CN103699873A

CN103699873A - Lower-limb flat ground walking gait recognition method based on GA-BP (Genetic Algorithm-Back Propagation) neural network

Info

Publication number: CN103699873A
Application number: CN201310433056.1A
Authority: CN
Inventors: 马玉良; 马云鹏; 佘青山; 张启忠; 孟明
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2013-09-22
Filing date: 2013-09-22
Publication date: 2014-04-02

Abstract

The invention discloses a lower-limb flat ground walking gait recognition method based on a GA-BP neural network. The method comprises the following steps: performing denoising smoothing and time domain feature parameters extraction to the acquired lower-limb continuous flat ground waking four-way surface electromyogram signals to obtain a feature value sample set; then optimizing the BP neural network with the GA to obtain a group of complete initial weight values and threshold values with the minimum BP neural network deviation; randomly dividing the extracted feature values into a training sample group and a test sample group, and using the training samples to train the GA optimized BP neural network; at last, inputting the test sample in the trained BP neural network classifier to perform recognition and classification. By virtue of the lower-limb flat ground walking gait recognition method, the time domain features of the electromyogram signals are easy to extract and obvious and has good expression capability.

Description

Gait recognition method for lower limbs walking on level ground based on GA-BP neural network

技术领域technical field

本发明涉及一种人体运动模式识别方法，特别涉及基于下肢平地行走时肌电信号特征值的GA-BP神经网络步态识别的方法。The invention relates to a human body motion pattern recognition method, in particular to a GA-BP neural network gait recognition method based on the eigenvalues of electromyography signals when the lower limbs walk on level ground.

背景技术Background technique

步态体现出下肢行走运动的姿态，是下肢行走状态统称。它与人体结构与功能、运动协调组织、行为及心理活动都有着重要的关系，是人体生命活动中最基本的动作。常步态（normal gait）是指健康人体下肢有自己感觉最自然、最舒适的姿态行走时的步态，它具有周期性以及协调和均衡的特点。Gait reflects the posture of the lower limbs walking movement, and is a general term for the walking state of the lower limbs. It has an important relationship with the structure and function of the human body, movement coordination organization, behavior and psychological activities, and is the most basic action in the life activities of the human body. Normal gait refers to the gait in which the lower limbs of a healthy human being walk in the most natural and comfortable posture. It has the characteristics of periodicity, coordination and balance.

人体表面肌电信号是一种低频的微弱生物信号，在本质上是一种具有非平稳、非高斯特性的生理信号，在拾取、调理、采集过程中，不可避免地会引入许多干扰，通过合适的信号消噪、特征值提取和模式识别方法可区分下肢平地行走的步态。The human body surface EMG signal is a low-frequency weak biological signal, which is essentially a non-stationary and non-Gaussian physiological signal. During the process of picking up, conditioning, and collecting, it will inevitably introduce a lot of interference. The signal denoising, eigenvalue extraction and pattern recognition methods can distinguish the gaits of lower extremity walking on flat ground.

人体下肢步态模式识别方法的研究从上世纪90年代开始，取得了许多成果。例如Mumse与Saka提出了一种时空相关匹配的方法用于区别不同的步态；南安普顿大学的Foster等提出采用区域度量的方法解决步态识别问题。随着人工神经网路的发展，国内有人使用神经网络对肌电信号提取出的特征值进行分类，最后得到了较好的效果。然而现在大部分方法的共同理论基础是经典统计学，采用的是研究样本数目趋于无穷大时的渐进理论。然而在实际问题中，样本数目往往有限，故这些在理论上有显著长处的分类方法在实际应用中的表现却可能不尽人意，例如传统的BP神经网络分类容易出现局部最小值以及分类效果不理想。本发明采用一种遗传算法优化之后的BP神经网络来更好地进行模式识别和分类。The research on gait pattern recognition methods of human lower limbs began in the 1990s, and many achievements have been made. For example, Mumse and Saka proposed a spatiotemporal correlation matching method to distinguish different gaits; Foster of the University of Southampton et al proposed to use the method of regional measurement to solve the problem of gait recognition. With the development of artificial neural networks, some people in China use neural networks to classify the eigenvalues extracted from EMG signals, and finally get better results. However, the common theoretical basis of most methods is classical statistics, which adopts the asymptotic theory when the number of research samples tends to infinity. However, in practical problems, the number of samples is often limited, so the performance of these classification methods with significant advantages in theory may not be satisfactory in practical applications. For example, the traditional BP neural network classification is prone to local minima and the classification effect is not good. ideal. The invention adopts a BP neural network optimized by a genetic algorithm to better perform pattern recognition and classification.

发明内容Contents of the invention

本发明就是针对传统BP神经网络分类的不足，采用GA（遗传算法）优化BP神经网络的初始权值和阈值，使优化后的BP神经网络对肌电信号提取出的特征值进行识别分类，从而提高正确识别率。The present invention aims at the deficiencies of traditional BP neural network classification, adopts GA (genetic algorithm) to optimize the initial weight and threshold of BP neural network, and enables the optimized BP neural network to identify and classify the feature values extracted from electromyographic signals, thereby Improve the correct recognition rate.

本发明的目的可以通过以下技术方案实现：The purpose of the present invention can be achieved through the following technical solutions:

本发明包括以下步骤：The present invention comprises the following steps:

步骤1.对采集到的下肢连续平地行走动作的四路表面肌电信号进行消噪滤波和时域特征值提取，得到其特征向量样本集。Step 1. Perform denoising filtering and time-domain eigenvalue extraction on the collected four-channel surface EMG signals of the continuous walking action of the lower limbs to obtain its eigenvector sample set.

