Summary of the invention
In view of the problems of the existing technology, the present invention provides a kind of wearable smart machine controlling parties that auxiliary is carriedMethod.
The invention is realized in this way a kind of control method for the wearable smart machine that auxiliary is carried, comprising:
Step 1: multi-channel surface myoelectric signal when index finger, middle finger, nameless activity is obtained, using Hidden MarkovModel obtains the wavelet coefficient of finger electromyography signal to surface electromyogram signal wavelet decomposition, which is utilized greatest hopeAlgorithm training, using gauss hybrid models, it is assumed that all wavelet coefficients are same distributions and have identical in same scaleState-transition matrix, the parameters of Hidden Markov Model are obtained by EM algorithm；
Step 2: obtain index finger, middle finger, nameless corresponding Hidden Markov Model parameter after, with removalThe wavelet coefficient of noise reconstructs the multi-channel surface myoelectric signal after being filtered；
Step 3: being based on generalized regression nerve networks, instrument synchronous acquisition ankle is captured with electromyograph(EMG and three-dimensional motion respectivelyThe outer flesh LG of flesh MG, sura, musculus peroneus longus PER and musculus soleus in tibialis anterior TA, sura around when flexion and extension is done in jointThe electromyography signal and ankle joint angle of five pieces of muscle of SOL, and feature extraction is carried out to ankle-joint electromyography signal；
Step 4: obtaining myoelectricity integral strength and myoelectrical activity space by the characteristic parameter of multi-channel surface myoelectric signalThe value indicative matrix decomposition is individual factor matrix Z and action mode matrix X, action mode by the multidimensional characteristic value matrix of distributionInput of the matrix X as pattern recognition classifier device indicates eigenvalue matrix y using symmetrical bilinear modelk=zTWkx；
In formula: zTWhat is indicated is individual factor part, and what x was indicated is action mode part, and Wk belongs to bilinear modelCoefficient matrix；
Defined feature value matrix
In formula:U ∈ (1~U), m ∈ (1~M)] what is indicated is that u-th of subject's execution m is dynamicMake multidimensional characteristic value matrix when n-th；
Step 5: obtaining eigenvalue matrix y of the new user under some movement, the action mode square of bilinear model is utilizedBattle array mean value and coefficient matrix, calculate new individual subscriber factor matrix:
Z=[[WX] [WX]CV]+ycv
By the surface electromyogram signal eigenvalue matrix y under different action modes, action mode matrix part x is obtained are as follows:
X '=[[WCVz]CV]+y′
Step 6: the musculus extensor digitorum entirety myoelectricity when index finger extracted with flexible electrode array, middle finger, nameless activity is believedNumber establish surface myoelectric amplitude, myoelectrical activity spatial distribution characteristic matrix bilinear model, carry out the knowledge of finger force levelNot；
Step 7: carrying out dimensionality reduction to ankle-joint myoelectricity data based on the numerical algorithm of principal component analysis, ankle-joint flesh is obtainedElectric principal component signal predicts ankle-joint angle track using GRNN algorithm based on ankle-joint myoelectricity principal component signal, uses golden sectionSearching algorithm determines the postfitted orbit parameter σ in GRNN, is filtered using wavelet Based on Denoising Algorithm to ankle-joint angle prediction locusTo improve precision of prediction；
Step 8: obtaining arm support portion faces electromyography signal, estimating for arm support position muscular exertion size is calculatedEvaluation Fe, the movement of arm support booster parts, the arm support are triggered according to the grade of arm support position muscular exertion sizePassive control of the action signal triggering finger booster parts of booster parts to finger；
Step 9: the feature point value of the ankle-joint myoelectricity principal component signal and ankle-joint angle prediction locus that will acquire is distinguishedIt is compared with preset ankle-joint myoelectricity principal component signal threshold value and ankle joint angle threshold value, when both greater than preset thresholdWhen, the signal by ankle joint angle prediction numerical value as triggering waist booster parts movement；
Step 10: leg booster parts are using single-degree-of-freedom exoskeleton system by wearer and leg power-assisted driving unit oneIt rises and ectoskeleton kinetic moment is provided, the output torque of motor is obtained according to ectoskeleton self information, and the output torque of motor is Ta=(1-α-1)G′(q)。
Further, a GRNN model prediction ankle joint angle is established based on ankle-joint electromyography signal, there are three input layersNeuron is three principal component signals of ankle-joint myoelectricity, and mode layer neuron number is ankle-joint myoelectricity number of training k,Summing, there are two neuron SD and SN1 for layer, only one neuron of output layer, is ankle joint angle, first closes myoelectricity and anklePreceding the 2/3 of section angle sample inputs the network as training data and is trained, and recycles rear the 1/3 of myoelectricity sample to be used as test numberIt is predicted that ankle-joint angle, in which:
In formula: YijFor training sample output vector YiJ-th of element.
