CN114931390A

CN114931390A - Muscle force estimation method based on fatigue analysis

Info

Publication number: CN114931390A
Application number: CN202210485153.4A
Authority: CN
Inventors: 夏侯士戟; 罗茜; 马敏
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2022-05-06
Filing date: 2022-05-06
Publication date: 2022-08-23
Anticipated expiration: 2042-05-06
Also published as: CN114931390B

Abstract

The invention discloses a muscle strength estimation method based on fatigue analysis, which comprises the steps of acquiring and preprocessing a surface electromyogram signal and a muscle strength signal of each sample object under a preset action, dividing the surface electromyogram signal into activity sections, extracting fatigue related characteristics in each activity section, and constructing a fatigue characteristic signal; constructing a multi-input LVN network, inputting surface myoelectric signals and M fatigue characteristic signals respectively, and outputting the signals as muscle strength signals; then, taking the surface electromyographic signals of all the sample objects and the corresponding fatigue characteristic signals as input, taking the corresponding electromyographic signals as expected output, and training the multi-input LVN network; when muscle strength estimation with the same action is needed, surface myoelectric signals are obtained after the surface myoelectric signals are collected and preprocessed by the same method as the sample object, fatigue characteristic signals are constructed and obtained, and the fatigue characteristic signals are input into a trained multi-input LVN network to obtain a muscle strength estimation result. The method combines the fatigue characteristics and the LVN network, and improves the accuracy and robustness of muscle strength estimation.

Description

Muscle force estimation method based on fatigue analysis

Technical Field

The invention belongs to the technical field of electromyographic data analysis, and particularly relates to a muscle force estimation method based on fatigue analysis.

Background

The brain activates muscles which contract to produce an electrical signal, thereby producing a mechanical force. Any kind of movement of the human body, including micro movements such as chewing and blinking, and large movements such as running, bouncing and lifting, needs to be realized through corresponding muscle contraction. Different exercises require different muscles, some exercises only require one muscle to participate, and some exercises require multiple muscles to participate together. Muscle strength, i.e. muscle contraction strength, is of great importance in many applications such as gait analysis, orthopedics, rehabilitation, ergonomic design, haptic technology, telesurgery and human-computer interaction.

Currently, myoelectricity-myoelectricity models are mostly adopted in the industry to estimate the myoelectricity. Myo-muscular force is a non-linear and dynamically varying relationship, the level of non-linearity being primarily dependent on the way the muscle fibres are combined when applying force, the time of contraction and the force level, and the dynamic relationship being due to muscle shortening effects and electrical time delay (i.e. the time delay of the myoelectric signal to generation). Therefore, whether the built myoelectric-myoelectric model is reliable depends on whether the model can capture the dynamic changes and non-linearity of the system. Besides the difficulty of expressing system dynamics and nonlinearity, the motion mode, muscle state, individual difference and the like all affect the muscle force estimation precision. Muscle fatigue is also an important and common influencing factor, but many experimental studies in the past have avoided the muscle fatigue problem. However, muscle fatigue seriously affects the muscle activation ability, contraction ability and dynamic relationship between myoelectric signals and force, and is a major difficulty which is difficult to ignore.

There are currently few studies on muscle force estimation in the fatigue state. Soo et al propose a force estimation model based on band technique, which is found to improve significantly compared to the conventional RMS-force model, the more fatigue is found. Na et al proposed a method of estimating the force under fatigue state combining the electrical and motor models of the surface muscles, and found that as the muscle fatigue deepens, the peak value of the motor model decreases and the contraction time increases. The Chenxiang professor team of the Chinese science and technology university corrects the polynomial model and the Hill model by using the fatigue trend, and the influence of fatigue on the muscle strength prediction is eliminated to a certain extent. Three sections of fatigue data are trained and predicted by using a Laguerre model in Asefi et al, the peak value and the median frequency of the Laguerre first-order kernel coefficient of the three sections of data are found to have a descending trend, the peak value and the second-order high-frequency component of the first-order kernel can be considered as indexes of fatigue generation, a myoelectricity-muscle force dynamic model affected by muscle fatigue under isometric contraction is realized, and the dynamic relationship of the myoelectricity-muscle force is identified. However, the above method has limited effect on fatigue analysis, and it is difficult to achieve substantial improvement of muscle strength estimation performance in practical application.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a muscle force estimation method based on fatigue analysis, and the accuracy and the robustness of muscle force estimation are improved by introducing fatigue characteristics and combining an LVN network.

In order to achieve the above object, the muscle strength estimation method based on fatigue analysis according to the present invention comprises the steps of:

s1: for K sample objects, acquiring a surface electromyogram signal and a muscle strength signal of each sample object under a preset action, preprocessing the surface electromyogram signals according to a preset method to obtain preprocessed surface electromyogram signals x _k (i) And muscle strength signal y _k (i) Wherein i ═ 1., N denote signal length;

s2: with a length L _win Electromyographic signal x of sliding window with sliding step length delta on surface _k (i) Performing sliding extraction on the signal segments, calculating the average instantaneous energy of each signal segment, and the average instantaneous energy E of the jth signal segment _k (j) The calculation formula of (a) is as follows:

wherein j is 1,2, …, N _part ，N _part Representing the number of the divided signal segments;

setting an energy threshold

When in use

And is

The start of the (j + 1) th signal segment is taken as the start of the active segment when

And is provided with

Taking the terminal point of the jth signal segment as the terminal point of the active segment, and performing no operation in other cases, thereby obtaining the surface electromyogram signal x _k (i) Dividing the active segment;

note surface electromyography signal x _k (i) The number of active segments obtained is D _k And the d-th activity segment is represented as

Respectively representing surface electromyographic signals x _k (i) The original sampling point sequence numbers of the beginning and end of the ith active segment, D ═ 1,2, …, D _k ；

S3: g characteristics related to fatigue in the surface electromyogram signals are set according to actual needs, and then each surface electromyogram signal x is subjected to _k (i) G characteristics f of each activity segment are respectively extracted _k,d,g ，g＝1,2,…,G；

S4: for each surface electromyographic signal x according to active segment division _k (i) Respectively constructing G fatigue characteristic signals F thereof _k,g (i) I.e. as the sampling point

The fatigue characteristic signal F corresponding to the g-th characteristic _k,g (i)＝f _k,d,g Otherwise F _k,g (i)＝0；

S5: construction of multiple outputsEntering an LVN network, inputting surface myoelectric signals and M fatigue characteristic signals respectively, and outputting the signals as muscle strength signals; then the surface electromyographic signal x of each sample object is measured _k (i) The corresponding fatigue characteristic signal is used as the input of the multi-output LVN network, the corresponding electromyographic signal is used as the expected output, and the multi-input LVN network is trained;

s6: when muscle force estimation with the same action is required, a surface electromyographic signal x ' (i) is acquired and preprocessed by the same method as the sample object, G characteristics related to fatigue of each activity section are extracted after activity section division is carried out, and G fatigue characteristic signals F ' are generated ' _g (i) Then, the surface electromechanical signal x '(i) and G fatigue characteristic signals F' _g (i) Inputting the trained multi-input LVN network to obtain an estimated muscle force signal.

