CN114931390B

CN114931390B - Muscle strength estimation method based on fatigue analysis

Info

Publication number: CN114931390B
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: 2023-06-13
Anticipated expiration: 2042-05-06
Also published as: CN114931390A

Abstract

The invention discloses a muscle strength estimation method based on fatigue analysis, which comprises the steps of acquiring and preprocessing surface electromyographic signals and muscle strength signals of each sample object under preset actions, dividing the surface electromyographic signals into active sections, extracting fatigue related features in each active section, and constructing fatigue feature signals; constructing a multi-input LVN network, wherein the inputs are respectively surface electromyographic signals and M fatigue characteristic signals, and the signals are output as muscle strength signals; then, taking the surface electromyographic signals and corresponding fatigue characteristic signals of each sample object as inputs, taking the corresponding electromyographic signals as expected outputs, and training a multi-input LVN (Linear variable network); when the muscle force estimation of the same action is required, the surface electromyographic signals are acquired and preprocessed by adopting the same method as the sample object, then 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 force estimation result. The invention combines the fatigue characteristics and the LVN network, and improves the accuracy and the robustness of muscle strength estimation.

Description

Muscle strength estimation method based on fatigue analysis

Technical Field

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

Background

The brain may activate muscles, which contract to produce electrical signals, thereby producing mechanical forces. Any kind of motion of human body, including small motions such as chewing, blinking, etc., and large motions such as running, bouncing, lifting, etc., are required to be realized by corresponding muscle contraction. Different exercises require different muscles, some require only one muscle to participate, and some require multiple muscles to participate together. Muscle strength, i.e., strength of muscle contraction, related studies are of great importance in many applications such as gait analysis, orthopedics, rehabilitation, ergonomic design, haptic technology, teleoperation, and human-machine interaction.

Currently, myoelectricity-myodynamia models are mostly adopted in the industry to perform myodynamia estimation. Myoelectricity-muscle force is a nonlinear and dynamically changing relationship, with the degree of nonlinearity being largely dependent on the manner in which muscle fibers are combined, the time of contraction, and the degree of force when applied, while dynamic relationship is due to the muscle shortening effect and electrical time delay (i.e., the time delay of the myoelectricity signal to the generation). Thus, whether the built myoelectric-myoforce model is reliable depends on whether the model captures dynamic changes of the system and nonlinearities. Besides the difficulty of expressing the dynamics and nonlinearity of the system, the motion mode, the muscle state, the individual variability and the like can influence the muscle strength estimation precision. Muscle fatigue is also one of the important and common influencing factors, but many experimental studies have avoided the problem of muscle fatigue. However, muscle fatigue affects the muscle activation ability, contraction ability, and the dynamic relationship between myoelectric signals and force seriously, which is a major difficulty that is difficult to neglect.

There is currently less research on muscle strength estimation under fatigue. Soo et al propose a band-technique based force estimation model that finds the greater the degree of fatigue, the more pronounced the improvement effect of the model compared to the conventional RMS-force model. Na et al propose a method for estimating force in a fatigue state combining surface myoelectricity and a motor model, and found that as muscle fatigue deepens, the peak value of the motor model decreases and the contraction time increases. The old aroma professor team of Chinese science and technology university corrects the polynomial model and Hill model by utilizing fatigue trend, and eliminates the influence of fatigue on muscle strength prediction to a certain extent. The Lagueri et al uses Lagueri model to train and predict three-section fatigue data respectively, finds that the peak value and the median frequency of Lagueri first-order nuclear coefficient of the three-section data have descending trend, considers that the first-order nuclear peak value and the second-order high-frequency component can be used as fatigue generation indexes, realizes a myoelectricity-myodynamia dynamic model influenced by muscle fatigue under isometric contraction, and identifies the dynamic relationship of myoelectricity-myodynamia. However, the above method has limited effect on fatigue analysis, and it is difficult to achieve substantial improvement of the muscle strength estimation performance in practical application.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a muscle strength estimation method based on fatigue analysis, which improves the accuracy and the robustness of muscle strength estimation by introducing fatigue characteristics and combining with an LVN (Linear virtual network).

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

s1: for K sample objects, acquiring a surface electromyographic signal and a muscle force signal of each sample object under a preset action, and preprocessing according to a preset method to obtain a preprocessed surface electromyographic signal x _k (i) And muscle strength signal y _k (i) Where i=1,..n, N represents the signal length;

s2: by a length L _win Sliding window with sliding step length delta is arranged on surface electromyographic signal x _k (i) Performing sliding extraction on the signal segments, calculating average instantaneous energy of each signal segment, and calculating average instantaneous energy E of jth signal segment _k (j) The calculation formula of (2) is as follows:

wherein j=1, 2, …, N _part ，N _part Representing the number of signal segments obtained by division;

setting an energy threshold

When->

And->

The start of the j+1th signal segment is taken as the start of the active segment when +.>

And->

The end point of the jth signal segment is taken as the end point of the active segment, and the rest conditions do not perform any operation, thereby obtaining the surface electromyographic signal x _k (i) Is divided into movable segments;

surface myoelectric signal x _k (i) The number of the obtained movable segments is D _k The d-th active segment is expressed as

Respectively represent the surface electromyographic signals x _k (i) Original sample point sequence numbers of start and end of the D-th active segment, d=1, 2, …, D _k ；

S3: setting G characteristics related to fatigue in the surface electromyographic signals according to actual needs, and then for each surface electromyographic signal x _k (i) Extracting G features f of each movable segment _k,d,g ，g＝1,2,…,G；

S4: dividing each surface electromyographic signal x according to the activity section _k (i) Respectively constructing G fatigue characteristic signals F thereof _k,g (i) I.e. when sampling points

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

S5: constructing a multi-input LVN network, wherein the inputs are respectively surface electromyographic signals and M fatigue characteristic signals, and the signals are output as muscle strength signals; then the surface electromyographic signals x of each sample object _k (i) And the corresponding fatigue characteristic signals are used as the input of the multi-output LVN, the corresponding electromyographic signals are used as expected output, and the multi-input LVN is trained;

s6: when the muscle force estimation of the same action is required, the surface electromyographic signals are acquired by adopting the same method as the sample object and are preprocessed to obtain the surface electromyographic signals x' (i), G characteristics of each active segment relevant to fatigue are extracted after the active segment is divided, and G fatigue are generated Characteristic signal F' _g (i) Then the surface electromechanical signal x '(i) and G fatigue characteristic signals F' _g (i) And inputting the trained multi-input LVN network to obtain an estimated muscle strength signal.

