CN101564328B - Laptop artificial limb multi-movement-mode identifying method based on support vector data description - Google Patents

Abstract

The invention relates to a laptop artificial limb multi-movement-mode identifying method based on support vector data description. All present myoelectric signal classification algorithms have defects. By the support vector data description method, the invention provides a dynamic model with multi-mode characteristic extracting capability and realizes the self-adaptive adjustment of a multi-mode characteristic space. The method first obtains body lower limb myoelectric signal sample data, then establishes a support vector data description multi-type classifier, and then judges the affiliation of a test sample, and finally carries out the support vector data description incremental learning, including sample addition and sample deletion. The method well satisfies the requirement of multi-movement-mode identification during laptop lower artificial limb control, and overcomes the defects that a support vector data description off-line training method cannot effectively treat the sample data reflecting object characteristic change and the like. The method has wide application prospect in the multi-movement-mode identification of intelligent artificial limb control.

Description

Laptop artificial limb multi-movement-mode identifying method based on Support Vector data description

Technical field

The invention belongs to area of pattern recognition, relate to a kind of electromyographic signal mode identification method, particularly a kind ofly be used for the multi-locomotion mode recognition methods that the control of above-knee artificial leg has adaptivity.

Background technology

Wearing artificial limb is amputee rehabilitation's important way, can recover its normal condition to some extent on profile and mobility, helps improving individuals with disabilities's quality of life and communal participation ability comprehensively.At the kneed above knee amputation person of disappearance, the coordination control of above-knee artificial leg is the key point that guarantees to adorn oneself with movable nature and safety.In real life, on the one hand the motion of human body lower limbs can be divided into away, run, multiple motor patterns such as crouching, up/down steps, posture feature difference according to human motion, each motor pattern further was subdivided into a plurality of stages again in a natural period, thereby needed artificial limb to have the identification ability of multiple motor pattern; On the other hand, the electromyographic signal of the same action of not only different experimenters is had any different, the same action that same experimenter does at different time also can be variant, and the state of wearing of artificial limb also can cause the variation of electromyographic signal, therefore require every kind of motor pattern to have adaptivity, can follow the changing features of new data and dynamically adjust its feature space.

Electromyographic signal is as a kind of important sport biomechanics information source, and the relatedness of various degrees between the active state of muscle and the functional status, obtained research in fields such as sports medical science, rehabilitation medical.By analysis and feature extraction to electromyographic signal, when individuals with disabilities's intention is made action or can not completely be made action, obtain the active wish of motion, the traditional method compared to only gathering sport dynamics information such as attitude, speed has huge advantage.Fleischer etc. utilize electromyographic signal and attitude information to predict lower extremity movement, and specific mode of operation has been obtained effect preferably.The many groups of employings such as Jin lower limb electromyographic signal to soon at a slow speed, situation such as stair activity carried out identification.Guo Xin etc. have studied the extraction and the analysis of lower limb electromyographic signals, have pointed out the application prospect of electromyographic signal in artificial leg control.Because the complexity of lower extremity movement pattern, the multi-locomotion mode identification of lower limb electromyographic signal has become one of core of artificial leg technical research.

What in fact the multi-locomotion mode identification of electromyographic signal solved is classification problem more than, and present existing electromyographic signal sorting algorithm mainly comprises following several:

1, based on Bayesian method

Consider that electromyographic signal is a kind of stochastic signal with certain feature, redundancy at the multichannel electromyographic signal, method with statistics is classified to the pairing mode of operation of musculation, reasonable recognition effect can be obtained, the information fusion method of Bayes (Bayes) statistical decision method can be selected as multi-locomotion mode identification.For every kind of mode of operation, need carry out test of many times in advance, record the priori probability density of its eigenvalue, to find the solution simultaneously and measure relatively difficulty of conditional probability that eigenvalue belongs to various mode of operations, these problems have all been brought limitation to practical application.At present, the Bayes sorting algorithm is discerned the accuracy rate that reaches in the application in electromyographic signal and is still waiting raising.

2, based on the method for hidden Markov model

(Hide Markov Model HMM) has very strong time series modeling ability to hidden Markov model.In the training stage, set up HMM at every kind of motor pattern, the initial parameter of setting model at first utilizes the Baum-Welch algorithm to carry out the HMM parameter after sample training obtains revaluation and then the HMM parameter library of all model parameter construction systems then; At cognitive phase, the characteristic quantity that extracts from electromyographic signal is input to the HMM of every kind of motor pattern, utilizes Viterbi algorithm computation output probability then, provides classification results by comparing the output probability size.Present achievement in research shows that the discrimination of discerning the lower extremity movement pattern with HMM is also not high, can also improve discrimination by improved model, obtains the higher model of precision such as the HMM model training method of using instead based on maximum mutual information.

