Electromyographic signal real-time decomposition method based on attention mechanism and convolutional network

Abstract

The invention provides an attention mechanism and convolutional network-based electromyographic signal real-time decomposition method, which comprises the following steps of: acquiring high-density surface electromyographic signals with different contraction intensities at a plurality of moments, and constructing an offline data set by using a sliding window method; decomposing the high-density surface electromyogram signals in the offline data set into a motion unit issuing time sequence based on a convolution kernel compensation algorithm; constructing a convolutional neural network with an attention mechanism; training a convolutional neural network with an attention mechanism according to the high-density surface electromyogram signals in the offline data set and a corresponding motion unit issuing time sequence to obtain a decomposition model; high-density surface electromyogram signals collected in real time are input into the decomposition model, and a corresponding motion unit issuing time sequence is output. According to the method, the attention mechanism is added in the deep convolutional neural network, so that the calculation cost of the training model is reduced while the decomposition accuracy of the high-density surface electromyogram signals is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioelectric signal processing, and in particular to a real-time decomposition method of electromyographic signals based on an attention mechanism and a convolutional network. Background Art

[0002] The human body's motor control is related to the activation of motor units, which consist of a motor neuron and the muscle fibers it innervates. When the human body controls the body, the motor cortex transmits nerve signals to the motor units to generate force and movement. When measured on the skin surface using a flexible two-dimensional electrode array, high-density surface electromyographic signals can be obtained.

[0003] Non-invasive identification of EMG signals can decompose them into motor unit action potential (MUAP) and motor unit release time series (MUST). Since muscle action potential is generated before muscle force is generated, causing EMG delay (225+50ms), high-density EMG signals can be decomposed to predict movement intention and converted into commands before actual movement. This requires us to quickly and accurately decompose high-density surface EMG signals within the range of EMG delay, provide a seamless connection between machines and humans, and realize real-time, reliable autonomous control between machines and humans.

[0004] At present, the main method for decomposing high-density surface EMG signals into a large number of related motor unit release time series is blind source separation method, including convolution kernel compensation method (CKC), gradient convolution kernel compensation method (gCKC), fast independent component analysis method (fastICA) or the combination and improvement of these methods. The main core of these methods is to mathematically estimate the separation vector of the EMG signal and optimize the separation vector using gradient descent or fixed-point iteration algorithm. However, these methods are all offline and require additional preprocessing operations such as delay extension or whitening of the EMG signal, resulting in a time delay of more than 300ms, which does not meet the requirements of real-time decomposition.

[0005] Recently, some algorithms for real-time decomposition of high-density surface EMG signals have been developed and verified. They are divided into two stages: one is the offline training stage, which obtains the separation vector of the motor unit on a known EMG signal, and the other is the online decomposition stage, which uses the trained separation vector to decompose the new EMG signal in real time. However, these decomposition methods still require complex preprocessing of the new EMG signal in the online decomposition stage. Summary of the invention

[0006] According to the technical problems mentioned in the above background technology, a real-time decomposition method of electromyographic signals based on attention mechanism and convolutional network is provided.

[0007] The technical means adopted by the present invention are as follows:

[0008] A real-time decomposition method of electromyographic signals based on attention mechanism and convolutional network includes the following steps:

[0009] S1, collect high-density surface electromyographic signals of different contraction intensities at several moments, and use the sliding window method to construct an offline data set;

[0010] S2, decomposing the high-density surface electromyographic signal in the offline data set into a motor unit firing time series based on a convolution kernel compensation algorithm;

[0011] S3, constructing a convolutional neural network with an attention mechanism; training the convolutional neural network with an attention mechanism according to the high-density surface electromyography signals and the corresponding motor unit release time series in the offline data set to obtain a decomposition model;

[0012] S4. Input the high-density surface electromyography signal collected in real time into the decomposition model, and output the corresponding motor unit firing time series.

[0013] Furthermore, in S1, the offline data set is constructed using a sliding window technique, including the following steps:

[0014] The high-density surface electromyographic signals at several moments are divided into several window surface electromyographic signal samples by the sliding window method. The window size of the sample represents the number of sampling points contained in each sample, and the step size of the sliding window represents the number of sampling points that the window moves each time in the time series.

