464413 Nonlinear SVM-Based Feature Selection for Fault Detection and Diagnosis of Continuous Processes

Wednesday, November 16, 2016: 9:08 AM
Monterey II (Hotel Nikko San Francisco)
Chris A. Kieslich1,2, Yannis A. Guzman1,2,3, Melis Onel1,2 and Christodoulos A. Floudas1,2, (1)Artie McFerrin Department of Chemical Engineering, Texas A&M University, College Station, TX, (2)Texas A&M Energy Institute, Texas A&M University, College Station, TX, (3)Department of Chemical and Biological Engineering, Princeton University, Princeton, NJ

Dimensionality reduction is a key task in most data-driven applications, in areas such as multi-scale systems engineering [1], where vast amounts of data must be reduced to an essential subset that is used to provide actionable insights. Advances in machine learning and pattern recognition techniques, such as support vector machines (SVMs) [2-6], have allowed the use of very high dimensional feature sets for tasks such as classification and regression; however, it is often desired to reduce a high-dimensional feature space to a minimal set of features in order to gain practical insight, reduce overhead of data collection, and to improve accuracy. For the case of linear SVMs models, vector weights can be used to rank the contribution of features [7]. However, for nonlinear SVMs, feature ranking is nontrivial since the kernel trick obfuscates vector weights. Accurate feature ranking for nonlinear SVM models is highly desired, since nonlinear SVMs models are often necessary to describe the complexities of many systems of interest. In order to rank features of nonlinear SVM models, we have derived novel kernel-dependent SVM feature rank criteria [8] derived from sensitivity analysis of the dual SVM objective. Additionally, we have developed a platform for support vector-based feature selection that is powered by the newly derived criteria, and utilizes existing and novel greedy algorithms to rank features.

Process monitoring of continuous processes represents a key potential application of these theoretical and methodological advances in nonlinear support vector-based feature selection. Fault detection and diagnosis are specific aspects of process monitoring that can be addressed by using SVM classification and will benefit from the developed platform for feature selection. To this end, specific two-class SVM models are trained to detect known faults, while one-class support vector data descriptors are used to characterize normal operations. In these models, the manipulated and measured variables of the process compose our input feature space, and instances of normal and faulty operation are used as training samples for SVM models. The developed platform for support vector-based feature selection is then used to improve the accuracy of fault detection models, as well as to perform fault diagnosis. We present results for the Tennessee Eastman [9,10] process as a case study and compare our approach to existing approaches for fault detection and diagnosis.

[1]. Floudas, C. A., Niziolek, A. M., Onel, O., & Matthews, L. R. (2016). Multi‐Scale Systems Engineering for Energy and the Environment: Challenges and Opportunities. AIChE Journal, 62(3), 602-623.

[2]. Vapnik, V. (1995). The Nature of Statistical Learning Theory. Springer.

[3]. Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3), 273-297.

[4]. Schölkopf, B., Smola, A. J., Williamson, R. C., & Bartlett, P. L. (2000). New support vector algorithms. Neural Computation, 12(5), 1207-1245.

[5]. Schölkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J., & Williamson, R. C. (2001). Estimating the support of a high-dimensional distribution. Neural Computation, 13(7), 1443-1471.

[6]. Tax, D. M., & Duin, R. P. (2004). Support vector data description. Machine Learning, 54(1), 45-66.

[7]. Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for cancer classification using support vector machines. Machine learning,46(1-3), 389-422.

[8]. Kieslich, C. A., Tamamis, P., Guzman, Y. A., Onel, M., & Floudas, C. A. (2016). Highly Accurate Structure-Based Prediction of HIV-1 Coreceptor Usage Suggests Intermolecular Interactions Driving Tropism. PloS one,11(2), e0148974.

[9]. Downs, J. J., & Vogel, E. F. (1993). A plant-wide industrial process control problem. Computers & chemical engineering, 17(3), 245-255.

[10]. Chiang, L. H., Braatz, R. D., & Russell, E. L. (2001). Fault Detection and Diagnosis in Industrial Systems. Springer Science & Business Media.

Extended Abstract: File Not Uploaded