Abstract: Accurate recognition of upper limb action intention in stroke patients is a key step towards efficient rehabilitation training. In order to improve the accuracy of upper limb action recognition based on surface electromyography (sEMG), a method is proposed that combines pre-trained models and support vector machine (SVM) classification. This method fully considers the correlation between channels and converts the preprocessed time-domain signal into corresponding spectrograms through short time Fourier transform (STFT), and concatenates the spectrograms of all channels in the vertical direction. Two fine-tuning pre-training models, VGG16 and Resnet50, are used to extract features from the EMG images. Three upper limb action recognition schemes are considered separately：using only fine-tuning pre-trained models for recognition, a single fine-tuning pre-trained model extracts features and uses SVM for recognition, and two fine-tuning pre-trained models extract feature concatenation and use SVM for recognition. The experimental results show that the proposed method achieves a recognition accuracy of over 90% on the collected subject EMG signal dataset, which can effectively differentiate between different upper limb action.

Key words: upper limb action recognition, surface electromyography (sEMG), short-time fourier transform (STFT), pre-trained models, support vector machine (SVM)

摘要： 准确识别脑卒中患者上肢运动意图是实现高效康复训练的关键步骤。为了提高基于表面肌电信号（surface electromyography，sEMG）的上肢动作识别精度，提出了一种结合预训练模型和支持向量机（support vector machine，SVM）的肌电动作识别方法。该方法充分考虑通道之间的关联性，将预处理后的时域信号通过短时傅里叶变换（short-time Fourier transform，STFT）转换为对应频谱图，并将所有通道的频谱图沿竖直方向拼接。利用两种微调的预训练模型VGG16和Resnet50对肌电图像提取特征，分别考虑三种上肢动作识别方案：仅使用微调的预训练模型进行识别、单个微调预训练模型提取特征后使用SVM进行识别、两个微调预训练模型提取特征拼接后使用SVM进行识别。实验结果表明，所提出的方法在采集的受试者肌电信号数据集上均达到90%以上的识别精度，可有效区分不同的上肢动作。

关键词: 上肢动作识别, 表面肌电信号（sEMG）, 短时傅里叶变换（STFT）, 预训练模型, 支持向量机（SVM）