Further, the absolute average of electromyography signalThe variance of electromyography signalAverage frequency
In formula, xijFor the numerical values recited of j-th of sampled point in i-th of timeslice,For signal in i-th of timesliceAverage value, fjFor Frequency point discrete on i-th of timeslice power spectrum, P (fj) it is discrete point in frequency fjCorresponding power, eachThere is n sampled point in timeslice.
Further, lumbar surface electromyography signal is obtained, the absolute average, variance, average frequency three of electromyography signal are chosenA characteristic parameter constitutes the input vector of direction of error Propagation Neural Network, show that neural network parameter matrix calculates hand after trainingThe estimated value of arm support zone muscular exertion size；
Input variable and output variable are blurred, waist muscle size of exerting oneself sets fuzzy language according to numerical values recitedFuzzy language for several grades, the corresponding practical grip size of waist is also set to several grades, for output variable, motorRevolving speed fuzzy language is set as several grades.
The present invention is based on the myoelectricity of GRNN prediction ankle-joint angle Trajectory Arithmetics can quickly and accurately predict that ankle-joint acts,Ankle joint angle, the ankle-joint myoelectricity principal component signal that will acquire and the prediction of ankle-joint angle are predicted by the electromyography signal of ankle-jointThe feature point value of track is compared with preset ankle-joint myoelectricity principal component signal threshold value and ankle joint angle threshold value respectively,When being both greater than preset threshold, the signal acted by ankle joint angle prediction numerical value as triggering waist booster parts is helpedThe action logic of power component is more scientific and reasonable, can realize the power-assisted effect of booster parts to a greater extent.It will be based on wavelet fieldHidden Markov Model method is applied in the de-noising filtering processing of electromyography signal, utilizes the aggregation of wavelet coefficient between adjacent scaleProperty and the characteristics of duration construct Hidden Markov Model, the wavelet coefficient of actual signal is obtained using Bayesian Estimation, is led toIt crosses signal reconstruction and effectively removes noise, by carrying out timeslice segmentation to signal using Short Time Fourier Transform thought, andSeveral representative electromyography signal parameters are had chosen to the signal analysis in timeslice, and with the neural network pair after trainedIt is analyzed and processed, and then estimates corresponding muscular exertion size.The resulting power of myoelectricity force-touch sensor is touched simultaneouslyFeel that signal and muscular exertion high low signal input fuzzy controller, the velocity of rotation of driving motor is to control grip size, benefitIt is acted with the grade triggering arm support booster parts of arm support position muscular exertion size, the arm support booster partsAction signal triggers passive control of the finger booster parts to finger, and the action logic of booster parts is more scientific and reasonable, can be moreThe power-assisted effect of booster parts is realized in big degree.
Specific embodiment
In order to further understand the content, features and effects of the present invention, the following examples are hereby given, and cooperate attached drawingDetailed description are as follows.
Structure of the invention is explained in detail with reference to the accompanying drawing.
A kind of control method for the wearable smart machine that auxiliary is carried, comprising:
Multi-channel surface myoelectric signal when S101, acquisition index finger, middle finger, nameless activity, using Hidden Markov mouldType obtains the wavelet coefficient of finger electromyography signal to surface electromyogram signal wavelet decomposition, which is calculated using greatest hopeMethod training, using gauss hybrid models, it is assumed that all wavelet coefficients are same distributions and have identical in same scaleState-transition matrix obtains the parameters of Hidden Markov Model by EM algorithm；
S102, obtain index finger, middle finger, nameless corresponding Hidden Markov Model parameter after, made an uproar with removalThe wavelet coefficient of sound reconstructs the multi-channel surface myoelectric signal after being filtered；
As shown in Fig. 2, entire electromyography signal filtering includes wavelet decomposition, training Hidden Markov Model (the maximum phaseHope value-based algorithm), four part of Bayesian Estimation and wavelet reconstruction, this method does not need any free parameter undetermined, has fineAdaptivity, the noise in electromyography signal can be effective filtered out and remain the detailed information in signal.