[0015] The number of usable samples generated using the sliding window technique is calculated as:

[0016]

[0017] Among them, N represents the total length of the time series, W represents the window size, and S represents the step size.

[0018] Furthermore, according to the method for real-time decomposition of electromyographic signals based on attention mechanism and convolutional network according to claim 1, it is characterized in that the high-density surface electromyographic signals in the offline data set are decomposed into motor unit firing time series based on the convolution kernel compensation algorithm, comprising the following steps:

[0019] S201. Mathematically model the multi-channel high-density surface electromyography signals as a multi-input and output system, wherein the high-density surface electromyography signals of each channel are represented by a convolution mixture of multiple motor unit release time series and corresponding motor unit action potentials:

[0020]

[0021] Among them, xi (n) represents the electromyographic signal of the i-th channel and the n-th sampling point; h ij (l) represents the action potential of the jth motor unit recorded on the i-th channel; s j (n) represents the firing time series of the jth motor unit; ω i (n) represents the noise on the i-th channel; L represents the sampling length of the action potential; J and M represent the number of active motor units and the number of EMG signal channels, respectively;

[0022] The convolution kernel compensation algorithm compensates for the unknown convolution kernel h in S201. ij (l), based on the minimum mean square error method, the firing time series s of the j-th motor unit j (n) is expressed as:

[0023]

[0024] Among them, C xx Represents the cross-correlation matrix; based on the surface electromyography signal, the cross-correlation matrix is ​​calculated as:

[0025] C xx =E(x(n)x T (n);

[0026] in, Represents the cross-correlation vector; based on the surface electromyography signal and the j-th motor unit release time series, the cross-correlation vector is calculated as:

[0027]

[0028] S202, expanding the multi-channel high-density surface electromyographic signal based on a preset expansion factor K; the expanded electromyographic signal is expressed as:

[0029]

[0030] S203, initializing the mutual correlation vector based on the activity index and iteratively updating the mutual correlation vector using the natural gradient descent algorithm; the activity index is initialized as:

[0031]

[0032] The gradient function is set as:

[0033]

[0034] Among them, s j (n) represents the firing time series of the jth motor unit, and the learning rate of the gradient function is set to η = 0.01;

[0035] S204, calculating the motor unit firing time series corresponding to the high-density surface electromyographic signal based on the cross-correlation matrix and the cross-correlation vector, and using the K-mean clustering algorithm to perform cluster analysis on the firing time series of each motor unit to obtain the final firing time series.

[0036] Further, the S204 includes the following steps:

[0037] S2041: Divide the motor unit firing time series into k clusters and define them as C = {C1, C2, .., C k ,}, the classification rule is to minimize the following objective function:

[0038]

[0039] Where x represents the firing time of the motor unit, μ i Represents cluster C i The centroid of , also called the mean vector, is defined as

[0040] S2042: After being divided into k clusters, the silhouette coefficient is used as a relevant indicator to evaluate the clustering effect. The silhouette system is expressed as:

[0041]

[0042] Among them, S i represents the silhouette coefficient of the i-th cluster; a i represents the average distance from the i-th sample data to other sample data in the cluster to which the sample data belongs; b i It represents the minimum average distance from the i-th sample to all sample data in the nearest adjacent cluster.

[0043] Furthermore, S3 comprises the following steps:

[0044] S301, constructing a convolutional neural network with an attention mechanism; the convolutional neural network with an attention mechanism includes: a convolutional layer, a maximum pooling layer, an attention module, a smoothing layer, and a fully connected layer;

[0045] S302, using the obtained final motor unit release time series as the true label, and performing offline training and verification on the convolutional neural network with the attention mechanism based on a three-fold cross-validation strategy to obtain a decomposition model;

[0046] S303, optimizing the model parameters by minimizing the loss function and maximizing the decomposition accuracy, and obtaining the decomposition model with the best performance; the decomposition accuracy is calculated as:

[0047]

[0048] in, TP j FN represents the number of pulses accurately identified by the jth motor unit. j FP represents the number of pulses fired by the jth motor unit that were not recognized. j Represents the number of pulses fired by the jth motor unit that were incorrectly identified.

