預測維護是醫療器材與工業器材智慧化管理的重要技術，其核心目標在於利用設備運轉資料、感測訊號與歷史維修資訊，於故障發生前推估設備狀態、故障風險與剩餘使用壽命。對醫療器材而言，設備可靠度直接影響臨床流程、病患安全與維修資源配置；對工業器材而言，停機時間、備品管理與產線穩定性則是導入預測維護的主要動機。因此本研究將預測維護定位為跨醫療與工業設備的序列資料分析問題，並以多資料集、多模型與多指標評估方式，探討不同序列模型在退化建模與部署效率上的適用性。 本研究以Mamba、Transformer、GRU、TCN 與LSTM五種序列模型為比較對象，並進一步提出整合式的新模型結構。資料集涵蓋AI4I2020、C-MAPSS、XJTU-SY與IEEE PHM 2012（PRONOSTIA）四類資料集上進行分析。分別代表結構化工業級機台資料、航空引擎退化資料，與軸承 run-to-failure 振動資料。由於目前公開醫療器材預測維護資料較少，故本研究以多種設備類型與不同退化訊號特性的公開資料作為方法驗證基礎，以提高模型對跨設備感測資料的泛化能力，並建立未來延伸至如醫療幫浦、影像設備、呼吸器、透析設備與高耗能醫療系統之資料分析流程。 研究中沿用既有 Python/PyTorch程式架構，對不同資料集分別進行資料清理、標準化、滑動視窗序列建構、生成故障horizon標籤、剩餘使用壽命(Remaining Useful Life， RUL)標註、模型訓練、效能評估與效率分析。再以MAE、RMSE、參數量、模型大小與推論延遲作為主要評估指標。初步結果顯示，單一模型沒有跨所有資料集均最優的情況。 例如Transformer在C-MAPSS與XJTU-SY的RUL誤差上具有優勢，GRU在效率與跨資料穩定性上具有實務價值，Mamba在部分資料上展現長序列建模的潛力，TCN對振動訊號的局部退化模式具有合理歸納偏置。基於上述的初步結果，本研究提出整合式的新模型，嘗試整合局部衝擊、長期退化與輕量記憶三種模型優勢，作為本論文後續主要貢獻與比較對象。 關鍵字：預測維護、剩餘使用壽命、故障預測、序列模型、整合式模型、醫療器材、工業物聯網;Predictive maintenance is a crucial technology for the intelligent management of medical and industrial equipment. Its core objective is to estimate equipment status， failure risk， and remaining service life before failure occurs， using operational data， sensor signals， and historical maintenance records. For medical equipment， equipment reliability directly impacts clinical workflows， patient safety， and maintenance resource allocation; for industrial equipment， downtime， spare parts management， and production line stability are the primary motivations for implementing predictive maintenance. Therefore， this study positions predictive maintenance as a sequence data analysis problem across medical and industrial equipment， and explores the applicability of different sequence models in degradation modeling and deployment efficiency using multi-dataset， multi-model， and multi-index evaluation methods. This study compares five sequence models—Mamba， Transformer， GRU， TCN， and LSTM—and further proposes a new integrated model. The datasets analyzed include AI4I2020， C-MAPSS， XJTU-SY， and IEEE PHM 2012 (PRONOSTIA). These represent data from structured industrial machinery， aircraft engine degradation， and bearing run-to-failure vibration， respectively. Although publicly available data on predictive maintenance of medical devices is currently limited， this study uses such data from various equipment types and with different degradation signal characteristics to validate the method. This aims to improve the model′s ability to generalize to cross-device sensing data and establish a data analysis workflow for future extensions to medical pumps， imaging equipment， ventilators， dialysis equipment， and high-energy-consuming medical systems. The study leverages the existing Python/PyTorch programming architecture to perform data cleaning， standardization， sliding-window sequence construction， fault-horizon label generation， remaining useful life (RUL) annotation， model training， performance evaluation， and efficiency analysis across different datasets. MAE， RMSE， number of parameters， model size， and inference latency are used as the main evaluation metrics. Preliminary results show that no single model is optimal across all datasets. The Transformer model has advantages in RUL error for C-MAPSS and XJTU-SY， the GRU model has practical value in terms of efficiency and cross-data stability， the Mamba model shows potential for long-sequence modeling on some data， and the TCN model has a reasonable inductive bias for local degradation modes of vibration signals. Based on the above preliminary results， this study proposes a new integrated model that attempts to integrate the advantages of three models: local shock， long-term degradation， and lightweight memory. This model will serve as the main contribution and comparison object for the subsequent work of this paper. Keywords: Predictive Maintenance， Remaining Useful Life， Failure Prediction， Sequence Modeling， Mamba， Transformer， GRU， TCN， LSTM， Integrated Model， Medical Equipment， Industrial Internet of Things
