(703g) Visualization and Data-Driven Monitoring of Batch Processes

Visualization and Data-Driven Monitoring of Batch Processes
Ray Wang, Michael Baldea, and Thomas F. Edgar

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

The operation of batch processes differs significantly from that of continuous processes. In continuous processes, one or more steady states are well-defined and are typically maintained. On the other hand, in batch processes, the final product and its desired quality parameters are well-known, but are reached via an inherently transient path. This feature poses specific challenges in the monitoring of batch processes, and has led, e.g., to batch-specific modifications of projection-based data-driven monitoring and fault detection tools such as dynamic principal component analysis (DPCA).

In this paper, we present a novel, geometric, method for data-driven monitoring of batch processes. Our work relies on an extension of the time-explicit Kiviat (radial) plot visualization and fault detection framework that we have previously developed for continuous processes [Wang et al., AIChE J., submitted]. In this framework, each sample of a multivariate time series data set collected from process operations is represented in a radial plot. The radial plots are stacked vertically in the order in which the samples were acquired, resulting in a time-explicit representation of the multivariate time series. The geometric properties of this representation allow for establishing an elliptic confidence region for the centroids of the polygons corresponding to each data sample, which in turn can be used to define the normal steady-state and faulty operating states of the corresponding process.

In order to extend these ideas to batch processes, we redefine confidence ellipses: rather than computing a single confidence region using multiple samples collected from steady-state operation, we calculate separate confidence ellipses at every sample point, based on data acquired from multiple batches. This allows us to establish a “funnel” for batch trajectories, with the expectation that the trajectories of successful batches would evolve within the funnel, while faulty batches exit the funnel at some time point before completion.  We study the effect of batch alignment on our proposed framework, and discuss its online implementation. Finally, we present an application to an industrial case study, along with a comparison of the fault detection performance of our methodology with that of other methods for batch process monitoring available in the literature.