A high-density surface electromyography signal real-time decomposition method, system and terminal

Abstract

The application discloses a high-density surface electromyogram real-time decomposition method, which comprises the following steps: collecting high-density surface electromyogram, constructing an offline data set by using the high-density surface electromyogram; decomposing the high-density surface electromyogram in the offline data set to obtain a motor unit firing sequence; constructing a multi-label deep residual shrinkage network, training the multi-label deep residual shrinkage network to obtain a decomposition model; collecting high-density surface electromyogram to be decomposed, dividing the high-density surface electromyogram to be decomposed into multiple analysis windows according to a collection time axis, inputting each analysis window into the decomposition model, outputting a motor unit firing sequence corresponding to each analysis window, and splicing all the motor unit firing sequences to obtain a final motor unit firing sequence. According to the synergistic effect between different motor unit discharge modes, the data preprocessing step is simplified, and the accuracy and real-time performance of surface electromyogram decomposition are improved.

Description

Technical Field

[0001] The present invention relates to the field of bioelectric signal processing technology, and in particular to a method, system, terminal and computer-readable storage medium for real-time decomposition of high-density surface electromyography signals. Background Art

[0002] Neural interfaces implemented at the central or peripheral nervous system level are crucial for understanding motor neurophysiology and developing human-computer interaction systems. They are a key direction for many major future developments in academia and industry. Compared to direct neural interfaces between the brain and neural tissue, neural interfaces based on non-invasive surface electromyography (sEMG) have the potential to surpass existing solutions in terms of information transfer rate and other aspects.

[0003] The ability to simultaneously record dozens or even hundreds of channels of high-density surface electromyography (HD-sEMG) with rich spatial information through two-dimensional flexible electrode arrays has promoted the development and implementation of sEMG decomposition algorithms. Blind source separation (BSS) algorithms, including convolution kernel compensation (CKC), independent component analysis, and fast spatial pulse detection, have been developed and verified to achieve accurate sEMG decomposition.

[0004] However, surface EMG signal decomposition is currently mainly based on the blind source separation algorithm, which requires signal expansion and whitening transformations with high computational complexity. Although preprocessing steps such as expansion and whitening reduce the computational complexity of gradient descent or fixed-point algorithms in the offline training phase, they lead to large delays in the online decomposition phase and cannot meet the real-time requirements of sEMG decomposition technology in applications such as neural interfaces. In addition, the BSS algorithm is inherently limited by current signal processing knowledge and ignores hidden relationships in the high-dimensional space of sEMG that are difficult for humans to identify, making it difficult to achieve accurate HD-sEMG real-time decomposition.

[0005] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0006] The main purpose of the present invention is to provide a method, device and computer-readable storage medium for real-time decomposition of high-density surface electromyography signals, aiming to solve the technical problem that the surface electromyography signal decomposition method in the existing technology is still in the offline processing stage, cannot meet the real-time requirements in the process of modern scientific and technological development, and has large application limitations.

[0007] To achieve the above object, the present invention provides a method for real-time decomposition of high-density surface electromyographic signals, the method comprising the following steps:

[0008] Collecting high-density surface electromyographic signals during muscle contraction at several moments, and constructing an offline data set using the high-density surface electromyographic signals;

[0009] Decomposing the high-density surface electromyographic signals in the offline data set based on a convolutional blind source separation algorithm to obtain a motor unit firing sequence;

[0010] Constructing a multi-label deep residual shrinkage network, and performing model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signal in the offline data set and the corresponding motor unit firing sequence to obtain a decomposition model;

[0011] The high-density surface electromyographic signals to be decomposed are collected in real time, and the high-density surface electromyographic signals to be decomposed are divided into multiple analysis windows according to the acquisition time axis. Each analysis window is input into the decomposition model, and the motor unit discharge sequence corresponding to each analysis window is output. All the motor unit discharge sequences are spliced ​​to obtain the final motor unit discharge sequence.

[0012] Optionally, in the real-time decomposition method of high-density surface electromyography signals, the step of decomposing the high-density surface electromyography signals in the offline data set based on a convolution blind source separation algorithm comprises:

[0013] The high-density surface electromyography signal in the offline data set is subjected to Butterworth bandpass filtering, and the filtered high-density surface electromyography signal is expanded to obtain an expanded signal:

[0014] x(k)=[x1(k),x2(k),K,x i (k),x C (k)] T ;

[0015]

[0016] Where x(k) represents the high-density surface electromyographic signal when the muscle contracts at time k, i is a positive integer, representing the number of the high-density surface electromyographic signal channel, C represents the total number of high-density surface electromyographic signal channels, and x i (k) represents the surface electromyographic signal of the i-th channel at time k, x C (k) represents the surface electromyographic signal of the last channel at time k, T represents the matrix transpose, Indicates extended signal, R indicates extended parameter;

[0017] De-averaging the extended signal to obtain a de-averaged extended signal;

[0018] A zero-phase analysis whitening transformation is performed on the de-averaged extended signal to obtain a whitened extended signal.

[0019] Optionally, in the high-density surface electromyography signal real-time decomposition method, the steps of de-averaging the extended signal to obtain a de-averaged extended signal; and performing a zero-phase analysis whitening transformation on the de-averaged extended signal to remove the channel spatial correlation of the de-averaged extended signal to obtain a whitened extended signal include:

[0020] De-averaging the extended signal to obtain a de-averaged extended signal, and calculating a covariance matrix of the de-averaged extended signal:

[0021]

[0022] in, represents the de-averaged extended signal, represents the de-averaged extended signal The covariance matrix of , E represents the mathematical expectation, represents the de-averaged extended signal The transposed matrix of

[0023] For the covariance matrix Perform eigenvalue decomposition and calculate the whitened extended signal:

[0024]

[0025]

[0026]

[0027] Among them, U represents the eigenvector matrix, D S represents the eigenvalue diagonal matrix, U T represents the transposed matrix of the eigenvector matrix U, W represents the whitening matrix, and z(k) represents the whitened extended signal.

[0028] Optionally, in the real-time decomposition method of high-density surface electromyography signals, the step of decomposing the high-density surface electromyography signals in the offline data set based on a convolution blind source separation algorithm to obtain a motor unit firing sequence includes:

[0029] Preset the total number of loop iterations to execute the convolution blind source separation algorithm;

[0030] If the number of cycles of the current convolution blind source separation algorithm is less than or equal to the total number of times the convolution blind source separation algorithm is iteratively executed, performing an iterative cycle of the fixed point algorithm to obtain a separation vector;

[0031] An initialization coefficient of variation is defined, a source signal is estimated according to the coefficient of variation and the separation vector, and a motor unit firing sequence is obtained by a K-means clustering algorithm.