步骤2.用GA对BP神经网络进行优化，得到BP神经网络误差最小的一组完整初始权值和阈值。Step 2. Use GA to optimize the BP neural network to obtain a complete set of initial weights and thresholds with the smallest error in the BP neural network.

步骤3.将步骤1中提取的特征值随机分成训练样本和测试样本两组，并用训练样本来训练GA优化之后的BP神经网络；用测试样本输入训练好的BP神经网络分类器，进行识别分类。Step 3. Randomly divide the eigenvalues extracted in step 1 into two groups of training samples and test samples, and use the training samples to train the GA-optimized BP neural network; use the test samples to input the trained BP neural network classifier for identification and classification .

其中步骤1中四路肌电信号的消噪滤波采用空域相关滤波，具体步骤归结如下：Among them, the denoising filter of the four channels of EMG signals in step 1 adopts spatial correlation filtering, and the specific steps are summarized as follows:

（1）对含噪信号进行离散小波变换，对采集的原始肌电信号进行5层小波分解，基小波选用双正交样条小bior 1.5，得到尺度j上位置n处的含噪信号f的离散小波变换Wf(j,n)。(1) Discrete wavelet transform is performed on the noise-containing signal, and 5-layer wavelet decomposition is performed on the collected original EMG signal. The base wavelet is biorthogonal spline small bior 1.5, and the noise-containing signal f at position n on the scale j is obtained. Discrete wavelet transform Wf(j,n).

（2）求取各尺度与其相邻尺度的相关系数Corr₂(j,n)＝Wf(j,n)Wf(j+1,n)。(2) Calculate the correlation coefficient Corr ₂ (j,n)=Wf(j,n)Wf(j+1,n) between each scale and its adjacent scales.

（3）将Corr₂(j,n)归一化到Wf(j,n)的能量上去，得到归一化后的相关系数NewCorr₂(j,n)。计算方法为：(3) Normalize Corr ₂ (j,n) to the energy of Wf(j,n) to obtain the normalized correlation coefficient NewCorr ₂ (j,n). The calculation method is:

$New new {Corr Corr}_{22} ((j j,, n no)) = = {Corr Corr}_{22} ((j j,, n no)) \sqrt{{P P}_{W W} ((j j)) / / {P P}_{{Corr Corr}_{22}} ((j j))},, n no = = 1,2 1,2,, \cdot \cdot \cdot \cdot \cdot \cdot,, N N$

其中， $P_{{Corr}_{2}} (j) = Σ_{n = 1}^{N} {Corr}_{2} {(j, n)}^{2};$ $P_{W} (j) = Σ_{n = 1}^{N} Wf {(j, n)}^{2}$ in, $P_{{Corr}_{2}} (j) = Σ_{no = 1}^{N} {Corr}_{2} {(j, no)}^{2};$ $P_{W} (j) = Σ_{no = 1}^{N} w {(j, no)}^{2}$

（4）若NewCorr₂(j,n)＞Wf(j,n)，则认为n点处的小波系数值是有信号产生，将Wf(j,n)的值替换Wf_new(j,n)的相应位置，并将Wf(j,n)置零，Corr₂(j,n)置零；否则认为Wf(j,n)是由噪声产生的，保留Wf(j,n)。(4) If NewCorr ₂ (j,n)>Wf(j,n), it is considered that the wavelet coefficient value at point n is generated by a signal, and the value of Wf(j,n) is replaced by Wf _new (j,n) and set Wf(j,n) to zero, and Corr ₂ (j,n) to zero; otherwise, Wf(j,n) is considered to be generated by noise, and Wf(j,n) is retained.

（5）重复步骤(3)和(4)，直到P_W(j)满足某一噪声能量阈值（在空域相关滤波消噪中需要设定某一噪声阈值，本发明在实验中以信号的80个只含有噪声的点估计噪声在各层的方差，并以此10倍作为噪声能量阈值。为获取80个只含噪声的点，在各相关肌肉放松时采样80个点作为参考，认为此时信号为噪声）。这时Wf_new(j,n)保留了去除噪声后的小波系数。(5) Repeat steps (3) and (4) until P _W (j) satisfies a certain noise energy threshold (a certain noise threshold needs to be set in the spatial correlation filter denoising, the present invention uses 80% of the signal in the experiment A point that only contains noise estimates the variance of the noise in each layer, and 10 times this is used as the noise energy threshold. In order to obtain 80 points that only contain noise, 80 points are sampled as a reference when each relevant muscle is relaxed. It is considered that at this time signal to noise). At this time, Wf _new (j, n) retains the wavelet coefficients after noise removal.

（6）对Wf_new(j,n)进行小波重构就得到空域相关滤波后的信号。(6) Perform wavelet reconstruction on Wf _new (j,n) to obtain the signal after spatial correlation filtering.

提取的时域特征值为积分肌电值和绝对值方差。两个特征值的定义及提取如下：The extracted time-domain feature values are integral EMG value and absolute value variance. The definition and extraction of the two eigenvalues are as follows:

积分肌电值(Integrate EMG)，其计算式为：Integral EMG value (Integrate EMG), its calculation formula is:

${x x}_{iemg iemg} = = \frac{11}{N N} {Σ Σ}_{i i = = 00}^{N N - - 11} | | {x x}_{((i i))} | |$

其中i为每组的采样点数，x_(i)为表面肌电信号采样的数据点值，N为每组采样点数。Among them, i is the number of sampling points in each group, x _(i) is the data point value of surface electromyographic signal sampling, and N is the number of sampling points in each group.

