Coca-Cola Develops Virtual Pressure Sensor with Machine Learning to Improve Beverage Dispenser Diagnostics

With the help of MathWorks, the team was able to reduce the footprint of this code so that it will fit nicely in the ARM-Cortex M microprocessor. It has transformed the flow control module into a smart component.

Key Outcomes

  • Transformed a standard flow control module into a diagnostics-capable smart component
  • Eliminated the need to retrofit thousands of existing dispensers with costly sensors
  • Achieved up to 91% accuracy in pressure predictions

Coca-Cola Freestyle beverage dispensers allow consumers to select from hundreds of different beverages via a touchscreen interface. A key component of the dispenser is its flow control module (FCM), which includes a solenoid actuated valve to regulate water flow. With no physical pressure sensor in the machine’s water line, field technicians were unable to distinguish between an FCM failure and an upstream pressure loss, leading to unnecessary FCM replacements.

To improve field diagnostics, Coca-Cola engineers used MATLAB® and Simulink® to develop a machine learning–based virtual sensor and deploy it to the dispenser’s resource-constrained microprocessor.

The team began by collecting data from FCMs via a hardware-in-the-loop testing process. They used Simulink to model and generate code for a simple controller, which they then downloaded to a dispenser control board used to capture valve pressure and electric current measurements. Working in MATLAB, they then developed functions for performing feature extraction and multivariable regression based on the collected data. They incorporated these functions into a Simulink model that predicted valve pressure based on valve current. Working with MathWorks engineers, the team reduced the footprint of the model before generating code for the FCM’s ARM Cortex-M microprocessor.

Before deploying the virtual sensor in the field, the team verified its accuracy by collecting and analyzing data from more than 3,000 tests on 10 different FCMs with two different control boards.