基于MLFDA的化工过程故障模式分类方法
Multiblock local Fisher discriminant analysis for chemical process fault classification

摘要： Fisher判别分析(FDA)是一种有效的化工过程故障模式分类方法,但是其忽视了数据局部结构信息的挖掘。针对该问题,提出一种多块局部Fisher判别分析(MLFDA)方法,以更有效地识别化工过程故障。从变量和样本两个维度来分析数据的局部结构特性。针对变量维度的局部信息挖掘问题,设计了一种基于变量与数据集主元空间的相关度的变量分块方法,将全局过程变量划分为多个局部变量块。进一步考虑到样本维度的局部结构特性,应用基于局部权重因子的局部Fisher判别分析(LFDA)为每个局部变量块构建分类器。提出一种基于分类性能加权的多分类器集成方法,以融合不同分类器的决策结果。在Tennessee Eastman过程上的仿真结果说明,MLFDA方法具有比传统的FDA和LFDA方法更低的故障误分类率。
Abstract: Fisher discriminant analysis(FDA)was an effective chemical process fault classification method. However, the local data structure information was not investigated within traditional FDA method. To deal with this problem, a multiblock local Fisher discriminant analysis(MLFDA)method was proposed for more effective chemical process fault recognition. This method analyzed the local data structure characteristics from the variable-dimension and sample-dimension. To mine the local information in the variable-dimension, a variable block division method was designed based on the relevancy between the variables and the principal component subspace of the dataset, with which all the variables could be divided into several local variable blocks. Furthermore, considering the characteristics of local sample structure, the local FDA(LFDA)using local weighting factors was applied to construct classifier for each local variable block. An integrating strategy based on weighting classification performance weighting was presented to combine the results from different classifiers. Simulation results on Tennessee Eastman process showed that the proposed MLFDA method had a lower misclassification rate than traditional FDA and LFDA methods