Towards AI Lifecycle Management in Manufacturing Using the Asset Administration Shell (AAS)

Driven by the digital transformation, manufacturing companies face the challenge of managing, but more importantly, enabling rapid operationalization of AI to achieve the full advantage of the exponential data growth. Heterogeneous data structures, the continuously growing variety of implementation frameworks, and the lack of standards for the semantic description of AI solution components, the effort required to manage and share datasets and models between stakeholders impedes efficient and reproducible progression. This paper addresses the current challenges in industrial AI applications currently hindering their acceptance and widespread adoption in manufacturing. Based on an overview of the AI application lifecycle, we present our approach for AI asset meta-data management utilizing the technical concept of the Asset Administration Shell (AAS). Following the definition of the AAS as a reference implementation of the digital twin that provides a digital representation of physical assets and their properties, we propose an AAS for AI assets. The AI AAS maps relevant properties of an AI model together with properties of the corresponding dataset and learning algorithm in order to integrate the AI lifecycle in the Industry4.0 ecosphere.