CN102073881A - Denoising, feature extraction and pattern recognition method for human body surface electromyography signals - Google Patents
Denoising, feature extraction and pattern recognition method for human body surface electromyography signals Download PDFInfo
- Publication number
- CN102073881A CN102073881A CN 201110009548 CN201110009548A CN102073881A CN 102073881 A CN102073881 A CN 102073881A CN 201110009548 CN201110009548 CN 201110009548 CN 201110009548 A CN201110009548 A CN 201110009548A CN 102073881 A CN102073881 A CN 102073881A
- Authority
- CN
- China
- Prior art keywords
- wavelet
- electromyographic signal
- node
- noising
- coefficient
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses a denoising, feature extraction and pattern recognition method for human body surface electromyography signals, which comprises the following steps: performing best wavelet packet node denoising on the human body surface electromyography signals; performing wavelet decomposition on the human body surface electromyography signals after denoising, and extracting maximum absolute values of all levels of wavelet high-frequency coefficients as feature vectors; and inputting the extracted feature vectors into a back propagation network for performing training and pattern recognition. With the adoption of the best wavelet packet leaf node denoising method provided by the invention, noise carried in the electromyography signals can be effectively removed, and useful information can be retained; the maximum absolute values of the wavelet high-frequency coefficients can well reflect features of the human body surface electromyography signals, thereby having obvious advantages in comparison with the existing feature value method; and the relatively high recognition effect can be obtained by searching optimal hidden neurons and training errors of the back propagation network.
Description
Technical field
The invention belongs to bio signal and handle and area of pattern recognition, relate to a kind of human body surface myoelectric signal noise silencing and feature extracting method.
Background technology
Surface electromyogram signal (surface electromyography (SEMG) signals) is a kind of bioelectrical signals, because it has easy collection, undamaged advantage, be widely used in clinical medicine and sports medical science, particularly in human action identification field, electromyographic signal is its main research object.And the de-noising of electromyographic signal and feature extraction are the key issues of human action identification.
The SEMG signal is the muscle electrical activity that obtains from the human body skin surface, and it comes from the electrical activity [1] of the nerve fibre that detects muscle.The SEMG signal is a kind of faint electric signal, and peak-to-peak value generally has only 0～10mV, and the frequency range of useful signal is 0～500HZ[2].Electromyographic signal is accompanied by intrinsic noise and neighbourhood noise etc. [3] in the various noises, electronic equipment of inside of human body tissue in the process of obtaining.Because the amplitude of surface electromyogram signal is little, signal to noise ratio (S/N ratio) is low, easily is submerged in the various noises, and noise therefore how effectively to remove surface electromyogram signal becomes the emphasis of research.
Surface electromyogram signal is the non-stationary bioelectrical signals, and wavelet analysis is to analyze the strong mathematical tool of non-stationary signal as a kind of multiresolution time and frequency zone analytical approach.Because the wavelet packet noise-eliminating method not only decomposes the low frequency part of signal, the while is also analyzed HFS, compares with wavelet noise to have higher time frequency resolution, has therefore obtained using widely.But these noise-eliminating methods mostly adopt global threshold, and promptly same scale adopts identical threshold value, perhaps HFS are adopted unified threshold value de-noising [4], perhaps optimal wavelet inclusion point are adopted identical thresholding algorithm [5].And mixed various noise of different nature in the electromyographic signal, and show that the character of each node coefficient is also different, therefore, adopt identical threshold value to be difficult to the de-noising effect that reaches desirable to different nodes.So, seek a kind of adaptive mode and come that each node of optimal wavelet bag is carried out noise reduction and have very necessary.
The key of pattern-recognition is choosing of eigenwert, and it directly influences the effect of identification.At present, in the human action identification that with SEMG is research object, explored a series of effective eigenwert choosing methods, and obtained good application and checking.Temporal signatures for example: autoregressive model (AR) [6], integration myoelectricity value [7]; Frequency domain character: power spectrum, double-spectrum analysis [8]; Because time domain or frequency domain can only extract the single characteristic of field of electromyographic signal, can lose a part of feature, therefore introduce the time and frequency zone feature, for example, short time discrete Fourier transform, Wigner-Ville conversion and wavelet transformation etc. wherein become effective most with the research of wavelet coefficient as eigenwert.Cai Liyu etc. [9] with the wavelet coefficient of each yardstick maximum be eigenwert to human upper limb four actions discern, accuracy has reached more than 80%; [10] such as Zhao Jing-dong small echo details component is carried out svd and with it as eigenwert, the average recognition rate based on six actions of doing evil through another person has been reached 90%; People such as Luo Zhizeng come the structural attitude value with the energy of each frequency range of wavelet package transforms, to stretch wrist, bend wrist, the discrimination of exhibition fist and 4 kinds of hand motion patterns such as clench fist reaches 92.5%[11].In recent years, a lot of scholars propose to utilize the Nonlinear Dynamical Characteristics of electromyographic signal to carry out action recognition, the largest Lyapunov exponent that Liu Nangeng etc. [12] utilize the coefficients at different levels after the wavelet decomposition to upper limbs four actions discern, the discrimination of each action is all more than 93%.But, in the medical science of recovery therapy field, require higher degree of accuracy, therefore how reaching higher discrimination is the final goal that each researcher pursues.
Backpropagation (Back Propagation, BP) neural network is a kind of sorter that is most widely used in the electromyographic signal pattern-recognition, still, implicit neuronic number selects also not form at present a complete theoretical direction.Nineteen ninety, R.C.Eberhart and R.WDobbins set forth " implicit neuronic selection is a kind of art " like this in their works " Neural Network PC tools ".As seen, the recognition effect of neural network is closely related with implicit neuronic number.We find also that under study for action recognition effect also has contact closely with training error, the parameter difference of selection, and last recognition effect has very big difference.Therefore, seeking optimum neural network parameter also is the key of electromyographic signal action recognition.
The above-mentioned list of references of mentioning is as follows:
[1]Constable?R?and?R.J.Thornbill，“Using?the?Discrete?WaveletTransform?for?Time-Frequency?Analysis?of?the?Surface?EMG?Signal，”ISA，Vol.16，pp.121-127，1993.
[2]Kilby?J?and?H.Gholam?Hosseini，“Wavelet?Analysis?ofElectromyography?Signal，”IEEE.vol.1，pp.384-387，2004
[3]Kale?S.N?and?S.V.Dudul，“Intelligent?Noise?Removal?fromEMG?Signal?Using?Focused?Time-Lagged?Recurrent?Neural?Network，”Hindawi?Publishing?Corporation?Applied?Computational?Intelligence?andSoft?Computing.Volume?2009，Article?ID?129761，12?pages
[4]Xin?Guo，Peng?Yang，Ying?Li，“The?sEMG?analysis?for?thelower?limb?prosthesis?using?wavelet?transformation，”[C]Proceedings?ofthe?26th?Annual?International?Conference?of?the?IEEE?EMBS，SanFrancisco，pp.341-344，2004.
