Page 75 - Vol.45
P. 75

TSMC / FACILITY PUBLISHED
                VOL.45
                                                         廠務季刊


                                                          URL.http://nfjournal/



                機械學習診斷冰機冷媒系統方法與實例

                Chiller Refrigeration Cycle Anomaly Detection with Machine Learning methodology
                and operation


                                                                        1
                                                                                  3
                                                                                               2 3
                                                                             2
                                                                                    1
                                                                  文│蔡昆憲  鄭凱元  謝子宸  │ 廠務數位發展部  十四廠暨廠務自動化整合部│
                摘要 / Abstract
                    本案為廠務數位發展與廠務CIM合作為台積電廠務首度成功推廣至全公司使用機器學習方法達到超越專家能力的異常偵測系
                統。主要應用於缺乏代表性感測元件的冰機冷媒密閉迴路,在不額外增加廠務工程與維護成本條件下,藉由現有資訊進行解析與
                運算,進而取代專家進行預警與判讀。本案已於2021年10月在北中南晶圓廠廠務運轉超過兩年的300mm廠區完成導入。解決密閉
                冷媒負壓洩漏量偵測、冷卻水質惡化偵測、冷媒含水偵測等中央空調傳統上難以判讀的異常偵測。可提早14天~57天異常偵測,
                讓細小到人類難以察覺的風險由人工智慧代理監控並量化品質管理,達成智能廠務冰機系統維運應用。

                關鍵詞 / 冰機系統、資料科學、機器學習、模型辨別預警、冷媒循環

                    FACDD cooperate with FDCIM to successfully develop the anomaly detection system that surpass the capabilities of experts,
                and this is the first time we promote to whole facility division of TSMC. It is mainly used in the closed loop of chiller refrigeration
                cycle that lacks representative sensing elements. To use the existing information for analysis and calculation to replace experts for
                early warning that without additional factory engineering and maintenance costs. This project has been completed in October
                2021 in the 300mm fab that has been operating for more than two years in the Taiwan. Solved the traditionally difficulty to
                interpret abnormal detection of chiller system, such as leakage in the negative pressure system detection, bad condenser water
                quality detection and water in the refrigerant detection. Abnormal detection can be carried out 14 days to 57 days in advance,
                so that the anomalies that are too small to be detected by humans can be monitored, quantified, and managed by this AI agent.
                Intelligent maintenance and application of facility chiller system can be achieved.

                Keywords / Chiller System, Data Science, Machine Learning, Pattern Recognition Alert, Refrigeration Cycle



                1.  前言

                    隨半導體世代演進,製程設備愈來越精密且複雜,製程                       案,在此,我們借重機器學習(Machine Learning)的技術來找
                參數也跟著指數型增加,面對上千上百的製程參數,人類的專                        出細節中的魔鬼。依據經驗,大多數的冰機異常問題屬於多
                注力已不足應付,所以有了自動警報系統來幫助工程師捕捉異                        維度異常,意旨單一冰機異常往往和多個運轉參數相關,故
                常。然而現行製程參數的管理局限於SPC/ICCI等單維度統計檢                    本研究致力於多維度異常分析,先使用機器學習方法來學習
                測,此方法單純從數學統計出發,將每個參數獨立管理,未加                        冰機正常運轉基線,再利用基線偵測異常並加以異常分類。
                以考慮參數的物理性以及參數間的相依性,如此容易與運轉工                        我們不希望工程師去猜警報可能異常的原因,而是收集和統
                程師的管理產生隔閡。統計上的偏差在專家工程師眼中可能只                        計300mm廠區過往案例,提供運轉同仁查案的標準流程並透
                是物理參數相依性影響,例如:功率的變化與冷凍噸和外氣焓                        過連線會議與北中南同仁分享成果。未來只要透過此「智能
                值相依,屬於正常現象,但往往卻不是如此。                               冰機冷凝壓力異常偵測系統」,簡稱「AI冰機冷凝偵測」,
                    魔鬼藏在細節中,只要對數據做正確的分析才會有答                        就可以清楚判斷「缺乏代表性感測器」的異常是否出現,精


                                                                                             FACILITY JOURNAL        03  2022  73
   70   71   72   73   74   75   76   77   78   79   80