A method and system for electromyography signal recognition and classification

Abstract

This invention discloses a method and system for electromyography (EMG) signal recognition and classification, relating to the field of signal recognition and classification technology. It collects EMG signal recognition and classification data, extracts features from the EMG signals, and uses a Transformer model to fuse the time-domain, frequency-domain, and time-frequency two-dimensional spatial dynamic feature data of the EMG signals, outputting the correlation coefficients between various features. Through an EMG signal recognition and classification model and an EMG signal recognition and classification result calibration model, it achieves comprehensive recognition and classification of EMG signals based on feature correlation, improving the accuracy of EMG signal recognition and classification results. This invention solves the problems of insufficient comprehensiveness and accuracy caused by the single classification basis and lack of calibration mechanism in traditional methods, enhancing the intelligence level in the EMG signal recognition and classification process.

Description

Technical Field

[0001] This invention relates to the field of signal recognition and classification technology, specifically to a method and system for electromyography (EMG) signal recognition and classification. Background Technology

[0002] With the convergence of artificial intelligence and biomedical engineering, electromyography (EMG) signals, as bioelectrical signals generated during muscle activity, have shown great application potential in fields such as rehabilitation medicine, human-computer interaction, and sports science. Traditional EMG signal recognition relies on manually designed features, but such methods suffer from problems such as strong subjectivity in feature extraction, high computational complexity, and weak generalization ability, making it difficult to meet the needs of complex movements and significant individual differences. The development of deep learning has brought breakthroughs to EMG signal processing, capturing temporal dependencies, but it still faces challenges such as insufficient data, redundant model parameters, and poor real-time performance. In addition, EMG signals themselves are nonlinear and non-stationary, and are susceptible to electrode displacement and environmental noise interference, further increasing the difficulty of accurate classification. Therefore, it is urgent to study efficient EMG signal recognition and classification methods and systems, improve recognition accuracy and robustness by optimizing feature extraction strategies and improving model architecture, and promote their application in scenarios such as intelligent prosthetic control, rehabilitation robots, and virtual reality interaction.

[0003] Traditional electromyography (EMG) signal recognition and classification techniques rely heavily on feature data to categorize EMG signals. However, these categorizations are based on a single criterion and lack calibration of the results, leading to a lack of comprehensiveness and accuracy in EMG signal recognition and classification. Therefore, this invention aims to address the problem of how to combine EMG signal feature data with other multi-dimensional data to analyze EMG signals and thereby improve the comprehensiveness and accuracy of EMG signal recognition and classification. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for electromyography signal recognition and classification to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, an electromyography (EMG) signal recognition and classification method, comprising the following steps:

[0006] The acquisition of electromyographic signal recognition and classification data, including electromyographic signal data, inertial measurement unit data, and near-infrared spectral data, lays the data foundation for subsequent steps.

[0007] Feature extraction is performed on electromyographic signals to obtain time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of the electromyographic signals;

[0008] By combining electromyography (EMG) signal intensity, inertial measurement unit (IMU) data, and near-infrared spectral data with a multiple linear regression algorithm, an EMG signal recognition and classification result calibration model is constructed, and then the corresponding EMG signal recognition and classification result calibration coefficients are output.

[0009] Based on the principle of the self-attention mechanism algorithm, a Transformer model is constructed, which integrates the time-domain feature data, frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data of electromyography signals, and then outputs the correlation coefficient between various features;

[0010] By combining the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals, and the correlation coefficients between various features, a convolutional neural network algorithm is used to construct an EMG signal recognition and classification model, and output the corresponding EMG signal recognition and classification coefficients.

[0011] The electromyography (EMG) signal recognition and classification coefficients are calibrated using calibration coefficients based on the EMG signal recognition and classification results. This improves the accuracy of the EMG signal recognition and classification results. The EMG signal recognition and classification results are then output by issuing corresponding commands.

[0012] A further improvement to the technical solution of this invention lies in that the process of acquiring electromyographic signal recognition and classification data includes:

[0013] Different types of acquisition devices are deployed to collect electromyographic signal data, inertial measurement unit data, and near-infrared spectral data. These acquisition devices include electromyographic sensors, electromyographic biofeedback devices, accelerometers, gyroscopes, and Train.RED wireless muscle oxygenation monitors.

[0014] Electromyographic signal data includes electromyographic signals and their intensity; inertial measurement unit data includes acceleration and angular velocity of limb movement; near-infrared spectral data includes oxyhemoglobin concentration and deoxyhemoglobin concentration in muscle tissue;

[0015] Specifically, an electromyography (EMG) sensor is used, with a sampling frequency set between 500-2000Hz, to collect EMG signals. The intensity of the EMG signals is collected using a biofeedback device. The Train.RED wireless muscle oxygenation monitor, based on near-infrared spectroscopy technology, emits near-infrared light to penetrate muscle tissue. Combined with a modified Beer-Lambert law and spatially resolved spectroscopy, the concentrations of oxyhemoglobin and deoxyhemoglobin in the muscle tissue are collected.

