CN104107042A - Electromyographic signal gait recognition method based on particle swarm optimization and support vector machine - Google Patents

Abstract

The invention relates to an electromyographic signal gait recognition method based on particle swarm optimization and a support vector machine. A particle swarm optimization algorithm is utilized to optimize a penalty parameter and a kernel function parameter of the support vector machine so that the performance of the support vector machine can be optimized, and effective recognition and classification are achieved. Firstly, wavelet modulus maximum denoising is carried out on collected lower limb electromyographic signals; secondly, time domain feature extraction is conducted on the electromyographic signals after denoising is carried out to obtain feature samples; thirdly, parameter optimization is carried out on the support vector machine by means of the particle swarm optimization algorithm to obtain a set of optimal parameters with minimal errors, and a classifier is constructed; at last, a feature sample set of the electromyographic signals is input to the classifier, and then classification and recognition are conducted on gait states. According to the method, both accuracy and adaptivity of classification are taken into consideration, the computational process is simple and efficient, and the method has broad application prospects in the field of lower limb motion state recognition.

Description

Electromyographic signal gait recognition method based on particle group optimizing-support vector machine

Technical field

The invention belongs to area of pattern recognition, relate to a kind of electromyographic signal recognition methods, particularly a kind of electromyographic signal gait recognition method for lower limb walking.

Background technology

Body gait refers to the attitude showing when people walks.Under the normal walking states hocketing at two lower limbs, gait has periodicity, harmony, balanced feature.In a gait cycle, the situation that contacts according to vola with ground, can be divided into gait and support phase (foot contacts to earth) and swing phase (foot is liftoff).Two phase places can be subdivided into again again and support early stage, the mid-term of support and support the later stage, and swing early stage and totally 5 stages in later stage of swing.Gait is the factors comprehensive outward manifestations when walking such as people self physiological structure, motor function, health status, behavioural habits.

Gait motion mainly relies on leg muscle to coordinate to drive lower limb muscles-skeletal system to complete, and electromyographic signal (EMG) is the action potential sequence producing during by muscle contraction, it with the active state of muscle and the functional status of motion between exist to a certain degree associated, the difference between different limb actions can embody by the feature of electromyographic signal.To the analysis of these features, likely, for distinguishing the lower limb different gaits in when walking, this is particularly important in the field such as sports medical science and rehabilitation medicine.

Lower limb electromyographic signal gait motion Study of recognition starts from the nineties in last century, has now obtained extensive concern.Existing electromyographic signal algorithm for pattern recognition comprises the method based on neutral net, the method based on hidden Markov model and based on Bayesian method etc.The people such as Abel E.W. have used LVQ Networks, self-organizing feature map neural network, and improved BP neutral net is carried out pattern recognition to human action, people's researchs such as Rabiner L.A. show that hidden markov model approach changes and provides a good probability description, the people such as Luo Zhizeng to utilize bayes method to upper limb electromyographic signal discriminator seasonal effect in time series.These researchs have all obtained certain achievement, but electromyographic signal is as a kind of small-signal of non-stationary, how to strengthen discriminator ability, improve the discrimination in gait action, remain the difficult problem that this research faces.

Along with the development of support vector machine (SVM), the feature that people bring into use support vector machine to extract electromyographic signal is classified, and its recognition effect is better than traditional method.Abroad, the people such as Naik G.R. extract the FRACTAL DIMENSION feature of all kinds of action electromyographic signals of hand, and then each action of combination supporting vector machine identification hand, has obtained higher discrimination.The method domestic, the people such as Yang Peng utilize hidden Markov model to combine with support vector machine, has effectively identified the motor pattern of standing and walking.The advantage of support vector machine is, has overcome traditional neutral net and has easily been absorbed in local minimum and the undesirable shortcoming of classifying quality.