According to the muscle strength estimation method based on fatigue analysis, surface electromyographic signals and muscle strength signals of each sample object under preset actions are acquired and preprocessed, fatigue-related characteristics in each active segment are calculated after the surface electromyographic signals are subjected to active segment division, and fatigue characteristic signals corresponding to the surface electromyographic signals are constructed; constructing a multi-input LVN network, wherein the inputs are respectively surface electromyographic signals and M fatigue characteristic signals, and the signals are output as muscle strength signals; then taking the surface electromyographic signals and the corresponding fatigue characteristic signals of each sample object 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 the muscle force estimation of the same action is required, the surface electromyographic signals are acquired and preprocessed by adopting the same method as the sample object, then 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 force estimation result.

The invention has the following technical effects:

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

2) When the characteristics with stronger correlation with the fatigue degree are selected from the alternative characteristics and used as fatigue-related characteristics, the characteristic description of the surface electromyographic signals is more accurate, so that the accuracy of muscle strength estimation is improved;

3) According to the invention, the specific structure of the LVN is researched, and an optimization mode of fusing fatigue characteristic signals and punishing local sparsity of the fusion weight of the fatigue characteristic signals is provided, so that the contribution of MDF and MPF characteristics with larger correction effect to a model is highlighted while the characteristic weight can be adaptively adjusted, and the accuracy of muscle strength estimation is further improved;

4) The training method of the LVN adopts the continuous domain ant colony algorithm, reduces the training time, improves the global optimizing performance, and ensures that the LVN obtained by training has better performance so as to ensure accurate muscle strength estimation.

Drawings

FIG. 1 is a flowchart of an embodiment of a method for estimating muscle strength based on fatigue analysis according to the present invention;

Fig. 2 is a flowchart of preprocessing of surface electromyographic signals in the present embodiment;

FIG. 3 is a flow chart of preprocessing of the muscle force signal in the present embodiment;

FIG. 4 is a flow chart of determining fatigue related features in the present embodiment;

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

fig. 6 is a flowchart of a multi-input LVN network training based on a continuous domain ant colony training algorithm in this embodiment;

FIG. 7 is a graph of muscle strength estimation metrics for five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparative methods;

FIG. 8 is a graph of test index significance analysis versus five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparison methods;

FIG. 9 is a graph of muscle strength estimation metrics for five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparative methods;

FIG. 10 is a graph of test index significance analysis versus five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparison methods;

FIG. 11 is a graph of the test myoelectric signal of subject 4 in the brachiocephalic lift force pulse experiment in this example;

FIG. 12 is a graph of muscle strength estimates for different fatigue stages in a brachiocephalic lift force pulse test for the present invention and four comparative methods;

FIG. 13 is a graph of muscle strength estimation metrics for five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparative methods;

FIG. 14 is a graph of test index significance analysis versus five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparison methods;

FIG. 15 is a graph of five stage index comparisons of lift data for four sets of brachioradial muscles for the present invention and four comparative methods.

Detailed Description

The following description of the embodiments of the invention is presented in conjunction with the accompanying drawings to provide a better understanding of the invention to those skilled in the art. It is to be expressly noted that in the description below, detailed descriptions of known functions and designs are omitted here as perhaps obscuring the present invention.

Examples

FIG. 1 is a flowchart of an embodiment of a method for estimating muscle strength based on fatigue analysis according to the present invention. As shown in fig. 1, the method for estimating muscle strength based on fatigue analysis of the present invention comprises the following specific steps:

s101: collecting a sample signal:

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

For example, when the force mode of the preset motion is a constant force pulse or a lifting force pulse, the surface electromyographic signal may be selected from a brachioradial surface electromyographic signal or a palmaris longus surface electromyographic signal, and the muscle force signal may be acquired by a grip sensor.

Fig. 2 is a flowchart of preprocessing of the surface electromyographic signals in the present embodiment. As shown in fig. 2, the specific steps of the preprocessing of the surface electromyographic signals in this embodiment include:

s201: denoising:

since the surface electromyographic signals may have motion artifacts and unwanted components, denoising the surface electromyographic signals is required.

The surface electromyographic signals are filtered using a 15Hz to 400Hz band pass filter in this embodiment to remove noise.

S202: sliding average:

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

S203: normalization:

to eliminate the variability of data between different subjects and different sets of data for the same subject, the surface electromyographic signals after the moving average need to be normalized to zero mean and unit variance.

S204: downsampling:

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

S205: correction smoothing:

research shows that the surface electromyographic signal obeys the mean value of 0 and the variance of sigma ² And the Gaussian distribution is adopted, so that the surface electromyographic signals after the downsampling can be processed by adopting a correction smoothing method, and the surface electromyographic signals after the correction smoothing can be obtained. Specific principles and methods of modified smoothing can be referred to in the literature "Hayashi H, furui A, kurita Y, et al A variance distribution model of surface emg signals based on inverse gamma 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 force signal in the present embodiment. As shown in fig. 3, the specific steps of preprocessing the muscle force signal in this embodiment include:

s301: sliding average:

in this embodiment, since the muscle force signal acquisition device is provided with the filtering processing device, the acquired muscle force signal is directly subjected to the moving average.

S302: normalization:

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

S303: downsampling:

and similarly, the normalized muscle force signal is downsampled according to a preset frequency by adopting the same downsampling frequency of the surface electromyographic signal.

S102: surface electromyographic signal activity segment analysis:

Since the muscle activity is not continuous, there is an active segment and a rest segment of the surface electromyographic signal, and no analysis is necessary for the rest segment signal, so that the individual surface electromyographic signals x need to be first of all _k (i) The method for analyzing the activity section comprises the following steps:

by a length L _win Sliding window with sliding step length delta is arranged on surface electromyographic signal x _k (i) Performing sliding extraction on the signal segments, calculating average instantaneous energy of each signal segment, and calculating average instantaneous energy E of jth signal segment _k (j) The calculation formula of (2) is as follows:

wherein j=1, 2, …, N _part ，N _part Representing the number of signal segments divided.

