In industry, data-driven techniques have revolutionized manufacturing by collecting huge amounts of information during production and turning it into valuable information for process optimization. While machine learning (ML) is a key technology and the main contributing factor for many recent success stories, we witness the transition of ML moving from the “virtual world” into “the wild”; this includes prominent applications in autonomous navigation, the Internet of Things, and Industry 4.0 applications. Evidently, this transition opens several real-world challenges for ML that need to be addressed for closing the gap between both worlds.

We focus on data-driven machine condition modeling, process optimization and forecasting. A crucial requirement for the widespread acceptance of ML-based condition modeling is to not only work accurately but to work reliably in every imaginable situation and to provide interpretations and uncertainty measurements of the model behavior. In real-world situations, a manifold of disturbances and environmental influences can occur that need to be accounted for. Particular requirements for real-world system are: first, robustness in the presence of outliers, domain shifts, and corrupted data, second, learning and transferring knowledge from similar problems to counteract the limited availability of labeled data, and third, being aware of the model’s limits; finally, in safety-critical systems, it is equally important to achieve accurate predictions and to understand the behavior of a model.

Simply scaling conventional modeling approaches shows diminishing improvements, which underlines the need for tackling the following three research areas to make significant progress in data-driven condition monitoring:

1. Robust representation: We exploit deep neural networks for learning representations to avoid manual feature engineering. Furthermore, they are used for outlier detection, data augmentation, and semi-supervised learning to counteract limited, unlabeled, and noisy data and to improve the generalization ability of the models.
2. Model learning and uncertainty estimation: We introduce Bayesian models to provide uncertainty estimates for the predictions, we develop methods to conquer domain shifts and we consider transfer learning for knowledge exploitation across related applications.
3. Explainability and process optimization: We will develop explainable AI techniques to understand the root cause for the prediction of the black-box ML models. Furthermore, we consider model adaptation and continual learning over the model exploitation phase.

We aim to focus in this CD Laboratory on data-driven process optimization and condition modelling with applications in refractory modeling, rail vehicle modelling and forecasting. These applications have similar fundamental research challenges in common, where a well-performing surrogate model (i.e. digital twin) using ML techniques is the primary research focus. The developed methods are versatile and can be used in various applications.