(719e) Root Cause Analysis of Process Faults Using Penalized Piecewise Linear Multiple Birth Support Vector Machine (pPWL-MBSVM)

Early and accurate root cause diagnosis of faults in chemical processes is crucial in reducing process downtime, improving safety and reducing manufacturing costs [1]. The difficulty to develop accurate first principle models and the overwhelming amount of stored data have made data-based techniques one of the most successfully applied tools for root cause diagnosis in chemical industry. One of the most widely used data-based method for root cause diagnosis is support vector machine (SVM) [2, 3]. SVM is a data-based binary classification method which builds a hyperplane between two classes to distinguish between two root causes. In order to perform diagnosis in presence of multiple root causes, SVM has been extended for multiclass classification as multiple birth SVM (MBSVM). It builds nonparallel hyperplanes such that the constructed hyperplane for a class is farthest to points in that class but proximal to rest of the classes [4]. However, training of MBSVM suffers from data imbalance as the number of data points in a class is much less than the combined number of data points in the rest of classes. Additionally, nonlinearity between classes in SVM is handled by using kernel transformations which leads to exponential computational load in the presence of high number of classes [5]. The poor training of MBSVM due to data imbalance and high complexity (due to nonlinear relationships) leads to misclassification error, that is, ineffective root cause analysis.

In order to handle the disadvantages of MBSVM, this work proposes penalized piecewise linear multiple birth support vector machine (pPWL-MBSVM) which simultaneously incorporates a penalty factor and the piecewise linear (PWL) classifiers to the MBSVM framework to account for data imbalance in training and high complexity, respectively. First, the classification model builds a hyperplane for a class by weighing the data point for the rest of classes by a penalty factor, which is the ratio of the number of data points in the particular class to that in the rest of classes. Hence, the penalty factor balances the datapoints on both sides of the hyperplane. Second, instead of using kernel transformation for nonlinear classes, a PWL hyperplane is built with an optimal number of planes to avoid overfitting. The effectiveness of the proposed method, pPWL-MBSVM, has been demonstrated with a case study of the Tennessee Eastman process. Hence, this work presents a novel methodology for root cause analysis involving multiple root causes while handling data imbalance in model training and high complexity due to nonlinear relationships between classes.

Keywords: support vector machine, piecewise linear classifier, data imbalance

References:

1. R. Maurya, R. Rengaswamy, V. Venkatasubramanian, A signed directed graph and qualitative trend analysis-based framework for incipient fault diagnosis, Chemical Engineering Research and Design, 85, 10, 2007, pp. 1407 - 1422
2. Onel, C. A. Kieslich, Y. A. Guzman, C. A. Floudas, E. N. Pistikopoulos, Big data approach to batch process monitoring: Simultaneous fault detection and diagnosis using nonlinear support vector machine-based feature selection, Computers & Chemical Engineering, 116, 2018, pp. 503 - 520
3. Ding, X. Zhao, J. Zhang, A review on multi-class TWSVM, Artificial Intelligence, 52, 2019, pp. 775 - 801
4. Zhang, S. Ding, Y. Xue. An improved multiple birth support vector machine for pattern classification, Neurocomputing, 225, 2017, pp. 119 - 128
5. Chen, B. Tang, T. Song, L. Li, Multi-fault diagnosis study on roller bearing based on multi-kernel support vector machine with chaotic particle swarm optimization, Measurement, 47, 2014, pp. 576 - 590