Setting an energy threshold

When->

And->

And->

The end point of the jth signal segment is taken as the end point of the active segment, and the rest conditions do not perform any operation, thereby obtaining the surface electromyographic signal x _k (i) Is divided into active segments.

Surface myoelectric signal x _k (i) The resulting activityThe number of segments is D _k The d-th active segment is expressed as

Respectively represent the surface electromyographic signals x _k (i) Original sample point sequence numbers of start and end of the D-th active segment, d=1, 2, …, D _k 。

S103: extracting fatigue related characteristics of myoelectricity:

setting G characteristics related to fatigue in the surface electromyographic signals according to actual needs, and then for each surface electromyographic signal x _k (i) Extracting G features f of each movable segment _k,d,g ，g＝1,2,…,G。

Generally, characteristics of the surface electromyographic signals are classified into time domain characteristics, such as time domain characteristics including integrated electromyographic values, amplitude absolute value averages, slopes of amplitude absolute value averages, root mean square, waveform length, differences of signal absolute averages, sample variances, zero crossing rates, wilson amplitudes, simple square accumulation, slope sign changes, and the like, frequency domain characteristics including peak frequencies, average frequencies, total powers, median frequencies, average instantaneous frequencies, spectral moment parameters, and the like, and non-gaussian characteristics including kurtosis, negative entropy, and the like. In practical application, the fatigue-related characteristics can be screened from the characteristics by technical means such as experience, qualitative analysis, quantitative analysis and the like.

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

s401: determining alternative features:

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

S402: extracting alternative features:

for each surface electromyographic signal x _k (i) Extracting G' features f of each movable segment _k,d,g′ G '=1, 2, …, G', resulting in each surface electromyographic signal x _k (i) Characteristic sequence corresponding to g' th characteristic in (3)

S403: obtaining a fatigue value sequence:

because the longer the action time is, the larger the fatigue value is, in this embodiment, the sequence number of the movable segment is used as the fatigue value and normalized to obtain the electromyographic signals x of each surface _k (i) Fatigue value sequence o= {1/D _k ,2/D _k ,…,1}。

S404: calculating a determination coefficient:

for each surface electromyographic signal x _k (i) The fatigue value is used as an independent variable and the characteristic value is used as a dependent variable, and the fatigue value is respectively used for each characteristic sequence

And fatigue value sequence o= {1/D _k ,2/D _k Linear regression of …,1 to obtain the decision coefficient +.>

Determining coefficient->

The linear correlation of the g' th feature with the fatigue level is expressed, and the degree of the feature changing with the fatigue level can be measured. Determining coefficient->

The larger the ratio, the higher the ratio of the g' th feature affected by the fatigue degree. And vice versa.

For K surface electromyographic signals x _k (i) Determination coefficient of g' th feature in (3)

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

S405: calculating the pearson correlation coefficient:

for each surface electromyographic signal x _k (i) Each characteristic sequence is calculated separately

And fatigue value sequence o= {1/D _k ,2/D _k Pearson correlation coefficient P of …,1} _k,g′ . Then to K surface electromyographic signals x _k (i) Pearson correlation coefficient P of g' th feature in (b) _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 conversely, the lower.

S406: calculating SVR coefficients:

since some myoelectric characteristics are nonlinear in change, and the decision coefficient and the pearson correlation coefficient can only measure the linear correlation of the characteristics, the SVR (sensitivity to variability) coefficient proposed in the literature "Rogers D R, macIsaac D T.A comparison of emg-based muscle fatigue assessments during dynamic contra" is also introduced in this embodiment, so as to evaluate the characteristics with the change of the fatigue degree more accurately when the fatigue trend is nonlinear. 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

And fatigue value sequence o= {1/D _k ,2/D _k ,…,1} performing second order fitting, wherein a fitting polynomial is as follows:

wherein a is ₀ ,a ₁ ,a ₂ For coefficients to be fitted, o _d The d-th fatigue value in the fatigue value sequence O is represented.

Then, a polynomial obtained by fitting is adopted to calculate each characteristic f _k,d,g′ Fitting characteristic values of (a)

The surface electromyographic signal x is calculated by adopting the following formula _k (i) SVR coefficient SVR of g' th feature in (3) _k,g′ ：

/>

Wherein max _k,g′ 、min _k,g′ Representing the surface electromyographic signal x _k (i) Maximum and minimum of the g' th feature in all active segments.

For K surface electromyographic signals x _k (i) SVR coefficient SVR of g' th feature in (3) _k,g′ Averaging to obtain SVR coefficient of g' th feature

S407: screening fatigue-related features:

screening SVR coefficients for G' optional features

Greater than a preset threshold T _SVR Determining coefficient->

Greater than a preset threshold->

And the pearson correlation coefficient->

Greater than a preset threshold T _P As fatigue-related features.

Through the three parameters, the characteristics with larger relevance to the fatigue degree and more sensitive 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 selected by the above screening method and combining qualitative analysis and experience include 11 features of integrated myoelectricity value (Integration of absolute of EMG signal, IEMG), amplitude absolute value Mean (Mean Absolute Value, MAV), root Mean Square (RMS), difference of signal absolute average (Difference absolute Mean value of signal, DAMV), variance (Variance of signal, VAR), simple Square accumulation (Simple Square integral of signal, SSI), total Power (Total Power, TP), average Frequency (Mean Spectral Frequency, MPF), median Frequency (MDF), negative entropy NEGEN, and 5-order spectral distance parameter FInsm 5.

S104: and (3) constructing a characteristic signal:

dividing each surface electromyographic signal x according to the activity section _k (i) Respectively constructing G fatigue characteristic signals F thereof _k,g (i) I.e. when sampling points

Fatigue characteristic signal F corresponding to the mth characteristic _k,g (i)＝f _k,d,g Otherwise F _k,g (i) =0. That is, the characteristic signals corresponding to the active segment are filled by the characteristic values calculated by the active segment, and the characteristic signals corresponding to the rest segment are 0, so that the characteristic signals only contain myoelectric characteristics during action, and the surface myoelectric signals are more accurately represented. And the characteristic signal and the surface electromyographic signal x thus obtained _k (i) Equal length.