But in practical problem, the performance quality of support vector machine depends primarily on the structure of kernel function and the selection of parameter thereof.Conventionally the way of parameter selection method is, within the scope of certain parameter, each parameter, according to certain interval value, is carried out to the different horizontal combination of parameters, form the alternative parameter combinations of many groups, select to make one group of expected risk upper bound minimum as optimal value of the parameter.The shortcoming of this parameter selection method, the one, be subject to the restriction of data scale, the 2nd, optimization method is quite consuming time, and is difficult to accurately find optimized parameter.Therefore, the present invention proposes a kind of particle group optimizing (PSO) algorithm, and self adaptation particle, according to its global search of practical situation dynamic equilibrium and local search ability, can find the optimized parameter of support vector machine rapidly and accurately, effectively carries out Gait Recognition.

Summary of the invention

The object of the invention is for the existing support vector machine deficiency that parameter is selected in electromyographic signal Gait Recognition, adopt PSO particle swarm optimization algorithm, punishment parameter and the kernel functional parameter of Support Vector Machines Optimized, the eigenvalue that uses the support vector machine after optimizing to extract electromyographic signal carries out discriminator, thus the correct recognition rata of gait while improving walking.

In order to realize above object, the inventive method comprises the following steps:

Step (1), the sample acquisition of lower limb electromyographic signal.Press muscle group effect and the contribution of gait action different phase when walking, and the sensitivity of effects on surface electromyographic signal collection equipment, select representative four muscle on thigh, the original electromyographic signal of muscle when the lower limb that pick up by myoelectricity Acquisition Instrument walking.Adopt Wavelet Modulus Maxima denoising method, first electromyographic signal is carried out to wavelet decomposition, then according to the singularity of wavelet coefficient, utilize signal from noise mode maximum the different variation characteristics on wavelet scale, isolate signal and noise, the electromyographic signal sample data after last reconstruct de-noising.Realize carrying out denoising Processing containing noisy electromyographic signal, remove white noise and retain singular point information.Concrete grammar is as follows:

First, signals and associated noises is carried out to wavelet transform.First the original electromyographic signal gathering is carried out to 4 layers of wavelet decomposition, base small echo is selected tight support biorthogonal wavelet sym8, obtains the wavelet transform Wf (s, x) of the signals and associated noises f at x place, the upper position of yardstick s.

Secondly, ask for the Lip index of signals and associated noises.The singular point of signal is exactly the catastrophe point in signal, and Lip index is that the one of the local singular point feature of characterization signal is measured, and is defined as follows:

If positive integer n,, if there is positive integer A > 0 and polynomial of degree n p in n≤α≤n+1 _n(x), make

|f(x)-p _n(x-x ₀)|≤A|x-x ₀| ^α (1)

For x ∈ (x ₀-δ, x ₀+ δ) set up, claim f (x) at x ₀point is Lip α.α is larger, and the smoothness of this point is higher; α is less, and the singularity of this point is larger.

Again, remain with and use signaling point.The Lip index of signal f (x) and Wavelet Modulus Maxima (Wavelet Modulus Maxima need meet | Wf (s, x) |≤| Wf (s, x ₀), x ₀for the local model maximum value point of wavelet transformation under yardstick s) need to meet

${\log }_2\left|W_{2^j}f\left(t\right)\right|\≤{\log }_{2}k+\mathrm{j\α}---\left(2\right)$

Wherein, t is the time, and j is wavelet scale, k ∈ R ⁿ.

To general signal α >=0, the modulus maximum of wavelet transformation increases the increase along with yardstick j; And for white noise α < 0, its modulus maximum reduces along with the increase of yardstick j.Therefore utilize the rule of wavelet modulus maxima conversion between different scale, the point (extreme point of corresponding noise) that removal amplitude reduces with the increase of yardstick, the point (extreme point of corresponding useful signal) that reservation amplitude increases with yardstick.

Finally, to Wf (s, x ₀) carry out wavelet reconstruction, obtain the electromyographic signal sample after Wavelet Modulus Maxima denoising.

