What do pastry chefs and the pharmaceutical industry have in common?

Many substances we need on a day-to-day basis, such as chemical products and medications, have their origins in the process industry. Their manufacture involves chemical reactions, and these complex processes can’t be stopped at just any point. That means they will sometimes keep running even if something’s wrong. The sooner deviations are detected, the better the chance of taking action in good time. Artificial Intelligence makes this possible.

Daniel Labisch, Project Manager Advanced Technologies at Siemens Process Automation, and his team have designed an algorithm to prevent critical situations like this in the process industry. It avoids the need to shut down systems quite so often. To make this happen, the algorithm learns the normal workflow for a manufacturing process, and triggers an alarm if any relevant deviations are detected. This new diagnostic method will soon be available to customers as an app. For his contributions to new products, Labisch is being recognized as Inventor of the Year 2020 in the “Talents” category.

“The process industry is different from industries that create discrete objects, like the automobile industry,” comments Labisch. Recipe-based production processes are used in many processing facilities in the chemical and pharmaceutical industry. “You can imagine recipe-based processes as being a little like baking a cake. There’s a recipe that describes every individual step: add the ingredients, stir, wait. Some steps can also run in parallel: preheating the oven while the dough rises, for example.” For Labisch, one fact makes such recipe-based processes particularly fascinating: “In any given step, the system usually operates dynamically. So, to continue the cake analogy, while the yeast reacts in the dough, causing it to increase in volume, the temperature in the oven increases.”

The chemical reactions in the industrial process are much more complex than with leavened dough, however. “There are hundreds of variables in play,” says Labisch. “And the connections can be so complex that we as humans can no longer keep track of them all.” The algorithm he and his team have designed is able to manage this level of complexity, and delivers specific, transparent results to the customer. Starting with good production data, artificial intelligence learns the combinations in which a given behavior is normal, and at which point it needs to trigger an alarm. “For this learning process, we take all the measurements from a system and use them to train a neural network so the algorithm can recognize the parameters which represent “good” production. On that basis it can then determine when something isn’t right and it needs to trigger an alarm.”

Labisch considers this a significant benefit for the customer. An experienced plant operator can draw the correct conclusions from this notification and the fact there’s a deviation from the “expected normal,” and take the appropriate measures in good time. In addition, however, the algorithm can be used for analyzing historical data. The target group Labisch has in mind for this aspect is process engineers who would like to improve the production process. “Our diagnostic method will soon be used in an app to monitor the performance of a technical system. It’s a user-friendly way of bringing the physical and digital worlds together.” The application is currently in the pilot stage.

Labisch describes the pathway from idea to pilot application as a positive example of teamwork. “Once again it’s become clear that, for some approaches, the time has to be right.” The methodology underlying the algorithm is known in technical-speak as “self-organizing maps.” He says it’s something his colleague Thomas Bierweiler was working on more than ten years ago. At the time, the solution didn’t achieve the desired effect, so it was put on the back-burner. “Then, about five years ago, we worked together to adapt it as the starting point for our current developments, and now we have a product as the result. This clearly shows once again that you have to keep working on things for the long term, and be persistent if you want a good solution at the end. You can’t just give up.”