Emerging Industry 4.0 architectures deploy datadriven applications and artificial intelligence services across multiple locations under varying ownership, and require specific data protection and privacy considerations to not expose confidential data to third parties. For this reason, federated learning provides a framework for optimizing machine learning models in single manufacturing facilities without requiring access to training data. In this paper, we propose a self-adaptive architecture for federated learning of industrial automation systems. Our approach considers the involved entities on the different levels of abstraction of an industrial ecosystem. To achieve the goal of global model optimization and reduction of communication cycles, each factory internally trains the model in a self-adaptive manner and sends it to the centralized cloud server for global aggregation. We model a multi-assignment optimization problem by dividing the dataset into a number of subsets equal to the number of devices. Each device chooses the right subset to optimize the model at each local iteration. Our initial analysis shows the convergence property of the algorithm on a training dataset with different numbers of factories and devices. Moreover, these results demonstrate higher model accuracy with our self-adaptive architecture than the federated averaging approach for the same number of communication cycles.

With the rise of Industry 4.0, businesses are increasingly turning to Machine Learning to leverage data for improving quality and productivity. However, one open challenge when embracing Machine Learning in this context is the integration of cloud infrastructures, as well as the heterogeneity of data, interfaces, and protocols in the production environment. To address this, we are developing a framework that aims to simplify the adoption of Machine Learning techniques for heterogeneous industrial automation systems. One of the core features of this framework is the ability to handle data about production devices -a scenario that is naturally suited to the use of Asset Administration Shells. However, the implementation of a system that uses Asset Administration Shells comes with its own set of challenges, such as the abstraction of details from users and the representation of device topologies. Thus, this paper introduces the concepts and implementation of a Metadata Manager component in the aforementioned framework that uses Asset Administration Shells as its basis. We further examine the Metadata Manager's current structure with unit testing, derive planned extensions, and discuss future directions from the Industry 4.0 perspective.

No abstract