[0032] Optionally, in the high-density surface electromyography signal real-time decomposition method, the iterative loop step of the fixed point algorithm includes:

[0033] Preset the total number of loop iterations N to execute the fixed point algorithm, define the number of loops of the current fixed point algorithm as n = 0, and define the initialization separation vector ω i (n) and ω i (n-1), define the initial separation matrix B and the transposed separation matrix B of the separation matrix B T ;

[0034] when When , the number of cycles of the current fixed point algorithm is accumulated n = n + 1, where Represents the separation vector ω i The transposed matrix of (n), ω i (n-1) represents the calculation result of the i-th separation vector (n-1), and Tolx represents the tolerance;

[0035] The formula of the fixed point algorithm is:

[0036] ω i (n) = E{z(k)g[ω i (n-1) T z(k)]}-E{z(k)g′[ω i (n-1) T z(k)]}ω i (n-1);

[0037]

[0038] Where E represents the mathematical expectation, z(k) represents the whitened extended signal, ω i (n-1) T Represents the separation vector ω i The transposed matrix of (n-1), g and g′ represent function operations, G(x) represents the contrast function, g(x) represents the derivative function of G(x), g′(x) represents the derivative function of g(x), cosh(x) represents the hyperbolic cosine function, and e represents a natural constant;

[0039] For the separation vector ω i (n) Orthogonalize and normalize to get the n-th calculation result of the current i-th separation vector:

[0040] ω i (n) = ω i (n)-BB T ω i (n);

[0041]

[0042] when Or when the number of cycles of the current fixed point algorithm n>N, end the cycle of the current fixed point algorithm and obtain the i-th separation vector ω i =ω i (n).

[0043] Optionally, in the high-density surface electromyography signal real-time decomposition method, the steps of defining an initialization coefficient of variation, estimating a source signal based on the coefficient of variation and the separation vector, and obtaining a motor unit firing sequence using a K-means clustering algorithm include:

[0044] The total number of loop iterations to perform source signal estimation is pre-set as L, the number of loop iterations of the current source signal estimation is defined as m = 0, and the coefficient of variation CoV of the mth and m-1th firing sequences of the initialization of the i-th motor unit is defined. i,m and CoV i,m-1 , define the preset threshold TH SIL ;

[0045] When CoV i,m <CoV i,m-1 When , the number of cycles of the current source signal estimation is accumulated m = m + 1, and the i-th source signal is estimated:

[0046]

[0047]

[0048] Among them, ω i (m) represents the ith separation variable of the mth calibration, represents the estimate of the i-th source signal;

[0049] Performing peak detection on the estimate of the i-th source signal to obtain a peak vector, and dividing the elements in the peak vector into class a vector and class b vector based on a K-means clustering algorithm, wherein the class a vector has the largest element mean and the class b vector has the smallest element mean;

[0050] Assigning values ​​to the non-zero elements in the a-type vector to obtain the m-th estimated i-th motor unit firing sequence PT i,m ;

[0051] Order CoV i,m =CoV i,m-1 , calculate the motor unit firing sequence PT i,m The coefficient of variation of the inter-peak interval ISI (CoV) i,m ;

[0052]

[0053] t j ={t:PT i,m (t) = 1};

[0054]

[0055] Among them, SD ISI Mean represents the standard deviation of all inter-peak intervals ISI, ISI represents the mean of all inter-peak intervals ISI, j is a positive integer, t j represents the time point of the jth discharge, and J represents the total number of discharges;

[0056] When CoV i,m ≥CoV i,m-1 Or when the number of cycles of the current source signal estimation m>L, the cycle of the current source signal estimation is ended, and the silhouette coefficient SIL is calculated according to the a-type vector and the b-type vector;

[0057] When SIL>TH SIL When the separation vector ω i =ω i (m+1) is saved into the separation matrix B, and all the motor unit firing sequences are extracted according to the separation matrix B;

[0058] When SIL <TH SIL When the number of cycles of the current convolution blind source separation algorithm is accumulated and the next cycle is performed, until the number of cycles of the current convolution blind source separation algorithm is greater than the total number of times the convolution blind source separation algorithm is iteratively executed, and all the motion unit emission sequences are extracted according to the separation matrix B.

[0059] Optionally, in the high-density surface electromyography signal real-time decomposition method, the step of constructing a multi-label deep residual shrinkage network according to the motion release sequence includes:

[0060] Taking the motion unit emission sequence as the true label, constructing a multi-label deep residual contraction network, wherein the multi-label deep residual contraction network includes a batch normalization layer, a linear rectification function, a residual contraction module, a fully connected layer and a sigmoid activation function;

[0061] The step of performing model training on the multi-label deep residual shrinkage network to obtain a decomposition model includes:

[0062] Offline training and validation of the multi-label deep residual shrinkage network are performed based on a five-fold cross-validation strategy to obtain a decomposition model;

[0063] The model parameters are optimized by minimizing the average loss of all motor unit firing sequences to obtain the decomposition model with the best performance.

[0064] In addition, to achieve the above-mentioned object, the present invention further provides a high-density surface electromyography signal real-time decomposition system, wherein the high-density surface electromyography signal real-time decomposition system comprises:

[0065] A data acquisition module is used to collect high-density surface electromyographic signals during muscle contraction at several moments, and to construct an offline data set using the high-density surface electromyographic signals;

[0066] An offline decomposition module is used to decompose the high-density surface electromyographic signals in the offline data set based on a convolution blind source separation algorithm to obtain a motor unit firing sequence;

[0067] A decomposition model training module is used to construct a multi-label deep residual shrinkage network, and perform model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signals in the offline data set and the corresponding motor unit firing sequences to obtain a decomposition model;

[0068] The result output module is used to collect the high-density surface electromyographic signals to be decomposed in real time, divide the high-density surface electromyographic signals to be decomposed into multiple analysis windows according to the acquisition time axis, input each analysis window into the decomposition model, output the motor unit discharge sequence corresponding to each analysis window, and splice all the motor unit discharge sequences to obtain the final motor unit discharge sequence.

[0069] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a high-density surface electromyography signal real-time decomposition program stored on the memory and runnable on the processor, and when the high-density surface electromyography signal real-time decomposition program is executed by the processor, the steps of the high-density surface electromyography signal real-time decomposition method as described above are implemented.

[0070] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a high-density surface electromyography signal real-time decomposition program, and when the high-density surface electromyography signal real-time decomposition program is executed by the processor, the steps of the high-density surface electromyography signal real-time decomposition method as described above are implemented.

