A muscle strength estimation method based on a GRA-ICS-SVR model and MMG signals

Abstract

This invention discloses a muscle strength estimation method based on the GRA-ICS-SVR model and MMG signals. Muscle strength is divided into eight levels, and three-channel MMG signals and corresponding muscle strength signals are collected for each level. Stable muscle strength signal segments and their corresponding MMG signal segments are selected. MMG signal features are extracted from the time domain, frequency domain, time-frequency domain, and nonlinear dynamic characteristics. The GRA method is used to filter effective features from the MMG signal features to construct sample data. The SVR model is used for muscle strength estimation, and an improved Cuckoo Search algorithm (ICS algorithm) is employed to optimize the key parameters of the SVR model, finding the optimal solution for these parameters and establishing a muscle strength estimation model based on ICS-SVR. This invention achieves higher accuracy in muscle strength estimation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knee joint extension force estimation, and particularly relates to a muscle force estimation method based on a GRA-ICS-SVR model and MMG signals. BACKGROUND

[0002] Skeletal muscles, as the power part of the human movement system, are under the control of the nervous system, and together with bones and joints, complete basic limb behaviors such as standing, sitting, walking and jumping, as well as more complex actions. Although these behaviors do not directly affect people's survival, they are directly related to the independent living ability of humans. Movement disorders in muscles, bones and joints have a serious impact on the free movement ability of the human body. In order to help these people with limited movement regain the ability to live independently, rehabilitation assistive devices have received more and more attention.

[0003] Although rehabilitation assistive devices in some structured environments can obtain kinetic information during limb movement through sensors or the mechanical principles of the devices themselves, they can only achieve simple and repetitive tasks. Rehabilitation devices in unstructured environments need to introduce force sensing to increase the ability of flexible interaction. However, the current rehabilitation devices for force control and interaction are far behind the expected effect, and they cannot reflect the movement intention of the user, nor can they make the rehabilitation assistive devices run flexibly and compliantly. Therefore, the force of the human body limbs should be accurately predicted and estimated to help rehabilitation assistive devices and assistive devices improve naturalness and flexibility, respond in time, appropriately assist the user, and make the user obtain effective and comfortable training.

[0004] In human rehabilitation training and assistive assistance, the lower limb knee joint is one of the objects that are widely concerned at present, which determines human walking, jumping and other movements. Knee joint movement is an important part of completing lower limb action, and knee joint force is crucial for completing lower limb action and interacting with the external environment. Knee joint force is a biomechanical measure used to infer the load of the knee joint structure, and timely and accurate estimation of lower limb force information is the key to the output power and force information of lower limb rehabilitation assistive devices, and is also the key to making lower limb knee joint control devices more compliant and flexible.

[0005] However, the complexity of neural information, the time-varying and nonlinear mapping between neural signals and human motion intention bring challenges to intention decoding and subsequent human-computer interaction. Compared with the traditional method of measuring muscle strength, the surface biological signal of human body is convenient to measure and does not damage the human body during the measurement process. Therefore, the estimation of muscle strength based on the surface biological signal of human body is increasingly concerned by researchers. The surface biological signal of human body, such as electromyography (sEMG) or muscle sound signal (MMG). In recent years, scholars have begun to use neural network models, support vector machines and other methods to explore the use of electromyography (sEMG) to estimate or predict human kinetic parameters. However, during the electromyography (sEMG) test, the electromyography sensor not only needs a low-noise environment, but also needs skin surface treatment and appropriate placement conditions. The influence of these conditions often leads to unstable actual test signals, which not only brings difficulties to actual application, but also limits the actual application.

[0006] Relevant studies have shown that muscle activity will not only be accompanied by the generation of sEMG signals, but also MMG signals, i.e. muscle sound signals. MMG signals are generated by the mechanical vibration of muscle fibers when the muscle contracts. Studies have shown that MMG signals can provide information reflecting the recruitment of motor units and firing rates of muscle activity characteristics. Muscle sound signals (MMG) are similar to electromyography (sEMG) and also contain a large amount of information about muscle activity status, movement patterns and movement intentions. At present, there are many types of sensors for testing muscle sound signals, and the placement of the sensor does not need to be very accurate and can be placed on clothes or paper for testing. In addition, muscle sound signals are mechanical signals and are not affected by changes in skin impedance caused by sweating. However, due to the characteristics of weak signal, low frequency and strong randomness of muscle sound signals, it also brings challenges to us to decode muscle sound signals and use muscle sound signals to estimate and predict muscle strength. Using MMG signals to estimate muscle strength needs to solve two key problems, one is to effectively extract features that can reflect muscle activity, and the other is to build a suitable muscle strength estimation model for muscle strength estimation.

[0007] The relevant research shows that there is a nonlinear relationship between the MMG signal and muscle strength, indicating that the MMG signal can be used for muscle strength estimation. Beck studied the reliability relationship between the MMG signal amplitude of the vastus lateralis muscle and muscle strength. Youn and Kim extracted the root mean square amplitude (RMS) and zero-crossing (ZC) features from the MMG signal and input them into a neural network algorithm (ANN) to estimate the elbow flexion strength. Lei extracted the root mean square amplitude (RMS) and frequency domain variance (FV) features from the MMG signal and input them into a neural network algorithm (ANN) to estimate muscle strength. With further research, researchers found that there is a nonlinear relationship between the MMG signal and muscle strength, indicating that the MMG signal can be used to estimate muscle strength, and time domain features and frequency domain features are extracted from the MMG signal and input into a deep learning model to construct a nonlinear relationship between "MMG-muscle strength". Although the existing research has achieved certain results in muscle strength estimation, there are still defects such as few features representing muscle activity and low prediction accuracy of the model. Although good results and accuracy have been achieved, the feature selection still cannot fully represent muscle activity.

[0008] At present, the MMG signal features mainly include time domain features, frequency domain features and time-frequency domain features. Since the MMG signal has nonlinear chaotic characteristics, these nonlinear dynamic features should also be considered to reflect muscle activity. However, when these features are input into a deep learning model, the computational complexity and operation time will increase due to too many features, which may lead to model overfitting and reduce the estimation accuracy. Therefore, feature dimension reduction methods such as PCA, ICA and LDA are widely used by scholars to obtain features related to the mapping model. However, these methods often destroy the original structure of the data, making the reduced data lose the original physical meaning.

[0009] Support vector regression (SVR) modeling is a statistical learning technique, which is a derivative model of support vector machine (SVM). SVR relies on statistical supervised learning theory and is especially suitable for complex and nonlinear regression problems. It has a perfect theoretical basis, strong fitting ability, powerful generalization ability and strong robustness, and is considered to be superior to BP and random forest regression methods. However, in the actual application of the SVR model, the selection of key parameters such as the penalty coefficient and the kernel function width parameter will directly affect the prediction accuracy of the model. Usually, trial-and-error method and grid search method are used to select SVM-related parameters, but these methods are often time-consuming and the results are not good. Therefore, intelligent algorithms are often used to optimize support vector regression (SVR) to build muscle strength estimation models. Although some algorithms such as particle swarm optimization algorithm (PSO) and genetic optimization algorithm (GA) can optimize SVR, they all have the disadvantages of being easy to fall into local optimum or slow in calculation speed, resulting in unsatisfactory optimization model effect. SUMMARY

[0010] In order to overcome the defects in the prior art, the present application provides a muscle strength estimation method based on a GRA-ICS-SVR model and MMG signals, selects MMG signal features to estimate muscle strength, constructs a muscle strength estimation model, and improves the accuracy of muscle strength estimation.

