Open Access

Philipp Mascha

• In recent years, machine learning (ML) has steadily found its way into everyday life. Countless applications—such as image filters, translation services, and chatbots—have become so commonplace that they often go unnoticed. In high‑throughput manufacturing, however, ML applications frequently fail to progress beyond the proof‑of‑concept (PoC) stage into continuous live operation. This is particularly evident in visual quality control, where potential is abundant and performance promising, yet concerns regarding viability, stability, and trust hinder full deployment. This thesis outlines the domain of manufacturing and its unique requirements, demonstrating how these constraints affect ML performance, especially with respect to long‑term stability. A practical framework, Visual Machine Learning Control (VMLC), is subsequently introduced to address these challenges through an observer–controller architecture. The framework incorporates an autoencoder‑based metric that measures theIn recent years, machine learning (ML) has steadily found its way into everyday life. Countless applications—such as image filters, translation services, and chatbots—have become so commonplace that they often go unnoticed. In high‑throughput manufacturing, however, ML applications frequently fail to progress beyond the proof‑of‑concept (PoC) stage into continuous live operation. This is particularly evident in visual quality control, where potential is abundant and performance promising, yet concerns regarding viability, stability, and trust hinder full deployment. This thesis outlines the domain of manufacturing and its unique requirements, demonstrating how these constraints affect ML performance, especially with respect to long‑term stability. A practical framework, Visual Machine Learning Control (VMLC), is subsequently introduced to address these challenges through an observer–controller architecture. The framework incorporates an autoencoder‑based metric that measures the dissimilarity between current measurements and the training data underlying the classification model. By applying statistical process control (SPC) to this metric, control limits can be established to ensure stability and enable responsive intervention. To address violations of these limits, several self‑healing methods are presented. These methods restore system performance through active learning, sensor self‑calibration, and improved explanations of faulty classifications. The framework is thoroughly evaluated using real‑world manufacturing data, and its viability for live operation is demonstrated through a case study.…