Profit-oriented dynamic scheduling for disassembly flexible job shops: dealing with structural uncertainty and operation skipping

Disassembly scheduling for End-of-Life (EOL) products, particularly EV batteries, exhibits dynamic structural uncertainty and physical degradation characteristics. However, existing scheduling models are predominantly static and fail to fully capture the distinctive stochastic features and operation skipping patterns of such disassembly systems. To address these gaps, we formulate the process as a Markov Decision Process and propose a Hierarchical Multi-Agent Deep Reinforcement Learning (H-MADRL) framework. To enhance practical applicability, a Hierarchical State Embedding module acts as a digital perception layer to quantify physical risks. Meanwhile, a Physics-Informed Dual-Agent mechanism coordinates task-machine assignments, and a Constraint-Compliant Masking layer ensures strict adherence to Standard Operating Procedures. Averaged over multiple independent runs, the framework achieves a 45.6% increase in profit rate and a 16.9% reduction in makespan compared to standard reinforcement learning and heuristic baselines. Ultimately, this approach secures shop-floor adaptability and significantly advances managerial precision in balancing economic density with operational throughput under stochastic constraints.