[0011] To achieve the above object, the present application adopts the following technical solutions, comprising:

[0012] A muscle strength estimation method based on a GRA-ICS-SVR model and MMG signals comprises the following steps:

[0013] S1, collecting MMG signals and corresponding muscle strength signals of a subject;

[0014] S2, filtering the collected MMG signals and muscle strength signals respectively;

[0015] S3, extracting features from the filtered MMG signals, extracting MMG signal features; taking the MMG signal features and the muscle strength signals corresponding to the MMG signals, i.e. labels, as sample data; obtaining a certain number of sample data to generate a sample set;

[0016] S4, constructing a muscle strength estimation model, the input of the muscle strength estimation model being MMG signal features and the output being muscle strength estimation values; dividing the sample set into a training set and a test set, training the estimation model using the training set, and constructing the muscle strength estimation model;

[0017] S5, using the muscle strength estimation results obtained in the muscle strength estimation model using the test set to evaluate the performance of the muscle strength estimation model.

[0018] Further, in step S1, three acceleration sensors are used to collect MMG signals at the rectus femoris muscle, the vastus lateralis muscle and the vastus medialis muscle respectively, and three-channel MMG signals are obtained; a force sensor is used to collect muscle strength signals at the lower leg;

[0019] In step S3, the muscle strength signals and the three-channel MMG signal features corresponding to the muscle strength signals are combined as sample data.

[0020] Further, in step S2, a fourth-order Butterworth filter is used to band-pass filter the collected MMG signals to obtain MMG signals with a frequency of 5-100 Hz; a third-order Butterworth filter is used to low-pass filter the collected muscle strength signals to obtain muscle strength signals with a frequency of less than 2 Hz;

[0021] A stable muscle strength signal segment and an MMG signal segment corresponding to the stable muscle strength signal segment are selected; the stable muscle strength signal segment refers to a muscle strength signal segment with a fluctuation range less than a set fluctuation threshold.

[0022] In step S3, feature extraction is performed on the selected MMG signal segment corresponding to the stable muscle force signal segment, and the stable muscle force signal segment and the MMG signal feature corresponding to the stable muscle force signal segment are taken as sample data.

[0023] Further, the muscle force is divided into 8 levels according to the maximum voluntary contraction force MVC, which are 10% MVC, 20% MVC, 30% MVC, 40% MVC, 50% MVC, 60% MVC, 70% MVC and 80% MVC respectively.

[0024] In step S3, MMG signal and muscle force signal are collected for 8 different levels of muscle force signal, and MMG signal corresponding to 8 different levels of muscle force signal is extracted for feature extraction, which is used to construct a sample set.

[0025] Further, in step S3, a sliding window is used to extract MMG signal features, and the following features are extracted for MMG signal in each window:

[0026] The root mean square, kurtosis, standard deviation, slope sign change, average absolute value, zero crossing rate and waveform length of the MMG signal are selected as time domain features.

[0027] The average frequency and median frequency of the MMG signal are selected as frequency domain features.

[0028] The MMG signal is decomposed by three layers of wavelet packet, the wavelet packet energy and the energy of each frequency band are calculated, and the wavelet packet energy and the energy of each frequency band are taken as time-frequency domain features.

[0029] The complexity, sample entropy, approximate entropy, fuzzy entropy, distribution entropy, box number fractal dimension and maximum Lyapunov exponent of the MMG signal are selected as nonlinear dynamic features.

[0030] Further, the grey relational analysis method, i.e. GRA method, is used to select effective features highly related to muscle force from MMG signal features, which are as follows:

[0031] F={F(k)|k=1,2,…,n} is a muscle force sequence;

[0032] Wherein, F(k) is the muscle force signal in the kth sample data; n is the total number of sample data;

[0033] X i ={x i (k)|k=1,2,…,n;i=1,2,…,m} is an MMG signal feature sequence;

[0034] Wherein, x i(k) is the i-th feature of the MMG signal in the k-th sample data, i.e. the i-th MMG signal feature in the k-th sample data; m is the total number of MMG signal features;

[0035] the correlation coefficient ξ between the muscle strength signal in the k-th sample data and the i-th MMG signal feature i The calculation formula of (k) is:

[0036]

[0037] Wherein, ρ is a resolution coefficient, and the value range of ρ is (0, 1); Δ i (k) = |F(k)-x i (k)|, Δ i (k) represents the difference between the muscle strength signal F(k) in the k-th sample data and the i-th MMG signal feature x i (k) in the k-th sample data;

[0038] The calculation formula of the correlation degree r i between the i-th MMG signal feature and the muscle strength signal is:

[0039]

[0040] If the correlation degree r i between the i-th MMG signal feature and the muscle strength signal exceeds a set threshold, the i-th MMG signal feature is taken as an effective feature highly correlated with the muscle strength size, and is screened out;

[0041] In step S4, the screened effective feature highly correlated with the muscle strength size is used as the input of the muscle strength estimation model.

[0042] Further, in step S4, the muscle strength is estimated by using the SVR model, and the specific process is as follows:

[0043] S41, the sample set is divided into a training set and a test set;

[0044] S42, the key parameters of the SVR model are optimized by using the improved cuckoo search algorithm (ICS algorithm) and the training set, the optimal solution of the key parameters of the SVR model is found, and the optimal solution of the key parameters is substituted into the SVR model to construct a muscle strength estimation model;

[0045] The key parameters of the SVR model include a penalty parameter C and a kernel function parameter σ;

[0046] S43, the muscle strength is estimated by using the constructed muscle strength estimation model, the input of the model is the MMG signal feature, and the output is the muscle strength estimation value;

[0047] In step S42, the improved cuckoo search algorithm (ICS algorithm) is used to optimize the key parameters of the SVR model, and the specific process is as follows:

[0048] S421, set the value range of the key parameters of the SVR model;

[0049] S422, set the ICS algorithm parameters and the total number of bird nest positions, i.e., the population size;

[0050] S423, initialize the bird nest positions;

[0051] S424, set the objective function for calculating the fitness value of each bird nest position;

[0052] wherein the bird nest position represents the solution of the key parameters of the SVR model;

[0053] The mean square error (MSE) is selected as the objective function for calculating the fitness value of the bird nest position, and the expression is as follows:

[0054]

[0055] wherein, is the muscle strength estimation value of the i-th sample data, y i is the expected value of the muscle strength, i.e., the true value of the muscle strength, and M is the total number of sample data in the training set; the smaller the mean square error (MSE) value, the better the fitness;

[0056] For the initialized bird nest positions, the objective function is used to calculate the fitness value of each initialized bird nest position, and the bird nest position with the best fitness is selected as the initialized optimal bird nest position;

[0057] S425, Lévy (λ) search is performed using the formula to update the bird nest position; wherein, represents the j-th bird nest position of the t+1-th iteration; represents the j-th bird nest position of the t-th iteration; represents the dot product; α represents the step control quantity for controlling the step search range, and the value of α is subject to normal distribution;

[0058] wherein, the step control quantity α is updated according to the iteration number, and the expression of α is as follows:

