Industry 4.0 is transforming traditional production systems into dynamic and adaptive value networks. However, this shift reveals significant challenges in achieving seamless interoperability, workflow optimization, and effective value chain analysis within complex production environments. This thesis contributes to addressing these challenges through a structured approach, supporting data-driven and adaptive decision-making to optimize operations in dynamic production value networks.

The research commences by architecting microservice-oriented systems that facilitate the integration of legacy and brownfield technologies with Industry 4.0-compliant environments. By leveraging the Eclipse Arrowhead framework, the thesis demonstrates how diverse systems can exchange data and collaborate at runtime, establishing the foundation for cohesive and interoperable production networks.

Building on this interoperable structure, the thesis explores AI-driven optimizations across key areas, including workflow optimization, predictive maintenance, and demand response. These approaches support operational efficiency and adaptability in production value networks. Case studies showcase collaborative learning models for condition monitoring and an edge-based framework to optimize energy use, demonstrating tangible improvements in efficiency and resilience.

Finally, a significant contribution of this thesis is the introduction of tools for visualization and analysis of value chains. Using the Reference Architectural Model for Industry 4.0 (RAMI 4.0), the research provides methods to map and evaluate value creation within dynamic production networks. By integrating activity-based costing with microservice architectures, it offers granular insights into cost and value dynamics at runtime, enabling agile and informed decision-making in complex industrial environments.

Through these contributions, the thesis advances the understanding and implementation of data-driven optimizations in production value networks, supporting agility and sustainability while contributing to the transition to circular business models by enabling value chain analysis across the product lifecycle. The thesis serves as a bridge between theoretical frameworks and practical applications, providing valuable insights for both academia and industry, and paving the way for more efficient and sustainable production ecosystems.

The growth in cyber-physical systems (CPS), the industrial internet of things (IIoT) and integrations of machine learning (ML) models have enabled Industry 4.0 automation and intelligence in industrial System of Systems (SoS). However, scalable automation frameworks, dynamic interoperability for smooth communication among heterogeneous systems, and the integration of ML models particularly for online collaborative intelligence in distributed architectures, are open challenges that must be addressed. This thesis structures its research as a progressive investigation, with each identified challenge leading to the next. First, local cloud-based automation is explored using the Eclipse Arrowhead framework to propose digitalization frameworks for industrial use cases, such as predictive maintenance in wind energy systems and smart manufacturing. The primary objective is to bridge the digital divide in SoSs environments, ensuring seamless intercommunication between IoT-connected devices. Significant engineering effort is dedicated to developing dynamically scalable automation solutions that integrate heterogeneous CPS, providing real-time adaptability and efficiency. This leads to the second research challenge addressed in this thesis. To investigate the challenge of semantic interoperability among heterogeneous IIoT devices, this research explores ontology alignment techniques through Natural Language Processing (NLP) and deep learning models. Extension of an existing language model (BERT_Intereaction) is proposed for ontology graphs to facilitate seamless communication between heterogeneous IIoT devices. It is designed using a language encoder to develop knowledge of the text to understand the labels or entities and a structural encoder to understand the context or semantics behind the text. This proposed model consistently outperforms cross-lingual tasks over the state-of-theart techniques with an error reduction of 2.1% on benchmark datasets DBP15KZH−EN, DBP15KJA−EN, and DBP15KFR−EN.

The third challenge involves scaling collaborative intelligence across distributed IioT systems. To address this, a local cloud-based collaborative learning (CCL) model is designed for the service-oriented architecture (SOA) and a decentralized ML model to digitalize IIoTs while enabling ML-based optimizations for IIoT tasks across cloud and edge nodes. The CCL model integrates machine learning-as-a-service (MLaaS) into the distributed cloud architectures. CCL offers scalable, privacy-driven, self-contained local

clouds for every CPS in the system of systems model. The local clouds enable distributed ML deployment across the IIoT SoS, where devices collaborate to share their knowledge representations. The model uses unsupervised dictionary learning, allowing IIoT nodes to share compressed, optimized learning representations. Furthermore, this thesis also highlights a culminating issue for designing decentralized ML-enabled IIoT solutions that is the information overload and redundancy at the edge and cloud. To mitigate this challenge, the CCL+ model is proposed, integrating coherence-based dictionary refinement

with Bayesian optimization. The model is tested on the condition monitoring task using data from an automated farm of six wind turbines using the CCL model. Implementing redundancy-aware strategies in the CCL+ optimized bandwidth usage and reduced communication overhead, especially for the resource-constrained IIoT devices. In the simulation experiments, over one year, the propagated learned dictionary size at a single wind turbine exceeded 1 petabyte. In contrast, in comparison, using the CCL+ model, for the same duration, the learned dictionary remained at 18 MB, significantly enhancing communication and computational efficiency without losing essential information.

Various potential future research directions accompany the findings presented in this thesis. For instance, to strengthen the ablation study on the semantic interoperability among heterogeneous SoS challenge, a generalized IIoT ontology that is designed for any IoT device (beyond sensors), such as the smart applications reference ontology (SAREF) can be tested for ontology alignment. This work provides a step towards enabling translation between heterogeneous IoT sensor devices. The proposed BERT Intereaction model can be further extended to a translation module using the generalized ontology graphs. Then investigations can be conducted to test if the model can interpret the messages transmitted across ontologically different devices in two scenarios: a) where the ontology graphs of both devices are a subset of the generalized ontology graph, and b) they are overlapping graphs and may contain different nodes and relations but they are semantically the same. Furthermore, exploring the utility of CCL+ model for extended

large-scale SoS with multiple parallel tasks to test the collaborative learning concept across the heterogeneous cyber-physical system of systems (CPSoS).