[0016] The collected electromyography (EMG) signal intensity, inertial measurement unit (IMU) data, and near-infrared spectral data were cleaned and normalized. The EMG signals underwent denoising, rectification, and normalization. Specifically, a bandpass filter was used to remove power frequency interference and motion artifacts from the EMG signals, and a notch filter was used to further suppress noise in specific frequency bands. The denoised EMG signals were then subjected to full-wave rectification to convert bipolar signals into unipolar signals, facilitating subsequent feature extraction. Combined with normalization, the EMG signal amplitude was mapped to the [0, 1] interval, eliminating individual differences and electrode contact impedance variations among EMG signal samples during subsequent feature extraction. The preprocessed EMG signal intensity, IMU data, and near-infrared spectral data were integrated to generate an EMG signal recognition and classification dataset.

[0017] A further improvement to the technical solution of this invention lies in the fact that the process of extracting features from electromyographic signals to obtain time-domain and frequency-domain feature data of the electromyographic signals includes:

[0018] The time-domain characteristics of electromyography (EMG) signals include root mean square (RMS), mean absolute value, zero-crossing rate, waveform length, and variance; the frequency-domain characteristics of EMG signals include median frequency and average power frequency.

[0019] Using temporal feature extraction technology, the segmentation window length was set to 200ms. Based on the set window length, the denoised, rectified, and normalized electromyographic (EMG) signal was divided into several windows. The sampling rate was set, and the number of EMG signal samples contained in each window was determined. The EMG signal samples of each window were obtained. Based on the EMG signal samples of each window and the number of EMG signal samples contained therein, the mean value of the EMG signal was calculated. Combining the EMG signal samples of each window, the mean value of the EMG signal and the number of EMG signal samples contained therein, and the zero-crossing rate indicator function, the root mean square value, mean absolute value, zero-crossing rate, waveform length, and variance of the EMG signal were calculated respectively.

[0020] The power spectral density curve of the electromyography (EMG) signal is obtained by using Fourier transform technology. The power spectral density function of the EMG signal is obtained by combining Welch method and fast Fourier transform. Then, the median frequency and average power frequency of the EMG signal are calculated respectively.

[0021] The time-domain and frequency-domain feature data of the acquired electromyographic signals are integrated into the electromyographic signal recognition and classification dataset.

[0022] A further improvement to the technical solution of this invention lies in the fact that the process of extracting features from electromyographic signals to obtain time-frequency two-dimensional spatial dynamic feature data of electromyographic signals includes:

[0023] The time-frequency two-dimensional spatial dynamic feature data includes the centroid frequency, bandwidth, and time-frequency entropy of the energy distribution on the two-dimensional time-frequency plot of the electromyographic signal;

[0024] Using time-frequency analysis technology, electromyographic signals are converted into two-dimensional time-frequency graphs. Combined with wavelet transform, the time, frequency and corresponding energy density are extracted from the two-dimensional time-frequency graphs.

[0025] The time, frequency and corresponding energy density of the electromyographic signal are extracted from the two-dimensional time-frequency graph. The centroid frequency, bandwidth and time-frequency entropy of the energy distribution on the two-dimensional time-frequency graph of the electromyographic signal are calculated respectively. The centroid frequency, bandwidth and time-frequency entropy of the energy distribution on the two-dimensional time-frequency graph of the electromyographic signal are integrated into the electromyographic signal recognition and classification dataset.

[0026] A further improvement to the technical solution of this invention lies in the process of constructing a calibration model for electromyography (EMG) signal recognition and classification results, and then outputting the corresponding calibration coefficients for EMG signal recognition and classification results, which includes:

[0027] The electromyography signal intensity, inertial measurement unit data, and near-infrared spectral data are extracted from the electromyography signal recognition and classification dataset. The extracted data are then converted into a first training set and a first test set, with a ratio of 8:2.

[0028] Combining the first training set data and the multiple linear regression algorithm, the first training set data is used as input, and the calibration coefficient of the electromyography signal recognition and classification result is used as output. The nonlinear relationship between the electromyography signal intensity, inertial measurement unit data, near-infrared spectral data and the calibration coefficient of the electromyography signal recognition and classification result is learned, and the electromyography signal recognition and classification result calibration model is trained.

[0029] The first test set data is input into the calibration model for electrical signal recognition and classification results. The regression coefficients and intercept terms of the calibration model are adjusted to optimize its performance. The final calibration model for electrical signal recognition and classification results is obtained. Combined with the current electromyographic signal intensity, inertial measurement unit data, and near-infrared spectral data, the corresponding calibration coefficients for electromyographic signal recognition and classification results are output.

