利用人工智慧建立石油煉製業壓縮機之異常診斷機制

Anomaly Detection of Compressors in Oil Refinery Industry byArtificial Intelligence

呂政芳(Jeng-Fan Leu)；趙仁德(Ren-Der Chao)；陳彥銘(Yenming J. Chen)

壓縮機 ； 高斯混合馬可夫隨機場 ； 稀疏空間理論 ； 集成學習 ； 專家評分 ； 人工智慧 ； Compressor ； Gaussian Markov Random Field Mixtures ； Sparse Space ； Ensemble Learning ； Expert Assessment ； Artificial Intelligence

石油季刊

58卷1期（2022 / 03 / 01）

83 - 103

繁體中文

本研究針對石油煉製業的關鍵設備，如大型壓縮機，為其建立故障預警模型。但是大型設備可供經驗累積的故障模態極少發生，也使得正式的故障分析極難實現。為了克服資料不足的問題，本研究發展出對應的處理過程，最主要的核心是假設產生故障的因子，並不是直接觀察到的變數，而是以另一空間的隱藏變數取代之。此外，再輔以稀疏空間理論及機器學習方法，將縮減過的複合式有效特徵預測因子（例如10個以內），導入高斯混合馬可夫隨機場以估計出隱藏變數，最後再利用隨機森林等集成學習方法做出判斷。本研究發現，採用稀疏隱藏空間的選取及轉換過的特徵，可以提升集成預測模型的準確率，結合專家對設備健康度評分的反饋可以循環性地透過再學習來強化預警模型。本研究提供建立石油煉製業關鍵設備之狀態診斷、預警機制與工廠維護策略之參考，以降低工安環保事故與跳俥損失，提升公司競爭力。

This study aims to establish an anomaly detection model for key equipment in oil refinery industry, ex. mega compressor. However its experienced failure modes are so rare that formal fault analysis is extremely difficult to achieve and the amount of information is far from the basic amount of machine learning, so in practice can only temporarily use this near-speculative way. The challenge of insufficient data must be solved first. In order to overcome the data problem, this study has developed a series of approaches to increase the accuracy of predictions. The core of the approach is to assume that the factors produced the failure is not the variable we observe directly, but the variable hidden in another space. In addition, we will be complemented by sparse space theory and machine learning methods, will be reduced by the represented predictors (e.g. less than ten), import Gaussian hybrid Markov model, estimate hidden variables, and then use random forests and other ensemble learning methods to make a final judgment. We use the novel approach in hidden spaces to fill the problem of fault mode deficiency. We found that the selection and transformation of features of sparse hidden space can improve the accuracy of the integrated predictive model and combine expert ratings on device health feedback from reinforces the early warning model in a recurrent way by learning again. This study provides a reference for establishing the state diagnosis, early warning mechanism and plant maintenance strategy of key equipment in oil refinery industry, in order to reduce the loss of environmental accidents and factory downtime and enhance the competitiveness of the company.

1. 王立志,范書愷,丁慶榮,郭財吉,林春成,許嘉裕(2018)。生產系統於先進智能製造之展望。管理與系統，25(3)
   連結：
2. 郭志恩,沈育霖,曹常成,張純明,高振山,許世希,邱俊憲,陳振和,蔡瑜潔,梁勝富(2018)。應用於石化業關鍵設備之集成式智慧預知維護系統。勞動及職業安全衛生研究季刊，26(3)
   連結：
3. 郭志恩,張純明,高振山,許世希,陳振和,蔡瑜潔,梁勝富(2018)。先期開發與評估應用於石化關鍵設備之智慧預知維護方法。勞動及職業安全衛生研究季刊，26(1)
   連結：
4. (1999).Machinery Failure Analysis and Troubleshooting, Practical Machinery Management for Process Plants.Elsevier Inc..
5. Anderson, T. W.(2003).An Introduction to Multivariate Statistical Analysis.Wiley-Interscience.
6. Blythe, D.,von Bunau, P.,Meinecke, F.,Muller, K.-R.(2012).Feature extraction for change-point detection using stationary subspace analysis.IEEE Transactions on Neural Networks and Learning Systems,23(4),631-643.
7. Campos, G. O.,Zimek, A.,Sander, J.,Campello, R. J.,Micenkov´a, B.,Schubert, E.,Assent, I.,Houle, M. E.(2016).On the evaluation of unsupervised outlier detection: measures, datasets, and an empirical study.Data Mining and Knowledge Discovery,30(4),1-37.
8. Kampa, Simon, “5 Steps to Implementing Predictive Maintenance at Scale”, IMPO. (https://www.impomag.com/maintenance/article/13246724/5-steps-to-implementing-predictive-maintenance-at-scale), 2018.
9. Perez, Robert X.,Conkey, Andrew P.(2016).Troubleshooting Rotating Machinery: Including Centrifugal Pumps and Compressors, Reciprocating Pumps and Compressors, Fans, Steam Turbines, Electric Motors, and More.John Wiley & Sons, Inc..
10. Xiong, L.,Chen, X.,Schneider, J.(2011).Direct robust matrix factorizatoin for anomaly detection.Data Mining (ICDM), 2011 IEEE 11th International Conference
11. 陳昇瑋,溫怡玲(2019).人工智慧在台灣,產業轉型契機與挑戰.