Summary of the invention
For realizing operator's wrist translational speed and the correct identification of motor pattern, the present invention in Teleoperation Systems
Propose a kind of L-Z complexity and fractal dimension and the electromyographic signal recognition methods of maximum fractal length.First from related muscles group
The corresponding surface electromyogram signal of upper collection, then extracts the L-Z complexity of electromyographic signal and fractal dimension as characteristic vector, finally
It is characterized vector input K arest neighbors model incremental learning algorithm grader, it is achieved electromyographic signal with L-Z complexity and fractal dimension
Upper limbs multi-pattern recognition, extracts the maximum fractal length of electromyographic signal as controlled quentity controlled variable, the speed of identification upper extremity exercise.
In order to realize object above, the inventive method mainly comprises the steps that
Step (1). obtain human upper limb electromyographic signal sample data, specifically: first pass through electromyographic signal collection instrument and pick up
Take human upper limb electromyographic signal, then use signal noise silencing method based on Wavelet Energy Spectrum entropy to the electromyographic signal containing interference noise
Carry out de-noising.
Step (2). the electromyographic signal that step (1) obtains is carried out feature extraction, and the L-Z obtaining this electromyographic signal is complicated
Degree, fractal dimension and maximum fractal length.
Described L-Z complexity, specific algorithm is as follows:
Lempel-Ziv complexity be by Lempel and Ziv propose a kind of for describing the non-linear of the random degree of sequence
Index, the algorithm calculating generally employing Kaspar Yu Schusyer proposition of its value c (n).
Assuming that the symbolism sequence of original signal is s1s2...sn.Start to add s from empty string1, by replicating and adding
Operation realizes the connection of complete sequence.If having generated prefix s1s2∧sr-1, r < n, and next symbol srIt is to have operated with interpolation
Become, be designated as:
s1s2∧sr-1→s1s2∧sr-1sr(1)
Here at srAfter mark " " reflect srGeneration process-interpolation.Below for the detailed process realized:
Make S=s1s2...sr, Q=sr+1, SQ represents total character string that S, Q are spliced into, SQ π represent in SQ last
The character string of gained left out in character, observes whether Q can obtain from certain symbol clone method of SQ π.If Q can not be from SQ π
In certain substring replicate and obtain, just with adding operation plus sr+1, and marking " ".If Q can from SQ π certain symbol
Duplication obtains, then continuing to observe increases character sr+2sr+3... string, check whether reproducible obtains for it, until can not, do and add
Operation, and marking " ", repetition said process to sequence last character sn.The number m reflection of mark " " in sequence
Take to add the number of times of operation." complexity " c (n) can be tried to achieve by following formula:
C (n)=(m.log2n)/n (2)
Wherein, m is the number of times adding operation, and n is the length of sequence.This Complexity Measurement c of Kaspar-Schuster
N () is simple to operate, it is easy to accomplish.
Described fractal dimension and maximum fractal length, specific algorithm is as follows:
Fractal dimension calculates and uses Tokyo Univ Japan Higuchi T to teach the method proposed.Assuming that equal interval sampling and
The time series obtained is: X=(x1,x2,...,xN), constructor sequence set
In formula, int rounds under representing, and k, m are integer.Take an integer k can obtain k group sequence to appointing.
DefinitionA length of
In formula, N-1/ (int ((N-m)/k) .k) is the normalization factor of Time Sub-series length, and corresponding to the length of k
< L (k) > is defined as Lm(k) average.Such as L (k)=Ck-D, then D is seasonal effect in time series fractal dimension, takes from so about peer-to-peer
Logarithm, has
Ln L (k)=ln C-D ln k (4)
Analyze understand, ln k and ln L (k) be slope be the linear relation of-D, if ln k and ln L (k) can be tried to achieve, then use
These points of least square fitting, i.e. can get Fractal dimensions.
Length < L (k) > during smallest dimension k is defined as by the Naik of Univ Melbourne Australia, Ganesh R etc.
Maximum fractal length (maximum fractal length MFL).
Step (3). the L-Z complexity and the fractal dimension that obtain using step (2) increase as characteristic vector input K arest neighbors model
Amount learning algorithm grader, it is thus achieved that recognition result.
The present invention develops on the basis of proposing a kind of K arest neighbors (kNN) method, uses K that clustering technique improves
Neighborhood Model Incremental Learning Algorithm.
The data preparation of this algorithm, uses C-means clustering algorithm, using central point as the representative point of KNN algorithm, algorithm
Step is described as follows.
(1) C-means clustering algorithm is used respectively the sample point in each classification to be carried out automatic cluster, it is assumed that each class
Sample in not is polymerized to m bunch the most respectively.
(2) all samples in same cluster are simply calculated, set up the four-tuple of model.
In the application of multimode recognition, correct pattern recognition result should can be used as follow-up pattern-recognition foundation, therefore,
Algorithm should possess adjust and improve bunch in model learn and receive increase newly sample data, reach the effect of incremental learning.Right
In the sample point being verified through pattern-recognition, when number does not reaches a certain amount of accumulation, use C-mean cluster similar
Criterion be integrated in some bunch of class, and to bunch four-tuple data be adjusted.When the sample points after checking reaches
During a certain amount of accumulation, again cluster with C-means clustering algorithm the most again, set up new model cluster.
