Dynamic Deployment of Fault Detection Models

This paper describes an approach for integrating a Machine Learning (ML) model for fault detection into an Asset Administration Shell (AAS), focusing on the seamless deployment of updated versions of this model during operation without causing downtime in production processes. To this end, a generic submodel was developed that conforms to the AAS metamodel definition and can support different types of ML models and application scenarios. In addition, three deployment strategies were identified that allow switching between model versions without downtime. Evolved concepts were evaluated with a proof-of-concept to illustrate the applicability and suitability of the approach. The evaluation shows that the most appropriate deployment strategy depends on the specific use case.