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

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

Based on GA-BP neural network lower limb level walking gait recognition method

Technical field

The present invention relates to a kind of human motion mode identification method, the method for the GA-BP neural network Gait Recognition of electromyographic signal eigenwert during particularly based on lower limb level walking.

Background technology

Gait embodies the attitude of lower limb walking movement, is lower limb walking states general designation.It and organization of human body and function, motor coordination tissue, behavior and psychological activity have important relation, are actions the most basic in human life activity.Normal gait (normal gait) refers to gait when healthy human body lower limb have the attitude walking of oneself feeling the most natural, the most comfortable, and it has periodically and coordination and balanced feature.

Human body surface myoelectric signal is a kind of weak biological signal of low frequency, a kind of physiological signal with non-stationary, non-Gaussian feature in itself, picking up, in conditioning, gatherer process, inevitably can introduce many interference, by suitable signal noise silencing, eigenwert extraction and mode identification method, can distinguish the gait of lower limb level walking.

The research of human body lower limbs gait pattern recognition methods, since the nineties in last century, has obtained many achievements.A kind of method that for example Mumse has proposed temporal and spatial correlations coupling from Saka is for distinguishing different gaits; The Foster of University of Southampton etc. propose to adopt the method for region tolerance to solve Gait Recognition problem.Along with the development of artificial neural network, the domestic people of having uses the eigenwert that neural network extracts electromyographic signal to classify, and has finally obtained good effect.Yet the common theoretical foundation of present most of method is the classical theory of statistics, employing be the progressive theory of research number of samples while being tending towards infinity.Yet in practical problems, number of samples is often limited, therefore these sorting technique that has in theory remarkable strong point performances in actual applications but may be unsatisfactory, for example traditional BP neural network classification easily occur local minimum and classifying quality undesirable.The present invention adopts the BP neural network after a kind of genetic algorithm optimization to carry out better pattern-recognition and classification.

Summary of the invention

The present invention is exactly the deficiency for traditional BP neural network classification, employing GA(genetic algorithm) initial weight and the threshold value of Optimized BP Neural Network, the eigenwert that BP neural network after optimizing is extracted electromyographic signal is carried out discriminator, thereby improves correct recognition rata.

Object of the present invention can be achieved through the following technical solutions:

The present invention includes following steps:

Four road surfaces electromyographic signals of the continuous level walking action of lower limb that step 1. pair collects carry out de-noising filtering and temporal signatures value is extracted, and obtain its proper vector sample set.

Step 2. is optimized BP neural network with GA, obtains one group of complete initial weight and the threshold value of BP neural network error minimum.

Step 3. is divided into two groups of training sample and test sample books at random by the eigenwert of extracting in step 1, and trains the BP neural network of GA after optimizing with training sample; The BP neural network classifier training with test sample book input, carries out discriminator.

Wherein the de-noising filtering of step 1 Zhong Si road electromyographic signal adopts spatial correlation filtering, and concrete steps are summed up as follows:

(1) signals and associated noises is carried out to wavelet transform, the original electromyographic signal gathering is carried out to 5 layers of wavelet decomposition, base small echo is selected the little bior 1.5 of bi-orthogonal spline, obtains the wavelet transform Wf (j, n) of the signals and associated noises f at n place, the upper position of yardstick j.

(2) ask for the related coefficient Corr that each yardstick is adjacent yardstick ₂(j, n)=Wf (j, n) Wf (j+1, n).

