CN103440498A - Surface electromyogram signal identification method based on LDA algorithm - Google Patents

Abstract

The invention discloses a surface electromyogram signal identification method based on an LDA algorithm. The surface electromyogram signal identification method is used for identifying up to eight kinds of grabbing gestures. According to the surface electromyogram signal identification method, only two electromyogram electrodes are utilized to collect surface electromyogram signals of corresponding gestures from related muscles of a forearm of a tester at first, then original electromyogram signals are segmented in an overlapping windowing mode, and absolute mean values, variances and 4-order AR coefficients are extracted from various windows to serve as original electromyogram characteristics; the LDA algorithm is utilized to carry out dimensionality reduction on the original electromyogram characteristics, redundant information is removed to the maximum degree, useful information is kept, and characteristics after the dimensionality reduction are obtained; the mean value of dimensionality reduction characteristics of front and back adjacent windows is computed and is inputted to an LDA classifier, and effective identification for the eight kinds of grabbing gestures is achieved. According to the surface electromyogram signal identification method, the electromyogram signal identification rate for the various kinds of gestures is high, the whole signal processing process is simple in computation and low in time consumption, and the requirement for the real-time performance of an electromyogram control system is met.

Description

Surface electromyogram signal recognition methods based on the LDA algorithm

Technical field

The present invention relates to area of pattern recognition, particularly grasp in the situation of gesture in multiclass, the judgement of effects on surface electromyographic signal identification, can be applicable to control EMG-controlling prosthetic hand and other Man Machine Interfaces.

Background technology

Surface electromyogram signal (surface electromyography, sEMG) is a kind of bioelectrical signals movable relevant to neuromuscular.When movement instruction conducts to relevant muscle fibre via the nerve center system, can cause the myofibrillar contraction of the concurrent life of potential change on muscle fibre, the stack of this potential change on skin surface place time of origin and space and form surface electromyogram signal, can collect by the surface myoelectric electrode.The information such as the pattern that surface electromyogram signal has comprised contraction of muscle and contraction intensity, the electromyographic signal that different limb actions is corresponding different, just can determine the corresponding concrete pattern of this signal by analyzing surface electromyogram signal.Therefore surface electromyogram signal is widely used in the fields such as medical diagnosis, athletic rehabilitation, and especially in apery myoelectricity artificial hand is controlled, surface electromyogram signal drives to do evil through another person as the control source and makes various grasping gestures and obtained and study widely and pay close attention to.Such as employing variance, the zero passages such as Huang count, AR model coefficient and spectrum estimate as feature, utilize the BP network to attempt identification 8 class gestures: three fingers are pinched and are got, side is pinched and got, hooks up, brute force captures, cylinder captures, center captures, palm stretches and wrist flexion, the discrimination of acquisition off-line test average 85% and the discrimination of on-line testing average 71%.Yang great Peng is only used waveform length (W/L) as the myoelectricity feature, utilizes SVM to identify successfully to have identified cylinder crawl, spherical crawl, side not to pinch to get with three fingers and pinches and get four classes and grasp gestures, obtains online discrimination more than 95%.The utilizations such as Matrone rhetorical question PCA algorithm has been identified powerful crawl preferably, three fingers are pinched to get with side and pinched and get three class gestures, obtain average 94% discrimination.Visible when the gesture classification is less, discrimination is higher; When the gesture classification is more, it is more that discrimination descends.Therefore be necessary to explore and a kind ofly can accurately identify the myoelectricity recognizer that multiclass grasps gesture with higher discrimination.

Summary of the invention

For above-mentioned technical matters, the present invention is intended to solve the problems of the technologies described above at least to a certain extent.

