請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90741
標題: | 心電圖身分辨識演算法與電路設計 Algorithm and Circuit Designs for ECG-based Biometrics |
作者: | 陳奕達 Yi-Ta Chen |
指導教授: | 吳安宇 An-Yeu Wu |
關鍵字: | 極限學習機,深度學習,心電圖,身分辨識, Extreme Learning Machine,Deep Learning,Electrocardiography,User Identification, |
出版年 : | 2023 |
學位: | 博士 |
摘要: | 隨著第四次工業革命的到來,工業物聯網(IIoT)中的設備在製造過程中產生大量機密數據,因此需要嚴格保護。用戶身份識別成為確保這些數據安全的必要條件。在各種用戶識別方法中,心電圖已逐漸成為眾所皆知的生物識別(Biometrics)方式,因為它能夠提供受測者活著的證明並且難以被竊取。本論文著重於心電圖生物識別中的用戶識別(User Identification)與身份驗證(Identity Verification)。面對以邊緣為中心的用戶識別任務中,所需要面對新的員工隨時會加入的挑戰,加上邊緣端的電池限制,我們亟需一個支援線上類別增量學習(O-CIL)且高能源效率的系統。面對雲端為中心的身分驗證任務,我們亟需優化系統的準確度與通用性。
本論文提出了一種可擴展的極限學習機 (S-ELM) 演算法,以克服新用戶於以邊緣為中心的基於心電圖用戶識別任務中的挑戰。我們選擇極限學習機(ELM)模型作為分類器,是因為它具有更快的學習速度。此外,極限學習機當中的解析解相對於深度學習分類器中的迭代解更適合使用於O-CIL場景下。因此本論文提出“其他”類概念賦予ELM的O-CIL的能力。為了在邊緣裝置上實現S-ELM,我們提出了可擴展的 QR 分解 ELM (S-QRD-ELM)並且將其使用TSMC 40nm製程來實現。在S-QRD-ELM引擎中,我們採用折疊(folding)和硬體共享(hardware-sharing)技術來設計,以提高整體引擎的能效。此外,本論文借助聯邦學習的概念,將聯邦平均算法(Federated Averaging)與 ELM結合,提出聯邦極限學習機系統(Fed-ELMS),使得邊緣裝置上的ELM模型能夠藉由相互合作來獲得更好的全局模型。 另一方面,本論文提出了一種可擴展的N-pair loss深度心電圖(SNL-Deep-ECG)系統,用於雲端上的身份驗證任務,以獲得更好的用戶體驗、準確度和通用性。單心電圖波形預處理方式縮短了訊號採集時間,提升了用戶體驗。使用N-pair loss訓練Deep-ECG模型相對於原始的交叉熵損失可以獲得更高的準確度。最後,本論文提出了混合數據庫驗證方法來驗證SNL-Deep-ECG的通用性。 綜上所述,本論文針對基於心電圖的邊緣裝置用戶識別和雲端身份驗證系統提出了多項前瞻性的設計,並希望我們所提出的技術能成為未來基於心電圖的生物識別系統中的關鍵技術。 As the fourth industrial revolution arrives, smart devices in the Industrial Internet of Things (IIoT) generate massive amounts of confidential data which require strict protection and user identification comes in handy. Among the user identification methods, ECG signals have emerged as an attractive biometric modality due to their ability to provide intrinsic liveness proof and their difficulty in spoofing. In this dissertation, we focus on the ECG-based user identification and identity verification task. For edge-centric user identification tasks, facing the fact that new users would enter the smart factory and the battery limitation on edge devices, the system requires online class incremental learning (O-CIL) capability with high energy efficiency. On the other hand, for cloud-based identity verification system, the accuracy and generalizability need to be improved. This dissertation presents a scalable extreme learning machine (S-ELM) algorithm to conquer the challenges of the edge-centric ECG-based user identification task. We select the extreme learning machine (ELM) model because it is lightweight and has a faster learning speed. “The others” class concept is proposed to enable the O-CIL capability of ELM. To implement both the S-ELM on edge devices, we propose the scalable QR-decomposition ELM (S-QRD-ELM) and its implementation with the TSMC 40nm technology. The S-QRD-ELM engine is designed with folding and hardware-sharing techniques for better energy efficiency. Finally, to collaboratively learn a good global model with local ELM models on the edge device, a federated ELM system (Fed-ELMS) is proposed. On the other hand, this dissertation presents a scalable N-pair loss Deep-ECG (SNL-Deep-ECG) system for deep-learning-based identity verification tasks on the cloud for better user experience, performance, and generalizability. The single-pulse preprocessing method shortens the signal collection time, thus improving the user experience. The N-pair loss is used to train the Deep-ECG model instead of the original cross-entropy loss for better performance. Finally, the mixed database validation scheme is proposed to validate the generalizability of the SNL-Deep-ECG. In summary, this dissertation presents advanced designs for ECG-based edge-centric user identification and cloud-based identity verification tasks. We expect the proposed frameworks to be the essential techniques for future ECG-based biometric systems. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90741 |
DOI: | 10.6342/NTU202300690 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 5.21 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。