應用多智能體分解與合成之動態派工與預防保養方法

Abstract

本研究提出MILPD(Mixed Integer Linear Programming Decomposition)模型，利用分派產品數與產能配置的概念降低多產品多機台問題的複雜度，該模型能夠克服大型馬可夫(Markov decision process)問題所造成的維度詛咒(Curse of dimensionality)，將大規模問題分解成若干個小問題，減少問題建模時的產品維度以降低求解的時間。接著，將工作站視為多智能體系統，利用MADP(Multi-agent based Dynamic dispatching and Preventive Maintenance)方法求解若干個小問題，得到每個單機台問題的動態與預防保養決策，並組合每個小問題的解以成為原問題的近似解。因此利用MILPD模型與MADP方法能夠彌補動態規劃方法所建構的DDPM(Dynamic dispatching and Preventive Maintenance model)模型無法實際運用於大規模系統的缺點。 另外，本研究運用深度學習(Deep Learning)建立了基於兩產品單機台架構的深度神經網路 (Deep Neuron Network，DNN)模型，能夠克服動態規劃模型專一性的問題，避免繁覆的建模過程，便能快速預測單機台的動態決策而獲得整個工作站的動態派工及預防保養策略。本研究亦開發組合多產品單機台動態策略的啟發式組合法SMP(Synthetic Multi-Products approach)，基於DNN模型的SMP方法能夠組合出多產品單機台問題的動態近似策略，以提高多產品系統在實務上運用本研究方法之可能性。 本研究模擬多個大規模系統問題，比較MADP方法與其他派工方法在不同績效指標下的表現。實驗結果顯示MADP方法的總完工時間能顯著的優於其他派工方法，並隨著系統利用率上升有更好的表現。另外，本研究欲探討系統內機台個數與產品種類數的改變對MADP方法造成的影響，分析結果為MADP方法在總產量和總完工時間的表現，皆不受機台數與產品總類數的變化而有所影響，代表運用本研究方法於不同的系統均能有穩健的效果。

This research proposes the MILPD (Mixed Integer Linear Programming Decomposition) model, which reduces the complexity of multi-product multi-machine problems by decomposing a large-scale problem into several small-scale problems. The model can overcome the dimensional curse caused by the large Markov decision process. Then, we can regarded a workstation as a multi-agent system, and use MADP (Multi-Agent based Dynamic dispatching and Preventive Maintenance) approach to solve each single-machine problem and combine all the solutions into an approximated solution of the original problem. Therefore, the MILPD and MADP methods improve the disadvantages of not being able to use DDPM model in large-scale systems. In addition, we use Deep Learning to construct a Deep Neuron Network (DNN) model to predict the dynamic strategy of two products and single-machine system, which can avoiding using dynamic programming to remodeling single-machine problems. Then, each agent in a workstation can use DNN model to predict single-machine solution quickly. Moreover, we also develop a method of synthesizing predicted results of DNN model to get multi-product single machine dynamic strategies, called Synthetic Multi-product approach (SMP), so agent can use SMP method to obtain a synthetic multi-product dynamic strategy. We simulate several large-scale systems and validate the performance of MADP approach. The experimental results show that the MADP approach is significantly better than other dispatching methods under the KPI of total completion time. It has better performance as utilization rate increases. In addition, our analyzed results also show that the MADP approach can perform robustly in different large-scale system.