Muscle movement recognition method and surface electromyogram signal collection device

Abstract

The application relates to a muscle movement recognition method and a surface electromyogram signal collection device, and the method comprises the following steps: collecting historical surface electromyogram signals by using the surface electromyogram signal collection device; processing the historical surface electromyogram signals to obtain a surface electromyogram signal feature subset; constructing a convolutional neural network model; training the convolutional neural network model by using the surface electromyogram signal feature subset to obtain a muscle movement recognition model; and recognizing muscle movement according to the muscle movement recognition model. The device comprises an electromyogram signal sensor, a low-pass filter, an operational amplifier, an analog-to-digital converter and a wireless transmission module. The surface electromyogram signal feature subset is obtained by processing historical surface electromyogram signals, and the muscle movement recognition model is obtained by training the surface electromyogram signal feature subset; therefore, the muscle movement can be recognized, and the recognition process is faster and more accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of muscle movement detection, in particular to a muscle movement recognition method and a surface electromyography signal acquisition device. BACKGROUND

[0002] Surface electromyography (sEMG) is the superposition of motor unit action potentials in muscle fibers in time and space, and the nervous system controls the activity of muscles (contraction or relaxation). Different muscle fiber motor units on the surface of the skin produce different signals at the same time, which contains a series of rich pathological / physiological information. At present, sEMG is widely used in clinical medicine, human-machine ergonomics, rehabilitation medicine, sports science, rehabilitation prosthesis application, gesture action recognition and other fields. However, silver and other rigid electrodes have the problems of large skin electrode contact impedance, easy to be interfered by power frequency signals, uneven skin surface contact area and the like in the process of collecting surface electromyography signals; in addition, the collected sEMG signals have strong external interference, multi-channel signal processing is cumbersome, and feature vector selection is difficult, which limits the further development and large-scale application of sEMG signals. SUMMARY

[0003] The purpose of the present application is to provide a muscle movement recognition method and a surface electromyography signal acquisition device to quickly and accurately collect surface electromyography signals and recognize muscle movement.

[0004] To achieve the above-mentioned purpose, the present application provides the following scheme:

[0005] A surface electromyography signal acquisition device, comprising:

[0006] An electromyography signal sensor for collecting an initial surface electromyography signal;

[0007] A low-pass filter connected with the electromyography signal sensor for filtering the initial surface electromyography signal to obtain a filtered signal;

[0008] An operational amplifier connected with the low-pass filter for amplifying the filtered signal to obtain an amplified signal;

[0009] An analog-to-digital converter connected with the operational amplifier for converting the amplified signal into a digital signal to obtain a surface electromyography signal;

[0010] A wireless transmission module connected with the analog-to-digital converter for transmitting the surface electromyography signal.

[0011] Optionally, the electromyography signal sensor comprises, from top to bottom, a touch surface layer, a motion trajectory suppression layer, a fabric conductive layer, and a substrate; the touch surface layer is used to contact a skin surface to be collected and collect an initial surface electromyography signal; the motion trajectory suppression layer is used to buffer the interaction between the touch surface layer and the skin surface to be collected; and the fabric conductive layer is used to transmit the initial surface electromyography signal.

[0012] Optionally, the touch surface layer is a knitted silver conductive cloth, the motion trajectory suppression layer is a conductive sponge, and the substrate is an elastic nylon material.

[0013] Optionally, the electromyography signal sensor further comprises:

[0014] An insulating shielding layer is arranged on both sides of the touch surface layer and is used to isolate the skin surface to be collected from the fabric conductive layer; the insulating shielding layer is an insulating fabric material with a hollow center.

[0015] Optionally, the low-pass filter is an RC low-pass filter; the operational amplifier and the analog-to-digital converter are ADS1299 chips; and the wireless transmission module is a WIFI module integrated in a CC3200 chip.