[0059]

[0060] wherein, is the optimal bird nest position of the t-th iteration, i.e., the corresponding optimal fitness; α0 is the step factor for controlling the step search range;

[0061] S426, evaluate the bird nest position obtained in this iteration in step S425, select the current optimal bird nest position, and the specific way is:

[0062] Using the objective function, the fitness value of all bird nest positions obtained in step S425 is calculated, and the bird nest position with the optimal fitness value in this iteration is selected, that is, the optimal bird nest position in this iteration;

[0063] If the fitness of the optimal bird nest position in this iteration is better than that of the optimal bird nest position in the last iteration, the optimal bird nest position updated in this iteration is used as the current optimal bird nest position; otherwise, the optimal bird nest position in the last iteration is used as the current optimal bird nest position;

[0064] S427, judge the probability P that the bird nest position obtained in step S425 is found by the host bird a , the bird nest position obtained in this iteration is updated, and the specific way is as follows:

[0065] A random number ε is generated, ε∈(0,1), and it obeys uniform distribution; if ε>P a , it means that the bird nest position obtained in this iteration is found by the host bird, so the bird nest position obtained in this iteration is discarded, and a new bird nest position is updated to replace the bird nest position obtained in this iteration; if ε≤P a , it means that the bird nest position obtained in this iteration is not found by the host bird, so the bird nest position obtained in this iteration is retained;

[0066] Wherein, the probability P that the host bird finds the parasitic egg a According to the number of iterations, the probability P is updated as follows:

[0067] S428, using the objective function, the fitness value of all bird nest positions updated in step S427 is calculated, and the bird nest position with the optimal fitness value is selected to obtain the optimal bird nest position after probability updating;

[0068] The optimal bird nest position after probability updating is compared with the current optimal bird nest position obtained in step S426, and the bird nest position with the optimal fitness value is retained;

[0069] S429, continue to iterate according to the way of steps S425-S428 until the iteration is finished, output the retained current optimal bird nest position, and obtain the final optimal bird nest position, which is the optimal solution of the SVR model key parameters. The optimal solution is substituted into the SVR model to construct the muscle strength estimation model.

[0070] Further, in step S423, Tent chaotic mapping is used for population initialization to obtain the initialized bird nest position, and the specific way is as follows:

[0071] Randomly generate a 0-1 value, and substitute the value into the Tent chaotic mapping to generate H values, that is:

[0072] {q j |j=1,2,…H}={q1,q2,…,q H};

[0073] Wherein, q j represents the jth value generated by the Tent chaotic mapping;

[0074] Map the H values generated by the Tent chaotic mapping into the value range of the key parameters of the SVR model to obtain the initialized H bird nest positions:

[0075]

[0076] Wherein, the superscript 0 represents the current iteration number is 0, that is, the initial value; represents the jth initialized bird nest position; Ub and Lb are the upper limit and lower limit of the value of the key parameters of the SVR model, respectively.

[0077] Further, the values of the parameters P a and a0 in the ICS algorithm are adaptively controlled, and the values of the parameters P a and a0 are as follows:

[0078]

[0079] P a =P amin +(P amax -P amin )((200-t) / T) 2

[0080] Wherein, t and T are the current iteration number and the total iteration number, respectively; a min is the minimum value of the parameter a0; a max is the maximum value of the parameter a0; P amin is the minimum value of the parameter P a ; and P amax is the maximum value of the parameter P a .

[0081] Further, in step S5, the muscle strength estimation results obtained by the test set in the muscle strength estimation model are used for performance evaluation, and the root mean square error RMSE, the mean absolute percentage error MAPE, and the correlation coefficient R are used as performance evaluation indexes, which are as follows:

[0082]

[0083]

[0084]

[0085] wherein, is the muscle strength estimation value of the i th sample data, y i is the muscle strength expectation value of the i th sample data, i.e., the muscle strength true value, D(·) is used for calculating variance, Cov(·) is used for calculating covariance, and N is the total number of sample data in the test set;

[0086] The root mean square error RMSE is used for measuring the deviation between the muscle strength estimation value and the muscle strength true value, the mean absolute percentage error MAPE is used for measuring the average value of the relative error between the muscle strength estimation value and the muscle strength true value in the test set, and the correlation coefficient R is used for measuring the degree of deviation of the muscle strength estimation value from the muscle strength true value;

[0087] The smaller the value of RMSE is, the smaller the value of MAPE is, and the closer the value of R is to 1, which indicates that the performance of the muscle strength estimation model is better, i.e., the accuracy of muscle strength estimation is higher.

[0088] The advantages of the present application are that:

[0089] (1) The precision of muscle strength estimation is the most critical factor for determining the working flexibility and comfort of rehabilitation equipment and powered prostheses, and since the MMG signal acquisition has relatively more portability, reliability, safety and stability than the sEMG signal acquisition, the present application selects the MMG signal for muscle strength estimation.

[0090] (2) The present application uses an acceleration sensor to collect MMG signals at the rectus femoris muscle, the vastus lateralis muscle and the vastus medialis muscle, respectively, uses a force sensor to collect muscle strength signals at the lower leg, carries out filtering processing on the collected MMG signals and muscle strength signals, respectively, divides the muscle strength into 8 levels, collects three-channel MMG signals under each muscle strength level, selects stable muscle strength signal segments and corresponding MMG signal segments for feature extraction, and takes the stable muscle strength signal segments and the MMG signal features corresponding to the stable muscle strength signal segments as sample data, so as to fully ensure the reliability of the sample data.

[0091] (3) The present application extracts MMG signal features capable of comprehensively reflecting muscle activity from the time domain, the frequency domain, the time-frequency domain and the nonlinear dynamics features, which is conducive to improving the estimation precision of the model.

[0092] (4) The application adopts a grey correlation analysis method, i.e., a GRA method, to screen out effective features highly correlated with muscle strength from MMG signal features, effectively avoiding the problems of difficult feature selection and long time spent on feature selection test. The grey correlation analysis method is used to select or reject MMG features to achieve the purpose of dimension reduction, which is beneficial to reduce model calculation load, prevent model overfitting, and improve the accuracy and generalization ability of the model.

[0093] (5) The application uses an SVR model to estimate muscle strength, and the SVR model is suitable for complex and nonlinear regression problems, has a perfect theoretical basis, strong fitting ability, powerful generalization ability, and strong robustness.

[0094] (6) The application uses an improved cuckoo search algorithm, i.e., an ICS algorithm, to optimize key parameters of the SVR model, find the optimal solution of the key parameters of the SVR model, and establish a muscle strength estimation model based on ICS-SVR. The ICS algorithm further improves the optimization performance by using chaotic optimization initial population and adaptive control parameters. The ICS algorithm can efficiently find the optimal solution of the key parameters of the SVR model without searching all parameter points. In the case of wide distribution of the key parameters of the SVR model, the ICS algorithm can greatly improve the optimization efficiency and thus improve the learning performance of the SVR model.

[0095] (7) In the embodiments, the ICS-SVR model is compared with classical machine learning algorithms such as BPNN, ELM and CS-SVR, and finally verifies that the ICS-SVR model performs best in knee extension force estimation, which can realize higher-precision knee extension force estimation. BRIEF DESCRIPTION OF DRAWINGS

[0096] Figure 1 It is a muscle strength estimation method flow chart based on a GRA-ICS-SVR model and MMG signals of the application.

