|Place of Conferral||中国科学院自动化研究所|
|Keyword||脉冲神经网络 表面肌电信号 手部康复动作识别 学习算法|
我国有超过70%的脑卒中患者面临手部运动功能障碍问题，利用手部康复机器人辅助患者进行手部康复治疗具备广泛的应用前景。依据康复医学指导，手部康复机器人应当能够以表面肌电信号 (sEMG) 作为人机接口信号，辅助患者进行康复训练。基于上述要求，本文研究了计算能力强、运行功耗低、适合处理生理信息的脉冲神经网络 (SNN)，探索了SNN在处理sEMG信号领域的应用。研究内容主要包括SNN的模型、编码方法和学习算法，通过SNN实现高精度的手部康复动作识别。本文的主要工作和创新点归纳如下：
3. 提出了脉冲时间的快速搜索算法，通过压缩搜索脉冲区间，使得脉冲神经网络的迭代步数降低了40%。提出了多脉冲梯度下降学习算法，进一步提高了脉冲神经网络的性能，在手部康复动作识别任务中取得了 97.4%的准确率。本文针对现有脉冲神经网络的一些不足，例如运行速度慢和网络训练速度缓慢等问题，提出了一些改进算法。在提高脉冲神经网络的运行速度方面，本文提出了一种快速脉冲时间搜索算法，通过分解膜电势函数，压缩了搜索脉冲的区间，加速了脉冲神经网络的运行效率。在学习算法方面，本文放松了近似梯度下降学习算法的单脉冲约束条件，并引入变学习率方法，加速了训练过程。另外，本文基于自编码器结构，对sEMG信号的每个特征设计了子网络，缓解了神经元不激活现象。
In China, more than 70% of stroke patients are faced with hand dysfunction, and the use of hand rehabilitation robot to assist patients in hand rehabilitation has a broad application prospects. According to the guidance of rehabilitation medicine, the hand rehabilitation robot should be able to use the sEMG signal as the human-machine interface signal to assist patients in rehabilitation training. Based on the above requirements, this thesis investigates the spiking neural network (SNN), which has the strong computing capability, low running power consumption and is suitable for processing physiological information. Then the application of SNN in sEMG signal processing is explored. The research content mainly includes the spiking neural network model, coding method and learning algorithm, and realizes high-precision hand rehabilitation gesture recognition through the spiking neural network. The main contents and innovations of this thesis are summarized as follows:
1. The sEMG signal feature extraction method based on the three-dimensional reservoir structure spiking neural network is designed. The hand rehabilitation gesture recognition is realized by using linear regression, RO learning algorithm and PSD learning algorithm, with the highest recognition accuracy of 93.13%. This thesis firstly designs the AER encoding method for the sEMG signal. Then a 10*10*10 three-dimensional reservoir is designed, which can unsupervised learn features in the sEMG signal and realize the automatic extraction of sEMG signal features. Finally, the STDP learning rule of biological neural network are introduced into the reservoir to further improve the feature extraction ability of the network.
2. The hand rehabilitation gesture recognition method based on the fully connected spiking neural network has been studied in this thesis, and five kinds of time-frequency features of sEMG signals are extracted. The accuracy of hand rehabilitation gesture recognition is 96.5%. In order to further improve the recognition accuracy of the hand rehabilitation gesture, this thesis studies the multi-layer fully connected spiking neural network. Five kinds of time-frequency features of sEMG signals that perform well in hand gesture recognition are firstly extracted in this thesis and encoded the above features into spike sequences by using the population coding method. Finally, a three-layer fully connected ispiking neural network is trained by the approximate gradient descent learning algorithm, and the hand rehabilitation gesture recognition is realized.
3. A rapid spike time search algorithm is proposed, which compresses the spike search interval and reduces the number of iteration steps of the spiking neural network by 40%. A multi-spike gradient descent learning algorithm is proposed to further improve the performance of the spiking neural network, , and achieved an accuracy of 97.4% in the hand rehabilitation gesture recognition task. In this thesis, some improved algorithms are proposed to overcome the shortcomings of the existing spiking neural network, such as slow running speed and slow network training rate. In order to improve the running speed of the spiking neural network, this thesis proposes a rapid spike time search algorithm, which compresses the interval of search spike by decomposing the membrane potential function and accelerates the operation speed of the spiking neural network. In terms of learning algorithm, this thesis relaxes the single spike constraint condition of the approximate gradient descent learning algorithm, and introduces variable learning rate method to accelerate the training process. In addition, based on the autoencoder structure, this thesis designs a sub-network for each sEMG feature, which alleviates the phenomenon of neuron inactivation.
|刘洋. 基于脉冲神经网络的手部康复动作识别[D]. 中国科学院自动化研究所. 中国科学院大学,2019.|
|Files in This Item:|
|Recommend this item|
|Export to Endnote|
|Similar articles in Google Scholar|
|Similar articles in Baidu academic|
|Similar articles in Bing Scholar|
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.