[0071] In the present invention, high-density surface electromyography signals (SEMs) are collected at several moments of muscle contraction, and an offline data set is constructed using the SEMs. The high-density SEMs in the offline data set are decomposed based on a convolution blind source separation algorithm to obtain a motor unit firing sequence. A multi-label deep residual contraction network is constructed, and a model is trained on the multi-label deep residual contraction network according to the high-density SEMs in the offline data set and the corresponding motor unit firing sequence to obtain a decomposition model. The high-density SEMs to be decomposed are collected in real time, and the high-density SEMs to be decomposed are divided into multiple analysis windows according to the acquisition time axis. Each of the analysis windows is input into the decomposition model, and the motor unit firing sequence corresponding to each analysis window is output. All the motor unit firing sequences are spliced ​​to obtain the final motor unit firing sequence. The present invention simplifies the data preprocessing steps based on the synergy between the discharge modes of different motor units, and improves the accuracy and real-time performance of surface electromyography signal decomposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of a preferred embodiment of the method for real-time decomposition of high-density surface electromyographic signals of the present invention;

[0073] Figure 2 Schematic diagram of the multi-label deep residual shrinkage network structure of a preferred embodiment of the high-density surface electromyography signal real-time decomposition method of the present invention;

[0074] Figure 3 1 is a schematic diagram showing the principle of a preferred embodiment of the method for real-time decomposition of high-density surface electromyographic signals of the present invention;

[0075] Figure 4 Schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0077] The real-time decomposition method of high-density surface electromyography signal described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the high-density surface electromyography signal real-time decomposition method includes the following steps:

[0078] Step S10: collecting high-density surface electromyographic signals during muscle contraction at several moments, and constructing an offline data set using the high-density surface electromyographic signals.

[0079] Specifically, a surface electromyographic signal acquisition device and a high-density electrode array are used to collect a high-density surface electromyographic signal from the belly of the muscle to be tested when the muscle contracts at time k, and the high-density surface electromyographic signal is recorded as x(k) = [x1(k), x2(k), K, x i (k),x C (k)] T , where i is a positive integer, representing the number of the high-density surface electromyography signal channel, C represents the total number of high-density surface electromyography signal channels, and x i (k) represents the surface electromyographic signal of the i-th channel at time k, x C (k) represents the surface electromyographic signal of the last channel at time k, T represents the matrix transpose, represents the extended signal, and R represents the extended parameter (ie, the delay factor of the signal extension).

[0080] In the embodiment, S subjects were recruited, each subject lying on a dynamometer equipped with a torque sensor, and performing three maximum isometric plantar flexions with the dominant leg, lasting 3-5 seconds, with a rest of 120 seconds in between. The maximum value obtained from the 250ms moving average window was used as the peak torque MVC (Maximal Voluntary Contraction). Three isometric plantar flexion tasks were performed with the knee fully extended and the ankle angle at 10° (0° means the foot is perpendicular to the calf) at different contraction intensities (10%, 30% and 50% MVC). High-density surface electromyographic signals of the medial gastrocnemius muscle were collected, with the number of channels being C and the center spacing of the electrodes in the electrode array being d. For example, high-density surface electromyographic signals from a public dataset can be used, where s = 9, C = 64, and d = 8 mm. The subjects are recorded as S1-S9, the collected data are saved as an experimental dataset, and the data used for subsequent decomposition model training is divided into an offline dataset.

[0081] Step S20: Decompose the high-density surface electromyographic signals in the offline data set based on a convolutional blind source separation algorithm to obtain a motor unit firing sequence.

[0082] Furthermore, Butterworth bandpass filtering is performed on the high-density surface electromyography signal in the offline data set, and the filtered high-density surface electromyography signal is expanded to obtain an expanded signal:

[0083] x(k)=[x1(k),x2(k),K,x i (k),x C (k)] T ;

[0084]

[0085] Where x(k) represents the high-density surface electromyographic signal when the muscle contracts at time k, i is a positive integer, representing the number of the high-density surface electromyographic signal channel, C represents the total number of high-density surface electromyographic signal channels, and x i (k) represents the surface electromyographic signal of the i-th channel at time k, x C (k) represents the surface electromyographic signal of the last channel at time k, T represents the matrix transpose, Indicates extended signal, R indicates extended parameter;

[0086] De-averaging the extended signal to obtain a de-averaged extended signal;

[0087] A zero-phase analysis whitening transformation is performed on the de-averaged extended signal to obtain a whitened extended signal.

[0088] Furthermore, the extended signal is de-averaged to obtain a de-averaged extended signal, and the covariance matrix of the de-averaged extended signal is calculated:

[0089]

[0090] in, represents the de-averaged extended signal, represents the de-averaged extended signal The covariance matrix of , E represents the mathematical expectation, represents the de-averaged extended signal The transposed matrix of

[0091] For the covariance matrix Perform eigenvalue decomposition and calculate the whitened extended signal:

[0092]

[0093]

[0094]

[0095] Among them, U represents the eigenvector matrix, D S represents the eigenvalue diagonal matrix, U T represents the transposed matrix of the eigenvector matrix U, W represents the whitening matrix, and z(k) represents the whitened extended signal.

[0096] Specifically, the extended signal The signal value of each channel in the extended signal is subtracted from the signal mean of the channel. De-average is performed to obtain the de-averaged extended signal Then, perform zero-phase analysis and whitening transformation to remove the mean-extended signal. The spatial correlation between channels, the specific steps include: calculating the covariance matrix of the de-averaged extended signal Covariance matrix Perform eigenvalue decomposition, expressed as Get the eigenvector matrix U and the eigenvalue diagonal matrix D S ; Calculate the whitening matrix Calculate the whitened extended signal

[0097] Furthermore, the total number of times the convolution blind source separation algorithm is executed in a loop is preset;

[0098] If the number of cycles of the current convolution blind source separation algorithm is less than or equal to the total number of times the convolution blind source separation algorithm is iteratively executed, performing an iterative cycle of the fixed point algorithm to obtain a separation vector;

[0099] An initialization coefficient of variation is defined, a source signal is estimated according to the coefficient of variation and the separation vector, and a motor unit firing sequence is obtained by a K-means clustering algorithm.

[0100] Specifically, the total number of times N of loop iterations to execute the fixed point algorithm is pre-set, n is defined as the number of loops of the current fixed point algorithm and initialized to n = 0, and the separation vector ω is defined and initialized i (n) and ω i (n-1), define the initial separation matrix B and the transposed separation matrix B of the separation matrix B T , and initialize the separation matrix to an empty matrix to store the separation vector.