Accurately identify eight classes for the discrimination with higher and grasp gestures, comprise that cylinder captures, hooks up, side is pinched and got, sensing, spherical crawl, three fingers pinch and get, accurately pinch and get and loosen attitude, the present invention proposes a kind of electromyographic signal method of discrimination based on the LDA algorithm.Hundred first utilize two pieces of bipolar differential electrodes of modular collection surface electromyographic signal from corresponding forearm muscle; Then cut apart original electromyographic signal in the mode of overlapping windowing, extract absolute mean (MAV), variance (VAR) and 4 rank AR coefficients as original myoelectricity feature from each window; Recycling LDA algorithm carries out dimensionality reduction to original myoelectricity feature, removes to greatest extent redundant information and retains useful information, feature after the acquisition dimensionality reduction; Then the dimensionality reduction feature of front and back adjacent window apertures is averaged, then input the LDA sorter, realize that eight classes grasp effective identification of gesture.

The present invention adopts following technical scheme:

Surface electromyogram signal recognition methods based on the LDA algorithm, the method comprises the steps:

Step 1, cleaning skin, strike off the fine hair of selected muscle place epidermis, with clear water, cleans and dip medicinal alcohol wiping skin with cotton swab;

Step 2, gather experimenter's forearm surface electromyogram signal, the experimenter makes that cylinder captures, hooks up, side is pinched and got, sensing, spherical crawl, three fingers are pinched and got, accurately pinch to get and loosen attitude eight classes and grasp gestures, uses two pieces of electromyographic electrodes to obtain myoelectricity data the storage of each gesture.

Step 3 pair electromyographic signal is carried out windowing and is cut apart, and the mode of the overlapping windowing of the original myoelectricity the data of each gesture is cut apart, and obtains the window sample.Length of window is defined as 250ms herein, and the window increment is 50ms.

Step 4, the myoelectricity feature of calculation window sample, the myoelectricity feature of selection absolute mean, variance and 4 rank AR coefficient calculations window samples, described absolute mean, variance and 4 rank AR coefficient formulas are as follows respectively:

Absolute mean: $\mathrm{MAV}=\frac{1}{N}{\mathrm{\Σ}}_{i=1}^N\left|X_i\right|$

Variance: $\mathrm{VAR}=\frac{1}{N-1}{\mathrm{\Σ}}_{i=1}^N{x}_i^2$

4 rank AR coefficients: $x_i={\mathrm{\Σ}}_{k=1}^4{a}_k{x}_{i-k}+w_i$

Wherein N=250, be data point number in window, a _k(k=1,2,3,4) are the AR coefficient, w _ifor the white noise residual error;

Step 5, carry out dimension-reduction treatment to the myoelectricity feature, utilizes the LDA algorithm to carry out dimension-reduction treatment to the myoelectricity characteristic series vector of trying to achieve in step 3;

Step 6, ask for the mean value of the characteristic series vector after each gesture dimensionality reduction, in the vector space after dimensionality reduction, characterizes the categorization vector of this gesture;

Step 7, calculate self-test data a window sample the myoelectricity feature and carry out dimension-reduction treatment, the Euclidean distance of itself and each gesture categorization vector relatively in the input sorter, judge which kind of gesture is this myoelectricity feature belong to.If it is to the Euclidean distance minimum of i class categorization vector, it belongs to the i class.The present invention adopts first to the dimensionality reduction of front and back adjacent window apertures sample, rear the characteristic series vector is averaging, then mean value is inputted to sorter and identify the scheme judged, has improved recognition accuracy.

Further, described dimension-reduction treatment is based on finding a suitable projection matrix W, make raw data set M (P * q) obtain the expression under new coordinate system space after the W conversion, can effectively reduce the dimension of raw data set, and distinguish preferably the concentrated inhomogeneity data of raw data, concrete steps are:

Step 501, hash matrix S between compute classes _w

$S_W={\mathrm{\Σ}}_{i=1}^C{\mathrm{\Σ}}_{j=1}^{N_i}(m_j-u_i){(m_j-u_i)}^T;$

Step 502, hash matrix S in compute classes _b

$S_B={\mathrm{\Σ}}_{i=1}^C{N}_i(u_i-u){(u_i-u)}^T;$

Step 503, calculate projection matrix W

$J\left(W\right)=\frac{\det \left(W^T{S}_{B}W\right)}{\det \left(W^T{S}_{W}W\right)},$

