Probabilistic Fault Diagnosis Based on Monte Carlo and Nested-Loop Fisher Discriminant Analysis for Industrial Processes

详细信息    查看全文

• 作者：Chunhui ZhaoWei Wang ; Furong Gao
• 刊名：Industrial & Engineering Chemistry Research
• 出版年：2016
• 出版时间：December 21, 2016
• 年：2016
• 卷：55
• 期：50
• 页码：12896-12908
• 全文大小：557K
• ISSN：1520-5045

文摘

In this study, a new discriminant analysis (DA)-based fault diagnosis method is proposed based on the idea of variable selection to overcome the drawbacks of the existing methods. It addresses three issues. One is that a prejudgment strategy should be developed that can determine whether it is suitable to perform DA on the fault data. Second, it should solve the dependence problem of faulty variable-selection performance on the modeling samples. Third, it may not clearly distinguish between different faults using simple Boolean algebra for reconstruction-based diagnosis. To achieve the above purposes, a Monte Carlo (MC)-based evaluation method is developed to quantify the bias of fault center and judge whether discriminant analysis can be conducted. Next, a nested-loop Fisher discriminant analysis (NeLFDA) algorithm is used to extract meaningful directions by which to distinguish between normal and abnormal classes. Iterative faulty variable selection is conducted using MC and NeLFDA in which a stability index of fault-relevant variable significance is defined to isolate the significant faulty variables and reduce the influences of modeling samples. Next, 2-fold fault diagnosis models are then developed for general variables and faulty variables to explore their different variable correlations relative to the normal case and fault reconstruction is performed for both variable subsets. A probability index is calculated using Bayes’ rule for each fault sample to quantitatively evaluate the probability that it belongs to each fault type and to further judge the fault affiliation. The performance of the proposed method is illustrated by applying it to the cigarette-production industrial process.