Gesture recognition method, system and terminal based on attention residual pulse neural network

Abstract

The present invention discloses a gesture recognition method, system and terminal based on an attention residual pulse neural network. The method comprises: collecting high-density surface electromyographic signal data of a target individual performing a gesture action, preprocessing the high-density surface electromyographic signal data to obtain a plurality of activity segment signals; decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motor unit pulse sequence corresponding to each activity segment signal; building an attention residual pulse neural network, training and testing the attention residual pulse neural network to obtain an electromyographic gesture recognition model; inputting all motor unit pulse sequences into the electromyographic gesture recognition model for recognition, and outputting the electromyographic gesture of the target individual; the present invention improves the accuracy of gesture recognition based on the attention residual pulse neural network and reduces energy consumption, and can promote the practical application of gesture recognition technology in the fields of neural interface and human-computer interaction.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a gesture recognition method, system, terminal and computer-readable storage medium based on an attention residual pulse neural network. Background Art

[0002] Surface electromyogram (sEMG) is acquired through an electrode array or sensor placed on the surface of the skin. It is non-invasive, multi-channel and cost-effective, and is a key physiological signal in neuromuscular research. EMG is composed of the action potentials generated by motor units (MUs) activated by central control, which are superimposed in time and space. Among them, MUs are composed of an α motor neuron and all the muscle fibers it controls. They are the basic components and smallest functional units of the peripheral neuromuscular system and can be used to describe the neural control of muscle contraction and movement formation.

[0003] Gesture recognition is one of the core technologies of myoelectric neural interface or myoelectric control method. Since traditional machine learning algorithms cannot process discrete binary data, the existing gesture recognition method based on attention residual spike neural network based on motor unit discharge information needs to convert the discrete motor unit pulse sequence into continuous data such as numerical features or motor unit action potential waveform, and then extract and filter features from it to achieve gesture recognition. This processing method may cause a large amount of motor unit discharge information to be lost, and fail to fully utilize the rich information in the motor unit pulse sequence, thereby limiting the improvement of classification performance.

[0004] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0005] The main purpose of the present invention is to provide a gesture recognition method, system, terminal and computer-readable storage medium based on an attention residual pulse neural network, aiming to solve the problem in the prior art that traditional machine learning methods cannot process discrete data and cannot fully utilize the timing and spatial dependencies in the motor unit pulse sequence, resulting in insufficient accuracy of gesture recognition based on the attention residual pulse neural network.

[0006] To achieve the above object, the present invention provides a gesture recognition method based on an attention residual spiking neural network, and the gesture recognition method based on an attention residual spiking neural network comprises the following steps:

[0007] Collecting high-density surface electromyography signal data of a target individual performing a gesture action, and preprocessing the high-density surface electromyography signal data to obtain a plurality of activity segment signals;

[0008] Decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal;

[0009] Building an attention residual spiking neural network, training and testing the attention residual spiking neural network, and obtaining an electromyographic gesture recognition model;

[0010] All the motor unit pulse sequences are input into the myoelectric gesture recognition model for recognition, and the myoelectric gesture of the target individual is output.

[0011] Optionally, the gesture recognition method based on the attention residual pulse neural network, wherein the high-density surface electromyography signal data of the target individual performing gesture actions is collected, and the high-density surface electromyography signal data is preprocessed to obtain multiple activity segment signals, specifically including:

[0012] Through the electromyography acquisition equipment and high-density electrodes, high-density surface electromyography signal data containing multiple surface electromyography signals are collected at the target individual's muscles to be tested, performing gestures at time t :

[0013] ;

[0014] in, Indicates the number of the high-density surface electromyography signal channel. Represents the total number of channels of high-density surface electromyography signals, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Represents the surface electromyographic signal of the last channel at time t, Represents matrix transpose;

[0015] Using an n-order Butterworth bandpass filter and a notch filter group, the high-density surface electromyography signal data is preprocessed according to the root mean square value of the high-density surface electromyography signal data to obtain multiple activity segment signals under each channel, wherein the first Active segment signal It is expressed as:

[0016] ;

[0017] in, Indicates the number of the active segment signal, Indicates the first channel Active segment signal, Indicates the second channel Active segment signal, Indicates The first Active segment signal, Indicates the last channel An active segment signal.

[0018] Optionally, the gesture recognition method based on the attention residual pulse neural network, wherein the decomposition of the plurality of activity segment signals by the convolution blind source separation algorithm to obtain the motion unit pulse sequence corresponding to each activity segment signal specifically includes:

[0019] For Active segment signal Expand and get the extended signal , for the extended signal De-average is performed to obtain the de-averaged extended signal ;

[0020] The de-mean extended signal is decomposed by covariance matrix and eigenvalue decomposition. Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal ;

[0021] Define and initialize the separation matrix , No. The separation vector The result of the first loop iteration and the The result of the next loop iteration is based on the current whitening expansion signal. , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. ;

[0022] Initialize the result of the mth correction of the kth separation vector , according to the result of the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , and separate the current correction vector that meets the preset conditions Save to separation matrix middle;

[0023] For The convolution blind source separation algorithm is repeatedly executed on the active segment signal until the preset convolution blind source separation algorithm cycle number K is reached, and the target separation matrix is ​​obtained. :

[0024] ;

[0025] in, represents the corrected separation vector obtained in the first cycle, represents the corrected separation vector obtained in the second cycle, Represents the corrected separation vector obtained in the last cycle;

[0026] Sequentially obtain the separation matrices corresponding to all active segment signals to form a set of separation matrices for all active segment signals :

[0027] ;

[0028] in, Indicates the total number of active segment signals, Indicates The separation matrix corresponding to the active segment signal is Indicates The separation matrix corresponding to the active segment signal is Represents the separation matrix corresponding to the last active segment signal;

[0029] According to the separation matrix set The source signal corresponding to each activity segment signal is calculated from the whitened extended signal corresponding to each activity segment signal, and peak detection and clustering are performed on each of the source signals to obtain a motion unit pulse sequence corresponding to each activity segment signal.