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

LVN (Laguerre-Volterra Network, laguerre-Volterra) Network is a Network model constructed based on Volterra model (Volterra Mode), laguerre extension (Laguerre Expasion Technique, LET), and standard dynamic basis (Principal Dynamic Modes, PDM), the specific principle of which can be referred to in the literature "Geng K, mararelis V Z.methodology of recurrent Laguerre-Volterra Network for modeling nonlinear dynamic systems [ J ]]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 multi-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) the outputs from the respective filter banks are

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

Is expressed as follows:

wherein n=1,..n; j (j) _n ＝0,...,L _n -1，L _n The number of Laguerre functions input for the nth group; t=1, a.i., T, T is the length of the data point; m is memory length;

representing the j-th Laguerre function, the expression is as follows:

wherein, the liquid crystal display device comprises a liquid crystal display device,m=0..m-1, α is the decay exponent, determining the convergence of the lager function.

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

Second layer output z _h (t) is defined as follows:

where h=1,..h, H represents the number of standard dynamic bases. c _q,h Representing a non-linear function f _h The coefficient of (-),

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

Substitution output

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

thus, the h standard dynamic base corresponding to the n-th input can be expressed as:

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

the signals used for assessing muscle strength in the present invention include and are extracted from surface electromyographic signalsThe obtained fatigue characteristic signals are input into the multi-input LVN network, so that the input is a surface electromyographic signal and M fatigue characteristic signals respectively, and the output is a muscle strength signal. Then the surface electromyographic signals x of each sample object _k (i) And the corresponding fatigue characteristic signals are used as the input of the multi-output LVN network, the corresponding electromyographic signals are used as expected outputs, and the multi-input LVN network is trained.

In practical application, in order to simplify the structure of the LVN network and improve the muscle strength estimation performance of the LVN network, one preferred mode is to set the LVN network as a two-input LVN network, wherein one input is a surface electromyographic signal, and the other input is according to a preset weight w _m And the M characteristic signals are weighted and summed to obtain a fusion characteristic signal, so that the number of input signals is reduced under the condition that the number of the characteristics is not reduced, the complexity of an LVN network is reduced, and the muscle strength estimation efficiency is improved. During training of the two-input LVN network, the weight w can be calculated _m As parameters to be trained in order to make the fused characteristic signals more effectively characterize the surface electromyographic signals.

In multi-input LVN network training, a normalized root mean square error (Normalized Mean Square Error, NMSE) may be employed as a cost function. When the mode that M characteristic signals are fused into one path of fused characteristic signals is adopted, sparsity penalty for the characteristic signal weight can be added into the cost function, so that training effect is improved. Namely, the calculation formula of the cost function J of the two-input LVN network under the scheme is as follows:

J＝NMSE+λ||W|| ₁

Wherein, NMSE represents the root mean square error of estimated muscle force signal and true muscle force signal, lambda represents the punishment intensity coefficient that presets, W represents punishment weight vector, its constitution method is: is set according to the requirement in G characteristic signals

The individual characteristic signal is taken as a penalty characteristic signal, < >>

Will punish eachWeight construction corresponding to penalty characteristic signals to obtain +.>

A row vector of the dimension, namely a penalty weight vector W; I.I ₁ The L1 norm is obtained.

According to the above description, when the sparsity penalty is performed on the feature signal weights, the sparsity penalty may be performed on all the feature signal weights, or only a part of the feature signal weights may be selected to perform the local sparsity penalty, and in practical application, the sparsity penalty may be selected according to practical situations. In the embodiment, through experimental comparison, the screened characteristic signals of 11 characteristics are fused to be used as one path of input, and then local sparsity penalty is carried out on characteristic weights except for average frequency MPF and median frequency MDF in a cost function, so that the performance of the obtained LVN is better. For ease of description, this LVN network is referred to as a LVN-pS network.

For a specific training method, the method can be selected 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 method, the continuous domain ant colony training algorithm can obtain an approximate global optimal solution in a short training time so as to reduce training expenditure and improve training efficiency. Fig. 6 is a flowchart of a multi-input LVN network training based on the continuous domain ant colony training algorithm in this 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 Laguerre radix L of N-way input _n N=1, 2, …, N, standard dynamic base number H, model order Q, etc.

Taking LVN-pS network as an example, the input of the LVN-pS network is 2 paths, and the coefficient to be estimated comprises the attenuation index alpha of 2 paths of input Laguerre coefficients ₁ ,α ₂ Output signal offset constant y ₀ From coefficient { c _q,h Q=1,..q, Q, h=1,..h, H } and a matrix C of Q rows and H columns consisting of coefficients

L of composition ₁ Matrix W1 of row H and column, composed of coefficients

L of composition ₂ Matrix W2 of row H and column, weight vector W composed of weights at 11 eigenvectors fusion _fea The total number of coefficients to be estimated np=3+q×h+l ₁ *H+L ₂ *H+11。

S602: initializing parameters of a continuous domain ant colony algorithm:

setting related parameters of a continuous domain ant colony algorithm, including the number A of solution combinations in a solution file SOL, generating a new number B of solution combinations each time, 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 as follows:

wherein, gamma represents a preset adjustment parameter. The smaller the gamma is, the larger the solution weight which is ranked at the front is, the faster the convergence is, otherwise, the more uniform the weight distribution is, the stronger the global optimizing capability is provided for the algorithm, and the slower the convergence is.

S603: initializing a solution file:

the solution profile SOL is a matrix of rows np+1 columns, each row including a solution combination and a cost function value corresponding to the solution combination. When initializing a solution file SOL, the solution combination is made into a parameter combination of the LVN network, A initial solution combinations are firstly generated, then the LVN network is configured based on each solution combination, a training sample is input into the LVN network to obtain an estimated muscle force signal, and a cost function value is calculated according to a cost function calculation formula by combining an expected muscle force signal.