[0030] A further improvement to the technical solution of this invention lies in the process of constructing a Transformer model, fusing time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography signals, and then outputting the correlation coefficients between various features, including:

[0031] The correlation coefficients between various features include the correlation coefficient between time-domain feature data and frequency-domain feature data, the correlation coefficient between time-domain feature data and time-frequency two-dimensional spatial dynamic feature data, and the correlation coefficient between frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data;

[0032] The time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals are extracted from the EMG signal recognition and classification dataset. The extracted data are then mapped to the corresponding dimension vector space through a fully connected layer to obtain the time-domain feature sequence, frequency-domain feature sequence, and time-frequency two-dimensional spatial dynamic feature sequence of EMG signals. This achieves the fusion of the time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of EMG signals.

[0033] Based on the self-attention mechanism, the time-domain feature sequence, frequency-domain feature sequence, and time-frequency two-dimensional spatial dynamic feature sequence of the electromyography (EMG) signal are taken as input, and the correlation coefficients between each feature are taken as output. The query matrix, key matrix, and value matrix are generated through linear transformation to obtain attention weights. Combined with the self-attention mechanism, the correlations between time-domain feature data and frequency-domain feature data, the correlations between time-domain feature data and time-frequency two-dimensional spatial dynamic feature data, and the correlations between frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data are learned respectively. A Transformer model is constructed, which combines the time-domain feature sequence, frequency-domain feature sequence, and time-frequency two-dimensional spatial dynamic feature sequence of the current EMG signal to output the corresponding correlation coefficients between each feature. The correlation coefficients between each feature are then integrated into the EMG signal recognition and classification dataset.

[0034] A further improvement to the technical solution of this invention lies in the process of constructing an electromyography (EMG) signal recognition and classification model and outputting corresponding EMG signal recognition and classification coefficients, which includes:

[0035] The time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data, and correlation coefficients between various features of the electromyography signal recognition and classification dataset are extracted. The extracted data is then converted into a second training set and a second test set, with the ratio of the second training set to the second test set being 7:3.

[0036] A convolutional neural network (CNN) architecture is constructed using a CNN algorithm. This CNN architecture includes an input layer, a convolutional layer, and a fully connected layer. The input layer receives time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals from a second training set. The convolutional layer analyzes the correlation coefficients between various features in the second training set data to obtain the correlation between time-domain feature data and frequency-domain feature data, the correlation between time-domain feature data and time-frequency two-dimensional spatial dynamic feature data, and the correlation between frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data. The fully connected layer outputs the corresponding EMG signal recognition and classification coefficients to train the EMG signal recognition and classification model.

[0037] The second test set data is input into the electromyography (EMG) signal recognition and classification model to evaluate its performance. The parameters of the EMG signal recognition and classification model are adjusted and optimized to obtain the final EMG signal recognition and classification model. Combining the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data of the EMG signals in the current EMG signal recognition and classification dataset, as well as the correlation coefficients between various features, the corresponding EMG signal recognition and classification coefficients are output.

[0038] A further improvement to the technical solution of this invention lies in that the electromyography (EMG) signal recognition and classification coefficient is calibrated using the calibration coefficient of the EMG signal recognition and classification result, and the process of evaluating the EMG signal recognition and classification result includes:

[0039] When the calibration coefficient of the electromyography (EMG) signal recognition and classification result is below 0.3, no calibration is performed on the EMG signal recognition and classification coefficient; when the calibration coefficient of the EMG signal recognition and classification result is between 0.3 and 0.5, calibration is performed through... The classification coefficients for electromyography (EMG) signal recognition are calibrated; when the calibration coefficient for EMG signal recognition is greater than 0.5, it is then... The classification coefficients for electromyographic signal recognition were calibrated, among which, The calibrated electromyography (EMG) signal recognition classification coefficients are defined as follows: QC and EP0 are the calibration coefficients and classification coefficients of the EMG signal recognition results, respectively, thus obtaining the calibrated EMG signal recognition classification coefficients.

[0040] When the calibrated electromyography (EMG) signal recognition classification coefficients are between 0~0.1, 0.1~0.2, 0.2~0.3, 0.3~0.4, 0.4~0.5, 0.5~0.6, 0.6~0.7, 0.7~0.8, and 0.8~1, they correspond to resting state EMG signal, systolic EMG signal, voluntary EMG signal, non-voluntary EMG signal, normal EMG signal, abnormal EMG signal, mild fatigue EMG signal, moderate fatigue EMG signal, and severe fatigue EMG signal, respectively. The EMG signal recognition classification results are then obtained.

