444199 Detecting Changes in the Catalyst Degradation Rate of a Fixed Bed Reactor

Tuesday, April 12, 2016: 10:37 AM
335B (Hilton Americas - Houston)
Jim Sturnfield and Daniel W. Trahan, Engineering & Process Sciences, The Dow Chemical Company, Freeport, TX

Analyzing process data to understand production issues has numerous challenges. There is a large amount of data with a variety of sources of measurement variations and errors. In order to determine the cause of an undesired change in the production process, the data often needs to be analyzed for corresponding changes in the other variables. This can be especially difficult when the exact time of the change is hidden by the variability within the data set. The underlying physics can provide insights to the analysis. These physics suggest underlying transformed variables that often reduce the noise and suggest parameters that can be fitted to enforce the physics on the process. The fitting of these parameters is complicated due to the limit range of process variables that are often found in production operations.

This presentation examines the degradation of catalyst in a fixed bed reactor and explores methods to detect the event that is connected with the change in the catalyst degradation pattern of the reactor bed. This degradation results in changes in the pressure drop. For the reactors being considered, a steady increase in this pressure drop is inherent, but the rate of pressure drop growth increases for unknown reasons. The pressure drop is also dependent on flow rate and process conditions, so it can difficult to identify the actual time of the change, making it difficult to determine potential causes for the changes in the underlying degradation.

A catalyst production run can last for years, although the current work has been limited to analyzing about 60 days in each study. In these studies, the data set involves about ninety sensors with 1 minute data intervals, so over 7 million data points are being considered. There are also a variety of sources of data errors with some being random error, others being systematic or measurement drifts, and even total sensor failure. The non-random error variables can sometimes be managed by considering a physical balance equation that involves multiple sensors. Knowledge of the physics can also be used to predict other variables that can be useful in analyzing the data.

For the fixed bed reactor example, the complexity of the physics and the lack of relevant information resulted in a number of competing simplifying assumptions that could be used to build a degradation process model.  Each of these potential models was considered in the data analysis in order to separate the process variability from the degradation rate changes. The comparison of these different fits and an analysis of the residue of these fits provide better identification of the time of the changes and provide insights into underlying physical changes. The analysis of changes in the data is then used to identify potential events that might be the source of the change in the degradation.

Extended Abstract: File Not Uploaded
See more of this Session: Big Data Analytics and Fundamental Modeling II
See more of this Group/Topical: Topical A: 2nd Big Data Analytics