The muscle force estimation method based on fatigue analysis comprises the steps of acquiring and preprocessing a surface electromyographic signal and a muscle force signal of each sample object under a preset action, dividing the activity section of the surface electromyographic signal, calculating fatigue-related characteristics in each activity section, and constructing a fatigue characteristic signal corresponding to the surface electromyographic signal; constructing a multi-input LVN network, inputting surface myoelectric signals and M fatigue characteristic signals respectively, and outputting the signals as muscle strength signals; then, taking the surface electromyographic signals of each sample object and the corresponding fatigue characteristic signals as the input of the multi-input LVN network, taking the corresponding electromyographic signals as expected output, and training the multi-input LVN network; when muscle strength estimation with the same action is needed, surface myoelectric signals are obtained after the surface myoelectric signals are collected and preprocessed by the same method as the sample object, fatigue characteristic signals are constructed and obtained, and the fatigue characteristic signals are input into a trained multi-input LVN network to obtain a muscle strength estimation result.

The invention has the following technical effects:

1) the LVN network is applied to muscle strength estimation for the first time, and the accuracy and the robustness of the muscle strength estimation are improved by combining fatigue characteristics compared with a method for estimating the muscle strength by only adopting surface electromyographic signals;

2) when the characteristics with strong correlation with the fatigue degree are screened out from the alternative characteristics as fatigue correlation characteristics, the characteristic description of the surface electromyographic signals is more accurate, and the accuracy of muscle strength estimation is improved;

3) the specific structure of the LVN network is researched, and an optimization mode of fusing fatigue characteristic signals and punishing local sparsity on the fusion weight of the fatigue characteristic signals is provided, so that the characteristic weight can be adjusted in a self-adaptive mode, the contribution of MDF and MPF characteristics with large correction effect to a model is highlighted, and the accuracy of muscle strength estimation is further improved;

4) according to the LVN network training method, the continuous domain ant colony algorithm is adopted, so that the training time is reduced, the global optimization is improved, the LVN network obtained through training has better performance, and the accurate muscle strength estimation is ensured.

Drawings

FIG. 1 is a flow chart of an embodiment of a muscle strength estimation method based on fatigue analysis according to the present invention;

FIG. 2 is a flowchart of the preprocessing of the surface myoelectric signal in the present embodiment;

FIG. 3 is a flowchart of the preprocessing of the muscle strength signal in the present embodiment;

FIG. 4 is a flowchart for determining fatigue-related characteristics in the present embodiment;

FIG. 5 is a block diagram of a multiple input LVN network;

FIG. 6 is a flow chart of training a multi-input LVN network based on a continuous domain ant colony training algorithm in the present embodiment;

FIG. 7 is a graph comparing muscle strength estimates for five stages of fatigue under constant force pulse data for brachioradialis according to the present invention and four comparison methods;

FIG. 8 is a comparison graph of significance analysis of five fatigue stages of testing criteria under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods;

FIG. 9 is a graph comparing muscle force estimates for five stages of fatigue under constant force pulse data for brachioradialis according to the present invention and four comparison methods;

FIG. 10 is a comparison graph of significance analysis of five fatigue stages of test markers under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods;

FIG. 11 is a diagram showing a test electromyographic signal of the subject 4 in the experiment of the ascending force pulse of the brachioradial muscle in the present embodiment;

FIG. 12 is a comparison graph of muscle force estimates for different stages of fatigue in the pulse test of the ascending force of the brachioradialis muscle for the present invention and four comparison methods;

FIG. 13 is a graph comparing muscle force estimates for five stages of fatigue under constant force pulse data for brachioradialis according to the present invention and four comparative methods;

FIG. 14 is a comparison graph of significance analysis of five fatigue stages of testing criteria under constant force fatigue pulse data of brachioradialis for the present invention and four comparison methods;

figure 15 is a graph comparing five phase indices for four brachioradial muscle ascent force data for the present invention and four comparison methods.

Detailed Description

The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.

Examples

FIG. 1 is a flow chart of an embodiment of the muscle strength estimation method based on fatigue analysis according to the present invention. As shown in fig. 1, the muscle strength estimation method based on fatigue analysis of the present invention specifically includes the steps of:

s101: collecting a sample signal:

for K sample objects, acquiring a surface electromyogram signal and a muscle strength signal of each sample object under a preset action, preprocessing the surface electromyogram signals according to a preset method to obtain preprocessed surface electromyogram signals x _k (i) And muscle strength signal y _k (i) Where i 1.. T, T denotes the signal length.

For example, when the force generation pattern of the preset action is a constant force pulse or an ascending force pulse, the surface electromyographic signal may be a brachioradial muscle surface electromyographic signal or a palmaris longus surface electromyographic signal, and the muscle strength signal may be acquired by the grip strength sensor.

Fig. 2 is a flowchart of preprocessing the surface myoelectric signal in the present embodiment. As shown in fig. 2, the specific steps of the pre-processing of the surface myoelectric signal in this embodiment include:

s201: denoising:

because the surface electromyogram signal has motion artifacts and useless components, the surface electromyogram signal needs to be denoised.

In the embodiment, the surface myoelectric signal is filtered by a band-pass filter of 15Hz to 400Hz to remove noise.

S202: moving average:

and carrying out moving average on the denoised surface electromyographic signals to remove some abnormally prominent peaks and avoid muscle force estimation errors.

S203: normalization:

in order to eliminate the difference between data of different objects and data of different groups of the same object, the surface electromyogram signals after the moving average needs to be normalized to zero mean and unit variance.

S204: down-sampling:

in order to reduce the time cost of subsequent muscle strength evaluation model training, the normalized surface electromyographic signals are subjected to down-sampling according to a preset frequency.

S205: and (3) correcting and smoothing:

research shows that the surface electromyogram signal obeys 0 mean value and sigma variance ² And Gaussian distribution, so that the surface electromyogram signal after the down sampling can be processed by adopting a correction smoothing method to obtain the surface electromyogram signal after the correction smoothing. A specific principle and method of modified smoothing may be found in the references "Hayashi H, Furui A, Kurita Y, et al. A variance distribution model of surface elements based on inverse variance distribution [ J].IEEE Transactions on Biomedical Engineering,2017,64(11):2672–2681.DOI:10.1109/TBME.2017.2657121.”

Fig. 3 is a flowchart of preprocessing the muscle strength signal in this embodiment. As shown in fig. 3, the specific steps of the pre-processing of the muscle strength signal in this embodiment include:

s301: moving average:

in this embodiment, the muscle strength signal acquisition device is provided with the filtering processing device, so that the acquired muscle strength signal is directly subjected to the sliding average.

S302: normalization:

and normalizing the muscle force signal after the sliding average by using the maximum muscle force value of the sample object under the preset action.

S303: and (3) down-sampling:

similarly, the normalized muscle force signal is down-sampled at a preset frequency by using the same down-sampling frequency of the surface muscle electric signal.