[0049] Furthermore, the numbers of convolution kernels are 128, 128, 128, and 64 respectively.

[0050] Compared with the prior art, the present invention has the following advantages:

[0051] The present invention provides a real-time decomposition method of electromyographic signals based on an attention mechanism and a convolutional network. The method uses a convolutional neural network with an attention mechanism to directly decompose high-density surface electromyographic signals that have not been preprocessed into the release time series of each motor unit, so that the electromyographic signal decomposition time within a window is within the electromyographic delay range, thereby achieving real-time decomposition of electromyographic signals. In addition, the attention mechanism added to the deep convolutional neural network improves the accuracy of high-density surface electromyographic signal decomposition while reducing the computational cost of the training model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0053] Figure 1 This is a flow chart of a real-time decomposition method of electromyographic signals based on attention mechanism and convolutional network of the present invention.

[0054] Figure 2 This is a schematic diagram of the structure of a convolutional neural network with an attention mechanism for a real-time decomposition method of electromyographic signals based on an attention mechanism and a convolutional network of the present invention.

[0055] Figure 3 This is a result diagram of a preferred embodiment of the real-time decomposition method of electromyographic signals based on attention mechanism and convolutional network of the present invention. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0058] like Figure 1-3 As shown, the present invention provides a real-time decomposition method of electromyographic signals based on an attention mechanism and a convolutional network, comprising the following steps:

[0059] S1, collecting high-density surface electromyographic signals of different contraction intensities at several moments, and constructing an offline data set using a sliding window method; in S1, constructing an offline data set using a sliding window technique includes the following steps:

[0060] The high-density surface electromyographic signals at several moments are divided into several window surface electromyographic signal samples by the sliding window method. The window size of the sample represents the number of sampling points contained in each sample, and the step size of the sliding window represents the number of sampling points that the window moves each time in the time series.

[0061] The number of usable samples generated using the sliding window technique is calculated as:

[0062]

[0063] Among them, N represents the total length of the time series, W represents the window size, and S represents the step size.

[0064] S2, based on the convolution kernel compensation algorithm, decomposing the high-density surface electromyography signal in the offline data set into a motor unit firing time series; based on the convolution kernel compensation algorithm, decomposing the high-density surface electromyography signal in the offline data set into a motor unit firing time series, comprising the following steps:

[0065] S201. Mathematically model the multi-channel high-density surface electromyography signals as a multi-input and output system, wherein the high-density surface electromyography signals of each channel are represented by a convolution mixture of multiple motor unit release time series and corresponding motor unit action potentials:

[0066]

[0067] Among them, x i (n) represents the electromyographic signal of the i-th channel and the n-th sampling point; h ij (l) represents the action potential of the jth motor unit recorded on the i-th channel; s j (n) represents the firing time series of the jth motor unit; ω i (n) represents the noise on the i-th channel; L represents the sampling length of the action potential; J and M represent the number of active motor units and the number of EMG signal channels, respectively;

[0068] The convolution kernel compensation algorithm compensates for the unknown convolution kernel h in S201. ij (l), based on the minimum mean square error method, the firing time series s of the j-th motor unit j (n) is expressed as:

[0069]

[0070] Among them, C xx Represents the cross-correlation matrix; based on the surface electromyography signal, the cross-correlation matrix is ​​calculated as:

[0071] C xx =E(x(n)x T (n);

[0072] in, Represents the cross-correlation vector; based on the surface electromyography signal and the j-th motor unit release time series, the cross-correlation vector is calculated as:

[0073]

[0074] S202, expanding the multi-channel high-density surface electromyographic signal based on a preset expansion factor K; the expanded electromyographic signal is expressed as:

[0075]

[0076] S203, initializing the mutual correlation vector based on the activity index and iteratively updating the mutual correlation vector using the natural gradient descent algorithm; the activity index is initialized as:

[0077]

[0078] The gradient function is set as:

[0079]

[0080] Among them, s j (n) represents the firing time series of the jth motor unit, and the learning rate of the gradient function is set to η = 0.01;

[0081] S204, based on the cross-correlation matrix and the cross-correlation vector, the motor unit firing time series corresponding to the high-density surface electromyographic signal is calculated, and the firing time series of each motor unit is clustered and analyzed using the K-mean clustering algorithm to obtain the final firing time series. The S204 includes the following steps:

[0082] S2041: Divide the motor unit firing time series into k clusters and define them as C = {C1, C2, .., C k ,}, the classification rule is to minimize the following objective function:

[0083]

[0084] Where x represents the firing time of the motor unit, μ i Represents cluster C i The centroid of , also called the mean vector, is defined as

[0085] S2042: After being divided into k clusters, the silhouette coefficient is used as a relevant indicator to evaluate the clustering effect. The silhouette system is expressed as:

[0086]

[0087] Among them, S i represents the silhouette coefficient of the i-th cluster; a i represents the average distance from the i-th sample data to other sample data in the cluster to which the sample data belongs; b i It represents the minimum average distance from the i-th sample to all sample data in the nearest adjacent cluster.

[0088] S3, constructing a convolutional neural network with an attention mechanism; training the convolutional neural network with an attention mechanism according to the high-density surface electromyography signals and the corresponding motor unit release time series in the offline data set to obtain a decomposition model; S3 includes the following steps:

[0089] S301, constructing a convolutional neural network with an attention mechanism; the convolutional neural network with an attention mechanism includes: a convolutional layer, a maximum pooling layer, an attention module, a smoothing layer, and a fully connected layer;

[0090] S302, using the obtained final motor unit release time series as the true label, and performing offline training and verification on the convolutional neural network with the attention mechanism based on a three-fold cross-validation strategy to obtain a decomposition model;

[0091] S303, optimizing the model parameters by minimizing the loss function and maximizing the decomposition accuracy, and obtaining the decomposition model with the best performance; the decomposition accuracy is calculated as:

[0092]

[0093] in, TP j FN represents the number of pulses accurately identified by the jth motor unit. j FP represents the number of pulses fired by the jth motor unit that were not recognized. j Represents the number of pulses fired by the jth motor unit that were incorrectly identified.

[0094] S4. Input the high-density surface electromyography signal collected in real time into the decomposition model, and output the corresponding motor unit firing time series.

[0095] Preferably, in the present application, the number of convolution kernels are 128, 128, 128, and 64 respectively.

[0096] Example 1

[0097] The present invention provides a real-time decomposition method of electromyographic signals based on an attention mechanism and a convolutional network, comprising the following steps:

[0098] Step 1: Collect high-density surface electromyographic signals of different contraction intensities at several moments, and use the sliding window method to construct an offline data set. The number of available samples generated is:

[0099]

[0100] Among them, N represents the total length of the time series, W represents the window size, and S represents the step size.

[0101] Step 2: Decompose the high-density surface electromyographic signals in the offline data set into motor unit firing time series based on a convolution kernel compensation algorithm.

[0102] Step 2.1: The electromyographic signals of J motor units obtained by M channels can be expressed as:

[0103]

[0104] Among them, x i (n) represents the electromyographic signal of the i-th channel and the n-th sampling point; h ij(l) represents the action potential of the jth motor unit recorded on the i-th channel; s j (n) represents the firing time series of the jth motor unit; ω i (n) represents the noise on the i-th channel; L represents the sampling length of the action potential; J and M represent the number of active motor units and the number of EMG signal channels, respectively;

[0105] Step 2.2: Based on the minimum mean square error method, calculate the firing time series of the jth motor unit:

[0106]

[0107] Based on the surface electromyography signal, calculate the cross-correlation matrix C xx :

[0108] C xx =E(x(n)x T (n);

[0109] Based on the surface electromyographic signal and the time series of the jth motor unit, the cross-correlation vector is calculated.