[0016] A muscle movement recognition method using the surface electromyography signal collection device, comprising:

[0017] Obtaining a historical surface electromyography signal;

[0018] Processing the historical surface electromyography signal to obtain a surface electromyography signal feature subset;

[0019] Constructing a convolutional neural network model;

[0020] Training the convolutional neural network model using the surface electromyography signal feature subset to obtain a muscle movement recognition model;

[0021] Recognizing muscle movement according to the muscle movement recognition model.

[0022] Optionally, the processing of the historical surface electromyography signal to obtain a surface electromyography signal feature subset specifically comprises:

[0023] Normalizing the historical surface electromyography signal to obtain a normalized signal;

[0024] Filtering the normalized signal to obtain a filtered signal;

[0025] Extracting the time domain features of the filtered signal to construct a feature matrix;

[0026] Filtering the feature matrix to obtain a filtered feature matrix;

[0027] According to the filtered feature matrix, a surface electromyogram signal feature subset is obtained by using a random forest-recurrent feature elimination algorithm.

[0028] Optionally, the filtering of the normalized signal to obtain a filtered signal is specifically:

[0029] Discrete wavelet transformation is performed on the normalized signal to obtain high-frequency signal coefficients and low-frequency signal coefficients;

[0030] According to a preset threshold, the high-frequency signal coefficients are filtered to obtain filtered high-frequency signal coefficients;

[0031] Signal reconstruction is performed by using the low-frequency signal coefficients and the filtered high-frequency signal coefficients to obtain a filtered signal.

[0032] Optionally, the filtering of the feature matrix is specifically performed by using a formula:

[0033]

[0034] The feature matrix is filtered; wherein y i is an i th value in the filtered feature matrix, N is a filter order, f j is a j th value extracted from the feature matrix, and the cumulative sum of f j is 1, q is a proportional coefficient.

[0035] Optionally, the filter order and the proportional coefficient are calculated by using a particle swarm algorithm; wherein the particle swarm algorithm calculation process is:

[0036] The filter order and the proportional coefficient are used to initialize the number of particles;

[0037] The fitness value of the particle is calculated according to the position of the particle;

[0038] The optimal position of the particle, the optimal position of the particle swarm, and the speed and position of the particle are updated according to the fitness value of the particle, the optimal position of the particle passed through, and the optimal position of the particle swarm passed through under the number of particles;

[0039] The final particle is output according to the updated optimal position of the particle swarm; the final particle is the filter order and the proportional coefficient.

[0040] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0041] The application discloses a muscle movement recognition method and a surface electromyogram signal collection device, and the method comprises the following steps: collecting historical surface electromyogram signals by using the surface electromyogram signal collection device; processing the historical surface electromyogram signals to obtain a surface electromyogram signal feature subset; constructing a convolutional neural network model; training the convolutional neural network model by using the surface electromyogram signal feature subset to obtain a muscle movement recognition model; and recognizing muscle movement according to the muscle movement recognition model. The device comprises: an electromyogram signal sensor, which is used for collecting initial surface electromyogram signals; a low-pass filter, which is connected with the electromyogram signal sensor and is used for filtering the initial surface electromyogram signals to obtain filtered signals; an operational amplifier, which is connected with the low-pass filter and is used for amplifying the filtered signals to obtain amplified signals; an analog-to-digital converter, which is connected with the operational amplifier and is used for converting the amplified signals into digital signals to obtain surface electromyogram signals; and a wireless transmission module, which is connected with the analog-to-digital converter and is used for transmitting the surface electromyogram signals. The surface electromyogram signal feature subset is obtained by processing the historical surface electromyogram signals, and the muscle movement recognition model is obtained by training the surface electromyogram signal feature subset; therefore, the muscle movement can be recognized by using the muscle movement recognition model, and the recognition process is faster and more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0043] Figure 1 The surface electromyogram signal collection device structure diagram provided for the embodiment 1 of the present application is shown in the figure.