Step (2), the feature extraction of lower limb electromyographic signal.The electromyographic signal of obtaining for step (1), by extracting its integration myoelectricity value, absolute value variance temporal signatures, obtains the sample set of characteristic vector.Concrete grammar is as follows:

The calculating formula of integration myoelectricity value is

$I_i=\frac{1}{N}\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\left|x_{\left(i\right)}\right|---\left(3\right)$

Here, x (i), i=0,1,2 ..., N-1 is the electromyographic signal time samples sequence of a length N, i is every group of sampling number.

Absolute value variance is defined as follows:

$N_i=\frac{1}{N-1}\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{(x_{\left(i\right)}-I_i)}^2---\left(4\right)$

Wherein I _ifor the integration myoelectricity value of electromyographic signal.

Thus, obtain eight characteristic parameters of four tunnel electromyographic signals, form a stack features vector Z.

Step (3), the grader of structure based on PSO-SVM.SVM is carried out to optimization of parameter choice with PSO, obtain one group of punishment parameter and the kernel functional parameter of SVM error minimum.Concrete grammar is as follows:

One, particle group optimizing (PSO) algorithm.PSO gains enlightenment from biotic population behavior characteristics and for solving-optimizing problem, in algorithm, each particle represents a potential solution of problem, with position, speed and three index expression particle characteristicses of fitness value.The corresponding fitness value being determined by fitness function of each particle.The speed of particle has determined direction and the distance that particle moves, and speed is dynamically adjusted with the mobile experience of self and other particles, thereby realizes individual optimizing in can solution space.

In iterative process each time, particle upgrades speed and the position of self by individual extreme value and global extremum, and more new formula is as follows:

$V_{\mathrm{id}}^{k+1}=w{V}_{\mathrm{id}}^k+c_1{r}_1(P_{\mathrm{id}}^k-X_{\mathrm{id}}^k)+c_2{r}_2(P_{\mathrm{gd}}^k-X_{\mathrm{id}}^k)$

(5)

$X_{\mathrm{id}}^{k+1}=X_{\mathrm{id}}^k+V_{\mathrm{id}}^{k+1}$

In formula, w is inertia weight, and the scope that larger w is suitable for separating is detected on a large scale, and less w is suitable for solution scope to carry out less detecting; D=1,2 ..., D is space dimensionality, i=1, and 2 ..., n is population; K is current iteration number of times; be illustrated in the flight speed of i particle, represent the position of i particle; with represent respectively individual optimal solution and the globally optimal solution of i particle.Here the change in location scope of d dimension is velocity variations scope is if X in iteration _idthe value of being beyond the boundary, is made as boundary value-x _maxor x _max.

C ₁and c ₂for the study factor (or accelerated factor), represent respectively the weights of each particle being pushed to the statistics acceleration term of individual extreme value and global extremum position, can be in the interior middle value of [0,4] scope; r ₁and r ₂for being distributed in the random number between [0,1]; Population size is the selection of population quantity size, generally adopts empirical method to get 20～40, can get 100～200 to problem more difficult or particular category; Maximum iteration time maxgen is optimizing end condition.

Two, support vector machine (SVM).The basic thought of SVM is by the more feature space of higher-dimension of DUAL PROBLEMS OF VECTOR MAPPING to, sets up a face that makes different types of data point interval maximum, i.e. largest interval hyperplane in this space.By setting suitable kernel function K (z _i, z) carry out nonlinear transformation the input space is transformed to a higher dimensional space, then in this new space, ask for optimum linearity classifying face.The output category decision function of SVM is:

$f\left(Z\right)=\mathrm{sgn}(\mathrm{wz}+b)=\mathrm{sgn}(\underset{i=1}{\overset{s}{\mathrm{\&Sigma;}}}{a}_i{y}_{i}K(z_i,z)+b),0\&le;a_i\&le;C---\left(6\right)$

A _ifor the corresponding Lagrangian coefficient of each training sample; K (z _i, z) be core letter (K (z _i, z)=exp (| z _i-z|^2/g^2)); C is punishment parameter, and b ∈ R is biasing.