3, based on neural network method

Because neutral net has good study, self adaptation and non-linear mapping capability, be applied to the pattern classification field of electromyographic signal in recent years, wherein Chang Yong neural network classifier comprises that error propagates (Back Propagation backward, BP) neutral net and RBF (Radius Basis Function, RBF) neutral net.Add hidden layer between the input and output layer of neutral net, (Multi-Layer Perceptron, MLP), and it has unique learning algorithm, i.e. BP algorithm thereby constitute multilayer perceptron.In the BP algorithm, the working signal forward-propagating, the error signal back propagation, the continuous correction by weights makes the actual output of network more near desired output.In the RBF neutral net, with RBF as " base " of hidden unit, constitute between hidden layer in, make that the linear inseparable problem in lower dimensional space can be divided in the higher dimensional space internal linear.The RBF neural network structure is simple, training is succinct and the study fast convergence rate, can approach any nonlinear function.But, still there is the local optimum space in existing neural network model, hidden layer carries out conversion to input vector, and the input data conversion of low-dimensional is slow to higher dimensional space, pace of learning, generalization difference and be difficult to handle difficulty such as complex patterns information is restricted practical application.

4, based on the method for support vector machine

Support vector machine (Support Vector Machine, SVM) be core content in the Statistical Learning Theory, it is theoretical and structural risk minimization principle based on VC dimension, has overcome problems such as dimension disaster during conventional machines is learnt and local minimum to a great extent.When handling small sample problem, the generalization ability of SVM is fine, has the incomparable performance of traditional machine learning method.Classical SVM algorithm is a kind of two class graders of supervised learning function, when handling the multiclass problem, must transform it, comprise following several method: (1) is converted into the multiclass problem one to one a plurality of or the one-to-many problem is handled, but promote the error unbounded, and in training, reuse sample, thereby can't dynamically adjust; (2) tree classification method need be found the solution a lot of quadratic programming problems, and amount of calculation is big; (3) k class SVM method needs all data of single treatment, and constraints increases, and the quadratic programming of classifying is huge, the dimension-limited of data; (4) (Decision Direct Acyclic Graph, DDAG) method is not considered the influence of sample unbalanced data to classification speed to the circular chart of decision-directed, does not consider the influence of classification error transmission to follow-up generation yet.

In sum, all there is weak point in existing electromyographic signal sorting algorithm, and mostly adopts the off-line training learning style.In fact, the electromyographic signal of the same action of not only different experimenters is had any different, and the same action that same experimenter does at different time also can be variant, these situations require to improve the stability of electromyographic signal eigenvalue on the one hand, require each class action pattern to have adaptivity on the other hand, merge the changing features of new data.The present invention adopts Support Vector data description, and (dynamic model that proposition has multi-mode feature extraction ability is realized the spatial self adaptation adjustment of various pattern features for Support Vector DataDescription, SVDD) method.

Summary of the invention

Purpose of the present invention is exactly at the deficiencies in the prior art, provides a kind of based on data field above-knee artificial leg multi-locomotion mode recognition methods that describe, that have adaptivity.

Multiple motor patterns such as the motion of human body lower limbs can be divided into away, runs, crouching, up/down steps, according to the posture feature difference of human motion, each motor pattern can further be subdivided into a plurality of stages in a natural period, promptly segments motor pattern.The different phase of each motor pattern has visibly different feature, can not directly discern it, and the lower limb attitude could distinguish human motion exactly by identification segmentation pattern the time.The motor pattern identification of mentioning among the present invention refers to the identification to the segmentation motor pattern.Core concept of the present invention is according to the single classification learning method of SVDD, SVDD multi-categorizer and Incremental Learning Algorithm thereof have been designed, can dynamically add emerging motor pattern, but self adaptation is adjusted the feature space of each motor pattern to reflect emerging changing features, and keeping realizing the interpolation of new samples and the deletion of old sample under a certain amount of number of samples situation.

In order to realize above purpose, the inventive method mainly may further comprise the steps:

Step (1) is obtained human body lower limbs electromyographic signal sample data, specifically: at first pick up the human body lower limbs electromyographic signal by the electromyographic signal collection instrument, adopt spatial domain correlation filtering method that the electromyographic signal that contains interference noise is carried out de-noising again, electromyographic signal after adopting multiscale analysis method to de-noising is then carried out feature extraction, and the vector that the relative energy average of the wavelet coefficient after the de-noising on each yardstick constituted is as the feature T of electromyographic signal _l,

$T_l=\frac{\sqrt{{\mathrm{\Σ}}_{n=1}^{N_s}W(l,n)}}{N_s},$ l＝1，2，…，L

Wherein, W (l, the n) wavelet coefficient after the last position n place de-noising of expression yardstick l, N _sBe sampling number, the yardstick of L for decomposing;

After M kind segmentation motor pattern carried out feature extraction, the characteristic of correspondence value constituted an eigenvectors under every kind of segmentation motor pattern, and every eigenvectors is as a kind of segmentation motor pattern corresponding sample x _i ^k, k is the sequence number of segmentation motor pattern, k=1, and 2 ..., M, i are the sequence number of sample, i=1, and 2 ..., N _k, N _kIt is the number of sample in the k kind segmentation motor pattern.