In above-mentioned three formulas, C means the classification of data, N _ithe sample number that means each classification, m _jmeaning each data in each class, is a column vector, u _imean mean value of all categories, u means the population mean of all categories data.In the process that solves projection matrix W, hundred first obtain S _w ^-1s _beigenwert, form the W matrix and get final product by then getting front K eigenwert characteristic of correspondence vector after descending sort;

Step 504, carry out projective transformation by raw data set through matrix W

Y=w ^TM

The new matrix of Y in formula (R * S) for obtaining after projection, the dimension of its each column data is down to R (R≤p, R≤C-l) by p.This has alleviated the computation burden of follow-up sorter to a certain extent.

Further, described myoelectricity Feature Dimension Reduction dimension is 7 dimensions.Make characteristic series vector after dimensionality reduction can farthest retain the information of primitive character column vector, improve last recognition accuracy and reduce the burden of sorter.

Eight classes to be identified grasp bending and the stretching, extension that gesture relates generally to thumb and all the other four fingers, for the rationality that guarantees that electrode position is placed, make it to extract the related muscles electric signal as far as possible, get rid of irrelevant muscle electric signal, must be familiar with the muscle relevant with hand motion and distribute.Distribute and corresponding function according to forearm muscle in human anatomy, select respectively long flexor muscle of thumb and musculus flexor digitorum sublimis place respectively to place one piece of electromyographic electrode.

Before starting to gather electromyographic signal, experimenter's right forearm skin need to be processed, thereby to reduce the surface electromyogram signal of the impedance acquisition high s/n ratio between electrode and skin.Hundred first strike off the fine hair at electrode and skin attachement place, with clear water, clean, and then use cotton swab to dip in medicinal alcohol skin is carried out disinfection, then electrode is close to skin is fixed on experimenter's forearm.Approximately after 5 minutes, treat that electrode and skin surface are fitted fully and start again electromyographic signal collection.

The experimenter makes corresponding gesture under the prompting of visual signal, continues 5 seconds, and sample frequency lKHz obtains the initial surface electromyographic signal.For the calculated amount that reduces the myoelectricity feature and the requirement of real-time that meets electromyographic signal identification, need to carry out windowing to the initial surface electromyographic signal and cut apart.Here select common rectangular window, length of window 250ms, the increment of window between window is 50ms.

Select absolute mean (MAV), variance (VAR) and 4 rank AR coefficients as the myoelectricity feature, to these features of each window calculation of dividing in step B, can obtain the primitive character column vector of one 12 dimension.To each window signal x (t), the computing formula of these features is as follows respectively:

Absolute mean: $\mathrm{MAV}=\frac{1}{N}{\mathrm{\Σ}}_{i=1}^N\left|X_i\right|$

Variance: $\mathrm{VAR}=\frac{1}{N-1}{\mathrm{\Σ}}_{i=1}^N{x}_i^2$

4 rank AR coefficients: $x_i={\mathrm{\Σ}}_{k=1}^4{a}_k{x}_{i-k}+w_i$

Wherein N=250 is data point number in window, a _k(k=1,2,3,4) are the AR coefficient, w _ifor the white noise residual error.

