Insight into CO2 fluctuations at a soft drink manufacturer

Predicting the future and thereby improving processes and resolving issues?

Those are the type of innovative projects that we at Batenburg Bellt love to work on. We focus on making our clients' production, cultivation, and real estate operations more efficient, safer, and sustainable. We also aim to encourage this approach among the new generation through challenging graduation projects. Thomas Bos had the opportunity to carry out such an innovative project at one of the largest soft drink manufacturers in the Netherlands: Domain-Knowledge-Driven Explainable Product Quality Prediction.

Many of the beverages produced in this factory contain carbon dioxide, which requires each filled bottle to meet strict regulations. The amount of CO2 in the bottles is measured, and if the measurements fall outside the tight margins, it hampers the production process. Additionally, if research reveals that too many bottles are outside the margins, the entire batch cannot be used anymore. Naturally, they want to avoid this situation, and since fluctuations in the measured amount of CO2 in the bottles were observed, they enlisted the help of Thomas Bos!

Past data helps predict the future

By investigating the sources of fluctuations in the measured amount of CO2 in filled bottles, a solution to reduce downtime on a filling line was sought. For this purpose, all historical data stored in the factory in Bunnik was utilized. Based on this data, it was possible to determine the reasons for the fluctuations in the measured CO2 quantity and, more importantly, if a solution could be found. Often, artificial intelligence and machine learning are used for such investigations. However, it is often unclear what these algorithms actually do under the hood. As this is crucial information for the engineers in the factory to solve the problem, they explored the concept of Explainable AI. This methodology attempts to provide understandable quality predictions, with the quality in this case being the CO2 quantity in the bottle. It involves using domain knowledge from the engineers to make the predictions more plausible. The combination of historical data with the process knowledge of the engineers offers the desired insights.

Innovative and challenging research

What makes this project unique is the use of so-called Counterfactual Explanations (CFE), which has not been previously employed in Machine Learning for the process industry. The starting point is a Random Forest model based on the current process state. A CFE essentially provides an alternative process state based on the desired final state. It indicates which process parameters (such as temperature or pressure) need to be changed to transition from the current process state (e.g., low CO2) to the desired process state (CO2 within the margin). Since a Random Forest model is inherently explainable, a CFE can be used to explain what is necessary to achieve the optimal process state. This information proved to be highly valuable for their engineers as it elucidates the reasons behind certain occurrences in the process.

There has been relatively little research on the use of AI and Machine Learning in the process industry. It is mostly applied to predictive maintenance rather than process optimization. Consequently, Thomas had to experiment with various methodologies and algorithms to determine what works and what doesn't. He also conducted simulations to further test and verify his methodology, in addition to using the existing dataset from the factory.

Future prospects

The project investigated the reasons for fluctuations in the measured CO2 quantity in filled bottles. The results now provide this soft drink manufacturer with insights to mitigate these fluctuations in the future. Reduced or lower fluctuations mean that the factory can stay within the specified CO2 margins more frequently, resulting in less downtime for the filling line and reduced waste due to fewer discarded batches. Overall, this leads to a more efficient production process!

Batenburg Industrial Automation

We focus every day on making the production, cultivation and property management processes of our clients smarter, safer and more sustainable. We do this by supplying smarter products, installing these and by developing advanced software to control processes and systems. This includes digital twins, predictive maintenance and augmented and virtual reality solutions. In close cooperation with clients, partners, universities and research institutes, we make the latest technologies applicable in practice.