[0030] Optionally, the gesture recognition method based on the attention residual pulse neural network, wherein the de-mean extended signal is decomposed by the covariance matrix and the eigenvalue decomposition. Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal , specifically including:

[0031] Calculate the extended signal The covariance matrix of , for the covariance matrix Perform eigenvalue decomposition to obtain the whitened matrix , according to the whitening matrix Calculate the current whitening extension signal :

[0032] ;

[0033] ;

[0034] ;

[0035] in, represents the mathematical expectation, represents the eigenvector matrix, represents the eigenvalue diagonal matrix.

[0036] Optionally, the gesture recognition method based on the attention residual pulse neural network, wherein the whitening expansion signal is , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. , specifically including:

[0037] The first The result of the next loop iteration Assign to The iteration result of the loop is obtained, and the iteration result after assignment is obtained , and expand the signal according to the current whitening , separation matrix And the iteration result after the assignment Calculate the separation vector for this loop iteration :

[0038] ;

[0039] in, , , , represents the number of the separation vector, represents the result of the nth fixed point algorithm loop iteration for the kth separation vector, represents the result of the (n-1)th fixed point algorithm iteration of the kth separation vector, G(x) represents the contrast function, (x) represents the derivative function of g(x), g(x) represents the derivative function of G(x), and x represents the input of the function;

[0040] For the separation vector Orthogonalize and obtain the orthogonal separation vector ;

[0041] ;

[0042] For the orthogonalized separation vector Standardize and get the standardized separation vector , and the standardized separation vector Assign the value of ;

[0043] ;

[0044] in, represents the L1 norm;

[0045] Repeat The result of the next loop iteration The assignment process is continued until the convergence condition of the fixed point algorithm or the preset first maximum number of loop iterations N is reached, and the initial separation vector is obtained. ;

[0046] Among them, the convergence condition of the fixed point algorithm is: , Indicates tolerance.

[0047] Optionally, the gesture recognition method based on the attention residual pulse neural network, wherein the result of the m-th correction is and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , specifically including:

[0048] According to the result of the mth correction and the current whitening extended signal Calculate the source signal :

[0049] ;

[0050] The source signal is calibrated according to the preset minimum distance between adjacent peaks. Perform peak detection to obtain a peak vector p, use the K-means clustering algorithm to divide the elements in the peak vector p into a first-class vector and a second-class vector, and calculate the silhouette coefficient SIL, select a class of vectors with the largest element mean from the first-class vectors and the second-class vectors as the target vector, and obtain the k-th motor unit pulse sequence of the m-th correction according to the target vector ;

[0051] Calculate the motor unit pulse train The coefficient of variation is calculated based on the inter-peak interval ISI. :

[0052] ;

[0053] in, is the standard deviation of all inter-peak intervals ISI, is the mean of all inter-peak intervals ISI;

[0054] According to the kth motor unit pulse sequence Calculate the time point of the vth discharge , according to the time point Calculate the kth separation vector The result of the calibration :

[0055] ;

[0056] Where v represents the number of discharges, V represents the total number of discharge pulses, Indicates at a point in time The whitening extended signal when

[0057] If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is less than that of the previous cycle, and the number of iterations of this cycle does not reach the preset second maximum number of iterations M, then the first The result of the calibration Assign to The result of the calibration , use the result after assignment to continue loop iteration;

[0058] If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is not less than that of the previous cycle, or the number of iterations of this cycle reaches the preset second maximum number of iterations M, then the first The result of the calibration As the correction separation vector .

[0059] Optionally, in the gesture recognition method based on the attention residual pulse neural network, the preset condition is that the silhouette coefficient SIL is greater than a preset first threshold and the coefficient of variation is smaller than a preset second threshold.

[0060] Optionally, the gesture recognition method based on the attention residual spiking neural network, wherein the attention residual spiking neural network comprises: a spiking convolution layer, a spiking batch normalization layer, four attention spiking residual blocks, a membrane potential update module, an adaptive average pooling layer, a random dropout layer and a fully connected layer;

[0061] The membrane potential update module is used to implement the core mechanism of leaky integrate-trigger neurons, which includes membrane potential accumulation, leakage, pulse emission and reset behavior.

[0062] In addition, to achieve the above-mentioned purpose, the present invention also provides a gesture recognition system based on an attention residual spiking neural network, wherein the gesture recognition system based on an attention residual spiking neural network comprises:

[0063] An electromyographic signal acquisition module is used to collect high-density surface electromyographic signal data of a target individual performing a gesture action, and pre-process the high-density surface electromyographic signal data to obtain a plurality of activity segment signals;

[0064] An activity segment signal decomposition module, used for decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal;

[0065] A recognition model training module is used to build an attention residual spiking neural network, train and test the attention residual spiking neural network, and obtain an electromyographic gesture recognition model;

[0066] The myoelectric gesture recognition module is used to input all the motor unit pulse sequences into the myoelectric gesture recognition model for recognition, and output the myoelectric gesture of the target individual.

[0067] 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 gesture recognition program based on an attention residual pulse neural network stored in the memory and executable on the processor, wherein the gesture recognition program based on the attention residual pulse neural network implements the steps of the gesture recognition method based on the attention residual pulse neural network as described above when executed by the processor.

[0068] 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 gesture recognition program based on an attention residual pulse neural network, and when the gesture recognition program based on the attention residual pulse neural network is executed by a processor, the steps of the gesture recognition method based on the attention residual pulse neural network as described above are implemented.