S103, generalized regression nerve networks are based on, capture instrument synchronous acquisition ankle with electromyograph(EMG and three-dimensional motion respectively and closesThe outer flesh LG of flesh MG, sura, musculus peroneus longus PER and musculus soleus SOL in tibialis anterior TA, sura around when section does flexion and extensionThe electromyography signal and ankle joint angle of five pieces of muscle, and feature extraction is carried out to ankle-joint electromyography signal；
S104, myoelectricity integral strength and myoelectrical activity space point are obtained by the characteristic parameter of multi-channel surface myoelectric signalThe value indicative matrix decomposition is individual factor matrix Z and action mode matrix X, action mode square by the multidimensional characteristic value matrix of clothInput of the battle array X as pattern recognition classifier device, indicates eigenvalue matrix y using symmetrical bilinear modelk=zTWkx；
In formula: zTWhat is indicated is individual factor part, and what x was indicated is action mode part, and Wk belongs to bilinear modelCoefficient matrix；
Defined feature value matrix
In formula:U ∈ (1~U), m ∈ (1~M)] what is indicated is that u-th of subject's execution m is dynamicMake multidimensional characteristic value matrix when n-th；
S105, eigenvalue matrix y of the new user under some movement is obtained, utilizes the action mode matrix of bilinear modelMean value and coefficient matrix calculate new individual subscriber factor matrix:
Z=[[WX] [WX]CV]+ycV
By the surface electromyogram signal eigenvalue matrix y under different action modes, action mode matrix part x is obtained are as follows:
X '=[[WCVz]CV]+y′
The steady section sEMG signal subsection by filtering processing is calculated into root mean square (time window H=256 in MATLABSampled point, each time window do not overlap), it is segmented myoelectrical activity intensity of the average value as current record channel of root mean square, with handRefer to that the percentage of maximal voluntary contractile force amount indicates；
S106, the musculus extensor digitorum entirety electromyography signal with the index finger of flexible electrode array extraction, middle finger, the third finger when movableEstablish surface myoelectric amplitude, myoelectrical activity spatial distribution characteristic matrix bilinear model, carry out the identification of finger force level；
S107, dimensionality reduction is carried out to ankle-joint myoelectricity data based on the numerical algorithm of principal component analysis, obtains ankle-joint myoelectricityPrincipal component signal is predicted ankle-joint angle track using GRNN algorithm based on ankle-joint myoelectricity principal component signal, is searched with golden sectionRope algorithm determines the postfitted orbit parameter σ in GRNN, use wavelet Based on Denoising Algorithm to ankle-joint angle prediction locus be filtered withImprove precision of prediction；
S108, arm support portion faces electromyography signal is obtained, calculates the estimation of arm support position muscular exertion sizeValue Fe, the movement of arm support booster parts is triggered according to the grade of arm support position muscular exertion size, which helpsPassive control of the action signal triggering finger booster parts of power component to finger；
Arm support portion faces electromyography signal is obtained, absolute average A, the variance S, average frequency of electromyography signal are chosenThree characteristic parameters constitute the input vector of direction of error Propagation Neural NetworkNeural network is obtained after trainingParameter matrix W1And W2, use W1And W2Calculate the estimated value Fe of arm support position muscular exertion size；
The absolute average of electromyography signalThe variance of electromyography signalAverage frequency
In formula, xijFor the numerical values recited of j-th of sampled point in i-th of timeslice,For signal in i-th of timesliceAverage value, fjFor Frequency point discrete on i-th of timeslice power spectrum, P (fj) it is discrete point in frequency fjCorresponding power, eachThere is n sampled point in timeslice.
After obtaining arm support portion faces electromyography signal, using Hidden Markov Model to the small wavelength-division of surface electromyogram signalSolve the wavelet coefficient of finger electromyography signal；By the wavelet coefficient using EM algorithm training, using Gaussian Mixture mouldType, it is assumed that all wavelet coefficients are same distributions and have identical state-transition matrix in same scale；By the maximum phaseAlgorithm is hoped to obtain the parameters of Hidden Markov Model, after obtaining the parameter of corresponding Hidden Markov Model, with removalThe wavelet coefficient of noise reconstructs the arm support portion faces electromyography signal after being filtered.
Estimate that the firmly degree of measured's arm, first design are real by the neural network of training error backpropagationIt tests, myoelectricity acquisition electrode is fitted on the musculus flexor carpi ulnaris position of measured, while finger pressing when measured's wrist flexion being allowed to surveyPower device can measure the firmly size F of measured's wrist muscle while acquiring electromyography signal in this way.