Step (2) is set up Support Vector data description (SVDD) multicategory classification device: adopt the SVDD algorithm to train respectively to M kind segmentation motor pattern corresponding sample, set up M different classes of minimum and comprise hypersphere S _k, k=1,2 ..., M uses centre of sphere a _kAnd radius R _kDescribe minimum and comprise hypersphere S _k, each minimum comprises hypersphere as the class in the grader;

$a_k=\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^k\mathrm{\φ}\left(x_i^k\right),$ k＝1，2，…，M，

${R_k}^2=K(x_q^k,x_q^k)-2\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{k}K(x_q^k,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^k{\mathrm{\α}}_j^{k}K(x_i^k,x_j^k),$ k＝1，2，…，M，

α _i ^kAnd α _j ^kBe the Lagrange multiplier, φ (x _i ^k) be sample x _i ^kNonlinear mapping from the lower dimensional space to the higher dimensional space; x _q ^kBe to be positioned at minimum to comprise hypersphere S _kOn point, q is the sequence number of sample, q ∈ 1,2 ..., N _k;

K (x _i ^k, x _j ^k) be kernel function, $K(x_i^k,x_j^k)=\mathrm{\φ}\left(x_i^k\right)\·\mathrm{\φ}\left(x_j^k\right),$ J is the sequence number of sample, j=1, and 2 ..., N _k

Step (3) is judged the ownership of test sample book: the test sample book x that calculates input respectively is mapped to the ratio DR that distance that each minimum comprises the hypersphere centre of sphere and this minimum comprise the hypersphere radius in feature space _k(x) $D{R}_k\left(x\right)=\frac{D_k\left(x\right)}{R_k},$ K=1,2 ..., M, wherein D _k(x) be that test sample book x branch is clipped to the distance that each minimum comprises the hypersphere centre of sphere; Get DR _k(x) minimum of minima correspondence comprises the ownership of hypersphere as test sample book, finishes the identification of this test sample book.

Step (4) SVDD incremental learning: online support vector classification (OSVC) algorithm so that people such as Lau propose, designed the SVDD Incremental Learning Algorithm.This is mainly based on following 2 considerations: 1) work sample concentrate the less relatively support vector of quantity to represent whole training set; 2) after the sample of running counter to the KKT condition in the increment sample adds the work sample collection, just change historical training result, need train minimum to comprise hypersphere again.

The SVDD Incremental Learning Algorithm comprises that mainly sample adds algorithm and deletes algorithm two parts, and wherein the former is used for adding new sample and enters the training set, and the latter is used to delete old sample to keep the limited memory size of computer.Concrete steps comprise:

1. sample adds: be positioned at minimum and comprise support vector set on hypersphere outside or the sphere ${\mathrm{SV}}_k=\left\{x_i^k\right|{\mathrm{\α}}_i^k>0\},$ I=1,2 ..., N _k, and the non-support vector set that is positioned at hypersphere inside ${\mathrm{NSV}}_k=\left\{x_i^k\right|{\mathrm{\α}}_i^k=0\},$ i＝1，2，…，N _k。When a new samples joined one of them, some other existing sample can be because be subjected to the influence of new samples, from SV _kAnd NSV _kIn one be updated in another and go, therefore need to upgrade these two set, satisfy enclosed pasture-Plutarch (Karush-Kuhn-Tucker, KKT) condition to keep all samples.Concrete steps are as follows:

A. for k type games pattern, at first select two training samples, forming the initialization sample set is T ₂ ^k, $T_2^k=\{x_1^k,x_2^k\},$ Adopt the SVDD algorithm to obtain corresponding minimum and comprise hypersphere S ₂ ^k, two samples of this moment are support vector, and they finish the initialization of support vector set and the set of non-support vector, promptly ${\mathrm{SV}}_2^k=\{x_1^k,x_2^k\},$ ${\mathrm{NSV}}_2^k\∪{\mathrm{SV}}_2^k=T_2^k;$

B. obtain a new training sample x ₃ ^kAfter, if it is mapped to the ratio DR (x that distance that minimum comprises the hypersphere centre of sphere and minimum comprise the hypersphere radius in feature space ₃ ^k) less than 1, show x ₃ ^kComprise hypersphere inside in this minimum, and the adding of new samples can not change training result, therefore need not to train new minimum to comprise hypersphere, and order $S_3^k=S_2^k,$ ${\mathrm{SV}}_3^k={\mathrm{SV}}_2^k;$

If DR is (x ₃ ^k) more than or equal to 1, show x ₃ ^kComprise hypersphere S in minimum ₂ ^kOn outside or the sphere, and x ₃ ^kBe a support vector, it is joined in the work sample set, promptly $T_3^k={\mathrm{SV}}_2^k{\∪x}_3^k,$ Again train the minimum that makes new advances to comprise hypersphere S according to step (2) method ₃ ^k, determine that promptly minimum comprises hypersphere S ₃ ^kThe centre of sphere and radius.

