I was reading the great Control magazine article, Data Analytics in Batch Operations, by Lubrizol’s Robert Wojewodka and Emerson’s Terry Blevins. It describes the challenges in applying on-line analytics in batch production for early fault detection and prediction of end of batch quality parameters. If done right, this can help avoid out of spec batches, waste and rework and the opportunity cost of the lost batch. This can be critically important in specialty chemical and pharmaceutical manufacturing, which in the words of the authors, “depend heavily on batch processing to produce low-volume, high-value product.”
The reward is great if you can make these on-line analytics work but the challenges are great. One of the biggest is the time differences between batches. Operators and events can halts and restarts to process. These may be for manual additive additions, waiting on common equipment to become available or abnormal conditions that may develop.
Another challenge is that online measurement of quality parameters “may not be technically feasible or economically justified.” These lab samples must be available to online analytics toolset to perform these analytics.
Other challenges they noted included variations in feedstocks, varying operating conditions, concurrent batches and the assembly and organization of the data.
The concept of “Golden Batch” has been around a long time, which is the concept of comparing batches in progress, or just completed with ideal ones from the past. The authors point out two big weaknesses with this approach. The first is conditions indicated by each measurement may affect the product quality in different ways. Secondly, this is a univariate approach to a multivariate problem–no knowledge is gained of the relationships of process variations.
In a prior post, I’ve discussed some of the analytic tools, PCA and PLS. By applying these online, “…changes can be made in the batch to correct for detected faults or deviations in the predicted value of key quality parameters.”
I mentioned to Terry that I gave this article a close read and was working on a blog post. He told me the really innovative thing they were able to do was to apply dynamic time warping (DTW). The article describes its use:
…allows such [batch time length] variations to be addressed by synchronizing batch data automatically using key characteristics of a reference trajectory.
This normalization process is covered in detail in the Greg McMillan and Mike Boudreau’s book, New Directions in Bioprocess Modeling and Control and discussed in the article, PAT Tools for Accelerated Development and Improvement.
It’s hard to give the article justice in a short blog post, so if you have 18 minutes and 30 seconds to spare, watch to the interview Walt Boyes did with Bob, Terry and Philippe Moro at last year’s Emerson Exchange.