S102: analyzing the surface electromyogram signal activity section:

because the muscle activity is not continuous, the surface electromyogram signals have an active section and a rest section, and the rest section signals are not necessary to be analyzed, so that each surface electromyogram signal x needs to be firstly analyzed _k (i) The activity segment analysis is carried out by the following specific method:

with a length L _win Electromyographic signal x of sliding window with sliding step length delta on surface _k (i) Performing sliding extraction on the signal segments, calculating the average instantaneous energy of each signal segment, and the average instantaneous energy E of the jth signal segment _k (j) The calculation formula of (a) is as follows:

wherein j is 1,2, …, N _part ，N _part Indicating the number of divided signal segments.

Setting an energy threshold

When in use

And is

And is provided with

Taking the terminal point of the jth signal segment as the terminal point of the active segment, and performing no operation in other conditions to obtain the surface electromyogram signal x _k (i) The active segment of (1).

Respectively representing surface electromyographic signals x _k (i) The original sampling point sequence numbers of the beginning and end of the ith active segment, D ═ 1,2, …, D _k 。

S103: extracting fatigue-related features of myoelectricity:

g characteristics related to fatigue in the surface electromyogram signals are set according to actual needs, and then each surface electromyogram signal x is subjected to _k (i) Respectively extracting G characteristics f of each activity section _k,d,g ，g＝1,2,…,G。

In general, the features of the surface electromyogram signal are classified into time-domain features, frequency features, and non-gaussian features, for example, the time-domain features include an integral electromyogram value, an amplitude absolute value mean value, a slope of the amplitude absolute value mean value, a root mean square, a waveform length, a difference of the signal absolute mean value, a sample variance, a zero crossing rate, a wilson amplitude, a simple square accumulation, a slope sign change, and the like, the frequency-domain features include a peak frequency, an average frequency, a total power, a median frequency, an average instantaneous frequency, a spectral moment parameter, and the like, and the non-gaussian features include a kurtosis, a negative entropy, and the like. In practical application, the fatigue-related characteristics can be screened from the above characteristics through technical means such as experience, qualitative analysis and quantitative analysis.

Fig. 4 is a flowchart for determining fatigue-related characteristics in the present embodiment. As shown in fig. 4, the specific steps of determining the fatigue-related characteristic in this embodiment include:

s401: determining alternative features:

and G' alternative characteristics are determined according to actual needs.

S402: extracting alternative features:

for each surface electromyographic signal x _k (i) Respectively extracting G' features f of each activity segment _k,d,g′ G '1, 2, …, G', each surface electromyography signal x is obtained _k (i) The g' th characteristic in the sequence of characteristics

S403: obtaining a fatigue value sequence:

since the fatigue value increases as the action time is longer, in this embodiment, the serial number of the active segment is taken as the fatigue value and normalized to obtain each surface electromyogram signal x _k (i) Fatigue value series O ═ 1/D _k ,2/D _k ,…,1}。

S404: calculating a decision coefficient:

for each surface electromyographic signal x _k (i) Fatigue value as independent variable and characteristic value as dependent variable, respectively for each characteristic sequence

And fatigue value sequence O ═ 1/D _k ,2/D _k …,1} is subjected to linear regression to obtain the coefficient of determination

Determining coefficients

The linear correlation of the g' th feature with the fatigue level can be measured. Determining coefficients

The larger the value, the higher the proportion of the g' th feature affected by the degree of fatigue. Otherwise the lower.

For K surface electromyographic signals x _k (i) Coefficient of determination of the g' th feature

Averaging to obtain the coefficient of determination of the g' th feature

S405: calculating the correlation coefficient of the pearson:

for each surface electromyographic signal x _k (i) Separately computing each signature sequence

And fatigue value sequence O ═ 1/D _k ,2/D _k Pearson's correlation coefficient P of …,1 _k,g′ . Then K surface electromyographic signals x _k (i) Pearson's correlation coefficient P of the g' th feature _k,g′ Averaging to obtain the Pearson correlation coefficient of the g' th feature

The larger the value of the pearson correlation coefficient, the higher the proportion of the feature affected by the degree of fatigue, and vice versa.

S406: calculating SVR coefficients:

since some changes in electromyographic features are non-linear, and both the determinant coefficient and the pearson correlation coefficient can only measure the linear correlation of features, the SVR (sensitivity-to-variability of the trend) coefficient proposed in the document Rogers D R, MacIsaac D T.A composition of emg-based muscle fault assessment and along dynamic contract is also introduced in this embodiment, in order to more accurately evaluate the change of features with the degree of fatigue when the trend of fatigue is non-linear. The larger the SVR coefficient value, the more sensitive the feature is. The calculation method of the SVR coefficient is as follows:

for each surface electromyographic signal x _k (i) For each characteristic sequence separately

And fatigue value sequence O ═ 1/D _k ,2/D _k …,1} is fitted a second order, fitting polynomial as follows:

wherein, a ₀ ,a ₁ ,a ₂ As coefficient of fit, o _d Representing the d fatigue value in the fatigue value sequence O.

Then, each characteristic f is calculated by adopting a polynomial obtained by fitting _k,d,g′ Characteristic value of fitting (2)

Calculating to obtain the surface myoelectric signal x by adopting the following formula _k (i) SVR coefficient SVR of the g' th characteristic _k,g′ ：

Therein, max _k,g′ 、min _k,g′ Representing surface electromyographic signals x _k (i) The maximum and minimum values of the g' th feature in all active segments.

For K surface electromyographic signals x _k (i) SVR coefficient SVR of the g' th characteristic _k,g′ Averaging to obtain the SVR coefficient of the g' th feature

S407: screening for fatigue-related characteristics:

screening out SVR coefficients for G' alternative characteristics

Greater than a predetermined threshold T _SVR Determining the coefficients

Greater than a predetermined threshold

And Pearson's correlation coefficient

Greater than a predetermined threshold T _P As fatigue-related characteristics.

Through the three parameters, the characteristics which are relatively high in correlation with the fatigue degree and relatively sensitive in response to the change of the fatigue degree can be screened out, so that the accuracy of muscle strength estimation is improved. In this embodiment, the features screened by the above screening method, combined with qualitative analysis and experience, include 11 features in Total, i.e., integrated myoelectric Value (IEMG), Mean Absolute Value of amplitude (MAV), Root Mean Square (RMS), Difference Absolute Value of signal (DAMV), Variance (Variance of signal, VAR), Simple Square Integration (SSI), Total Power (TP), Mean Frequency (MPF), Median Frequency (MDF), negative entropy gen, and 5 spectrum distance parameter FInsm 5.

S104: constructing a characteristic signal:

for each surface electromyographic signal x according to active segment division _k (i) Respectively constructing G fatigue characteristic signals F thereof _k,g (i) I.e. as the sampling point

The fatigue characteristic signal F corresponding to the mth characteristic _k,g (i)＝f _k,d,g Otherwise F _k,g (i) 0. That is to say, the characteristic signal corresponding to the active segment is filled by the characteristic value calculated by the active segment, and the characteristic signal corresponding to the rest segment is 0, so that the characteristic signal only contains the myoelectric characteristics during action, and the surface myoelectric signal is more accurately characterized. And the characteristic signal and the surface electromyogram signal x thus obtained _k (i) The lengths are equal.

