Keeping your cool by trusting the models
There's nothing like a cold beer on a hot summer day, so it's no wonder beer is the world's third most popular drink behind water and tea. At Deschutes Brewery, their mission is to offer a plethora of delicious brews to keep consumers satiated. It's easy to pull a Deschutes beer (or two!) from the fridge, pop the top, and enjoy, but before it reaches a fridge, it's been through a rigorous quality control process at the company's headquarters in Bend, Oregon. This includes monitoring diacetyl levels in the beer as well as a rapid cooling process using glycol chillers to achieve product stability during filtration and packaging.
Unfortunately, Deschutes' backup glycol chiller was on the verge of retirement. Faced with a $750,000 price tag to install a new one, Deschutes turned to the combination of the PI System and machine learning models to predict diacetyl levels, beer cooling cycles and reduce demand for the backup chiller-all while working to win the trust of the operations team.
As part of the brewing process, beer is poured into fermentation tanks where it will ferment for approximately 18 hours. During fermentation, beer reaches temperatures of 60-65 degrees Fahrenheit. As sugars are consumed, the yeast produces enzymes to reduce compounds, including diacetyl. When diacetyl levels move below a certain threshold, they turn on the glycol chiller jackets, seal the tank, and increase the pressure to crash cool the vessel to approximately 32-30 degrees Fahrenheit.
The primary chiller has plenty of capacity for the average load, but at peak load times the antiquated rotary backup chiller is needed to crash cool multiple fermentation vessels at once. To prevent this, they stagger fermentation cycles. However, quality technicians take manual samples for the batches in the morning, resulting in multiple tanks crossing the diacetyl threshold at the same time, demand for the chillers to spike, and the backup chiller to kick on.
Deschutes already uses the PI System to monitor its production process and predict fermentation rates, so it was natural to leverage PI System data to predict and optimize cooling cycles. Through a closed loop analysis, Deschutes pushed PI System data to Azure Machine Learning in the cloud to run models that would predict diacetyl levels and cooling times. Once the models were trained and working, that information was brought back into the PI System so operators could visualize the cooling curves in PI Vision.
“The key to this project was developing a way that operators could interact with the both the algorithm as well as our control system to initiate an automated start of cooling for the respective fermentation vessel,” said Kyle Kotiach, Senior Data Analyst at Deschutes Brewery during the 2020 PI World digital conference.
The team used inductive automation to bridge the gap between the predictive outputs that are being stored in the PI System and the Delta V control system, enabling them to stagger cooling times until there was enough capacity available in the primary chiller. Operators even had the option to check a box to use the predictive models to automatically cool the tanks.
While the solution seemed simple and effective to those who trained the models and understood the data, the operations team felt differently. They questioned the stability of the program and wanted to know more about the data that underpinned the predictive models. The data team needed to earn their trust-and that wouldn't happen overnight. “It was really important for us to explain that what the algorithm was doing was using much more data to predict than a human could possibly use to judge,” Kotiach said.
In response, Kotiach and his team worked diligently with the brewing staff to understand their questions and concerns. From there, they developed a plan to integrate their feedback into the process to ensure they felt comfortable enough to check the box to use the predictive models.
As a result of their efforts to get the entire team on board, Deschutes has optimized its cooling process while ensuring product quality. The backup chiller has not been used for production load since July 2, 2019 and the number of tanks waiting to cool has dropped by 90 percent. The diacetyl predictions have been used 38 times but have saved 273 hours of fermentation. In addition, the company has been able to put off the $750,000 backup chiller project for all of 2019 and 2020.
To learn more about how Deschutes won the trust of the team to successfully predict and optimize cooling times, watch the full 2020 PI World online presentation here.