It is conducive to the application of sEMG signals in helping disabled people through combining wearable devices with deep learning. Therefore, design of sEMG gesture recognition system using deep learning based on wearable device is proposed in this paper. The system is mainly consisted of wearable sEMG acquisition device and sEMG gesture recognition method based on deep learning. In the wearable sEMG acquisition device, the sEMG signal sensor is mainly used to convert the human bioelectrical signal into an analog electrical signal. Then it can be acquired using an analog to digital converter. We also use 2.4 GHz wireless communication for data transmission, and use the micro-controller as the core of system control and data processing. In the sEMG gesture recognition method, we designed a model of sEMG signal gesture classification based on convolutional neural network (CNN). It can avoid omission of important feature information and improve accuracy of recognition, effectively. In the experimental part, we collected the sEMG signals of three different gestures using our own wearable sEMG acquisition device. Then, we trained and evaluated on the designed sEMG gesture recognition model using these data. A recognition accuracy of about 79.43% can be achieved in three gestures. Finally, we trained and tested the sEMG gesture recognition model on the Ninapro DB5 dataset and can reach about 74.51% accuracy on 52 gestures. In the case that there are more types of gestures recognized, our accuracy is still 5.02%, 6.61%, and 2.58% higher than Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Long Short Term Memory-CNN (LCNN), respectively. Also, the accuracy rate is 5.47% higher than SVM and Random Forests.

通过将可穿戴设备与深度学习相结合，有助于将sEMG信号应用于帮助残疾人。因此，本文提出了基于可穿戴设备的基于深度学习的sEMG手势识别系统设计。该系统主要由可穿戴式sEMG采集设备和基于深度学习的sEMG手势识别方法组成。在可穿戴式sEMG采集设备中，sEMG信号传感器主要用于将人体生物电信号转换为模拟电信号。然后可以使用模数转换器获取它。我们还将2.4 GHz无线通信用于数据传输，并将微控制器用作系统控制和数据处理的核心。在sEMG手势识别方法中，我们基于卷积神经网络（CNN）设计了sEMG信号手势分类模型。它可以有效地避免遗漏重要特征信息，提高识别精度。在实验部分，我们使用自己的可穿戴sEMG采集设备收集了三个不同手势的sEMG信号。然后，我们使用这些数据对设计的sEMG手势识别模型进行了训练和评估。三个手势可以实现约79.43％的识别精度。最后，我们在Ninapro DB5数据集上训练并测试了sEMG手势识别模型，在52个手势上可以达到约74.51％的准确度。在识别出更多类型的手势的情况下，我们的准确度仍然比线性判别分析（LDA），支持向量机（SVM），5.01％，6.61％和2.58％更高，和长期短期记忆CNN（LCNN）。而且，准确率比SVM和随机森林高5.47％。