(3) by Corr ₂the energy that (j, n) normalizes to Wf (j, n) gets on, and obtains the related coefficient NewCorr after normalization ₂(j, n).Computing method are:

$\mathrm{New}{\mathrm{Corr}}_2(j,n)={\mathrm{Corr}}_2(j,n)\sqrt{P_W\left(j\right)/P_{{\mathrm{Corr}}_2}\left(j\right)},n=\mathrm{1,2},\·\·\·,N$

Wherein, $P_{{\mathrm{Corr}}_2}\left(j\right)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{Corr}}_2{(j,n)}^2;$ $P_W\left(j\right)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}\mathrm{Wf}{(j,n)}^2$

(4) if NewCorr ₂(j, n) > Wf (j, n), thinks that the wavelet coefficient values at n point place is to have signal to produce, and the value of Wf (j, n) is replaced to Wf _newthe relevant position of (j, n), and by Wf (j, n) zero setting, Corr ₂(j, n) zero setting; Otherwise think that Wf (j, n) is produced by noise, retain Wf (j, n).

(5) repeating step (3) and (4), until P _w(j) (in spatial correlation filtering de-noising, needs are set a certain noise threshold, and the present invention only contains noisy point estimation noise in the variance of each layer with 80 of signal in experiment, and usings these 10 times as noise energy threshold value to meet a certain noise energy threshold value.For obtaining the point of 80 Noises, when loosening, each related muscles samples 80 points as a reference, think that now signal is noise).At this moment Wf _new(j, n) retained the wavelet coefficient after removal noise.

(6) to Wf _new(j, n) carries out wavelet reconstruction and just obtains the signal after spatial correlation filtering.

The temporal signatures value of extracting is integration myoelectricity value and absolute value variance.The definition of two eigenwerts and extraction are as follows:

Integration myoelectricity value (Integrate EMG), its calculating formula is:

$x_{\mathrm{iemg}}=\frac{1}{N}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\left|x_{\left(i\right)}\right|$

The sampling number that wherein i is every group, x _(i)for the data dot values of surface electromyogram signal sampling, N is every group of sampling number.

Absolute value variance (Variance): to the operation that takes absolute value of original electromyographic signal, then ask for the variance of gained signal, variance is in such cases defined as follows again:

$\mathrm{VAR}=\frac{1}{N-1}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}{(x_{\left(i\right)}-x_{\mathrm{iemg}})}^2$

Totally eight characteristic parameters of four road signals build a proper vector.

Wherein the GA in " BP neural network is optimized with GA, obtains one group of complete initial weight and the threshold value of BP neural network error minimum " described in step 2 is a kind of parallel random search optimization method.It introduces the theory of biologic evolution of nature " survival of the fittest; the survival of the fittest " in the coding series connection colony of Optimal Parameters formation, according to selected fitness function and by the selection in heredity, crossover and mutation, individuality is screened, good individuality is retained to make fitness value, the poor individuality of fitness is eliminated, the information of previous generation had both been inherited by new colony, was better than again previous generation.Iterative cycles like this, until satisfy condition.Specifically describe as follows:

Initialization of population:

Individual coding method is real coding, and each individuality is a real number string, is connected that weights, hidden layer threshold value, hidden layer are connected weights with output layer and output layer threshold value 4 parts form by input layer with hidden layer.Individuality has comprised the whole weights of neural network and threshold value, in the situation that network structure is known, just can form a structure, weights, the definite neural network of threshold value.

Determining of fitness function:

According to individuality, obtain initial weight and the threshold value of BP neural network, with prognoses system output after training data training BP neural network, the Error Absolute Value between prediction output and desired output and E are as ideal adaptation degree value F, and computing formula is:

$F=k\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{abs}(y_i-o_i)\right)$

In formula, n is network output node number; y _idesired output for i node of BP neural network; o _ibe the prediction output of i node; K is coefficient.

Select operation:

Genetic algorithm selects operation to have the several different methods such as roulette method, tournament method, and the present invention selects roulette method, i.e. the selection strategy based on fitness ratio, the selection Probability p of each individual i _icomputing formula be:

f _i＝k/F _i； $p_i=\frac{f_i}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{f}_i}$

In formula, F _ifor the fitness value of individual i, because fitness value is the smaller the better, so fitness value is asked to reciprocal before individual choice; K is coefficient; N is population at individual number.

