Abstract
The advancement of smart manufacturing, driven by the Industry 4.0 paradigm and the emerging vision of Industry 5.0, has fundamentally transformed the architecture of modern production systems. At the center of this transformation lies the development of Flexible Smart Manufacturing Systems (FSMS)—reconfigurable, data-rich, and highly automated platforms designed to respond rapidly to volatile market demands. While FSMS promise enhanced productivity, adaptability, and resilience, their realization is impeded by persistent gaps in real-time system modeling, performance evaluation, and integrated decision-making under uncertainty.
This dissertation presents a comprehensive modeling–analysis–control framework aimed at improving the responsiveness, efficiency, and intelligence of FSMS. Grounded in physics-informed modeling and data-enabled learning, the framework addresses critical limitations in existing methodologies by integrating system knowledge, real-time data streams, and advanced machine learning techniques to solve representative prediction and control problems across different FSMS architectures.
Level 1: Data-Enabled Modeling, Analysis, and Control of Variable Cycle Time FSMS
A data-enabled mathematical model is developed for FSMS with variable machine cycle times, treating disruption events as stochastic inputs. This model enables real-time system performance evaluation and introduces two interpretable indicators—Permanent Production Loss (PPL) and Opportunity Window (OW)—which facilitate causal attribution and predictive monitoring. A hybrid control strategy is designed by combining OW-based feedback logic with deep reinforcement learning (DRL) to optimize machine on/off decisions and cycle time adjustments.
Level 2: Integrated Process–System Modeling, Analysis, and Multi-Agent Control
While Level 1 focuses on system throughput, it overlooks process-level attributes such as quality. To incorporate product quality alongside productivity, the framework is extended to an integrated process–system environment. A recursive model captures quality and system-level interactions, supporting real-time evaluation and PPL attribution. Building on this, a multi-agent reinforcement learning (MARL) control framework is formulated, including two novel algorithms that embed system properties into credit assignment and policy learning to enhance coordination and robustness.
Level 3: Modeling, Planning, and Generative Control in Robot-Operated FSMS
At the highest level of complexity, FSMS are equipped with multi-skilled mobile robots, introducing dynamic reconfiguration and market-driven planning needs. A system identification method is used to extract key dynamic features. These are used within a Hierarchical Integrated Planning–Scheduling (H-IPS) framework, which integrates long-horizon planning with real-time robot assignment. To overcome structural constraints, a hybrid graph–diffusion planning framework is proposed. By combining a Heterogeneous Graph Neural Network (HGNN) with a diffusion model, system configurations are generated directly from demand inputs and refined through Ideal Clean Configuration (ICC) principles.
This dissertation establishes a coherent and extensible architecture for real-time intelligent decision-making in FSMS. By unifying modeling, analysis, and control across different layers of system complexity, the proposed framework enables scalable, interpretable, and adaptive solutions for the design and operation of next-generation manufacturing systems.