[5]Sun?Chengkui，Ye?Min，Mei?pingao，“SEMG?denoising?basedon?best?wavelet?package?analysis”.MECHANICAL?&?ELECTRICALENGINERRING?MAGAZINE[J]，Vol.25.No.8.Aug.2008
[6]Chen?XB(Chen?Xinben)，Yang?GY(Yang?Guangying)，“TheProcessing?of?Electromyography?Signal?Based?on?Wavelet?NeuralNetwork”.2009?INTERNATIONAL?SYMPOSIUM?ON?WEBINFORMATION?SYSTEMS?AND?APPLICATIONS，PROCEEDINGS.pp.529-532，2009
[7]Katsutoshi?Kuribayashi，“A?discrimination?system?usingneural?network?for?emg?controlled?prostheses?integral?type?of?emg?signalprocessing，”Proceedings?of?the?1993?IEEE/RSJ?International?Conferenceon?Intelligent?Robots?and?Systems?Yokohama，Japan，pp.26-30，July.1993.
[8] Yang Jun, Lei Min, " based on the surface electromyogram signal pattern-recognition of double-spectrum analysis, " automated manufacturing [J], 2009 03 phases.
[9]Cai?Liyu，Wang?Zhizhong，Zhang?Haihong，“Surface?EMGSignal?Classification?Method?Based?on?Wavelet?Transform，”Journal?ofData?Acquisition?&?Processing[J]，Vol.15?No.2?Jun.2000
[10]Jingdong?Zhao，Zongwu?Xie，Li?Jiang，Hegao?Cai.“EMGControl?for?a?Five-fingered?Prosthetic?Hand?Based?on?Wavelet?Transformand?Autoregressive?Model，”Proceedings?of?the?2006?IEEE?InternationalConference?on?Mechatronics?and?Automation.China，pp.25-28，June.2006.
[11] Mei Pingao, Luo Zhizeng. " electromyographic signal based on wavelet packet analysis and Elman network is handled ". electromechanical engineering [J], the 25th the 1st phase of volume of in January, 2008.
[12]Liu?Nangeng，Lei?Min，“Characterization?of?surfaceelectromyography?signal?based?on?wavelet?analysis?and?non-linearexponent，”Journal?of?Clinical?Rehabilitative?Tissue?EngineeringResearch[J]，April.22.2008，Vol.12，No.17.
Summary of the invention
The purpose of this invention is to provide a kind of human body surface myoelectric signal noise silencing, feature extraction and mode identification method, can overcome the problem and shortage that exists in noise reduction, Feature Selection and the pattern-recognition of present electromyographic signal, can carry out noise reduction to each node of optimal wavelet bag adaptively, discrimination height, the recognition effect of action are good.
To achieve these goals, the invention provides a kind of human body surface myoelectric signal noise silencing, feature extraction and mode identification method, comprise the steps:
(1) treat that recognized action is corresponding and gather many group electromyographic signals after, in described many group electromyographic signals, select training data electromyographic signal and test data electromyographic signals;
(2), adopt the small echo threshold values corresponding to carry out de-noising respectively to each destination node of its optimal wavelet Bao Shu, according to the tree of the optimal wavelet bag after de-noising reconstruct training data electromyographic signal with this terminal note coefficient to each training data electromyographic signal;
(3) choose wavelet basis function and wavelet decomposition progression, the training data electromyographic signal of each reconstruct is carried out wavelet decomposition;
(4) extract small echo high frequency coefficients at different levels in the training data electromyographic signal after each wavelet decomposition, with the maximum value of small echo high frequency coefficients at different levels eigenwert, the feature value vector of the eigenwert composing training data electromyographic signal of all training data electromyographic signals as the training data electromyographic signal;
(5) to each test data electromyographic signal, itself and each training data electromyographic signal same treatment mode are carried out de-noising, reconstruct, wavelet decomposition and extraction eigenwert, obtain the eigenwert of test data electromyographic signal, the eigenwert of all test data electromyographic signals constitutes the feature value vector of test data electromyographic signal;
(6) feature value vector of training data electromyographic signal and the feature value vector input reverse transmittance nerve network of test data electromyographic signal are carried out pattern-recognition.
In one embodiment of the invention, described step (2) is specially:
(21) each training data electromyographic signal is carried out WAVELET PACKET DECOMPOSITION, obtain the wavelet packet tree;
(22) obtain the optimal wavelet Bao Shu that the wavelet packet of each training data electromyographic signal is set;
(S23) obtain the coefficient of each destination node of the optimal wavelet Bao Shu of each training data electromyographic signal, the coefficient of the non-destination node of optimal wavelet Bao Shu is made as zero;
(24) to all destination nodes of the optimal wavelet Bao Shu of each training data electromyographic signal, obtain small echo threshold values corresponding and de-noising mode with the coefficient of this terminal note, according to this small echo threshold values and de-noising mode this terminal note is carried out de-noising, obtain the destination node coefficient after the new de-noising;
(25) to the optimal wavelet Bao Shu of the electromyographic signal after each de-noising, begin to handle from first wavelet packet node of afterbody, whether the father node coefficient of judging this node correspondence is zero, if, this father node of node coefficient reconstruct by after another child node de-noising of the node coefficient after this node de-noising and this father node obtains the father node coefficient of reconstruct; If not, keep this father node coefficient, again even higher level of node is handled equally after every grade of node processing finishes, finish, obtain the training data electromyographic signal of reconstruct until first node of the first order.
In another embodiment of the present invention, described step (6) is specially:
(61) feature value vector of training data electromyographic signal is changed after, input BP neural network;
(62) implicit neuron being set increases one by one from 1 to H, and the beginning training network is wherein for each implicit neuronic each network training, according to from 10
-1-10
-GThe training error that reduces 10 times one by one carries out network training successively, obtains the BP neural network that trains successively, the output layer output expectation value of the BP neural network that trains, and wherein H, G all get any positive integer;
(63) feature value vector of test data electromyographic signal is changed after, import the BP neural network that each trains;
(64) according to the correct identification number of the expectation value statistical test data electromyographic signal of the output valve of BP neural network and BP neural network, the correct identification number of all test data electromyographic signals and the number of samples of test data electromyographic signal liken discrimination to into each network training, write down the training error value and the implicit neuron number of this discrimination correspondence.