C. proceed to n-1 (n 〉=4) during the step according to step b method, the support vector set that current minimum comprises the hypersphere correspondence is SV _N-1 ^k,

Be SV _N-1 ^kIn support vector, | SV _N-1 ^k| express support for the number of vector.At n training sample x _n ^kAfter obtaining, still use DR (x _n ^k) judge whether it can join the work sample collection.If it can not need not to train again as new training sample set, and order $S_n^k=S_{n-1}^k,$ ${\mathrm{SV}}_n^k={\mathrm{SV}}_{n-1}^k;$ Otherwise, order $T_n^k={\mathrm{SV}}_{n-1}^k{\∪x}_n^k,$ Again training obtains new minimum and comprises hypersphere S _n ^kAnd vector set SV _n ^kAnd NSV _n ^k

D. repeat above-mentioned steps c, online treatment is finished all N _kIndividual sample.

2. sample is deleted: when carrying out online training with the SVDD Incremental Learning Algorithm, along with the operation of algorithm, sample can get more and more, and the computer of operation algorithm has only limited memory size, therefore must design delete the algorithm of redundant samples.

For k type games pattern, if its sample number runs up to a certain degree, new samples of so every interpolation it's time to train set, just needs deletion one Geju City sample.Need to be moved out of training set fashionable whenever Geju City sample, and some other existing sample also may be affected, from SV _kAnd NSV _kOne in the set is updated in another and goes, and satisfies the KKT condition to keep all samples.Concrete steps are as follows:

E. n training sample x _n ^kAfter obtaining, if n＞N _Max, select the sample of deleting: at first calculate former working set T _N-1 ^kIn each sample at the distance D (x that minimum comprises the hypersphere centre of sphere that is mapped to of feature space _i ^k), $x_i^k\∈T_{n-1}^k,$ Select its middle distance minima then $(D_{\min }^k=\min \left(D\left(x_i^k\right)\right|x_i^k\∈T_{n-1}^k))$ Corresponding sample is old sample x to be deleted _Old

If f. old sample x _OldAt set NSV _N-1 ^kIn, then it is shifted out the training set, it so it is smaller to the influence of model to delete such sample, does not therefore need to do any adjustment to the not contribution of finding the solution of SVDD model;

If old sample x _OldAt S set V _N-1 ^kIn, x _OldFrom the training sample set, shift out, train the minimum that makes new advances to comprise hypersphere again, determine that promptly minimum comprises the centre of sphere and the radius of hypersphere according to step (2) method.

G. the training sample number is updated to $\hat{n}=n-1,$ Repeating step e, f, g.

The present invention compares with existing many multi-locomotion mode recognizers based on limb electromyographic signal in the above-knee lower limb vacation, has following characteristics:

1, can dynamically add emerging motor pattern

The motion of human body lower limbs is rich and varied, except this basic exercise of walking, also comprise up/down steps, race, rise and sit back and wait common motor pattern, and every kind of motor pattern can be subdivided into a plurality of stages again in a natural period.In above-knee prosthesis control, it is a process of learning gradually and accumulating that the test data of these motor patterns obtains, and is not easy to obtain comprehensive type testing data by once or several times testing, and therefore is difficult to set up complete multi-locomotion mode set.This just requires grader can dynamically add new motor pattern, to satisfy the multiformity of artificial leg motion.

Different with the SVM method is, the SVDD method is not to seek a hyperplane, come the distribution of data is described but comprise the hypersphere border by the minimum that calculating comprises one group of data, fall into suprasphere and belong to object set, the training sample of determining only to rely on object set of hypersphere with interior training data.Every kind of motor pattern can be learnt separately, obtains corresponding hypersphere respectively, and promptly various patterns are set up a grader respectively.This thought meets human learning process, at first in drops learns and accumulates, and makes a strategic decision on the basis of accumulation then.

2, can every kind of motor pattern characteristic of correspondence of dynamically adapting change

Because the electromyographic signal of the same action of different experimenters is had any different, and the same action that same experimenter does at different time also can be variant, therefore need be according to experimenter's different characteristics, feature space (being the SVDD hypersphere) to every kind of motor pattern is dynamically adjusted, to adapt to emerging changing features.

3, the SVDD incremental learning can be realized the interpolation of new samples and the deletion of old sample

The SVDD Incremental Learning Algorithm of design makes full use of existing training result, increases new samples and learn on the basis of original study, significantly reduces subsequent training time, reduces the requirement of algorithm to memory space.When sample constantly adds and reaches in the sample size in limited time, the replacement policy that comprises hypersphere centre of sphere distance according to sample to its feature space minimum, the old sample that the working set middle distance is less eliminates sample set, guarantees that simultaneously the existing sample that training sample is concentrated satisfies the KKT condition.Can under the situation that as far as possible keeps original sample information, in the multi-locomotion mode model of cognition, introduce the newly-increased sample that to represent new work information like this.

The inventive method can satisfy the multi-locomotion mode identification requirement in the above-knee artificial leg control preferably, the learning capacity of dynamically extracting the multiclass pattern feature is better than the single classification learning method of traditional SVDD, also overcome SVDD off-line training mode and can't effectively handle the shortcomings such as sample data of reflection object characteristic changing, this method has broad application prospects in the multi-mode identification of intelligent artificial limb control.

Description of drawings

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

Fig. 2 is that multi-locomotion mode identification realizes the system model sketch map among Fig. 1;

Fig. 3 is a SVDD Incremental Learning Algorithm flow chart among Fig. 2.