[0044] Figure 2 The RC low-pass circuit diagram provided for the embodiment 1 of the present application is shown in the figure.

[0045] Figure 3 The electromyogram signal sensor structure diagram provided for the embodiment 1 of the present application is shown in the figure.

[0046] Figure 4 The muscle movement recognition method flow chart provided for the embodiment 2 of the present application is shown in the figure.

[0047] Figure 5 The flow chart for obtaining the surface electromyogram signal feature subset provided for the embodiment 2 of the present application is shown in the figure.

[0048] Figure 6 The flow chart for filtering the normalized signals provided for the embodiment 2 of the present application is shown in the figure.

[0049] Figure 7A signal diagram before filtering provided for the embodiment 2 of the present application;

[0050] Figure 8 A signal diagram after filtering provided for the embodiment 2 of the present application;

[0051] Figure 9 A particle swarm algorithm process schematic diagram provided for the embodiment 2 of the present application;

[0052] Figure 10 A random forest-recursive feature elimination algorithm flow chart provided for the embodiment 2 of the present application.

[0053] Symbol explanation: 1-touch surface layer, 2-movement trajectory inhibition layer, 3-substrate, 4-insulating shielding layer, 5-fabric conductive layer. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be apparently and completely described below with reference to the drawings in the embodiments of the present application, obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0055] The purpose of the present application is to provide a muscle movement recognition method and a surface electromyogram signal acquisition device to quickly and accurately acquire surface electromyogram signals and recognize muscle movements.

[0056] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0057] Embodiment 1

[0058] Figure 1 A surface electromyogram signal acquisition device structure diagram provided for the embodiment 1 of the present application is shown as Figure 1 The surface electromyogram signal acquisition device comprises:

[0059] The electromyogram signal sensor is used to acquire initial surface electromyogram signals.

[0060] The low-pass filter is connected with the electromyogram signal sensor and is used to filter the initial surface electromyogram signals to obtain filtered signals.

[0061] The operational amplifier is connected with the low-pass filter and is used to amplify the filtered signals to obtain amplified signals.

[0062] The analog-to-digital converter is connected with the operational amplifier and is used to convert the amplified signals into digital signals to obtain surface electromyogram signals.

[0063] The wireless transmission module, connected to the analog-to-digital converter, is used to transmit surface electromyography signals.

[0064] In this embodiment, the surface electromyography (EMG) signal acquisition device further includes a controller connected to the EMG signal sensor, low-pass filter, operational amplifier, analog-to-digital converter, and wireless transmission module. The controller is used to control the operation of the EMG signal sensor, low-pass filter, operational amplifier, analog-to-digital converter, and wireless transmission module. The control module is a CC3200 chip.

[0065] In this embodiment, the operational amplifier and analog-to-digital converter are ADS1299 chips. The wireless transmission module is a WIFI module integrated within the CC3200 chip. The low-pass filter is an RC low-pass filter. Figure 2 The diagram shows an RC low-pass circuit provided in Embodiment 1 of the present invention.

[0066] Figure 3 This is a structural diagram of the electromyography signal sensor provided in Embodiment 1 of the present invention, as shown below. Figure 3 As shown, the electromyography (EMG) signal sensor includes, from top to bottom, a contact layer 1, a motion-inhibiting layer 2, a conductive fabric layer 5, and a substrate 3. The contact layer 1 is used to contact the skin surface to collect initial surface EMG signals. The motion-inhibiting layer 2 buffers the interaction between the contact layer 1 and the skin surface. The conductive fabric layer 5 transmits the initial surface EMG signals and also shields against high-frequency interference signals. Specifically, the EMG signal sensor also includes an insulating shielding layer 4, disposed on both sides of the contact layer 1, to isolate the skin surface from the conductive fabric layer 5.