With LVN-pS network asFor example, its solution combination may be expressed as { α } ₁ ,α ₂ ,y ₀ ,C,W1,W2,W _fea }. Then alpha is at the time of initializing the solution profile in this embodiment ₁ ,α ₂ At [0,1]Uniformly taking values in the range, and the rest parameters are within the range of < -1,1 [ -1]And uniformly taking values in the range. Then according to alpha ₁ And the 1 st input in the training sample to obtain the corresponding V ₁ According to the weight W _fea And fatigue characteristic signals in training samples, calculating the 2 nd input, and then combining alpha ₂ Calculating to obtain corresponding V ₂ Then calculate the intermediate output u=v of the LVN-pS network ₁ *W ₁ +V ₂ *W ₂ ,U2＝U*U ^T Finally, calculating to obtain an estimated muscle strength signal y of the LVN-pS network _e ＝U*C(1,:) ^T +U2*C(2,:) ^T . And according to the estimated muscle force signal and the expected muscle force signal of the training sample, calculating by adopting a cost function calculation formula based on local sparsity penalty to obtain a cost function value corresponding to the solution combination.

S604: let iteration number τ=1.

S605: generating a random sample value in the range of [0,1], if the random sample value is smaller than the preset threshold, proceeding to step S606, otherwise proceeding to step S607.

S605: screening the guided solution combinations and generating new solution combinations:

sorting A solution combinations in the current solution archive SOL according to the cost function value from small to large, and determining the selection probability p of each solution combination according to the sorting _a . Then at [0,1]And randomly generating B sampling values, judging the accumulated selection probability interval corresponding to the solution combination which the sampling values fall into for each sampling value, and combining the corresponding solution combination into a guide solution combination. 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, taking the value of the parameter to be estimated in the guide solution combination as a mean value mu, and then calculating to obtain a standard deviation sigma by adopting the following formula:

wherein SOL _a And the value of the parameter to be estimated in the a-th solution combination in the current solution archive SOL is represented, and the value of the value is larger, so that the slower the convergence, the slower the worse the solution is forgotten.

And sampling the Gaussian distribution with the mean value of mu and the standard deviation of 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 the LVN based on each new solution combination, inputting a training sample into the LVN to obtain an estimated muscle force signal, and calculating according to a cost function calculation formula by combining the expected muscle force signal 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 the LVN based on each new solution combination, inputting a training sample into the LVN to obtain an estimated muscle force signal, and calculating a cost function of each new solution combination according to a cost function calculation formula by combining the expected muscle force signal. The overall solution combination is probabilistically adopted as a guide solution to generate a new solution combination, so that the global optimizing performance can be improved.

S608: generating a new solution file:

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

S609: judging the iteration times tau < tau _max If yes, step S610 is entered, otherwise step S611 is entered.

S610: let τ=τ+1, return to step S605.

S611: determining LVN network parameters:

and combining the solution with the optimal cost function value in the current solution file SOL into LVN network parameters, and configuring the LVN network.

S107: estimating muscle strength:

when it is neededWhen the muscle strength estimation of the same action is carried out, the surface electromyographic signals are acquired by adopting the same method as the sample object and are preprocessed to obtain the surface electromyographic signals x ' (i), G characteristics of each movable section relevant to fatigue are extracted after the movable section is divided, and G fatigue characteristic signals F ' are generated ' _g (i) Then the surface electromyographic signals x '(i) and G fatigue characteristic signals F' _g (i) And inputting the trained multi-input LVN network to obtain an estimated muscle strength signal.

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

In the experimental verification, the surface electromyographic signals and grip strength signals of the sample object are collected. The experimental equipment mainly comprises a host MP150, an myoelectricity acquisition module EMG100C, an analog signal acquisition device DA100C and a holding force sensor TSD121C of BIOPAC company in America. Subjects included 9 healthy right-handicapped subjects (6 of young men, 3 of young women, all 25 years old). The experiments comprise a maximum autonomous contractile force experiment and a constant force pulse experiment of the brachial radial muscle and the palmaris longus in the process, and the specific conditions are as follows:

1) Maximum autonomous contractile force experiment: firstly, the experimental object needs to implement the maximum autonomous contractile force (Maximum Voluntary Contraction, MVC) twice, each time the force is exerted to hold the maximum force of the experimental object as much as possible, the time lasts for 4 to 5 seconds, and the interval between the two times is more than 2 minutes, so that the muscle fatigue is avoided. MVC takes the average of the two experimental forces.

2) Constant force pulse experiment: after the maximum autonomous contractility experiment is finished, the full rest is needed for more than 5 minutes. The subjects began a constant force pulse experiment and a pulse force of 60% mvc was applied. The force-generating mode is as follows: force is applied for 5s, rest for 5s and cycle and reciprocation are carried out until the experimental object can not apply force to 60% MVC continuously twice.

The experiments are one group, and each subject needs to repeat two groups of experiments with more than 24 hours between the two groups due to the requirement of model evaluation (one group is used for training and one group is used for testing), so that the fatigue feeling of the previous experiment is completely eliminated.

In order to illustrate the effectiveness of the fatigue characteristic signals introduced in the invention, 6 LVN networks are set for comparison in the experiment, and the 6 LVN networks are respectively as follows:

lvn_1: the LVN network uses a single input LVN network with only surface electromyographic signals.

Lvn_2MPF: the LVN network adopts a two-input LVN network, one input is a surface electromyographic signal, and the other input is a fatigue characteristic signal corresponding to an average frequency (MPF) characteristic.

Lvn_2MDF: the LVN network adopts a two-input LVN network, one input is a surface electromyographic signal, and the other input is a fatigue characteristic signal corresponding to the median frequency (MDF) characteristic.

Lvn_3: the LVN network adopts a three-input LVN network, one input is a surface electromyographic signal, and the other two inputs are fatigue characteristic signals corresponding to average frequency (MPF) characteristics and median frequency (MDF) characteristics respectively.

The cost functions of the above 4 LVN networks all use normalized root mean square error NMSE.

Lvn_s: the LVN network adopts a two-input LVN network, one input is a surface electromyographic signal, and the other input is a fused characteristic signal obtained by fusing fatigue characteristic signals corresponding to 11 characteristics screened in the embodiment. The cost function of the network adopts a cost function fused with NMSE and sparsity penalty, and takes 11 fatigue characteristic signals as penalty characteristic signals.