[0041] A further improvement to the technical solution of this invention lies in that the process of outputting electromyographic signals to identify and classify results by issuing corresponding instructions includes:

[0042] By combining voice feedback commands with text display commands, the voice broadcast and the text display on the terminal interface are controlled respectively, thereby realizing the output of electromyography signal recognition and classification results;

[0043] When the electromyography (EMG) signal identification and classification result is a resting state EMG signal, the voice announces that the current EMG signal is in a resting state, and the terminal interface displays the text "resting state"; when the EMG signal identification and classification result is a contraction state EMG signal, the voice announces that the current EMG signal is in a contraction state, and the terminal interface displays the text "contraction state".

[0044] When the electromyography (EMG) signal identification and classification result is voluntary EMG signal, the voice will announce that the current EMG signal is in a voluntary state, and the terminal interface will display the voluntary state text. When the EMG signal identification and classification result is involuntary EMG signal, the voice will announce that the current EMG signal is in a non-voluntary state, and the terminal interface will display the non-voluntary state text.

[0045] When the electromyography (EMG) signal identification and classification result is a normal EMG signal, the voice announces that the current EMG signal is in a normal state, and the terminal interface displays the text indicating that the signal is normal; when the EMG signal identification and classification result is an abnormal EMG signal, the voice announces that the current EMG signal is in an abnormal state, and the terminal interface displays the text indicating that the signal is abnormal.

[0046] When the electromyography (EMG) signal identification and classification result is a mild fatigue state EMG signal, the voice announces that the current EMG signal is in a mild fatigue state, and the terminal interface displays the text "mild fatigue state"; when the EMG signal identification and classification result is a moderate fatigue state EMG signal, the voice announces that the current EMG signal is in a moderate fatigue state, and the terminal interface displays the text "moderate fatigue state"; when the EMG signal identification and classification result is a severe fatigue state EMG signal, the voice announces that the current EMG signal is in a severe fatigue state, and the terminal interface displays the text "severe fatigue state".

[0047] Secondly, an electromyography (EMG) signal recognition and classification system is provided to implement the aforementioned EMG signal recognition and classification method. The system includes a recognition and classification data acquisition module, a feature extraction module, a feature fusion analysis module, an EMG signal recognition and classification module, and an instruction output module, wherein the modules are communicatively connected.

[0048] The identification and classification data acquisition module collects electromyographic signal identification and classification data, which includes electromyographic signal data, inertial measurement unit data, and near-infrared spectral data, providing a data foundation for subsequent identification, classification, and calibration of electromyographic signals;

[0049] The feature extraction module extracts features from electromyographic signals, acquiring time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data, thus enabling in-depth analysis of electromyographic signals.

[0050] The feature fusion analysis module uses a self-attention mechanism to construct a Transformer model, which fuses the time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography signals, and then outputs the correlation coefficients between various features.

[0051] The electromyography (EMG) signal recognition and classification module is divided into a recognition and classification unit and a result calibration unit, which are used to recognize, classify and calibrate EMG signals to improve the accuracy of EMG signal recognition and classification results.

[0052] The identification and classification unit combines the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals, and the correlation coefficients between various features. It then uses a convolutional neural network algorithm to construct an EMG signal identification and classification model and outputs the corresponding EMG signal identification and classification coefficients.

[0053] The calibration unit uses electromyography (EMG) signal intensity, inertial measurement unit (IMU) data, and near-infrared spectral data, combined with a multiple linear regression algorithm, to construct an EMG signal recognition and classification result calibration model, output corresponding EMG signal recognition and classification result calibration coefficients, calibrate the EMG signal recognition and classification coefficients, and then evaluate the EMG signal recognition and classification results.

[0054] The instruction output module outputs the electromyographic signal recognition and classification results by issuing corresponding instructions.

[0055] The beneficial effects of this invention are as follows: Compared with traditional electromyography (EMG) signal recognition and classification methods and systems, the method and system of this invention closely integrates multi-source data acquisition technology, multi-dimensional feature extraction technology, recognition classification and calibration technology with modern information technology. It accurately acquires the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data, and correlation coefficients between various features of the EMG signal. Using a convolutional neural network, an EMG signal recognition and classification model is constructed, outputting corresponding EMG signal recognition and classification coefficients. Combined with a multiple linear regression algorithm, an EMG signal recognition and classification result calibration model is constructed to calibrate the EMG signal recognition and classification coefficients, thereby evaluating the EMG signal recognition and classification results. This improves the accuracy of the EMG signal recognition and classification results, achieving comprehensive recognition and classification of EMG signals based on feature correlation. It solves the problems of insufficient comprehensiveness and accuracy caused by the strong subjectivity of feature extraction, single classification basis, and lack of calibration mechanism in traditional methods. This ensures that the method of this invention can refine the dynamic monitoring standards for EMG signal recognition and classification within a more precise range, making the monitored data more accurate indicators under the same conditions. The development and application of this method has significantly enhanced the intelligence level of the electromyography signal recognition and classification process. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0057] Figure 1 This is a flowchart of an electromyography signal recognition and classification method according to the present invention;

[0058] Figure 2This is a block diagram of an electromyography signal recognition and classification system according to the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1, as Figure 1 As shown, the present invention provides a method for electromyography signal recognition and classification, which consists of the following steps:

[0061] Step 101: Collect electromyography (EMG) signal recognition and classification data. The EMG signal recognition and classification data includes EMG signal data, inertial measurement unit (IMU) data, and near-infrared spectral data, which lays the data foundation for the implementation of subsequent steps.