绝对值方差(Variance)：对原始肌电信号进行取绝对值操作,而后再求取所得信号的方差,此种情况下的方差定义如下：Absolute value variance (Variance): The absolute value operation is performed on the original EMG signal, and then the variance of the obtained signal is calculated. In this case, the variance is defined as follows:

$VAR VAR = = \frac{11}{N N - - 11} {Σ Σ}_{i i = = 00}^{N N - - 11} {(({x x}_{((i i))} - - {x x}_{iemg iemg}))}^{22}$

四路信号共八个特征参数来构建一个特征向量。A total of eight characteristic parameters of the four signals are used to construct a characteristic vector.

其中步骤2所述的“用GA对BP神经网络进行优化，得到BP神经网络误差最小的一组完整初始权值和阈值”中的GA是一种并行随机搜索最优化方法。它把自然界“优胜劣汰,适者生存”的生物进化原理引入优化参数形成的编码串联群体中,按照所选择的适应度函数并通过遗传中的选择、交叉和变异对个体进行筛选,使适应度值好的个体被保留,适应度差的个体被淘汰,新的群体既继承了上一代的信息,又优于上一代。这样反复循环,直至满足条件。具体描述如下：Among them, the GA in the "optimize the BP neural network with GA to obtain a complete set of initial weights and thresholds with the smallest error of the BP neural network" described in step 2 is a parallel random search optimization method. It introduces the biological evolution principle of "survival of the fittest and survival of the fittest" in nature into the coded series population formed by optimized parameters, and screens individuals according to the selected fitness function through selection, crossover and mutation in genetics, so that the fitness value Good individuals are retained, and individuals with poor fitness are eliminated. The new group not only inherits the information of the previous generation, but also is better than the previous generation. This cycle is repeated until the condition is met. The specific description is as follows:

种群初始化：Population initialization:

个体编码方法为实数编码,每个个体均为一个实数串,由输入层与隐含层连接权值、隐含层阈值、隐含层与输出层连接权值以及输出层阈值4部分组成。个体包含了神经网络全部权值和阈值,在网络结构已知的情况下,就可以构成一个结构、权值、阈值确定的神经网络。The individual encoding method is real number encoding, and each individual is a real number string, which consists of four parts: input layer and hidden layer connection weight, hidden layer threshold, hidden layer and output layer connection weight, and output layer threshold. The individual contains all the weights and thresholds of the neural network. When the network structure is known, a neural network with a certain structure, weights and thresholds can be formed.

适应度函数的确定：Determination of the fitness function:

根据个体得到BP神经网络的初始权值和阈值,用训练数据训练BP神经网络后预测系统输出,把预测输出和期望输出之间的误差绝对值和E作为个体适应度值F,计算公式为：According to the initial weight and threshold of the BP neural network obtained by the individual, the BP neural network is trained with the training data to predict the output of the system, and the absolute value of the error between the predicted output and the expected output is taken as the individual fitness value F, and the calculation formula is:

$F f = = k k (({Σ Σ}_{i i = = 11}^{n no} abs abs (({y the y}_{i i} - - {o o}_{i i}))))$

式中,n为网络输出节点数;y_i为BP神经网络第i个节点的期望输出;o_i为第i个节点的预测输出;k为系数。In the formula, n is the number of network output nodes; y _i is the expected output of the i-th node of the BP neural network; o _i is the predicted output of the i-th node; k is the coefficient.

选择操作：Choose an action:

遗传算法选择操作有轮盘赌法、锦标赛法等多种方法,本发明选择轮盘赌法,即基于适应度比例的选择策略,每个个体i的选择概率p_i的计算公式为：The genetic algorithm selection operation includes multiple methods such as the roulette method and the tournament method. The present invention selects the roulette method, which is a selection strategy based on the fitness ratio. The calculation formula for the selection probability _p of each individual i is:

f_i＝k/F_i； $p_{i} = \frac{f_{i}}{Σ_{j = 1}^{N} f_{i}}$ f _i =k/F _i ; $p_{i} = \frac{f_{i}}{Σ_{j = 1}^{N} f_{i}}$

式中,F_i为个体i的适应度值,由于适应度值越小越好,所以在个体选择前对适应度值求倒数;k为系数;N为种群个体数目。In the formula, F _i is the fitness value of individual i, since the smaller the fitness value is, the better, so the reciprocal of the fitness value is calculated before individual selection; k is the coefficient; N is the number of individuals in the population.

交叉操作：Cross operation:

由于个体采用实数编码,所以交叉操作方法采用实数交叉法,a_kj为第k个染色体a_k的第j位，a_kj为第l个染色体a_l的第j位，两者在第j位交叉操作方法如下:Since the individual is encoded with real numbers, the crossover operation method adopts the real number crossover method, a _kj is the jth position of the kth chromosome a _k , a _kj is the jth position of the lth chromosome a _l , and the two are crossed at the jth position The operation method is as follows:

$\{\begin{matrix} {a a}_{kj kj} = = {a a}_{kj kj} ((11 - - b b)) + + {a a}_{ij ij} b b \\ {a a}_{lj lj} = = {a a}_{lj lj} ((11 - - b b)) + + {a a}_{kj kj} b b \end{matrix}$

式中,b是[0,1]间的随机数。In the formula, b is a random number between [0,1].