[0101] when When , the number of cycles of the current fixed point algorithm is accumulated, that is, n = n + 1, which represents the result of the n-th calculation of the i-th separation vector (the separation vector currently calculated), where Represents the separation vector ω i The transposed matrix of (n), ω i (n-1) represents the n-1th calculation result of the ith separation vector (the separation vector obtained by the last calculation), and Tolx represents the tolerance.

[0102] ω i (n) is assigned to ω i (n-1), perform the fixed point algorithm according to the following formula:

[0103] ω i (n) = E{z(k)g[ω i (n-1) T z(k)]}-E{z(k)g′[ω i (n-1) T z(k)]}ωi (n-1);

[0104]

[0105] For the separation vector ω i (n) orthogonalize ω i (n)-BB T ω i (n) is assigned to ω i (n):

[0106] ω i (n) = ω i (n)-BB T ω i (n);

[0107] And standardize it, Assign to ω i (n), get the n-th calculation result of the current i-th separation vector:

[0108]

[0109] Where E represents the mathematical expectation, z(k) represents the whitened extended signal, ω i (n-1) T Represents the separation vector ω i where g and g′ represent function operations, G(x) represents a contrast function, g(x) represents the derivative of G(x), g′(x) represents the derivative of g(x), cosh(x) represents the hyperbolic cosine function, and e represents a natural constant.

[0110] Repeat the adjacent separation vector (ω i (n) and ω i (n-1)) similarity is judged, when Or when the number of cycles of the current fixed point algorithm n>N, end the cycle of the current fixed point algorithm and obtain the i-th separation vector ω i =ω i (n).

[0111] Furthermore, the total number of loop iterations of source signal estimation is pre-set, L, m is defined as the number of loop iterations of the current source signal estimation and initialized to m = 0, and the coefficient of variation CoV of the m-th and m-1-th estimations of the i-th motor unit firing sequence is defined and initialized. i,m and CoV i,m-1 , define the preset threshold TH SIL .

[0112] When the coefficient of variation CoV i,m <CoV i,m-1When , the number of cycles of the current source signal estimation is accumulated, that is, m = m + 1, and the i-th source signal is estimated:

[0113]

[0114]

[0115] Among them, ω i (m) represents the ith separation variable of the mth calibration, represents the estimate of the i-th source signal;

[0116] The square of the i-th source signal Perform peak detection to obtain a peak vector p. Based on the K-means clustering algorithm, the elements in the peak vector p are divided into two categories and stored in two vectors of the same length as the peak vector p. Each vector only retains elements belonging to the same category of peaks (the position remains unchanged), and the remaining elements are assigned to 0; the vector with the largest element mean is recorded as vector a, and the vector with the smallest element mean is recorded as vector b; the elements in vector a that are not 0 are assigned to 1 to obtain the m-th estimated i-th motor unit firing sequence PT i,m .

[0117] Order CoV i,m =CoV i,m-1 , calculate the motor unit firing sequence PT i,m The coefficient of variation CoV of the inter-peak interval ISI (i.e. the distance between two adjacent non-zero elements) i,m ;

[0118] Calculate the coefficient of variation:

[0119] The time point of the jth discharge: t j ={t:PT i,m (t) = 1};

[0120] The i-th separation variable of the m+1-th correction:

[0121] Among them, SD ISI Mean represents the standard deviation of all inter-peak intervals ISI, ISI represents the mean of all inter-peak intervals ISI, j is a positive integer, t j represents the time point of the jth discharge, and J represents the total number of discharges.

[0122] Repeated coefficient of variation (CoV) of adjacent separation vectors i,m and CoV i,m-1 ) to judge, when CoV i,m ≥CoV i,m-1Or when the number of loops m of the current source signal estimation is greater than L, the loop of the current source signal estimation is ended.

[0123] Further, the silhouette coefficient SIL is calculated based on the a-type vector and the b-type vector, using vector a, vector b and source signal Calculate the silhouette coefficient SIL for K-means clustering.

[0124] When SIL>TH SIL When receiving the i-th source signal The separation vector ω i =ω i (m+1) is added and saved into the separation matrix B, and all the motion unit firing sequences are extracted according to the separation matrix B.

[0125] When SIL <TH SIL When , the number of cycles of the current convolution blind source separation algorithm is accumulated, that is, i = i + 1, and the next cycle is performed until the number of cycles of the current convolution blind source separation algorithm is greater than the total number of loop iterations of the convolution blind source separation algorithm, that is, i>M. Finally, the separation matrix B corresponding to the high-density surface electromyography signal in the offline data set is obtained, which is used to extract the firing sequence of all activated motor units.

[0126] In this embodiment, an offline decomposition algorithm verified by simulation signals and experimental signals, namely the convolution blind source separation algorithm, is used to identify the motor units contained in the surface electromyography signal in the offline data set for supervised training of the subsequent decomposition model. For example, the range of the Butterworth bandpass filter is 20-500Hz, and the expansion parameter (i.e., the delay factor of the signal expansion) R is selected as 16; in order to obtain more separation vectors, the number of iterations M is generally large, for example, M=300; when n=0, initialize ω i (n) = ω i (n-1) is the 64 signal values ​​at the i-th peak position of the sum of the squares of all channel signals of the whitened extended signal z(k); the tolerance Tolx is set to 10 -8 , the comparison function is set to G(x) = log(cosh(x)); the number of iterations of the fixed point algorithm N is set to 50; in order to ensure that the first cycle can proceed smoothly, initialize CoV i,m-1 is infinite inf,CoV i,m is 1; the number of source signal estimations L is set to 50; the custom threshold TH SIL The value of the decomposition coefficient was set to 0.9. To ensure the accuracy of the decomposition, only the motor unit firing sequences with a peak signal-to-noise ratio (PNR) > 30 dB were retained. All decomposition steps were implemented in MATLAB R2022b.

[0127] Step S30: construct a multi-label deep residual shrinkage network, and perform model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signal in the offline data set and the corresponding motor unit firing sequence to obtain a decomposition model.