[0097] Figure 2 It is a muscle strength estimation data transmission schematic diagram based on a GRA-ICS-SVR model and MMG signals of the application.

[0098] Figure 3 It is a schematic diagram of filtered signals and selected stable muscle strength signal segments.

[0099] Figure 4 It is a correlation degree result schematic diagram of different MMG signal features and muscle strength size based on a GRA method.

[0100] Figure 5 It is an estimation result schematic diagram of the muscle strength estimation model of the application. DETAILED DESCRIPTION

[0101] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0102] As shown in Figure 1 and Figure 2 , a muscle strength estimation method based on a GRA-ICS-SVR model and MMG signals according to the present application comprises the following steps:

[0103] S1, signal acquisition:

[0104] The subject is required to sit comfortably on a test chair, with the right leg fixed and bent at a 90-degree angle;

[0105] MMG signals and muscle strength signals are collected, and the sampling frequency is 1 kHz;

[0106] MMG signals are collected by using an acceleration sensor; three acceleration sensors are bound to the rectus femoris muscle, the vastus lateralis muscle, and the vastus medialis muscle respectively by using an elastic bandage, and the MMG signals of the rectus femoris muscle (RF-MMG signals), the MMG signals of the vastus lateralis muscle (VL-MMG signals), and the MMG signals of the vastus medialis muscle (VM-MMG signals) are collected respectively; the model of the acceleration sensor is ADXL335;

[0107] Muscle strength signals are collected by using a force sensor; the model of the force sensor is DYLF-30; the force sensor is placed on the lower leg to measure the extension force of the lower leg, which is used as the label of the muscle strength estimation model.

[0108] S2, signal processing:

[0109] In order to reduce the influence of motion and noise, the original MMG signals are subjected to band-pass filtering by using a fourth-order Butterworth filter to obtain MMG signals with a frequency of 5-100 Hz; the original muscle strength signals are subjected to low-pass filtering by using a third-order Butterworth filter to obtain muscle strength signals with a frequency of less than 2 Hz.

[0110] Due to the fluctuation of force in the signal acquisition process, a stable muscle strength signal segment with a fluctuation range of less than 5% of the relative value is selected as experimental data. The filtered signals and the selected stable muscle strength signal segment are as shown in Figure 3 .

[0111] A stable muscle strength signal segment and the MMG signals corresponding to the stable muscle strength signal segment are selected.

[0112] The force conditions, i.e., muscle strength, are divided into 8 levels according to the maximum voluntary contraction force MVC, which are 10% MVC, 20% MVC, 30% MVC, 40% MVC, 50% MVC, 60% MVC, 70% MVC and 80% MVC.

[0113] In this embodiment, the stable force signal obtained by the muscle strength test under one level is 6s, the MMG signal and the muscle strength signal of 6s are subjected to feature extraction and label extraction, and a total of 50 samples are obtained, 400 samples are obtained under 8 levels of force conditions, the 400 samples are shuffled and then divided, of which 360, i.e., 90%, are training sets, and 40, i.e., 10%, are test sets.

[0114] S3, feature extraction:

[0115] S31, a sliding window is used for MMG signal feature extraction, the window length is 1000 data points, and the step length is 100 data points, 25 dimensional features are extracted from the MMG signal of each window, and the specific contents are as follows:

[0116] The root mean square (RMS), kurtosis, standard deviation (SD), slope sign change (SSC), mean absolute value (MAV), zero-crossing rate (ZC) and waveform length (WL) of the MMG signal are selected as time domain features;

[0117] The mean frequency (MPF) and median frequency (MDF) of the MMG signal are selected as frequency domain features;

[0118] The MMG signal is subjected to three-layer wavelet packet decomposition, the wavelet packet energy (WPE) and the energy of each frequency band (WP1, WP2, WP3, WP4, WP5, WP6, WP7, WP8) are calculated, and are used as time-frequency domain features;

[0119] The complexity (LZC), sample entropy (SampEn), approximate entropy (ApEn), fuzzy entropy (FuzzyEn), distribution entropy (DistEn), box number fractal dimension (FD) and maximum Lyapunov exponent (LyapExp) of the MMG signal are selected as nonlinear dynamic features.

[0120] Thus, a 25-dimensional feature set is formed; three-channel MMG signals, i.e., RF-MMG signal, VL-MMG signal and VM-MMG signal, form a 75-dimensional feature set.

[0121] For 8 different levels of muscle strength, MMG signal and muscle strength signal collection is carried out respectively, the stable muscle strength signal segment is extracted, the stable muscle strength signal segment and the MMG signal corresponding to the stable muscle strength signal segment are selected to construct a sample set, and the three-channel MMG signal feature data obtained is merged as a kind of sample data, 8 different levels of muscle strength obtain 400 sample data in total, and constitute a sample set.

[0122] S32, because the units of MMG signal features and muscle strength signal are different and the numerical range difference is large, in order to avoid experimental error caused by the difference between different samples, the feature data in the sample data is normalized, and the Z-score method is used for data normalization processing, and the expression is as follows:

[0123]

[0124] Among them, m=75;x i represents the i-th MMG signal feature;

[0125] By normalizing the MMG signal feature data, overfitting of the ICS-SVR model can also be avoided to some extent.

[0126] S33, the grey correlation analysis method, that is, the GRA method, is used to screen the effective MMG features highly associated with muscle strength:

[0127] Selecting the optimal feature dimension is crucial in pattern recognition, because it determines the accuracy and generalization ability of the classifier, 25-dimensional features are extracted from each MMG signal, and 75-dimensional features are obtained from the three-channel MMG signal, among these features, there are irrelevant or low-correlation features, if these features are input into the model, it will not only increase the model calculation load, but also lead to model overfitting. Therefore, it is necessary to screen the effective features highly associated with muscle strength from the MMG signal features.

[0128] There are two main data dimension reduction methods, one is to destroy the original structure of the data and extract the main features of the data to achieve data dimension reduction, for example, the principal component analysis method PCA, the other is to analyze the correlation of the data and calculate the correlation degree of the data according to certain rules to extract the main features of the data to achieve data dimension reduction, for example, the grey correlation analysis method GRA. In actual engineering problems, the data collected by instruments and equipment has very important significance, and it is not recommended to destroy the original structure of the data when modeling unless necessary. Therefore, the grey correlation analysis method, that is, the GRA method, is used to screen the effective features highly associated with muscle strength from the MMG signal features, and the specific method is as follows:

[0129] F = {F(k) | k = 1, 2, …, n} is a muscle force sequence, i.e., a knee joint extension force sequence;

[0130] X i = {x i (k) | k = 1, 2, …, n; i = 1, 2, …, m} is a MMG signal feature sequence;

[0131] wherein F(k) is a muscle force signal in the kth sample data; n is a total number of sample sets; x i (k) is the ith feature of a MMG signal in the kth sample data, i.e., the ith MMG signal feature; and m is a total number of MMG signal features;

[0132] A calculation formula of the correlation coefficient between the muscle force sequence, i.e., the reference sequence, and the MMG feature sequence, i.e., the comparison sequence, is as follows:

[0133]

[0134] wherein p is a resolution coefficient;

[0135] Let Δ i (k) = |F(k) - x i (k) |, then a difference sequence is as follows:

[0136] Δ i (k) = (Δ i (1), Δ i (2), …, Δ i (n))

[0137] After the calculation formula of the correlation coefficient is simplified, the following is obtained:

[0138]

[0139] The smaller the resolution coefficient p is, the higher the resolution is. In general, the value range of p is (0, 1), and when p ≤ 0.5463, the obtained result is optimal, and p = 0.5 is usually taken.