变异操作：Mutation operation:

选取第i个个体的第j个基因a_ij进行变异,变异操作方法如下:Select the j-th gene a _ij of the i-th individual to mutate, and the mutation operation method is as follows:

${a a}_{ij ij} = = \{\begin{matrix} {a a}_{ij ij} + + (({a a}_{ij ij} - - {a a}_{max max})) * * f f ((g g)) & r r &GreaterEqual; &Greater Equal; 0.5 0.5 \\ {a a}_{ij ij} + + (({a a}_{min min} - - {a a}_{ij ij})) * * f f ((g g)) & r r < < 0.5 0.5 \end{matrix}$

式中,a_max为基因a_ij的上界;a_min为基因a_ij的下界;f(g)＝r₂(1-g/G_max);r₂为一个随机数;g为当前迭代次数;G_max是最大进化次数;r为[0,1]间的随机数。这样，经过以上遗传运算，就得到了BP神经网络误差最小的一组完整初始权值和阈值。In the formula, a _max is the upper bound of gene a _ij ; a _min is the lower bound of gene a _ij ; f(g)=r ₂ (1-g/G _max ); r ₂ is a random number; g is the current iteration number ; G _max is the maximum number of evolutions; r is a random number between [0,1]. In this way, through the above genetic operations, a complete set of initial weights and thresholds with the smallest error of the BP neural network are obtained.

其中步骤2、3中所述的BP神经网络即误差反向传播的神经网络，其算法基本思想是梯度下降法。它采用梯度搜索技术，以期使网络的输出值与期望输出值的误差均方值为最小。对优化之后的BP神经网络进行训练，以方差值样本为例，用于训练BP网络包括输入层神经元、隐含层输出神经元和输出神经元。The BP neural network described in steps 2 and 3 is the neural network of error backpropagation, and the basic idea of its algorithm is the gradient descent method. It uses gradient search technology to minimize the mean square value of the error between the output value of the network and the expected output value. The optimized BP neural network is trained. Taking the variance value sample as an example, it is used to train the BP network including input layer neurons, hidden layer output neurons and output neurons.

BP网络的训练过程如下：正向传播是输入信号从输入层经隐含层传向输出层，若输出层得到了期望的输出，则学习算法结束；否则，转至反向传播。The training process of the BP network is as follows: Forward propagation is the input signal from the input layer to the output layer through the hidden layer. If the output layer obtains the desired output, the learning algorithm ends; otherwise, it goes to backpropagation.

网络学习算法如下：The network learning algorithm is as follows:

（1）前向传播：计算网络的输出(1) Forward propagation: calculate the output of the network

w_ij为输入层第i个神经元和隐含层第j个神经元的联接权值，x_i为输入层第i个神经元向输出层第j个神经元的输出，隐含层神经元的输入x_j为所有x_i的加权之和：w _ij is the connection weight of the i-th neuron in the input layer and the j-th neuron in the hidden layer, x _i is the output from the i-th neuron in the input layer to the j-th neuron in the output layer, and the hidden layer neuron The input x _j of is the weighted sum of all x _i :

${x x}_{j j} = = \underset{i i}{Σ Σ} {w w}_{ij ij} {x x}_{i i}$

隐含层层神经元的输出x_j′，采用S函数激发x_j：The output x _j ′ of neurons in the hidden layer uses the S function to excite x _j :

${x x}_{j j}^{' '} = = f f (({x x}_{j j})) = = \frac{11}{11 + + {e e}^{- - {x x}_{j j}}}$

则 $\frac{{&PartialD; x}_{j}^{'}}{{&PartialD; x}_{j}} = x_{j}^{'} (1 - x_{j}^{'})$ but $\frac{{&PartialD; x}_{j}^{'}}{{&PartialD; x}_{j}} = x_{j}^{'} (1 - x_{j}^{'})$

输出层神经元的输出x_l：The output x _l of the neurons in the output layer:

${x x}_{l l} = = \underset{j j}{Σ Σ} {w w}_{jl jl} {x x}_{j j}^{' '}$

网络第l个输出与相应理想输出

的误差e₁为：The lth output of the network and the corresponding ideal output

The error _e1 of is:

${e e}_{11} = = {x x}_{l l}^{00} - - {x x}_{l l}$

第p个样本的误差性能指标函数E_p为：The error performance index function E _p of the pth sample is:

${E E.}_{p p} = = \frac{11}{22} {Σ Σ}_{l l = = 11}^{N N} {e e}_{l l}^{22}$

其中N为网络输出层的个数；Where N is the number of network output layers;

（2）反向传播：采用梯度下降法，调整各层间的权值；权值的学习算法如下：(2) Backpropagation: Use the gradient descent method to adjust the weights between layers; the weight learning algorithm is as follows:

输出层第l个神经元和隐含层第j个神经元的联接权值w_jl学习算法为：The connection weight w _jl learning algorithm of the lth neuron in the output layer and the jth neuron in the hidden layer is:

${Δw Δw}_{jl jl} = = - - η η \frac{{&PartialD; &PartialD; E E.}_{p p}}{{&PartialD; &PartialD; w w}_{jl jl}} = = {ηe ηe}_{l l} \frac{{&PartialD; &PartialD; x x}_{l l}}{{&PartialD; &PartialD; w w}_{jl jl}} = = η η {e e}_{l l} {x x}_{j j}^{' '}$

w_ij(k+1)＝w_ij(k)+Δw_ij w _ij (k+1)=w _ij (k)+Δw _ij

其中w_ij(k)为第k次学习中w_ij的值，且where w _ij (k) is the value of w _ij in the kth learning, and

