DEEP Q-LEARNING FOR FIXED-LENGTH CONTROL IN INTEGRATED CONTINUOUS CASTING & ROLLING

This thesis investigates a Deep Reinforcement Learning (DRL)–based control strategy for the fixed-length (cutto-length) problem in integrated continuous casting and rolling (CCRI). The task is formulated as a MarkovDecision Process (MDP), and a Deep Q-Network (DQN) controller is developed to learn both optimal shear pointsand fine-tuned length offsets. The objective is to maximize material utilization and yield while minimizing shearfrequency, energy consumption, and secondary cutting losses. The approach leverages rich process-aware statefeatures—including billet temperature profiles, geometric dimensions, expansion and oxidation losses, andequipment constraints—encoded into a structured state representation.A high-fidelity, disturbance-rich digital twin simulation environment is constructed to mimic real productionscenarios such as billet length deviations, thermal drifts, and shear delays. Within this platform, the DQN agentis trained and evaluated under both nominal and perturbed conditions. Experimental results demonstrate that thelearned policies achieve competitive or superior performance compared to rule-based and optimization-basedbaselines, while maintaining robustness and adaptability. Additional analyses include sensitivity to rewardweighting, evaluation of architectural variants (Double/Dueling DQN, prioritized replay), and deploymentoriented metrics such as inference latency and online responsiveness.The findings indicate that value-based DRL, when combined with careful process modeling and reward shaping,provides a viable route toward robust and adaptive control in steel manufacturing. Beyond immediatecontributions, this thesis also aligns with the supervisor’s 2025–2027 research roadmap, which emphasizesprogressive algorithmic refinement, expansion toward multi-agent and hybrid control, and eventual pilot-scaleindustrial deployment.