[0128] Specifically, a multi-label deep residual contraction network is built. In this embodiment, Figure 3 As shown in the figure, the designed multi-label deep residual shrinkage network includes a one-dimensional convolutional layer (the number of output channels is 64, the kernel size is 3, and the padding parameter is 1), a batch normalization layer, a linear rectification function, eight residual shrinkage modules, a fully connected layer and a sigmoid activation function; the output channel numbers of the eight residual shrinkage modules are 64, 64, 128, 128, 256, 256, 512, and 512 respectively; the residual shrinkage module consists of a basic residual module and a small attention network; the residual module contains two batch normalization layers, two linear rectification functions, two convolutional layers and a An identity mapping is added to the residual module; a small sub-network combining soft thresholding and attention mechanism is added on the basis of the residual module; the small sub-network consists of two fully connected layers, a batch normalization layer, a linear rectification function and a sigmoid activation function; its input is a one-dimensional vector of the feature map output by the last batch normalization layer in the residual module after absolute operation, global average pooling layer and flattening layer; the batch normalization layer and the linear rectification function are located between the two fully connected layers; the output of the small sub-network is a feature map after soft thresholding and sigmoid activation function to scale the feature value to the range of [0, 1].

[0129] Furthermore, the multi-label deep residual shrinkage network is trained and verified offline to obtain a decomposition model; the present invention adopts a five-fold cross-validation strategy to train the decomposition model to avoid overfitting, improve the generalization ability of the model, and ensure consistent performance on different data sets. A sliding window with a window length of WS and a step size of SS is used to divide the surface electromyography signal in the offline data set into multiple analysis windows. The length of these analysis windows is WS and the width is the number of channels C. For example, the window length WS is set to 20 data points (9.77ms), the step size SS is set to 10 data points (4.88ms), and the learning rate lr is set to 10 -4 , weight decay w = 10 -6 , iteration number e = 100, batch size b = 64; network construction and model training are both implemented in Python 3.8 and PyTorch 1.11.0.

[0130] In this embodiment, the analysis window is randomly divided into five mutually exclusive subsets, three of which are used as training sets, one subset is used as validation set, and the remaining subset is used as test set; the mean and standard deviation of each data channel in all analysis windows in the training set are calculated, and each analysis window in the training set, validation set, and test set is normalized to zero mean to eliminate the deviation between different channels and different samples and accelerate the convergence of the decomposition model; the binary cross entropy loss function is used to measure the loss between the predicted motor unit emission sequence and the motor unit emission sequence obtained by the convolution blind source separation algorithm; by minimizing the average loss of all motor units, the loss function is used to calculate the loss between the predicted motor unit emission sequence and the motor unit emission sequence obtained by the convolution blind source separation algorithm; The model parameters are optimized by using the learning rate lr and weight decay w, and the adaptive random algorithm optimizer is used to update the network parameters. For example, by dynamically adjusting the learning rate, the optimization efficiency is improved to ensure that the model parameters converge quickly to the optimal solution. The number of iterations e and the batch size b are set, and the model training, verification and testing are performed in sequence. The model with the smallest average loss on the verification set is selected as the best model, and the performance of the best model is evaluated using the test set data. The model training is repeated five times to ensure that each subset is used as a test set once, ensuring the reliability and effectiveness of the model in practical applications, and finally obtaining five trained decomposition models.

[0131] Step S40: collect the high-density surface electromyographic signals to be decomposed in real time, and divide the high-density surface electromyographic signals to be decomposed into multiple analysis windows according to the acquisition time axis, input each of the analysis windows into the decomposition model, output the motor unit release sequence corresponding to each of the analysis windows, and splice all the motor unit release sequences to obtain the final motor unit release sequence.

[0132] Specifically, high-density surface electromyography signals during muscle contraction are collected in real time from the belly of the muscle to be tested through surface electromyography signal acquisition equipment and high-density surface electrode arrays, and the acquired surface electromyography signals are divided into analysis windows with a window length WS and a step length SS according to the time axis. In this embodiment, the test set in the five-fold cross-validation is used as a real-time data set, and the sliding window length WS and step length SS of the real-time decomposition are consistent with those during offline training to meet real-time requirements. The surface electromyography signals in the current analysis window are directly input into the five trained decomposition models, and all motor unit firing sequences corresponding to the current analysis window are output by the decomposition model, and the final motor unit firing sequence is obtained according to the majority voting principle; the firing sequences corresponding to each analysis window are spliced ​​and saved in chronological order until there is no new input, and the final prediction result of the motor unit firing sequence is obtained as ST = [st1, st2, K, st w ], st w represents the recognition result of the motor unit firing sequence corresponding to the w-th analysis window.

[0133] In order to quantitatively evaluate the effect of the present invention, the method of the present invention was compared with the existing real-time decomposition method based on convolutional neural networks using experimental data from 9 different subjects S1-S9. All data processing processes were run on MATLAB R2022b and Python 3.8. In the comparative experiment, the time required for the decomposition model obtained by the method of the present invention and the decomposition model obtained by the convolutional neural network method to process a segment of 20 data points of surface electromyographic signals in real time (randomly selecting the data of one subject for calculation) and the accuracy of the predicted release sequence were calculated respectively. Taking the decomposition results of the offline convolution blind source separation algorithm in the present invention as the standard, the accuracy indicators of the decomposition model include the number of correctly identified motor units, accuracy and sensitivity, where the number of correctly identified motor units is defined as the total number of motor units with an error rate of less than 10% between the predicted release sequence and the standard release sequence. The method of the present invention only takes 5.39±0.12ms to process a segment of 20 data points of surface electromyographic signals, while the method based on the convolutional neural network takes 8.40±0.29ms. As shown in Table 1, the real-time decomposition results of 9 subjects are shown, where each value is the average value under three different contraction intensities (10%, 30% and 50% MVC). It can be seen that the offline decomposition algorithm obtained a total of 16.07±6.17 motion units, the method of the present invention correctly identified 11.56±6.24 motion units after real-time processing, and the method based on convolutional neural network only correctly identified 1.30±1.22 motion units. Compared with the real-time decomposition method based on convolutional neural network, the multi-label deep residual shrinkage network provided by the present invention significantly improved the accuracy of motion unit recognition, with the sensitivity increased from 43.36±9.20% to 79.65±24.96%, and the accuracy increased from 58.40±10.46% to 87.56±18.33%.

[0134] The present invention trains a multi-label deep residual shrinkage network decomposition model offline and uses it to process real-time data sets, which can improve the accuracy and real-time performance of surface electromyography signal decomposition, and extract motor unit firing sequences from high-density surface electromyography signals more accurately and in real time, and has broad application prospects.