The specific embodiment

Describe the laptop artificial limb multi-movement-mode identifying method that the present invention is based on Support Vector data description in detail below in conjunction with accompanying drawing, Fig. 1 is implementing procedure figure, and Fig. 2 puies forward the realization system model of multi-locomotion mode recognition methods for the present invention.

As Fig. 1, the enforcement of the inventive method mainly comprises four steps: (1) obtains human body lower limbs electromyographic signal sample data, comprises collection, denoising Processing and the feature extraction of human body lower limbs electromyographic signal; (2) set up SVDD multicategory classification device, adopt the SVDD algorithm to train respectively, set up corresponding minimum and comprise hypersphere, described with the centre of sphere and radius to various motor pattern corresponding sample data; (3) calculate test sample book respectively and in feature space, be mapped to the ratio that distance that each minimum comprises the hypersphere centre of sphere and this minimum comprise the hypersphere radius, judge the ownership of test sample book according to the ratio size; (4) SVDD incremental learning is realized the interpolation of new samples and the deletion of old sample, and its flow chart as shown in Figure 3.

Multiple motor patterns such as the motion of human body lower limbs can be divided into away, runs, crouching, up/down steps, posture feature difference (as single foot support, both feet support etc.) according to human motion, each motor pattern further is subdivided into a plurality of stages in a natural period, promptly segments motor pattern.With the level walking is example, it is divided into 5 segmentation motor patterns such as supporting early stage, the mid-term of support, support later stage, swing early stage and swing later stage, so the gesture mode that needs to identify under the level walking situation has 5 kinds.At first, gather the electromyographic signal of 5 kinds of pattern correspondences, after data are carried out de-noising and feature extraction, obtain 5 groups of sample datas, therefrom select 2 samples to form the initialization collection of every kind of pattern respectively; Then, with SVDD every kind of pattern corresponding sample is carried out stand-alone training after, obtain the distribution characteristics space (minimum comprises the hypersphere hypersphere) of each pattern; Then, the classification that comprises the ratio in judgement test sample book of the hypersphere centre of sphere with the distance that is mapped to each hypersphere centre of sphere of test sample book in feature space and this minimum; At last, by all samples of SVDD incremental learning method online treatment, on the basis of existing feature distributed intelligence,, adaptively the feature space of every kind of pattern is adjusted according to learning outcome by upgrading the new sample information of operate learning.

Below one by one to each step be elaborated (as Fig. 2).

Step 1: obtain human body lower limbs electromyographic signal sample data

(1) electromyographic signal of collection human body lower limbs.Select the electromyographic signal source of the most representative several muscle on the thigh for use, such as the vastus lateralis of thigh front side, the semitendinosus m. of thigh rear side, the tensor fasciae latae that thigh links to each other with crotch, the adductor longus m. of femoribus internus as human body lower limbs.In the experiment, adopt MyoTrace 400 electromyographic signal collection instrument to pick up the thigh signal, gather the corresponding electromyographic signal of 4 muscle groups (vastus lateralis, semitendinosus m., tensor fasciae latae and adductor longus m.) on the thigh simultaneously.

(2) adopt spatial domain correlation filtering method to carry out denoising Processing.The spatial domain correlation filtering is based on a kind of noise-eliminating method of wavelet analysis, utilize the signal correlation properties different with noise, electromyographic signal is carried out wavelet transform after, the wavelet coefficient of each yardstick is carried out relevant treatment, in the hope of keeping the details of primary signal to greatest extent, remove noise jamming.In experiment, the original electromyographic signal of gathering is carried out 5 layers of wavelet decomposition, basic small echo is selected Bi-orthogonal Spline Wavelet Transformation bior1.5 for use.In the correlation filtering de-noising of spatial domain, need to set a certain noise threshold, with 80 variances that only contain the point estimation noise of noise at each layer of signal, and with these 10 times as the noise energy threshold value.

(3) electromyographic signal after adopting multiscale analysis method to de-noising is carried out feature extraction.After signal was through the spatial domain correlation filtering, the vector that the relative energy average of the wavelet coefficient after the de-noising on each yardstick constituted was shown below as the feature Tl of electromyographic signal.

$T_l=\frac{\sqrt{{\mathrm{\Σ}}_{n=1}^{N_s}W(l,n)}}{N_s},$ l＝1，2，…，L (1)

Wherein, W (l, the n) wavelet coefficient after the last position n place de-noising of expression yardstick l, N _sBe sampling number, the yardstick of L for decomposing.In experiment, utilize 5 layers of wavelet decomposition in the correlation filtering method of spatial domain, the base small echo is selected Bi-orthogonal Spline Wavelet Transformation bior1.5 for use, then selects the correlation coefficient of first three yardstick behind the correlation analysis for use, calculates the eigenvalue that respectively segments under the motor pattern according to formula (1).Because every road electromyographic signal characteristic of correspondence value has 3, gathers 4 muscle groups on the thigh simultaneously, therefore one is segmented down corresponding 12 eigenvalues of motor pattern, and each sample is made of the characteristic vector of one group of 12 dimension.