Utilize the LDA algorithm to carry out dimension-reduction treatment to the original feature vector of extracting.Its basic thought is to find a suitable projection matrix W, make raw data set M (P * q) obtain the expression under new coordinate system space after the W conversion, can effectively reduce the dimension of raw data set, and distinguish preferably the concentrated inhomogeneity data of raw data, concrete steps are:

Step 501, hash matrix S between compute classes _w

$S_W={\mathrm{\Σ}}_{i=1}^C{\mathrm{\Σ}}_{j=1}^{N_i}(m_j-u_i){(m_j-u_i)}^T;$

Step 502, hash matrix S in compute classes _b

$S_B={\mathrm{\Σ}}_{i=1}^C{N}_i(u_i-u){(u_i-u)}^T;$

Step 503, calculate projection matrix W

$J\left(W\right)=\frac{\det \left(W^T{S}_{m}W\right)}{\det \left(W^T{S}_{W}W\right)},$

In above-mentioned three formulas, C means the classification of data, N _ithe sample number that means each classification, m _jmeaning each data in each class, is a column vector, u _imean mean value of all categories, u means the population mean of all categories data.In the process that solves projection matrix W, hundred first obtain S _w ^-1s _b, eigenwert, form the W matrix and get final product by then getting front K eigenwert characteristic of correspondence vector after descending sort;

Step 504, carry out projective transformation by raw data set through matrix W

Y=w ^TM

The new matrix of Y in formula (R * S) for obtaining after projection, the dimension of its each column data is down to R (R≤p, R≤C-l) by p.This has alleviated the computation burden of follow-up sorter to a certain extent;

Using the surface electromyogram signal that gathers half as training data, it is completed to windowing, feature extraction and Feature Dimension Reduction processes, feature samples after the acquisition dimensionality reduction, the feature samples of each class gesture presents higher separability, feature samples to all kinds of gestures is averaged, and just can obtain a R dimension categorization vector that characterizes such gesture in vector space.

At great majority, utilize during LDA carries out the work of electromyographic signal classification, myoelectricity feature to be sorted is directly sent into the LDA sorter, by calculating myoelectricity feature to be sorted to the Euclidean distance that characterizes each gesture categorization vector, judge which kind of gesture is this myoelectricity feature belong to.If it is to the Euclidean distance minimum of i class categorization vector, it belongs to the i class.But this way can not obtain higher recognition accuracy.The present invention finds in the reduction process of step 5, and the myoelectricity proper vector after dimensionality reduction is the state of stochastic distribution in vector space.According to these characteristics, the present invention proposes j proper vector and j+l proper vector after dimensionality reduction are first averaged, and can be expected that the categorization vector that this mean value can more close corresponding gesture.

The present invention compares with existing various electromyographic signal recognition methodss, and the advantage had and good effect are: the present invention only need to gather two channel surface electromyographic signals, and equipment is simple, and the calculated signals amount is less; In LDA Feature Dimension Reduction process, choose suitable dimension R, can preserve to greatest extent the essential information of original myoelectricity feature; The first average treatment that the present invention proposes is the algorithm of LDA classification again, can significantly improve the discrimination of eight class gestures; Whole signal processing of the present invention calculates simple, consuming time few, sends the corresponding sports instruction from experimenter's brain and accurately identifies in this signal 150ms of being controlled at consuming time to the myoelectricity system, meets the requirement of real-time of myoelectric control system.

The accompanying drawing explanation

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

Fig. 2 stack window scheme of attaching most importance to.

Embodiment

Below in conjunction with drawings and Examples, the present invention is further detailed explanation.

It is example that eight classes to be identified of now take grasp gesture, and the technical scheme proposed in conjunction with the present invention, provide detailed operation steps and concrete recognition result.

Step 1, cleaning skin, strike off the fine hair of selected muscle place epidermis, with clear water, cleans and dip medicinal alcohol wiping skin with cotton swab, according to the muscle distribution situation of human body forearm, selects corresponding muscle.Relate generally to the crooked of thumb and all the other four fingers and stretch because this eight class grasps gesture, therefore selecting long flexor muscle of thumb and the musculus flexor digitorum sublimis collection source as this surface electromyogram signal;

