以可靠度為中心與以機器學習為依據之預測性維護研究比較

Abstract

近年來，隨著科技發展，大數據分析變得更為容易，人工智慧的發展也帶出工業4.0、智慧機械等生產製造業的發展需求，其中，預測性維護(predictive maintenance, PdM)常為工程師留意且重視的議題，針對該議題，有人想到「以可靠度為中心之維護」(reliability-centered maintenance, RCM)研究；也有人透過「機器學習」(machine learning, ML)相關方法探討問題；本論文旨在探討由「以可靠度為中心」與「以機器學習為依據」不同思維發展出來的兩種分析方法，在系統或設備預測性維護所獲結果之異同，並探討兩種方法之優劣。本研究分析一批與飛機引擎健康度監控有關的大數據，一方面依照RCM思維，建立一套以指數模型為依據的健康指標，而後依據貝葉斯理論處理監控到之即時數據，以預測引擎之殘餘使用壽命(remaining useful life, RUL)；另一方面，依據ML思維，針對監控數據，建立一套長短期記憶(long short-term memory, LSTM)模型來處理提取的特徵並捕捉機械系統或設備的退化過程，並引入多頭注意力機制(multi-head attention)加重數據序列中重要部分，以提高RUL預測的準確度。本研究分析結果顯示，兩種方法都能讓我們依據監控數據預測引擎之RUL，但ML方法較能精確捕捉引擎健康退化程度，有較佳預測；然而，RCM在操作的可解釋性和規則明確性方面則具有較大優勢。

Along with technological developments including artificial intelligent, the ability of computer for treating big data has progressed tremendously in recent years. The development of artificial intelligence has also brought about the development needs of production and manufacturing industries such as Industry 4.0 and smart machinery. Among these, ‘predictive maintenance’ (PdM) is a topic of significant interest and importance to engineers. Some researchers focus on ‘reliability-centered maintenance’ (RCM) while others explore the issue through ‘machine learning’ (ML) methodologies. This study aims to investigate the similarities and differences in the results of system or equipment predictive maintenance obtained from the two analytical approaches based on RCM and ML perspectives, as well as to discuss the strengths and weaknesses of both methods. This study analyzes a batch of big data related to the health monitoring of aircraft engines. Following the RCM approach, a set of health indicators based on the exponential model is established, and then the monitored real-time data is processed according to Bayesian theorem to predict the remaining useful life (RUL) of the engines. On the other hand, based on the ML approach, the long short-term memory (LSTM) is used to process the extracted features and capture the degradation process of the mechanical system or equipment, and a multi-head attention mechanism is introduced that emphasizes important parts of the data sequence to improve the accuracy of RUL prediction. The results of this study show that both methods allow us to predict the RUL of the engine based on monitored data. However, the machine learning (ML) method is more accurate in capturing the degradation of engine health, providing better predictions. On the other hand, RCM has a greater advantage in terms of interpretability and clarity of operation rules.