Development of an Automation Framework for Data-Driven Modeling and Adaptive Learning in Industrial Processes

Data-driven modeling and adaptive learning are increasingly used in industrial cyber-physical systems to support monitoring, prediction, and process optimization. In practice, recurring cycles of data acquisition, data preparation, model updating, and feedback integration are frequently realized with manual handovers. As a result, traceability and reproducibility can be impaired, and the reliability of model-based decisions can be reduced. In this paper, an end-to-end automation architecture and an executable workflow are presented that operationalize the full pipeline from industrial data acquisition and automated data handling to Bayesian surrogate modeling and safety-aware feedback execution. The proposed approach is designed to reduce human-induced errors, to increase transparency of the end-to-end traceability from data to decisions, and to accelerate iteration cycles in data-driven development. The framework is demonstrated in an application-oriented high-throughput chemical setting, where non-stationary pH behavior is learned with Bayesian active learning using a non-stationary Gaussian process surrogate. Automated experiment proposals are generated, validated under explicit safety checks, and returned to the production environment to enable efficient model improvement and optimization.