(65) with the discrimination of the highest discrimination, with the implicit neuron number of the highest discrimination correspondence and the combination of training error, as the parameter of BP neural network as described action to be identified.
In an embodiment more of the present invention, the WAVELET PACKET DECOMPOSITION progression in the described step (21) is elected 3 grades as.Because through repeatedly testing and testing, to the de-noising result and the recognition effect of human body surface myoelectric signal, choose 3 grades of WAVELET PACKET DECOMPOSITION progression, the effect that reaches is best.Need to prove, be not decomposed class multiple-effect fruit is good more more, and WAVELET PACKET DECOMPOSITION progression is chosen too conference and is increased operand.
Compared with prior art, inventor surface electromyographic signal de-noising, feature extraction and mode identification method have following advantage:
1. in step (2) and step (5), each node to the optimal wavelet Bao Shu of each training data electromyographic signal and test data electromyographic signal adopts the small echo threshold values corresponding with this node to carry out de-noising respectively, this noise-eliminating method is from operand, because the principle of optimal wavelet bag is to delete unnecessary wavelet packet node, therefore the computational data amount more required than traditional method of wavelet packet lacked (when decomposed class is selected 3 grades, wavelet decomposition progression than existing most of electromyographic signal lacks, and has further reduced the computational data amount); Select from threshold value, each node to optimal wavelet Bao Shu adopts the small echo threshold values corresponding with this node coefficient to carry out de-noising respectively, therefore the node of each optimal wavelet bag adopts different threshold size and threshold mode respectively, this de-noising mode has adaptive characteristic, and de-noising more flexibly, accurately;
2. in step (4) and step (5) with the maximum value of small echo high frequency coefficients at different levels as the eigenwert of training data electromyographic signal and the eigenwert of test data electromyographic signal, compare with existing various eigenvalue methods (svd of the variance of the greatest coefficient of AR parameter, small echos at different levels, small echo high frequency coefficients at different levels, small echo high frequency coefficients at different levels, Nonlinear Dynamical Characteristics or the like), improved the discrimination of human action greatly, near 100%;
3. introduce the search of optimum neuron number and optimum training error in step (62), by seeking the highest discrimination, determine the implicit neuron number and the training error of BP neural network, finally reach the highest discrimination, recognition effect is good.
By following description also in conjunction with the accompanying drawings, it is more clear that the present invention will become, and these accompanying drawings are used to explain embodiments of the invention.
Description of drawings
Fig. 1 is the process flow diagram of the electromyographic signal de-noising of inventor surface, feature extraction and mode identification method.
Fig. 2 is the detail flowchart of de-noising in human body surface myoelectric signal noise silencing shown in Figure 1, feature extraction and the mode identification method.
Fig. 3 is the process flow diagram of pattern-recognition in human body surface myoelectric signal noise silencing shown in Figure 1, feature extraction and the mode identification method.
Fig. 4 (a) and Fig. 4 (b) are respectively wavelet packet tree graph and optimal wavelet bag tree graph.
Fig. 5 (a), Fig. 5 (b) and Fig. 5 (c) are respectively each destination node coefficient of optimal wavelet bag, each the destination node coefficient of optimal wavelet bag after the de-noising, the optimal wavelet bag reconstruct node coefficient of not de-noising.
Fig. 6 (a), Fig. 6 (b), Fig. 6 (c) and Fig. 6 (d) are respectively and are original electromyographic signal waveform, wavelet noise design sketch, wavelet packet de-noising effect figure, optimal wavelet bag destination node de-noising effect figure.
Embodiment
With reference now to accompanying drawing, describe embodiments of the invention, the similar elements label is represented similar elements in the accompanying drawing.
Move-stretch tight pin and hook with two of gathering below the electromyographic signal de-noising of inventor surface, feature extraction and mode identification method are described from human body lower limbs shank position.It should be noted that the identification of other actions and similar recognition methods are also in interest field of the present invention.
Wherein, adopt a passage to carry out to the electromyographic signal collection of stretch tight pin and two actions of hook, 40 groups of electromyographic signals are gathered in each action, and every group of electromyographic signal collection frequency is 1024HZ, and it is 201 points that the collection of each action is counted
With reference to figure 1, the following describes the electromyographic signal de-noising of inventor surface, feature extraction and mode identification method.This method comprises the steps:
Step S1, the 40 groups of electromyographic signals and the hook action of do gathering at the foot-propelled that stretches tight are gathered in 40 groups of electromyographic signals, and each selects 20 groups as the training data electromyographic signal, and remaining 20 groups as the test data electromyographic signal.Training data electromyographic signal and test data electromyographic signal are write two four-matrix A1=K * L * M * N and A2=K * L * M * N respectively, wherein K is the collection of each action count (length of electromyographic signal), K=201 in the present embodiment, the port number of L for gathering, L=1 in the present embodiment, M is the group number that the action of training data or test data correspondence is gathered, M=20 in the present embodiment (training data or test data are 20 groups), N trains or the recognized action number N=2 in the present embodiment for waiting.So A1=A2=201 * 1 * 20 * 2;
Step S2 chooses a untreated training data electromyographic signal or choose a untreated test data electromyographic signal in test data matrix A 2 in training data matrix A 1;
Step S3, each destination node to the optimal wavelet Bao Shu of the training data electromyographic signal selected or test data electromyographic signal adopts the small echo threshold values corresponding with this terminal note coefficient to carry out de-noising respectively, according to tree reconstruct training data electromyographic signal of the optimal wavelet bag after the de-noising or test data electromyographic signal;
Step S4, choose 3 grades of wavelet basis function ' sym3 ' and decomposed classes, to the training data electromyographic signal of each reconstruct or test data electromyographic signal carry out wavelet decomposition (essence of wavelet decomposition be wavelet basis function on different decomposition progression with the convolution of signal, therefore to select wavelet packet basis functions and decomposed class could carry out wavelet decomposition to signal, different wavelet basis functions and decomposed class have very big influence to the effect of wavelet noise and identification, ' Symlets ' adopted in most of research to electromyographic signal, ' Daubechies ', ' Coiflets ' wavelet basis function, effect is best, this experiment is through repeatedly test, when wavelet basis function was 3 grades for ' sym3 ' and decomposed class, wavelet packet de-noising and recognition effect were preferably);
Step S5 extracts training data electromyographic signal after each wavelet decomposition or the small echo high frequency coefficient cd at different levels in the test data electromyographic signal
i(i=1,2 ... lev) (each training data electromyographic signal or test data electromyographic signal all have three small echo high frequency coefficient cd in the present embodiment
1, cd
2, cd
3), the small echo high frequency coefficients at different levels of each training data electromyographic signal or test data electromyographic signal are carried out dimensionality reduction (formula λ by the maximal value that takes absolute value
i=max (abs (cd
i)) (i=1,2, ... lev)), (eigenwert of selected electromyographic signal (foot-propelled that stretches tight in the test sample book is made the 13rd group of data) is C1=[λ as the eigenwert of training data electromyographic signal or test data electromyographic signal with the maximum value of small echo high frequency coefficients at different levels
3=22.8807, λ
2=99.1590, λ
1=146.5688] '), (lev is capable, the L row for the feature value vector D1=lev * L * M * N of the eigenwert composing training data electromyographic signal of all training data electromyographic signals in the training data matrix A 1, the M group, the N class), D1=3 in the present embodiment * 1 * 20 * 2 (3 row, 1 row, 20 groups, 2 classes), feature value vector D2=3 * 1 * 20 * 2 (3 row, 1 row of the eigenwert of all test data electromyographic signals formation test data electromyographic signal in the test data matrix A 2,20 groups, 2 classes);
Whether step S6 has untreated electromyographic signal in training of judgement data matrix A1 or the test data matrix A 2, if, change step S2, if not, continue next step,
Step S7 carries out pattern-recognition with the feature value vector D1 of training data electromyographic signal and the feature value vector D2 input BP neural network of test data electromyographic signal.
