Modern manufacturing systems face increasingly complex, uncertain, and dynamic demands, making the achievement of system-level autonomy a critical challenge for competitive and resilient operations. A fundamental barrier to this goal is the inherent h...
Modern manufacturing systems face increasingly complex, uncertain, and dynamic demands, making the achievement of system-level autonomy a critical challenge for competitive and resilient operations. A fundamental barrier to this goal is the inherent hierarchical and distributed structure of manufacturing control, which spans from high-level planning to low-level machine execution. Existing approaches that often rely on fragmented automation or decoupled optimization, fail to manage this complexity. This results in localized, suboptimal solutions that are incapable of system-wide coordination and ignore the critical hierarchical context, leading to poor adaptability and resilience.
This dissertation proposes an integrated framework to overcome this limitation by introducing the Autonomous Digital Twin (ADT). First, the ADT is formally defined from a function-centric perspective, possessing an autonomous regulation cycle and seven core functional capabilities. Second, this dissertation proposes a distributed autonomous control framework that reinterprets the ANSI/ISA-95 standard by mapping intelligent decision-making units, termed DT-agents, to each functional layer. Third, a Context-Aware Hierarchical Multi-Agent Reinforcement Learning (CAH-MARL) architecture has been developed to enable cooperative decision-making within this structure. The core of this architecture is a context propagation mechanism, where the strategic intent of higher-level agents (e.g., a production schedule) is explicitly encoded as context to guide the tactical execution of lower-level agents (e.g., AGV routing).
The framework’s efficacy was empirically validated through two complementary, complex simulation case studies; a two-level vertical co-optimization system and a three-level vertical-horizontal integrated optimization system. Experimental results demonstrated that the proposed ADT framework achieved superior operational effectiveness and efficiency in stable scenarios compared to the conventional decoupled optimization approach. In dynamic disruption scenarios, such as demand shifts or resource failures, it showed remarkable adaptability and resilience, characterized by significantly smaller performance drops and substantially faster recovery times. These results were driven by a high rate of goal achievement across hierarchical layers, confirming that the CAH-MARL context propagation mechanism within the ADT framework successfully orchestrated system-wide coordination.
Consequently, this research provides a novel paradigm that leverages the inherent hierarchical structure of manufacturing systems not as a constraint, but as a mechanism to achieve scalable, system-level autonomy.