Visualizing High-dimensional Operations Data
 

The ease and low cost of on-line measurement of common process variables have led to chemical plants that are replete with operational data. However, there has been no corresponding increase in information or knowledge available to process operators and engineers. In order to excel in this era of agile manufacturing, global sourcing and increased automation, chemical companies need to derive value and knowledge from available data. Agile manufacturing requires plants to produce multiple products (or grades) and switch between them frequently and quickly. Global sourcing and the inevitable differences in raw material properties from different vendors necessitate changes in process unit operating conditions are necessary, in order to maintain product quality within specifications. Also, as the level of automation in the plant increases, operations personnel find it increasingly difficult to understand the actions of complex automation. We have developed techniques that provide visualization of the processes, similar to a “digital cockpit” in an aircraft, and enable superior operation.

Figure 1: SOM of a polyethylene reactor operation plotted from 25 variables 
measured online. Each product grade is shown as a cluster in SOM. 


Figure 2: SOM of a bioreactor operation with 14 variables. 
Different operating steps show different profiles.

Self Organizing Maps (SOMs) are neural networks that use unsupervised learning to automatically reduce high-dimensional data to two-dimensions. Visualization of the SOM using vector quantization and projection then provides insight into the process such as (1) the current position of a unit in the operations landscape, (2) the nature (steady state, transition, etc) of the current operation, (3) periods of extensive change in a unit, and (4) disturbance/transition propagation front through various units. These can help operators to take corrective control actions at the right time. Figure 1 shows an example of an SOM developed to visualize the operation of a Polythelene reactor. Online data from 6 months of operation is plotted. The resulting SOM shows clear clusters in the SOM space corresponding to each product grade. Deviations from normal operation is projected outside the cluster; this can be used as an additional online quality indicator, particularly since product quality cannot be measured online in this process. Figure 2 shows the time evolution of a biochemical reactor plotted from 14 easily measurable process variables. As can be seen from this figure, different types of normal and abnormal process operations can also be easily visualized. 

Our current research is directed at exploiting SOM data reduction and visualization properties to process control, monitoring and, supervision, which is of great interest to our industrial collaborators, such as Honeywell.