本文為瑞典查爾姆斯理工大學（作者：JOHAN BORGLIN）的訊號系統碩士論文，共68頁。

人工神經網路ANNs用於手臂肌電訊號（EMG）的分類。使用SmartHand專案的訊號放大器採集卡，從患者手臂採集16通道EMG訊號並進行濾波。從特徵提取中將肌電訊號進行時域分類後，利用簡單的反向傳播網路進行訓練。在訓練過程中，病人按預定義的手勢移動手指；訓練結束後，患者可以通過複製訓練過程中的動作來移動假手。

每根手指的運動對應一個神經網路的訓練，在任何給定的時間，當只有其中一根手指伸縮時，都可以獲得很好的效果。在大多數的測試中，患者能夠以這種方式使用所有的手指。（不能讓多根手指同時運動！！！）

額外的測試是在同一時間使用幾根手指，這種測試僅僅在某些情況下成功完成。

研究結果表明，人工神經網路可以作為類似SmartHand專案的分類器，但是如果需要完成高階運動，那麼在訓練網路時必須非常小心。

這項工作於2010秋季在瑞典隆德大學工程學院電氣測量系（LTH）進行。

Artificial neural networks (ANNs) an arm. Using a amplifiercard from the SmartHand project, 16-channel EMG signals were collected from thepatients arm and filtered. After time-domain were used to classify EMG signalsfrom feature extraction, simple back-propagation training was used to train thenetworks. During the training the patient moved his fingers according to apredefined patter. After the training, the patient could move an artificialhand by duplicating the movements made during training. One network wasimplemented for each finger, and good results were achieved when only onefinger was contracted at any given time. The patient was able to use all the fingersin this way in a majority of the tests. Additional tests were made with usingseveral fingers at the same time. This was done successfully only in some ofthe tries. The results show that ANNs are a possible candidate as a classifierfor a project like the SmartHand project, but that great care has to be takenwhen training the networks, if advanced motions are desired. The work wascarried out at the Department for Electrical Measurements, the Faculty ofEngineering (LTH), Lund University during thefall of 2010.

1 引言

2 專案背景

3 問題定義及侷限性

4 理論研究

5 方法設計

6 實驗結果及討論

7 結論

附錄A C++程式碼

下載英文原文地址：

更多精彩文章請關注微訊號：