Interlace operation:

Because individuality adopts real coding, so interlace operation method adopts real number bracketing method, a _kjbe k chromosome a _kj position, a _kjbe l chromosome a _lj position, both are as follows in j position interlace operation method:

$\left\{\begin{array}{c}{a}_{\mathrm{kj}}=a_{\mathrm{kj}}(1-b)+a_{\mathrm{ij}}b\\ a_{\mathrm{lj}}=a_{\mathrm{lj}}(1-b)+a_{\mathrm{kj}}b\end{array}\right.$

In formula, b is the random number between [0,1].

Mutation operation:

Choose i j individual gene a _ijmake a variation, mutation operation method is as follows:

$a_{\mathrm{ij}}=\left\{\begin{array}{cc}{a}_{\mathrm{ij}}+(a_{\mathrm{ij}}-a_{\max })*f\left(g\right)& r\&GreaterEqual;0.5\\ a_{\mathrm{ij}}+(a_{\min }-a_{\mathrm{ij}})*f\left(g\right)& r<0.5\end{array}\right.$

In formula, a _maxfor gene a _ijthe upper bound; a _minfor gene a _ijlower bound; F (g)=r ₂(1-g/G _max); r ₂it is a random number; G is current iteration number of times; G _maxit is maximum evolution number of times; R is the random number between [0,1].Like this, through above hereditary computing, just obtained one group of complete initial weight and the threshold value of BP neural network error minimum.

Wherein the BP neural network described in step 2,3 is the neural network of error back propagation, and its algorithm basic thought is gradient descent method.It adopts gradient search technology, to making the output valve of network and the error mean square value of desired output for minimum.BP neural network after optimizing is trained, take variance yields sample as example, for BP network, comprise input layer, hidden layer output neuron and output neuron.

The training process of BP network is as follows: forward-propagating is that input signal is transmitted to output layer from input layer through hidden layer, if output layer has obtained the output of expectation, learning algorithm finishes; Otherwise, go to backpropagation.

Learning Algorithms is as follows:

(1) propagated forward: the output of computational grid

W _ijfor i neuron of input layer and hidden layer j neuronic connection weights, x _ifor i neuron of input layer is to output layer j neuronic output, the input x of hidden layer neuron _jfor all x _iweighting sum:

$x_j=\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}{x}_i$

Implicit neuronic output x layer by layer _j', adopt S function to excite x _j:

$x_j^{\&prime;}=f\left(x_j\right)=\frac{1}{1+e^{-x_j}}$

? $\frac{{\&PartialD;x}_j^{\&prime;}}{{\&PartialD;x}_j}=x_j^{\&prime;}(1-x_j^{\&prime;})$

The neuronic output of output layer x _l:

$x_l=\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{jl}}{x}_j^{\&prime;}$

L output of network and corresponding desirable output

error e ₁for:

$e_1=x_l^0-x_l$

The error performance target function E of p sample _pfor:

$E_p=\frac{1}{2}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_l^2$

Wherein N is the number of network output layer;

(2) backpropagation: adopt gradient descent method, adjust the weights of each interlayer; The learning algorithm of weights is as follows:

L neuron of output layer and hidden layer j neuronic connection weight w _jllearning algorithm is:

${\mathrm{\&Delta;w}}_{\mathrm{jl}}=-\mathrm{\&eta;}\frac{{\&PartialD;E}_p}{{\&PartialD;w}_{\mathrm{jl}}}={\mathrm{\&eta;e}}_l\frac{{\&PartialD;x}_l}{{\&PartialD;w}_{\mathrm{jl}}}=\mathrm{\&eta;}{e}_l{x}_j^{\&prime;}$