(1) error punishment parameters C, divides sample proportion and algorithm complex to trade off to mistake, in definite proper subspace, regulates study machine fiducial range and empiric risk ratio, makes generalization ability (to the prediction judgement of data) best.It is chosen by concrete problem and determines, and depends on the quantity of noise in data;

(2) kernel functional parameter, different IPs function has impact to classification performance, and identical kernel function different parameters also has impact.For SVM, thereby the change of nuclear parameter has impliedly changed mapping function and has changed the complexity (dimension) of sample in data Subspace Distribution, a given kernel function (kernel function type and kernel functional parameter are determined), it is a corresponding data subspace with definite dimension just, also just limits optimum SVM corresponding to this subspace.

Three, the SVM parameter optimization based on PSO.The present invention selects radial basis kernel function, and it is exactly punishment parameters C and the kernel function radius parameter g of Support Vector Machines Optimized that PSO optimizes SVM, makes the SVM after optimizing carry out better discriminator.

Four, PSO optimization SVM flow process is as follows:

(1) the random initial value that generates m particle in given parameters space; The particle generating is SVM parameter, selects certain sample input to determine the SVM model of parameter;

(2) determine the categorised decision functional value of SVM;

(3) be the calculating evaluation that particle carries out fitness to selected parameter;

(4) if do not meet end condition, adopt position, the speed of PSO algorithm to particle to carry out iteration renewal, generate particle of new generation, turn to (2);

(5) if meet end condition, export optimized parameter, again train SVM, carry out discriminator as final grader.

Like this, through above PSO optimized algorithm, the punishment parameters C of SVM and the optimal value of kernel function radius parameter g are just obtained, for training and the classification prediction of SVM.

Step (4), the Gait Recognition of lower limb electromyographic signal.The electromyographic signal feature samples collection that step (2) is extracted, is divided into two groups of training sample and test sample books at random.Training set is for training the SVM after PSO optimizes to set up model.Test set is used for inputting svm classifier device to be identified, and obtains recognition result.

The present invention, compared with existing gait recognition method, has following features:

1, avoided study, local minimum.Neutral net is widely used at electromyographic signal area of pattern recognition, but it still there is local optimum space, pace of learning is slow, generalization is poor and be difficult to process the difficulties such as complex patterns information, and practical application is restricted.Support vector machine has overcome the shortcoming of local minimum in neutral net, in solution small sample, non-linear and high dimensional pattern problem, has advantage.

2, parameter optimization process adaptive is good, simple, efficient.In particle swarm optimization algorithm, particle, according to its global search of practical situation dynamic equilibrium and local search ability, has adaptivity.Utilize particle swarm optimization algorithm, punishment parameter and the kernel functional parameter of Support Vector Machines Optimized, improved parameter optimization speed, takes into account accuracy and the adaptivity of Gait Recognition.

Brief description of the drawings

Fig. 1 is implementing procedure figure of the present invention;

Fig. 2 is comparison diagram before and after electromyographic signal de-noising;

Fig. 3 is support vector machine structure chart;

Fig. 4 is that PSO optimizes SVM flow chart;

Fig. 5 is fitness (accuracy rate) curve chart that PSO finds optimal parameter.

Detailed description of the invention

Below in conjunction with accompanying drawing, embodiments of the invention are elaborated: the present embodiment is implemented under taking technical solution of the present invention as prerequisite, provided detailed embodiment and concrete operating process.But protection scope of the present invention is not limited to following embodiment.

