基于深度神经网络的肌电信号降维与分类方法

发布时间：2018-04-14 09:23

本文选题：表面肌电信号 + 神经网络　；参考：《东华大学》2017年硕士论文

【摘要】：表面肌电信号(sEMG)是通过电极引导记录下来的神经肌肉系统活动时的生物电信号,能够实时反应肌肉活动状态,被广泛应用于助残、康复医学、运动医学等领域。由于肌电信号维度很高,在表面肌电信号的模式识别的相关研究中,需要对数据进行降维。本研究采集了7200条6类手部动作的表面肌电信号,用于研究表面肌电信号的降维和动作识别问题。本文从神经网络层数、隐含层节点数、激活函数、分类模型等方面进行了详细的讨论,在此基础上构建了一个5层神经网络模型对表面肌电信号进行了降维与分类。该模型将维度为3000的原始数据作为神经网络的输入,依次将数据从3000维度降维至500维、100维、6维。为了解决高维度输入下算法溢出问题,在模型最后一层采用6个二元分类器,完成对信号的分类,实验表明分类准确率达到96.2%。论文使用了第二范数正则项解决了过拟合的问题,同时详细说明了如何设定合适的参数使得算法能够正常收敛,并达到一个较好的泛化能力。在模型的参数求解过程中,由于输入层的数据维度较高且存在大量的矩阵运算,导致参数求解的计算过程过慢。论文提取了参数求解过程中相同的计算步骤(矩阵点乘与激活函数运算),运用矩阵运算并行化的思想对样本数据进行分割,将样本数据划分到相应的计算节点中,并行运行两个计算过程,再对各个计算节点的计算结果进行归并,从而加快了计算速度。论文最后对该神经网络模型分层提取特征值的过程进行了详细分析并利用类间离散度与类内离散度这两个指标对模型分层提取的特征值进行了评估,实验结果表明随着模型层数的增加,类间离散度与类内离散度的比值也不断增大,这表明该模型提取的特征值是有利于分类的。
[Abstract]:Surface electromyography (SEMG) is a bioelectric signal recorded in the neuromuscular system under the guidance of electrodes, which can respond to the state of muscle activity in real time. It is widely used in the fields of disability, rehabilitation medicine, sports medicine and so on.Because of the high dimension of EMG signal, it is necessary to reduce the dimension of the data in the research of pattern recognition of surface EMG signal.In this study, the surface electromyography (EMG) signals of 7200 hand movements of 6 kinds were collected, which were used to study the demotion and motion recognition of SEMG signals.In this paper, the number of neural network layers, the number of hidden layer nodes, the activation function and the classification model are discussed in detail. On the basis of this, a five-layer neural network model is constructed to reduce the dimension and classify the surface EMG signals.In this model, the original data with dimension 3000 is taken as the input of neural network, and the data is reduced from 3000 dimension to 100D / 100D / 6D in turn.In order to solve the problem of algorithm overflow in high dimensional input, six binary classifiers are used in the last layer of the model to classify the signals. The experimental results show that the classification accuracy reaches 96.22.In this paper, the second norm canonical term is used to solve the problem of over-fitting. At the same time, how to set appropriate parameters to make the algorithm converge normally and achieve a better generalization ability is explained in detail.In the process of solving the parameters of the model, the calculation process of the parameters is too slow because of the high data dimension in the input layer and the existence of a large number of matrix operations.In this paper, the same calculation steps (matrix point multiplication and activation function operation) are extracted in the process of parameter solving, and the sample data is divided into corresponding computing nodes by using the idea of parallelization of matrix operation.Two computing processes are run in parallel, and the results of each computing node are merged, thus speeding up the calculation speed.At the end of the paper, the process of extracting the eigenvalues from the hierarchical model is analyzed in detail, and the eigenvalues extracted from the hierarchical models are evaluated by using the two indexes of inter-class dispersion and intra-class dispersion.The experimental results show that the ratio of inter-class dispersion to intra-class dispersion increases with the increase of the number of model layers, which indicates that the eigenvalues extracted by the model are favorable for classification.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TN911.7

【参考文献】