With reference to figure 2, described step S3 is specially:
Step S31, to each training data electromyographic signal or test data electromyographic signal, utilize the function " wpdec " among the matlab to carry out WAVELET PACKET DECOMPOSITION, obtain the wavelet packet tree, to each wavelet packet tree, utilize " wpcoef " function among the matlab to obtain the wavelet packet coefficients at different levels of wavelet packet tree (as utilizing wpcoef (t, 2
1) obtain the first order coefficient that wavelet packet is set t, utilize wpcoef (t, 2
2) obtain the second level coefficient of wavelet packet tree t);
Step S32, wavelet packet tree to each training data electromyographic signal or test data electromyographic signal, calculate the length of wavelet packet coefficients at different levels and the node sum of wavelet packet: to each training data electromyographic signal, selecting wavelet packet basis functions waveneme is ' sym3 ', the length of ' sym3 ' wavelet filter is 6, length K=201 of training data electromyographic signal in the integrating step (1), training data electromyographic signal and wavelet filter are carried out convolution, length is 206 after the convolution, descend sampling then, obtain first order wavelet packet low frequency and high fdrequency component, so the length p of first order wavelet packet coefficient
1Be 103, again first order wavelet packet coefficient and wavelet filter carried out convolution, length is 108 after the convolution, and sampling down obtains second level wavelet packet low frequency and high fdrequency component then, so the length p of second level wavelet packet node coefficient
2Be 54, with sampling under second level wavelet packet coefficient and the wavelet filter convolution, obtain the length p of third level wavelet packet coefficient again
3Be 29; Selecting WAVELET PACKET DECOMPOSITION progression lev is 3, and the node of wavelet packet adds up to q=2
Lev+1-2=14; The calculating of the node sum of the length of wavelet packet coefficients at different levels and wavelet packet is identical with the calculating of top training data electromyographic signal in the wavelet packet tree of test data electromyographic signal, here training data matrix A 1=K * L * M * N and test data matrix A 2=K * L * M * N comprise L * M * N electromyographic signal, its length of each electromyographic signal is K, so calculating the length of the wavelet packet coefficients at different levels of L * M * N electromyographic signal is correspondent equal, the node sum of wavelet packet also equates when selecting same WAVELET PACKET DECOMPOSITION progression;
Step S33, wavelet packet tree to each training data electromyographic signal or test data electromyographic signal, utilize the function " besttree " among the matlab to calculate its optimal wavelet Bao Shu (calculating employing prior art of optimal wavelet Bao Shu, the electromyographic signal difference, it is also different to calculate resulting optimal wavelet Bao Shu.Accompanying drawing 4 (a) (b) is respectively the wavelet packet tree graph and the optimal wavelet bag tree graph of a training data electromyographic signal (foot-propelled that stretches tight in the test sample book is made the 13rd group of data), the optimal wavelet bag tree graph of noting each electromyographic signal is different, only provides one in the example and describes);
Step S34, (destination node of the optimal wavelet Bao Shu in the accompanying drawing 4 (b) is { (4) to the destination node of the optimal wavelet Bao Shu of each training data electromyographic signal or test data electromyographic signal, (5), (7), (8), (13), (14) }), utilize function " wpcoef " among the matlab directly to obtain each destination node coefficient of optimal wavelet Bao Shu, all the other non-destination node coefficients are made as zero; Set up reconstruction coefficients matrix B 1 '=P * Q, wherein the reconstruction coefficients matrix B ' in each element B '
I, jThe sample value of each the destination node coefficient of the optimal wavelet Bao Shu when representing not de-noising, i refers to sample point, j refers to the site position, and the span of i is the length 103 of the first order wavelet packet coefficient that obtains from 1 to step S32, and the span of j is the 1 node sum 14 to wavelet packet; With each destination node coefficient of optimal wavelet Bao Shu deposit in reconstruction coefficients matrix B 1 '=P * Q the 14th, 13,8,7,5,4 row, with non-destination node coefficient deposit in reconstruction coefficients matrix B 1 '=respective column of P * Q, wherein the length of every row be the wavelet packet coefficients at different levels of the correspondence that calculates among the step S32 length (reconstruction coefficients B1 '=every row of P * Q preserve the corresponding destination node coefficient of the optimal wavelet Bao Shu of not denoising, reconstruction coefficients matrix B 1 '=P * Q in the coefficient length of the corresponding destination node of length of each row coefficient, the size of the coefficient of the big or small corresponding destination node of coefficient, non-destination node coefficient all is made as zero, Fig. 5 (a) has showed the data among reconstruction coefficients B1 '=P * Q, horizontal ordinate is represented site position (columns of corresponding B1 '=P * Q respectively), ordinate is represented the length (the coefficient length of corresponding B1 '=every row of P * Q respectively) of coefficient, and ordinate is represented the size (size of the coefficient value among corresponding B1 '=P * Q in every row) of coefficient);
Step S35 chooses a untreated destination node at the optimal wavelet Bao Shuzhong of each training data electromyographic signal or test data electromyographic signal, promptly the reconstruction coefficients matrix B from step S34 ' in choose a destination node;
Step S36 obtains small echo threshold values corresponding with the coefficient of this terminal note and de-noising mode, according to this small echo threshold values and de-noising mode, utilize the function " wthresh " among the matlab, this terminal note is carried out de-noising (the optimal wavelet bag tree node { (2,1) among Fig. 4 (b), (2,2), (3,0), (3,1), (3,6), (3,7) acquiescence wavelet threshold } is respectively { (38.2401), (28.5267), (35.9210), (98.8202), (64.0582), (21.3267) }, default threshold de-noising mode is respectively { s, s, s, s, s, s}, to all destination nodes of optimal wavelet Bao Shu, get wavelet threshold corresponding and de-noising mode and carry out de-noising with the coefficient of this destination node), obtain the destination node coefficient after the new de-noising; Set up reconstruction coefficients matrix B "=P * Q, each element B ' ' in the reconstruction coefficients matrix B ' ' after the de-noising after the de-noising