$\frac{{&PartialD; &PartialD; x x}_{l l}}{{&PartialD; &PartialD; w w}_{ij ij}} = = \frac{{&PartialD; &PartialD; x x}_{l l}}{{&PartialD; &PartialD; x x}_{j j}^{' '}} \cdot &Center Dot; \frac{{&PartialD; &PartialD; x x}_{j j}^{' '}}{{&PartialD; &PartialD; x x}_{j j}} \cdot &Center Dot; \frac{{&PartialD; &PartialD; x x}_{j j}}{{&PartialD; &PartialD; w w}_{ij ij}} = = {w w}_{jl jl} \cdot &Center Dot; \frac{{&PartialD; &PartialD; x x}_{j j}^{' '}}{{&PartialD; &PartialD; x x}_{j j}} \cdot &Center Dot; {x x}_{i i} = = {w w}_{jl jl} \cdot &Center Dot; {x x}_{j j}^{' '} ((11 - - {x x}_{j j}^{' '})) \cdot &Center Dot; {x x}_{i i}$

考虑上次权值对本次权值变化的影响，须加入动量因子a，此时的权值为：Considering the impact of the last weight on this weight change, the momentum factor a must be added, and the weight at this time is:

w_jl(k+1)＝w_jl(k)+Δw_jl+a(w_jl(k)-w_jl(k-1))w _jl (k+1)＝w _jl (k)+Δw _jl +a(w _jl (k)-w _jl (k-1))

w_ij(t+1)＝w_ij(t)+Δw_ij+a(w_ij(t)-w_ij(t-1))w _ij (t+1)＝w _ij (t)+Δw _ij +a(w _ij (t)-w _ij (t-1))

其中，η为学习速率，a为动量因子，k，t为学习次数，η∈[0,1],a∈[0,1]；Among them, η is the learning rate, a is the momentum factor, k, t is the number of learning, η∈[0,1], a∈[0,1];

训练好之后，输入测试样本进行识别分类。After training, input test samples for recognition and classification.

与现有技术相比，本发明具有以下优点：Compared with the prior art, the present invention has the following advantages:

1.肌电信号的时域特征容易提取、特征明显、具有良好表达能力。1. The time-domain features of EMG signals are easy to extract, with obvious features and good expressive ability.

2.采用GA优化之后的BP神经网络进行步态识别，识别精度高，误差率小。2. The GA-optimized BP neural network is used for gait recognition, with high recognition accuracy and low error rate.

附图说明Description of drawings

图1为肌电信号消噪前后对比；Figure 1 is the comparison before and after denoising the EMG signal;

图2为GA优化BP神经网络流程；Figure 2 is the process of GA optimizing BP neural network;

图3为BP神经网络结构示意图；Fig. 3 is the schematic diagram of BP neural network structure;

图4为本发明流程框图；Fig. 4 is a flow chart of the present invention;

图5为识别误差进化曲线；Fig. 5 is the recognition error evolution curve;

图1中(a)、(b)分别表示消噪前后的肌电信号，横轴为采样点数，纵轴为电压(uV)。从图1中可以看出空域相关滤波消噪后的肌电信号的信噪比明显得到了提高，同时肌电信号的边缘特征被较好的保留了下来，这为特征提取和模式识别率的提高创造了良好的条件。(a) and (b) in Figure 1 represent the EMG signals before and after denoising respectively, the horizontal axis is the number of sampling points, and the vertical axis is the voltage (uV). It can be seen from Fig. 1 that the signal-to-noise ratio of the EMG signal after denoising by spatial correlation filtering has been significantly improved, and the edge features of the EMG signal have been well preserved, which is an important factor for the improvement of feature extraction and pattern recognition rate. Improvement creates good conditions.

图2中左部分为GA算法部分，右部分为BP神经网络训练学习过程。The left part in Figure 2 is the GA algorithm part, and the right part is the BP neural network training and learning process.

图3中,X₁,X₂,…,X_m是BP神经网络的输入值，Y₁,Y₂,…,Y_n是BP神经网络的预测值,ω_ij和ω_jk为BP神经网络权值。In Figure 3, X ₁ , X ₂ ,...,X _m are the input values of BP neural network, Y ₁ , Y ₂ ,...,Y _n are the predicted values of BP neural network, ω _ij and ω _jk are the weights of BP neural network value.

具体实施方式Detailed ways

见图4，本发明一种GA优化BP神经网络对下肢平地行走步态识别方法，该方法包括以下步骤：See Fig. 4, a kind of GA optimization BP neural network of the present invention is to lower limb level ground walking gait recognition method, and this method comprises the following steps:

步骤1.对采集到的下肢连续平地行走动作的四路表面肌电信号进行消噪滤波和时域特征值提取得到其特征向量样本集。Step 1. Perform denoising filtering and time-domain eigenvalue extraction on the collected four-channel surface electromyography signals of the continuous walking action of the lower limbs to obtain its eigenvector sample set.

步骤2.用GA对BP神经网络进行优化，优化流程如图2所示，得到BP神经网络误差最小的一组完整初始权值和阈值。Step 2. Use GA to optimize the BP neural network. The optimization process is shown in Figure 2, and a complete set of initial weights and thresholds with the smallest error of the BP neural network are obtained.

步骤3.将步骤1中提取的特征值随机分成训练样本和测试样本两组，并用训练样本来训练GA优化之后的BP神经网络。用测试样本输入训练好的BP神经网络分类器，进行识别分类。Step 3. Randomly divide the feature values extracted in step 1 into two groups of training samples and test samples, and use the training samples to train the BP neural network after GA optimization. Use the test samples to input the trained BP neural network classifier for recognition and classification.

