Transformer-based deep reinforcement learning and small language models for production decision-making in remanufacturing systems

Remanufacturing systems face unique challenges in production decision-making due to the highly non-standard and uncertain nature of end-of-life (EOL) products, including variable arrivals, rework loops, machine breakdowns, stochastic processing times, and seasonal demand fluctuations. Process planning must adapt to diverse product conditions, while scheduling must operate in dynamic environments with these uncertainties. Traditional physics-based or expert-driven approaches often struggle to handle such complexity, leading to inefficiencies and inconsistent decision quality. This study proposes an integrated decision-making framework that combines small language model (SLM)-based process planning with Transformer-based deep reinforcement learning (DRL) scheduling. The SLM is trained on technical documents and maintenance logs to automatically generate adaptive process plans, reducing hardware requirements and minimising human bias. Building on these plans, the Transformer-based DRL algorithm learns robust scheduling policies capable of handling both static and dynamic shop floor conditions, including machine breakdowns and rework loops modelled via stochastic rewards. Experimental evaluations demonstrate that the proposed framework achieves higher decision quality and faster computation compared to priority dispatching rules, conventional fuzzy methods, and baseline DRL models. The results highlight the potential of integrating compact language models with advanced DRL architectures to enhance decision-making in complex remanufacturing environments.