I, jThe sample value of each the destination node coefficient of the optimal wavelet Bao Shu after the expression de-noising, i refers to sample point, j refers to the site position, and the span of i is the length 103 of the first order wavelet packet coefficient that obtains from 1 step S32, and the span of j is the node sum 14 from 1 wavelet packet; " correspondence position of=P * Q, non-destination node coefficient are made as zero (coefficient after each destination node de-noising is shown in Fig. 5 (b)) to deposit the coefficient of the destination node after the de-noising in after the de-noising reconstruction coefficients matrix B;
Step S37 judges whether optimal wavelet Bao Shuzhong has node to be untreated, if, change step S35, if not, continue next step;
Step S38, optimal wavelet Bao Shu to the electromyographic signal after each de-noising, begin to handle from first wavelet packet node of afterbody, whether the father node coefficient of judging this node correspondence is zero, if, by the node coefficient after another child node de-noising of the node coefficient after this node de-noising (the reconstruction coefficients matrix B after de-noising " obtain=P * Q) and this father node (the reconstruction coefficients matrix B after de-noising "=obtain P * Q) this father node of reconstruct, obtain the father node coefficient of reconstruct; If not, keep this father node coefficient, again even higher level of node is handled equally after every grade of node processing finishes, finish, obtain the training data electromyographic signal of reconstruct until first node of the first order;
With reference to figure 3, wherein, described step S7 is specially:
Step S71 is converted to 2 dimension training characteristics matrix E1 (lev * L is capable, M * N row) with the feature value vector D1 of training data electromyographic signal, as the input layer of BP neural network;
Step S72 will imply neuron num and be initialized as 1, be i.e. num=1;
Step S73, goal is initialized as goal with training error, i.e. num=10
-1
Step S74, carry out network training: enter 2 dimension training characteristics matrix E1 of input layer each implicit neuron through hidden layer, arrive output layer, if the average error of two-dimentional training characteristics matrix E1 is less than training error goal, then output layer is exported expectation value (end of this time network training), continue step S75, otherwise calculate output error (output error is the poor of actual output and expectation value), and output error is fed back to each implicit neuron (neural network weight is revised) of hidden layer;
Step S75 is converted to 2 dimension test feature matrix E2 (lev * L is capable, M * N row), the BP neural network that input trains with the feature value vector D2 of test data electromyographic signal;
Step S76, obtain discrimination: (2 numbers of tieing up the column vector of test feature matrix E2 are exactly the number of action to be identified as the input layer of the BP neural network that trains with 2 dimension test feature matrix E2, each characteristic series vector all is identified as an action), output to the BP neural network of test data, whether the expectation value of judging the BP neural network that output valve and step S74 obtain is identical, if identical then identification is correct, if difference is identification error then, can count the correct identification number of test data electromyographic signal thus, (number of samples of test data electromyographic signal is by the experiment decision for the correct identification number of all test data electromyographic signals and the number of samples of test data electromyographic signal, generally get the number identical, get empirical value and get final product for tens groups with training sample) the ratio discrimination of BP neural network for this reason; Deposit discrimination in matrix F ((N+1 is capable, G row, H group)), wherein N represents action kind to be identified, by the test objective decision, deposits the discrimination of average recognition rate and each action during N+1 is capable in, totally 2 classes; G row corresponding different training errors; The implicit neuronic number of H group expression;
Step S77 reduces 10 times with training error goal, i.e. goal=goal * 0.1;
Step S78, whether training of judgement error goal is less than 10
-G(G=9) (wherein G gets any positive integer, and through test of many times, G gets empirical value 9 because in the experiment along with the increase of G, subtract behind G＞9, discrimination kept stable after identification takes the lead in increasing), if not, change step S74, if continue next step;
Step S79 will imply neuron num and add 1, be i.e. num=num+1;
Step S791 judges that whether (wherein H gets any positive integer to implicit neuron num, as required, desirablely wants the scope of observing greater than H (H=20), this experiment is in order to find best implicit neuron number, and H gets 20), if not, change step S73, if continue next step;
Step S792, from matrix F, the highest discrimination of search with the final discrimination of the highest discrimination as described action to be identified, with the implicit neuron number of the highest discrimination correspondence and the combination of training error, as the parameter of BP neural network, finishes.
In step S74, expectation value needs a few class actions of identification according to wanting the recognized action type to determine, the residing element that just which class of desired output moved is made as 1, and the element of other positions is made as 0; Such as action 1, so its desired output vector is [1,0], and 2 its desired output vectors that move are [0,1].
Need to prove, Fig. 6 (a) (b) (c) (d) provides waveform after waveform after this electromyographic signal original waveform, the wavelet noise, the waveform after the wavelet packet de-noising, the de-noising of optimal wavelet bag destination node respectively, here utilize prior art that electromyographic signal is carried out wavelet noise and wavelet packet de-noising, come the de-noising effect of comparative illustration step S36 optimal wavelet of the present invention bag destination node de-noising.Can see that from Fig. 6 (b) wavelet noise has only kept the global characteristics of electromyographic signal, detail section is filtered substantially; Fig. 6 (c) though the wavelet packet de-noising has kept the global characteristics and the detail section of electromyographic signal, has comprised too much details composition, can't represent the actual characteristic of electromyographic signal accurately; Fig. 6 (d), the de-noising of optimal wavelet bag destination node had both kept global characteristics, had kept part details component again, with Fig. 6 (a) (b) (c) compare and played reasonable de-noising effect.
Below step S38 is specifically described.