[0069] In the present invention, high-density surface electromyography signal data of a target individual performing a gesture action is collected, and the high-density surface electromyography signal data is preprocessed to obtain a plurality of activity segment signals; the plurality of activity segment signals are decomposed by a convolution blind source separation algorithm to obtain a motor unit pulse sequence corresponding to each activity segment signal; an attention residual pulse neural network is constructed, and the attention residual pulse neural network is trained and tested to obtain an electromyography gesture recognition model; all the motor unit pulse sequences are input into the electromyography gesture recognition model for recognition, and the electromyography gesture of the target individual is output; the present invention overcomes the limitation that traditional machine learning methods cannot process discrete data, improves the accuracy of gesture recognition based on the attention residual pulse neural network and reduces energy consumption, and can promote the practical application of gesture recognition technology in the fields of neural interfaces and human-computer interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flow chart of a preferred embodiment of the gesture recognition method based on the attention residual pulse neural network of the present invention;

[0071] Figure 2 It is a schematic diagram of the original electromyographic signal of a gesture in a region in the gesture recognition method based on the attention residual pulse neural network of the present invention;

[0072] Figure 3 It is a schematic diagram of the original electromyographic signal of a gesture in another area in the gesture recognition method based on the attention residual pulse neural network of the present invention;

[0073] Figure 4 It is a schematic diagram of dividing the root mean square value activity segment signal in the gesture recognition method based on the attention residual pulse neural network of the present invention;

[0074] Figure 5 It is a structural schematic diagram of an attention residual spiking neural network in a gesture recognition method based on an attention residual spiking neural network of the present invention;

[0075] Figure 6 It is an evaluation index diagram of three subjects under the basic spiking neural network provided in the gesture recognition method based on the attention residual spiking neural network of the present invention;

[0076] Figure 7 It is a graph of evaluation indicators of three subjects under the attention residual spiking neural network provided in the gesture recognition method based on the attention residual spiking neural network of the present invention;

[0077] Figure 8 It is a structural diagram of a preferred embodiment of a gesture recognition system based on an attention residual spike neural network of the present invention;

[0078] Fig. 9Schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0079] The present application provides a gesture recognition method, system and terminal based on an attention residual pulse neural network. In order to make the purpose, technical solution and effect of the present application clearer and more specific, the present application is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0080] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0081] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0082] The gesture recognition method based on the attention residual pulse neural network described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the gesture recognition method based on the attention residual pulse neural network includes the following steps:

[0083] Step S10: collecting high-density surface electromyography signal data of the target individual performing gesture actions, and preprocessing the high-density surface electromyography signal data to obtain multiple activity segment signals.

[0084] Specifically, through the electromyography acquisition equipment and high-density electrodes, high-density surface electromyography signal data containing multiple surface electromyography signals are collected at the target individual's muscles to be tested when the gesture is performed at time t. :

[0085] ;

[0086] in, Indicates the number of the high-density surface electromyography signal channel. Represents the total number of channels of high-density surface electromyography signals, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Represents the surface electromyographic signal of the last channel at time t, Represents matrix transpose.

[0087] It can be understood that the high-density surface electromyographic signal data is collected by electromyographic acquisition equipment and N high-density electrodes (for example, two 32-channel flexible high-density electrodes), and the number of channels of each high-density electrode is , high-density surface electromyographic signal data of G gestures (for example, 35 fine gestures, each gesture lasting 5 seconds, repeated 5 times, and the rest time between two adjacent gestures is 5 seconds) performed at time t on the muscles to be tested of P subjects (for example, on the belly of the radial extensor carpi and the ulnar flexor carpi).

[0088] Further, using an n-order Butterworth bandpass filter and a notch filter group, the high-density surface electromyography signal data is preprocessed according to the root mean square value of the high-density surface electromyography signal data to obtain multiple activity segment signals under each channel, wherein the first Active segment signal It is expressed as:

[0089] ;

[0090] in, Indicates the number of the active segment signal, Indicates the first channel Active segment signal, Indicates the second channel Active segment signal, Indicates The first Active segment signal, Indicates the last channel An active segment signal.

[0091] like Figure 2 , Figure 3 and Figure 4 As shown in the figure, the high-density surface electromyographic signals collected by the two electrodes (respectively Figure 2 and Figure 3The collected original electromyographic signals were saved separately, and the eight-order Butterworth bandpass filter of 10-500 Hz and the notch filter of 50 Hz and its harmonics were used to pre-process the two groups of high-density surface electromyographic signals. The segmentation point of the active segment signal ( Figure 4 ), and use the split point to divide the active segment signal.

[0092] Step S20: Decompose the plurality of activity segment signals by using a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal.

[0093] Specifically, for Active segment signal Expand and get the extended signal , for the extended signal De-average is performed to obtain the de-averaged extended signal .

[0094] Further, the de-meaned extended signal is decomposed by covariance matrix and eigenvalue decomposition. Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal .

[0095] The de-meaning extended signal is decomposed by covariance matrix and eigenvalue Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal , specifically including:

[0096] Calculate the extended signal The covariance matrix of , for the covariance matrix Perform eigenvalue decomposition to obtain the whitened matrix , according to the whitening matrix Calculate the current whitening extension signal : ; ; ;in, represents the mathematical expectation, represents the eigenvector matrix, represents a diagonal matrix of eigenvalues.

[0097] Furthermore, define and initialize the separation matrix , No. The separation vector The result of the first loop iteration and the The result of the next loop iteration is based on the current whitening expansion signal. , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. .

[0098] The whitening extension signal is based on the current , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. , specifically including:

[0099] The first The result of the next loop iteration Assign to The iteration result of the loop is obtained, and the iteration result after assignment is obtained , and expand the signal according to the current whitening , separation matrix And the iteration result after the assignment Calculate the separation vector for this loop iteration :

[0100] ;in, , , , represents the number of the separation vector, represents the result of the nth fixed point algorithm loop iteration for the kth separation vector, represents the result of the (n-1)th fixed point algorithm iteration of the kth separation vector, G(x) represents the contrast function, (x) represents the derivative function of g(x), g(x) represents the derivative function of G(x), and x represents the input of the function.

