請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88371
標題: | 應用U-Net與Transformer遮罩器於單通道表面肌電訊號除噪 Single-Channel Surface Electromyography Denoising Using U-Net with Transformer Masker |
作者: | 王貫蓁 Kuan-Chen Wang |
指導教授: | 葉丙成 Ping-Chen Yeh |
關鍵字: | 表面肌電訊號,表面肌電訊號除噪,單通道,深度類神經網路,U-Net,Transformer, surface electromyography,sEMG noise removal,single channel,deep neural network,U-Net,Transformer, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 表面肌電訊號(surface electromyography,下稱sEMG)是一種量測人體肌肉活動的生醫訊號。由於其非侵入式的特性,sEMG被廣泛應用於許多領域中,比如臨床醫療或人機界面等。在這些sEMG應用中,訊號除噪的程序非常關鍵,因為sEMG相當容易受雜訊干擾而失真。現今已存在諸多類型的單通道sEMG除噪演算法,不過這些方法仍有限制,比如除噪過程使sEMG失真或除噪效果限於特定雜訊類型或訊噪比等。有鑑於此,本論文選擇以深度學習為基礎開發單通道sEMG除噪演算法,期望藉由深度學習類型演算法較佳的非線性映射能力與資料驅動的特性,使除噪方法擁有更強大的訊號還原效果、強建性以及泛用性,並突破既存方法的限制。
本論文基於除噪自動編碼器的概念,以遮罩式的架構結合U-Net與Transformer編碼器,並提出了深度學習類型的單通道sEMG除噪演算法——U-Net+Transmask。該方法採用了除噪自動編碼器的架構,並使用多規模(multi-scale)的特徵進行sEMG訊號的還原。此外U-Net+Transmask的遮罩式設計(masking)提升了對於窄頻帶雜訊的除噪效果。實驗資料方面,本論文採用「非侵入性及可適性義肢資料庫」(Non-Invasive Adaptive Prosthetics database)中所收錄的資料作為乾淨的sEMG,並以五種常見的sEMG雜訊與乾淨的sEMG合成含噪sEMG。為了全面地評估演算法的除噪效果及對後端應用的影響,本實驗採用三個訊號品質類型指標與兩個訊號特徵的萃取誤差,並將U-Net+Transmask與現有的六種深度學習模型架構和五種既有的sEMG除噪演算法進行比較。 實驗結果顯示,本研究所提出的U-Net+Transmask是這些方法中最適合應用於sEMG除噪的深度學習模型。經與其他六種類神經網路架構比較,U-Net+Transmask除了整體除噪能力最佳以外,還擁有Transformer能夠平行運算的優勢。此外U-Net+Transmask的除噪能力明顯優於既有的單通道sEMG除噪演算法,不論於廣泛的輸入訊噪比或各式雜訊類型等不同除噪條件下均能維持其優勢。因此,本論文所提出的U-Net+Transmask提供了sEMG相關應用一個更有力且強健的除噪選項。 Surface electromyography (sEMG) is a type of biomedical signal that measures human muscle activity. sEMG is widely used in many applications because it is a non-invasive method. For these sEMG applications, the process of signal denoising is very critical, because sEMG can be easily distorted by noise, interference or artifacts. There are many types of single-channel sEMG denoising algorithms, but these methods still have certain limitations, such as the signal may be distorted during the denoising process or the denoising effect is limited to specific noise types and signal-to-noise-ratios. To handle these difficulties, this paper utilizes deep learning to develop a single-channel sEMG denoising algorithm. Deep-learning-based methods have powerful nonlinear mapping capability and data-driven characteristics, so the denoising algorithms based on deep learning can usually have a better, more robust and generalized denoising performance, which may conquer the problems of the existing sEMG denoising algorithms. This paper proposes a single-channel sEMG denoising algorithm, U-Net+Transmask, based on deep learning. It adopts the structure of denoising autoencoder and combines U-Net and a Transformer encoder by a masking architecture. The multi-scale features are beneficial for signal reconstruction, and the masking architecture further improves denoising performance on narrow-band noise. In terms of experimental data, this paper uses the sEMG in the Non-Invasive Adaptive Prosthetics database as clean sEMG, and the noisy sEMG data is synthesized by superimposing five common sEMG noise types to clean sEMG. In order to comprehensively evaluate the denoising effect of the algorithms and their impact on back-end applications, this experiment adopts three and two evaluation criteria related to signal quality and feature extraction errors respectively. Besides, six other deep learning model and five existing sEMG denoising algorithms are implemented for comparison with U-Net+Transmask. The experimental results show that U-Net+Transmask is the most suitable deep learning model for sEMG denoising among these methods. Compared with the six other types of neural network structure, U-Net+Transmask not only has the best noise removal capability overall, but it also has the advantage of parallel computations from Transformer. In addition, the denoising ability of U-Net+Transmask is significantly better than that of the existing single-channel sEMG denoising algorithms, and it can maintain its advantages under various denoising conditions such as a wide range of input signal-to-noise ratios or various noise types. Therefore, the proposed U-Net+Transmask provides a powerful and robust denoising option for sEMG applications. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88371 |
DOI: | 10.6342/NTU202302087 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 此日期後於網路公開 2024-07-27 | 3.23 MB | Adobe PDF |
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。