Combining Digital Twins and Game Theory for Precomputing Energy-Aware Reconfiguration Strategies for Cyber-Physical Production Systems (TwinGears)

• Applicants: Professor Dr.-Ing. Alexander Fay; Professor Dr. Malte Lochau
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 562026346

Modern cyper-physical productions systems (CPPS) must be able to dynamically adapt themselves to ever-changing contextual requirements. These requirements include functional capabilities in terms of system skills as well as non-functional self-optimization goals like reduction of latency and energy consumption. The required sequences of self-reconfiguration decisions are hardly computable for every context change on demand at run-time due to the high computational complexity of the underlying multi-objective constrained optimization problem. In contrast, precomputing all self-reconfiguration decisions potentially emerging at run-time already at design time to be able to utilize theoretically unlimited amounts of computational resources is also not feasible due to the exponentially growing search space in the number of configuration options and the inherent uncertainty about contextual changes. The goal of the proposed project is to counter-act this dilemma by a new software development methodology to enable CPPS to efficiently reason about effective context-dependent self-reconfiguration decisions for energy-aware self-optimization. We employ the digital twin paradigm to encapsulate our model-based approach for automating the computation of self-reconfiguration decisions into a separated component interacting with the physical part of the CPPS. The knowledge required for effective decision-making is captured in an integrated self-reconfiguration model combining a CSS-based ontology model for relevant domain knowledge and a context-feature model formalizing validity constraints of reconfiguration steps. This model facilitates staged decision-making by combining offline precomputation and online completion of context-dependent reconfiguration steps. For the offline part, we extract from the knowledge model two separate behavioral models, one describing valid system reconfigurations and one for the changes of contextual requirements, both represented as probabilistic timed energy game-automata. This representation allows us to apply algorithms from game theory to synthesize winning strategies for the system player defining for each context move the presumably most effective reconfiguration step with respect to functional and non-functional requirements. The probabilistic parts of the model denote contextual uncertainty which not only enables reactive reconfiguration moves for observed contextual changes, but also proactive reconfiguration moves for predicted future contextual changes. In the online part, we map the precomputed abstract decisions to corresponding system capabilities to operationalize self-reconfiguration decisions at skill level by completing missing context knowledge with run-time observations. Finally, we close the self-adaptive feedback loop by applying context learning techniques to consecutively reduce uncertainty in the knowledge model and to define criteria for refining and/or recomputing reconfiguration decisions.

• DFG Programme: Research Grants