S105: constructing and training a multi-input LVN network:

the LVN (Laguerre-Volterra Network, Laguerre-Volterra) Network is a Network model constructed based on a Volterra model, Laguerre Extension Technique (LET) and a standard Dynamic base (PDM), and the specific principle can be referred to the literature "Geng K, Marmarmarares V Z]IEEE Transactions on Neural Networks and Learning Systems,2017,28(9): 2196-2208. DOI:10.1109/TNNLS.2016.2581141. The LVN model comprising a plurality of input signals is a multi-input LVN network. Fig. 5 is a block diagram of a multiple input LVN network. As shown in FIG. 5, the number of input signals of the multi-input LVN network is recorded as N, N inputs x _n (t) outputs via respective filter banks are

For the nth input x _n (t) and (j) _n Laguerre function

Is expressed as follows:

wherein N is 1.., N; j is a function of _n ＝0,...,L _n -1，L _n The number of the Laguerre functions input for the nth group; t1, T is the data point length; m is the memory length;

represents the j laguerre function, the expression is as follows:

where M is 0.., M-1, and α is a decay exponent, which determines the convergence of the laguerre function.

Respectively representing coefficients calculated from combinations of v values taken from m and j.

Second layer output z _h (t) is defined as shown in the following formula:

wherein, H is 1. c. C _q,h Representing a non-linear function f _h The coefficient of the (c) is,

representing L from each filter bank _n The connection coefficients of the outputs to the H standard dynamic bases.

Substituted output

Second layer output z _h The expression of (t) is modified as follows:

thus, the h standard dynamic basis for the n input can be expressed as:

finally, the output of the N-input LVN model is:

the signals for evaluating the muscle force comprise surface electromyogram signals and fatigue characteristic signals extracted from the surface electromyogram signals, so that the input of the multi-input LVN network is the surface electromyogram signals and the M fatigue characteristic signals respectively, and the output is the muscle force signals. Then the surface electromyographic signal x of each sample object is measured _k (i) And taking the corresponding fatigue characteristic signal as the input of the multi-output LVN network, taking the corresponding myoelectric signal as the expected output, and training the multi-input LVN network.

In practical application, in order to simplify the structure of the LVN network and improve the muscle strength estimation performance of the LVN network, a preferable mode is to set the LVN network as a two-input LVN network, wherein one input is a surface myoelectric signal, and the other input is according to a preset weight w _m And the M characteristic signals are subjected to weighted summation to obtain a fusion characteristic signal, so that the number of input signals is reduced under the condition of not reducing the number of characteristics, the complexity of an LVN (linear variable network) is reduced, and the muscle strength estimation efficiency is improved. The weight w may be used in the two-input LVN network training _m As a parameter to be trained, so as to make the fused feature signal more effectively represent the feature of the surface electromyogram signal.

In the multi-input LVN network training, a Normalized Mean Square Error (NMSE) may be used as the cost function. When a mode of fusing M characteristic signals into one path of fused characteristic signals is adopted, sparsity punishment on the weight of the characteristic signals can be added into the cost function so as to improve the training effect. Namely, the calculation formula of the cost function J of the two-input LVN network under the scheme is:

J＝NMSE+λ||W|| ₁

the NMSE represents the root mean square error of the estimated muscle force signal and the real muscle force signal, the lambda represents a preset penalty intensity coefficient, the W represents a penalty weight vector, and the forming method comprises the following steps: set up as required in G characteristic signals

The individual signature signal is used as a penalty signature signal,

constructing the weight corresponding to each punishment characteristic signal to obtain

A dimensional row vector, namely a penalty weight vector W; i | · | purple wind ₁ Indicating that the norm of L1 is found.

According to the above description, when performing sparsity punishment on the feature signal weights, it may perform sparsity punishment on all the feature signal weights, or may perform local sparsity punishment only by selecting a part of the feature signal weights, and may be selected according to actual conditions in practical applications. Experiments in this embodiment show that the characteristic signals of the 11 screened characteristics are fused and used as one path of input, and then local sparsity punishment is performed on characteristic weights except for average frequency MPF and median frequency MDF in a cost function, so that the obtained LVN network has better performance. For convenience of description, the LVN network is referred to as LVN-pS network.

As for a specific training method, selection can be performed according to actual needs, such as a simulated annealing training algorithm and a continuous domain ant colony training algorithm, and experiments show that, for the continuous domain ant colony training algorithm, an approximate global optimal solution can be obtained in a short training time, so that training overhead is reduced, and training efficiency is improved. Fig. 6 is a flowchart of training the multi-input LVN network based on the continuous domain ant colony training algorithm in the embodiment. As shown in fig. 6, the specific steps of the multi-input LVN network training based on the continuous domain ant colony training algorithm in this embodiment include:

s601: initializing LVN network parameters:

initializing LVN network parameters, including N-way input Laguerre radix L _n N is 1,2, …, N, the number of standard dynamic bases H, the model order Q, etc.

Taking LVN-pS network as an example, the input is 2 paths, and the coefficient to be estimated comprises the attenuation index alpha of 2 paths of input Laguerre coefficients ₁ ,α ₂ The output signal is offset by a constant y ₀ By a coefficient { c _q,h A matrix C of Q rows and H columns, Q1, Q, H1, Q

Of composition L ₁ Matrix W1 of rows and columns of H, formed by coefficients

Of composition L ₂ A matrix W2 of rows and columns and a weight vector W formed by weights when fusing 11 eigenvectors _fea Therefore, the total number Np of coefficients to be estimated is 3+ Q H + L ₁ *H+L ₂ *H+11。

S602: initializing parameters of the continuous-domain ant colony algorithm:

setting relevant parameters of the continuous domain ant colony algorithm, including solution combination quantity A in the solution file SOL, generating new solution combination quantity B in each circulation, and calculating the selection probability p of A solutions _a The calculation formula is as follows:

β _a the weight of each solution combination is represented by the following calculation formula:

where γ represents a preset adjustment parameter. The smaller the gamma is, the larger the solution weight at the front of the sequence is, the faster the convergence is, otherwise, the more uniform the weight distribution is, the stronger the global optimization capability of the algorithm is, and the slower the convergence is.

S603: initializing a solution file:

the solution file SOL is a matrix of a rows Np +1, each row including a solution combination and a cost function value corresponding to the solution combination. Therefore, when the solution archive SOL is initialized, the solutions are combined into parameter combinations of the LVN network, A initial solutions are generated firstly, then the LVN network is configured based on each solution, a training sample is input into the LVN network to obtain an estimated muscle strength signal, and a cost function value is calculated according to a cost function calculation formula by combining an expected muscle strength signal.

Using the LVN-pS network as an example, the solution combination can be expressed as { alpha } ₁ ,α ₂ ,y ₀ ,C,W1,W2,W _fea }. Then alpha is used in the embodiment to initialize the solution file ₁ ,α ₂ In [0,1]]The range is uniform, and the rest parameters are [ -1, 1]]The values are uniformly taken in the range. Then according to alpha ₁ And the 1 st input in the training sample obtains a corresponding V ₁ According to the weight W _fea And the fatigue characteristic signal in the training sample calculates the 2 nd path input, and then combines alpha ₂ Calculating to obtain corresponding V ₂ Then, calculating the intermediate output U ═ V of the LVN-pS network ₁ *W ₁ +V ₂ *W ₂ ,U2＝U*U ^T Finally, the estimated muscle force signal y of the LVN-pS network is obtained through calculation _e ＝U*C(1,:) ^T +U2*C(2,:) ^T . And calculating to obtain a cost function value corresponding to the solution combination by adopting a cost function calculation formula based on local sparsity punishment according to the estimated muscle force signal and the expected muscle force signal of the training sample.