Lvn_ps: the LVN network adopts a two-input LVN network, one input is a surface electromyographic signal, and the other input is a fused characteristic signal obtained by fusing fatigue characteristic signals corresponding to 11 characteristics screened in the embodiment. The cost function of the network adopts a cost function fused with NMSE and sparsity penalty, and takes 9 fatigue characteristic signals except average frequency (MPF) characteristic and median frequency (MDF) as penalty characteristic signals.

The normalized root mean square error (NMSE%) versus Fitness (Fitness%) for the 6 LVN networks was statistically compared. Table 1 is a comparative table of training and testing metrics for the 6 LVN networks of this example.

TABLE 1

As shown in table 1, the performance of the 5 LVN networks with the fatigue characteristic signals is superior to that of the LVN network with the surface electromyographic signals as input, wherein lvn_ps is the most, and the normalized root mean square error and the fitness of the predicted muscle force and the true muscle force are optimal on the network training result and the network testing result.

Next, the lvn_ps model was taken as a representative network structure of the present invention, and the muscle strength estimation method based on four models of POL (polynominal Polynomial simulation model), FOS (Fast Orthogonal Search, fast orthogonal search model), PCI (Parallel Cascade Identification, parallel cascade model), LET (Laguerre Expasion Technique, lager expansion technique) was taken as a comparison method, to compare the performance of the present invention. The technical indicators for comparison include mean square error MSE%, normalized root mean square error NMSE% and Fitness Fitness%. The model can show excellent muscle strength estimation performance under different muscles and different stress modes by designing a brachiocephalic constant force fatigue pulse experiment, a brachiocephalic lifting force fatigue pulse experiment and a palmar longus lifting force fatigue pulse experiment. In order to show the muscle strength estimation conditions of the model in different fatigue stages in various data, the test data are divided into five stages of a non-fatigue state seg1, a slight fatigue state seg2, a medium fatigue state seg3, a serious fatigue state seg4 and an extreme fatigue state seg5, and the trained model is utilized for testing in each stage.

Experimental results based on brachiocephalic constant force fatigue pulse data

FIG. 7 is a graph of muscle strength estimation metrics for five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparative methods. Fig. 7 shows the mean and standard deviation of three technical indices for muscle strength estimation on five phases of data of different degrees of fatigue of the brachiocephalic muscle, with the abscissa representing the five fatigue phases and the ordinate representing the corresponding index. As can be seen from fig. 7, the five methods all show the situation that MSE% and NMSE% decrease and Fitness% increase with the progressive fatigue, and it is assumed that the models used in the five methods capture more myoelectricity-myoelectricity relationships when the fatigue is heavy during training of the full pulse data, so that the test data has higher estimation accuracy for the three phases with higher fatigue degrees, namely, the medium fatigue phase seg3, the severe fatigue phase seg and the extreme fatigue phase seg 5. In the five methods, MSE% and NMSE% of the LVN_pS model in five different fatigue stages are obviously lower than those of the other four models, and the muscle force Fitness Fitness% is obviously higher than that of the other models, which indicates that the LVN_pS model has stronger capturing capability on myoelectricity-muscle force relationship in the stages of the muscles in five fatigue states, and shows excellent muscle force estimation performance. The lvn_ps model was calculated to reach 5.25%, 2.52% and 77.68% for the average test MSE, NMSE, fitness% for the five phases, respectively.

To further illustrate whether there were significant differences between the lvn_ps model of the invention and the POL, FOS, LET model at five different fatigue stages, a significance analysis was performed. FIG. 8 is a graph of test index significance analysis versus five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparison methods.

Fig. 8 (a), (b), and (c) show the results of the significance analysis between the five models under the three evaluation indexes of MSE, NMSE%, and Fitness%, respectively, from left to right, the results of the five fatigue stages seg1, seg2, seg3, seg4, and seg5, respectively. The middle gray NS indicates that p >0.05 between the two models corresponding to the abscissa and the ordinate has no significant difference. Light grey indicates p <0.05, dark grey indicates p <0.01, all indicate significant differences between models, and the differences in dark grey are more pronounced. In combination with the optimal lvn_ps three indices in fig. 6, 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 at a level of p <0.05 under the MSE% index, the lvn_ps model of the present invention is significantly better than the POL, FOS, PCI and LET four models at a level of p <0.01 for the three index variability analysis of five different fatigue phases. However, there was no significant difference between the POL, FOS, PCI and LET models, indicating that the muscle strength estimation effect between the four models was similar.

Experimental results based on the fatigue pulse of the lifting force of the brachiocephalus muscle

Performance comparisons of the five methods over the brachiocephalic lift five different fatigue level stage data are then shown. FIG. 9 is a graph of muscle strength estimation metrics for five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparative methods. As shown in fig. 9, as the fatigue levels of seg1 to seg5 are increased, the indexes MSE% and NMSE% of the five methods show a decreasing trend, and the index Fitness shows a gradually increasing trend. That is, as the fatigue of the subject increases, the effect of estimating the muscle strength tends to increase gradually. This may occur because model training is based on a set of full pulse data that includes five fatigue phases, and the model may capture more information of the more tiring and heavier phases during the training process. Compared with POL, FOS, PCI and LET models, the MSE% and NMSE% of LVN_pS are obviously reduced, and the muscle strength Fitness Fitness is obviously improved, namely, the MSE% and NMSE% are estimated in a non-fatigue stage seg1, a mild fatigue stage seg2, a moderate fatigue stage seg3, a severe fatigue stage seg4 and an extreme fatigue stage seg5, which have lower muscle strength than the rest four models, and the muscle strength Fitness Fitness is higher.

FIG. 10 is a graph of test index significance analysis versus five fatigue stages under brachial radial muscle constant force fatigue pulse data for 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 indexes of the first stage are 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 different degrees of significant difference 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 p <0.01, and the Fitness% of the lvn_ps model is better than the POL model by p < 0.01. The lvn_ps model was calculated to achieve 7.5%, 3.26% and 73.32% for the average test MSE, NMSE, fitness% for the five phases.