As Fig. 1, the enforcement of the inventive method mainly comprises four steps:

Step 1, the sample acquisition of lower limb electromyographic signal.Press muscle group effect and the contribution of gait action different phase when walking, and the sensitivity of effects on surface electromyographic signal collection equipment, the most representative four muscle on thigh selected: vastus medialis, semitendinosus m., adductor longus m., tensor fasciae latae.These 4 muscle are distributed in the zones of different of thigh, on position and signaling zone calibration, all have typicality.Experimenter carries out level walking with normal speed (1.5m/s), and staff utilizes MyoTrace400 myoelectricity Acquisition Instrument, gathers above-mentioned 4 tunnel muscle surface electromyographic signal data in periodicity gait motion.

Adopt Wavelet Modulus Maxima denoising method, first electromyographic signal is carried out to wavelet decomposition, then according to the singularity of wavelet coefficient, utilize signal from noise mode maximum the different variation characteristics on wavelet scale, isolate signal and noise, the electromyographic signal sample data after last reconstruct de-noising.As shown in Figure 2, Fig. 2 (a), (b) are respectively the forward and backward electromyographic signal of de-noising to result, and transverse axis is sampling number, and the longitudinal axis is voltage (uV).From comparison diagram, can find out, the electromyographic signal after Wavelet Modulus Maxima denoising, useful information is remained well, and signal to noise ratio obviously improves.Realize carrying out denoising Processing containing noisy electromyographic signal, removed white noise and retain singular point information.

Step 2, the feature extraction of lower limb electromyographic signal.To the electromyographic signal after step 1 denoising Processing, extract integration myoelectricity value and absolute value variance temporal signatures, obtain characteristic vector sample set.Specific implementation process is as follows:

The integration myoelectricity value I of 4 road surfaces electromyographic signals while asking for i state _ij, absolute value variance V _ij, i=1,2 ..., N is status switch number; J=1,2,3,4 corresponding four muscle, build temporal signatures vector Z _i={ I _i1, V _i1, I _i2, V _i2, I _i3, V _i3, I _i4, V _i4, can obtain the characteristic vector sample set that 8 combination of eigenvectors become.

Step 3, structure PSO-SVM grader.SVM is carried out to optimization of parameter choice with PSO, obtain making one group of punishment parameters C and the kernel functional parameter g of SVM error minimum, Optimizing Flow is as Fig. 4.Svm classifier device after the electromyographic signal feature samples set pair optimization that step 2 is extracted is trained, is tested, and carries out discriminator.Detailed description of the invention is as follows:

First, the initial parameter of PSO algorithm is set.Research with reference to people such as Pan Feng to PSO algorithm, establishes inertia weight w=0.8, meets w ∈ [0.2,1] scope, study factor c ₁=1.5, c ₂=1.7, meet the span of [0,4].Particle scale is that population quantity is set to 20, and maximum iteration time maxgen is just made as 100, as the stopping criterion for iteration of PSO algorithm.

As Fig. 5, when PSO meets stopping criterion for iteration, optimizing process finishes, and the fitness (accuracy rate) that PSO finds optimal parameter reaches 99.7086%, output optimized parameter C _best=969.0311, g _best=1, all meet C ∈ [0.1,10000], g ∈ [0.1,100] scope.Punish parameter and kernel functional parameter for one group that after optimization, has obtained SVM error minimum.

Step 4, the Gait Recognition of lower limb electromyographic signal.The electromyographic signal feature samples collection that step 2 is extracted is divided into two groups of training sample and test sample books at random, and SVM after optimizing with training sample training PSO sets up model.With test sample book input svm classifier device, obtain recognition result, support vector machine structure is as shown in Figure 3.

Test target carries out according to each gait cycle, be divided into successively for: support early stage mid-term the later stage, swing early stage later stage, totally 5 status recognitions.Under the normal gait obtaining according to step 2,2760 groups altogether of the feature samples of electromyographic signal, choose arbitrarily 2000 groups of sample datas as training set, and residue 760 stack features data, as test set, are sent into svm classifier device and identified.If recognition result is consistent with test target, illustrates that correct classification has been carried out in the action of test, otherwise be wrong classification.Contrast recognition result, as shown in following table 1 and table 2, can be found out, adopts the result after PSO-SVM parameter optimization, aspect recognition accuracy and generalization ability, is all better than the SVM recognition result without parameter optimization.