S604: let the iteration number τ be 1.

S605: a random sample value is generated in the range of [0,1], if the random sample value is less than the preset threshold, the step S606 is proceeded, otherwise, the step S607 is proceeded.

S605: screening guide solution combination and generating new solution combination:

sorting A solution combinations in the current solution file SOL from small to large according to cost function values, and determining the selection probability p of each solution combination according to the sorting _a . Then in [0,1]]B sampling values are randomly generated, and for each sampling value, the solution combination in which the sampling value falls is judgedAnd selecting the corresponding probability interval accumulatively, and combining the corresponding solutions into a guide solution combination. And respectively generating a new solution combination according to each guide solution combination, wherein the specific method comprises the following steps:

for each parameter to be estimated of the LVN network, taking the value of the parameter to be estimated in the guiding solution combination as a mean value mu, and then calculating by adopting the following formula to obtain a standard deviation sigma:

wherein, SOL _a And xi represents a preset forgetting factor, and the larger the value of the parameter to be estimated in the a-th solution combination in the current solution file SOL is, the slower convergence is, and the worse solution can be forgotten more slowly.

And sampling the Gaussian distribution with the mean value mu and the standard deviation sigma to obtain the value of the parameter to be estimated in the new solution, thereby generating 1 new solution combination.

And for the generated B new solution combinations, configuring an LVN network based on each new solution combination, inputting the training samples into the LVN network to obtain estimated muscle force signals, and calculating according to a cost function calculation formula by combining the expected muscle force signals to obtain a cost function value of each new solution combination.

S607: sequentially as a guide solution combination and generating a new solution combination:

and sequentially taking A solution combinations in the current solution file SOL as guide solution combinations to generate A new solution combinations. And then configuring an LVN network based on each new solution combination, inputting the training samples into the LVN network to obtain estimated muscle force signals, and calculating according to a cost function calculation formula by combining the expected muscle force signals to obtain a cost function of each new solution combination. The global optimization performance can be increased by probabilistically using the ensemble of solution combinations as a guide solution to generate a new solution combination.

S608: and (3) generating a new solution file:

combining A solution combinations in the current solution file SOL and all generated new solution combinations, sorting the solution combinations from small to large according to the cost function values, and screening the former A solution combinations to form a new solution file SOL.

S609: judging the iteration number tau < tau _max If so, the process proceeds to step S610, otherwise, the process proceeds to step S611.

S610: let τ be τ +1, return to step S605.

S611: determining LVN network parameters:

and combining the solutions with the optimal cost function values in the current solution files SOL into LVN network parameters, and configuring the LVN network.

S107: and (3) muscle strength estimation:

when muscle force estimation with the same action is required, a surface electromyographic signal x ' (i) is acquired and preprocessed by the same method as the sample object, G characteristics related to fatigue of each activity section are extracted after activity section division is carried out, and G fatigue characteristic signals F ' are generated ' _g (i) Then the surface electromyographic signal x '(i) and G fatigue characteristic signals F' _g (i) Inputting the trained multi-input LVN network to obtain an estimated muscle force signal.

In order to better illustrate the technical scheme of the invention, the invention is experimentally verified by adopting specific examples.

In the experimental verification, the collected signals are the surface electromyographic signals and the grip strength signals of the sample object. The experimental device mainly comprises a host MP150 of BIOPAC company, an electromyography acquisition module EMG100C, an analog signal collector DA100C and a grip strength sensor TSD 121C. The subjects included 9 healthy dextrorotatory subjects (6 young men and 3 young women, all 25 years old). The experiment comprises a maximum autonomous contractility experiment and a constant force pulse experiment of brachioradialis and palmaris longus in the process, and the specific conditions are as follows:

1) maximum spontaneous contractility test: firstly, the experimental subject needs to apply Maximum Voluntary Contraction force (MVC) twice, each time of exerting force holds the Maximum force of the experimental subject as much as possible for 4 to 5 seconds, and the interval between the two times of exerting force is more than 2 minutes, so that muscle fatigue is avoided. MVC was the average of the forces of two experiments.

2) Constant force pulse experiment: after the experiment of the maximum autonomic contraction force is finished, the patient needs to take a full rest for more than 5 minutes. The subject started a constant force pulse experiment and applied a pulse force of 60% MVC. The force application mode is as follows: the force is applied for 5s, the rest is performed for 5s, and the process is repeated until the experimental object cannot apply force to 60% MVC continuously twice.

The experiments are one group, and due to the needs of model evaluation (one group is used for training and the other group is used for testing), each object needs to repeat two groups of experiments, and the interval between the two groups is more than 24 hours so as to ensure that the fatigue feeling of the previous experiment completely disappears.

In order to illustrate the effectiveness of introducing the fatigue characteristic signal, 6 types of LVN networks are set in the experiment for comparison, and the 6 types of LVN networks are as follows:

LVN _ 1: the LVN network adopts a single-input LVN network and only has surface electromyogram signals.

LVN _2 MPF: the LVN network adopts a two-input LVN network, one input is a surface electromyogram signal, and the other input is a fatigue characteristic signal corresponding to an average frequency (MPF) characteristic.

LVN — 2 MDF: the LVN network adopts a two-input LVN network, one input is a surface electromyogram signal, and the other input is a fatigue characteristic signal corresponding to a median frequency (MDF) characteristic.

LVN _ 3: the LVN network adopts a three-input LVN network, one input is a surface electromyogram signal, and the other two inputs are fatigue characteristic signals corresponding to an average frequency (MPF) characteristic and a median frequency (MDF) characteristic respectively.

The cost functions of the 4 types of LVN networks adopt normalized root mean square error NMSE.

LVN _ S: the LVN network adopts a two-input LVN network, one input is a surface electromyogram signal, and the other input is a fusion characteristic signal obtained by fusing fatigue characteristic signals corresponding to the 11 kinds of characteristics screened in this embodiment. The cost function of the network adopts a cost function which integrates NMSE and sparsity punishment, and 11 fatigue characteristic signals are all used as punishment characteristic signals.

LVN _ pS: the LVN network adopts a two-input LVN network, one input is a surface electromyogram signal, and the other input is a fusion characteristic signal obtained by fusing fatigue characteristic signals corresponding to the 11 kinds of characteristics screened in this embodiment. The cost function of the network adopts a cost function which integrates NMSE and sparsity punishment, and 9 fatigue characteristic signals except average frequency (MPF) characteristic and median frequency (MDF) are used as punishment characteristic signals.

The normalized root mean square error (NMSE%) and Fitness (Fitness%) of the 6 LVN networks were statistically compared. Table 1 is a comparison table of training and testing indexes of 6 LVN networks in this embodiment.

