應用機器學習於預測維護診斷之研究

APPLICATION OF MACHINE LEARNING IN PREDICTIVE MAINTENANCE

洪鈺欣(Y.-S. Hung) ； 賴淑玲(S.-L. Lai) ； 張瑞益(R.-I. Chang)

《中國造船暨輪機工程學刊》 38卷2期 (2019/05) Pp. 53-59

摘要

近年來綠色能源意識抬頭，許多離岸發電、太陽能發電產業鏈也如雨後春筍般出現，國內外皆有導入工業物聯網至綠能產業中，舉天澤智雲公司為例，該公司將工業物聯網解決方案導入離岸風機監控。臺灣沿海為優質風場，適合發展離岸發電，然而離岸風機發電初期投入成本高昂，風機設備的維護成本較陸地風場高，隨著人工智慧與機器學習等資料分析技術推陳出新，近年來廣泛應用於工業物聯網（industrial internet of things，IIoT），在工業物聯網架構下，巨量資料可藉由數據採集與監控系統（supervisory control and data acquisition，SCADA）進行資料傳輸與設備監控，如何應用巨量資料預測設備之故障，對於設備的維護成本有著舉足輕重之影響，本研究目標在於導入資料分析技術至預測性維護分析（predictive maintenance，PdM），我們採用UCI PdM資料集進行演算法訓練分析，探討集成式演算法對於相關資料集之運算結果，其實驗結果指出不同集成式學習演算法均達96%準確度、平均AUC亦達98%，集成式演算法對於預測性分析有不錯的準確度，可供未來風機設備預測性維護分析之參考。

關鍵字

預測性維護 ； 資料分析 ； 機器學習 ； 集成式學習

並列摘要

Green energy industries are important for creating opportunities for the development of sustainable energy to flourish. Sustainable technologies such as wind turbines and solar panels have been extensively used in harnessing renewable energy resources. The Internet of Things (IoT) has brought transformations in various fields, including the green energy field, and the industrial Internet of Things (IIoT) is regarded as an industrial framework through which a large number of equipment can be synchronized by applying sensor, gateway, and cloud computing in machine-to-machine (M2M) and human-to-machine (H2M) communication. CyberInsight company proposed the WindInsight solution, the IIoT is used to predict the failure of wind turbines and maintenance of the equipment in this case. Taiwan has one of the best off-shore wind farms, and it is suitable to develop wind power industries. How-ever, the initial investment cost is very high for the off-shore wind power industries, and the wind turbines used in off-shore wind power industries have a higher maintenance cost than those used in plants. Recently, artificial intelligence and machine learning have been integrated in IIoT to develop value-added services, including Predictive Maintenance (PdM). PdM is used to determine when in-service equipment need maintenance to prevent costly operational interruptions resulting from equipment failures. It relies on real-time equipment data collection through the Supervisory Control and Data Acquisition (SCADA). The end-user can analyze the collected data to build failure models and make predictions for equipment maintenance. The application of PdM in IIoT has attracted considerable attention from various industries, and it is beneficial for reducing downtime cost and optimizing asset availability. This aim of this study is to apply data analytics technologies in PdM. Consequently, the UCI PdM dataset was used as the training dataset. The ensemble learning algorithms were employed to predict the failure of manufacturing processes. The results have a 96% accuracy on average, and the average AUC value is 98%. The higher computation performance demonstrated that ensemble learning algorithms offered the appropriate tools needed for the users to analyze their equipment data and develop maintenance strategies.

並列關鍵字

Predictive maintenance (PdM) ； Data analytics ； Machine-learning ； Ensemble-learning