近年來，智慧工廠為熱門的研究項目之一。預診斷健康管理系統在智慧工廠應用中扮演相當重要的角色，可位機器或元件提供不同層次的預診斷，如錯誤預測和剩餘使用壽命(Remaining Useful Life，RUL)預估。本篇論文強化卷積神經網路(Convolutional Neural Network, CNN)用於預診斷健康管理的可用壽命預估，意指在機器經過一定時間的運作之後，預估其在無法運作之前所剩餘的可用壽命。透過預估機器可用壽命，可使維護人員能夠提早進行機器的維護或替換，以確保系統的永續性。 本論文提出強化卷積神經網路(Enhanced Convolutional Neural Networks, ECNN)進行機器剩餘可用壽命預估。ECNN與傳統CNN的不同在於，傳統CNN大多數用於圖像處理，而ECNN則應用於時序資料進行剩餘可用壽命預估上。ECNN透過更細緻的資料前處理、更好的適應時機預估(adaptive moment estimation, Adam)優化器、更適合的softplus激活函數以及疊加圖層的方式，將時序資料視為雙色圖像，擷取長時間序列資料的有效特徵，準確預估機器剩餘可用壽命。 本論文在實驗階段以NASA C-MAPSS(Commercial Modular Aero-Propulsion System Simulation)的模擬資料集作為驗證對象，其資料集內容為航太引擎資料之時序資料及運作狀態。我們將驗證結果與其他文獻所提出的方法進行比較，如MLP、SVR、RVR和CNN，可發現ECNN優於其他方法。 ;The smart factory becomes a hot research topic recently. Prognostic health management (PHM) plays a critical role in smart factory applications to produce different level of prognostics, such as failure prediction and remaining useful life (RUL) estimation, for machines or components. This thesis enhances the convolutional neural network (CNN) deep learning for RUL estimation in smart factory applications. A CNN is a special type of deep neural networks (DNNs) used in deep learning for analyzing image data for the applications of image recognition and video recognition. It has convolution layers, pooling layers, and fully connected layers. A convolution layer contains many filters to abstract features from input data, and a pooling layer can reduce data dimensionality without losing features. The CNN deep learning has been applied in an earlier study for RUL estimation. This thesis enhances the learning by applying more sophisticated data pre-processing, a better optimizer, namely the adaptive moment estimation (Adam) method, and a proper activation function, namely the softplus function. The enhanced CNN deep learning is applied to NASA C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) data set to estimate the RUL of aero-propulsion engines. Its performance is evaluated by a scoring function that can measure the RUL estimation accuracy. The evaluation results are compared with those of other methods using the multi-layer perceptron (MLP), support vector regression (SVR), relevance vector regression (RVR) and traditional CNN. We find that the enhanced CNN deep learning method is superior to other methods.
