應用於動態生產及預防保養規劃的多智能體決策融合框架

Abstract

科技的迅猛進步使得製造業的競爭愈加激烈。為了保持產業領先優勢，製造商在努力滿足個人化需求的同時，還需有效壓縮生產成本以保持強勁獲利，這構成了雙重挑戰。這種微妙的平衡要求製造商具備迅速開發和實施高效的生產規劃(production planning，PP) 方法，以達到最佳且可行的工作分配。此外，製造商務必綜合考慮設備的保養策略，以提高設備性能。

製造環境天然呈現動態特性。需求變動、訂單更改或取消和機台故障及其他不確定事件將使最初的生產計劃不符合生產需要或是無法使用。因此，務必基於現實狀態的變動及時調整生產計劃。然而，由於求解的複雜性，求解動態控製策略是相當困難的，尤其是複雜製造系統。

為了解決求解的挑戰，本研究引入了多智能體決策融合（multi-agent decision fusion，MADF）框架來獲取多工作多機台系統的動態策略。此外，將 MADF 框架與深度神經網路（deep neural network，DNN）相結合，實現即時生成策略來解決頻繁重新排程或更新生產計劃的挑戰。

第 3 章利用馬可夫決策過程（MDP）製定了生產計畫（PP）模型和生產計畫與預防保養（production planning and preventive maintenance，PPPM）模型。求解PP 或 PPPM 問題的複雜度關於機台數量和工作類型數量指數增長。此外，也研究了求解 PP 問題價值疊代算法的單調性。

第 4 章介紹了 MADF 框架以降低求解複雜度並產生整個系統的策略。在MADF 框架中，透過工作分配（job allocation，JA）模型將原問題轉變為多個單機問題。機台健康狀態的平穩分佈可以通過 JA 模型推導出來。在 PP 問題中，系統將處於不穩定狀態一旦實際總到達率超過 JA 模型產生的最大達到率。整個系統的生產控制策略可以由每個被分解機台的最優策略產生。這些最優策略通過有兩個工作的單機系統的最優解進行多數表決（majority vote，MV）來產生分解後單機系統的策略。所有的獨立生產由機台智能體透過即時調度規則進行協調，以提升系統效率。MADF 求解複雜度的增長與機台數量呈線性增長，與工作類型數量呈二次增長。

第 5 章建立深度神經網絡模型。該模型在由一機台和兩個工作的系統的最優解所建構的訓練空間上進行訓練。MADF 和 DNN 的結合（MADF-DNN）實現了即時產生大規模系統的生產決策。

第 6 章通過數值實驗針對 PPPM 問題驗證本研究方法的有效性，並與文獻中常用策略進行比較。對於兩個工作兩個機台系統，就平均完工時間而言，PPPM與 MADF 之間的平均差距在 6% 以內。與求解 PPPM 相比，求解時間平均降低了70.93%。在重負載生產系統，具有十個機台和十個工作類型且無法解決的大型系統中，與其他策略相比，MADF 在平均完工時間和生產量方面都具備顯著優勢。MADF 框架將平均完工時間縮短了 90% 以上。與不考慮機台之間的協調相比，MADF 框架將平均完工時間縮短了 26.89%。

The rapid advancement of technology has brought about fierce market competition for manufacturing firms. In pursuit of industry leadership, manufacturers are confronted with the dual challenge of meeting customized demand while simultaneously reducing production costs to maintain a robust profit. This delicate balance requires the urgent development and implementation of efficient production planning (PP) methodologies, enabling optimal capacity planning. Additionally, manufacturers must prioritize comprehensive maintenance planning strategies to improve equipment performance.

The manufacturing environment is inherently dynamic. Uncertainties, such as fluctu ations in demand, order cancellations, and machine failures will render the initial produc tion planning impracticable or unworkable. Consequently, updating production planning is required in accordance with changes in the actual situation. However, efficiently gener ating the optimal dynamic control policies is difficult due to the computational complexity, especially in large-scale manufacturing systems.

To address the computational challenge, this dissertation proposes a multi-agent de cision fusion (MADF) framework to obtain the dynamic policies in multiple-machine and multiple-job systems. Furthermore, the MADF framework is combined with deep neural network (DNN) to address the challenges of frequent rescheduling or updating of planning by instantly generating policies.

Chapter 3 formulates the PP model and the joint production planning and preventive maintenance (PPPM) model by Markov decision processes (MDP). The computational complexity in addressing PP or PPPM problems experiences exponential growth as the numbers of machines or job types increase. Additionally, the monotonicity properties of the value iteration algorithm in solving PP problems are investigated.

Chapter 4 introduces the MADF framework to reduce the computational complexity and generate policies for the entire system. A large-scale problem is decomposed into sev eral single-machine problems through a job allocation (JA) model. The JA model enables derivation of the stationary distribution of machine health. In PP problems, instability occurs if the actual total arrival rate exceeds the maximal total arrival rate generated by the JA model. The policies of the entire system can be generated by the optimal poli cies of each decomposed machine. These optimal policies are formed through majority vote (MV) among optimal solutions for systems with two jobs. Independent productions are coordinated through a real-time dispatching rule by machine agents to enhance sys tem efficiency. The computational complexity of solving dynamic control problems with MADF grows linearly with machine numbers and quadratically with job-type numbers.

Chapter 5 constructs a DNN model, which is trained on the knowledge space built by optimal policies of systems with one machine and two jobs. The combination of MADF and DNN (MADF-DNN) can instantly generate the policies to a large-scale system.

Chapter 6 evaluates the effectiveness of the proposed MADF framework in PPPM problems through comparisons with other commonly used policies from literatures. For two-job and two-machine systems, the average cycle time has a variance within the range of 6% when comparing PPPM with MADF. The MADF framework can achieve an re duction on computation time 70.93% on average compared to solving PPPM policy. In heavily loaded production systems with ten machines and ten job types, which are unsolv able, MADF shows significant superiority over alternative policies in both average cycle time and throughput. The MADF framework reduces average cycle time by over 90% on average. Compared with no coordination, the MADF framework reduces average cycle time by 26.89% on average.