[0062] Step 102: Extract features from the electromyographic signal to obtain time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of the electromyographic signal;

[0063] Step 103: Using electromyography signal intensity, inertial measurement unit data, and near-infrared spectral data, combined with a multiple linear regression algorithm, construct a calibration model for electromyography signal recognition and classification results, and then output the corresponding calibration coefficients for electromyography signal recognition and classification results.

[0064] Step 104: Based on the principle of self-attention mechanism algorithm, construct a Transformer model, integrate the time domain feature data, frequency domain feature data and time-frequency two-dimensional spatial dynamic feature data of electromyography signal, and then output the correlation coefficient between various features;

[0065] Step 105: Combining the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data, and correlation coefficients between various features of electromyography (EMG) signals, a convolutional neural network algorithm is used to construct an EMG signal recognition and classification model and output the corresponding EMG signal recognition and classification coefficients.

[0066] Step 106: The electromyography (EMG) signal recognition classification coefficient is calibrated using the calibration coefficient of the EMG signal recognition classification result. The EMG signal recognition classification result is evaluated, which improves the accuracy of the EMG signal recognition classification result. The EMG signal recognition classification result is output by issuing corresponding instructions.

[0067] In step 101, the process of acquiring electromyographic signal recognition and classification data includes:

[0068] Different types of acquisition devices are deployed to collect electromyographic signal data, inertial measurement unit data, and near-infrared spectral data. These acquisition devices include electromyographic sensors, electromyographic biofeedback devices, accelerometers, gyroscopes, and Train.RED wireless muscle oxygenation monitors.

[0069] Electromyographic signal data includes electromyographic signals and their intensity; inertial measurement unit data includes acceleration and angular velocity of limb movement; near-infrared spectral data includes oxyhemoglobin concentration and deoxyhemoglobin concentration in muscle tissue;

[0070] Specifically, an electromyography (EMG) sensor is used, with a sampling frequency set between 500-2000Hz, to collect EMG signals. The intensity of the EMG signals is collected using a biofeedback device. The Train.RED wireless muscle oxygenation monitor, based on near-infrared spectroscopy technology, emits near-infrared light to penetrate muscle tissue. Combined with a modified Beer-Lambert law and spatially resolved spectroscopy, the concentrations of oxyhemoglobin and deoxyhemoglobin in the muscle tissue are collected.

[0071] The collected electromyography (EMG) signal intensity, inertial measurement unit (IMU) data, and near-infrared spectral data were cleaned and normalized. The EMG signals underwent denoising, rectification, and normalization. Specifically, a bandpass filter was used to remove power frequency interference and motion artifacts from the EMG signals, and a notch filter was used to further suppress noise in specific frequency bands. The denoised EMG signals were then subjected to full-wave rectification to convert bipolar signals into unipolar signals, facilitating subsequent feature extraction. Combined with normalization, the EMG signal amplitude was mapped to the [0, 1] interval, eliminating individual differences and electrode contact impedance variations among EMG signal samples during subsequent feature extraction. The preprocessed EMG signal intensity, IMU data, and near-infrared spectral data were integrated to generate an EMG signal recognition and classification dataset.

[0072] Step 102, the process of extracting features from the electromyographic signal to obtain its time-domain and frequency-domain feature data, includes:

[0073] The time-domain characteristics of electromyography (EMG) signals include root mean square (RMS), mean absolute value, zero-crossing rate, waveform length, and variance; the frequency-domain characteristics of EMG signals include median frequency and average power frequency.