[0140] The correlation coefficient ξ i (k) represents a correlation degree value of the muscle force sequence and the MMG feature sequence, and an average value of the correlation coefficients at each time is usually taken to represent the correlation degree between the muscle force sequence and the MMG feature sequence, and the calculation method of the correlation degree r i is as follows:

[0141]

[0142] If the ith MMG signal feature and the muscle force signal have a correlation degree r i .If the i-th MMG signal feature exceeds the set threshold, the i-th MMG signal feature is taken as an effective feature highly related to muscle strength size, and is screened out; the screened effective feature related to muscle strength size is taken as an input of a muscle strength estimation model.

[0143] S4, constructing an ICS-SVR model, i.e., constructing a muscle strength estimation model:

[0144] Support vector machine (SVM) is considered as one of the most powerful supervised machine learning algorithms based on the principle of structural risk minimization. As an effective tool, SVM can overcome traditional difficulties such as "dimension disaster" and "over-learning", has a reliable statistical theoretical basis compared with neural networks, has a high prediction accuracy for high-dimensional and small sample data, and is particularly suitable for nonlinear problems and has been successfully applied to classification and prediction fields.

[0145] Support vector regression (SVR) is a kind of support vector machine for processing fitting regression problems. For nonlinear prediction problems, a kernel function is used to map data to a high-dimensional space for linear regression, and a nonlinear relationship between a prediction vector and a support vector is established. SVR has good fitting ability, robustness and fault tolerance for nonlinear and unstable data. Since the kernel function also affects the accuracy of the SVR model prediction, the present application selects an RBF kernel function as the kernel function. However, there are many parameters to be selected and set when using SVR, and the key parameters that have a greater impact on the model results, such as the penalty parameter C and the kernel function parameter sigma, are difficult to select. Random selection of model key parameters or improper selection of model key parameters will have a great impact on the estimation accuracy of the SVR model. Traditional parameter optimization methods such as grid search, particle swarm optimization algorithm PSO and genetic optimization algorithm GA have the disadvantages of falling into local optimum or slow calculation speed. Therefore, the improved cuckoo search algorithm ICS is used to optimize the key parameters of the SVR model to find the optimal solution of the key parameters of the SVR model, and to avoid over-learning or under-learning phenomenon caused by improper parameter selection.

[0146] Cuckoo search (CS) is a new intelligent algorithm inspired by the incubation parasitism behavior of cuckoo birds, which was proposed by YANG et al. in 2009 when studying the selection of the most suitable nest for laying eggs by cuckoo birds. Cuckoos find suitable nests for laying eggs in a random way. In order to simulate the way of cuckoos finding nests for incubating and parasitizing, the CS algorithm sets three ideal assumptions as follows:

[0147] 1. Each cuckoo lays one egg at a time and puts it into a randomly selected host nest; the egg represents a possible solution to the problem in the study;

[0148] 2. The best host nest with the best quality eggs will be passed on to the next generation;

[0149] 3. The number of available host nests is fixed, and the host bird finds the cuckoo's egg with a probability P a ∈(0,1). In this case, the host bird either destroys the egg or abandons the old nest and builds a new one.

[0150] The CS algorithm, like other intelligent optimization algorithms, also has some shortcomings: slow convergence speed at the beginning, insufficient search activity in the later period, and easy to fall into local optimum.

[0151] In step S4, the muscle strength is estimated by using the SVR model, and the key parameters of the SVR model are optimized by using the improved cuckoo search algorithm (ICS algorithm) to find the optimal solution of the key parameters of the SVR model, and an optimal muscle strength estimation model is constructed. The specific process of step S4 is as follows:

[0152] S41, divide the sample set into a training set and a test set;

[0153] S42, use the improved cuckoo search algorithm (ICS algorithm) and the training set to optimize the key parameters of the SVR model, find the optimal solution of the key parameters of the SVR model, and substitute the optimal solution of the key parameters into the SVR model to construct a muscle strength estimation model;

[0154] The key parameters of the SVR model include the penalty parameter C and the kernel function parameter σ;

[0155] S43, use the constructed muscle strength estimation model to estimate the muscle strength, and the input of the model is the MMG signal feature, and the output is the muscle strength estimation value.

[0156] In step S42, the key parameters of the SVR model are optimized by using the improved cuckoo search algorithm (ICS algorithm), and the specific process is as follows:

[0157] S421, set the value range of the key parameters of the SVR model;

[0158] S422, set the parameters of the ICS algorithm and the population size, i.e. the total number of nest positions;

[0159] S423, initialize the nest positions;

[0160] S424, set the objective function for calculating the fitness value of each nest position;

[0161] Where, the nest position represents the solution of the key parameters of the SVR model;

[0162] The mean square error (MSE) is selected as the objective function for calculating the fitness value of the nest position, and the expression is as follows:

[0163]

[0164] wherein, is the muscle strength estimation value of the i-th sample data, y i is the muscle strength expectation value of the i-th sample data, i.e., the muscle strength true value, M is the total number of sample data in the training set; the smaller the mean square error (MSE) value, the better the fitness;

[0165] For the initialized bird nest position, the fitness value of each initialized bird nest position is calculated using the objective function, and the bird nest position with the best fitness is selected as the initialized optimal bird nest position.

[0166] S425, Lévy (λ) search is performed using the formula to update the bird nest position; wherein, represents the bird nest position obtained by the j-th bird in the t+1-th iteration, i.e., the j-th bird nest position in the t+1-th iteration; represents the bird nest position obtained by the j-th bird in the t-th iteration, i.e., the j-th bird nest position in the t-th iteration; represents the dot product; α represents a step control quantity, which is used to control the step search range, and the value of α is subject to a normal distribution;

[0167] wherein, the step control quantity α is updated according to the iteration number, and the expression of α is:

[0168]

[0169] wherein, is the optimal bird nest position in the t-th iteration, i.e., the corresponding optimal fitness; α0 is a step factor for controlling the step search range;

[0170] S426, the bird nest position obtained in step S425 is evaluated, and the current optimal bird nest position is selected, in particular:

[0171] The fitness values of all bird nest positions obtained in step S425 are calculated using the objective function, and the bird nest position with the best fitness in the current iteration is selected, i.e., the optimal bird nest position in the current iteration;

[0172] If the fitness of the optimal bird nest position in the current iteration is better than that of the optimal bird nest position in the last iteration, the optimal bird nest position updated in the current iteration is used as the current optimal bird nest position; otherwise, the optimal bird nest position in the last iteration is used as the current optimal bird nest position.