w _ij(k+1)＝w _ij(k)+Δw _ij

W wherein _ij(k) be w in the k time study _ijvalue, and

$\frac{{\&PartialD;x}_l}{{\&PartialD;w}_{\mathrm{ij}}}=\frac{{\&PartialD;x}_l}{{\&PartialD;x}_j^{\&prime;}}\&CenterDot;\frac{{\&PartialD;x}_j^{\&prime;}}{{\&PartialD;x}_j}\&CenterDot;\frac{{\&PartialD;x}_j}{{\&PartialD;w}_{\mathrm{ij}}}=w_{\mathrm{jl}}\&CenterDot;\frac{{\&PartialD;x}_j^{\&prime;}}{{\&PartialD;x}_j}\&CenterDot;x_i=w_{\mathrm{jl}}\&CenterDot;x_j^{\&prime;}(1-x_j^{\&prime;})\&CenterDot;x_i$

Consider the impact that last time, weights changed these weights, must add factor of momentum a, weights are now:

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, η is learning rate, and a is factor of momentum, k, and t is study number of times, η ∈ [0,1], a ∈ [0,1];

After training, input test sample carries out discriminator.

Compared with prior art, the present invention has the following advantages:

1. the temporal signatures of electromyographic signal easily extracts, feature is obvious, have good representation ability.

2. adopt GA to optimize BP neural network afterwards and carry out Gait Recognition, accuracy of identification is high, and error rate is little.

Accompanying drawing explanation

Fig. 1 is contrast before and after electromyographic signal de-noising;

Fig. 2 is GA Optimized BP Neural Network flow process;

Fig. 3 is BP neural network structure schematic diagram;

Fig. 4 is FB(flow block) of the present invention;

Fig. 5 is identification error evolution curve;

In Fig. 1, (a), (b) represent respectively the electromyographic signal of de-noising front and back, and transverse axis is sampling number, and the longitudinal axis is voltage (uV).As can be seen from Figure 1 the signal to noise ratio (S/N ratio) of the electromyographic signal after spatial correlation filtering de-noising is obviously improved, and the edge feature of electromyographic signal is kept down preferably simultaneously, and good condition has been created in this raising for feature extraction and pattern-recognition rate.

In Fig. 2, left part is divided into GA algorithm part, and right part is divided into BP neural metwork training learning process.

In Fig. 3, X ₁, X ₂..., X _mthe input value of BP neural network, Y ₁, Y ₂..., Y _nthe predicted value of BP neural network, ω _ijand ω _jkfor BP neural network weight.

Embodiment

See Fig. 4, a kind of GA Optimized BP Neural Network of the present invention is to lower limb level walking gait recognition method, and the method comprises the following steps:

Four road surfaces electromyographic signals of the continuous level walking action of lower limb that step 1. pair collects carry out de-noising filtering and the extraction of temporal signatures value obtains its proper vector sample set.

Step 2. is optimized BP neural network with GA, and Optimizing Flow as shown in Figure 2, obtains one group of complete initial weight and the threshold value of BP neural network error minimum.

Step 3. is divided into two groups of training sample and test sample books at random by the eigenwert of extracting in step 1, and trains the BP neural network of GA after optimizing with training sample.The BP neural network classifier training with test sample book input, carries out discriminator.

Clearer for what the object, technical solutions and advantages of the present invention were expressed, below in conjunction with drawings and the specific embodiments, the present invention is further described in detail again.

Main thought of the present invention is by GA Optimized BP Neural Network, has given one group of complete initial weight and the threshold value of BP neural network error minimum.On this basis in conjunction with electromyographic signal temporal signatures value extract easily, eigenwert obviously a little, build the proper vector with good representation ability.BP neural network after optimizing is trained and identified.

Step 1: embodiment is as follows:

Aspect, signal source, the MyoTrace400 electromyographic signal collection instrument of the electromyographic signal collection Yi Shi U.S. Noraxon company that the present invention is used.The surface electromyogram signal of the most representative 4 muscle on thigh while gathering level walking, the tensor fasciae late muscle that they are respectively the vastus medialises of thigh front side, the semitendinosus of thigh rear side, thigh are connected with crotch and the long adductor muscle of femoribus internus.These 4 muscle are distributed in the zones of different of thigh, on position and signaling zone calibration, all have typicalness.