[0067] The contact layer 1 is made of knitted silver conductive cloth, which overcomes the problem of unevenness in the contact area of ​​rigid electrodes and has good conductivity and extremely high sensitivity. The motion trajectory suppression layer 2 is made of conductive sponge. The substrate 3 is made of elastic nylon material. The insulating shielding layer 4 is an insulating fabric material with a hollow center. It has a certain dimensional difference from the contact layer 1, forming a supporting plane to prevent the skin surface to be sampled from contacting the conductive fabric layer 5 and generating interference signals.

[0068] Example 2

[0069] Figure 4 The flowchart of the muscle movement recognition method provided in Embodiment 2 of the present invention is as follows: Figure 4 As shown, muscle movement recognition methods include:

[0070] Step 101: Acquire historical surface electromyography (EMG) signals. These historical EMG signals are acquired using the surface EMG signal acquisition device described in Example 1.

[0071] Step 102: Process the historical surface electromyography (EMG) signals to obtain a subset of surface EMG signal features.

[0072] Step 103: Construct a convolutional neural network model.

[0073] Step 104: Train the convolutional neural network model using the subset of surface electromyography signal features to obtain a muscle movement recognition model.

[0074] Step 105: Identify muscle movements based on the muscle movement recognition model.

[0075] Figure 5 The flowchart for obtaining a subset of surface electromyography signal features provided in Embodiment 2 of the present invention is as follows: Figure 5 As shown, step 102 specifically includes:

[0076] Step 1021: Normalize the historical surface electromyography signal to obtain a normalized signal; specifically, using the formula:

[0077]

[0078] The historical surface electromyography signals are normalized; wherein, X i ' is the normalized surface electromyography signal, X' i For historical surface electromyography signals, For X i The mean, i = 1, 2, ..., n, where i is the i-th value, n is the total number, and σ 2 For variance,

[0079] Step 1022: Filter the normalized signal to obtain a filtered signal; specifically: perform discrete wavelet transform on the normalized signal to obtain high-frequency signal coefficients and low-frequency signal coefficients. Filter the high-frequency signal coefficients according to a preset threshold to obtain filtered high-frequency signal coefficients. Reconstruct the signal using the low-frequency signal coefficients and the filtered high-frequency signal coefficients to obtain the filtered signal.

[0080] Figure 6 This is a flowchart of filtering the normalized signal provided in Embodiment 2 of the present invention, as shown below. Figure 5 As shown, in space V j =V j-1 +W j-1 The above represents the normalized surface electromyography signal X. i As shown in equation (2), for each in V j Signals in space can be represented by two basis functions.

[0081]

[0082] where A1(k) and D1(k) are two coefficients of scale metric space j-1, the decomposition of coefficient A0 from j space, c is a constant, φ j,k (t) is X i the decomposition quantity on space j, φ j-1,k (t) and ω j-1,k (t) is the decomposition quantity on space j-1. A0(k) is the coefficient on the entire space.

[0083] The process of decomposing A0(k) into coefficients A1(k) and D1(k) is as follows:

[0084]

[0085] where A1(k) corresponds to the accurate coefficient, D1(k) corresponds to the coarse coefficient, h0 is the low-pass filter coefficient, h1 is the high-pass filter coefficient, n represents the filter order, and k represents the kth data.

[0086] The threshold ξ is set for the decomposed coefficient D1(k), as shown in formula (4):

[0087]

[0088] where N is the total number of signals, m = Median(|D1(k)|) / 0.6745, m is the coefficient, and Median() is the median function.

[0089] D1(k) is filtered, and when the value in D1(k) is less than the threshold ξ, it is set to zero, and when the value in D1(k) is greater than or equal to the threshold ξ, it is kept unchanged, that is:

[0090]

[0091] The filtered signal is reconstructed according to formula (2) again to obtain the filtered signal. The comparison chart of the signals before and after filtering is shown in Figures 7-8 .