Thus, except that the lvn_ps model of the present invention is similar to the LET model in terms of performance in estimating muscle force on the non-tired phase seg1 of the humeral radial lift data, the lvn_ps of the present invention is significantly better than the POL, FOS, PCI and LET models in the mild, moderate, severe and extreme tired phases seg2, seg3, seg4, seg 5. Moreover, the LVN_pS model has the most obvious performance improvement compared with the FOS model.

To demonstrate the muscle force estimation effect of the five methods on the different fatigue phases of the brachioradial muscle lift force pulse, experiments were performed on the muscle force estimation performance of the subject 4 at the different fatigue phases. Fig. 11 is a graph of the test myoelectric signal of subject 4 in the brachiocephalic lift pulse experiment in this example. As shown in fig. 11, the myoelectric signal of the test data of the subject 4 has a tendency to gradually rise from seg1 to seg4, and then has a tendency to slightly decrease in myoelectric amplitude to seg 5. FIG. 12 is a graph of muscle strength estimates for different fatigue stages in the brachiocephalic lift pulse test for the present invention and four comparative methods. As shown in FIG. 12, it can be seen that the LVN_pS model of the present invention had less bias in the five methods, but was superior to the remaining models in the five stages of different fatigue levels.

In summary, for the above-mentioned three methods, the lvn_ps model of the present invention shows the best training and testing effects for the above-mentioned three methods, and the muscle strength estimation effects for the above-mentioned three methods are significantly better than those of POL, FOS, PCI and LET models.

Results based on palmaris longus lift fatigue pulse data

FIG. 13 is a graph of muscle strength estimation metrics for five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparative methods. As shown in fig. 13, three indices of the lvn_ps model of the present invention were significantly lower than the other four models in five different fatigue phases, and all indices of the other four models were overlapped to some extent. All five models have the trend that MSE percent and NMSE percent are reduced and Fitness is increased along with the increase of fatigue degree, and the trend of four models except the LVN_pS model is more obvious. Wherein, the LET and PCI models have a tendency of gradually deteriorating three indexes in the severe fatigue stage seg4 and the extreme fatigue stage seg 5. The average test MSE, NMSE% and Fitness% of the LVN_pS model of the present invention at five stages was calculated to be 9.08%, 3.79% and 70.35%.

FIG. 14 is a graph of test index significance analysis versus five fatigue stages under brachial radial muscle constant force fatigue pulse data for the present invention and four comparison methods. As shown in fig. 14, for the three evaluation indexes, there is no significant difference in the estimated muscle strength performance between POL, FOS, PCI and the LET model under the data verification of five different fatigue stages, but the estimated muscle strength performance of the lvn_ps model of the present invention in the five different fatigue stages is significantly different from that of the POL and FOS models, i.e., the estimated muscle strength performance of the lvn_ps model of the present invention in the different stages is significantly better than that of the POL and FOS models. However, the test MSE, NMSE% and three indexes of the LVN_pS model in the non-fatigue stage seg1 and the extreme fatigue stage seg5 are not significantly different from those of the PCI and LET models, and the standard deviation of the PCI and the LET in the seg1 and the seg5 is larger in combination with the standard deviation in fig. 12, so that the range of the LVN_pS model is covered, and the estimated performance of the PCI, the LET and the LVN_pS model in the seg1 and the seg5 for muscle strength of a certain object is relatively similar, so that significant differences are lost. The three models have similar muscle strength estimation performance in the non-fatigue stage seg1 and the extreme fatigue stage seg5, because the palmaris longus information of individual subjects is weak, and fatigue information can not be provided for the LVN_pS model for reference through myoelectric fatigue characteristics. Although there is an individual variability problem, the average performance of the lvn_ps model of the present invention is still better than that of the PCI and LET models.

In summary, for the fatigue pulse data of the lift force of the palmaris longus, although the lvn_ps model of the present invention has no significant difference from the PCI and LET under the indexes of MSE and NMSE%, the average test performance is still better than the two models. In the five fatigue data stages with different degrees, the LVN_pS model has no obvious difference between the test effect of the seg1 and the test effect of the seg5 and the test effect of the PCI and LET models, but the average muscle strength estimation performance of the LVN_pS model is still superior to that of the two models. Therefore, under the palm long muscle lifting force fatigue data, the test muscle force estimation performance of the LVN_pS model in the full pulse training test and different fatigue degree stages is obviously superior to that of the POL and FOS models, and the LVN_pS model is also obviously superior to that of the PCI and LET models in terms of average muscle force performance.

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

And then, performing muscle strength estimation performance verification of different groups of data among the same subjects based on fatigue data of the same humerus radial muscle lifting force experiment of two subjects with different sexes on different five days. For two-bit subjects, five sets of experimental data five days after each other were named D1, D2, D3, D4, and D5 data. Each set of data includes myoelectric and muscle strength signals. D1 data was used for training and the last four sets of data were used for testing. The training and testing results from the full pulse and different fatigue phases are then shown.

To analyze the muscle force estimation performance of each model in the four sets of test data for each of five different fatigue phases (seg 1-seg 5), the model pair D2-D5 trained on the D1 data described above was tested for five different phases. FIG. 15 is a graph of five stage index comparisons of lift data for four sets of brachioradial muscles for the present invention and four comparative methods. Fig. 15 (a), (b), (c) and (D) show the test results of the data sets D2, D3, D4 and D5, respectively, and the coordinate ranges of the same index in different charts are set to be consistent and convenient for comparison. Therefore, in the LVN_pS model, in three groups of data D2, D3 and D5, five stages from the non-fatigue stage seg1, the light fatigue stage seg2 and the extreme fatigue stage seg5 have the phenomena that MSE and NMSE% are obviously reduced and Fitness% is obviously improved compared with the other four models. And in the D4 data, the other four models perform better.

In order to comprehensively evaluate the muscle strength estimation performance of the five models for different fatigue stages, the mean value and 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 the groups D2 to D4.

TABLE 2

As shown in Table 2, the LVN_pS model of the present invention had 9.6% and 4.15% reduction in average tested MSE% and NMSE% for the four data sets at different fatigue stages, 13.05% improvement in Fitness% compared to the optimal LET for the remaining models, 12.11% reduction in MSE% and NMSE% and 13.05% improvement in Fitness% compared to the FOS model with the worst muscle strength estimation performance in the remaining models.

