基于K-means聚类的TE过程故障诊断与识别

Fault diagnosis and identification based on K-means clustering of tennessee eastman process

准确地诊断与识别化工系统的故障对保障系统的长期安全运行和高质量生产具有重要的意义。利用K-means聚类方法对标准田纳西-伊斯曼(TE)过程故障进行诊断,并通过主元分析(PCA)方法识别了故障发生的原因。首先,选择正常工况数据与某一故障工况数据组成新数据集,并用z-score标准化方法预处理新数据集,初始聚类中心数量为新数据集包含的工况数,通过分类性能指标F1-score(精确率和召回率的加权平均值)评价K-means聚类方法的故障诊断能力。其次,针对每种故障工况的数据集,采用PCA方法计算数据集中每个变量的统计量(T^2和SPE),统计量越大的变量越有可能引起故障。研究结果表明,K-means方法对TE过程的故障1、2、6和18能够100%诊断,主元分析对故障原因的识别结果与TE过程知识完全符合。与使用PCA方法和支持向量机方法故障诊断的结果相比,K-means方法对二者难以诊断的故障3、9和15有更好的诊断能力。

It is of great importance to accurately diagnose and identify faults in chemical systems for their long-term safety operation and high-quality production.The K-means clustering method was used to diagnose faults on the benchmark Tennessee-Eastman(TE)process.Principal component analysis(PCA)was applied in identifying fault causes.Firstly,the new data set was composed of normal operation process data and process data from a faulty operation of TE process.New data set was pre-processed by the z-score standardization method.The count of initial clustering centers was the number of operations in new data set.The performance of fault diagnosis of K-means clustering was evaluated by the F1-score(weighted average of precision rate and recall rate).Secondly,PCA was applied to calculate statistics of every variable for each faulty operation process data set,including T^2 and SPE.The larger the statistics,the more likely the variable was to cause fault.The results show that K-means method can diagnose the faults 1,2,6 and 18 in TE process with one hundred percent.The fault causes identified by PCA have a good agreement with the knowledge of TE process.The F1-score of K-means clustering of faults 3,9 and 15 are better than PCA and the support vector machine method.