[0135] Table 1: Decomposition results comparison table

[0136]

[0137] Further, if Figure 3 As shown, based on the above-mentioned high-density surface electromyography signal real-time decomposition method, the present invention also provides a high-density surface electromyography signal real-time decomposition system, wherein the high-density surface electromyography signal real-time decomposition system includes:

[0138] A data acquisition module 51 is used to collect high-density surface electromyographic signals during muscle contraction at several moments, and to construct an offline data set using the high-density surface electromyographic signals;

[0139] An offline decomposition module 52 is used to decompose the high-density surface electromyographic signals in the offline data set based on a convolution blind source separation algorithm to obtain a motor unit firing sequence;

[0140] A decomposition model training module 53 is used to construct a multi-label deep residual shrinkage network, and perform model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signals in the offline data set and the corresponding motor unit firing sequences to obtain a decomposition model;

[0141] The result output module 54 is used to collect the high-density surface electromyographic signals to be decomposed in real time, and divide the high-density surface electromyographic signals to be decomposed into multiple analysis windows according to the acquisition time axis, input each of the analysis windows into the decomposition model, output the motor unit release sequence corresponding to each of the analysis windows, and splice all the motor unit release sequences to obtain the final motor unit release sequence.

[0142] Further, if Figure 4 As shown, based on the above-mentioned high-density surface electromyography signal real-time decomposition method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 4 Only some of the components of the terminal are shown, but it should be understood that implementation of all of the shown components is not required, and more or fewer components may be implemented instead.

[0143] In some embodiments, the memory 20 can be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 can also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Furthermore, the memory 20 can also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 can also be used to temporarily store data that has been output or is to be output. In one embodiment, a high-density surface electromyography signal real-time decomposition program 40 is stored on the memory 20, and the high-density surface electromyography signal real-time decomposition program 40 can be executed by the processor 10, thereby realizing the high-density surface electromyography signal real-time decomposition method in the present application.

[0144] In some embodiments, the processor 10 can be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the high-density surface electromyography signal real-time decomposition method.

[0145] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0146] In one embodiment, when the processor 10 executes the high-density surface electromyography signal real-time decomposition program 40 in the memory 20, the following steps are implemented:

[0147] Collecting high-density surface electromyographic signals during muscle contraction at several moments, and constructing an offline data set using the high-density surface electromyographic signals;

[0148] Decomposing the high-density surface electromyographic signals in the offline data set based on a convolutional blind source separation algorithm to obtain a motor unit firing sequence;

[0149] Constructing a multi-label deep residual shrinkage network, and performing model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signal in the offline data set and the corresponding motor unit firing sequence to obtain a decomposition model;

[0150] The high-density surface electromyographic signals to be decomposed are collected in real time, and the high-density surface electromyographic signals to be decomposed are divided into multiple analysis windows according to the acquisition time axis. Each analysis window is input into the decomposition model, and the motor unit discharge sequence corresponding to each analysis window is output. All the motor unit discharge sequences are spliced ​​to obtain the final motor unit discharge sequence.

[0151] The step of decomposing the high-density surface electromyography signal in the offline data set based on the convolution blind source separation algorithm includes:

[0152] The high-density surface electromyography signal in the offline data set is subjected to Butterworth bandpass filtering, and the filtered high-density surface electromyography signal is expanded to obtain an expanded signal:

[0153] x(k)=[x1(k),x2(k),K,x i (k),x C (k)] T ;

[0154]

[0155] Where x(k) represents the high-density surface electromyographic signal when the muscle contracts at time k, i is a positive integer, representing the number of the high-density surface electromyographic signal channel, C represents the total number of high-density surface electromyographic signal channels, and x i (k) represents the surface electromyographic signal of the i-th channel at time k, x C (k) represents the surface electromyographic signal of the last channel at time k, T represents the matrix transpose, Indicates extended signal, R indicates extended parameter;

[0156] De-averaging the extended signal to obtain a de-averaged extended signal;

[0157] A zero-phase analysis whitening transformation is performed on the de-averaged extended signal to obtain a whitened extended signal.

[0158] The steps of de-averaging the extended signal to obtain a de-averaged extended signal, and performing a zero-phase analysis whitening transform on the de-averaged extended signal to remove channel spatial correlation of the de-averaged extended signal to obtain a whitened extended signal include:

[0159] De-averaging the extended signal to obtain a de-averaged extended signal, and calculating a covariance matrix of the de-averaged extended signal:

[0160]

[0161] in, represents the de-averaged extended signal, represents the de-averaged extended signal The covariance matrix of , E represents the mathematical expectation, represents the de-averaged extended signal The transposed matrix of

[0162] For the covariance matrix Perform eigenvalue decomposition and calculate the whitened extended signal:

[0163]

[0164]

[0165]

[0166] Among them, U represents the eigenvector matrix, D S represents the eigenvalue diagonal matrix, U Trepresents the transposed matrix of the eigenvector matrix U, W represents the whitening matrix, and z(k) represents the whitened extended signal.

[0167] The step of decomposing the high-density surface electromyography signal in the offline data set based on the convolution blind source separation algorithm to obtain the motor unit firing sequence includes:

[0168] Preset the total number of loop iterations to execute the convolution blind source separation algorithm;

[0169] If the number of cycles of the current convolution blind source separation algorithm is less than or equal to the total number of times the convolution blind source separation algorithm is iteratively executed, performing an iterative cycle of the fixed point algorithm to obtain a separation vector;

[0170] An initialization coefficient of variation is defined, a source signal is estimated according to the coefficient of variation and the separation vector, and a motor unit firing sequence is obtained by a K-means clustering algorithm.

[0171] The iterative loop steps of the fixed point algorithm include:

[0172] Preset the total number of loop iterations N to execute the fixed point algorithm, define the number of loops of the current fixed point algorithm as n = 0, and define the initialization separation vector ω i (n) and ω i (n-1), define the initial separation matrix B and the transposed separation matrix B of the separation matrix B T ;

[0173] when When , the number of cycles of the current fixed point algorithm is accumulated n = n + 1, where Represents the separation vector ω i The transposed matrix of (n), ω i (n-1) represents the calculation result of the i-th separation vector (n-1), and Tolx represents the tolerance;

[0174] The formula of the fixed point algorithm is:

[0175] ω i (n) = E{z(k)g[ω i (n-1) T z(k)]}-E{z(k)g′[ω i (n-1) T z(k)]}ω i (n-1);

[0176]

[0177] Where E represents the mathematical expectation, z(k) represents the whitened extended signal, ω i (n-1)T Represents the separation vector ω i The transposed matrix of (n-1), g and g′ represent function operations, G(x) represents the contrast function, g(x) represents the derivative function of G(x), g′(x) represents the derivative function of g(x), cosh(x) represents the hyperbolic cosine function, and e represents a natural constant;

[0178] For the separation vector ω i (n) Orthogonalize and normalize to get the n-th calculation result of the current i-th separation vector:

[0179] ω i (n) = ω i (n)-BB T ω i (n);

[0180]

[0181] when Or when the number of cycles of the current fixed point algorithm n>N, end the cycle of the current fixed point algorithm and obtain the i-th separation vector ω i =ω i (n).