TABLE 1

As shown in table 1, the performance of the 5 LVN networks with the fatigue signature signal introduced therein is better than that of the LVN network with only the surface electromyogram signal as an input, and most of them, the LVN _ pS is the most suitable for the normalized root mean square error and fitting degree of the predicted muscle strength and the real muscle strength in both the network training result and the network testing result.

Next, taking the LVN _ pS model as a representative network structure of the present invention, a muscle strength estimation method based on four models, namely POL (multinomial Polynomial simulation model), FOS (Fast Orthogonal Search model), PCI (Parallel Cascade model), LET (Laguerre expedition Technique) is used as a comparison method to compare with the performance of the present invention. The technical indexes of comparison comprise mean square error MSE%, normalized mean square error NMSE% and fitting degree Fitness%. The brachioradial constant force fatigue pulse experiment, the brachioradial rising force fatigue pulse experiment and the palmaris longus rising force fatigue pulse experiment are designed to verify that the model can show more excellent muscle force estimation performance under different muscles and different force applying modes. In order to show the muscle force estimation conditions of different fatigue stages of the model in various types of data, the test data are divided into five stages, namely a non-fatigue state seg1, a slight fatigue state seg2, a medium fatigue state seg3, a severe fatigue state seg4 and an extreme fatigue state seg5, and the stages of the trained model are used for testing.

Experimental results based on constant force fatigue pulse data of brachioradial muscle

Fig. 7 is a comparison graph of muscle strength estimation indexes of five fatigue stages under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods. Fig. 7 shows the mean values and standard deviations of three technical indicators for muscle strength estimation of data of five stages of brachioradialis at different fatigue degrees, wherein the abscissa represents the five fatigue stages and the ordinate represents the corresponding indicators. It can be seen from fig. 7 that the five methods all show the situation that MSE% and NMSE% decrease and Fitness% increase with increasing fatigue, and it is presumed that the myoelectric-muscle force relationship when the fatigue is more severe is captured in the full-pulse data training process by the models used in the five methods, so that the three phases with higher fatigue degree, i.e. the medium fatigue phase seg3, the severe fatigue phase seg and the extreme fatigue phase seg5, of the test data have higher estimation accuracy. In the five methods, the MSE% and NMSE% of the LVN _ pS model at five different fatigue stages are obviously lower than those of the rest four models, and the fitting degree Fitness% of the muscle force is obviously higher than those of the rest models, so that the LVN _ pS model has stronger capturing capability on the myoelectricity-muscle force relationship and shows excellent muscle force estimation performance when the muscle is in five fatigue states. The mean test MSE%, NMSE%, Fitness% for the LVN _ pS model over the five stages was calculated to reach 5.25%, 2.52% and 77.68%, respectively.

To further illustrate whether there were significant differences between the LVN _ pS model of the present invention and the POL, FOS, LET models at five different fatigue stages, a significance analysis was performed. Fig. 8 is a comparison graph of significance analysis of five fatigue stages of test indexes under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods.

The three graphs of fig. 8(a), (b), and (c) respectively show the results of significance analysis among five models under three evaluation indexes of MSE%, NMSE%, and Fitness%, and from left to right, the results of the five fatigue stages seg1, seg2, seg3, seg4, and seg5, respectively. Middle gray NS indicates p >0.05 between the two models corresponding to the abscissa and ordinate, with no significant difference. Light grey indicates p <0.05 and dark grey indicates p <0.01, both indicating significant differences between models and differences in dark grey are more pronounced. With the three indexes of LVN _ pS in fig. 6 being optimal, it can be seen from fig. 7 that, except that the LVN _ pS model of the present invention is significantly better than the LET model by a degree of p <0.05 under the MSE% index, the LVN _ pS model of the present invention is significantly better than the four models of POL, FOS, PCI and LET by a degree of p <0.01 for the differential analysis of the three indexes at five different fatigue stages. The POL model, the FOS model, the PCI model and the LET model have no significant difference, which shows that the muscle force estimation effects of the four models are similar.

Experimental results based on fatigue pulse of ascending force of brachioradial muscle

Performance comparisons of the five methods on data of five different fatigue levels of the ascending force of the brachioradialis are then shown. Fig. 9 is a comparison graph of muscle strength estimation indexes of five fatigue stages under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods. As shown in fig. 9, as the fatigue degree of seg1 to seg5 increases, the indexes MSE% and NMSE% of the five methods show a decreasing trend, and the index Fitness shows a rising trend. That is, as the fatigue feeling of the subject becomes more severe, the effect of estimating the muscle strength tends to be gradually increased. This may occur because the model training is based on a full pulse set of data including five fatigue phases, and the model may capture more information about the more severe phases of fatigue during the training process. Compared with POL, FOS, PCI and LET models, MSE% and NMSE% of LVN _ pS are obviously reduced, and the muscle force fitting degree Fitness is obviously improved, namely the MSE% and NMSE% estimated in the non-fatigue stage seg1, the light fatigue stage seg2, the medium fatigue stage seg3, the heavy fatigue stage seg4 and the extreme fatigue stage seg5 are lower than those of the rest four models, and the muscle force fitting degree Fitness is higher than those of the rest four models.

Fig. 10 is a comparison graph of significance analysis of five fatigue stages of test indexes under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods. As shown in fig. 10, there is no significant difference between the four models of POL, FOS, PCI and LET, except that the three first-stage samples show no significant difference between the LVN _ pS model of the present invention and the LET model, the LVN _ pS model of the present invention has significant differences of different degrees under different indexes compared with the other four models, wherein the three indexes of the LVN _ pS model of the present invention are more significantly better than the FOS model by a degree of p <0.01, and the fixness% of the LVN _ pS model is better than the POL model by a degree of p < 0.01. The mean test MSE%, NMSE%, Fitness% for the LVN _ pS model over the five stages was calculated to reach 7.5%, 3.26% and 73.32%.

Therefore, except that the performance of the LVN _ pS models of the invention and LET models to estimate muscle force on the non-fatigue stage seg1 of the brachioradialis ascension force data was similar, the LVN _ pS models of the invention were significantly superior to POL, FOS, PCI and LET models in the light fatigue stage seg2, the medium fatigue stage seg3, the severe fatigue stage seg4 and the extreme fatigue stage seg 5. In addition, compared with an FOS model, the LVN _ pS model has the most obvious performance improvement.

In order to show the muscle force estimation effects of the five methods on different fatigue stages of the ascending force pulse of the brachioradialis, the muscle force estimation performances of the experimental object 4 at different fatigue stages were tested. Fig. 11 is a test electromyogram of the subject 4 in the experiment of the ascending force pulse of brachioradial muscle in the present embodiment. As shown in fig. 11, the myoelectric signals of the test data of the subject 4 have a tendency of gradually increasing from seg1 to seg4 and a tendency of slightly decreasing in myoelectric amplitude from seg 5. Fig. 12 is a comparison graph of muscle force estimates for different stages of fatigue in the pulse test of the ascending force of the brachioradial muscle for the present invention and four comparison methods. As shown in FIG. 12, it can be seen that in the five methods, although there are small deviations in the LVN _ pS model of the present invention, the five different fatigue stages are superior to the rest of the models.