Signal noise silencing filtering adopts spatial correlation filtering (the electromyographic signal waveform before and after de-noising is as shown in Figure 1), and it is as follows that eigenwert is afterwards extracted specific implementation process:

Ask for the integration myoelectricity value I of four road surfaces electromyographic signals of i kind gait _ij, variance V _ij, i=1,2,3,4,5; J=1,2,3,4.

Build temporal signatures vector X _i={ I _i1, V _i1, I _i2, V _i2, I _i3, V _i3, I _i4, V _i4.

Each proper vector composition characteristic vector sample set drawing from above.

Step 2: by BP neural network design three layers of BP network structure (BP neural network structure schematic diagram as shown in Figure 3), because integration myoelectricity value and two eigenwerts of variance are all extracted in the 4 every roads of tunnel electromyographic signal, input end has 8 neurons, output layer has 5 neurons, hidden layer neuron number rule of thumb formula and repeatedly operating analysis we be made as 15.

n in formula ₁for hidden layer neuron number, n is input layer number, and m is output layer neuron number, constant a=1～10.

Step 3: all getting learning rate η during training is 0.15, and factor of momentum a is 0.05.Output uses respectively (00001) to represent to support early stage, and (00010) represents to support mid-term, and (00100) represents to support the later stage, and (01000) represents to swing early stage, and (10000) represent to swing the later stage.1100 groups altogether, the proper vector sample collecting in step 1, therefrom random choosing 800 groups be used for training, remaining 300 groups is used for testing, discrimination as shown in table 2 all reaches more than 98%, and the least error of the neural network that when genetic algorithm evolved to for the 50th generation, its sample obtains by evolution test is 0.00827, as shown in Figure 5.Compare better with the BP neural network recognization effect (shown in table 1) of not optimizing, this recognition effect is in place, lower limb Gait Recognition field higher level simultaneously.

The recognition result of table 1 BP neural network to gait

The recognition result of table 2 genetic algorithm optimization BP neural network to gait

Claims (5)

1. based on GA-BP neural network lower limb level walking gait recognition method, it is characterized in that the concrete steps of the method are:

Four road surfaces electromyographic signals of the continuous level walking action of lower limb that step 1. pair collects carry out de-noising filtering and temporal signatures value is extracted, and obtain its proper vector sample set;

Step 2. is optimized BP neural network with GA, obtains one group of complete initial weight and the threshold value of BP neural network error minimum;

Step 3. is divided into two groups of training sample and test sample books at random by the eigenwert of extracting in step 1, and trains the BP neural network of GA after optimizing with training sample; The BP neural network classifier training with test sample book input, carries out discriminator.

2. according to claim 1 based on GA-BP neural network lower limb level walking gait recognition method, it is characterized in that: the de-noising filtering in step 1 adopts spatial correlation filtering, specifically:

(1) signals and associated noises is carried out to wavelet transform, the original electromyographic signal gathering is carried out to 5 layers of wavelet decomposition, base small echo is selected the little bior 1.5 of bi-orthogonal spline, obtains the wavelet transform Wf (j, n) of the signals and associated noises f at n place, the upper position of yardstick j;

(2) ask for the related coefficient Corr that each yardstick is adjacent yardstick ₂(j, n)

Corr ₂(j,n)＝Wf(j,n)Wf(j+1,n)；

(3) by Corr ₂the energy that (j, n) normalizes to Wf (j, n) gets on, and obtains the related coefficient NewCorr after normalization ₂(j, n):

$\mathrm{New}{\mathrm{Corr}}_2(j,n)={\mathrm{Corr}}_2(j,n)\sqrt{P_W\left(j\right)/P_{{\mathrm{Corr}}_2}\left(j\right)},n=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N$