[0110]

[0111] The cross-correlation vector is initialized based on the activity index and iteratively updated using the natural gradient descent algorithm, where:

[0112] The activity index is initialized to:

[0113]

[0114] The gradient function is set as:

[0115]

[0116] The learning rate of the gradient function is set to η = 0.01. The motor unit firing time series corresponding to the high-density surface electromyography signal is calculated through the cross-correlation matrix and cross-correlation vector.

[0117] Step 2.3: Based on the K-mean clustering algorithm, cluster analysis is performed on the firing time series of each motor unit to calculate the final firing time series, and the silhouette coefficient is used as a relevant indicator to evaluate the clustering effect. The silhouette system is expressed as:

[0118]

[0119] Among them, S i represents the silhouette coefficient of the i-th cluster, a i represents the average distance from the i-th sample data to other sample data in the cluster to which the sample data belongs, bi It represents the minimum average distance from the ith sample to all sample data in the nearest adjacent cluster. Only when the silhouette coefficient is greater than or equal to 0.9, the firing time series of the motor unit is considered valid.

[0120] Step 3: Construct a convolutional neural network with an attention mechanism; train the convolutional neural network with an attention mechanism according to the high-density surface electromyography signals and the corresponding motor unit release time series in the offline data set to obtain a decomposition model. Figure 2 shown.

[0121] Step 4: input the real-time collected high-density surface electromyography signal into the decomposition model, and output the corresponding motor unit firing time series.

[0122] The comparison of the offline and real-time decomposition results of the high-density surface electromyographic signals of muscle contraction intensities of 10%, 30%, and 50% is shown in the figure. Figure 3 As shown. The method used for offline decomposition is the convolution kernel compensation algorithm, and the decomposition model of the present invention is used for real-time decomposition. The sampling rate of the electromyographic signal is 2048Hz, one sampling point represents a moment, the sliding window size is 120 sampling points, and the window sliding step is 20 sampling points. 10% MVC surface electromyographic signals decompose 13 motor units, with an average accuracy of 88.5.1% for offline decomposition and 90% for real-time decomposition; 30% of the surface electromyographic signals decompose 10 motor units, with an average accuracy of 82.0% for offline decomposition and 86.7% for real-time decomposition; for 50% of the surface electromyographic signals, 9 motor units are decomposed, with an average accuracy of 80.8% for offline decomposition and 83.5% for real-time decomposition.

[0123] In summary, for the decomposition of high-density surface electromyographic signals, the accuracy of the real-time electromyographic signal decomposition method based on the attention mechanism and convolutional network proposed in the present invention is higher than that of the classic convolution kernel compensation algorithm. In addition, the delay time of the decomposition method proposed in the present invention is 80ms, which is within the boundary of human electromyographic delay, which improves the decomposition accuracy while ensuring that it can be used for real-time decomposition.

[0124] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time decomposition method of electromyographic signals based on attention mechanism and convolutional network, characterized in that: The following steps are involved: S1, collect high-density surface electromyographic signals of different contraction intensities at several moments, and use the sliding window method to construct an offline data set; S2, decomposing the high-density surface electromyographic signal in the offline data set into a motor unit firing time series based on a convolution kernel compensation algorithm; S3, constructing a convolutional neural network with an attention mechanism; training the convolutional neural network with an attention mechanism according to the high-density surface electromyography signals and the corresponding motor unit release time series in the offline data set to obtain a decomposition model; S4. Input the high-density surface electromyography signal collected in real time into the decomposition model, and output the corresponding motor unit firing time series.

2. According to claim 1, a real-time decomposition method of electromyographic signals based on attention mechanism and convolutional network is characterized in that: In S1, the offline data set is constructed using the sliding window technology, including the following steps: The high-density surface electromyographic signals at several moments are divided into several window surface electromyographic signal samples by the sliding window method. The window size of the sample represents the number of sampling points contained in each sample, and the step size of the sliding window represents the number of sampling points that the window moves each time in the time series. The number of usable samples generated using the sliding window technique is calculated as: Among them, N represents the total length of the time series, W represents the window size, and S represents the step size.

