Abnormal situations arise in a process plant due to deviations from the normal operating envelope of the plant. Although most plants are automated, many important operational tasks continue to require human intervention. Human operators are largely responsible for process monitoring and fault diagnosis and can be occasionally overwhelmed due to alarm floods and inadequate time for troubleshooting. This motivates the need for automated, accurate fault detection and diagnosis technologies. Data-driven process monitoring methods have become popular in recent years. These use the information in the data collected during the operation of the process to define regimes that can be characterized as normal or abnormal. In this paper, we seek to develop a novel, robust, data-driven technique that utilizes the strengths of multiple fault detection techniques.

Various statistical and machine learning methods like principal component analysis, decision trees, and neural networks have been used for fault detection and diagnosis. In literature, it has been observed that when fault detection is performed using any one technique can be highly accurate for some faults and inaccurate for others. The set of faults for which a classifier offers good performance depends on the classifier; that is, each classifier has its expertise and inadequacies. Hence, a combination of multiple classifier systems has been proposed along with a suitable strategy to combine their individual results through decision fusion. This is found to increase the overall performance of classification. Decision fusion methods can be classified into two categories utility-based and evidence-based methods. Utility-based methods (eg: Majority voting, Min, Max, Product) do not use any prior knowledge of the predictions, whereas evidence-based methods (eg: Naive-Bayes, Dempster-Shafer, Decision Template) use prior knowledge of the strengths and weaknesses of the method. Ghosh et al. (2011) performed evidence-based fusion strategies like weighted voting, Bayesian, and Dempster–Shafer on the Tennessee Eastman case study. They demonstrated that fusion offers complete fault coverage and an increase in overall fault recognition rate vis-a-vis single classifiers. In this work, we consider a decision fusion technique called Decision Templates, which has been successfully used in many domains and evaluate its applicability to process monitoring.

Decision template-based techniques use the weights assigned to a class by each classifier to make a classification decision. Without loss of generality, assume that the output from a classifier is a fuzzy number quantifying the likelihood of a class. These weights given by each classifier to the particular class can be organized in a (p x q) matrix, called the decision profile, where p is the number of classifiers and q is the number of classes. Hence for each class, there exists a characteristic decision template. The characteristic Decision Template (DTi) is formed in the training stage by taking the weighted average of the decision profiles belonging to a particular class (Ci). Once all the decision templates are trained, they are stored in memory. In the testing stage, a new sample is fed to all the classifiers, and a decision profile for that sample prepared from their outputs. The obtained decision profile is then compared with the stored decision templates of each class based on similarity measures, and the sample assigned to the class whose decision template offers the maximum similarity value.

In this paper, we use the proposed decision template-based fusion method on the well-known Tennessee Eastman challenge process (Downs and Vogel, 1993). We apply the proposed method using several classifiers such as Support Vector Machine, Neural Networks, and Quadratic Discriminant Analysis. The obtained results are then compared with those of the individual classifiers. A preliminary analysis showed an increase in classification accuracy of over 5% by the proposed method. This can be increased further by using appropriate combinations of classifiers, whose decision boundaries are different from each other. In this paper, we will describe the proposed decision template-based method and report the results from fault diagnosis on the Tennessee Eastman challenge problem.

References

1. Ghosh, K., Ng, Y. S., Srinivasan, R., Evaluation of decision fusion strategies for effective collaboration among heterogeneous fault diagnostic methods, Computers and Chemical Engineering, 35, 342–35 (2016)
2. Kuncheva, L.I., Bezdek, J.C., Duin, R.P.W., Decision Templates for Multiple Classifier Fusion: An Experimental Comparison. Pattern Recognition 34(2), 299–314. (2001)
3. Downs, J. J. and E. F. Vogel, A Plant-wide Industrial Process Control Problem,” Computers and Chemical Engineering, 17, 245-255. (1993)
4. Venkatasubramanian, V., Rengaswamy, R., Yin, K., Kavuri, S. N., A review of process fault detection and diagnosis. Part I: Quantitative model-based methods. Computers and Chemical Engineering, 27, 293–311. (2003)