近年來，發展智慧工廠成為全球製造業的目標，結合了人工智慧、物聯網、大數據和雲端運算等先進技術，希望以機器人、智慧製造和智慧服務等面向做到「智動化」，有助於提升生產的效率和商品的良率，並解決人力短缺、市場需求變化大的問題。本篇論文著重於預測機器剩餘可用壽命(Remaining Useful Life, RUL)，屬於機器健康預診斷(Prognosis)的應用，可用於預估機器剩餘可用壽命，在機器停止運作前，提早進行維修或更換，以降低機器突然停機所造成的損害，提高系統的運行可靠性。 本論文提出一個深度學習(Deep Learning)的方法，建構深度神經網路(Deep Neural Network)以預估機器剩餘可用壽命。所提之方法基於時間遞歸神經網路(Recurrent Neural Network, RNN)中的長短期記憶(Long Short-Term Memory, LSTM)模式。LSTM比傳統RNN更適合於處理和預測時間序列中間隔和延遲非常長的重要相關資訊，可以有效找出時間序列中的間隔相當長的相關資訊特徵。我們希望使用LSTM長記憶的特性，準確預估機器剩餘可用壽命。 為了驗證所提方法的效能，以NASA C-MAPSS(Commercial Modular Aero-Propulsion System Simulation)包含二百多組引擎模擬資料的資料集做驗證，並且與文獻中的MLP、SVR、RVR和CNN方法做比較。結果顯示，無論是在均方根差(Root Mean Squared Error, RMSE)還是在資料集本身定義的Scoring Function上，所提的方法都是最佳的。本論文最後並提出實作上的觀察和所提方法未來可能的應用場景。 ;In recent years, it is a worldwide goal to develop smart factories by integrating the artificial intelligence, Internet of Things and cloud computing technologies. Smart factories can achieve higher yield rates and better quality; they can also mitigate the problems of labor shortage and react properly to the dynamically changing of market. This thesis focuses on Remaining Useful Life (RUL) estimation, which is a part of the prognosis application. By accurate RUL estimation, machines or components can be repaired or replaced before they malfunction to cause the production line or the system to stop unexpectedly. This can reduce the damage caused by an unexpected shutdown, and reduce the cost of management. In this paper, we propose a deep learning method to construct deep neural networks for the RUL estimation. The proposed method is based on the Long Short-Term Memory (LSTM) model, which belongs to the category of Recurrent Neural Networks (RNNs). LSTM is more suitable for dealing with long-sequenced data of time series than general RNNs, and it can effectively extract and memorize significant relationship of data items which are apart from one another in the time series. It is believed that the memory characteristic in LSTM is useful for predicting RUL. The NASA C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) data set of hundreds of propulsion engines is applied to the proposed method for performance evaluation. The evaluation results are compared with those of the MLP, SVR, RVR and CNN methods proposed in the literature. The comparisons indicate that the proposed method is the best among all compared methods in terms of the Root Mean Squared Error (RMSE) and the Scoring Function. At the end of this thesis, we describe some observations and possible application scenarios of the proposed method.