As three grades of WAVELET PACKET DECOMPOSITION of Fig. 4 (a), its wavelet packet node adds up to 14, is 4 as the destination node of Fig. 4 (b) optimal wavelet Bao Shu, 5,7,8,13,14, the restructuring matrix B after the de-noising then is " among=P * Q the corresponding the 4th; 5,7,8; 13,14 row are respectively the coefficients after these several destination node de-noisings, and the coefficient of the row of all the other non-destination node correspondences is 0; " the N row are preserved the coefficient of N node to the restructuring matrix B after the de-noising among=P * Q, and length is the length of the coefficients at different levels of calculating in the step (21).The process of step S38 reconstruct electromyographic signal is as follows:
The third level is from B "=P * Q the 7th row begin search, and the upper level father node 3 of the 7th node, coefficient are zero, then by the 7th node and the 8th node coefficient reconstruct upper level father node 3, reconstruct node coefficient are kept at B "=P * Q the 3rd row cover original zero;
See B again "=P * Q the 9th row; its upper level father node 4; coefficient is non-vanishing; illustrate that 4 nodes are destination nodes of wavelet packet tree; keep 4 node coefficients " at B=P * Q the 4th row is used for the search of second level wavelet packet node, here can not be with 9 and 10 node reconstruct, 4 nodes, because this moment, these two nodes were zero, if the reconstruct meeting covers the coefficient of destination node 4;
" coefficient is non-vanishing for=P * Q the 11st row, its upper level father node 5, keeps 5 node coefficients at B1 '=P * Q the 5th row to see B again;
"=P * Q the 13rd row, its upper level father node 6, coefficient are zero, then by the 6th node coefficient of the 13rd and the 14th node coefficient reconstruct, deposit B in, and "=P * Q the 6th is listed as to see B again.So far, third level wavelet packet node search finishes, by the search of third level wavelet packet node, reconstruct partial wavelet packet node coefficient;
The second level is from B " the 3rd row search of=P * Q, its upper level father node 1, coefficient are zero, then by the 1st node coefficient of the 3rd and the 4th node coefficient reconstruct, deposit B in, and "=P * Q the 1st is listed as;
" " the 2nd row of=P * Q have been used the coefficient of the 4th node that keeps and the coefficient of the 5th node here to the 5th row of=P * Q, its upper level father node 2 coefficients are zero, by the 2nd node coefficient of the 5th and the 6th node reconstruct, deposit B in to see B again;
First order search is with the 1st row coefficient and the 2nd row coefficient of B " row of the 1st among=P * Q and the 2nd be listed as (be the reconstruction coefficients matrix B after the de-noising ") the original electromyographic signal of reconstruct is at this moment.The reconstruction coefficients matrix B " in store all reconstruct node coefficients, shown in Fig. 5 (c).
The validity of the eigenvalue method that extracts in order to illustrate in the technology of the present invention, this example extracts the autoregressive model (AutoRegressive after de-noising of optimal wavelet bag and the wavelet noise respectively, AR) maximum value of the wavelet coefficient of parameter, largest Lyapunov exponent, absolute value maximums at different levels, small echo high frequency coefficients at different levels is as eigenwert, the human body lower limbs hook and two actions of pin of stretching tight are discerned, wherein, 20 groups of training samples, 20 groups of test sample books.The recognition effect contrast (after the de-noising of optimal wavelet inclusion point) that its acceptance of the bid 1 of its contrast and experiment such as table 1 and table 2. is various eigenwerts.Fig. 2 is the recognition effect contrast (after wavelet noise) of various eigenwerts
Table 1
Table 2
Contrast table 1 and table 2, pass through the de-noising of optimal wavelet bag as can be seen after, the recognition effect of various eigenwerts has all had obvious increase, the electromyographic signal disposal route based on the de-noising of optimal wavelet inclusion point that visible the present invention proposes is effective.As can be seen from Table 1, get the maximum value of small echo high frequency coefficients at different levels, recognition effect is best.
For the validity of searching method that BP neural network used among the present invention is described, this example provided under different neuron hidden layer numbers and the different training errors average recognition effect relatively.Table 3 has been showed the influence to discrimination of hidden neuron number and training error.Wherein the training and testing proper vector is the maximum value of the small echo high frequency coefficient that extracts after the de-noising of optimal wavelet leafy node.
Table 3
As can be seen from Table 3, hidden neuron number and the training error network training result that can affect the nerves finally influences discrimination.Therefore, reach than higher identification, it is vital choosing suitable neuron number and training error.From table, can see, when neuron number is that 7 training errors are when being 0.01, and neuron number is that 12 training errors are 0.01 o'clock, it is the highest by 97.5% all to reach discrimination, but, take all factors into consideration net training time and frequency of training, selecting neuron is 7, and training error is 0.01 parameter as the BP neural network.
Above invention has been described in conjunction with most preferred embodiment, but the present invention is not limited to the embodiment of above announcement, and should contain various modification, equivalent combinations of carrying out according to essence of the present invention.
Claims (4)
1. a human body surface myoelectric signal noise silencing, feature extraction and mode identification method comprise the steps:
(1) treat that recognized action is corresponding and gather many group electromyographic signals after, in described many group electromyographic signals, select training data electromyographic signal and test data electromyographic signals;
(2), adopt the small echo threshold values corresponding to carry out de-noising respectively to each destination node of its optimal wavelet Bao Shu, according to the tree of the optimal wavelet bag after de-noising reconstruct training data electromyographic signal with this terminal note coefficient to each training data electromyographic signal;
(3) choose wavelet basis function and wavelet decomposition progression, the training data electromyographic signal of each reconstruct is carried out wavelet decomposition;
(4) extract small echo high frequency coefficients at different levels in the training data electromyographic signal after each wavelet decomposition, with the maximum value of small echo high frequency coefficients at different levels eigenwert, the feature value vector of the eigenwert composing training data electromyographic signal of all training data electromyographic signals as the training data electromyographic signal;
(5) to each test data electromyographic signal, itself and each training data electromyographic signal same treatment mode are carried out de-noising, reconstruct, wavelet decomposition and extraction eigenwert, obtain the eigenwert of test data electromyographic signal, the eigenwert of all test data electromyographic signals constitutes the feature value vector of test data electromyographic signal;
(6) feature value vector of training data electromyographic signal and the feature value vector input reverse transmittance nerve network of test data electromyographic signal are carried out pattern-recognition.