That is, for multiple sets of brachioradial muscle lift force fatigue data of the same subject, the lvn_ps model of the present invention exhibits more stable average muscle force estimation performance than the remaining four models, and the average muscle force estimation performance is superior to the remaining four models at five different fatigue level stages.

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

While the foregoing describes illustrative embodiments of the present invention to facilitate an 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, but is to be construed as protected by the accompanying claims insofar as various changes are within the spirit and scope of the present invention as defined and defined by the appended claims.

Claims

1. The muscle strength estimation method based on fatigue analysis is characterized by comprising the following steps of:

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

setting an energy threshold

When->

And->

And->

S3: setting the surface electromyographic signals to be related to fatigue according to actual needsThen for each surface electromyographic signal x _k (i) Extracting G features f of each movable segment _k,d,g ，g＝1,2,…,G；

S5: constructing a multi-input LVN network, wherein the inputs are respectively surface electromyographic signals and M fatigue characteristic signals, and the signals are output as muscle strength signals; then the surface electromyographic signals x of each sample object _k (i) And the corresponding fatigue characteristic signals are used as the input of the multi-output LVN, the corresponding electromyographic signals are used as expected output, and the multi-input LVN is trained; the LVN adopts a training method based on a continuous domain ant colony training algorithm, and comprises the following steps:

s5.1: initializing LVN network parameters;

s5.2: setting related parameters of a continuous domain ant colony algorithm, including the number A of solution combinations in a solution file SOL, generating a new number B of solution combinations each time, 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 as follows:

Wherein, gamma represents a preset adjustment parameter;

s5.3: initializing a solution file SOL: the solution combination is made into a parameter combination of the LVN, A initial solution combinations are firstly generated, then the LVN is configured based on each solution combination, a training sample is input into the LVN to obtain an estimated muscle force signal, and a cost function value is calculated according to a cost function calculation formula by combining an expected muscle force signal;

s5.4: let iteration number τ=1;

s5.5: randomly generating a sampling value in [0,1], if the sampling 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 archive SOL according to the cost function value from small to large, and determining the selection probability p of each solution combination according to the sorting _a The method comprises the steps of carrying out a first treatment on the surface of the Then at [0,1]Randomly generating B sampling values, judging the cumulative selection probability interval corresponding to the solution combination which the sampling values fall into for each sampling value, and combining the corresponding solution combination into a guide solution combination; generating a new solution combination according to each guide solution combination, wherein the specific method comprises the following steps:

Wherein SOL _a Representing the value of the parameter to be estimated in the a-th solution combination in the current solution archive SOL, and xi represents a preset forgetting factor;

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

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

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

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

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

s5.11: combining the solution with the optimal cost function value in the current solution file SOL into LVN network parameters, and configuring the LVN network;

s6: when the muscle force estimation of the same action is required, the surface electromyographic signals are acquired by adopting the same method as the sample object and are preprocessed to obtain the surface electromyographic signals x ' (i), G characteristics of each active segment relevant to fatigue are extracted after the active segment is divided, 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) And inputting the trained multi-input LVN network to obtain an estimated muscle strength signal.

2. The method according to claim 1, wherein the method for preprocessing the surface electromyographic signals in step S1 comprises the steps of:

1) Denoising the surface electromyographic signals;

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

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

4) Downsampling the normalized surface electromyographic signals according to a preset frequency;

5) Processing the down-sampled surface electromyographic signals by adopting a correction smoothing method to obtain corrected and smoothed surface electromyographic signals;

the preprocessing method of the muscle strength signal comprises the following steps:

1) Performing a sliding average on the muscle strength signal;

2) Normalizing the muscle strength signal after the moving average by adopting the maximum value of the muscle strength of the sample object under the preset action;

3) And adopting the same downsampling frequency of the surface electromyographic signals to downsample the normalized muscle force signals according to a preset frequency.

3. The method according to claim 1, wherein the method for determining the characteristics in step S3 comprises the steps of:

s3.1: g' alternative features are determined according to actual needs;

s3.2: for each surface electromyographic signal x _k (i) Extracting G' features f of each movable segment _k,d,g′ G '=1, 2, …, G', resulting in each surface electromyographic signal x _k (i) Characteristic sequence corresponding to g' th characteristic in (3)

S3.3: taking the sequence number of the movable segment as a fatigue value and normalizing to obtain each surface electromyographic signal x _k (i) Fatigue value sequence o= {1/D _k ,2/D _k ,…,1}；

S3.4: for each surface electromyographic signal x _k (i) The fatigue value is used as an independent variable and the characteristic value is used as a dependent variable, and the fatigue value is respectively used for each characteristic sequence

For K surface electromyographic signals x _k (i) The determination coefficient of the g' th feature +.>

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

S3.5: for each surface electromyographic signal x _k (i) Each characteristic sequence is calculated separately

And fatigue value sequence o= {1/D _k ,2/D _k Pearson correlation coefficient P of …,1} _k,m′ The method comprises the steps of carrying out a first treatment on the surface of the Then to K surface electromyographic signals x _k (i) Pearson correlation coefficient P of g' th feature in (b) _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

And fatigue value sequence o= {1/D _k ,2/D _k …,1} a second order fit is performed, the fitting polynomial being as follows:

wherein a is ₀ ,a ₁ ,a ₂ To be treatedFitting coefficient o _d Representing the d-th fatigue value in the fatigue value sequence O;

Wherein max _k,g′ 、min _k,g′ Representing the surface electromyographic signal x _k (i) Maximum and minimum values of the g' th feature in all the active segments;

S3.7: screening SVR coefficients for G' optional features

Greater than a preset threshold T _SVR Determining coefficient->

Greater than a preset threshold->

And the pearson correlation coefficient->

Greater than a preset threshold T _P As fatigue-related features.

4. The method as claimed in claim 1, wherein the multi-input LVN network in the step S6 is a two-input LVN network, wherein one input is a surface electromyographic 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 fusion characteristic signal.

5. The method of claim 4, wherein the cost function J of the two-input LVN network is calculated by the following formula:

J＝NMSE+λ||W|| ₁

Constructing weights corresponding to the penalty characteristic signals to obtain +.>

A row vector of the dimension, namely a penalty weight vector W; I.I ₁ The L1 norm is obtained. />