Input variable and output variable are blurred, arm support position muscular exertion size FeAccording to numerical values recited by mouldPaste language is set as several grades, the corresponding practical grip size F in arm support positionhFuzzy language be also set to it is several etc.Grade, for output variable, motor speed S fuzzy language is set as several grades；
The output variable of fuzzy controller is that the speed S of arm closure namely arm support booster parts act motorThe locked-rotor torque of revolving speed, motor speed and motor is directly proportional, so realizing arm grip indirectly by the closing speed of armControl.
Input variable and output variable are blurred, wherein being directed to input variable, muscular exertion size FeFuzzy language is setIt is set to attonity, small, smaller, larger, 5 grades big；
The practical grip size F of armhFuzzy language be defined as it is 4 grades small, smaller, larger, big；
For output variable, motor speed S fuzzy language is set as quickly opening, middling speed is opened, open at a slow speed, attonity, at a slow speedIt closes, middling speed is closed, quickly 7 grades of pass, positive value indicate that motor rotates forward, i.e. the steering of arm closure, negative value expression motor reversal, i.e. armThe steering of opening.
The feature point value of S109, the ankle-joint myoelectricity principal component signal that will acquire and ankle-joint angle prediction locus respectively withPreset ankle-joint myoelectricity principal component signal threshold value and ankle joint angle threshold value are compared, when both greater than preset thresholdWhen, the signal by ankle joint angle prediction numerical value as triggering waist booster parts movement；
In the present embodiment, different ankle-joint myoelectricity principal component signal, ankle joint angle predicted value and waist is arranged to helpThe mapping of power component force size, realizes triggering of the ankle motion situation to waist booster parts；
S110, leg booster parts using single-degree-of-freedom exoskeleton system by wearer and leg power-assisted driving unit togetherEctoskeleton kinetic moment is provided, the output torque of motor is obtained according to ectoskeleton self information, and the output torque of motor is Ta=(1-α-1)G′(q)。
α is the angle between thigh and vertical direction in formula.In other no drive forces, only wearer provides powerWhen square, T=T at this timehw, q=G (T), sensitivity coefficient is
Compared with existing control mode, the present invention is by following advantages:
The practical grip that arm myoelectricity is measured is small and practical grip that finger myoelectricity is measured is big, illustrates that this thing only has finger dynamicWork, arm and attonity or movement very little, further explanation does not need arm at this time too big grip, so arm powered portionPart does not need to act, once and the practical grip measured of practical grip and finger myoelectricity that arm myoelectricity is measured is mostly big, explanationArm powered component needs to act, and waist stress is big when user takes a step during bearing a heavy burden and advancing needs power-assisted, is closed by ankleThe electromyography signal of section predicts ankle joint angle, the spy of the ankle-joint myoelectricity principal component signal and ankle-joint angle prediction locus that will acquireSign point value is compared with preset ankle-joint myoelectricity principal component signal threshold value and ankle joint angle threshold value respectively, when bothWhen greater than preset threshold, the signal acted by ankle joint angle prediction numerical value as triggering waist booster parts, booster partsAction logic is more scientific and reasonable, can realize the power-assisted effect of booster parts to a greater extent.
The present invention will be applied in the de-noising filtering processing of electromyography signal based on wavelet domain concealed Markov model method, benefitWith Hidden Markov Model is constructed the characteristics of the aggregation and duration of wavelet coefficient between adjacent scale, using Bayesian EstimationThe wavelet coefficient of actual signal is obtained, noise is effectively removed by signal reconstruction, by being thought using Short Time Fourier TransformWant to carry out signal timeslice segmentation, and several representative electromyography signals are had chosen to the signal analysis in timeslice and are joinedNumber, and processing is analyzed it with the neural network after training, and then estimate corresponding muscular exertion size.Based on GRNNMyoelectricity prediction ankle-joint angle Trajectory Arithmetic can quickly and accurately predict ankle-joint act, it is more applicable compared with BP neural networkOn-line prediction in ankle-joint angle track.Simultaneously by the resulting power haptic signal of myoelectricity force-touch sensor and muscular exertion sizeSignal inputs fuzzy controller, and the velocity of rotation of driving motor is used to control grip size using arm support position muscleThe grade triggering arm support booster parts movement of power size, the action signal of the arm support booster parts trigger finger power-assistedPassive control of the component to finger, arm support booster parts action signal trigger the movement of waist booster parts, realize power-assisted portionThe action logic of part is more scientific and reasonable, can realize the power-assisted effect of booster parts to a greater extent.
The above is only the preferred embodiments of the present invention, and is not intended to limit the present invention in any form,Any simple modification made to the above embodiment according to the technical essence of the invention, equivalent variations and modification, belong toIn the range of technical solution of the present invention.