Increasingly flexible and efficient industrial processes and automation systems are developed by integrating computational systems and physical processes, thereby forming large heterogeneous systems of cyber-physical systems. Such systems depend on particular data models and payload formats for communication, and making different entities interoperable is a challenging problem that drives the engineering costs and time to deployment. Interoperability is typically established and maintained manually using domain knowledge and tools for processing and visualization of symbolic metadata, which limits the scalability of the present approach. The vision of next generation automation frameworks, like the Arrowhead Framework, is to provide autonomous interoperability solutions. In this thesis the problem to automatically establish interoperability between cyber-physical systems is reviewed and formulated as a mathematical optimisation problem, where symbolic metadata and message payloads are combined with machine learning methods to enable message translation and improve system of systems utility. An autoencoder based implementation of the model is investigated and simulation results for a heating and ventilation system are presented, where messages are partially translated correctly by semantic interpolation and generalisation of the latent representations. A maximum translation accuracy of 49% is obtained using this unsupervised learning approach. Further work is required to improve the translation accuracy, in particular by further exploiting metadata in the model architecture and autoencoder training protocol, and by considering more advanced regularization methods and utility optimization.

Industrial automation is transforming by ongoing digitalization efforts to create a flexible industrial internet of things, turning manufacturing facilities into large-scale systems of cyber-physical systems. This development requires addressing the challenging issue of making heterogeneous systems, data models, and standards interoperable, a core problem in designing sustainable service-oriented automation frameworks. This thesis reviews the problem of automatically establishing the interoperability of services exchanging heterogeneous message data. A machine-learning architecture is developed, where the optimization of message transcoders and the system of systems utility are mechanisms for establishing interoperability. By optimizing the transcoders using both service- and metadata, the aim is to ground the learned latent representations in the physical environment to improve generalization. Two physical simulation experiments were performed to investigate and evaluate the architecture by generating heterogeneous JSON messages from multiple heating and air conditioning services. The first experiment focuses on unsupervised learning via back-translation for transcoding engineered features of service messages, reaching a maximum translation accuracy of 49%. The second experiment focuses on supervised learning and a modular neural network (JSON2Vec) for automated encoding of the heterogeneous JSON messages, which enables correct message interpretation in terms of the expected system of systems behavior. These results suggest that machine learning is a viable direction in interoperability automation research which can benefit from both symbolic metadata and message data for reliable generalization and adaptation. Appropriate open datasets are required to consolidate the envisioned development and the migration of solutions to automation systems.