MARTIN GmbH explored innovative methods to capture and refine complex processes, thereby elevating the efficiency of its thermal facilities. In partnership with Zühlke, the global leader in thermal waste recovery solutions aimed to automatically detect and correct deviations from the optimal flame positioning. The strategy involved the use of a computer vision-based soft sensor and AI-powered software to automatically identify and optimise the flame's position during combustion in real-time. 

The collaboration was founded on a prototype developed by MARTIN, which already provided basic combustion data based on colour and brightness in the burning bed. The project team then faced the challenge of refining this approach to enable precise flame position detection using real-time video imagery. This detection identified deviations and subsequently optimised control to ensure more effective combustion. 

The interdisciplinary team from MARTIN and Zühlke employed an agile methodology, enabling continuous coordination and flexibility. The initial process step involved setting up a training and testing environment in the AWS Cloud. The proven and adaptable AWS infrastructure not only offered MARTIN a scalable and resource-efficient setting but also facilitated later integration and scaling. 

Our team, together with a specialised service provider, was responsible for segmenting selected portions of the imagery and programming the core application. The team developed a robust model trained with augmented images, including cropped, rotated, shifted, and mirrored versions, to prepare the application for various camera situations. This approach ensured the high robustness and reliability of the position detection, even in complex environments. 

The application was deployed using Docker images with docker compose scripts, enabling flexible integration into existing systems and easy scaling and maintenance. The training and testing environment was handed over to MARTIN as "Infrastructure as Code," ensuring easy implementation and a transparent, reproducible setting. The project's success was complemented by comprehensive technical documentation, allowing the team not only to meet technical requirements but also to provide a sustainable and well-documented solution for MARTIN's future challenges. 

The project not only showcased technological innovation through the use of computer vision but also demonstrated how precise control and regulation in complex thermal processes could lead to more efficient resource utilisation. The accurate real-time detection of flame positions quickly identified deviations, optimising combustion control. This resulted in safer and more stable thermal treatment and reduced unburned components in the mineral residues from combustion. 

The swift integration of the solution into ongoing operations allowed MARTIN to immediately benefit from the enhanced position detection. Furthermore, the project was already focused on future developments. Planned extensions to detect additional characteristics, especially for identifying anomalies in the burning bed, continued to provide MARTIN with a competitive edge through innovation. Thus, MARTIN GmbH not only achieved immediate benefits in efficient thermal utilisation but also established a foundation for future innovations and adaptability to new challenges.