(587b) Multivariate Control Charts in Radial Coordinates

McKetta Department of Chemical Engineering

The University of Texas at Austin, 1 University Station C0400, Austin, TX 78712

Mark Nixon, Willy Wojsznis, Ricardo Dunia

Emerson Process Management, Austin, TX

Email: mbaldea@che.utexas.edu

10B06 Process Monitoring and Fault Detection (https://aiche.confex.com/aiche/2014/10/sessions/sessioninfopopup.cgi?sessionid=26932)

For the past decades, the chemical industry has established itself as a prime source and collector of “big data” in the form of process information. These data are diligently stored on historian systems, with sample times that are on the order of seconds. However, they often remain dormant and underutilized, placing many branches of the chemical industry in the paradoxical situation of being data rich, but information poor.

In this paper, we focus on process fault detection as one of the key insights to be gained from historical and real-time data. We take a geometric approach based on the radial coordinate representation that we introduced recently [1]. In our previous work, we proposed representing each data sample on a Kiviat (“radar” or “star”) plot, with the plots of successive scans stacked along a time axis normal to the plotting plane. In this paper, we use this representation to define a multivariate control chart. We use the centroid of the radial representation of each sample as a condensed metric to define the state of the process at a given time instant. The centroid point can then be used to construct a multivariable control chart in the sense of MacGregor and Kourti [2] (where this concept has been developed for the two-dimensional case). The multivariate control chart is based on a confidence ellipse defined by the centroids computed for the normal operating state of the process. For fault detection, the real-time data samples whose centroids fall within the ellipse can be considered within specification, while data outside the ellipse are indicative of a faulty state.  

We apply these developments to an industrial data set concerning the operation of a compressor that is prone to surge as well as to fault data collected from simulating the Tennessee Eastman Challenge Process [3]. We show that the performance of the proposed method is at least comparable in terms of fault detection speed and missed detection rates to other approaches discussed in the literature [4,5]. In addition, we show that the method provides unique capabilities in detecting process-wide abnormal events, e.g., column flooding and compressor surge. Subsequently, we discuss an extension of our method to assessing the occurrence of faults during process transitions between operating states.

References:

[1] Ray C. Wang, Ricardo Dunia, Michael Baldea, Thomas F. Edgar, Visual Data Mining and Fault Detection in Radial Coordinates, American Institute of Chemical Engineers Annual Meeting, San

[2] J.F. MacGregor, T. Kourti, Statstical process control of multivariate processes, Control Eng. Pract. 3 (1995) 403–414.

[3]          J.J. Downs, E.F. Vogel, A plant-wide industrial process control problem, Comput. Chem. Eng. 17 (1993) 245–255.

[4]          E.L. Russell, L.H. Chiang, R.D. Braatz, Fault detection in industrial processes using canonical variate analysis and dynamic principal component analysis, Chemom. Intell. Lab. Syst. 51 (2000) 81–93.

[5]          Y. Zhang, Fault Detection and Diagnosis of Nonlinear Processes Using Improved Kernel Independent Component Analysis (KICA) and Support Vector Machine (SVM), Ind Eng Chem Res. 47 (2008) 6961–6971.