利用域对抗神经网络改进基于sEMG的经桡骨截肢者运动意图识别能力

Improving sEMG-Based Recognition of Motor Intent in Transradial Amputees Using Domain Adversarial Neural Networks

摘要: 上肢肢体功能丧失在很大程度上降低了截肢患者生活质量，为他们提供智能化、高性能的假肢有助于恢复患者所丧失的肢体功能。目前，肌电控制假肢技术已然成为研究的热点。然而，当前商用肌电假肢的弃用率较高，主要原因包括缺乏准确的意图识别、不具备直观和方便的特性。文章深入研究了基于肌电信号的上肢运动模式识别技术，特别关注于上肢截肢患者的需求，旨在设计面向这一特定群体的智能识别方法，以提高识别性能并促进肌电假肢系统的实际应用。并提出基于域对抗适应的动作识别算法，针对不同个体的识别率差异问题，通过域对抗训练和自适应批量归一化来最小化不同域之间的差异，从而提高模型在目标域上的泛化能力和动作识别的精度。实验结果表明，所设计算法优于微调和卷积神经网络，在截肢受试者的识别任务中识别准确率最高达94.70%，实时实验中平均动作完成率为83.48%。因此，该算法能够有效地提升面向截肢受试者的识别性能。

Abstract: Loss of upper limb function largely reduces the quality of life of amputation patients, and providing them with intelligent, high-performance prostheses helps to restore the lost limb function of patients. At present, myoelectric control prosthesis technology has become a research hotspot. However, the current commercial myoelectric prostheses have a high abandonment rate, and the main reasons include the lack of accurate intent recognition and intuitive and convenient features. This paper presents an in-depth study of upper limb motion pattern recognition based on EMG signals, with a special focus on the needs of upper limb amputee patients, aiming to design intelligent recognition methods for this specific group to improve recognition performance and facilitate the practical application of EMG prosthetic systems. An action recognition algorithm based on depth domain adversarial adaptation is proposed. Aiming at the problem of recognition rate difference between different individuals, the differences between different domains are minimized by domain adversarial training and adaptive batch normalization, so as to improve the generalization ability of the model on the target domain and the accuracy of action recognition. The experimental results show that the designed algorithm outperforms the fine-tuning and convolutional neural network, and the recognition accuracy reaches up to 94.70% in the recognition task of amputee subjects. The average action completion rate in real-time experiments was 83.48%, respectively. Therefore, the algorithm can effectively improve the recognition performance of amputee subjects.