[0101] For the separation vector Orthogonalize and obtain the orthogonal separation vector ;

[0102] ; For the orthogonal separation vector Standardize and get the standardized separation vector , and the standardized separation vector Assign the value of ; ;

[0103] in, represents the L1 norm.

[0104] Repeat The result of the next loop iteration The assignment process is continued until the convergence condition of the fixed point algorithm or the preset first maximum number of loop iterations N is reached, and the initial separation vector is obtained. .

[0105] Among them, the convergence condition of the fixed point algorithm is: , Indicates tolerance.

[0106] Further, the result of the mth correction of the kth separation vector is initialized , according to the result of the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , and separate the current correction vector that meets the preset conditions Save to separation matrix middle.

[0107] According to the result of the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , specifically including:

[0108] According to the result of the mth correction and the current whitening extended signal Calculate the source signal : .

[0109] The source signal is calibrated according to the preset minimum distance between adjacent peaks. Perform peak detection to obtain a peak vector p, use the K-means clustering algorithm to divide the elements in the peak vector p into a first-class vector and a second-class vector, and calculate the silhouette coefficient SIL, select a class of vectors with the largest element mean from the first-class vectors and the second-class vectors as the target vector, and obtain the k-th motor unit pulse sequence of the m-th correction according to the target vector .

[0110] Calculate the motor unit pulse train The coefficient of variation is calculated based on the inter-peak interval ISI. : ;in, is the standard deviation of all inter-peak intervals ISI, is the mean of all inter-peak intervals (ISI).

[0111] According to the kth motor unit pulse sequence Calculate the time point of the vth discharge , according to the time point Calculate the kth separation vector The result of the calibration : ; Where v represents the number of discharges, V represents the total number of discharge pulses, Indicates at a point in time The whitening extended signal at .

[0112] If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is less than that of the previous cycle, and the number of iterations of this cycle does not reach the preset second maximum number of iterations M, then the first The result of the calibration Assign to The result of the calibration , and continue the loop iteration using the assigned result.

[0113] If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is not less than that of the previous cycle, or the number of iterations of this cycle reaches the preset second maximum number of iterations M, then the first The result of the calibration As the correction separation vector .

[0114] It can be understood that the current correction separation vector is obtained After that, the current correction separation vector that meets the preset conditions Save to separation matrix When the preset condition is that the silhouette coefficient SIL is greater than the preset first threshold value, and the coefficient of variation is smaller than a preset second threshold.

[0115] Furthermore, for The convolution blind source separation algorithm is repeatedly executed on the active segment signal until the preset convolution blind source separation algorithm cycle number K is reached, and the target separation matrix is ​​obtained. :

[0116] ;

[0117] in, represents the corrected separation vector obtained in the first cycle, represents the corrected separation vector obtained in the second cycle, Represents the corrected separation vector obtained in the last cycle.

[0118] Sequentially obtain the separation matrices corresponding to all active segment signals to form a set of separation matrices for all active segment signals : ;in, Indicates the total number of active segment signals, Indicates The separation matrix corresponding to the active segment signal is Indicates The separation matrix corresponding to the active segment signal is Represents the separation matrix corresponding to the last active segment signal.

[0119] Furthermore, according to the separation matrix set The source signal corresponding to each activity segment signal is calculated from the whitened extended signal corresponding to each activity segment signal, and peak detection and clustering are performed on each of the source signals to obtain a motion unit pulse sequence corresponding to each activity segment signal.

[0120] It can be understood that in a specific embodiment, the extended parameter R is set to 31, and the tolerance Tolx is set to 10. -8 , the maximum loop iteration number N of the fixed point algorithm is set to 200, the maximum loop number M of the separation vector correction is set to 200, the loop number K of the convolution blind source separation algorithm is set to 992, and the threshold of the silhouette coefficient SIL is set to 0.80. All decomposition steps are implemented in Python 3.8. The motor unit pulse trains extracted from the same active segment of the two sets of high-density surface electromyography signals are finally merged into one pulse train as attention.

[0121] It should be noted that the convolution blind source separation algorithm in the present invention can also be replaced by other blind source separation algorithms, such as a convolution kernel compensation algorithm and a fast independent component analysis algorithm, which can be selected according to actual conditions.

[0122] Step S30: construct an attention residual spiking neural network, train and test the attention residual spiking neural network, and obtain an electromyographic gesture recognition model.

[0123] like Figure 5 As shown, specifically, the attention residual pulse neural network includes: a pulse convolution layer, a pulse batch normalization layer, four attention pulse residual blocks, a membrane potential update module, an adaptive average pooling layer, a random inactivation layer and a fully connected layer; wherein, the membrane potential update module is used to implement the core mechanism of leaky integral-trigger neurons, and the core mechanism includes membrane potential accumulation, leakage, pulse emission and reset behavior.

[0124] The core mechanism is expressed as follows:

[0125] ;

[0126] in, and are the time instant and neuron index, respectively. For the Neuron current time The membrane potential, is the membrane potential at the previous moment (t-1), indicating the cumulative part of the potential, For the current moment The input stimulus (i.e., the local spatiotemporal features extracted by the spike convolution layer), represents element-wise multiplication, For the current moment The pulse firing state. The pulse firing is determined by the step (Heaviside) function. Exceeding the activation threshold When the pulse ,otherwise . is the updated membrane potential, is the reset value of the membrane potential, represents the exponential decay of potential over time, is the time constant, , and are membrane resistance and membrane capacitance, respectively. is the accumulated membrane potential. This is to prevent the potential from accumulating again after the pulse is released.

