警報資訊在FCC製程失誤診斷

Fault Diagnosis Using Neural Networks with Alarm Information

994 年在英國Milford Haven 的Texaco 煉油廠發生的爆炸及火災，造成26 人受傷，約 4,800 萬英磅的工廠設備與生產損失，從英國健康與安全部（Health and Safety Executive，簡稱HSE）研究調查指出，其主要因素有三：1.有太多的警報設定而其重要性未適當的排序。2.在控制室的顯示畫面，無法幫助操作員了解發生的狀況。3.訓練不足，使操作員無法面對工廠內連續不穩定的壓力。在爆炸前的 11 分鐘，有275個警報等待著兩位操作員去確認與應變。因為，操作員可能每分或每秒都會收到很多由量測元件傳來的數據，所以一旦製程狀態從正常的控制範圍偏離，操作員將面臨許多警報，為降低監控的負擔，因此本文分析警報歷史，鑑認警報與特定事件之間的關係，充分利用警報資訊建立失誤源預測模式，可協助操作員在製程發生狀況時，正確判斷失誤源。以流體化觸媒裂解單元（fluid catalytic cracking unit，簡稱 FCCU）之案例研究，來說明嚴重的失誤源，可能帶來警報氾濫問題，例如進料閥阻塞。一個失誤源發生後，其所有事件的演變與相關警報的引發，都由虛擬工廠（shadow plant）來模擬。虛擬工廠是 Honeywell 發展出的操作員訓練系統，能夠執行分散式控制系統（distributed control system，簡稱DCS）的功能，包含控制迴路參數及所有警報系統之設定，不僅 FCCU 之閉迴路動態可藉由虛擬工廠作模擬，而且製程中所有警報與製程變數的趨勢，可分別被收集和貯存在事件日誌收集器（event journal collector，簡稱 EJC）和製程歷史資料庫（process historical database，簡稱 PHD）中，利用不連續警報狀態和連續製程數據等 2 種製程資訊，本文所提出的失誤預測模式對FCCＵ的警報失誤源預測有明顯的功效。

The 1994 explosion and fire at the Texaco Milford Haven refinery injured twenty-six people and caused damages around 48 million pounds and significant production loss. Key factors that emerged from the investigation of the Health and Safety Executive, UK were: 1) there were too many alarms and they were poorly prioritized, 2) the control room displays did not help the operators to understand what was happening, 3) there had been inadequate training for dealing with a stressful and sustained plant upset. In the last 11 minutes before the explosion the two operators had to recognize, acknowledge and act 275 alarms. Since the operators are capable of dealing with lots of sensor data that update every minute or even seconds, they may be in front of dozens of alarms once the process status deviating from the normal control limits. To reduce the loading of monitoring and aid this decision-making process of operators, a fault diagnosis model has been developed. By analyzing the alarm history, the logics between alarms to be triggered and the specific events can thus be identified and used to construct a fault diagnosis model. A case study of a Fluidized Catalytic Cracking Unit (FCCU) is used to illustrate the possible alarm flooding problem caused by a severe fault such as a sticky valve of the input stream flow. All scenarios of event propagation and correlated triggered alarms after a specific root cause or fault are simulated by Shadow Plant, an operator training system developed by Honeywell, which can also perform the Distributed Control System (DCS) functions including the control loop parameters as well as all the alarm system settings. Not only the FCCU closed-loop dynamics is simulated by Shadow Plant, but also all the alarms and the trends of process variables are collected and stored in the Event Journal Collector (EJC) and Process Historical Database (PHD), respectively. With these two types of data, discrete alarm status and continuous process values, the proposed architecture of logical process for FCCU alarm suppressing are preformed effectively.