[0173] S427, the probability P that the bird nest position obtained in step S425 is discovered by the host bird to lay eggs is determineda The bird nest position obtained in this iteration is updated by probability, and the specific process is as follows:

[0174] A random number epsilon is generated, epsilon belongs to (0, 1), and is subject to uniform distribution; if epsilon>P a , it indicates that the bird nest position obtained in this iteration is found by the host bird, the bird nest position obtained in this iteration is discarded, and a new bird nest position is updated to replace the bird nest position obtained in this iteration; if epsilon≤P a , it indicates that the bird nest position obtained in this iteration is not found by the host bird, and the bird nest position obtained in this iteration is retained.

[0175] Wherein, the probability P a is updated according to the number of iterations:

[0176] S428, using the objective function, the fitness value of all the bird nest positions updated by probability in step S427 is calculated, the bird nest position with the optimal fitness is selected, and the optimal bird nest position updated by probability is obtained;

[0177] The optimal bird nest position updated by probability is compared with the current optimal bird nest position obtained in step S426, and the bird nest position with the optimal fitness is retained.

[0178] S429, according to the mode of steps S425-S428, iteration is continued until iteration is ended, the current optimal bird nest position retained is output, and the final optimal bird nest position is obtained. The final optimal bird nest position is the optimal solution of the SVR model key parameters, the optimal solution is substituted into the SVR model, and a muscle strength estimation model is constructed.

[0179] In order to solve the shortcomings and deficiencies of the CS algorithm, the CS algorithm is improved from two aspects in the application, firstly the diversity of population, and secondly the parameter P a and alpha0 closely related to performance.

[0180] The two aspects of improving the CS algorithm are as follows:

[0181] 1. Chaos of population

[0182] Chaos is defined as highly unstable and unpredictable motion in a finite phase space, which often occurs in deterministic nonlinear dynamic systems. In order to overcome the randomness of the solution of the CS algorithm, reduce the non-uniformity of the distribution and improve the convergence ability of the CS algorithm, the diversity of the population needs to be increased. Therefore, the Tent chaotic mapping is used to improve the diversity of the cuckoo population, so as to avoid falling into local optimum in the initial stage. The Tent chaotic mapping is expressed as follows:

[0183]

[0184] wherein q n is a variable to be mapped, q n+1 is a variable after Tent chaotic mapping.

[0185] The population initialization is performed by using the Tent chaotic mapping, and the initialized nest positions are obtained in the following specific manner:

[0186] First, a value of 0-1 is randomly generated, and the value is substituted into the Tent chaotic mapping to generate H values, i.e.

[0187] {q j |j=1,2,…H}={q1,q2,…,q H};

[0188] wherein q j represents the jth value generated by the Tent chaotic mapping.

[0189] The H values generated by the Tent chaotic mapping are mapped into the value range of the key parameters of the SVR model to obtain the initialized H nest positions:

[0190]

[0191] wherein the superscript 0 represents that the current iteration number is 0, i.e. representing an initial value; represents the jth initialized nest position; Ub and Lb are respectively the upper limit and the lower limit of the value range of the key parameters of the SVR model.

[0192] In the present application, since the key parameters of the SVR model include the penalty parameter C and the kernel function parameter σ, the two key parameters are respectively subjected to population chaos, wherein the upper limit Ub C of the value range of the penalty parameter C is 500, the upper limit Ub σ of the value range of the kernel function parameter σ is 500, the lower limit Lb C of the value range of the penalty parameter C is 0.001, and the lower limit Lb σ of the value range of the kernel function parameter σ is 0.001.

[0193] 2. Adaptive control parameters, including: the probability P a of the host bird finding the parasitic egg, a step length factor α0 for controlling the range of random search step length and representing the local search ability.

[0194] The parameters P a and the parameter α0 in the cuckoo search algorithm CS are fixed, and in general cases, P aThe value of alpha 0 is 0.01. However, when facing different optimization problems, the parameters need to be adjusted according to personal experience to adapt to specific needs. The step length obtained by using the Lévy flight mechanism is blind and lacks adaptability, which cannot guarantee rapid convergence. When the value of alpha 0 is large, the global search ability of the algorithm is improved, but the search ability is reduced. If the value of alpha 0 is too small, the local search is good, but the total iteration number increases and the convergence is poor. Therefore, the value of alpha 0 should be gradually reduced during the iteration process. In addition, during the iteration optimization process, the strength of population evolution should be appropriately increased as the quality of the cuckoo individuals gradually improves, so as to avoid the algorithm falling into a local optimal solution. a The value of P should also be gradually reduced as the search proceeds. Therefore, in order to improve the performance of the cuckoo search algorithm CS, the adaptive control parameter strategy is adopted in the present application, and the specific embodiments are as follows:

[0195]

[0196] P a = P amin + (P amax -P amin )((200-t) / T) 2

[0197] Wherein, t and T are the current iteration number and the total iteration number respectively; a min = 0.001, a max = 0.2, P amin = 0.1, P amax = 0.70, T = 200.

[0198] The performance of the SVR model is greatly related to the selection of the key parameters C and sigma. The SVR model is optimized by using the improved cuckoo search algorithm CS (ICS algorithm) to find the optimal solution, and the muscle strength estimation model based on ICS-SVR is established. Unlike the grid search method, the ICS algorithm can find the optimal solution of the parameters without searching all the parameter points, which greatly improves the optimization efficiency and thus improves the learning performance of the SVR model.

[0199] S5, the muscle strength estimation results obtained by using the test set in the muscle strength estimation model are used to evaluate the performance of the muscle strength estimation model.

[0200] The performance evaluation index of the muscle strength estimation model adopted in the present application is the root mean square error RMSE, the mean absolute percentage error MAPE and the correlation coefficient R, which are respectively defined by the following formula:

[0201]

[0202]

[0203]

[0204] in, Let y be the muscle strength estimate for the i-th sample data. i Let be the expected value of muscle strength for the i-th sample data, i.e. the true value of muscle strength. D(·) is used to calculate the variance, Cov(·) is used to calculate the covariance, and N is the total number of sample data in the test set.

[0205] RMSE measures the deviation between estimated and actual muscle strength values ​​and is sensitive to outliers in the data; a smaller value indicates higher accuracy of the prediction model. MAPE considers not only the error between the estimate and the actual value but also the ratio between the error and the actual value, avoiding the problem of error cancellation and accurately reflecting the magnitude of the actual prediction error; a smaller MAPE value indicates better accuracy of the prediction model. MAPE is used to measure the relative error between the average test value and the actual value in the test set. The correlation coefficient R reflects the overall strength of the deviation between the estimate and the actual value; the closer R is to 1, the closer the estimate is to the actual value.

[0206] In this embodiment,

[0207] The GRA method was used to select MMG signal features and extract features highly correlated with knee extension force. First, based on the GRA method, the correlation between knee extension force (muscle strength) and different MMG signal features was calculated, such as... Figure 4 As shown in the figure. Generally, the grey relational classification is as follows: 0.8–1 indicates strong correlation, 0.6–0.8 indicates moderate correlation, and 0.6–0 indicates weak correlation. To obtain the optimal feature description of knee joint extension force, feature combination A was constructed using MMG signal features with a correlation of 0.6–1, feature combination B using MMG signal features with a correlation of 0.75–1, feature combination C using MMG signal features with a correlation of 0.8–1, feature combination D using MMG signal features with a correlation of 0.85–1, and feature combination E using MMG signal features with a correlation of 0.9–1. These features were then input into the ICS-SVR model for knee joint extension force estimation, and the results are shown in Table 1 below.