Step 2: set up SVDD multicategory classification device

Adopt the SVDD algorithm to train respectively to various segmentation motor pattern corresponding sample, set up different classes of minimum and comprise hypersphere, described with the centre of sphere and radius.Training sample set { the x of k type games pattern _i ^k,

I=1,2 ..., N _k, k=1,2 ..., M, the minimum that SVDD sought comprises hypersphere S _kAvailable centre of sphere a _kAnd radius R _kDescribe, the minimum that then satisfies condition comprises hypersphere and can try to achieve by following optimization problem

$\min F(R_k,a_k)={R_k}^2+C_k\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i^k$

$s.t.{\left|\right|\mathrm{\φ}\left(x_i^k\right)-a_k\left|\right|}^2\≤{R_k}^2+{\mathrm{\ξ}}_i^k---\left(2\right)$

${\mathrm{\ξ}}_i^k\≥0,$ i＝1，2，…，N _k

Wherein, φ (x _i ^k) be sample x _i ^kNonlinear mapping from the lower dimensional space to the higher dimensional space, slack variable ξ _i ^kRepresentative concentrates i sample to produce the wrong penalty term that divides to k class training sample, and ‖ ‖ is the Euclidean norm, and penalty factor C _kBeing used for the balance minimum comprises the size and the wrong ratio of dividing sample of hypersphere.

This has the quadratic programming problem of linear inequality constraint to adopt the Lagrange multiplier method to find the solution formula (2), obtains minimum and comprises hypersphere centre of sphere a _kAnd the dual problem of formula (2)

$a_k=\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^k\mathrm{\φ}\left(x_i^k\right)---\left(3\right)$

$\underset{\mathrm{\α}}{\max }\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{k}K(x_i^k,x_i^k)-\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^k{\mathrm{\α}}_j^{k}K(x_i^k,x_j^k)---\left(4\right)$

$s.t.\underset{i=1}{\overset{N_k}{\mathrm{\Σ}}}{\mathrm{\α}}_i^k=1,$ ${\mathrm{\α}}_i^k\∈[0,C],$ i＝1，2，…，N _k

Its KKT condition is:

${\mathrm{\α}}_i^k(R_k^2+{\mathrm{\ξ}}_i^k-{\left|\right|\mathrm{\φ}\left(x_i^k\right)-a_k\left|\right|}^2)=0---\left(5\right)$

${\mathrm{\β}}_i^k{\mathrm{\ξ}}_i^k=0$

Here, α _i ^kAnd β _i ^kBe the Lagrange multiplier.Can get by formula (5),, then work as if sample satisfies the KKT condition ${\mathrm{\&alpha;}}_i^k=0$ The time, sample is positioned at minimum and comprises hypersphere inside; When $0<{\mathrm{\&alpha;}}_i^k<C_k$ The time, sample is positioned at minimum and comprises on the hypersphere; When ${\mathrm{\&alpha;}}_i^k=C_k$ The time, sample is positioned at minimum and comprises hypersphere outside (may on sphere).

Kernel function $K(x_i^k,x_j^k)=\mathrm{\&phi;}\left(x_i^k\right)\&CenterDot;\mathrm{\&phi;}\left(x_j^k\right)$ Introducing avoided inner product operation in the higher dimensional space calculation of complex.The kernel function of present embodiment is chosen radially base nuclear of Gauss, sees formula (6).

$K(x_i^k,x_j^k)=\exp (-{\mathrm{\&gamma;}}_k{\left|\right|x_i^k-x_j^k\left|\right|}^2)---\left(6\right)$

γ wherein _kBe the nuclear parameter of gaussian radial basis function, then K (x, x)=1.At this moment, dual problem (4) is equivalent to

$\underset{\mathrm{\&alpha;}}{\max }1-\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_i^k-\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1,j\&NotEqual;i}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_j^{k}K(x_i^k,x_j^k)---\left(7\right)$

$s.t.\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k=1,$ ${\mathrm{\&alpha;}}_i^k\&Element;[0,C],$ i＝1，2，…，N _k

After training finished, for a given test sample book point x, if it satisfies following decision condition, then this test point was accepted, otherwise is rejected.

$K(x,x)-2\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{k}K(x,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_j^{k}K(x_i^k,x_j^k)$

$=1-2\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{k}K(x,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_j^{k}K(x_i^k,x_j^k)---\left(8\right)$

$\&le;{R_k}^2$

R wherein _kBe the radius that k minimum comprises hypersphere, comprise point on the hypersphere sphere by asking minimum $(0<{\mathrm{\&alpha;}}_i^k<C_k)$ The distance that comprises the hypersphere centre of sphere to this minimum obtains, and establishes x _q ^kBe to be positioned at minimum to comprise point on the hypersphere, then

${R_k}^2=K(x_q^k,x_q^k)-2\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{k}K(x_q^k,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_j^{k}K(x_i^k,x_j^k)---\left(9\right)$

$=1-2\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{k}K(x_l^k,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_j^{k}K(x_i^k,x_j^k)$

SVDD is used for classification learning, needs to select the nuclear parameter γ and the penalty factor C of gaussian radial basis function _kIn experiment, adopt lattice search or many times of sample crosscheck methods, obtain best parameter combination (C _k, γ _k), take into account the popularization ability of nicety of grading and training result.