Step 2, gather experimenter's forearm surface electromyogram signal, after dry skin, put electrode and with the skin close contact.Approximately after 5 minutes, treat that electrode and skin fits fully, after contact conditions is good, start to gather experimenter's surface electromyogram signal, the experimenter makes cylinder and captures, hook up, side is pinched and is got, point to, spherical crawl, three fingers are pinched and are got, accurately pinch to get and loosen attitude eight classes and grasp gesture, use two pieces of electromyographic electrodes to obtain myoelectricity data the storage of each gesture, the experimenter is according to cue, making respectively cylinder captures, hook up, side is pinched and is got, point to, spherical crawl, three fingers are pinched and are got, accurately pinch and get and loosen attitude eight class gestures, every class gesture duration of contraction 5s, respectively do 10 groups, between group and group, interval is 5 seconds, to prevent muscular fatigue.Adopt DELSYS Table top type two passage myoelectricity Acquisition Instruments carry out the collection surface electromyographic signal and complete storage in conjunction with data collecting card PCI-6220, sample frequency is lKHz, and collection obtains 80 groups of myoelectricity data after finishing altogether;

Step 3, carry out windowing to electromyographic signal and cut apart, and the mode of the overlapping windowing of the original myoelectricity the data of each gesture is cut apart, obtain the window sample, length of window is defined as 250ms herein, and the window increment is 50ms, what with cylinder, capture gesture once is punctured into example, can obtain altogether 5000 sampled points.Have and beat because data acquisition is stuck in the moment that starts to gather, therefore need to remove the some data that start most, to guarantee not affect follow-up Classification and Identification, remove 250 sampled points that start most herein.To the residue sampled point, adopt rectangular window to be cut apart, length of window 250ms, window increment 50ms, single shrinks and can obtain altogether 91 window samples, shrinks for 10 times and obtains altogether 910 window samples.All the other gestures also will produce 910 window samples separately.Following formula under overlapping windowing scheme, the method for calculation window sample size:

Window sample number=(data total length-length of window)/window increment+1;

Step 4, the myoelectricity feature of calculation window sample, by half of 910 window samples of each gesture, totally 455 * 8 window samples are as training data, extract its primitive character, calculate absolute mean, variance and 4 rank AR coefficients, each gesture will obtain the primitive character column vector of 455 12 dimensions, select the myoelectricity feature of absolute mean, variance and 4 rank AR coefficient calculations window samples, described absolute mean, variance and 4 rank AR coefficient formulas are as follows respectively:

Absolute mean: $\mathrm{MAV}=\frac{1}{N}{\mathrm{\Σ}}_{i=1}^N\left|X_i\right|---\left(1\right)$

Variance: $\mathrm{VAR}=\frac{1}{N-1}{\mathrm{\Σ}}_{i=1}^N{x}_i^2---\left(2\right)$

4 rank AR coefficients: $x_i={\mathrm{\Σ}}_{k=1}^4{a}_k{x}_{i-k}+w_i---\left(3\right)$

Wherein N=250, be data point number in window, a _k(k=1,2,3,4) are the AR coefficient, w _ifor the white noise residual error;

Step 5, the myoelectricity feature is carried out to dimension-reduction treatment, utilize the LDA algorithm to carry out dimension-reduction treatment to the myoelectricity characteristic series vector of trying to achieve, the matrix M (l2 * 455) that the training data of all kinds of gestures is formed, according to hash matrix S between formula (4) and (5) compute classes _wwith hash matrix S in class _b, continuing to solve S _w ^-1s _bin the eigenwert process of footwear, get by the corresponding proper vector of front 7 eigenwerts of descending sort and form projection matrix W (l2 * 7).The primitive character column vector is carried out to dimensionality reduction by formula (7), and each gesture training data matrix becomes Y (7 * 455).In the present invention, determine that best dimensionality reduction dimension is 7 dimensions,, when 12 dimension primitive character column vectors are down to 7 dimension, the characteristic series vector after dimensionality reduction can at utmost retain the information of primitive character column vector, and contributes to improve follow-up Classification and Identification accuracy rate;

