Vision-Based Machine Learning voor Chemische Procescontrole: Een Haalbaarheidsstudie over Meerdere Toepassingen
This research is motivated by the increasing complexity of chemical processes and the growing demand for robust, non-invasive monitoring and control solutions in the context of the ongoing digitalization of the chemical industry. Many critical aspects of chemical processes, such as phase behavior, flow patterns, and fouling, are inherently visual and therefore difficult to capture using traditional point measurements. By leveraging computer vision as a sensing modality, this thesis addresses a critical gap between observable process behavior and advanced, data-driven process control. The proposed methodology aims to support the transition towards more autonomous and intelligent operation by emphasizing robustness, interpretability, and seamless integration with existing industrial control frameworks, thereby facilitating the practical adoption of vision-based process control in the chemical sector. The research is structured around three primary objectives. The first objective focuses on the development of machine vision methods for classification and segmentation in chemical production environments characterized by continuous operation and high intrinsic safety standards. As a result, truly abnormal or failure-related events occur rarely, leading to highly imbalanced and limited datasets that pose significant challenges for conventional supervised learning approaches. Moreover, the creation of labeled datasets in the chemical sector is often prohibitively expensive, as it requires the involvement of domain experts to accurately interpret and annotate complex process phenomena. To address these constraints, two novel methods are proposed. The first method employs generative adversarial networks for anomaly detection, incorporating tailored cost functions and the structural similarity index to enable automated segmentation. This approach outperforms conventional supervised segmentation models trained for task-specific detection problems. The second method combines original and synthetically generated data to optimize classifier performance while quantitatively assessing generalization through an explainable AI framework. This strategy demonstrates superior performance compared to standard data augmentation techniques, increasing classification accuracy for the chemical foam classification task from 57% to 91%. The second objective focuses on developing a vision-based closed-loop control strategy for process management in the chemical sector. A laboratory setup simulating a chemical foaming production process, in which foam is continuously generated and a vision-based dosing system has been implemented to control foam volume, was established to develop a strategy capable of maintaining effectiveness even when precise control input accuracy cannot be guaranteed. The results demonstrate the feasibility of vision-based process control for automated anti-foaming agent dosing. Furthermore, a sensitivity analysis was conducted to evaluate the impact of the detection system's performance on the control solution. The analysis revealed that the precision of the detector has a limited effect on the system's ability to mitigate foam formation, whereas recall plays a more critical role: if recall dropped below 50%, the system was no longer able to effectively combat foam formation. The thesis concludes with the introduction of a generic framework for the development and deployment of machine vision-based control applications in chemical settings. This framework provides guidance on addressing data acquisition challenges, selecting suitable models, and integrating them into live production environments for automated decision-making. The framework was validated through four use cases at BASF Antwerpen, showcasing its applicability in real-world environments and gave way for several cost savings. The contributions of this thesis significantly advanced the field of vision-based process control, particularly within BASF Antwerpen. The findings emphasize the importance of robust model development, effective use of synthetic data, and the integration of machine vision systems into closed-loop control processes. These advancements offer tangible benefits for automation, efficiency, and cost reduction in chemical production environments and has been applied in four different use cases at BASF Antwerpen.