先期開發與評估應用於石化業關鍵設備之智慧預知維護方法

Analyzing the Systems for Certification of Personnel Competencies for Explosive Atmospheres

郭至恩(Chih-En Kuo) ； 張純明(Chwen-Ming Chang) ； 高振山(Chen-Shan (Jason) Kao) ； 許世希(Shih-His Hsu) ； 陳振和(Jhen-He Chen) ； 蔡瑜潔(Lih-Chuan Chen) ； 梁勝富(Sheng-Fu Liang)

《勞動及職業安全衛生研究季刊》 26卷1期 (2018/03) Pp. 1-8

摘要

壓縮機是石化製程中最關鍵的設備，石化廠壓縮機跳機損失動輒以千萬元計，而氣體的可壓縮性造成跳機時危險性相對較高，也是整個石化業者的競爭力的關鍵所在。在工業4.0的推動下，許多產業界機具的智慧狀態診斷與預知維護概念紛紛被提出中，預知維護（Predictive Maintenance，簡稱PdM）是以狀態為依據(Condition Based)來預測設備狀態未來的運作趨勢，據以預先制定性維護計劃，以預測結果來輔助決策。預知維護不僅能優化設備的運作時間和性能，並減少預防性維護的時間和人力成本，提升企業在製程安全管理的能力。本研究的目標是提出一套應用於石化業關鍵設備壓縮機之智慧預知維護方法，以預知跳機時間，避免危險性跳機，同時可以使停機時間縮短，將誤判跳機降至最低，進而提高生產效益，提升安全管理績效。我們使用國內石化廠壓縮機長度約連續81天的物聯網資料來開發預知維護方法，使用五種不同的數值分類器運用於氣體壓縮機上振動、壓力、溫度等三種不同類型的大數據感測資料上，以評估分類器的適用性與感測資料類型的重要性順序。此外，為了降低運算負擔與萃取有效的資訊，我們亦使用主成分分析(Principal Components Analysis, PCA)技術應用於壓縮機之感測資料上。本研究的結果顯示使用最近鄰居法(K-nearest neighbor, KNN)分類器在預測壓縮機狀態的平均準確性最高，達到99.68%。線性判別分析(Linear discriminant analysis, LDA)分類器的平均準確性則最低，只有86.15%。而在資料類型的重要性順序則依序是振動、壓力、溫度。未來研究方向須將短期感測資料與資料的頻譜特徵考慮進去，以得到更通用的人工智慧預知維護方法。

關鍵字

壓縮機 ； 物聯網 ； 大數據 ； 製程安全管理 ； 人工智慧

並列摘要

In the petrochemical industry, an unplanned stop causes extremely high costs. It results in an unscheduled downtime with no possibility to continue production, unplanned maintenance costs are a lot higher than planned maintenance, and often, sudden high investments in new machinery or machine parts must be done. A shutdown would cause production downtime, and therefore, forces the companies to buy the products from competitors, causing major costs. Through condition monitoring analyses, the lifespan and future maintenance of machines are determined. This ensures that the necessary small repairs do not grow into major repairs, which requires a prolonged production stoppage. Predicted maintenance is a condition-based maintenance through IOT gathering various parameters data (big data) during production and analyzing these data through data mining to determine the condition of production assets and to decide which action to be taken for the purpose of important process safety management decision. In this study, we use the data from petrochemical plant compressor which length of about 81 days to develop an artificial intelligence predictive maintenance methods. Five different numerical classifier were used in compressors on the vibration, pressure, temperature types of sensing data in order to assess the suitability of the classifier and the importance of the type of data. In addition, in order to reduce the computational burden and extract the effective information, we also use the principal component analysis (PCA) technology applied to the compressor of the sensing data. The results show that the average accuracy of the state of the compressor is predicted to be 99.68% using the K-nearest neighbor classifier. The average accuracy of the linear discriminant analysis classifier is the lowest, only 86.15%. While the order of importance in the data type is followed by vibration, pressure, and temperature. In future, in order to develop a generalized predictive maintenance model with high accuracy, more short-term data and frequency domain features should be considered and tested.

並列關鍵字

Compressor ； Internet of Things ； Big Data ； Process Safety Management ； Artificial Intelligence

參考文獻

Nolan, F,Heap, H.(1979).Reliability Centered Maintenance.National Technical Information Service Report.(National Technical Information Service Report).,未出版.

Sherwin, D(2000).A Review of Overall Models for Maintenance Management.Journal of Quality in Maintenance Engineering.6,138-64.

Northerton, D(2000).RCM Standard.Maintenance & Asset Management.15,12-20.

Mobley, RK(2002).An introduction to predictive maintenance.Butterworth-Heinemann.

El-thalji, Idriss,Jantunen, Erkki(2015).A summary of fault modelling and predictive health monitoring of rolling element bearings.Mechanical Systems and Signal Processing.60,252-72.