[0127] In this embodiment, the neuron activation threshold is exemplarily Set to 0.5, the reset value of the membrane potential Set to 0, exponential decay is set to 0.25, and the total time step TW of the leaky integrate-and-fire neuron is set to 2.

[0128] Furthermore, the attention residual pulse neural network is trained and tested to obtain an electromyographic gesture recognition model, which specifically includes:

[0129] Use a sliding window with a window length of ws and a step size of ss to convert the motor unit pulse train Divided into multiple analysis windows;

[0130] The length of the analysis window is ws, the width is the number of motion units, and the label of each analysis window is the gesture number corresponding to the activity segment signal;

[0131] Randomly dividing the plurality of analysis windows into three mutually exclusive data sets according to a preset ratio, wherein the data sets include a training set, a validation set, and a test set;

[0132] The cross entropy loss function is used to measure the loss between the predicted label and the true label, and the model parameters are optimized by minimizing the loss function value; the learning rate and weight decay are set, and the network parameters are updated using the optimizer; the number of iterations and batch size are set, and the model training, validation, and testing are performed in sequence. The model with the smallest loss function value on the validation set is selected as the best model, and the performance of the best model is evaluated using the test set data to obtain the electromyographic gesture recognition model.

[0133] In this embodiment, the window length ws of the sliding window is set to 200 and the step length ss is set to 100. The ratio of the training set, validation set and test set is 6:2:2. The cross entropy loss function is used to measure the loss between the predicted label and the true label. The network parameters are updated using the weighted adaptive random algorithm optimizer, and the learning rate is set to 10 -4 , weight decay is set to 10 -6 , momentum control parameter Set to 0.9, The number of iterations is set to 400 and the batch size is set to 32.

[0134] like Figure 6 and Figure 7 The test results show that the accuracy, recall rate, sensitivity and F1 score of the electromyography gesture recognition model are better than those of the existing spiking neural network.

[0135] Step S40: input all the motor unit pulse sequences into the myoelectric gesture recognition model for recognition, and output the myoelectric gesture of the target individual.

[0136] Specifically, all the motor unit pulse sequences are input into a terminal having a gesture recognition program based on an attention residual pulse neural network, myoelectric gestures are recognized through the myoelectric gesture recognition model, the myoelectric gestures of the target individual are input, and the myoelectric gestures are sent to corresponding staff members.

[0137] The beneficial effects of the present invention are as follows:

[0138] (1) Using motor unit pulse sequences as input to improve recognition accuracy: The present invention directly uses motor unit pulse sequences as input to the pulse neural network. Compared with the traditional pulse coding method, the motor unit pulse sequences extracted by the present invention are less susceptible to signal non-stationarity and can accurately capture the subtle dynamic changes of gesture movements. In addition, surface electromyography signal decomposition can cope with the sparsity and irregularity of the signal, accurately characterize the neural driving characteristics of different gesture movements, and improve the robustness of gesture recognition.

[0139] (2) Simulation of biological nervous system characteristics based on spiking neural networks: The spiking neural network designed by the present invention can significantly improve the ability to capture temporal and spatial features by combining residual connections and attention mechanisms. The spiking neural network directly simulates the information transmission method of biological neurons, uses motor unit pulse sequences as input, and can truly restore the neural signal transmission and processing mechanism, showing higher biological significance and technological innovation. The multi-dimensional synergy of the attention mechanism in time, space and channel dimensions enables the network to dynamically focus on the most important spatiotemporal features, further improving the accuracy of gesture recognition.

[0140] (3) Low power consumption and efficient computing, suitable for embedded and edge devices: The present invention utilizes the time-driven characteristics of pulse neural networks and performs calculations only when neurons emit pulses, further reducing power consumption and computational complexity. By optimizing convolution calculations into accumulation operations, efficient computing can be achieved in software and hardware. In addition, the membrane potential of background noise is dynamically regulated through the attention mechanism, so that only a small number of key neurons need to be activated, reducing the amount of calculation and system energy consumption.

[0141] It can be seen that the present invention realizes full-link optimization from signal processing, feature extraction to efficient computing through the innovation of technical solutions, which can significantly improve the accuracy, real-time and system applicability of gesture recognition based on attention residual pulse neural network, and can be widely used in wearable devices and other intelligent control fields.

[0142] Furthermore, if Figure 8 As shown, based on the above-mentioned gesture recognition method based on the residual attention pulse neural network, the present invention also provides a gesture recognition system based on the residual attention pulse neural network, wherein the gesture recognition system based on the residual attention pulse neural network includes:

[0143] The electromyographic signal acquisition module 51 is used to collect high-density surface electromyographic signal data of the target individual performing gesture actions, and pre-process the high-density surface electromyographic signal data to obtain multiple activity segment signals;

[0144] An activity segment signal decomposition module 52, used for decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal;

[0145] A recognition model training module 53 is used to build an attention residual pulse neural network, train and test the attention residual pulse neural network, and obtain an electromyographic gesture recognition model;

[0146] The myoelectric gesture recognition module 54 is used to input all the motor unit pulse sequences into the myoelectric gesture recognition model for recognition, and output the myoelectric gesture of the target individual.

[0147] Furthermore, if Fig. 9 As shown, based on the above-mentioned gesture recognition method and system based on attention residual pulse neural network, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Fig. 9 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0148] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may 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 (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 20 may 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 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a gesture recognition program 40 based on an attention residual pulse neural network is stored on the memory 20, and the gesture recognition program 40 based on the attention residual pulse neural network can be executed by the processor 10, thereby realizing the gesture recognition method based on the attention residual pulse neural network in the present application.

[0149] In some embodiments, the processor 10 may 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 gesture recognition method based on the attention residual pulse neural network.

[0150] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface.