[0092] Step 1023: Extracting the time domain features of the filtered signal, constructing a feature matrix; the time domain features include root mean square, variance, wavelength length, zero crossing point number, average absolute value, maximum fractal length, average energy and autoregressive coefficient.

[0093] The root mean square RMS calculation formula is as follows:

[0094]

[0095] where N is the total number of signals, Y i is the ith signal in the filtered signal Y.

[0096] The variance VAR calculation formula is as follows:

[0097]

[0098] The wavelength length WL is calculated according to the following formula:

[0099]

[0100] wherein Y i+1 is the ith signal in the filtered signal Y.

[0101] The zero-crossing point number ZC is calculated according to the following formula:

[0102]

[0103]

[0104] wherein τ is a threshold value for avoiding low-level noise, set to 0.5, and sgn(x) is a sign function, used to determine whether a single zero-crossing point exists.

[0105] The average absolute value MAV is calculated according to the following formula:

[0106]

[0107] The maximum fractal length MLF is calculated according to the following formula:

[0108]

[0109] The average energy AP is calculated according to the following formula:

[0110]

[0111] The autoregressive coefficient is calculated according to the following formula:

[0112]

[0113] wherein α is the autoregressive coefficient, p is the order of the autoregressive coefficient, e is the residual error, and Y k is a time series with a sequence number k, and k is the number of sequences.

[0114] Step 1024: filtering the feature matrix to obtain a filtered feature matrix.

[0115] In some actions of collecting electromyographic signals, it is inevitable to cause different electromyographic signal amplitudes of the same action due to inconsistent execution of the action and other disturbances, resulting in isolated points or outliers in different windows. To solve this problem, the embodiment proposes a filtering method based on feature information. That is, considering the existence of isolated points / outliers, each point is replaced by multiplying the previous and next N values by a time-decreasing proportional coefficient. Specifically, the formula is used:

[0116]

[0117] filtering the feature matrix; wherein, y i is the i th value in the filtered feature matrix, N is the filter order, f j is the j th value extracted from the feature matrix, f j the cumulative sum of which is 1, q is a proportional coefficient. The filter order and the proportional coefficient are optimized by using a particle swarm algorithm, and the specific process is as follows:

[0118] The number of particles is initialized by using the filter order and the proportional coefficient.

[0119] The particle fitness value is calculated according to the position of the particle.

[0120] The particle optimal position, the particle swarm optimal position, and the speed and position of the particle are updated according to the particle fitness value, the optimal position passed by the particle, and the optimal position passed by the particle swarm under the number of particles.

[0121] The final particle is output according to the updated particle swarm optimal position; the final particle is the filter order and the proportional coefficient.

[0122] Figure 9 The particle swarm algorithm process diagram provided by Example 2 of the present application is shown in FIG. 1, and the specific optimization process is as follows: Figure 9

[0123] Random particles are initialized, the maximum number of iterations is set to 800, the maximum speed of the particle Vmax is 5, the position information is the entire search space, the speed and position are randomly initialized in the speed interval and the search space, the particle swarm size is set to M (i.e., different combinations of the filter order and the proportional coefficient), and each particle is randomly initialized with a flying speed.

[0124] The updated speed and position of the particle are as follows:

[0125]

[0126] wherein, i is the i th particle, i = 1, 2,..., N1, N1 is the total number of particles, v i is the particle speed, rand() is a random number, x i is the particle position, c1 and c2 are learning factors, and are set to 2.

[0127] The particle update speed contains three parts:

[0128] (1) The first part is the previous speed of the particle.

[0129] ​(2) The second part is the cognitive part, which represents the thinking of the particle itself, i.e., the distance between the current position of the particle and the optimal position.

[0130] (3) The third part is the social part, which represents the information sharing and cooperation between particles, i.e., the optimal position between the particle and the group.

[0131] Calculate the fitness value of the particle, and the calculation formula is as follows:

[0132] Iny = (cos(2Πx[0])+cos(2Πx[1])) / 2-2.71289 (16)

[0133] Where y is the fitness value, and x is the particle position.