为使本发明的目的、技术方案和优点表达的更加清楚明白，下面结合附图及具体实施例对本发明再做进一步详细的说明。In order to make the object, technical solution and advantages of the present invention more clearly, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

本发明主要思想是通过GA优化BP神经网络，赋予了BP神经网络误差最小的一组完整初始权值和阈值。在此基础上结合肌电信号时域特征值提取容易、特征值明显有点，构建具有良好表达能力的特征向量。对优化后的BP神经网络进行训练和识别。The main idea of the present invention is to optimize the BP neural network through GA, and endow the BP neural network with a complete set of initial weights and thresholds with the smallest error. On this basis, it is easy to extract the time-domain eigenvalues of the EMG signal, and the eigenvalues are obvious, and the eigenvectors with good expressive ability are constructed. Train and identify the optimized BP neural network.

步骤1：具体实施方式如下：Step 1: The specific implementation method is as follows:

信号来源方面，本发明所用的肌电信号采集仪是美国Noraxon公司的MyoTrace400肌电信号采集仪。采集平地行走时大腿上最具代表性的4块肌肉的表面肌电信号，它们分别是大腿前侧的股内侧肌、大腿后侧的半腱肌、大腿与胯部相连的阔筋膜张肌与大腿内侧的长收肌。这4块肌肉分布在大腿的不同区域，在位置和信号区分度上都具有典型性。In terms of signal source, the electromyographic signal acquisition instrument used in the present invention is the MyoTrace400 electromyographic signal acquisition instrument of Noraxon Corporation in the United States. Collect the surface electromyographic signals of the four most representative muscles on the thigh when walking on flat ground, they are the vastus medialis muscle on the front side of the thigh, the semitendinosus muscle on the back side of the thigh, and the tensor fascia lata connected to the thigh and crotch And the adductor longus of the inner thigh. These 4 muscles are distributed in different regions of the thigh and are typical in both location and signal differentiation.

信号消噪滤波采用空域相关滤波（消噪前后的肌电信号波形如图1所示)，之后的特征值提取具体实现过程如下：The signal denoising filter adopts spatial correlation filtering (the EMG signal waveform before and after denoising is shown in Figure 1), and the specific implementation process of the subsequent feature value extraction is as follows:

求取第i种步态的四路表面肌电信号的积分肌电值I_ij、方差V_ij，i=1,2,3,4,5；j=1,2,3,4。Calculate the integral EMG value I _ij and the variance V _ij of the four-way surface EMG signals of the i-th gait, i=1,2,3,4,5; j=1,2,3,4.

则构建时域特征向量X_i＝{I_i1,V_i1,I_i2,V_i2,I_i3,V_i3,I_i4,V_i4}。Then construct time-domain feature vector X _i ={I _i1 , V _i1 , I _i2 , V _i2 , I _i3 , V _i3 , I _i4 , V _i4 }.

由上得出的各特征向量组成特征向量样本集。The eigenvectors obtained above form the eigenvector sample set.

步骤2：将BP神经网络设计三层BP网络结构(BP神经网络结构示意图如图3所示)，由于4路肌电信号每路都提取积分肌电值和方差两个特征值，则输入端有8个神经元，输出层具有5个神经元，隐含层神经元数目根据经验公式及多次运行分析我们设为15。Step 2: Design the BP neural network with a three-layer BP network structure (the schematic diagram of the BP neural network structure is shown in Figure 3). Since each of the 4 road myoelectric signals extracts two eigenvalues, the integral myoelectric value and the variance, the input end There are 8 neurons, the output layer has 5 neurons, and the number of hidden layer neurons is set to 15 according to the empirical formula and multiple running analysis.

式中n₁为隐含层神经元数目，n为输入层神经元数目，m为输出层神经元数目，常数a＝1～10。

In the formula, n ₁ is the number of neurons in the hidden layer, n is the number of neurons in the input layer, m is the number of neurons in the output layer, and the constant a=1~10.

步骤3：训练时均取学习速率η为0.15，动量因子a为0.05。输出分别用(00001)表示支撑前期，(00010)表示支撑中期，(00100)表示支撑后期，(01000)表示摆动前期，(10000)表示摆动后期。步骤1中采集到的特征向量样本一共1100组，从中随机选则800组用来训练，剩余300组用来测试，如表2所示识别率均达到98%以上，且遗传算法进化到第50代时其样本通过进化得到的神经网络测试的最小误差为0.00827，如图5所示。与未优化的BP神经网络识别效果（表1所示）相比更好，同时此识别效果在下肢步态识别领域处较高水平。Step 3: During training, the learning rate η is set to 0.15, and the momentum factor a is set to 0.05. The output uses (00001) to represent the early stage of support, (00010) to represent the middle stage of support, (00100) to represent the late stage of support, (01000) to represent the early stage of swing, and (10000) to represent the late stage of swing. A total of 1100 groups of feature vector samples were collected in step 1, 800 groups were randomly selected for training, and the remaining 300 groups were used for testing. As shown in Table 2, the recognition rate reached more than 98%, and the genetic algorithm evolved to the 50th The minimum error of the neural network test obtained through the evolution of its samples during the epoch is 0.00827, as shown in Figure 5. Compared with the unoptimized BP neural network recognition effect (shown in Table 1), the recognition effect is at a higher level in the field of lower limb gait recognition.