[0182] The steps of defining and initializing a coefficient of variation, estimating a source signal based on the coefficient of variation and the separation vector, and obtaining a motor unit firing sequence using a K-means clustering algorithm include:

[0183] The total number of loop iterations to perform source signal estimation is pre-set as L, the number of loop iterations of the current source signal estimation is defined as m = 0, and the coefficient of variation CoV of the mth and m-1th firing sequences of the initialization of the i-th motor unit is defined. i,m and CoV i,m-1 , define the preset threshold TH SIL ;

[0184] When CoV i,m <CoV i,m-1 When , the number of cycles of the current source signal estimation is accumulated m = m + 1, and the i-th source signal is estimated:

[0185]

[0186]

[0187] Among them, ω i (m) represents the ith separation variable of the mth calibration, represents the estimate of the i-th source signal;

[0188] Performing peak detection on the estimate of the i-th source signal to obtain a peak vector, and dividing the elements in the peak vector into class a vector and class b vector based on a K-means clustering algorithm, wherein the class a vector has the largest element mean and the class b vector has the smallest element mean;

[0189] Assigning values ​​to the non-zero elements in the a-type vector to obtain the m-th estimated i-th motor unit firing sequence PT i,m ;

[0190] Order CoV i,m =CoV i,m-1 , calculate the motor unit firing sequence PT i,m The coefficient of variation of the inter-peak interval ISI (CoV) i,m ;

[0191]

[0192] t j ={t:PT i,m (t) = 1};

[0193]

[0194] Among them, SD ISI Mean represents the standard deviation of all inter-peak intervals ISI, ISI represents the mean of all inter-peak intervals ISI, j is a positive integer, t j represents the time point of the jth discharge, and J represents the total number of discharges;

[0195] When CoV i,m ≥CoV i,m-1 Or when the number of cycles of the current source signal estimation m>L, the cycle of the current source signal estimation is ended, and the silhouette coefficient SIL is calculated according to the a-type vector and the b-type vector;

[0196] When SIL>TH SIL When the separation vector ω i =ω i (m+1) is saved into the separation matrix B, and all the motor unit firing sequences are extracted according to the separation matrix B;

[0197] When SIL <TH SIL When the number of cycles of the current convolution blind source separation algorithm is accumulated and the next cycle is performed, until the number of cycles of the current convolution blind source separation algorithm is greater than the total number of times the convolution blind source separation algorithm is iteratively executed, and all the motion unit emission sequences are extracted according to the separation matrix B.

[0198] The step of constructing a multi-label deep residual shrinkage network according to the motion release sequence includes:

[0199] Taking the motion unit emission sequence as the true label, constructing a multi-label deep residual contraction network, wherein the multi-label deep residual contraction network includes a batch normalization layer, a linear rectification function, a residual contraction module, a fully connected layer and a sigmoid activation function;

[0200] The step of performing model training on the multi-label deep residual shrinkage network to obtain a decomposition model includes:

[0201] Offline training and validation of the multi-label deep residual shrinkage network are performed based on a five-fold cross-validation strategy to obtain a decomposition model;

[0202] The model parameters are optimized by minimizing the average loss of all motor unit firing sequences to obtain the decomposition model with the best performance.

[0203] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a high-density surface electromyography signal real-time decomposition program, and when the high-density surface electromyography signal real-time decomposition program is executed by a processor, the steps of the high-density surface electromyography signal real-time decomposition method described above are implemented.

[0204] In summary, the present invention provides a real-time decomposition method for high-density surface electromyography signals and related equipment, the method comprising: collecting high-density surface electromyography signals during muscle contraction at several moments, and using the high-density surface electromyography signals to construct an offline data set; decomposing the high-density surface electromyography signals in the offline data set based on a convolution blind source separation algorithm to obtain a motor unit release sequence; constructing a multi-label deep residual contraction network, and training the multi-label deep residual contraction network according to the high-density surface electromyography signals in the offline data set and the corresponding motor unit release sequence to obtain a decomposition model; collecting the high-density surface electromyography signals to be decomposed in real time, and dividing the high-density surface electromyography signals to be decomposed into multiple analysis windows according to the acquisition time axis, inputting each analysis window into the decomposition model, outputting the motor unit release sequence corresponding to each analysis window, and splicing all the motor unit release sequences to obtain the final motor unit release sequence. The present invention fully utilizes the synergy between the discharge patterns of different motor units through multi-label learning, and extracts high-dimensional spatiotemporal features in the original high-density surface electromyography signals based on a deep residual shrinkage network, simplifies data preprocessing steps with high computational complexity such as expansion and whitening, and improves the accuracy and real-time performance of surface electromyography signal decomposition.

[0205] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0206] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When the program is executed, it can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0207] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A real-time decomposition method for high-density surface electromyography signals, characterized in that: The high-density surface electromyography signal real-time decomposition method comprises: Collecting high-density surface electromyographic signals during muscle contraction at several moments, and constructing an offline data set using the high-density surface electromyographic signals; Decomposing the high-density surface electromyographic signals in the offline data set based on a convolutional blind source separation algorithm to obtain a motor unit firing sequence; The step of decomposing the high-density surface electromyographic signal in the offline data set based on the convolution blind source separation algorithm to obtain the motor unit firing sequence includes: Preset the total number of loop iterations to execute the convolution blind source separation algorithm; If the number of cycles of the current convolution blind source separation algorithm is less than or equal to the total number of times the convolution blind source separation algorithm is iteratively executed, performing an iterative cycle of the fixed point algorithm to obtain a separation vector; defining an initialization coefficient of variation, estimating a source signal based on the coefficient of variation and the separation vector, and obtaining a motor unit firing sequence using a K-means clustering algorithm; The steps of defining and initializing a coefficient of variation, estimating a source signal based on the coefficient of variation and the separation vector, and obtaining a motor unit firing sequence using a K-means clustering algorithm include: Preset the total number of loop iterations to perform source signal estimation , define the number of cycles of current source signal estimation as , define the initialization Motor unit firing sequence Second and The coefficient of variation and , define the preset threshold ; when When , the number of cycles of the current source signal estimation is accumulated , and estimate the Source signal: ； ； in, Indicates the The first correction Separate variables, Indicates the The estimate of the source signal, express k Whitening expansion signal at the moment; Regarding the Performing peak detection on the estimation of the source signal to obtain a peak vector, and dividing the elements in the peak vector into type A vectors and type B vectors based on the K-means clustering algorithm, wherein the element mean of the type A vector is the largest and the element mean of the type B vector is the smallest; Assign the non-zero elements in the a-type vector to obtain the first The second estimated Motor unit firing sequence ; make , calculate the motor unit firing sequence The inter-peak interval The coefficient of variation ; ； ； ； in, Represents all inter-peak intervals The standard deviation of Represents all inter-peak intervals The mean of is a positive integer, Indicates the The time point of the second discharge, t Represents a point in time on the timeline. Indicates the total number of discharges, express Whitening expansion signal at the moment; when Or the estimated number of cycles of the current source signal When the current source signal estimation cycle ends, the contour coefficient is calculated according to the a-type vector and the b-type vector. ; when When , the vector is separated Save to separation matrix and according to the separation matrix extracting all the motor unit firing sequences; when When the number of cycles of the current convolution blind source separation algorithm is accumulated and the next cycle is performed, until the number of cycles of the current convolution blind source separation algorithm is greater than the total number of cycles of the convolution blind source separation algorithm, and the separation matrix is ​​calculated according to the separation matrix. extracting all the motor unit firing sequences; Constructing a multi-label deep residual shrinkage network, and performing model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signal in the offline data set and the corresponding motor unit firing sequence to obtain a decomposition model; The high-density surface electromyographic signals to be decomposed are collected in real time, and the high-density surface electromyographic signals to be decomposed are divided into multiple analysis windows according to the acquisition time axis. Each analysis window is input into the decomposition model, and the motor unit discharge sequence corresponding to each analysis window is output. All the motor unit discharge sequences are spliced ​​to obtain the final motor unit discharge sequence.