[0074] Using temporal feature extraction technology, a segmentation window length of 200ms was set. Based on this window length, the denoised, rectified, and normalized electromyographic (EMG) signal was divided into several windows. A sampling rate was set, and the number of EMG signal samples in each window was determined. EMG signal samples for each window were acquired. Based on the number of EMG signal samples and their respective values ​​in each window, the mean EMG signal was calculated. Combining the number of EMG signal samples, the mean EMG signal, the number of EMG signal samples in each window, and the zero-crossing rate indicator function, the root mean square (RMS), mean absolute value, zero-crossing rate, waveform length, and variance of the EMG signal were calculated. The calculation process is as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] in, , , , and These are the root mean square (RMS), mean absolute value, zero-crossing rate, waveform length, and variance of the electromyographic (EMG) signal. Among them, the RMS and mean absolute values ​​of the EMG signal reflect the average force level of muscle contraction and are positively correlated with muscle activation intensity; the zero-crossing rate of the EMG signal reflects the frequency characteristics of the signal and is related to the speed of muscle contraction; the waveform length of the EMG signal describes the complexity of the signal and reflects the degree of synchronization of muscle fibers; and the variance of the EMG signal measures the degree of fluctuation of the signal. The variance is close to the noise level in the resting state and increases significantly during contraction. and Electromyography (EMG) signal samples for each window; and These represent the mean electromyography (EMG) signal value for each window and the number of EMG signal samples it contains. This is a zero-crossing rate indicator function;

[0082] The power spectral density curve of the electromyography (EMG) signal is obtained using Fourier transform technology. Combined with Welch's method and Fast Fourier Transform, the power spectral density function of the EMG signal is obtained. Then, the median frequency and average power frequency of the EMG signal are calculated. The calculation process includes:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] in, is the power spectral density function of the electromyographic signal; To define the length of the segmented window; This represents the number of split windows; For Fast Fourier Transform, there is one corresponding operation; Time-domain samples of electromyography signals for each window; For window functions; and These are the median frequency and average power frequency of the electromyographic signal, respectively. denoted as the frequency in the power spectral density function of the electromyographic signal.

[0088] The time-domain and frequency-domain feature data of the acquired electromyographic signals are integrated into the electromyographic signal recognition and classification dataset.

[0089] Step 102, the process of extracting features from the electromyographic signal to obtain the time-frequency two-dimensional spatial dynamic feature data of the electromyographic signal, includes:

[0090] The time-frequency two-dimensional spatial dynamic feature data includes the centroid frequency, bandwidth, and time-frequency entropy of the energy distribution on the two-dimensional time-frequency plot of the electromyographic signal;

[0091] Time-frequency analysis techniques are used to convert electromyography (EMG) signals into a two-dimensional time-frequency graph. Combined with wavelet transform, the time, frequency, and corresponding energy density are extracted from the two-dimensional time-frequency graph. It's important to note that EMG signals are typically non-stationary signals, with their frequency components varying over time. Therefore, time-frequency analysis is necessary to convert the one-dimensional EMG time-domain signal into a two-dimensional time-frequency graph for subsequent feature extraction. The process of extracting the time, frequency, and corresponding energy density from the two-dimensional time-frequency graph using wavelet transform includes: wavelet transform performs multi-scale analysis of the two-dimensional time-frequency graph of the EMG signal through scaling and translation of the mother wavelet function. The expressions involved include: ,in, , and This represents the time, frequency, and corresponding energy density of a two-dimensional time-frequency plot of an electromyographic signal. For the mother wavelet function, This is a time-domain sample of electromyography (EMG) signals;

[0092] The time, frequency, and corresponding energy density of the energy distribution on the two-dimensional time-frequency plot of the electromyography (EMG) signal are extracted. The centroid frequency, bandwidth, and time-frequency entropy of the energy distribution on the two-dimensional time-frequency plot of the EMG signal are calculated respectively. These parameters are then integrated into an EMG signal recognition and classification dataset. The formulas for calculating the centroid frequency, bandwidth, and time-frequency entropy of the energy distribution on the two-dimensional time-frequency plot of the EMG signal are as follows:

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] in, , and These represent the centroid frequency, bandwidth, and time-frequency entropy of the energy distribution on the two-dimensional time-frequency plot of the electromyographic signal, respectively. The probability distribution function of the two-dimensional time-frequency plot of the electromyographic signal; , and This represents the time, frequency, and corresponding energy density of a two-dimensional time-frequency plot of an electromyographic signal.

[0098] Step 103, the process of constructing a calibration model for electromyography (EMG) signal recognition and classification results, and then outputting the corresponding calibration coefficients for EMG signal recognition and classification results, includes:

[0099] The electromyography signal intensity, inertial measurement unit data, and near-infrared spectral data are extracted from the electromyography signal recognition and classification dataset. The extracted data are then converted into a first training set and a first test set, with a ratio of 8:2.

[0100] Combining the first training set data and the multiple linear regression algorithm, the first training set data is used as input, and the calibration coefficient of the electromyography signal recognition and classification result is used as output. The nonlinear relationship between the electromyography signal intensity, inertial measurement unit data, near-infrared spectral data and the calibration coefficient of the electromyography signal recognition and classification result is learned, and the electromyography signal recognition and classification result calibration model is trained.