[0151] In one embodiment, when the processor 10 executes the gesture recognition program 40 based on the attention residual spiking neural network in the memory 20, the following steps are implemented:

[0152] Collecting high-density surface electromyography signal data of a target individual performing a gesture action, and preprocessing the high-density surface electromyography signal data to obtain a plurality of activity segment signals;

[0153] Decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal;

[0154] Building an attention residual spiking neural network, training and testing the attention residual spiking neural network, and obtaining an electromyographic gesture recognition model;

[0155] All the motor unit pulse sequences are input into the myoelectric gesture recognition model for recognition, and the myoelectric gesture of the target individual is output.

[0156] The collecting of high-density surface electromyographic signal data of the target individual performing gesture actions and preprocessing of the high-density surface electromyographic signal data to obtain multiple activity segment signals specifically include:

[0157] Through the electromyography acquisition equipment and high-density electrodes, high-density surface electromyography signal data containing multiple surface electromyography signals are collected at the target individual's muscles to be tested, performing gestures at time t :

[0158] ;

[0159] in, Indicates the number of the high-density surface electromyography signal channel. Represents the total number of channels of high-density surface electromyography signals, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Represents the surface electromyographic signal of the last channel at time t, Represents matrix transpose;

[0160] Using an n-order Butterworth bandpass filter and a notch filter group, the high-density surface electromyography signal data is preprocessed according to the root mean square value of the high-density surface electromyography signal data to obtain multiple activity segment signals under each channel, wherein the first Active segment signal It is expressed as:

[0161] ;

[0162] in, Indicates the number of the active segment signal, Indicates the first channel Active segment signal, Indicates the second channel Active segment signal, Indicates The first Active segment signal, Indicates the last channel An active segment signal.

[0163] The step of decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal specifically includes:

[0164] For Active segment signal Expand and get the extended signal , for the extended signal De-average is performed to obtain the de-averaged extended signal ;

[0165] The de-mean extended signal is decomposed by covariance matrix and eigenvalue decomposition. Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal ;

[0166] Define and initialize the separation matrix , No. The separation vector The result of the first loop iteration and the The result of the next loop iteration is based on the current whitening expansion signal. , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. ;

[0167] Initialize the result of the mth correction of the kth separation vector , according to the result of the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , and separate the current correction vector that meets the preset conditions Save to separation matrix middle;

[0168] For The convolution blind source separation algorithm is repeatedly executed on the active segment signal until the preset convolution blind source separation algorithm cycle number K is reached, and the target separation matrix is ​​obtained. :

[0169] ;

[0170] in, represents the corrected separation vector obtained in the first cycle, represents the corrected separation vector obtained in the second cycle, Represents the corrected separation vector obtained in the last cycle;

[0171] Sequentially obtain the separation matrices corresponding to all active segment signals to form a set of separation matrices for all active segment signals :

[0172] ;

[0173] in, Indicates the total number of active segment signals, Indicates The separation matrix corresponding to the active segment signal is Indicates The separation matrix corresponding to the active segment signal is Represents the separation matrix corresponding to the last active segment signal;

[0174] According to the separation matrix set The source signal corresponding to each activity segment signal is calculated from the whitened extended signal corresponding to each activity segment signal, and peak detection and clustering are performed on each of the source signals to obtain a motion unit pulse sequence corresponding to each activity segment signal.

[0175] Wherein, the de-meaning extended signal is decomposed by covariance matrix and eigenvalue Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal , specifically including:

[0176] Calculate the extended signal The covariance matrix of , for the covariance matrix Perform eigenvalue decomposition to obtain the whitened matrix , according to the whitening matrix Calculate the current whitening extension signal :

[0177] ;

[0178] ;

[0179] ;

[0180] in, represents the mathematical expectation, represents the eigenvector matrix, represents the eigenvalue diagonal matrix.

[0181] Wherein, the whitening extension signal is based on the current , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. , specifically including:

[0182] The first The result of the next loop iteration Assign to The iteration result of the loop is obtained, and the iteration result after assignment is obtained , and expand the signal according to the current whitening , separation matrix And the iteration result after the assignment Calculate the separation vector for this loop iteration :

[0183] ;

[0184] in, , , , represents the number of the separation vector, represents the result of the nth fixed point algorithm loop iteration for the kth separation vector, represents the result of the (n-1)th fixed point algorithm iteration of the kth separation vector, G(x) represents the contrast function, (x) represents the derivative function of g(x), g(x) represents the derivative function of G(x), and x represents the input of the function;

[0185] For the separation vector Orthogonalize and obtain the orthogonal separation vector ;

[0186] ;

[0187] For the orthogonalized separation vector Standardize and get the standardized separation vector , and the standardized separation vector Assign the value of ;

[0188] ;

[0189] in, represents the L1 norm;

[0190] Repeat The result of the next loop iteration The assignment process is continued until the convergence condition of the fixed point algorithm or the preset first maximum number of loop iterations N is reached, and the initial separation vector is obtained. ;

[0191] Among them, the convergence condition of the fixed point algorithm is: , Indicates tolerance.

[0192] Wherein, the result according to the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , specifically including:

[0193] According to the result of the mth correction and the current whitening extended signal Calculate the source signal :

[0194] ;

[0195] The source signal is calibrated according to the preset minimum distance between adjacent peaks. Perform peak detection to obtain a peak vector p, use the K-means clustering algorithm to divide the elements in the peak vector p into a first-class vector and a second-class vector, and calculate the silhouette coefficient SIL, select a class of vectors with the largest element mean from the first-class vectors and the second-class vectors as the target vector, and obtain the k-th motor unit pulse sequence of the m-th correction according to the target vector ;

[0196] Calculate the motor unit pulse train The coefficient of variation is calculated based on the inter-peak interval ISI. :

[0197] ;

[0198] in, is the standard deviation of all inter-peak intervals ISI, is the mean of all inter-peak intervals ISI;

[0199] According to the kth motor unit pulse sequence Calculate the time point of the vth discharge , according to the time point Calculate the kth separation vector The result of the calibration :

[0200] ;

[0201] Where v represents the number of discharges, V represents the total number of discharge pulses, Indicates at a point in time The whitening extended signal when

[0202] If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is less than that of the previous cycle, and the number of iterations of this cycle does not reach the preset second maximum number of iterations M, then the first The result of the calibration Assign to The result of the calibration , use the result after assignment to continue loop iteration;

[0203] If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is not less than that of the previous cycle, or the number of iterations of this cycle reaches the preset second maximum number of iterations M, then the first The result of the calibration As the correction separation vector .