2. The method according to claim 1, characterized in that The step of decomposing the high-density surface electromyography signal in the offline data set based on the convolution blind source separation algorithm includes: The high-density surface electromyography signal in the offline data set is subjected to Butterworth bandpass filtering, and the filtered high-density surface electromyography signal is expanded to obtain an expanded signal: ； ； in, express High-density surface electromyographic signals during muscle contraction. is a positive integer, indicating the number of the high-density surface electromyography signal channel. Indicates the total number of channels of high-density surface electromyography signals, express Moment Surface electromyography signals of channels, express The surface electromyographic signal of the last channel at the moment, represents the matrix transpose, Indicates extended signal, Indicates extended parameters; De-averaging the extended signal to obtain a de-averaged extended signal; A zero-phase analysis whitening transformation is performed on the de-averaged extended signal to obtain a whitened extended signal.

3. The method according to claim 2, characterized in that The steps of de-averaging the extended signal to obtain a de-averaged extended signal; and performing a zero-phase analysis whitening transformation on the de-averaged extended signal to remove channel spatial correlation of the de-averaged extended signal to obtain a whitened extended signal include: De-averaging the extended signal to obtain a de-averaged extended signal, and calculating a covariance matrix of the de-averaged extended signal: ； in, represents the de-averaged extended signal, represents the de-averaged extended signal The covariance matrix of represents the mathematical expectation, represents the de-averaged extended signal The transposed matrix of For the covariance matrix Perform eigenvalue decomposition and calculate the whitened extended signal: ； ； ； in, represents the eigenvector matrix, represents the eigenvalue diagonal matrix, Represents the eigenvector matrix The transposed matrix of represents the whitening matrix, Represents the whitened extended signal.

4. The method according to claim 1, wherein The steps of the iterative loop of the fixed point algorithm include: Preset the total number of loop iterations to execute the fixed point algorithm , define the number of cycles of the current fixed point algorithm as , define the initialization separation vector and , define the initialization separation matrix and separation matrix The transposed matrix of ; when When , the number of cycles of the current fixed point algorithm is accumulated ,in Represents a separation vector The transposed matrix of Indicates the Separation vector The calculation results of times, Indicates tolerance; The formula of the fixed point algorithm is: ； ； in, represents the mathematical expectation, represents the whitened extended signal, Represents a separation vector The transposed matrix of and Represents function operation, represents the contrast function, express The derivative of express The derivative of represents the hyperbolic cosine function, represents a natural constant; For the separation vector Orthogonalize and standardize to get the current Separation vector The calculation results of times: ； ； when or the number of iterations of the current fixed point algorithm When , the current fixed point algorithm cycle ends and the first Separation vector .

5. The method according to claim 1, wherein The step of constructing a multi-label deep residual shrinkage network according to the motion emission sequence includes: Taking the motion unit emission sequence as the true label, constructing a multi-label deep residual contraction network, wherein the multi-label deep residual contraction network includes a batch normalization layer, a linear rectification function, a residual contraction module, a fully connected layer and a sigmoid activation function; The step of performing model training on the multi-label deep residual shrinkage network to obtain a decomposition model includes: Offline training and validation of the multi-label deep residual shrinkage network are performed based on a five-fold cross-validation strategy to obtain a decomposition model; The model parameters are optimized by minimizing the average loss of all motor unit firing sequences to obtain the decomposition model with the best performance.

6. A high-density surface electromyography signal real-time decomposition system, wherein the high-density surface electromyography signal real-time decomposition system is applied to the high-density surface electromyography signal real-time decomposition method according to any one of claims 1 to 5, characterized in that: The high-density surface electromyography signal real-time decomposition system comprises: A data acquisition module is used to collect high-density surface electromyographic signals during muscle contraction at several moments, and to construct an offline data set using the high-density surface electromyographic signals; An offline decomposition module is used to decompose the high-density surface electromyographic signals in the offline data set based on a convolution blind source separation algorithm to obtain a motor unit firing sequence; A decomposition model training module is used to construct a multi-label deep residual shrinkage network, and perform model training on the multi-label deep residual shrinkage network according to the high-density surface electromyography signals in the offline data set and the corresponding motor unit firing sequences to obtain a decomposition model; The result output module is used to collect the high-density surface electromyographic signals to be decomposed in real time, divide the high-density surface electromyographic signals to be decomposed into multiple analysis windows according to the acquisition time axis, input each analysis window into the decomposition model, output the motor unit discharge sequence corresponding to each analysis window, and splice all the motor unit discharge sequences to obtain the final motor unit discharge sequence.

7. A terminal, characterized in that: The terminal includes: a memory, a processor, and a high-density surface electromyography signal real-time decomposition program stored in the memory and runnable on the processor. When the high-density surface electromyography signal real-time decomposition program is executed by the processor, the steps of the high-density surface electromyography signal real-time decomposition method as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a high-density surface electromyography signal real-time decomposition program, which, when executed by a processor, implements the steps of the high-density surface electromyography signal real-time decomposition method according to any one of claims 1 to 5.

Images