[0101] The first test set data is input into the electrical signal recognition and classification result calibration model. The regression coefficients and intercept terms of the electrical signal recognition and classification result calibration model are adjusted to optimize its performance. The final electrical signal recognition and classification result calibration model is obtained. Combining the current electromyographic signal intensity, inertial measurement unit data, and near-infrared spectral data, the corresponding electromyographic signal recognition and classification result calibration coefficients are output. The expression of this electrical signal recognition and classification result calibration model is as follows:

[0102] ;

[0103] in, Calibrate coefficients for electromyography signal recognition and classification results. , , , and These are the regression coefficients for electromyographic signal intensity, acceleration and angular velocity of limb movement, and concentrations of oxyhemoglobin and deoxyhemoglobin in muscle tissue, respectively. , , , and These are the electromyographic signal intensity, the acceleration and angular velocity of limb movement, and the concentrations of oxyhemoglobin and deoxyhemoglobin in muscle tissue. and These are the intercept term and error term of the calibration model for the electrical signal recognition and classification results, respectively.

[0104] Step 104 involves constructing a Transformer model, fusing time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography signals, and then outputting the correlation coefficients between various features.

[0105] The correlation coefficients between various features include the correlation coefficient between time-domain feature data and frequency-domain feature data, the correlation coefficient between time-domain feature data and time-frequency two-dimensional spatial dynamic feature data, and the correlation coefficient between frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data;

[0106] The time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals are extracted from the EMG signal recognition and classification dataset. The extracted data are then mapped to the corresponding dimension vector space through a fully connected layer to obtain the time-domain feature sequence, frequency-domain feature sequence, and time-frequency two-dimensional spatial dynamic feature sequence of EMG signals. This achieves the fusion of the time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of EMG signals.

[0107] Based on the self-attention mechanism, the time-domain feature sequence, frequency-domain feature sequence, and time-frequency two-dimensional spatial dynamic feature sequence of the electromyography (EMG) signal are taken as input, and the correlation coefficients between each feature are taken as output. The query matrix, key matrix, and value matrix are generated through linear transformation to obtain attention weights. Combined with the self-attention mechanism, the correlations between time-domain feature data and frequency-domain feature data, the correlations between time-domain feature data and time-frequency two-dimensional spatial dynamic feature data, and the correlations between frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data are learned respectively. A Transformer model is constructed, which combines the time-domain feature sequence, frequency-domain feature sequence, and time-frequency two-dimensional spatial dynamic feature sequence of the current EMG signal to output the corresponding correlation coefficients between each feature. The correlation coefficients between each feature are then integrated into the EMG signal recognition and classification dataset.

[0108] Step 105, the process of constructing an electromyography (EMG) signal recognition and classification model and outputting the corresponding EMG signal recognition and classification coefficients, includes:

[0109] The time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data, and correlation coefficients between various features of the electromyography signal recognition and classification dataset are extracted. The extracted data is then converted into a second training set and a second test set, with the ratio of the second training set to the second test set being 7:3.

[0110] A convolutional neural network (CNN) architecture is constructed using a CNN algorithm. This CNN architecture includes an input layer, a convolutional layer, and a fully connected layer. The input layer receives time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals from a second training set. The convolutional layer analyzes the correlation coefficients between various features in the second training set data to obtain the correlation between time-domain feature data and frequency-domain feature data, the correlation between time-domain feature data and time-frequency two-dimensional spatial dynamic feature data, and the correlation between frequency-domain feature data and time-frequency two-dimensional spatial dynamic feature data. The fully connected layer outputs the corresponding EMG signal recognition and classification coefficients to train the EMG signal recognition and classification model.

[0111] The second test set data is input into the electromyography (EMG) signal recognition and classification model to evaluate its performance. The parameters of the EMG signal recognition and classification model are adjusted and optimized to obtain the final EMG signal recognition and classification model. Combining the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data of the EMG signals in the current EMG signal recognition and classification dataset, as well as the correlation coefficients between various features, the corresponding EMG signal recognition and classification coefficients are output.

[0112] In step 106, the process of calibrating the electromyography (EMG) signal recognition classification coefficient using the calibration coefficient of the EMG signal recognition classification result, and evaluating the EMG signal recognition classification result, includes:

[0113] When the calibration coefficient of the electromyography (EMG) signal recognition and classification result is below 0.3, no calibration is performed on the EMG signal recognition and classification coefficient; when the calibration coefficient of the EMG signal recognition and classification result is between 0.3 and 0.5, calibration is performed through... The classification coefficients for electromyography (EMG) signal recognition are calibrated; when the calibration coefficient for EMG signal recognition is greater than 0.5, it is then... The classification coefficients for electromyographic signal recognition were calibrated, among which, The calibrated electromyography (EMG) signal recognition classification coefficients are defined as follows: QC and EP0 are the calibration coefficients and classification coefficients of the EMG signal recognition results, respectively, thus obtaining the calibrated EMG signal recognition classification coefficients.