表1 BP神经网络对步态的识别结果Table 1 The recognition results of gait by BP neural network

表2 遗传算法优化BP神经网络对步态的识别结果Table 2 The recognition results of gait by genetic algorithm optimization BP neural network

Claims

1. A lower limb flat ground walking gait recognition method based on GA-BP neural network is characterized by comprising the following specific steps:

step 1, performing noise elimination filtering and time domain characteristic value extraction on four paths of collected surface electromyographic signals of the lower limb continuous and flat walking motion to obtain a characteristic vector sample set of the four paths of collected surface electromyographic signals;

step 2, optimizing the BP neural network by using GA to obtain a group of complete initial weights and thresholds with minimum errors of the BP neural network;

step 3, randomly dividing the characteristic values extracted in the step 1 into two groups of training samples and testing samples, and training the GA-optimized BP neural network by using the training samples; and inputting the test sample into the trained BP neural network classifier for identification and classification.

2. The GA-BP neural network-based lower limb flat walking gait recognition method of claim 1, wherein: the denoising and filtering in the step 1 adopts spatial domain correlation filtering, and specifically comprises the following steps:

(1) performing discrete wavelet transform on the noisy signal, performing 5-layer wavelet decomposition on the acquired original electromyographic signal, wherein a base wavelet selects a bior 1.5 of a bi-orthogonal spline to obtain discrete wavelet transform Wf (j, n) of the noisy signal f at a position n on a scale j;

(2) calculating the correlation coefficient Corr of each scale and its adjacent scales₂(j,n)

Corr₂(j,n)＝Wf(j,n)Wf(j+1,n)；

(3) Corr is prepared from₂Normalizing the (j, n) to the energy of Wf (j, n) to obtain a normalized correlation coefficient NewCorr₂(j,n)：

Wherein,

(4) if NewCorr₂(j, n) > Wf (j, n), the wavelet coefficient value at n point is considered to be signal generation, and the value of Wf (j, n) is substituted for Wf_new(j, n) and zero Wf (j, n), Corr₂(j, n) zeroing; otherwise, considering that Wf (j, n) is generated by noise, and reserving Wf (j, n);

(5) repeating steps (3) and (4) until P_W(j) Meets a certain noise energy threshold when Wf_new(j, n) the wavelet coefficients after the noise removal are retained;

(6) for Wf_newAnd (j, n) performing wavelet reconstruction to obtain a spatial correlation filtered signal.

3. The GA-BP neural network-based lower limb flat walking gait recognition method of claim 1, wherein: in the step 1, time domain characteristic values are extracted as an integral myoelectricity value and an absolute value variance, and the definition and extraction of the two characteristic values are as follows:

integral myoelectric value x_iemgThe calculation formula is as follows:

where i is the number of samples per group, x_(i)Sampling data point values of the surface electromyogram signals, wherein N is the number of sampling points in each group;

absolute value variance VAR: the operation of taking absolute value is carried out on the original electromyographic signal, then the variance of the obtained signal is obtained, and the variance under the condition is defined as follows:

the four signals have eight characteristic parameters to construct a characteristic vector.

4. The GA-BP neural network-based lower limb flat walking gait recognition method of claim 1, wherein: the optimization of the BP neural network by using the GA in the step 2 specifically comprises the following steps:

(1) population initialization: the individual coding method is real number coding, each individual is a real number string and consists of an input layer and hidden layer connection weight, a hidden layer threshold, a hidden layer and output layer connection weight and an output layer threshold 4; the individual comprises all weights and thresholds of the neural network, and under the condition that the network structure is known, the neural network with a determined structure, weight and threshold can be formed;

(2) fitness function determination

Obtaining initial weight and threshold of BP neural network according to individual, predicting system output after training BP neural network with training data, taking absolute value sum E of error between predicted output and expected output as individual fitness value F, and calculating formula as follows:

wherein n is the network outputNumber of nodes y_iExpected output of the ith node of the BP neural network, o_iK is a coefficient;

(3) selection operation

The genetic algorithm selection operation adopts a roulette method, namely a selection strategy based on fitness proportion and the selection probability p of each individual i_iThe calculation formula of (2) is as follows:

f_i＝k/F_i；

in the formula, F_iThe fitness value of the individual i is the smaller the better, so the reciprocal of the fitness value is calculated before individual selection, k is a coefficient, and N is the number of population individuals;

(4) crossover operation

Since the individuals adopt real number coding, the cross operation method adopts a real number cross method, a_kjIs the kth chromosome a_kJ-th bit of (a)_kjIs the l-th chromosome a_lThe operation method of the bit j of the first bit and the bit j of the second bit is as follows:

\{\begin{matrix} a_{kj} = a_{kj} (1 - b) + a_{ij} b \\ a_{lj} = a_{lj} (1 - b) + a_{kj} b \end{matrix}

wherein b is a random number between [0,1 ];

(5) mutation operation

Selecting the jth gene a of the ith individual_ijCarrying out mutation by the following operation method:

<math> <mrow> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>+</mo> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mi>max</mi> </msub> <mo>)</mo> </mrow> <mo>*</mo> <mi>f</mi> <mrow> <mo>(</mo> <mi>g</mi> <mo>)</mo> </mrow> </mtd> <mtd> <mi>r</mi> <mo>&GreaterEqual;</mo> <mn>0.5</mn> </mtd> </mtr> <mtr> <mtd> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>+</mo> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>min</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>)</mo> </mrow> <mo>*</mo> <mi>f</mi> <mrow> <mo>(</mo> <mi>g</mi> <mo>)</mo> </mrow> </mtd> <mtd> <mi>r</mi> <mo><</mo> <mn>0.5</mn> </mtd> </mtr> </mtable> </mfenced> </mrow> </math>

in the formula, a_maxIs gene a_ijA upper bound of_minIs gene a_ijF (g) r₂(1-g/G_max);r₂Is a random number, G is the current iteration number, G_maxIs the maximum number of evolutions, r is [0,1]]A random number in between; thus, a complete set of initial weights and thresholds with the minimum BP neural network error are obtained through the genetic operation.