(3) distance of sample to be sorted and each representative point (central point) is calculated by K arest neighbors (kNN) method.In ballot certainly
When determining the classification of sample to be identified, represent dot product one weight coefficient to eachAfter weighting
Data ballot determines the classification of sample to be identified.
Step (4). the maximum fractal length of the electromyographic signal obtained using step (2) is as controlled quentity controlled variable, it is achieved gripper of manipulator
Take the control of speed.
In order to the motion of main with the operator as far as possible hand of the motion making manipulator keeps Tong Bu, strengthen coming personally of remote operating
Sense, the sense of reality, the present invention captures speeds control to manipulator and is designed.Control is quantified according to surface electromyogram signal characteristic parameter
The responsiveness of robot processed, signal is the strongest, and characteristic ginseng value is the biggest, and action is the fastest, otherwise the slowest.The maximum of electromyographic signal
Fractal length is to characterize the suitable characteristic parameter of muscle activity intensity, and is easy to calculate, it is achieved convenient, the size of its value and crawl
Speed is monotone increasing relation.Thus, it is to control preferable reference input.Limited by mechanical hardware condition, dominant operator
Hand speed range of operation is taken as 0.1~0.5 (1/s).If with m=av2+ bv+c represents that service speed is fractal with flesh signal maximum
Relation between length, wherein: v is speed, m is maximum fractal length, then extensor is as follows with the functional relation on musculus flexor:
Extensor:
Musculus flexor:
The size of manipulator control input quantity is determined by the electromyographic signal maximum fractal length weighted sum in extensors and flexors:
Mx=aM1+bM2 (8)
Wherein, MxFor weighing the total maximum fractal length of proportional control factor, M in certain time period1For stretching in this time period
The maximum fractal length of flesh, M2For the maximum fractal length of musculus flexor in this time period, a, b are respectively weighted value.
The value of [a, b] is relevant with pattern.As under stretching wristing operation mode, directly related with this pattern is
Extensor, the impact of musculus flexor is the faintest, therefore [a, b] desirable [1,0];Under wrist flexion pattern, phase direct with this pattern
Close is musculus flexor, and the impact of extensor is the faintest, therefore [a, b] desirable [0,1];What the present invention was discussed is grasping movement pattern,
Directly related with this pattern is extensors and flexors, therefore [a, b] desirable [1/2,1/2].
The present invention, compared with existing many hand muscle signal of telecommunication action identification methods, has a characteristic that
Utilize the fractal dimension feature in the L-Z complexity index of flesh signal and fractal theory to realize Wrist-sport pattern
Identification, the pattern of remote floor-washing robot manipulator is controlled by result for operator.Movement recognition grader uses
The KNN model incremental learning algorithm that C-means clustering technology improves, not only inherits that KNN algorithm performance is stable, discrimination is high
Advantage, and possessed the ability of incremental learning, along with the increase of pattern-recognition sample, discrimination can improve further.Hand
Responsiveness depend on the activity intensity of arm muscles group, muscle activity intensity can be by the maximum fractal length table of electromyographic signal
Levy, therefore with the maximum fractal length of electromyographic signal for input controlled quentity controlled variable, it is achieved that the control of manipulator grasp speed, achieve
Comparatively ideal effect.
Detailed description of the invention
Below in conjunction with the accompanying drawings embodiments of the invention are elaborated: the present embodiment is being front with technical solution of the present invention
Put and implement, give detailed embodiment and concrete operating process.
As it is shown in figure 1, the present embodiment comprises the steps:
Step one, obtains human upper limb electromyographic signal sample data, specifically: first pass through the pickup of electromyographic signal collection instrument
Human upper limb electromyographic signal, then use signal noise silencing method based on Wavelet Energy Spectrum entropy that the electromyographic signal containing interference noise is entered
Row de-noising.
(1) electromyographic signal of human upper limb is gathered.Experimenter carries out clenching fist respectively, exrending boxing, wrist inward turning are moved with wrist outward turning 4
Make each 50 groups, select upper limbs musculus extensor carpi ulnaris and musculus flexor carpi ulnaris to originate as surface electromyogram signal.First divide with alcohol during experiment
On the musculus extensor carpi ulnaris and musculus flexor carpi ulnaris of experimenter, do not rub decontamination, to strengthen picking up signal ability, use
MyoTrace400 electromyographic signal collection instrument picks up the surface electromyogram signal that musculus extensor carpi ulnaris is corresponding with musculus flexor carpi ulnaris.
(2) use signal noise silencing method based on Wavelet Energy Spectrum entropy that the electromyographic signal containing interference noise is carried out de-noising.
Step 2, electromyographic signal step one obtained carries out feature extraction, obtain this electromyographic signal L-Z complexity,
Fractal dimension and maximum fractal length.
