The goal of batch manufacturing is to safely produce a maximum yield batch within product quality specifications in the shortest amount of time and with a minimum amount of waste. However, achieving that goal presents many challenges. Operators often work in a highly complex, interlinked and dynamic environment and must manage a large amount of data and information on a running unit. This makes it easy for batches to end up with undesirable processing events or inadequate end-of-batch quality. It also makes it easy for operators and engineers to miss atypical process variation relationships.
The proliferation of data being collected in companies today as well as the expansion of instrumentation and number of measurements are increasing the complexity of the situation. Coupling this with the drive by many companies to do more despite a smaller work force creates a situation where even more things may go unnoticed.
The traditional univariate (one variable at a time) approach to problem solving may provide some insight. However, it can't adequately deal with this complexity because interactions often occur among variables. Unfortunately, the shortcomings of applying univariate approaches are not always realized.
Some companies also determine the quality of a batch by comparing it to a "golden batch," i.e., one that has met quality standards and is considered ideal for a particular process. However, this approach of relating results to a single batch doesn't take into account interactions among inputs and process variables and output quality characteristics. It does not incorporate the concepts of acceptable and unacceptable variation. And, it does not provide any process insight into what is and is not important regarding inputs to the process, the processing conditions, and the effect that this variation has on product quality characteristics. It's a univariate approach to what in reality is a multivariate problem. Trying to replicate the golden batch often just leads to chasing after "fool's gold."
It's better to compare a batch to a multivariate-based model, or a "golden profile," that compiles many acceptable batches and accounts for process relationships and process variability. This allows a much more accurate analysis of that batch and, depending on the techniques used, fault detection and end-of-batch quality prediction. In addition, it affords process insight unobtainable from a golden batch analysis.
However, online analytics that takes into account relationships among variables and also characterizes typical and atypical process and product-quality variation have proven challenging in batch operations for many reasons, including process holdups, access to lab data, feedstock variations, unsteady operations, data organization and concurrent batches. When done however, the benefits achieved usually are immediate and substantial.
To advance these concepts, we have partnered to develop and incorporate online multivariate batch process analytics into the manufacturing process to aid operators and engineers in dealing with this complexity. A six-month field trial started in late 2009 at Lubrizol's plant in Rouen, France; the plant is continuing to use the online batch analytics.