彈性零工式排程問題（Flexible Job Shop Scheduling Problem， FJSP）須同時決定工序加工順序與機台指派，屬於高度複雜之組合最佳化問題。本研究以最小化最大完工時間（Makespan）為目標，提出一套結合異質圖神經網路、雙解碼器與自標籤優化機制之生成式智慧排程架構，並比較監督式學習與自標籤訓練策略之效能差異。 本研究首先將 FJSP 建構為包含工序節點與機台節點之異質圖，利用圖注意力網路第二版（GATv2）擷取工序先後關係、候選機台彈性及資源競爭等結構資訊。模型採用工序解碼器與機台解碼器所組成之雙解碼器架構，依序完成工作選擇與機台指派，並結合工件進度、候選加工時間、機台就緒時間及預估完工時間等動態狀態特徵，以提升排程決策之合理性與可行性。 在訓練策略方面，監督式學習使用 PyJobShop 約束編程求解器產生之專家解作為標籤；自標籤機制則針對同一實例平行抽樣多組候選解，並選取 Makespan 最小者作為偽標籤，反覆更新模型參數。實驗結果顯示，自標籤模型在不同規模與不同機台彈性之驗證資料集中，整體解品質均優於監督式學習，並可有效降低自迴歸推論過程中的誤差累積與曝光偏差。 在跨規模測試中，自標籤模型未經重新訓練即可直接應用於 30×10 與 40×10 大型實例，且僅需數秒即可完成單一實例之排程，展現良好的推論效率與規模泛化能力。於 Brandimarte 與 Hurink 公開基準資料集上，自標籤模型亦優於傳統派工法則及比較之深度強化學習方法。 綜合而言，本研究所提出之異質圖雙解碼器可有效處理 FJSP 中工序排序與機台指派之耦合決策，而自標籤優化機制能降低對高成本專家標籤之依賴，並在解品質、推論效率及跨規模泛化能力方面展現良好表現，可作為智慧製造即時排程決策之可行方法。;The Flexible Job Shop Scheduling Problem (FJSP) requires simultaneous decisions on operation sequencing and machine assignment and is therefore a complex combinatorial optimization problem. This study aims to minimize the makespan by proposing a generative scheduling framework that integrates a heterogeneous graph neural network， a dual-decoder architecture， and a self-labeling optimization mechanism. The performance of supervised learning and self-labeling strategies is also compared. The FJSP is represented as a heterogeneous graph containing operation and machine nodes. Graph Attention Network version 2 (GATv2) is used to capture operation precedence， routing flexibility， and resource competition. A job decoder and a machine decoder then sequentially determine job selection and machine assignment. Dynamic information， including job progress， processing times， machine ready times， and estimated completion times， is incorporated to improve scheduling feasibility and decision quality. For supervised learning， expert solutions generated by the PyJobShop constraint programming solver are used as training labels. In contrast， the self-labeling strategy generates multiple candidate schedules for each instance， selects the solution with the smallest makespan as a pseudo-label， and iteratively updates the model. Experimental results show that the self-labeling models consistently achieve better solution quality than the supervised models across datasets with different problem sizes and levels of machine flexibility. The self-labeling strategy also reduces exposure bias and error accumulation during autoregressive inference. In cross-scale testing， the trained self-labeling models can be directly applied to 30×10 and 40×10 instances without retraining and can generate schedules within seconds. The models also outperform conventional dispatching rules and the benchmark deep reinforcement learning method on the Brandimarte and Hurink datasets. Overall， the proposed framework effectively addresses the coupled decisions of operation sequencing and machine assignment while reducing dependence on costly expert labels. It also demonstrates favorable solution quality， inference efficiency， and cross-scale generalization， indicating its potential for real-time scheduling in smart manufacturing.