Step 3: the ownership of judging test sample book

Calculate the radius R that each minimum comprises hypersphere respectively _k, k=1,2 ..., M, and given test sample book point x is mapped to the distance D that each minimum comprises the hypersphere centre of sphere in feature space _k(x), k=1,2 ..., M obtains D then _k(x) and R _kRatio DR _k, shown in following formula

${\mathrm{DR}}_k=\frac{D_k\left(x\right)}{R_k},$ k＝1，2，…，M

${D_k}^2\left(x\right)=K(x,x)-\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K(x,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_{j}K(x_i^k,x_j^k)---\left(10\right)$

$=1-\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K(x,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_{j}K(x_i^k,x_j^k)$

Here, work as DR _k(x)＜1 o'clock, sample is positioned at hypersphere inside; Work as DR _k(x)=1 o'clock, sample is positioned on the hypersphere; Work as DR _k(x)＞1 o'clock, sample is positioned at the hypersphere outside.Get DR _k(x) minimum of minima correspondence comprises the ownership of hypersphere as test sample book, finishes the identification of this test sample book.

Step 4: SVDD incremental learning

In the incremental learning implementation process, select two training samples to form the initialization sample set earlier, constantly add new samples then, thereby the sample size that makes work sample concentrate is on the increase; After sample size reached the maximum sample number, the old sample of deletion added new samples more earlier, trains again according to the step 2 method then.All samples of online treatment are on the basis of existing feature distributed intelligence, by upgrading the new sample information of operate learning.

To be positioned at the support vector set that minimum comprises on hypersphere outside or the sphere is designated as ${\mathrm{SV}}_k=\left\{x_n^k\right|{\mathrm{\&alpha;}}_n^k>0\},$ N=1,2 ..., N _Max, the set that n sample is added the sample composition of the fashionable KKT of running counter to condition is designated as $E_n^k={\left\{x_i^k\right|\mathrm{DR}\left(x_i^k\right)>1\}}_{i=1}^n.$ Concrete steps are as follows:

(1) for k type games pattern, select 2 training samples to form the initialization sample set $T_2^k=\{x_1^k,x_2^k\},$ Adopt the SVDD algorithm to obtain corresponding minimum and comprise hypersphere S ₂ ^k, then ${\mathrm{SV}}_2^k=\{x_1^k,x_2^k\}),$ And initialization $\left|E_2^k\right|=0;$

(2) obtain a new training sample x ₃ ^kAfter, if its minimum that is mapped in feature space comprises the distance of the hypersphere centre of sphere and the ratio DR (x of hypersphere radius ₃ ^k) less than 1, need not to train new minimum to comprise hypersphere, and order $S_3^k=S_2^k,$ ${\mathrm{SV}}_3^k={\mathrm{SV}}_2^k;$ If DR is (x ₃ ^k) more than or equal to 1, x ₃ ^kJoin in the work sample set (promptly $T_3^k={\mathrm{SV}}_2^k{\&cup;x}_3^k$ ), training obtains new hypersphere S again ₃ ^k, to hold new samples;

(3) as n 〉=3 a training sample x _n ^kAfter obtaining, with DR (x _n ^k) judge whether it can join the work sample collection: if x _n ^kCan not need not to train again as new training sample set, and order $S_n^k=S_{n-1}^k,$ ${\mathrm{SV}}_n^k={\mathrm{SV}}_{n-1}^k,$ Jump to step (7); Otherwise, execution in step (4);

(4) judge that whether n is greater than the maximum sample number N _MaxIf: n＞N _Max, then jump to step (6), otherwise execution in step (5);

(5) new samples is joined the work sample collection, order $T_n^k={\mathrm{SV}}_{n-1}^k{\&cup;x}_n^k,$ Again training obtains new minimum and comprises hypersphere S _n ^kAnd support vector S set V _n ^k, and will $E_n^k={\left\{x_i^k\right|\mathrm{DR}\left(x_i^k\right)>1\}}_{i=1}^n$ Join in next time the training process as new samples.After executing this step, jump to step (7);

(6) calculate former working set T _N-1 ^kIn each sample at the distance D (x that minimum comprises the hypersphere centre of sphere that is mapped to of feature space _i ^k), $x_i^k\&Element;T_{n-1}^k,$ According to $D_{\min }^k=\min \left(D\left(x_i^k\right)\right|x_i^k\&Element;T_{n-1}^k)$ Select corresponding sample, and from working set, shift out, jump to step (5);

(7) the training sample number is updated to

$\hat{n}=n+1,$ Repeating step (3) circulation is gone down.