[0134] Initialize a group of particles (including random position and velocity).

[0135] Evaluate and assess the fitness value of each particle.

[0136] For each particle, compare its fitness value with its best position pbest it has passed, if better, then take it as the current best position pbest. (pbest is the historical best position)

[0137] For each particle, compare its fitness value with its best position gbest it has passed, if better, then take it as the current best position gbest. (gbest is the global best position)

[0138] Adjust the particle velocity and position according to formula (15).

[0139] Determine whether the end condition is reached. The end condition is to reach the maximum number of iterations 500, or the global best position meets the minimum limit J.

[0140] If not, re-evaluate and assess the fitness value of each particle; if yes, output the final particle, and the final particle is the filter order and the proportional coefficient.

[0141] Step 1025: According to the filtered feature matrix, a random forest-recursive feature elimination algorithm is used to obtain a surface electromyogram signal feature subset. Specifically:

[0142] A bootstrap resampling method is used to extract multiple samples from the original sample, a decision tree is constructed for each bootstrap sample, all decision trees constitute a random forest, and the feature importance is calculated in the regression model. At this time, the backward iterative feature evaluation is introduced, and after deleting the features with small feature importance, the random forest algorithm is used again to calculate the feature importance of the remaining features, until the last feature is only one. According to the correlation coefficient and the root mean square error, the most characteristic set is selected.

[0143] The detailed process is as follows:

[0144] (1) The filtered matrix S has n samples, bootstrap sampling is applied to randomly draw b sample subsets with replacement, and b regression trees are constructed according to the sample subsets; the samples not drawn each time of bootstrap sampling form b out-of-bag data, and the out-of-bag data form test samples of the random forest.

[0145] (2) The number of features of the original sample set is p, m variables (m<=p) are randomly drawn at each node of each regression tree as candidate variables, then the optimal branch is selected from the candidate variables according to certain criteria, and the maximum growth of each decision tree is determined in different cases;

[0146] (3) The b regression trees generated in (1) are integrated into a random forest regression model, the residual mean square MSE of the out-of-bag data prediction is used to evaluate the effect of the random forest regression model, and the calculation formula of the MSE is as follows:

[0147]

[0148] Wherein, s i is the actual value of the dependent variable in the out-of-bag data, is the predicted value of the random forest for the out-of-bag data.

[0149] (4) The average drop MSE value is calculated by the residual mean square of the out-of-bag data prediction, and the variable importance in the random forest regression can be measured by the average drop MSE value, and the greater the value, the more important the feature.

[0150] (5) After the average drop MSE value is calculated, according to the principle of backward iteration, the feature with the smallest feature importance is deleted first, then the remaining features are repeated (1) to (4), the features with small importance are gradually deleted, until the last feature is left, and the feature number with the smallest root mean square error and the largest correlation coefficient is selected as the feature selection result for forest biomass remote sensing estimation. Figure 10 The random forest-recursive feature elimination algorithm flowchart provided in embodiment 2 of the present application.

[0151] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0152] (1) The present application filters the signal twice, double denoising, greatly reduces the noise, and improves the signal-to-noise ratio.

[0153] (2) The application adopts RF-RFE (Random Forest-Recursive Feature Elimination algorithm), in the iterative process of feature selection, the current remaining feature set is re-evaluated, and the score of each feature is adjusted in the repeated iteration process, overcoming the defect that the feature selection result of single random forest needs repeated test.

[0154] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between various embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0155] The principles and implementation manners of the application are described by using specific examples in the present application, and the above embodiment description is only used to help understand the method and core idea of the application; meanwhile, for the general skilled in the art, the specific implementation manner and application range can be changed according to the idea of the application. In conclusion, the content of the specification should not be understood as the limitation of the application.