[0204] The preset condition is that the silhouette coefficient SIL is greater than a preset first threshold and the coefficient of variation is smaller than a preset second threshold.

[0205] Optionally, the gesture recognition method based on the attention residual spiking neural network, wherein the attention residual spiking neural network comprises: a spiking convolution layer, a spiking batch normalization layer, four attention spiking residual blocks, a membrane potential update module, an adaptive average pooling layer, a random dropout layer and a fully connected layer;

[0206] The membrane potential update module is used to implement the core mechanism of leaky integrate-trigger neurons, which includes membrane potential accumulation, leakage, pulse emission and reset behavior.

[0207] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a gesture recognition program based on an attention residual pulse neural network, and when the gesture recognition program based on the attention residual pulse neural network is executed by a processor, the steps of the gesture recognition method based on the attention residual pulse neural network as described above are implemented.

[0208] In summary, the present invention provides a gesture recognition method, system and terminal based on an attention residual pulse neural network. The method includes: collecting high-density surface electromyography signal data of a target individual performing a gesture action, preprocessing the high-density surface electromyography signal data to obtain a plurality of activity segment signals; decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motor unit pulse sequence corresponding to each activity segment signal; building an attention residual pulse neural network, training and testing the attention residual pulse neural network to obtain an electromyographic gesture recognition model; inputting all the motor unit pulse sequences into the electromyographic gesture recognition model for recognition, and outputting the electromyographic gesture of the target individual; the present invention overcomes the limitation that traditional machine learning methods cannot process discrete data, improves the accuracy of gesture recognition based on the attention residual pulse neural network and reduces energy consumption, and can promote the practical application of gesture recognition technology in the fields of neural interfaces and human-computer interaction.

[0209] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0210] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0211] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, 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 gesture recognition method based on attention residual spiking neural network, characterized in that: The gesture recognition method based on the attention residual pulse neural network includes: Collecting high-density surface electromyography signal data of a target individual performing a gesture action, and preprocessing the high-density surface electromyography signal data to obtain a plurality of activity segment signals; Decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal; Building an attention residual spiking neural network, training and testing the attention residual spiking neural network, and obtaining an electromyographic gesture recognition model; Inputting all the motor unit pulse sequences into the myoelectric gesture recognition model for recognition, and outputting the myoelectric gesture of the target individual; The attention residual spiking neural network includes: a spiking convolution layer, a spiking batch normalization layer, four attention spiking residual blocks, a membrane potential update module, an adaptive average pooling layer, a random dropout layer and a fully connected layer; Wherein, the membrane potential update module is used to implement the core mechanism of leaky integrate-trigger neurons, and the core mechanism includes membrane potential accumulation, leakage, pulse emission and reset behavior; The attention pulse residual block includes: a membrane potential update module, a pulse convolution layer, a pulse batch normalization layer and an attention module.

2. The gesture recognition method based on attention residual spiking neural network according to claim 1 is characterized in that: The collecting of high-density surface electromyography signal data of the target individual performing gesture actions and preprocessing the high-density surface electromyography signal data to obtain multiple activity segment signals specifically includes: Through the electromyography acquisition equipment and high-density electrodes, high-density surface electromyography signal data containing multiple surface electromyography signals are collected at the target individual's muscles to be tested, performing gestures at time t : ； in, Indicates the number of the high-density surface electromyography signal channel. Represents the total number of channels of high-density surface electromyography signals, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Indicates the time t Surface electromyography signals of channels, Represents the surface electromyographic signal of the last channel at time t, Represents matrix transpose; Using an n-order Butterworth bandpass filter and a notch filter group, the high-density surface electromyography signal data is preprocessed according to the root mean square value of the high-density surface electromyography signal data to obtain multiple activity segment signals under each channel, wherein the first Active segment signal It is expressed as: ； in, Indicates the number of the active segment signal, Indicates the first channel An active segment signal, Indicates the second channel An active segment signal, Indicates The first An active segment signal, Indicates the last channel An active segment signal.

3. The gesture recognition method based on attention residual spiking neural network according to claim 2 is characterized in that: Decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal specifically includes: For Active segment signal Expand and get the extended signal , for the extended signal De-average is performed to obtain the de-averaged extended signal ; The de-mean extended signal is decomposed by covariance matrix and eigenvalue decomposition. Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal ; Define and initialize the separation matrix , No. The separation vector The result of the first loop iteration and the The result of the next loop iteration is based on the current whitening expansion signal. , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. ; Initialize the result of the mth correction of the kth separation vector , according to the result of the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , and separate the current correction vector that meets the preset conditions Save to separation matrix middle; For The convolution blind source separation algorithm is repeatedly executed on the active segment signal until the preset convolution blind source separation algorithm cycle number K is reached, and the target separation matrix is ​​obtained. : ； in, represents the corrected separation vector obtained in the first cycle, represents the corrected separation vector obtained in the second cycle, represents the corrected separation vector obtained in the last cycle; Sequentially obtain the separation matrices corresponding to all active segment signals to form a set of separation matrices for all active segment signals : ； in, Indicates the total number of active segment signals, Indicates The separation matrix corresponding to the active segment signal is Indicates The separation matrix corresponding to the active segment signal is Represents the separation matrix corresponding to the last active segment signal; According to the separation matrix set The source signal corresponding to each activity segment signal is calculated from the whitened extended signal corresponding to each activity segment signal, and peak detection and clustering are performed on each of the source signals to obtain a motion unit pulse sequence corresponding to each activity segment signal.