2. human body surface myoelectric signal noise silencing as claimed in claim 1, feature extraction and mode identification method is characterized in that, described step (2) is specially:
(21) each training data electromyographic signal is carried out WAVELET PACKET DECOMPOSITION, obtain the wavelet packet tree;
(22) obtain the optimal wavelet Bao Shu that the wavelet packet of each training data electromyographic signal is set;
(S23) obtain the coefficient of each destination node of the optimal wavelet Bao Shu of each training data electromyographic signal, the coefficient of the non-destination node of optimal wavelet Bao Shu is made as zero;
(24) to all destination nodes of the optimal wavelet Bao Shu of each training data electromyographic signal, obtain small echo threshold values corresponding and de-noising mode with the coefficient of this terminal note, according to this small echo threshold values and de-noising mode this terminal note is carried out de-noising, obtain the destination node coefficient after the new de-noising;
(25) to the optimal wavelet Bao Shu of the electromyographic signal after each de-noising, begin to handle from first wavelet packet node of afterbody, whether the father node coefficient of judging this node correspondence is zero, if, this father node of node coefficient reconstruct by after another child node de-noising of the node coefficient after this node de-noising and this father node obtains the father node coefficient of reconstruct; If not, keep this father node coefficient, again even higher level of node is handled equally after every grade of node processing finishes, finish, obtain the training data electromyographic signal of reconstruct until first node of the first order.
3. human body surface myoelectric signal noise silencing as claimed in claim 1, feature extraction and mode identification method is characterized in that, described step (6) is specially:
(61) feature value vector of training data electromyographic signal is changed after, input BP neural network;
(62) implicit neuron being set increases one by one from 1 to H, and the beginning training network is wherein for each implicit neuronic each network training, according to from 10
-1-10
-GThe training error that reduces 10 times one by one carries out network training successively, obtains the BP neural network that trains successively, the output layer output expectation value of the BP neural network that trains, and wherein H, G all get any positive integer;
(63) feature value vector of test data electromyographic signal is changed after, import the BP neural network that each trains;
(64) according to the correct identification number of the expectation value statistical test data electromyographic signal of the output valve of BP neural network and BP neural network, the correct identification number of all test data electromyographic signals and the number of samples of test data electromyographic signal liken discrimination to into each network training, write down the training error value and the implicit neuron number of this discrimination correspondence.
(65) with the discrimination of the highest discrimination, with the implicit neuron number of the highest discrimination correspondence and the combination of training error, as the parameter of BP neural network as described action to be identified.
4. human body surface myoelectric signal noise silencing as claimed in claim 2, feature extraction and mode identification method is characterized in that, the WAVELET PACKET DECOMPOSITION progression in the described step (21) is elected 3 grades as.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110009548 CN102073881A (en) | 2011-01-17 | 2011-01-17 | Denoising, feature extraction and pattern recognition method for human body surface electromyography signals |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110009548 CN102073881A (en) | 2011-01-17 | 2011-01-17 | Denoising, feature extraction and pattern recognition method for human body surface electromyography signals |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102073881A true CN102073881A (en) | 2011-05-25 |
Family
ID=44032415
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201110009548 Pending CN102073881A (en) | 2011-01-17 | 2011-01-17 | Denoising, feature extraction and pattern recognition method for human body surface electromyography signals |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102073881A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102426651A (en) * | 2011-08-25 | 2012-04-25 | 武汉理工大学 | Human body forearm surface electromyogram signal acquisition and pattern recognition system |
| CN102622527A (en) * | 2012-04-13 | 2012-08-01 | 西南大学 | Taboo searching method for selection of galvanic skin response signal features |
| CN102622605A (en) * | 2012-02-17 | 2012-08-01 | 国电科学技术研究院 | Surface electromyogram signal feature extraction and action pattern recognition method |
| CN102930284A (en) * | 2012-09-13 | 2013-02-13 | 杭州电子科技大学 | Surface electromyogram signal pattern recognition method based on empirical mode decomposition and fractal |
| CN103345641A (en) * | 2013-07-16 | 2013-10-09 | 杭州电子科技大学 | Hand electromyographic signal motion recognition method based on wavelet entropy and support vector machine |
| CN105997064A (en) * | 2016-05-17 | 2016-10-12 | 成都奥特为科技有限公司 | Method for identifying human lower limb surface EMG signals (electromyographic signals) |
| CN106293057A (en) * | 2016-07-20 | 2017-01-04 | 西安中科比奇创新科技有限责任公司 | Gesture identification method based on BP neutral net |
| CN107169432A (en) * | 2017-05-09 | 2017-09-15 | 深圳市科迈爱康科技有限公司 | Biometric discrimination method, terminal and computer-readable recording medium based on myoelectricity |
| CN107368801A (en) * | 2017-07-13 | 2017-11-21 | 国网江苏省电力公司电力科学研究院 | Piler characteristic extracting method based on wavelet packet analysis |
| CN107714037A (en) * | 2017-10-12 | 2018-02-23 | 西安科技大学 | A kind of miner's fatigue identification method based on the mining helmet of brain-computer interface |
| CN107798343A (en) * | 2017-10-16 | 2018-03-13 | 南京邮电大学 | One kind is based on the improved SVM analog-circuit fault diagnosis methods of manifold structure |
| CN108629311A (en) * | 2018-05-02 | 2018-10-09 | 尚谷科技（天津）有限公司 | A kind of action identification method based on biological pulsation |
| CN110032987A (en) * | 2019-04-23 | 2019-07-19 | 福州大学 | A kind of surface electromyogram signal classification method based on CMAC Neural Network model |
| CN110827987A (en) * | 2019-11-06 | 2020-02-21 | 西安交通大学 | Myoelectricity continuous prediction method and system for wrist joint torque in multi-grabbing mode |
| CN112932508A (en) * | 2021-01-29 | 2021-06-11 | 电子科技大学 | Finger activity recognition system based on arm electromyography network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101625869A (en) * | 2009-08-11 | 2010-01-13 | 中国人民解放军第四军医大学 | Non-air conduction speech enhancement method based on wavelet-packet energy |
| CN101833671A (en) * | 2010-03-30 | 2010-09-15 | 西安理工大学 | Support vector machine-based surface electromyogram signal multi-class pattern recognition method |
| CN101859377A (en) * | 2010-06-08 | 2010-10-13 | 杭州电子科技大学 | Electromyographic signal classification method based on multi-kernel support vector machine |