Claims

1. A surface electromyogram signal acquisition device, characterized in that, include: Electromyography (EMG) signal sensor, used to acquire initial surface EMG signals; A low-pass filter, connected to the electromyography signal sensor, is used to filter the initial surface electromyography signal to obtain a filtered signal. An operational amplifier, connected to the low-pass filter, is used to amplify the filtered signal to obtain an amplified signal; An analog-to-digital converter, connected to the operational amplifier, is used to convert the amplified signal into a digital signal to obtain a surface electromyography signal; A wireless transmission module, connected to the analog-to-digital converter, is used to transmit the surface electromyography signals; A controller, connected to the electromyography (EMG) signal sensor, low-pass filter, operational amplifier, analog-to-digital converter, and wireless transmission module, is used to control the operation of the EMG signal sensor, low-pass filter, operational amplifier, analog-to-digital converter, and wireless transmission module; the controller is used to acquire historical surface EMG signals; and to process the historical surface EMG signals to obtain a subset of surface EMG signal features; Construct a convolutional neural network model; train the convolutional neural network model using the feature subset of surface electromyography signals to obtain a muscle movement recognition model; identify muscle movements based on the muscle movement recognition model; The process of processing the historical surface electromyography (EMG) signals to obtain a subset of surface EMG signal features includes: normalizing the historical surface EMG signals to obtain a normalized signal; and filtering the normalized signal to obtain a filtered signal, specifically: The normalized signal is subjected to discrete wavelet transform to obtain high-frequency signal coefficients and low-frequency signal coefficients. The high-frequency signal coefficients are filtered according to a preset threshold to obtain filtered high-frequency signal coefficients. The low-frequency signal coefficients and the filtered high-frequency signal coefficients are used to reconstruct the signal to obtain the filtered signal. Extract the time-domain features of the filtered signal to construct a feature matrix; the time-domain features include root mean square, variance, wavelength, number of zero crossings, mean absolute value, maximum fractal length, average energy, and autoregressive coefficients; filter the feature matrix to obtain a filtered feature matrix; and use the random forest-recursive feature elimination algorithm to obtain a subset of surface electromyography signal features based on the filtered feature matrix. The filtering process for the feature matrix to obtain the filtered feature matrix includes: replacing the current point with the product of N values ​​before and after the current point and a scaling factor; the current point is any point in the feature matrix; and the scaling factor decreases over time. The filtering process for the feature matrix is ​​specifically performed using the following formula: ； filtering the feature matrix; wherein is the i-th value in the filtered feature matrix, N is the filter order, is the j-th value extracted from the feature matrix, the cumulative sum of is 1, q is a proportional coefficient; The filter order and the scaling factor are calculated using a particle swarm optimization algorithm; wherein the particle swarm optimization algorithm calculation process is as follows: The particle number is initialized using the filter order and the scaling factor; Calculate the particle fitness value based on the particle's position; The optimal position of the particle, the optimal position of the particle swarm, and the velocity and position of the particle are updated based on the particle fitness value, the optimal position traversed by the particle, and the optimal position traversed by the particle swarm under the given number of particles. Output final particles according to the updated particle swarm optimal position; the final particles are the filter order and the proportional coefficient; The random forest-recursive feature elimination algorithm is specifically: bootstrap resampling method is used to extract multiple samples from original samples as bootstrap samples, a decision tree is constructed for each bootstrap sample, a random forest is constructed according to all the decision trees, feature importance is calculated in a regression model, a backward iterative feature evaluation is introduced, the feature corresponding to the minimum feature importance is deleted, and then the random forest algorithm is used again to calculate the feature importance of the remaining features, until the last feature is only one.

2. The sEMG signal acquisition apparatus according to claim 1, wherein, The myoelectric signal sensor comprises, from top to bottom, a touch surface layer, a motion trajectory suppression layer, a fabric conductive layer and a substrate; the touch surface layer is used to contact the surface of the skin to be collected to collect an initial surface myoelectric signal; The motion trajectory suppression layer is used to buffer the interaction between the touch surface layer and the surface of the skin to be collected; and the fabric conductive layer is used to transmit the initial surface myoelectric signal.