The recognition result of table 1SVM to gait

Table 2PSO optimizes the recognition result of SVM to gait

Claims (1)

1. the electromyographic signal gait recognition method based on particle group optimizing-support vector machine, is characterized in that the method comprises the steps:

Step (1), the sample acquisition of lower limb electromyographic signal; Press muscle group effect and the contribution of gait action different phase when walking, and the sensitivity of effects on surface electromyographic signal collection equipment, select representative four muscle on thigh, the original electromyographic signal of muscle when the lower limb that pick up by myoelectricity Acquisition Instrument walking; Adopt Wavelet Modulus Maxima denoising method, first electromyographic signal is carried out to wavelet decomposition, then according to the singularity of wavelet coefficient, utilize signal from noise mode maximum the different variation characteristics on wavelet scale, isolate signal and noise, the electromyographic signal sample data after last reconstruct de-noising; Realize carrying out denoising Processing containing noisy electromyographic signal, remove white noise and retain singular point information; Specific as follows:

First, signals and associated noises is carried out to wavelet transform; First the original electromyographic signal gathering is carried out to 4 layers of wavelet decomposition, base small echo is selected tight support biorthogonal wavelet sym8, obtains the wavelet transform Wf (s, x) of the signals and associated noises f at x place, the upper position of yardstick s;

Secondly, ask for the Lip index of signals and associated noises; The singular point of signal is exactly the catastrophe point in signal, and Lip index is that the one of the local singular point feature of characterization signal is measured, and is defined as follows:

If positive integer n,, if there is positive integer A > 0 and polynomial of degree n p in n≤α≤n+1 _n(x), make

|f(x)-p _n(x-x ₀)|≤A|x-x ₀| ^α (1)

For x ∈ (x ₀-δ, x ₀+ δ) set up, claim f (x) at x ₀point is Lip α; α is larger, and the smoothness of this point is higher; α is less, and the singularity of this point is larger;

Again, remain with signaling point; The Lip index of signal f (x) and Wavelet Modulus Maxima (Wavelet Modulus Maxima need meet | Wf (s, x) |≤| Wf (s, x ₀), x ₀for the local model maximum value point of wavelet transformation under yardstick s) need to meet

${\log }_2\left|W_{2^j}f\left(t\right)\right|\&le;{\log }_{2}k+\mathrm{j\&alpha;}---\left(2\right)$

Wherein, t is the time, and j is wavelet scale, k ∈ R ⁿ;

To general signal α >=0, the modulus maximum of wavelet transformation increases the increase along with yardstick j; And for white noise α < 0, its modulus maximum reduces along with the increase of yardstick j; Therefore utilize the rule of wavelet modulus maxima conversion between different scale, the point (extreme point of corresponding noise) that removal amplitude reduces with the increase of yardstick, the point (extreme point of corresponding useful signal) that reservation amplitude increases with yardstick;

Finally, to Wf (s, x ₀) carry out wavelet reconstruction, obtain the electromyographic signal sample after Wavelet Modulus Maxima denoising;

Step (2), the feature extraction of lower limb electromyographic signal; The electromyographic signal of obtaining for step (1), by extracting its integration myoelectricity value, absolute value variance temporal signatures, obtains the sample set of characteristic vector; Specific as follows:

The calculating formula of integration myoelectricity value is

$I_i=\frac{1}{N}\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\left|x_{\left(i\right)}\right|---\left(3\right)$

Here, x (i), i=0,1,2 ..., N-1 is the electromyographic signal time samples sequence of a length N, i is every group of sampling number;

Absolute value variance is defined as follows:

$N_i=\frac{1}{N-1}\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{(x_{\left(i\right)}-I_i)}^2---\left(4\right)$

