雲端和伺服器服務與人們生活已密不可分，儲存裝置發生故障可能會造成系統停機、資料遺失或是高花費修復。雖然S.M.A.R.T.是一種常見的硬碟自我檢測技術，但是多數故障硬碟，無法在S.M.A.R.T.發現異常。若能夠增加一種S.M.A.R.T.以外的硬碟檢測方法，盡早預測硬碟故障，能夠提升系統穩定性與可靠度。因此，本研究提出一種硬碟預測式維護系統，透過感測器蒐集硬碟數據，使用快速傅立葉轉換與線性預測倒頻譜係數擷取特徵，以此特徵訓練自編碼器與遞迴神經網路進行異常偵測。實驗結果得知，傳統使用ReLU的自編碼器準確率76.42%，改用GELU能夠學習更複雜特徵的自編碼器準確率77.33%，結合遞迴神經網路，能夠學習資料時間關係的RNN自編碼器準確率84.75%，可以學習資料長期關係的LSTM自編碼器準確率85.08 %，輕量化LSTM的GRU自編碼器準確率87.08%。本研究以低成本與體積小的微控制器與感測器，不需要大幅度變動既有設備，即可佈署硬碟預測式維護系統，提供一種硬碟異常偵測方法。;Cloud and server services have become inseparable from people′s lives. Failure of storage devices may cause system downtime, data loss or costly repairs. Although S.M.A.R.T. is a common hard disk drive self-testing technology, most faulty hard disk drive cannot detect failure through S.M.A.R.T.. If a hard disk drive detection method other than S.M.A.R.T. can be added to predict hard disk drive failures as early as possible, system stability and reliability can be improved. Therefore, this thesis proposes a hard disk drive predictive maintenance system that collects hard disk drive data through sensors and use FFT and LPCC to extract features, utilizes these features to train recurrent neural networks and autoencoder for anomaly detection. Experimental results show that the accuracy of traditional ReLU-Activated autoencoder is 76.42%. Switch ReLU to GELU, which can learn more complex features, the accuracy of GELU-Activated autoencoder is 77.33%. Import the time series neural network to help the model learn the time relationships and features in the data. The accuracy of RNN-Based autoencoder is 84.75%. The accuracy of LSTM-Based autoencoder that can help learn the long-term relationship of the data is 85.08%. GRU-Based autoencoder is not only a lightweight version of LSTM-autoencoder but also achieves an accuracy to 87.08%. This thesis employs low-cost and small-size microcontrollers and sensors to implement predictive maintenance for hard disk drive, which can provide a hard disk drive anomaly detection method. It is suitable existing systems without significantly changing the equipment.
