Unilever Optimizes Hellmann’s Mayo Fill Levels with AI-based Predictive Analytics

This edge-based approach allowed the company to analyze and predict potential drifts in equipment performance over time, enabling real-time adjustments at the edge. The predictive model was trained using both historical and live data, which helped in exploring the feature space and determining the relationship between input variables and the desired fill level. “Once we determine which variables are contributing in this use case, we use that prediction for the next jar and then determine, okay, you're about to overfill by, let's say three or four grams. We can make that adjustment and reduce it,” Resseguie said.Project specifications.

By reducing the standard deviation of fill amounts across different filler heads, the company was able to operate closer to the target set point, significantly minimizing waste. Notably, Ramos and Castro de Almeida had determined that all of the fill head nozzles were performing efficiently and comparably before upstream inputs were selected, thus controlling the filler heads themselves as variables. The process involved selecting key input variables through a correlation matrix and using them to train the model. The representative described the process: "You can drag and drop and select those inputs and then, from there, you train the model.”

The integration of LogixAI into the company's existing systems was facilitated by its compatibility with Rockwell’s Studio 5000 and ladder logic, allowing for seamless control and automation at the control layer. “You actually have variables that control when you train and when you calculate. So that's what allows you to truly automate this entire process at the control layer,” Resseguie said.

Ramos and Castro de Almeida's workflow for this application was as follows. They first identified the variable of interest—the target dosed weight. The next questions is, ‘what are the input variables that impact that outcome, thus can help us predict what it will be?’ That can be a slippery question to answer because there can be thousands of variables that can be produced as data tags in the PLC.

“So work that was done beforehand that at least helped us narrow down the scope of the data by doing that correlation matrix, helping us at least identify what are some of the top key variables that we want to utilize,” Resseguie said.

Unilever reduced finished goods overfill giveaway on the line by 50%. When extrapolating across millions of packages this drives significant savings and margin improvement.

The project is indicative of Rockwell’s recent practice of combining traditional soft sensing technology with AI and machine learning algorithms to analyze multiple input variables in real-time to make fill weight predictions to optimize fill accuracy. These insights are directly integrated back to a Logix controller to actuate and execute adjustments in real time. One the show floor, a technical demo built on new FT Optix HMI/SCADA/IoT software demonstrated LogixAI Perfect Fill impact on fill accuracy by toggling functionality on and off.