Wherein, $P_{{\mathrm{Corr}}_2}\left(j\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{Corr}}_2{(j,n)}^2;$ $P_W\left(j\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}\mathrm{Wf}{(j,n)}^2$

(4) if NewCorr ₂(j, n) > Wf (j, n), thinks that the wavelet coefficient values at n point place is to have signal to produce, and the value of Wf (j, n) is replaced to Wf _newthe relevant position of (j, n), and by Wf (j, n) zero setting, Corr ₂(j, n) zero setting; Otherwise think that Wf (j, n) is produced by noise, retain Wf (j, n);

(5) repeating step (3) and (4), until P _w(j) meet a certain noise energy threshold value, at this moment Wf _new(j, n) retained the wavelet coefficient after removal noise;

(6) to Wf _new(j, n) carries out wavelet reconstruction and just obtains the signal after spatial correlation filtering.

3. according to claim 1ly based on GA-BP neural network lower limb level walking gait recognition method, it is characterized in that: in step 1, temporal signatures value is extracted as integration myoelectricity value and absolute value variance the definition of two eigenwerts and extract as follows:

Integration myoelectricity value x _iemg, its calculating formula is:

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

The sampling number that wherein i is every group, x _(i)for the data dot values of surface electromyogram signal sampling, N is every group of sampling number;

Absolute value variance VAR: to the operation that takes absolute value of original electromyographic signal, then ask for the variance of gained signal, variance is in such cases defined as follows again:

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

Totally eight characteristic parameters of four road signals build a proper vector.

4. according to claim 1 based on GA-BP neural network lower limb level walking gait recognition method, it is characterized in that: BP neural network is optimized with GA described in step 2, specifically:

(1) initialization of population: individual coding method is real coding, each individuality is a real number string, is connected that weights, hidden layer threshold value, hidden layer are connected weights with output layer and output layer threshold value 4 parts form by input layer with hidden layer; Individuality has comprised the whole weights of neural network and threshold value, in the situation that network structure is known, just can form a structure, weights, the definite neural network of threshold value;

(2) fitness function determines

According to individuality, obtain initial weight and the threshold value of BP neural network, with prognoses system output after training data training BP neural network, the Error Absolute Value between prediction output and desired output and E are as ideal adaptation degree value F, and computing formula is:

$F=k\left(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{abs}(y_i-o_i)\right)$

In formula, n is network output node number; y _idesired output for i node of BP neural network; o _ibe the prediction output of i node; K is coefficient;

(3) select operation

Genetic algorithm selects operation to adopt roulette method, i.e. the selection strategy based on fitness ratio, the selection Probability p of each individual i _icomputing formula be:

f _i＝k/F _i； $p_i=\frac{f_i}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_i}$

In formula, F _ifor the fitness value of individual i, because fitness value is the smaller the better, so fitness value is asked to reciprocal before individual choice; K is coefficient; N is population at individual number;

(4) interlace operation

Because individuality adopts real coding, so interlace operation method adopts real number bracketing method, a _kjbe k chromosome a _kj position, a _kjbe l chromosome a _lj position, both are as follows in j position interlace operation method:

$\left\{\begin{array}{c}{a}_{\mathrm{kj}}=a_{\mathrm{kj}}(1-b)+a_{\mathrm{ij}}b\\ a_{\mathrm{lj}}=a_{\mathrm{lj}}(1-b)+a_{\mathrm{kj}}b\end{array}\right.$

In formula, b is the random number between [0,1];

(5) mutation operation

Choose i j individual gene a _ijmake a variation, mutation operation method is as follows:

$a_{\mathrm{ij}}=\left\{\begin{array}{cc}{a}_{\mathrm{ij}}+(a_{\mathrm{ij}}-a_{\max })*f\left(g\right)& r\&GreaterEqual;0.5\\ a_{\mathrm{ij}}+(a_{\min }-a_{\mathrm{ij}})*f\left(g\right)& r<0.5\end{array}\right.$

In formula, a _maxfor gene a _ijthe upper bound; a _minfor gene a _ijlower bound; F (g)=r ₂(1-g/G _max); r ₂it is a random number; G is current iteration number of times; G _maxit is maximum evolution number of times; R is the random number between [0,1]; Like this, through above hereditary computing, just obtained one group of complete initial weight and the threshold value of BP neural network error minimum.