Claims (1)

1. based on the laptop artificial limb multi-movement-mode identifying method of Support Vector data description, it is characterized in that this method may further comprise the steps:

Step (1) is obtained human body lower limbs electromyographic signal sample data, specifically: at first pick up the human body lower limbs electromyographic signal by the electromyographic signal collection instrument, adopt spatial domain correlation filtering method that the electromyographic signal that contains interference noise is carried out de-noising again, electromyographic signal after adopting multiscale analysis method to de-noising is then carried out feature extraction, and the vector that the relative energy average of the wavelet coefficient after the de-noising on each yardstick constituted is as the feature T of electromyographic signal ₁,

$T_1=\frac{\sqrt{{\mathrm{\&Sigma;}}_{n=1}^{N_s}W(l,n)}}{N_s},l=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,L$

Wherein, W (l, the n) wavelet coefficient after the n place de-noising of position on the expression yardstick 1, N _sBe sampling number, the yardstick of L for decomposing;

After M kind segmentation motor pattern carried out feature extraction, the characteristic of correspondence value constituted an eigenvectors under every kind of segmentation motor pattern, and every eigenvectors is as a kind of segmentation motor pattern corresponding sample

K is the sequence number of segmentation motor pattern, k=1, and 2 ..., M, i are the sequence number of sample, i=1, and 2 ..., N _k, N _kIt is the number of sample in the k kind segmentation motor pattern;

Step (2) is set up Support Vector data description multicategory classification device: adopt the Support Vector data description algorithm to train respectively to M kind segmentation motor pattern corresponding sample data, set up M different classes of minimum and comprise hypersphere S _k, k=1,2 ..., M uses centre of sphere a _kAnd radius R _kDescribe minimum and comprise hypersphere S _k, each minimum comprises hypersphere as the class in the grader;

$a_k=\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k\mathrm{\&phi;}\left(x_i^k\right),k=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,M,$

${R_k}^2=K(x_q^k,x_q^k)-2\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{k}K(x_q^k,x_i^k)+\underset{i=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\stackrel{N_k}{\underset{j=1}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^k{\mathrm{\&alpha;}}_j^{k}K(x_i^k,x_j^k),$

With

Be the Lagrange multiplier,

Be sample

Nonlinear mapping from the lower dimensional space to the higher dimensional space;

Be to be positioned at minimum to comprise hypersphere S _kOn point, q is the sequence number of sample, q ∈ 1,2 ..., N _k;

Be kernel function, J is the sequence number of sample, j=1, and 2 ..., N _k

Step (3) is judged the ownership of test sample book: the test sample book x that calculates input respectively is mapped to the ratio DR that distance that each minimum comprises the hypersphere centre of sphere and this minimum comprise the hypersphere radius in feature space _k(x)

D wherein _k(x) be that test sample book x branch is clipped to the distance that each minimum comprises the hypersphere centre of sphere; Get DR _k(x) minimum of minima correspondence comprises the ownership of hypersphere as test sample book, finishes the identification of this test sample book;

Step (4) Support Vector data description incremental learning, concrete steps comprise

1. sample adds, and concrete steps are as follows:

A. for k type games pattern, at first select two training samples, composition initialization sample set is

Adopt the Support Vector data description algorithm to obtain corresponding minimum and comprise hypersphere

Two samples of this moment are support vector, and they finish the initialization of support vector set and the set of non-support vector, promptly

Being positioned at the support vector set that minimum comprises on hypersphere outside or the sphere is expressed as

The non-support vector set that is positioned at hypersphere inside is expressed as ${\mathrm{NSV}}_k=\left\{x_i^k\right|{\mathrm{\&alpha;}}_i^k=0\},i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;N_k;$

B. obtain a new training sample

After, if its minimum that is mapped in feature space comprises hypersphere

The distance of the centre of sphere comprises hypersphere with minimum

The ratio of radius

Less than 1, order

$S_3^k=S_2^k,$ ${\mathrm{SV}}_3^k={\mathrm{SV}}_2^k;$

If

More than or equal to 1, it is joined in the work sample set, promptly

Again train the minimum that makes new advances to comprise hypersphere according to step (2) method Determine that promptly minimum comprises hypersphere The centre of sphere and radius;

C. proceed to n-1 during the step according to step b method, n 〉=4, the support vector set that current minimum comprises the hypersphere correspondence is

Be

In support vector,

Express support for the number of vector; At n training sample

After obtaining, if

Less than 1, order

$S_n^k=S_{n-1}^k,$ ${\mathrm{SV}}_n^k={\mathrm{SV}}_{n-1}^k$

If

More than or equal to 1, order

Again train the minimum that makes new advances to comprise hypersphere according to step (2) method And vector set

With

D. repeat above-mentioned steps c, online treatment is finished all N _kIndividual sample;

2. sample is deleted, concrete steps are as follows:

E. n training sample

After obtaining, if n＞N _Max, N _MaxBe the maximum sample number, select the sample of deleting: at first calculate former working set

In each sample in the distance that minimum comprises the hypersphere centre of sphere that is mapped to of feature space

Select its middle distance minima then

Corresponding sample is old sample x to be deleted _Old

If f. old sample x _OldIn set

In, then it is shifted out the training set;

If old sample x _OldIn set

In, x _OldFrom the training sample set, shift out, train the minimum that makes new advances to comprise hypersphere again, determine that promptly minimum comprises the centre of sphere and the radius of hypersphere according to step (2) method;

G. the training sample number is updated to Repeating step e, f, g.

Images