[0208] Table 1

[0209]

[0210]

[0211] From Table 1, it can be seen that the feature combination D, i.e. the feature combination with a correlation greater than 0.85, has the smallest RMSE, the smallest MAPE and the largest R of the estimated muscle strength of the subject, i.e. the best estimation result of the knee extension force. Therefore, by selecting the feature combination D, the best estimation result of the knee extension force can be obtained. Table 1 also shows that different feature combinations have certain influences on the estimation of the knee extension force, and further shows that some other features have no effect on the estimation result. Therefore, when estimating the knee extension force in the embodiment, by selecting the feature sequence of the feature combination D, overfitting of the MMG-muscle strength model caused by too many input features can be avoided.

[0212] Figure 5 is the estimation result of the knee extension force of the subject under the feature combination D, from which it can be seen that the estimation result is basically consistent with the actual observation result, which shows that the method is effective. Figure 5

[0213] In addition, different muscle strength estimation models are compared to verify the effectiveness of the ICS-SVR model designed in the application. The network structures of the BPNN and the ELM are 14-10-1 and 14-8-1 respectively. The estimation results of different models obtained after running the program for 10 times are shown in Table 2. From Table 2, it can be seen that the estimation result of the ICS-SVR model is better than that of the BPNN, the ELM and the CS-SVR. Moreover, compared with the estimation result fluctuations of the BPNN, the ELM and the CS-SVR, the estimation result fluctuation of the ICS-SVR model is the smallest, and the estimation result is the most stable.

[0214] Table 2

[0215] Regression model RMSE MAPE R BPNN 0.3952±0.3246 0.0361±0.0064 0.9954±0.0012 ELM 1.0464±0.6673 0.1071±0.0190 0.9681±0.0130 CS-SVR 0.1424±0.0274 0.0358±0.0026 0.9923±0.0008 ICS-SVR 0.1295±0.0021 0.0349±8.06E-06 0.9966±1.38E-06

[0216] ​The accuracy of muscle force estimation is the most critical factor to determine the flexibility and comfort of rehabilitation equipment and powered prostheses. Since the MMG signal acquisition is more portable than the sEMG signal, the reliability, safety and stability of its specific application are increasingly valued by people. Knee movement is an important part of completing lower limb action, and knee extension force plays a crucial role in completing lower limb action and interacting with the external environment. The present invention proposes a method of ICS-SVR model combined with GRA feature extraction. From the time domain, frequency domain, time-frequency domain and nonlinear dynamic characteristics, the characteristics that comprehensively reflect muscle activity are extracted, and these characteristics are obtained by the GRA method. The correlation characteristics have high relevance to the muscle force reference sequence, effectively avoiding the problem of difficult feature selection and spending a long time to select features. In addition, the Tent chaotic mapping and adaptive control parameters are introduced to improve the CS, which improves the global convergence ability of the CS algorithm. ICS is applied to the SVR model to search for key parameters, and high precision is obtained in the estimation of knee extension force. The test results of the subjects show that the RMSE and MAPE mean values obtained by using the ICS-SVR combined with the GRA method are significantly lower than those of BPNN, ELM and CS-SVR, and the R mean value is also closer to 1. In general, the ICS-SVR combined with the GRA method is a feasible and effective method in the estimation of knee extension force, and this method can also be expected to further improve the working performance of rehabilitation equipment and powered prosthetic systems.

[0217] The above is only the preferred embodiment of the present invention, and is not intended to limit the present invention. Any modification, equivalent replacement and improvement within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A muscle strength estimation method based on a GRA-ICS-SVR model and MMG signals, characterized by, The method comprises the following steps: S1, collecting MMG signals and corresponding muscle strength signals of a subject; S2, filtering the collected MMG signals and muscle strength signals respectively; S3, extracting features from the filtered MMG signals, and extracting MMG signal features; The MMG signal features and the muscle strength signal corresponding to the MMG signal, i.e. the label, are used as sample data; a certain number of sample data are obtained to generate a sample set; the grey correlation analysis method, i.e. the GRA method, is used to screen out effective features highly correlated with the muscle strength from the MMG signal features; S4, constructing a muscle strength estimation model, wherein the input of the muscle strength estimation model is the MMG signal feature, and the output is a muscle strength estimation value; The sample set is divided into a training set and a test set, the training set is used to train the estimation model, and the muscle strength estimation model is constructed; S5, using the muscle strength estimation result obtained by the test set in the muscle strength estimation model to evaluate the performance of the muscle strength estimation model; In step S4, the muscle strength is estimated by using the SVR model, and the specific process is as follows: S41, dividing the sample set into a training set and a test set; S42, using the improved cuckoo search algorithm, i.e. the ICS algorithm, and the training set to optimize the key parameters of the SVR model, finding the optimal solution of the key parameters of the SVR model, and substituting the optimal solution of the key parameters into the SVR model to construct the muscle strength estimation model; Key parameters of the SVR model include a penalty parameter and a kernel function parameter ; S43, using the constructed muscle strength estimation model to estimate the muscle strength, wherein the input of the model is the MMG signal feature, and the output is a muscle strength estimation value; In step S42, the key parameters of the SVR model are optimized by using the improved cuckoo search algorithm, i.e. the ICS algorithm, and the specific process is as follows: S421, setting the value range of the key parameters of the SVR model; S422, setting the ICS algorithm parameters and the total number of bird nest positions, i.e. the population size; S423, initializing the bird nest positions; S424, setting the objective function for calculating the fitness value of each bird nest position; The bird nest position represents the solution of the key parameters of the SVR model; The mean square error, i.e. the MSE, is selected as the objective function for calculating the fitness value of the bird nest position, and the expression is as follows: wherein, is the muscle strength estimate value of the i-th sample data, i is the muscle strength expectation value of the i-th sample data, i.e., the muscle strength true value, i is the total number of sample data in the training set; the smaller the average mean square error (MSE) value, the better the fitness.​​ For the initialized bird nest positions, the fitness values of the initialized bird nest positions are calculated by using the objective function, and the bird nest position with the optimal fitness is selected as the optimal bird nest position of initialization; S425, using the formula is performed searching, updating the bird nest position; wherein, denotes the t bird nest position of the th iteration of the th iteration of the t th iteration of the th iteration of the denotes the dot product; denotes a step length control quantity for controlling the step length search range, the value of the step length control quantity is subject to a normal distribution; Wherein, the step control quantity According to the number of iterations, the update is performed, The expression of is: wherein, is the optimal nest position for the t th iteration, i.e. the optimal fitness; is a step factor controlling the search range of the step. S426, evaluating the bird nest position obtained in this iteration in step S425, and selecting the current optimal bird nest position, and the specific method is as follows: The fitness values of all the bird nest positions obtained in this iteration in step S425 are calculated by using the objective function, and the bird nest position with the optimal fitness in this iteration, i.e. the optimal bird nest position of this iteration, is selected; If the fitness of the optimal bird nest position of this iteration is better than that of the optimal bird nest position of the last iteration, the optimal bird nest position updated in this iteration is used as the current optimal bird nest position; otherwise, the optimal bird nest position of the last iteration is used as the current optimal bird nest position. S427, judging the probability that the nest position obtained in this iteration is discovered by the host bird to lay eggs The nest position obtained in this iteration is updated as follows: generating a random number , , subject to uniform distribution; if , it indicates that the nest position obtained in this iteration is discovered by the host bird, discarding the nest position obtained in this iteration and updating a new nest position to replace the nest position obtained in this iteration; if , it indicates that the nest position obtained in this iteration is not discovered by the host bird, retaining the nest position obtained in this iteration; where the probability of a host bird finding a parasitic egg Update according to the number of iterations: S428, using the objective function, the fitness value of all the bird nest positions after the probability update in step S427 is calculated, and the bird nest position with the optimal fitness is selected to obtain the optimal bird nest position after the probability update; The optimal bird nest position after the probability update is compared with the current optimal bird nest position obtained in step S426, and the bird nest position with the optimal fitness is retained; S429, in the manner of steps S425-S428, the iteration is continued until the iteration is ended, and the retained current optimal bird nest position is output to obtain the final optimal bird nest position. The final optimal bird nest position is the optimal solution of the SVR model key parameters. The optimal solution is substituted into the SVR model to construct a muscle strength estimation model; In step S423, the Tent chaotic mapping is used for population initialization to obtain the initialized bird nest position, and the specific manner is as follows: A random number between 0 and 1 is generated, and this number is substituted into the Tent chaotic map to generate a number, i.e.: H A random number between 0 and 1 is generated, and this number is substituted into the Tent chaotic map to generate a number, i.e.: ； wherein, represents the number of values generated by the Tent chaotic mapping for the first Generated by the Tent chaotic map H Each value is mapped to the range of values ​​for the key parameters of the SVR model, resulting in the initial values. H Location of each bird's nest: wherein the superscript 0 indicates that the current iteration number is 0, i.e. represents the initial value; represents the initialized first bird's nest position; respectively represent the upper limit and the lower limit of the value of the key parameter of the SVR model. The values of parameters in the ICS algorithm are adaptively controlled and parameters The values of parameters and parameters are as follows: wherein, t and T are the current iteration number and the total iteration number, respectively; is the minimum value of the parameter ; is the maximum value of the parameter ; is the minimum value of the parameter P a ; is the maximum value of the parameter P a .

