Advances in industrial machine learning applications

Industry is adopting machine learning (ML) technologies for smart manufacturing at a rapid pace. However, the requirement for manual validation and a lack of trust in the predictions of ML models in safety-critical applications limit widespread adoption and scalability. The integration of domain knowledge is a key driver in the development of a fully automated end-to-end machine learning lifecycle management system for smart manufacturing. Such a system aims to facilitate the training and deployment of trustworthy, functionally safe models, while reducing human effort. Domain knowledge is encoded as constraints on the function shape of predictive models (e.g., positivity, monotonicity, convexity) and thus ensures that the resulting models adhere to relevant physical laws and fulfill safety requirements. Such knowledge is readily available in physical systems and thus equally applicable in the industrial context. This work describes an automated end-to-end machine learning lifecycle management system. In this system, domain knowledge is (i) automatically extracted from data, (ii) applied to automatically validate sensor data without the need for manual labels or established baselines, and (iii) integrated into the ML training pipeline to ensure that trained models adhere to expected behavior. The presented approaches are evaluated on synthetic benchmarks, implemented in a prototypical system, and validated on industrial case studies from smart manufacturing. The design and implementation of the system are described in detail in this thesis. Together, these contributions demonstrate that knowledge integration is a key enabler for fully automated, functionally safe ML in industrial applications. Machine learning should be an easy-to-use technology for domain experts. This thesis brings this goal one step closer, by integrating and adhering to the knowledge of domain experts.