Seasonal effect in time series coarse, is premise and the key calculating kmpel ziv complexity.The present invention uses one many
The coarse method of yardstick, uses four intervals, encodes former electromyographic signal with two binary systems, the first for rough segmentation position, " 1 "
Represent signal more than mean value, " 0 " otherwise, second is refinement position, and " 1 " represents that signal is in the top in interval, at " 0 " expression
In interval bottom.So, the reconstruct of sequence is the most no longer binaryzation, but many-valued coarse algorithm, represent with " 0~3 ".Adopt
By this its advantage of coarse method be original signal and reproducing sequence is man-to-man mapping relations, the reconstruct sequence of signal shown in Fig. 2
Row become ((23131020).
Electromyographic signal product complexity theory process is as follows:
1) local maximum of electromyographic signal, minimum are asked.The envelope up and down of signal is obtained by interpolating function.And to upper
Lower envelope line is averaging, and is designated as m (i).To substitute electromyographic signal original value x in corresponding sequence of points with value of symbol s (i) on envelope
I (), seeks the difference of signal s (i) and m (i), be designated as h (i)=s (i)-m (i).Above, i is the ordinal number of electromyographic signal, and its value is 1
It is the length of signal to N, N.
2) h (i) signal normalization: find out amount h of maximum absolute value in h (i)max=max | h (i) | i=1,2 ... N.?
After, by amplitude normalization:
EMG (i)=h (i)/hmaxI=1,2 ..., N
3) time series signal is carried out the multiple dimensioned many-valued coarse of four-range.
4) complexity of signal is calculated with Kaspar-Schusyer method.
In the algorithm seeking electromyographic signal complexity, the 1st step replaces original signal value to be method success with the value on envelope
Key.Reason is, identical pattern, and the fluctuation tendency of electromyographic signal is identical, and the motion of details and hand
Speed, dynamics size etc. are relevant, as directly asked for the complexity of signal with original signal, and the value variation of same pattern complexity
Scope can be bigger.Preferable effect is not reached for pattern-recognition.
Table 1 for respectively taking 50 groups of electromyographic signals electromyographic signal on musculus extensor carpi ulnaris with musculus flexor carpi ulnaris to four class patterns
The statistics of complexity.
Table 1 surface electromyogram signal complexity index statistics
In the fractal dimension for pattern-recognition calculates, the sampling number of electromyographic signal takes 2000, and the value of k takes respectively
20,21,…,28.Table 2 is to 50 groups of electromyographic signals of four class patterns electromyographic signal on musculus extensor carpi ulnaris with musculus flexor carpi ulnaris
The statistics of its fractal dimension.
Table 2 surface electromyogram signal dimension statistics
During little yardstick k, the numerical value of gained < L (k) > namely the maximum fractal length of electromyographic signal can reflect action well
The activity intensity of related muscles during state.Measure responsiveness under manipulator's grasp mode maximum with related muscles group electromyographic signal
Fractal length value relation.When Fig. 3 (a) and (b) are respectively experiment, the grasp speed that obtains and electromyographic signal on extensor, musculus flexor are
The curve that big fractal length respective value is described, abscissa uses normalizated velocity, namely operator's hand is from the beginning of straight opening
To holding the inverse of time used by spherical object, unit is 1/s.Extensor and musculus flexor when ordinate is for being converted into 2000 sampled points
Maximum fractal length (MFL) average of upper electromyographic signal, immeasurable just.It can be seen that the maximum of muscle groups electromyographic signal
Fractal length, along with the increase of grasp speed and monotone increasing, after speed reaches 0.6 (1/s), is held essentially constant.
Step 3. the L-Z complexity and the fractal dimension that obtain using step 2 input K arest neighbors model incremental as characteristic vector
Learning algorithm grader, it is thus achieved that recognition result.
When building grader at the beginning of tele-robotic system, first calculate on two groups of muscle the complexity of electromyographic signal and point
Dimension, constitutes four dimensional vector samples.The m value of C-means clustering algorithm takes 15, carries out the arrangement of sample points evidence, then to follow-up
Action carries out pattern-recognition.
Table 3 is the service condition after grader foundation.Part I is just to build after grader the mould to follow-up 50 groups of actions
Formula recognition result.Part II is to be added up by the discrimination continuing after 50 groups of incremental learnings of every class action to identify again.Result table
Bright, and accumulated by a certain amount of incremental learning, discrimination has further raising, and effect is the most ideal.
Table 3 grader recognition result
Step 4. the maximum fractal length of electromyographic signal obtained using step 2 is as controlled quentity controlled variable, it is achieved manipulator captures
The control of speed.
The electromyographic signal maximum fractal length of acquisition is converted into by a kind of linear corresponding relation the shape of PWM duty cycle
Formula, to reach to control the responsiveness of direct current generator.Fig. 4 is the remote control system dominant operator's hand designed by the present invention and machinery
Hand captures time consistency contrast test data and curves, and abscissa represents the grasp speed of dominant operator, and ordinate represents machinery
Hand puts in place relative error.Error is in the patient scope of dominant operator.