-
2011
- 2011-01-17 CN CN 201110009548 patent/CN102073881A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101625869A (en) * | 2009-08-11 | 2010-01-13 | 中国人民解放军第四军医大学 | Non-air conduction speech enhancement method based on wavelet-packet energy |
| CN101833671A (en) * | 2010-03-30 | 2010-09-15 | 西安理工大学 | Support vector machine-based surface electromyogram signal multi-class pattern recognition method |
| CN101859377A (en) * | 2010-06-08 | 2010-10-13 | 杭州电子科技大学 | Electromyographic signal classification method based on multi-kernel support vector machine |
Non-Patent Citations (3)
| Title |
|---|
| 《2010 Second WRI World Congress on Software Engineering》 20101220 YUAN Tingting et.al Study on Best Wavelet Packet Based Independent Threshold De-noising for MUAP 269 - 272 第1卷, * |
| 《数据采集与处理》 20000630 蔡立羽 等 基于小波变换的肌电信号识别方法研究 1-4 第15卷, 第2期 * |
| 《生物医学工程学杂志》 20080430 宋维翰 等 基于小波包变换的肌电信号去噪研究 323-326 1-4 第25卷, 第2期 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102426651A (en) * | 2011-08-25 | 2012-04-25 | 武汉理工大学 | Human body forearm surface electromyogram signal acquisition and pattern recognition system |
| CN102426651B (en) * | 2011-08-25 | 2014-04-16 | 武汉理工大学 | Human body forearm surface electromyogram signal acquisition and pattern recognition system |
| CN102622605A (en) * | 2012-02-17 | 2012-08-01 | 国电科学技术研究院 | Surface electromyogram signal feature extraction and action pattern recognition method |
| CN102622605B (en) * | 2012-02-17 | 2015-06-03 | 国电科学技术研究院 | Surface electromyogram signal feature extraction and action pattern recognition method |
| CN102622527A (en) * | 2012-04-13 | 2012-08-01 | 西南大学 | Taboo searching method for selection of galvanic skin response signal features |
| CN102930284A (en) * | 2012-09-13 | 2013-02-13 | 杭州电子科技大学 | Surface electromyogram signal pattern recognition method based on empirical mode decomposition and fractal |
| CN102930284B (en) * | 2012-09-13 | 2015-06-17 | 杭州电子科技大学 | Surface electromyogram signal pattern recognition method based on empirical mode decomposition and fractal |
| CN103345641A (en) * | 2013-07-16 | 2013-10-09 | 杭州电子科技大学 | Hand electromyographic signal motion recognition method based on wavelet entropy and support vector machine |
| CN105997064B (en) * | 2016-05-17 | 2018-10-23 | 成都奥特为科技有限公司 | A kind of discrimination method for human body lower limbs surface electromyogram signal |
| CN105997064A (en) * | 2016-05-17 | 2016-10-12 | 成都奥特为科技有限公司 | Method for identifying human lower limb surface EMG signals (electromyographic signals) |
| CN106293057A (en) * | 2016-07-20 | 2017-01-04 | 西安中科比奇创新科技有限责任公司 | Gesture identification method based on BP neutral net |
| CN107169432A (en) * | 2017-05-09 | 2017-09-15 | 深圳市科迈爱康科技有限公司 | Biometric discrimination method, terminal and computer-readable recording medium based on myoelectricity |
| CN107368801A (en) * | 2017-07-13 | 2017-11-21 | 国网江苏省电力公司电力科学研究院 | Piler characteristic extracting method based on wavelet packet analysis |
| CN107714037A (en) * | 2017-10-12 | 2018-02-23 | 西安科技大学 | A kind of miner's fatigue identification method based on the mining helmet of brain-computer interface |
| CN107798343A (en) * | 2017-10-16 | 2018-03-13 | 南京邮电大学 | One kind is based on the improved SVM analog-circuit fault diagnosis methods of manifold structure |
| CN108629311A (en) * | 2018-05-02 | 2018-10-09 | 尚谷科技（天津）有限公司 | A kind of action identification method based on biological pulsation |
| CN110032987A (en) * | 2019-04-23 | 2019-07-19 | 福州大学 | A kind of surface electromyogram signal classification method based on CMAC Neural Network model |
| CN110032987B (en) * | 2019-04-23 | 2022-07-08 | 福州大学 | Surface electromyographic signal classification method based on cerebellar neural network model |
| CN110827987A (en) * | 2019-11-06 | 2020-02-21 | 西安交通大学 | Myoelectricity continuous prediction method and system for wrist joint torque in multi-grabbing mode |
| CN110827987B (en) * | 2019-11-06 | 2021-03-23 | 西安交通大学 | Myoelectricity continuous prediction method and system for wrist joint torque in multi-grabbing mode |
| CN112932508A (en) * | 2021-01-29 | 2021-06-11 | 电子科技大学 | Finger activity recognition system based on arm electromyography network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102073881A (en) | Denoising, feature extraction and pattern recognition method for human body surface electromyography signals | |
| Tuncer et al. | Surface EMG signal classification using ternary pattern and discrete wavelet transform based feature extraction for hand movement recognition | |
| CN102622605B (en) | Surface electromyogram signal feature extraction and action pattern recognition method | |
| CN108714026B (en) | Fine-grained electrocardiosignal classification method based on deep convolutional neural network and online decision fusion | |
| CN102930284B (en) | Surface electromyogram signal pattern recognition method based on empirical mode decomposition and fractal | |
| Sapsanis et al. | EMG based classification of basic hand movements based on time-frequency features | |
| CN105997064B (en) | A kind of discrimination method for human body lower limbs surface electromyogram signal | |
| Phinyomark et al. | Feature reduction and selection for EMG signal classification | |
| CN102961203A (en) | Method for identifying surface electromyography (sEMG) on basis of empirical mode decomposition (EMD) sample entropy | |
| CN101859377A (en) | Electromyographic signal classification method based on multi-kernel support vector machine | |
| Wan et al. | A neural network to identify human subjects with electrocardiogram signals | |
| CN103345641A (en) | Hand electromyographic signal motion recognition method based on wavelet entropy and support vector machine | |
| CN113111831A (en) | Gesture recognition technology based on multi-mode information fusion | |
| Pancholi et al. | Intelligent upper-limb prosthetic control (iULP) with novel feature extraction method for pattern recognition using EMG | |
| CN113116361A (en) | Sleep staging method based on single-lead electroencephalogram | |
| Khezri et al. | An exploratory study to design a novel hand movement identification system | |
| CN114387668A (en) | Classification method based on multi-level neuromuscular coupling characteristic information fusion | |
| Wang et al. | Motor imagination eeg recognition algorithm based on dwt, CSP and extreme learning machine | |
| Bo et al. | Hand gesture recognition using semg signals based on cnn | |
| CN111985270A (en) | sEMG signal optimal channel selection method based on gradient lifting tree | |
| CN114569143A (en) | Myoelectric gesture recognition method based on attention mechanism and multi-feature fusion | |
| Boernama et al. | Multiclass classification of brain-computer interface motor imagery system: a systematic literature review | |
| Wang et al. | A novel approach for the pattern recognition of hand movements based on EMG and VPMCD | |
| CN114343679A (en) | Surface electromyogram signal upper limb action recognition method and system based on transfer learning | |
| Khadivi et al. | SEMG classification for upper-limb prosthesis control using higher order statistics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20110525 |