In summary, for the brachioradialis lift fatigue data, in the five methods, the LVN _ pS model of the invention shows the optimal training and testing effect, and the overall evaluation effect on the brachioradialis lift full-pulse data and the muscle force at five different fatigue degree stages is significantly better than that of POL, FOS, PCI and LET models.

Results based on palmaris longus lift fatigue pulse data

Fig. 13 is a comparison graph of muscle strength estimation indicators for five fatigue stages under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods. As shown in fig. 13, the three indicators of the LVN _ pS model of the invention were significantly lower than those of the remaining four models in five different fatigue phases, with some overlap of all indicators of the remaining four models. The five models have the trends that MSE% and NMSE% decrease and Fitness increases along with the increase of the fatigue degree, and the trends of the four models except the LVN _ pS model are more obvious. The LET and PCI models showed a tendency of gradual degradation in the three indexes of the severe fatigue stage seg4 and the extreme fatigue stage seg 5. The calculated average test MSE%, NMSE% and Fitness% of the LVN _ pS model of the invention in five stages can reach 9.08%, 3.79% and 70.35%.

Fig. 14 is a comparison graph of significance analysis of five fatigue stages of test indexes under constant force fatigue pulse data of brachioradialis according to the present invention and four comparison methods. As shown in fig. 14, for three evaluation indexes, there is no significant difference in muscle strength estimation performance among POL, FOS, PCI, and LET models under data verification of five different fatigue stages, while there is a significant difference between the muscle strength estimation performance of the LVN _ pS model of the present invention and both the POL and FOS models in five stages of different fatigue degrees, i.e., the muscle strength estimation performance of the LVN _ pS model of the present invention is significantly better than that of the POL and FOS models in different stages. However, the MSE%, NMSE% and three indexes of the LVN _ pS model in the non-fatigue stage seg1 and in the extreme fatigue stage seg5 are not significantly different from those of the PCI and LET models, and the ranges of the LVN _ pS models are covered by combining the larger standard deviation of the PCI and LET in seg1 and seg5 in fig. 12, which shows that the PCI, LET and LVN _ pS models have closer estimated performance for the muscle force of a certain subject in seg1 and seg5 stages, so that the significant difference is lost. The muscle force estimation performance of the three models in the non-fatigue stage seg1 and the extreme fatigue stage seg5 are similar, and the reason may be that the palm long muscle information of individual subjects is weak, and the fatigue information cannot be provided for the LVN _ pS model of the invention through the myoelectric fatigue characteristics for reference. Although the problem of individual difference exists, the average performance of the LVN _ pS model is still better than that of the PCI and LET models.

In conclusion, for the fatigue pulse data of the lifting force of the long palm muscle, although the LVN _ pS model has no significant difference with the indexes of PCI and LET in MSE% and NMSE%, the average test performance is still better than that of the latter two models. In five stages of fatigue data with different degrees, although the testing effect of the LVN _ pS model is not obviously different from that of the PCI and LET models in the stages of seg1 and seg5, the average performance of the muscle strength estimation of the LVN _ pS model is still better than that of the latter two models. Therefore, under the palm long muscle lift fatigue data, the LVN _ pS model of the invention has significantly better test muscle force estimation performance in the full pulse training test and different fatigue degree stages than the POL and FOS models, and also has significantly better average muscle force performance than the PCI and LET models.

Stability analysis based on multiple sets of data for the same object

And then, performing muscle force estimation performance verification of different groups of data among the same subjects based on fatigue data of the same brachial radial muscle ascending force experiment of two subjects with different sexes on different five days. For two subjects, five sets of experimental data, five days later, were named D1, D2, D3, D4, and D5 data. Each group of data comprises myoelectric and muscle force signals. The D1 data was used for training and the last four sets of data were used for testing. Training and test results from the full pulse training and test and training and test results from different fatigue phases are then shown.

In order to analyze the muscle force estimation performance of each model in the four groups of test data for five different fatigue stages (seg 1-seg 5), the model trained on the D1 data is tested for five different stages of D2-D5. Figure 15 is a graph comparing five phase indices for four brachioradial muscle ascent force data for the present invention and four comparison methods. FIGS. 15(a), (b), (c) and (D) show the results of the tests on the sets of data D2, D3, D4 and D5, respectively, and the coordinate ranges of the same indices in different panels are set to be consistent for comparison. It can be seen that in the three data sets D2, D3 and D5, the LVN _ pS model of the present invention has the phenomena of MSE%, NMSE% decrease and Fitness% increase significantly compared to the remaining four models in five stages from the non-fatigue stage seg1, the light fatigue stage seg2 to the extreme fatigue stage seg 5. While the remaining four models performed better in the D4 data.

In order to comprehensively evaluate the muscle force estimation performance of the five models in different fatigue stages, the mean value and the standard deviation of all five stage test indexes of the D2-D4 group data are calculated. Table 2 is a comparison table of the average test indexes of the invention and four comparison methods for different fatigue stages of the data of D2-D4.

TABLE 2

As shown in table 2, the LVN _ pS model of the invention showed 9.6% and 4.15% reduction in MSE% and 4.05% improvement in Fitness% compared to the optimal LET of the rest models for different fatigue stages under four sets of data, 12.11% and 5.21% reduction in MSE% and 13.05% improvement in Fitness% compared to the FOS model with the worst muscle force estimation performance in the rest models.

That is, for the multiple groups of brachioradial muscle lift fatigue data of the same object, the LVN _ pS model of the invention shows more stable average muscle force estimation performance compared with the remaining four models, and the average muscle force estimation performance is better than the remaining four models at five different stages of fatigue degree.

In conclusion, the invention improves the performance of muscle force estimation by introducing fatigue characteristics, and compared with the prior art, the invention has better performance on different types of surface electromyographic signals and fatigue degrees and stronger robustness.

Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims

1. A muscle force estimation method based on fatigue analysis is characterized by comprising the following steps:

s1: for K sample objects, each sample is collectedThe surface electromyogram signal and the muscle strength signal of the object under the preset action are preprocessed according to a preset method to obtain a preprocessed surface electromyogram signal x _k (i) And muscle strength signal y _k (i) Wherein i ═ 1., N denote signal length;

setting an energy threshold

When in use

And is provided with

And is provided with

Taking the terminal point of the jth signal segment as the terminal point of the active segment, and performing no operation in other conditions to obtain the surface electromyogram signal x _k (i) Dividing the active segment;

recording surface electromyography signal x _k (i) The number of active segments obtained is D _k And the d-th activity segment is represented as

Respectively representing surface electromyographic signals x _k (i) Original sampling point sequence numbers of the starting point and the end point of the D-th activity segment, D is 1,2, …, D _k ；

S4: for each surface electromyographic signal x according to active segment division _k (i) Respectively constructing G fatigue characteristic signals F thereof _k,g (i) When sampling points

S5: constructing a multi-input LVN network, inputting surface myoelectric signals and M fatigue characteristic signals respectively, and outputting the signals as muscle strength signals; then the surface electromyographic signal x of each sample object is measured _k (i) The corresponding fatigue characteristic signal is used as the input of the multi-output LVN network, the corresponding electromyographic signal is used as the expected output, and the multi-input LVN network is trained;

s6: when the muscle force estimation of the same action is required, a surface electromyogram signal is acquired and preprocessed by the same method as the sample object to obtain a surface electromyogram signal x ' (i), the activity segments are divided, G characteristics of each activity segment related to fatigue are extracted, and G fatigue characteristic signals F ' are generated ' _g (i) Then the surface electromechanical signal x '(i) and G fatigue characteristic signals F' _g (i) Inputting the trained multi-input LVN network to obtain an estimated muscle force signal.