5. A GA-BP neural network based lower limb flat walking gait recognition method according to any one of claims 1-4, characterized in that:

the BP neural network in the step 2 and the step 3 is a neural network with error back propagation, and the basic idea of the algorithm is a gradient descent method; it adopts gradient search technique to make the error mean square value of the output value of the network and the expected output value be minimum; training the optimized BP neural network, taking a variance value sample as an example, and using the variance value sample as an example for training the BP network to comprise input layer neurons, hidden layer output neurons and output neurons;

the training process of the BP network is as follows: the forward propagation is that the input signal is transmitted from the input layer to the output layer through the hidden layer, and if the output layer obtains expected output, the learning algorithm is ended; otherwise, turning to the reverse propagation;

the network learning algorithm is as follows:

(1) forward propagation: output of computing network

w_ijIs the connection weight, x, of the ith neuron of the input layer and the jth neuron of the hidden layer_iFor the output of the ith neuron of the input layer to the jth neuron of the output layer, the input x of the neuron of the hidden layer is hidden_jFor all x_iWeighted sum of (2):

output x of hidden layer neurons_j', exciting x with S function_j：

Then

<math> <mrow> <mfrac> <msubsup> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> <mo>′</mo> </msubsup> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> </msub> </mfrac> <mo>=</mo> <msubsup> <mi>x</mi> <mi>j</mi> <mo>′</mo> </msubsup> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <msubsup> <mi>x</mi> <mi>j</mi> <mo>′</mo> </msubsup> <mo>)</mo> </mrow> </mrow> </math>

Output x of output layer neurons_l：

Net ith output and corresponding ideal output

Error e of₁Comprises the following steps:

e_{1} = x_{l}^{0} - x_{l}

error performance indicator function E for the p-th sample_pComprises the following steps:

wherein N is the number of network output layers;

(2) and (3) back propagation: adjusting the weight between layers by adopting a gradient descent method; the learning algorithm of the weight is as follows:

connecting weight w of ith neuron of output layer and jth neuron of hidden layer_jlThe learning algorithm is as follows:

<math> <mrow> <msub> <mi>Δw</mi> <mi>jl</mi> </msub> <mo>=</mo> <mo>-</mo> <mi>η</mi> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>E</mi> </mrow> <mi>p</mi> </msub> <msub> <mrow> <mo>&PartialD;</mo> <mi>w</mi> </mrow> <mi>jl</mi> </msub> </mfrac> <mo>=</mo> <msub> <mi>ηe</mi> <mi>l</mi> </msub> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>l</mi> </msub> <msub> <mrow> <mo>&PartialD;</mo> <mi>w</mi> </mrow> <mi>jl</mi> </msub> </mfrac> <mo>=</mo> <mi>η</mi> <msub> <mi>e</mi> <mi>l</mi> </msub> <msubsup> <mi>x</mi> <mi>j</mi> <mo>′</mo> </msubsup> </mrow> </math>

w_ij(k+1)＝w_ij(k)+Δw_ij

wherein w_ij(k) For w in the kth learning_ijA value of (a), and

<math> <mrow> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>l</mi> </msub> <msub> <mrow> <mo>&PartialD;</mo> <mi>w</mi> </mrow> <mi>ij</mi> </msub> </mfrac> <mo>=</mo> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>l</mi> </msub> <msubsup> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> <mo>′</mo> </msubsup> </mfrac> <mo>·</mo> <mfrac> <msubsup> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> <mo>′</mo> </msubsup> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> </msub> </mfrac> <mo>·</mo> <mfrac> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> </msub> <msub> <mrow> <mo>&PartialD;</mo> <mi>w</mi> </mrow> <mi>ij</mi> </msub> </mfrac> <mo>=</mo> <msub> <mi>w</mi> <mi>jl</mi> </msub> <mo>·</mo> <mfrac> <msubsup> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> <mo>′</mo> </msubsup> <msub> <mrow> <mo>&PartialD;</mo> <mi>x</mi> </mrow> <mi>j</mi> </msub> </mfrac> <mo>·</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>=</mo> <msub> <mi>w</mi> <mi>jl</mi> </msub> <mo>·</mo> <msubsup> <mi>x</mi> <mi>j</mi> <mo>′</mo> </msubsup> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <msubsup> <mi>x</mi> <mi>j</mi> <mo>′</mo> </msubsup> <mo>)</mo> </mrow> <mo>·</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> </mrow> </math>

considering the influence of the previous weight value on the change of the current weight value, a momentum factor a needs to be added, and the weight value at this time is:

w_jl(k+1)＝w_jl(k)+Δw_jl+a(w_jl(k)-w_jl(k-1))

w_ij(t+1)＝w_ij(t)+Δw_ij+a(w_ij(t)-w_ij(t-1))

wherein eta is the learning rate, a is the momentum factor, k, t is the learning times, eta belongs to [0,1], and a belongs to [0,1 ];

after training, inputting a test sample for recognition and classification.