基於TSMamba深度學習模型的3D快閃記憶體異常偵測;Anomaly Detection for 3D NAND Flash Memory Based on the TSMamba Deep Learning Model

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/98321

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98321

題名:	基於TSMamba深度學習模型的3D快閃記憶體異常偵測;Anomaly Detection for 3D NAND Flash Memory Based on the TSMamba Deep Learning Model
作者:	林瑞庭;LIN, Jui-Ting
貢獻者:	資訊工程學系
關鍵詞:	異常偵測;3D NAND Flash
日期:	2025-07-24
上傳時間:	2025-10-17 12:37:37 (UTC+8)
出版者:	國立中央大學
摘要:	隨著儲存裝置朝向高密度與長壽命發展，NAND Flash 健康監控與異常預警需求日益迫切，現行多數監測機制仍仰賴被動式錯誤回報與統計閾值判斷，難以即時掌握非線性退化趨勢與高雜訊負載所造成之潛在風險。尤其在企業級儲存與嵌入式應用中，一旦錯誤擴大至無法修復階段，將導致重大資料損失與系統停機，進而衍生高額維運成本與營運中斷風險。為解決上述挑戰，本研究提出一套結合深度學習技術之 NAND Flash 健康預測系統，採用具備選擇性狀態建模能力之 TSMamba 架構，整合滑動視窗切片、嵌入式特徵轉換等技術，針對多維時序資料（如 ECC 錯誤數、P/E 次數與溫度）進行單點預測與異常分析。系統實作涵蓋韌體日誌解析、資料標準化、序列建構與異常信心評估，並透過 Sigmoid 函數結合統計（μ, σ）進行分層式門檻判定。實驗結果顯示，於 μ+ 2σ 至 μ + 3σ閾值設定下，可有效達成 FAR 低至 0.58%、Lead Time 降至 18.25 小時及 RUL 降至 20.9%，展現高度準確性與提前性。同時，本系統模型參數量僅 7.27M，推論延遲僅 4.57 ms，具備即時性與部署彈性。整體而言，本研究不僅提升 NAND Flash 故障預警效能，亦提供可因應應用場景調整之高適應性預測性維護解決方案，具備跨儲存裝置推廣潛力。 ;As storage devices continue to evolve toward higher density and longer lifespans, the demand for health monitoring and anomaly prediction in NAND Flash memory has become increasingly critical. Most existing monitoring methods still rely on passive error reporting and threshold-based statistical approaches, which struggle to detect nonlinear degradation trends and anomalies under high-noise, variable-load conditions. In enterprise and embedded applications, uncorrected faults can lead to severe data loss and system downtime, resulting in costly maintenance and operational risks. To address these challenges, this study proposes a NAND Flash health prediction system based on deep learning techniques. The system adopts the TSMamba architecture, which incorporates selective state-space modeling, sliding window segmentation, and embedded feature transformation to perform point-wise forecasting and anomaly analysis on multi-dimensional time series data, including ECC errors, P/E cycles, and temperature. The implementation includes firmware log parsing, data normalization, time series construction, and anomaly scoring through a sigmoid function combined with statistical norms (μ, σ) for threshold-based classification. Experimental results demonstrate that with thresholds set between μ + 1σ and μ + 2σ, the system achieves a false alarm rate (FAR) ≤ 3%, an average lead time of approximately 23 hours, and a remaining useful life (RUL) of at least 21%, indicating strong predictive accuracy and early warning capability. Moreover, with only 7.27M model parameters and an inference latency of 4.57 ms, the system exhibits real-time responsiveness and lightweight deployment potential. In summary, this work offers a highly adaptive and accurate predictive maintenance solution for NAND Flash memory and holds promise for broader deployment across various storage health management applications.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	14	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....