The development of predictive maintenance (PdM) solutions is one of the key challenges in the industry today. Manufacturing processes are usually well described by the law of physics and mathematical equations, but the irregularity and randomness of the asset degradation process make it a demanding task to model it. This makes physics-driven models insufficient for this kind of problem. On the other hand, data-driven models, mainly Artificial Intelligence (AI), are gaining much interest in research and applications due to their flexibility and robustness. A compromise between these two approaches are hybrid models that take into account the physics of the process and use modern AI methods to learn its behavior. The next challenge for AI models is to provide information on their reasoning to build understading and trustworthiness, which can be achieved through post-hoc Explainable AI (XAI) methods. In this paper, we use a Physics-Informed Autoencoder (PIAE) in a semi-supervised manner to learn the degradation process of work rolls in the cold- rolling process. We incorporate physics knowledge into the AI model by extending its input space and applying feature masking during the prediction phase. The results of the research show that such an architecture is capable of distinguising between low- and high-wear observations. Furthermore, we include the XAI layer in the model, which gives explanations for the prediction of the model through counterfactuals.

Anomaly detection is emerging trend in manufacturing processes and may be considered as part of the Industry 4.0 revolution. It can serve both as diagnostic tool in predictive maintenance task, as well as trace back mechanism for assessing quality of production or services. In this paper we describe and approach for explainable anomaly detection in industrial data which contains sequential and static features. We based our solution on modified autoencoder architecture with Long Short-Term Memory layers. To address a problem of explinability in deep learning and find origin of the anomalies we have engaged the SHAP method, which gives both local and global explanations of the model. Analysis of SHAP explanations allowed us to determine the source of majority of anomalies detected by deep learning model. We demonstrated the feasibility of our approach on synthetic, reproducible dataset and on real-life data gathered from hot rolling industrial process.