Wherein I _ifor the integration myoelectricity value of electromyographic signal;

Thus, obtain eight characteristic parameters of four tunnel electromyographic signals, form a stack features vector Z;

Step (3), the grader of structure based on PSO-SVM; SVM is carried out to optimization of parameter choice with PSO, obtain one group of punishment parameter and the kernel functional parameter of SVM error minimum; Specific as follows:

One, in iterative process each time, particle upgrades speed and the position of self by individual extreme value and global extremum, and more new formula is as follows:

$V_{\mathrm{id}}^{k+1}=w{V}_{\mathrm{id}}^k+c_1{r}_1(P_{\mathrm{id}}^k-X_{\mathrm{id}}^k)+c_2{r}_2(P_{\mathrm{gd}}^k-X_{\mathrm{id}}^k)$

(5)

$X_{\mathrm{id}}^{k+1}=X_{\mathrm{id}}^k+V_{\mathrm{id}}^{k+1}$

In formula, w is inertia weight; D=1,2 ..., D is space dimensionality, i=1, and 2 ..., n is population; K is current iteration number of times; be illustrated in the flight speed of i particle, represent the position of i particle; with represent respectively individual optimal solution and the globally optimal solution of i particle; Here the change in location scope of d dimension is velocity variations scope is if X in iteration _idthe value of being beyond the boundary, is made as boundary value-x _maxor x _max; c ₁and c ₂for the study factor (or accelerated factor), represent respectively the weights of each particle being pushed to the statistics acceleration term of individual extreme value and global extremum position, middle value in [0,4] scope; r ₁and r ₂for being distributed in the random number between [0,1];

Two, support vector machines; The basic thought of SVM is by the more feature space of higher-dimension of DUAL PROBLEMS OF VECTOR MAPPING to, sets up a face that makes different types of data point interval maximum, i.e. largest interval hyperplane in this space; By setting suitable kernel function K (z _i, z) carry out nonlinear transformation the input space is transformed to a higher dimensional space, then in this new space, ask for optimum linearity classifying face; The output category decision function of SVM is:

$f\left(Z\right)=\mathrm{sgn}(\mathrm{wz}+b)=\mathrm{sgn}(\underset{i=1}{\overset{s}{\mathrm{\&Sigma;}}}{a}_i{y}_{i}K(z_i,z)+b),0\&le;a_i\&le;C---\left(6\right)$

A _ifor the corresponding Lagrangian coefficient of each training sample; K (z _i, z) be core letter (K (z _i, z)=exp (| z _i-z|^2/g^2)); C is punishment parameter, and b ∈ R is biasing;

Three, the SVM parameter optimization based on PSO: select radial basis kernel function, it is exactly punishment parameters C and the kernel function radius parameter g of Support Vector Machines Optimized that PSO optimizes SVM, makes the SVM after optimizing carry out better discriminator;

Four, PSO optimization SVM flow process is as follows:

(1) the random initial value that generates m particle in given parameters space; The particle generating is SVM parameter, selects certain sample input to determine the SVM model of parameter;

(2) determine the categorised decision functional value of SVM;

(3) be the calculating evaluation that particle carries out fitness to selected parameter;

(4) if do not meet end condition, adopt position, the speed of PSO algorithm to particle to carry out iteration renewal, generate particle of new generation, turn to (2);

(5) if meet end condition, export optimized parameter, again train SVM, carry out discriminator as final grader;

Like this, through above PSO optimized algorithm, the punishment parameters C of SVM and the optimal value of kernel function radius parameter g are just obtained, for training and the classification prediction of SVM;

Step (4), the Gait Recognition of lower limb electromyographic signal; The electromyographic signal feature samples collection that step (2) is extracted, is divided into two groups of training sample and test sample books at random; Training set is for training the SVM after PSO optimizes to set up model; Test set is used for inputting svm classifier device to be identified, and obtains recognition result.