請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/87790
標題: | 使用類神經網路與理論模型對聚合物的預測及控制 Prediction and Control of Polymers with Artificial Neural Network and Theoretical Model |
作者: | 高衡山 Heng-Shan Kao |
指導教授: | 吳哲夫 Jeffrey D. Ward |
關鍵字: | 類神經網路,模糊控制器,模型預測控制器,高分子聚合物製程, Artificial neural network,Fuzzy controller,Model predictive control,Polymer process, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 在高分子製程中,對於產品性質管控上具有一定的標準,但品管指標的測量常需要以人工間隔數小時之採樣方式,無法線上進行量測,導致品質管控上之困難。此外,許多化工製程生產時有著不同規格之產品,因此常面臨產品規格間之轉換,現場通常以操作人員經驗進行調整,因此容易因操作人員之判斷疏忽、手法不一致等因素造成品別轉換的時間延長且有次級品的產生。
針對產品性質測量產生之問題,本研究嘗試兩種方法,第一種在品別資料充足的情況下,使用非線性循環類神經網絡模型於產品之動態預測,實現對品質之實時監控與未來趨勢之預測;第二種在資料不足情況下,使用Aspen Plus建立理論模型,便於減少找尋新品別之試誤時間和物料成本,未來亦可進而利用於節能研究。本研究利用虛擬PID控制器、虛擬模糊控制器與模型預測控制器三種控制算法,實現製程導航技術,用以實時提供現場人員操作變數之調控指引,改善品質震盪或產品規格轉換時間過長之問題。 本研究以高分子聚合物製程為例,其中高分子產品之熔融指數、平均重量分子量與平均數量分子量皆可作為品質指標,斷鏈劑添加量為品質控制之操作變數。結果顯示數據模型能透過過往資料準確預測,製程導航系統用於斷鏈劑添加操作指引,相比實廠手動控制策略,也能更快完成品別轉換,縮短其耗時與降低次級品的產生;理論模型則自行開發四種品別,提供資料量不足的工廠一個參考,並使用製程導航系統達到品別改制,同時比較不同虛擬控制系統的轉換時間與平均絕對誤差。 In polymer processes, there are certain standards for product quality. However, quality indicators often cannot be measured online, but are instead measured offline intervals of several hours. This results in difficulties in quality control. In addition, many polymer processes produce products with different grades, so grade transitions occur frequently. Some process variables are adjusted according to a recipe, but others are adjusted based on the experience of the operator, so it is common for the grade transition to take longer than necessary for excessive amounts of off-spec product to be produced. To address these problems, in this work two different models of polymerization process are developed. The first is based on industrial data and uses nonlinear recurrent neural network models to dynamically predict product properties to realize real-time monitoring of quality and the prediction of future trends. This approach requires sufficient industrial data for model training and validation. The second model is developed using Aspen Plus. Aspen Plus can be used to establish a theoretical model in the case of insufficient data, which is convenient because it reduces the trial and error time and material cost of modeling new grades. After the model is established, three control algorithms are developed: a virtual PID controller, a virtual fuzzy controller, and an RNN based virtual model predictive control controller. These are used to deal with operation problems such as variability in product quality and long grade transition times. In this work the melt index, weight average molecular weight and number average molecular weight of polymer products are quality indicators, and chain modifier feed flow rate is a manipulated variable for quality control. The results show that the data driven model can accurately predict the melt index. A process guidance system is used to guide the operators in adjusting the chain modifier feed flow rate. Compared with the manual control strategy in the actual factory, it can complete the grade transition faster and reduce the generation of off-spec products. The theoretical model is tested with four fictitious grades. It can provide a reference for factories that cannot develop a data-driven model due to insufficient data. The model is also used to compare the transition time and average absolute error of different virtual control systems. The results show that the model predictive controller performs better than the other controllers. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/87790 |
DOI: | 10.6342/NTU202300997 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 化學工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 此日期後於網路公開 2028-06-12 | 6.9 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。