4. The gesture recognition method based on attention residual pulse neural network according to claim 3 is characterized in that: The de-meaning extended signal is decomposed by covariance matrix and eigenvalue Perform zero-phase analysis whitening transformation to obtain the current whitening extended signal , specifically including: Calculate the extended signal The covariance matrix of , for the covariance matrix Perform eigenvalue decomposition to obtain the whitened matrix , according to the whitening matrix Calculate the current whitening extension signal : ； ； ； in, represents the mathematical expectation, represents the eigenvector matrix, represents a diagonal matrix of eigenvalues.

5. The gesture recognition method based on attention residual spiking neural network according to claim 4 is characterized in that: The whitening extension signal is based on the current , separation matrix , the said The result of the loop iteration and the The result of the loop iteration is to obtain the initial separation vector through the fixed point algorithm. , specifically including: The first The result of the next loop iteration Assign to The iteration result of the loop is obtained, and the iteration result after assignment is obtained , and expand the signal according to the current whitening , separation matrix And the iteration result after the assignment Calculate the separation vector for this loop iteration : ； in, , , , represents the number of the separation vector, represents the result of the nth fixed point algorithm loop iteration for the kth separation vector, represents the result of the (n-1)th fixed point algorithm iteration of the kth separation vector, G(x) represents the contrast function, (x) represents the derivative function of g(x), g(x) represents the derivative function of G(x), and x represents the input of the function; For the separation vector Orthogonalize and obtain the orthogonal separation vector ; ； For the orthogonalized separation vector Standardize and get the standardized separation vector , and the standardized separation vector Assign the value of ; ； in, represents the L1 norm; Repeat The result of the next loop iteration The assignment process is continued until the convergence condition of the fixed point algorithm or the preset first maximum number of loop iterations N is reached, and the initial separation vector is obtained. ; Among them, the convergence condition of the fixed point algorithm is: , Indicates tolerance.

6. The gesture recognition method based on attention residual spiking neural network according to claim 5, characterized in that: According to the result of the mth correction and the current whitening extended signal For the initial separation vector Perform correction to obtain the current correction separation vector , specifically including: According to the result of the mth correction and the current whitening extended signal Calculate the source signal : ； The source signal is calibrated according to the preset minimum distance between adjacent peaks. Perform peak detection to obtain a peak vector p, use the K-means clustering algorithm to divide the elements in the peak vector p into a first-class vector and a second-class vector, and calculate the silhouette coefficient SIL, select a class of vectors with the largest element mean from the first-class vectors and the second-class vectors as the target vector, and obtain the k-th motor unit pulse sequence of the m-th correction according to the target vector ; Calculate the motor unit pulse train The coefficient of variation is calculated based on the inter-peak interval ISI. : ； in, is the standard deviation of all inter-peak intervals ISI, is the mean of all inter-peak intervals ISI; According to the kth motor unit pulse sequence Calculate the time point of the vth discharge , according to the time point Calculate the kth separation vector The result of the calibration : ； Where v represents the number of discharges, V represents the total number of discharge pulses, Indicates at a point in time The whitening extended signal when If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is less than that of the previous cycle, and the number of iterations of this cycle does not reach the preset second maximum number of iterations M, then the first The result of the calibration Assign to The result of the calibration , use the result after assignment to continue loop iteration; If the coefficient of variation obtained in this round of iteration If the coefficient of variation of the previous cycle is not less than that of the previous cycle, or the number of iterations of this cycle reaches the preset second maximum number of iterations M, then the first The result of the calibration As the correction separation vector .

7. The gesture recognition method based on attention residual spiking neural network according to claim 6, characterized in that: The preset condition is that the silhouette coefficient SIL is greater than a preset first threshold and the coefficient of variation is smaller than a preset second threshold.

8. A gesture recognition system based on an attention residual spiking neural network, characterized in that: The gesture recognition system based on the attention residual spiking neural network includes: An electromyographic signal acquisition module is used to collect high-density surface electromyographic signal data of a target individual performing a gesture action, and pre-process the high-density surface electromyographic signal data to obtain a plurality of activity segment signals; An activity segment signal decomposition module, used for decomposing the plurality of activity segment signals by a convolution blind source separation algorithm to obtain a motion unit pulse sequence corresponding to each activity segment signal; A recognition model training module is used to build an attention residual spiking neural network, train and test the attention residual spiking neural network, and obtain an electromyographic gesture recognition model; The myoelectric gesture recognition module is used to input all the motor unit pulse sequences into the myoelectric gesture recognition model for recognition, and output the myoelectric gesture of the target individual; The attention residual spiking neural network includes: a spiking convolution layer, a spiking batch normalization layer, four attention spiking residual blocks, a membrane potential update module, an adaptive average pooling layer, a random dropout layer and a fully connected layer; Wherein, the membrane potential update module is used to implement the core mechanism of leaky integrate-trigger neurons, and the core mechanism includes membrane potential accumulation, leakage, pulse emission and reset behavior; The attention pulse residual block includes: a membrane potential update module, a pulse convolution layer, a pulse batch normalization layer and an attention module.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a gesture recognition program based on an attention residual pulse neural network stored in the memory and executable on the processor. When the gesture recognition program based on the attention residual pulse neural network is executed by the processor, the steps of the gesture recognition method based on the attention residual pulse neural network as described in any one of claims 1 to 7 are implemented.

Images