2. The muscle force estimation method according to claim 1, wherein the preprocessing method of the surface myoelectric signal in step S1 includes the steps of:

1) denoising the surface myoelectric signal;

2) carrying out moving average on the denoised surface electromyographic signals;

3) normalizing the surface electromyographic signals after the moving average to a zero mean value and a unit variance;

4) performing down-sampling on the normalized surface electromyographic signals according to a preset frequency;

5) processing the surface electromyogram signal after the down sampling by adopting a correction smoothing method to obtain a surface electromyogram signal after the correction smoothing;

the muscle force signal preprocessing method comprises the following steps:

1) performing a moving average on the muscle force signal;

2) normalizing the muscle force signal after the sliding average by adopting the maximum muscle force value of the sample object under the preset action;

3) and carrying out down-sampling on the normalized muscle force signal according to a preset frequency by adopting the same down-sampling frequency of the surface muscle force signal.

3. The muscle force estimation method according to claim 1, wherein the determination method of the characteristics in step S3 includes the steps of:

s3.1: determining G' alternative characteristics according to actual needs;

s3.2: for each surface electromyographic signal x _k (i) Respectively extracting G' features f of each activity segment _k,d,g′ G '1, 2, …, G', each surface electromyography signal x is obtained _k (i) The g' th characteristic in the sequence of characteristics

S3.3: taking the serial number of the active segment as a fatigue value and normalizing to obtain each surface electromyogram signal x _k (i) Fatigue value series O ═ 1/D _k ,2/D _k ,…,1}；

S3.4: for each surfaceElectromyographic signal x _k (i) Fatigue value as independent variable and characteristic value as dependent variable, respectively for each characteristic sequence

Averaging to obtain the coefficient of determination of the g' th feature

S3.5: for each surface electromyogram signal x _k (i) Separately computing each signature sequence

And fatigue value sequence O ═ 1/D _k ,2/D _k Pearson correlation coefficient P of …,1 _k,m′ (ii) a Then K surface electromyographic signals x _k (i) Pearson's correlation coefficient P of the g' th feature _k,g′ Averaging to obtain the Pearson correlation coefficient of the g' th feature

S3.6: for each surface electromyographic signal x _k (i) For each characteristic sequence separately

wherein, a ₀ ,a ₁ ,a ₂ As coefficient of fit, o _d Representing the d fatigue value in the fatigue value sequence O;

then, each characteristic f is obtained by adopting polynomial calculation obtained by fitting _k,d,g′ Fitting eigenvalue of

Therein, max _k,g′ 、min _k,g′ Representing the surface electromyogram signal x _k (i) The maximum value and the minimum value of the g' th characteristic in all the active segments;

S3.7: screening out SVR coefficients for G' alternative characteristics

Greater than a predetermined threshold T _SVR Determining the coefficients

Greater than a predetermined threshold

And Pearson's correlation coefficient

Greater than a predetermined threshold T _P As fatigue-related characteristics.

4. The muscle force estimation method according to claim 1, wherein the multi-input LVN network in the step S6 is a two-input LVN network, one input is a surface electromyography signal, and the other input is according to a preset weight w _m And carrying out weighted summation on the M characteristic signals to obtain a fused characteristic signal.

5. The muscle force estimation method according to claim 4, wherein the cost function J of the two-input LVN network is calculated by the formula:

J＝NMSE+λ||W|| ₁

the NMSE represents the root mean square error of the estimated muscle force signal and the real muscle force signal, the lambda represents a preset punishment intensity coefficient, the W represents a punishment weight vector, and the constitution method comprises the following steps: set up as required in G characteristic signals

The individual signature signal is used as a penalty signature signal,

A dimensional row vector, namely a penalty weight vector W; i | · | live through ₁ The norm of L1 is obtained.

6. The muscle force estimation method according to claim 1, wherein the LVN network in step S5 adopts a training method based on a continuous domain ant colony training algorithm, comprising the steps of:

s5.1: initializing LVN network parameters;

s5.2: setting relevant parameters of the continuous domain ant colony algorithm, including solution combination quantity A in the solution archive SOL, generating new solution combination quantity B in each cycle, and calculating selection probability p of A solutions _a The calculation formula is as follows:

β _a the weight of each solution combination is represented, and the calculation formula is as follows:

wherein γ represents a preset adjustment parameter;

s5.3: initializing a solution archive SOL: enabling the solution combination to be a parameter combination of the LVN network, firstly generating A initial solution combinations, then configuring the LVN network based on each solution combination, inputting a training sample into the LVN network to obtain an estimated muscle strength signal, and calculating according to a cost function calculation formula by combining an expected muscle strength signal to obtain a cost function value;

s5.4: let the iteration number τ be 1;

s5.5: randomly generating a sample value in [0,1], if the sample value is smaller than a preset threshold value, entering a step S5.6, otherwise, entering a step S5.7;

S5.6：sorting A solution combinations in the current solution file SOL from small to large according to cost function values, and determining the selection probability p of each solution combination according to the sorting _a (ii) a Then in [0,1]]Randomly generating B sampling values, judging the accumulative selection probability interval corresponding to the solution combination in which each sampling value falls, and combining the corresponding solution combinations into guide solution combinations; and respectively generating a new solution combination according to each guide solution combination, wherein the specific method comprises the following steps:

for each parameter to be estimated of the LVN network, taking the value of the parameter to be estimated in the guide solution combination as a mean value mu, and then calculating by adopting the following formula to obtain a standard deviation sigma:

wherein, SOL _a Representing the value of the parameter to be estimated in the a-th solution combination in the current solution file SOL, and ξ representing a preset forgetting factor;

sampling the Gaussian distribution with the mean value mu and the standard deviation sigma to obtain the value of the parameter to be estimated in the new solution, thereby generating 1 new solution combination;

for the B generated new solution combinations, configuring an LVN network based on each new solution combination, inputting training samples into the LVN network to obtain estimated muscle force signals, and calculating according to a cost function calculation formula by combining the expected muscle force signals to obtain a cost function value of each new solution combination;

s5.7: sequentially taking A solution combinations in the current solution file SOL as guide solution combinations to generate A new solution combinations; then configuring an LVN network based on each new solution combination, inputting the training samples into the LVN network to obtain estimated muscle force signals, and calculating according to a cost function calculation formula by combining the expected muscle force signals to obtain a cost function of each new solution combination;

s5.8: combining A solution combinations in the current solution file SOL and all generated new solution combinations, sorting the solution combinations from small to large according to cost function values, and screening the former A solution combinations to form a new solution file SOL;

s5.9: judging the iteration number tau < tau _max If yes, go to step S5.10, otherwise go to step S5.11;

s5.10: let τ be τ +1, return to step S5.5;

s5.11: and combining the solutions with the optimal cost function values in the current solution files SOL into LVN network parameters, and configuring the LVN network.