3. The sEMG signal acquisition apparatus according to claim 2, wherein, The touch surface layer is a knitted silver conductive cloth, the motion trajectory suppression layer is a conductive sponge, and the substrate is an elastic nylon material.

4. The sEMG signal acquisition apparatus according to claim 2, wherein, The myoelectric signal sensor further comprises: An insulating shielding layer is arranged on both sides of the touch surface layer and is used to isolate the surface of the skin to be collected from the fabric conductive layer; the insulating shielding layer is an insulating fabric material with a hollow center.

5. The sEMG signal acquisition apparatus according to claim 1, wherein, The low-pass filter is an RC low-pass filter; the operational amplifier and the analog-to-digital converter are ADS1299 chips; and the wireless transmission module is a WIFI module integrated in a CC3200 chip.

6. A method of recognizing a muscle movement, characterized by, The method comprises: acquiring historical surface myoelectric signals; processing the historical surface myoelectric signals to obtain a surface myoelectric signal feature subset, specifically: normalizing the historical surface myoelectric signals to obtain normalized signals; and filtering the normalized signals to obtain filtered signals, specifically: performing discrete wavelet transformation on the normalized signals to obtain high-frequency signal coefficients and low-frequency signal coefficients; filtering the high-frequency signal coefficients according to a preset threshold to obtain filtered high-frequency signal coefficients; and reconstructing signals by using the low-frequency signal coefficients and the filtered high-frequency signal coefficients to obtain filtered signals; extracting time domain features of the filtered signals to construct a feature matrix; the time domain features include root mean square, variance, wavelength length, zero-crossing point number, average absolute value, maximum fractal length, average energy and autoregressive coefficient; filtering the feature matrix to obtain a filtered feature matrix; using a filtering method based on feature information, that is, considering the existence of isolated points / abnormal values, replacing each point with the product of the previous and subsequent N values and a time-decreasing proportional coefficient; and obtaining a surface myoelectric signal feature subset by using a random forest-recursive feature elimination algorithm according to the filtered feature matrix; constructing a convolutional neural network model; The surface electromyogram feature subset is used to train the convolutional neural network model, so as to obtain a muscle movement recognition model; The muscle movement is recognized according to the muscle movement recognition model; The feature matrix is filtered to obtain a filtered feature matrix, including: using the product of N values before and after the current point and a proportional coefficient to replace the current point; the current point is any point in the feature matrix; the proportional coefficient decreases with time; The filtering process of the feature matrix is specifically using the formula: ； filtering the feature matrix; wherein is the i-th value in the filtered feature matrix, N is the filter order, is the j-th value extracted from the feature matrix, the cumulative sum of is 1, q is a proportional coefficient; The filter order and the proportional coefficient are calculated by using a particle swarm algorithm; the particle swarm algorithm calculation process is as follows: The filter order and the proportional coefficient are used to initialize the number of particles; The particle fitness value is calculated according to the position of the particle; The particle optimal position, the particle swarm optimal position, and the speed and position of the particle are updated according to the particle fitness value, the optimal position of the particle, and the optimal position of the particle swarm under the number of particles; The final particle is output according to the updated particle swarm optimal position; the final particle is the filter order and the proportional coefficient; The random forest-recursive feature elimination algorithm is specifically as follows: bootstrap resampling method is used to extract multiple samples from the original samples as bootstrap samples, a decision tree is constructed for each bootstrap sample, a random forest is constructed according to all the decision trees, the feature importance is calculated in the regression model, the backward iterative feature evaluation is introduced, the feature corresponding to the minimum feature importance is deleted, and then the random forest algorithm is used to calculate the feature importance of the remaining features, until the last feature is only one.

Images