標題: 使用遞迴型神經網路之肌電圖動態動作預測機制
An EMG-based Dynamic Motion Prediction Mechanism Using Recurrent Neural Networks
作者: 許佑綸
Yu-Lun Hsu
楊谷洋
Kuu-Young Young
電控工程研究所
關鍵字: 肌電圖;運動學;類神經網路;動作分析;人機介面;electromyogram;kinematics;artificial neural network;motion analysis;human-machine interface
公開日期: 2007
摘要: 由於肌電圖 (electromyography, EMG) 分析對於了解人體運動意圖非常有效,因此讓研究者們藉此開發了許多新型的EMG識別系統及EMG人機介面。在這些研究之中,有些已經可以利用EMG訊號成功分辨不同的固定姿態,並將這些技巧應用在操作義肢上。而在本研究中,我們計畫進一步辨識連續的動態動作,以建構一個可以接受更細微的運動命令的EMG人機介面,而截肢者也可以藉此操作義肢來執行與平常人一樣的流暢動作。研究的主要目標是識別在連續前臂運動的執行期間,EMG訊號和手臂運動學之間的關係,並藉由EMG訊號立即預測使用者所預定的前臂運動位置。為了學習帶有動態與非線性特性的EMG-運動學映射關係,我們提議使用一種稱為外部輸入非線性自動迴歸模型 (nonlinear autoregressive with exogenous input model, NARX model) 的動態遞迴型神經網路 (dynamic recurrent neural networks, DRNN) 與向量量化時間性聯想記憶 (vector-quantized temporal associative memory, VQTAM) 學習演算法。我們也執行了一個基於三個受測者與二自由度 (degree of freedom, DOF) 前臂動態運動的實驗,實驗結果則展示出在一系列的連續運動期間,所提議的方法的確能藉由EMG訊號立即建立目前的前臂位置。這個方法適合應用在靈巧的義肢控制上。
Electromyography (EMG) analysis is very effective for interpreting human motion intention. Therefore, many researchers develop novel EMG recognition systems and EMG human-machine interface. Among them, some studies have already succeeded in discriminating fixed postures using EMG signals and applied the techniques to manipulate the prosthesis. In this study, we plan to further recognize their continuous dynamic movements for constructing the EMG human-machine interface which can accept more delicate motion commands. In addition, the amputee can also use the prosthesis to execute smooth movements just like ordinary people. The goal of this study is to identify the relationship between EMG signals and arm kinematics during the execution of the continuous forearm movements, and predict the user’s intended forearm motion with EMG signals in a real time manner. To learn the EMG–kinematics mappings with the dynamic and nonlinear characteristics, we propose using the nonlinear autoregressive with exogenous input model (NARX model) which is a kind of dynamic recurrent neural network (DRNN) and the vector- quantized temporal associative memory (VQTAM) learning algorithm. We are carried out the experiment based on three subjects and two degrees of freedom (DOF) dynamic movements. Experiment results show that the proposed method is able to estimate the forearm positions using EMG signals during a series of continuous motions immediately. This method is ready for application on delicate prosthetic control.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009512537
http://hdl.handle.net/11536/38244
Appears in Collections:Thesis


Files in This Item:

  1. 253701.pdf