Step 6, ask for the mean value of the characteristic series vector after each gesture dimensionality reduction, in the vector space after dimensionality reduction, characterizes the categorization vector of this gesture;

Step 7, calculate self-test data a window sample the myoelectricity feature and carry out dimension-reduction treatment, the Euclidean distance of each gesture categorization vector in itself and step 6 relatively in the input sorter, judge which kind of gesture is this myoelectricity feature belong to.The Euclidean distance of characteristic series vector to eight categorization vector after the calculating dimensionality reduction, the gesture of test data representative is with it the gesture of nearest categorization vector representative.If recognition result is consistent with test target, illustrates test data has been carried out to correct classification, otherwise be exactly mis-classification.Table 1 has provided the identified off-line result of experimenter's 8 classes grasping gestures.

Table 1 eight classes grasp gesture identified off-line result

The identified off-line accuracy rate is higher as seen from the table, and eight class gesture average recognition rate can reach 96.59%, illustrates that training process is effective.Preserve the projection matrix W obtained and eight categorization vectors in training process, for next step real-time online gesture identification.

Described dimension-reduction treatment is based on finding a suitable projection matrix W, make raw data set M (P * q) obtain the expression under new coordinate system space after the W conversion, can effectively reduce the dimension of raw data set, and distinguish preferably the concentrated inhomogeneity data of raw data, concrete steps are:

Step 501, hash matrix S between compute classes _w

$S_W={\mathrm{\Σ}}_{i=1}^C{\mathrm{\Σ}}_{j=1}^{N_i}(m_j-u_i){(m_j-u_i)}^T;---\left(4\right)$

Step 502, hash matrix S in compute classes _b

$S_B={\mathrm{\Σ}}_{i=1}^C{N}_i(u_i-u){(u_i-u)}^T;---\left(5\right)$

Step 503, calculate projection matrix W

$J\left(W\right)=\frac{\det \left(W^T{S}_{B}W\right)}{\det \left(W^T{S}_{W}W\right)};---\left(6\right)$

In above-mentioned three formulas, C means the classification of data, N _ithe sample number that means each classification, m _jmeaning each data in each class, is a column vector, u _imean mean value of all categories, u means the population mean of all categories data.In (6) formula solves the process of projection matrix W, hundred first obtain S _w ^-1s _beigenwert, form the W matrix and get final product by then getting front K eigenwert characteristic of correspondence vector after descending sort;

Step 504, carry out projective transformation by raw data set through matrix W

Y=w ^TM (7)

The new matrix of Y in formula (R * S) for obtaining after projection, the dimension of its each column data is down to R (R≤p, R≤C-l) by p.

In the identifying of real-time online, input a window sample, complete feature extraction and Feature Dimension Reduction, then the direct distance of the characteristic series vector after dimensionality reduction and each categorization vector relatively.After having adopted the dimensionality reduction that front and back are adjacent in this example, the characteristic series vector is averaging, then the distance of this average and vector of all categories relatively.The ONLINE RECOGNITION rate that table 2 has provided average front and back changes.Visible the average treatment that example adopted before the LDA classification, can improve to a certain extent the online Real time identification rate of most of gesture, and improve on the whole average recognition rate, makes this myoelectricity identifying schemes more reliable in actual applications.

The contrast of ONLINE RECOGNITION rate before and after table 2 is average

The above embodiment of the present invention is only for example of the present invention clearly is described, and is not the restriction to embodiments of the present invention.For those of ordinary skill in the field, can also make other changes in different forms on the basis of the above description.Here exhaustive without also giving all embodiments.All any modifications of doing within the spirit and principles in the present invention, be equal to and replace and political affairs are spouted etc., within all should being included in the protection domain of the claims in the present invention.

