别的识别正确率.更多还原

【Abstract】 The typical imbalance characteristic in wastewater treatment seriously affects the effect of fault diagnosis,especially leads to the low accuracy of fault diagnosis.Aiming at this problem,one Stacking ensemble algorithm is proposed to establish the fault diagnosis model of wastewater treatment with a two-layer superimposed frame structure.On this basis,two improved frameworks are proposed.Firstly,support vector machine,relevant vector machine and weighted extreme learning machine are selected as meta-classifiers to construct an ensemble structure of multi-base classifiers with multiple classifications,and the final diagnosis result is obtained by weighted voting of multiple classifications,which reduces the possible deviation of single meta-classifier.Secondly,the output weight calculation formula of the base classifier is defined based on the G-mean value of the imbalanced classification performance index.The output results of the base classifier are corrected by the weight value and then fused to make the fusion results more stable.The simulation experiments on KEEL data sets and wastewater treatment data set,indicates that the performance of Stacking algorithm is better than other comparative algorithms,two improved frameworks outperform the conventional Stacking algorithm,which can effectively improve the G-mean value and the overall classification accuracy,especially increase the recognition accuracy of the faulty category.更多还原