Krones Applies CI/CD to AI-Based Process Control Software
MATLAB and Simulink Workflows Support Repeatable Building, Testing, and Packaging for Customer-Specific AI Controllers
“You get consistency because the same tests always run, and traceability has improved considerably with continuous integration.”
Key Outcomes
- Brought selected AI controller development steps into automated build and test workflows
- Standardized container creation and delivery steps, reducing dependence on local environments and manual handoffs
- Improved traceability and release readiness for customer-specific AI control software updates
Contiloop AI uses in-line light transmission measurements to monitor container quality and readjust heating and blow-molding parameters when needed.
For machinery OEMs developing customer-specific AI control software, release readiness can depend as much on repeatable testing, packaging, and delivery as on control algorithm development. At Krones, engineers addressed that challenge in Contiloop AI, an intelligent process control system for PET stretch blow molding.
Contiloop AI measures container quality in-line using light transmission and readjusts heating and blow-molding parameters when needed. During commissioning, a telemetry run varies process settings and collects bottle-quality measurements, which are sent to the Krones platform for machine learning processing.
Because bottles, preforms, ovens, and stretch blow-molding machines differ from customer to customer, Krones trains a new AI agent for each customer. The broader workflow spans data collection, preprocessing, environment-model training, AI agent training, control-routine generation, Docker® image creation, and deployment of the controller to an edge device.
For the engineering team, the scaling issue was ensuring the repeatability of the software workflow. Manual tests, locally built Docker images, differences among environments and software versions, limited toolbox availability, access permission restrictions, and local resource and network constraints made releases harder to coordinate. Krones integrated CI/CD pipelines into selected parts of the workflow to automate tests, builds, container creation, and delivery steps.
Within those pipelines, Krones uses MATLAB® to train AI models and Simulink® to integrate the trained models into control logic and generate code that has defined interfaces for embedded systems. The team uses different packaging approaches depending on the task: Docker images with MATLAB Runtime for standalone applications, or MATLAB Docker images when the workflow requires the full MATLAB environment and code generation capabilities.
