摘要：

目前, 大多数商用表面肌电(sEMG)信号采集系统存在价格昂贵、便携性和时效性无法满足应用需求的缺陷。为了解决该问题, 结合模拟采集电路的精度要求以及微控制器低功耗、高性能、灵活性等方面的需求, 设计一种嵌入式一体化肌电腕带, 可以用于实时手势识别。首先, 采用低成本、低噪声的精密放大器完成模拟采集电路设计, 并在仿真环境中对电路进行仿真验证, 保证信号采集质量; 其次, 在低功耗、高性能的微控制器ESP32-S3上提出轻量化卷积神经网络L-CNN以进行实时识别。L-CNN在预训练后进行剪枝和量化, 然后完成部署。通过剪枝算法去除模型中冗余的权重参数, 减小模型大小, 加快推理速度, 并微调到预训练模型中。量化将原有32位浮点数降到8位整数进行计算, 使模型尺寸降低以适用于嵌入式设备。实验结果表明, L-CNN的尺寸相比原模型大幅降低, 推理速度得到提升, 并且在实时手势识别中能达到95%左右的识别准确率, 验证了整个系统的可靠性。

关键词: 表面肌电信号, 手势识别, 微控制器, 模型剪枝, 模型量化

Abstract:

Currently, most commercial surface Electromyographic (sEMG) signal collection systems have the drawbacks of high cost, portability, and timeliness, which cannot meet application needs. To solve these problems, this study combines the accuracy requirements of analog acquisition circuits with low-power, high-performance, and flexibility requirements of microcontrollers to design an embedded integrated electromyographic wristband that can be used for real-time gesture recognition. First, a low-cost, low-noise precision amplifier was used to complete the design of the analog acquisition circuit, and the circuit was verified in a simulation environment to ensure the quality of signal acquisition. Second, a lightweight convolutional neural network L-CNN was developed for real-time recognition of the low-power and high-performance microcontroller ESP32-S3. A pruning algorithm was used to remove redundant weight parameters from the model, thereby reducing model size, which accelerates inference speed. After fine-tuning it into a pre-trained model, L-CNN was quantified prior to deployment. Quantification reduced the original 32 bit floating-point number to an 8 bit integer for calculation, thereby making the reduced model suitable for embedded devices. The experimental results show that the size of the L-CNN is significantly reduced compared to the original model, which improves inference speed, and a recognition accuracy of approximately 95% is achieved in real-time gesture recognition, verifying the reliability of the entire system.

Key words: surface Electromyographic (sEMG) signal, gesture recognition, microcontroller, model pruning, model quantification