基于正常人的肌电模式识别抗力变化的鲁棒性研究

Objective: In the practical application of electromyographic (EMG) pattern recognition, the change of force can affect the performance of the trained classifier, resulting in poor robustness. This paper proposes to combine the frequency domain and the spectral domain and extract four spectral domain features for classification to improve the robustness of the EMG pattern recognition classifier against force changes. Methods: In this experiment, 7 healthy subjects were recruited, and a cross-validation analysis was carried out on 15 fine hand movements at three different force levels. High-density EMG signals were used to extract four amplitude-independent features in the spectral domain in order to improve the robustness of EMG pattern recognition, and a comparative study was conducted with the method of extracting four traditional time-domain features. Results: By comparing the average classification results between the traditional time-domain features and the spectral domain features, the spectral domain features showed better robustness against force changes. The average classification result between the traditional time-domain features at the same force level was 89.31% ± 0.49%, and the average classification result between different force levels was 76.57% ± 3.59%; the average classification result between the spectral domain features at the same force level was 92.39% ± 0.26%, and the average classification result between different force levels was 86.94% ± 2.23%. Compared with the traditional time-domain features, the average classification result of the spectral domain features increased by 3.09% at the same force level and by 10.36% between different force levels. Especially when using the spectral domain features for low-intensity testing, the average classification result of the classifier was 70.97%, which was 20.45% higher than that of the traditional time-domain features. Conclusion: Compared with the traditional time-domain feature method, the spectral domain features have a significant comparative advantage. Extracting spectral domain features may become an effective method for motion intention recognition.

目的:在肌电模式识别的实际应用中,力的变化会影响训练分类器的性能,导致很差的鲁棒性。本文提出了联合频域和谱域,提取四个频谱域特征进行分类来提高肌电模式识别分类器抗力变化的鲁棒性。方法:本实验招募了7名健康受试者,对15个手部精细动作在三个不同力度水平下进行交叉验证分析。利用高密度肌电信号提取频谱域四个与幅值无关的特征,以便提高肌电模式识别的鲁棒性,并与提取四个传统时域特征方法进行对比研究。结果:通过比较传统时域特征和频谱域特征之间的平均分类结果,频谱域特征显示出对抗力变化更好的鲁棒性。传统时域特征同力度水平间平均分类结果为89.31%±0.49%,不同力度水平间平均分类结果为76.57%±3.59% ;频谱域特征同力度水平间平均分类结果为92.39%±0.26%,不同力度水平间平均分类结果为86.94%±2.23%。与传统时域特征相比,频谱域特征在同力度水平间平均分类结果增加了3.09%,在不同力度水平间平均分类结果增加了10.36%。特别当使用频谱域特征进行低强度测试时,分类器的平均分类结果为70.97%,比传统时域特征高20.45%。结论:与传统时域特征方法相比,频谱域特征具有显著的比较性优势。提取频谱域特征可能成为一种运动意图识别的有效方法。