Masters Matthew R, Smith Ryan J, Soares Alcimar B, Thakor Nitish V
Annu Int Conf IEEE Eng Med Biol Soc. 2014;2014:2577-80. doi: 10.1109/EMBC.2014.6944149.
Myoelectric control of prosthetic devices tend to rely on classification schemes of extracted features of EMG data. Those features however, may be sensitive to arm position resulting in decreased performance in real-world applications. The effect of varying limb position in a pattern recognition system have been illustrated by documenting the change in classification accuracy as the user achieves particular limb configurations. We continue to investigate this limb position effect by observing its impact on classification accuracy as well as through an analysis of how each extracted feature of the raw EMG varies in each position. Finally, LDA classification schemes are applied both to demonstrate the effect varying limb position has on classification accuracy and to increase classification accuracy without the use of additional hardware or sensors such as accelerometers as has been done in the past. It is shown that high classification accuracy can be achieved by (1) training an LDA classifier with data from many positions, as well as (2) by utilizing an extra position LDA classifier which can weigh the grasp classifiers appropriately. The classification accuracies achieved by these methods approached that of a model relying on a perfect knowledge of arm position.
假肢装置的肌电控制往往依赖于肌电数据提取特征的分类方案。然而，这些特征可能对手臂位置敏感，从而导致在实际应用中性能下降。通过记录用户达到特定肢体配置时分类准确率的变化，说明了模式识别系统中肢体位置变化的影响。我们继续通过观察其对分类准确率的影响以及分析原始肌电的每个提取特征在每个位置如何变化来研究这种肢体位置效应。最后，应用线性判别分析（LDA）分类方案来证明肢体位置变化对分类准确率的影响，并在不使用过去所采用的诸如加速度计等额外硬件或传感器的情况下提高分类准确率。结果表明，通过（1）使用来自多个位置的数据训练LDA分类器，以及（2）利用额外位置的LDA分类器来适当权衡抓握分类器，可以实现较高的分类准确率。这些方法所达到的分类准确率接近依赖于对手臂位置完全了解的模型的准确率。