According to described in any one in claim 1-4 based on GA-BP neural network lower limb level walking gait recognition method, it is characterized in that:

BP neural network described in step 2 and step 3 is the neural network of error back propagation, and its algorithm basic thought is gradient descent method; It adopts gradient search technology, to making the output valve of network and the error mean square value of desired output for minimum; BP neural network after optimizing is trained, take variance yields sample as example, for BP network, comprise input layer, hidden layer output neuron and output neuron;

The training process of BP network is as follows: forward-propagating is that input signal is transmitted to output layer from input layer through hidden layer, if output layer has obtained the output of expectation, learning algorithm finishes; Otherwise, go to backpropagation;

Learning Algorithms is as follows:

(1) propagated forward: the output of computational grid

W _ijfor i neuron of input layer and hidden layer j neuronic connection weights, x _ifor i neuron of input layer is to output layer j neuronic output, the input x of hidden layer neuron _jfor all x _iweighting sum:

$x_j=\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}{x}_i$

Implicit neuronic output x layer by layer _j', adopt S function to excite x _j:

$x_j^{\&prime;}=f\left(x_j\right)=\frac{1}{1+e^{-x_j}}$

? $\frac{{\&PartialD;x}_j^{\&prime;}}{{\&PartialD;x}_j}=x_j^{\&prime;}(1-x_j^{\&prime;})$

The neuronic output of output layer x _l:

$x_l=\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{jl}}{x}_j^{\&prime;}$

L output of network and corresponding desirable output

error e ₁for:

$e_1=x_l^0-x_l$

The error performance target function E of p sample _pfor:

$E_p=\frac{1}{2}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_l^2$

Wherein N is the number of network output layer;

(2) backpropagation: adopt gradient descent method, adjust the weights of each interlayer; The learning algorithm of weights is as follows:

L neuron of output layer and hidden layer j neuronic connection weight w _jllearning algorithm is:

${\mathrm{\&Delta;w}}_{\mathrm{jl}}=-\mathrm{\&eta;}\frac{{\&PartialD;E}_p}{{\&PartialD;w}_{\mathrm{jl}}}={\mathrm{\&eta;e}}_l\frac{{\&PartialD;x}_l}{{\&PartialD;w}_{\mathrm{jl}}}=\mathrm{\&eta;}{e}_l{x}_j^{\&prime;}$

w _ij(k+1)＝w _ij(k)+Δw _ij

W wherein _ij(k) be w in the k time study _ijvalue, and

$\frac{{\&PartialD;x}_l}{{\&PartialD;w}_{\mathrm{ij}}}=\frac{{\&PartialD;x}_l}{{\&PartialD;x}_j^{\&prime;}}\&CenterDot;\frac{{\&PartialD;x}_j^{\&prime;}}{{\&PartialD;x}_j}\&CenterDot;\frac{{\&PartialD;x}_j}{{\&PartialD;w}_{\mathrm{ij}}}=w_{\mathrm{jl}}\&CenterDot;\frac{{\&PartialD;x}_j^{\&prime;}}{{\&PartialD;x}_j}\&CenterDot;x_i=w_{\mathrm{jl}}\&CenterDot;x_j^{\&prime;}(1-x_j^{\&prime;})\&CenterDot;x_i$

Consider the impact that last time, weights changed these weights, must add factor of momentum a, weights are now:

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, η is learning rate, and a is factor of momentum, k, and t is study number of times, η ∈ [0,1], a ∈ [0,1];

After training, input test sample carries out discriminator.

Images