3. The method for real-time decomposition of electromyographic signals based on attention mechanism and convolutional network according to claim 1, characterized in that: Based on the convolution kernel compensation algorithm, the high-density surface electromyographic signal in the offline data set is decomposed into a motor unit firing time series, including the following steps: S201. Mathematically model the multi-channel high-density surface electromyography signals as a multi-input and output system, wherein the high-density surface electromyography signals of each channel are represented by a convolution mixture of multiple motor unit release time series and corresponding motor unit action potentials: Among them, x i (n) represents the electromyographic signal of the i-th channel and the n-th sampling point; h ij (l) represents the action potential of the jth motor unit recorded on the i-th channel; s j (n) represents the firing time series of the jth motor unit; ω i (n) represents the noise on the i-th channel; L represents the sampling length of the action potential; J and M represent the number of active motor units and the number of EMG signal channels, respectively; The convolution kernel compensation algorithm compensates for the unknown convolution kernel h in S201. ij (l), based on the minimum mean square error method, the firing time series s of the j-th motor unit j (n) is expressed as: Among them, C xx Represents the cross-correlation matrix; based on the surface electromyography signal, the cross-correlation matrix is ​​calculated as: C xx =E(x(n)x T (n); in, Represents the cross-correlation vector; based on the surface electromyography signal and the j-th motor unit release time series, the cross-correlation vector is calculated as: S202, expanding the multi-channel high-density surface electromyographic signal based on a preset expansion factor K; the expanded electromyographic signal is expressed as: S203, initializing the mutual correlation vector based on the activity index and iteratively updating the mutual correlation vector using the natural gradient descent algorithm; the activity index is initialized as: The gradient function is set as: Among them, s j (n) represents the firing time series of the jth motor unit, and the learning rate of the gradient function is set to η = 0.01; S204, calculating the motor unit firing time series corresponding to the high-density surface electromyographic signal based on the cross-correlation matrix and the cross-correlation vector, and using the K-mean clustering algorithm to perform cluster analysis on the firing time series of each motor unit to obtain the final firing time series.

4. The method for real-time decomposition of electromyographic signals based on attention mechanism and convolutional network according to claim 1, characterized in that: The S204 comprises the following steps: S2041: Divide the motor unit firing time series into k clusters and define them as C = {C1, C2, .., C k ,}, the classification rule is to minimize the following objective function: Where x represents the firing time of the motor unit, μ i Represents cluster C i The centroid of , also called the mean vector, is defined as S2042: After being divided into k clusters, the silhouette coefficient is used as a relevant indicator to evaluate the clustering effect. The silhouette system is expressed as: Among them, S i represents the silhouette coefficient of the i-th cluster; a i represents the average distance from the i-th sample data to other sample data in the cluster to which the sample data belongs; b i It represents the minimum average distance from the i-th sample to all sample data in the nearest adjacent cluster.

5. The method for real-time decomposition of electromyographic signals based on attention mechanism and convolutional network according to claim 1, characterized in that: The S3 comprises the following steps: S301, constructing a convolutional neural network with an attention mechanism; the convolutional neural network with an attention mechanism includes: a convolutional layer, a maximum pooling layer, an attention module, a smoothing layer, and a fully connected layer; S302, using the obtained final motor unit release time series as the true label, and performing offline training and verification on the convolutional neural network with the attention mechanism based on a three-fold cross-validation strategy to obtain a decomposition model; S303, optimizing the model parameters by minimizing the loss function and maximizing the decomposition accuracy, and obtaining the decomposition model with the best performance; the decomposition accuracy is calculated as: in, TP j FN represents the number of pulses accurately identified by the jth motor unit. j FP represents the number of pulses fired by the jth motor unit that were not recognized. j Represents the number of pulses fired by the jth motor unit that were incorrectly identified.

6. The method for real-time decomposition of electromyographic signals based on attention mechanism and convolutional network according to claim 1, characterized in that: The numbers of convolution kernels are 128, 128, 128, and 64 respectively.

Images