人工智慧於產線決策之應用－以航太放電加工產線為例

Abstract

“萬里之船，成於羅盤；千里之行，積於跬步”，一個有效率且準確的生產方式應植基於初始完善的規劃與安排，本研究以航空發動機擴散器放電加工智慧產線作為研究標的，分析企業新產線的評估流程，以及探討如何建置放電加工產線製程偏誤智慧預警模型，將人工智慧技術實際應用於航太放電加工智慧製造產線製程偏誤決策支援。 放電加工為非傳統製程，由於加工件需要浸泡在介電質溶液內，因此加工過程是否產生製程偏異難以使用目測法判別。本研究標的產品為航空專用部件，產品幾何形狀複雜，若使用rule based方式對工件放電加工過程之物理訊號進行分析，也難以定義出一套規則協助分析判斷，加上放電加工製程物理訊號資料量龐大，也不適用採用rule based方式進行製程偏異檢測。 鑒於以上之原因，本研究提出以深度學習之方法，檢測放電加工用於航太複雜部件中的物理訊號分析演算法，用於判斷加工過程是否產生偏異。本研究使用類神經網路、遞迴神經網路以及自動編碼器做為演算法基底，擷取放電加工設備控制器資料和放電迴路之物理訊號作為輸入，並給予標記或無標記，使其判斷加工之狀態，輸出則為所標記之種類。 本研究結果發現無論是採用類神經網路或遞迴神經網路，皆可準確偵測放電加工製程變異，而自動編碼器為無監督式學習之方法，雖然現階段尚無法準確分群，但其自動分類的特性，可協助研究人員找到一般人所認定規則以外的規則，是一項用於探索新規則時的利器。 此外，從整理本研究資料之過程可知，企業對於新產線之評估，除了必須由產品及市場角度切入從時間軸維度分析投入資金設置新產線是否能還能在產品生命週期內創造出經濟效益外，也必須評估新建之生產線是否具備產業長期競爭力，才不致於使新產線面臨淘汰或者被競爭者取代之風險。

An efficient and precise production method should be based on the initial perfect planning and arrangement. This study uses the intelligent production line of aero engine diffuser electrical discharge machining as a study Target, to research the evaluation process of the new production line introduction, and how to build a smart early warning model for the process error of the EDM production line, and apply artificial intelligence technology to the decision support of the EDM manufacturing line for aviation industry. Electrical discharge machining is a non-traditional process. Because the parts need to be immersed in a dielectric solution, the quality of the process result is difficult to check by visual inspection. The rule-based method is used to detect the physical signals of electrical discharge machining. Because the object of this research is an aviation components, the geometry is complex, and it is difficult to define a set of rules for the physical signals of electrical discharge machining. In addition, the physical signal data of electrical discharge machining is huge, so rule based method is not applicable. Based on the above reasons, this study proposes a deep learning method to come up with the physical signal analysis algorithms for electrical discharge machining apply to aviation components, and to judge the quality of the EDM process. In this study, neural network, recurrent neural network and auto-encoder are used as the basis of the algorithm. The controller data of EDM and the physical signal of the electrical discharge circuit are captured as input, and identify the process status by different marking. The output is the marking type. The results of this study found that both neural network and recurrent neural network get pretty good results. But the auto-encoder is an unsupervised learning method. Although it is not yet possible to accurately group at this stage, its automatic classification features can help researchers find rules other than those recognized , and is a good weapon for exploring new rules. In addition, through the research process, we found that the new production line introduction evaluation not only need to consider whether investing funds to set up a new production line can still create economic benefits within the product life cycle, but also consider whether the new production line has long-term competitiveness in the industry, so as not to expose the new production line to the risk of being eliminated or replaced by competitors.