2. The muscle strength estimation method based on GRA-ICS-SVR model and MMG signal according to claim 1, wherein, In step S1, three acceleration sensors are used to collect MMG signals at the rectus femoris, vastus lateralis and vastus medialis to obtain three-channel MMG signals; and a force sensor is used to collect muscle strength signals at the calf. In step S3, the muscle strength signal and the three-channel MMG signal features corresponding to the muscle strength signal are combined as sample data.

3. The muscle strength estimation method based on GRA-ICS-SVR model and MMG signal according to claim 1 or 2, characterized in that, In step S2, a fourth-order Butterworth filter is used for band-pass filtering of the collected MMG signals to obtain MMG signals with a frequency of 5-100Hz; and a third-order Butterworth filter is used for low-pass filtering of the collected muscle strength signals to obtain muscle strength signals with a frequency below 2Hz; A stable muscle strength signal segment and an MMG signal segment corresponding to the stable muscle strength signal segment are selected; the stable muscle strength signal segment refers to a muscle strength signal segment with a fluctuation range less than a set fluctuation threshold; In step S3, the selected MMG signal segment corresponding to the stable muscle strength signal segment is subjected to feature extraction, and the stable muscle strength signal segment and the MMG signal features corresponding to the stable muscle strength signal segment are used as sample data.

4. The muscle strength estimation method based on the GRA-ICS-SVR model and MMG signal according to claim 1 or 2, characterized in that, The muscle strength is divided into eight levels according to the maximum voluntary contraction MVC, i.e., 10%MVC, 20%MVC, 30%MVC, 40%MVC, 50%MVC, 60%MVC, 70%MVC and 80%MVC; In step S3, MMG signals and muscle strength signals are collected for the eight different levels of muscle strength signals, and MMG signals corresponding to the eight different levels of muscle strength signals are subjected to feature extraction for constructing a sample set.

5. The muscle strength estimation method based on GRA-ICS-SVR model and MMG signal according to claim 1 or 2, characterized in that, In step S3, a sliding window is used for MMG signal feature extraction, and the following features are extracted for the MMG signal in each window: The root mean square, kurtosis, standard deviation, slope sign change, average absolute value, zero-crossing rate and waveform length of the MMG signal are selected as time domain features; The average frequency and median frequency of the MMG signal are selected as frequency domain features; The MMG signal is subjected to three-layer wavelet packet decomposition, the wavelet packet energy and the energy of each frequency band are calculated, and the wavelet packet energy and the energy of each frequency band are used as time-frequency domain features; The complexity, sample entropy, approximate entropy, fuzzy entropy, distribution entropy, box number fractal dimension, and maximum Lyapunov exponent of the MMG signal are selected as the nonlinear dynamic characteristics.

6. The muscle strength estimation method based on GRA-ICS-SVR model and MMG signal according to claim 5, wherein, The grey relational analysis (GRA) method is used to screen the effective features highly related to the muscle strength from the MMG signal features, and the specific process is as follows: For muscle strength sequence; wherein, is the number of the first sample data; is the muscle strength signal in the first sample data; is the total number of sample data; is a sequence of features of the MMG signal; wherein, is the th feature of the MMG signal in the i th sample data, i.e., the th MMG signal feature in the i th sample data; is the total number of MMG signal features. The correlation coefficient between the muscle strength signal in the first sample data and the second MMG signal feature is calculated according to the following formula: i ​​​ in, For resolution coefficients, The range of values ​​for is (0,1); , Indicates the first Muscle force signals in a sample data With the i MMG signal characteristics The difference between them; The correlation degree of the first i MMG signal feature and the muscle strength signal The calculation formula is: If the first i Correlation between individual MMG signal features and muscle strength signals If the threshold is exceeded, then the [number]th [item] will be [affected / removed]. i Several MMG signal features were selected as effective features highly correlated with muscle strength and size. In step S4, the effective features highly related to the muscle strength are used as the input of the muscle strength estimation model.

7. The muscle strength estimation method based on GRA-ICS-SVR model and MMG signal according to claim 1, wherein, In step S5, the muscle strength estimation results obtained by the test set in the muscle strength estimation model are used for performance evaluation, and the root mean square error (RMSE), mean absolute percentage error (MAPE), and correlation coefficient (R) are used as the performance evaluation indexes, and the specific process is as follows: wherein, is the muscle strength estimate value of the i-th sample data, i is the muscle strength estimate value of the i-th sample data, is the muscle strength expected value of the i-th sample data, i.e., the muscle strength true value, i is the muscle strength expected value of the i-th sample data, i.e., the muscle strength true value, is used to calculate the variance, is used to calculate the covariance, N is the total number of sample data in the test set; The root mean square error (RMSE) is used to measure the deviation between the muscle strength estimation value and the actual muscle strength value; the mean absolute percentage error (MAPE) is used to measure the average value of the relative error between the muscle strength estimation value and the actual muscle strength value in the test set; and the correlation coefficient (R) is used to measure the degree of deviation of the muscle strength estimation value from the actual muscle strength value. The smaller the value of the RMSE and the MAPE, and the closer the value of the R to 1, the better the performance of the muscle strength estimation model, that is, the higher the accuracy of the muscle strength estimation.

Images