Claims (3)

1. the surface electromyogram signal recognition methods based on the LDA algorithm, is characterized in that, the method comprises the steps:

Step 1, cleaning skin, strike off the fine hair of selected muscle place epidermis, with clear water, cleans and dip medicinal alcohol wiping skin with cotton swab;

Step 2, gather experimenter's forearm surface electromyogram signal, the experimenter makes that cylinder captures, hooks up, side is pinched and got, sensing, spherical crawl, three fingers are pinched and got, accurately pinch to get and loosen attitude eight classes and grasp gestures, uses two pieces of electromyographic electrodes to obtain myoelectricity data the storage of each gesture.

Step 3, carry out windowing to electromyographic signal and cut apart, and the mode of the overlapping windowing of the original myoelectricity the data of each gesture is cut apart, and obtains the window sample.Length of window is defined as 250ms herein, and the window increment is 50ms.

Step 4, the myoelectricity feature of calculation window sample, the myoelectricity feature of selection absolute mean, variance and 4 rank AR coefficient calculations window samples, described absolute mean, variance and 4 rank AR coefficient formulas are as follows respectively:

Absolute mean: $\mathrm{MAV}=\frac{1}{N}{\mathrm{\Σ}}_{i=1}^N\left|X_i\right|$

Variance: $\mathrm{VAR}=\frac{1}{N-1}{\mathrm{\Σ}}_{i=1}^N{x}_i^2$

4 rank AR coefficients: $x_i={\mathrm{\Σ}}_{k=1}^4{a}_k{x}_{i-k}+w_i$

Wherein N=250, be data point number in window, a _k(k=1,2,3,4) are the AR coefficient, w _ifor the white noise residual error;

Step 5, carry out dimension-reduction treatment to the myoelectricity feature, utilizes the LDA algorithm to carry out dimension-reduction treatment to the myoelectricity characteristic series vector of trying to achieve;

Step 6, ask for the mean value of the characteristic series vector after each gesture dimensionality reduction, in the vector space after dimensionality reduction, characterizes the categorization vector of this gesture;

Step 7, calculate self-test data a window sample the myoelectricity feature and carry out dimension-reduction treatment, the Euclidean distance of each gesture categorization vector in itself and step 6 relatively in the input sorter, judge which kind of gesture is this myoelectricity feature belong to.

2. the surface electromyogram signal recognition methods based on the LDA algorithm according to claim 1, it is characterized in that: described dimension-reduction treatment is based on finding a suitable projection matrix W, make raw data set M (P * q) obtain the expression under new coordinate system space after the W conversion, can effectively reduce the dimension of raw data set, and distinguish preferably the concentrated inhomogeneity data of raw data, concrete steps are:

Step 501, hash matrix S between compute classes _w

$S_W={\mathrm{\Σ}}_{i=1}^C{\mathrm{\Σ}}_{j=1}^{N_i}(m_j-u_i){(m_j-u_i)}^T;$

Step 502, hash matrix S in compute classes _b

$S_B={\mathrm{\Σ}}_{i=1}^C{N}_i(u_i-u){(u_i-u)}^T;$

Step 503, calculate projection matrix W

$J\left(W\right)=\frac{\det \left(W^T{S}_{m}W\right)}{\det \left(W^T{S}_{W}W\right)},$

In above-mentioned three formulas, C means the classification of data, N _ithe sample number that means each classification, m _jmeaning each data in each class, is a column vector, u _imean mean value of all categories, u means the population mean of all categories data.In the process that solves projection matrix W, at first obtain S _w ^-1s _beigenwert, form the W matrix and get final product by then getting front K eigenwert characteristic of correspondence vector after descending sort;

Step 504, carry out projective transformation by raw data set through matrix W

V=w ^TM

The new matrix of Y in formula (R * S) for obtaining after projection, the dimension of its each column data is down to R (R≤p, R≤C-l) by p.

3. the surface electromyogram signal recognition methods based on the LDA algorithm according to claim 2 is characterized in that: described myoelectricity Feature Dimension Reduction dimension is 7 dimensions.

Images