[0114] When the calibrated electromyography (EMG) signal recognition classification coefficients are between 0~0.1, 0.1~0.2, 0.2~0.3, 0.3~0.4, 0.4~0.5, 0.5~0.6, 0.6~0.7, 0.7~0.8, and 0.8~1, they correspond to resting state EMG signal, systolic EMG signal, voluntary EMG signal, non-voluntary EMG signal, normal EMG signal, abnormal EMG signal, mild fatigue EMG signal, moderate fatigue EMG signal, and severe fatigue EMG signal, respectively. The EMG signal recognition classification results are then obtained.

[0115] Step 106, the process of outputting electromyographic signal recognition and classification results by issuing corresponding instructions, includes:

[0116] By combining voice feedback commands with text display commands, the voice broadcast and the text display on the terminal interface are controlled respectively, thereby realizing the output of electromyography signal recognition and classification results;

[0117] When the electromyography (EMG) signal identification and classification result is a resting state EMG signal, the voice announces that the current EMG signal is in a resting state, and the terminal interface displays the text "resting state"; when the EMG signal identification and classification result is a contraction state EMG signal, the voice announces that the current EMG signal is in a contraction state, and the terminal interface displays the text "contraction state".

[0118] When the electromyography (EMG) signal identification and classification result is voluntary EMG signal, the voice will announce that the current EMG signal is in a voluntary state, and the terminal interface will display the voluntary state text. When the EMG signal identification and classification result is involuntary EMG signal, the voice will announce that the current EMG signal is in a non-voluntary state, and the terminal interface will display the non-voluntary state text.

[0119] When the electromyography (EMG) signal identification and classification result is a normal EMG signal, the voice announces that the current EMG signal is in a normal state, and the terminal interface displays the text indicating that the signal is normal; when the EMG signal identification and classification result is an abnormal EMG signal, the voice announces that the current EMG signal is in an abnormal state, and the terminal interface displays the text indicating that the signal is abnormal.

[0120] When the electromyography (EMG) signal identification and classification result is a mild fatigue state EMG signal, the voice announces that the current EMG signal is in a mild fatigue state, and the terminal interface displays the text "mild fatigue state"; when the EMG signal identification and classification result is a moderate fatigue state EMG signal, the voice announces that the current EMG signal is in a moderate fatigue state, and the terminal interface displays the text "moderate fatigue state"; when the EMG signal identification and classification result is a severe fatigue state EMG signal, the voice announces that the current EMG signal is in a severe fatigue state, and the terminal interface displays the text "severe fatigue state".

[0121] Example 2, as Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: an electromyography (EMG) signal recognition and classification system for implementing the aforementioned EMG signal recognition and classification method, comprising a recognition and classification data acquisition module, a feature extraction module, a feature fusion analysis module, an EMG signal recognition and classification module, and an instruction output module, wherein the modules are communicatively connected.

[0122] The identification and classification data acquisition module collects electromyographic signal identification and classification data, which includes electromyographic signal data, inertial measurement unit data, and near-infrared spectral data, providing a data foundation for subsequent identification, classification, and calibration of electromyographic signals;

[0123] The feature extraction module extracts features from the electromyographic signal, obtaining time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of the electromyographic signal, thereby realizing in-depth analysis of the electromyographic signal;

[0124] The feature fusion analysis module uses a self-attention mechanism to construct a Transformer model, which fuses the time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional spatial dynamic feature data of electromyography signals, and then outputs the correlation coefficients between various features.

[0125] The electromyography (EMG) signal recognition and classification module is divided into a recognition and classification unit and a result calibration unit, which are used to recognize, classify and calibrate EMG signals to improve the accuracy of EMG signal recognition and classification results.

[0126] The identification and classification unit combines the time-domain feature data, frequency-domain feature data, time-frequency two-dimensional spatial dynamic feature data of electromyography (EMG) signals, and the correlation coefficients between various features. It then uses a convolutional neural network algorithm to construct an EMG signal identification and classification model and outputs the corresponding EMG signal identification and classification coefficients.

[0127] The calibration unit uses electromyography (EMG) signal intensity, inertial measurement unit (IMU) data, and near-infrared spectral data, combined with a multiple linear regression algorithm, to construct an EMG signal recognition and classification result calibration model, output corresponding EMG signal recognition and classification result calibration coefficients, calibrate the EMG signal recognition and classification coefficients, and then evaluate the EMG signal recognition and classification results.

[0128] The instruction output module outputs the electromyographic signal recognition and classification results by issuing corresponding instructions.

[0129] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of myoelectric signal recognition classification, the method comprising: The method comprises the following steps: Collecting electromyogram signal recognition classification data, which includes electromyogram signal data, inertial measurement unit data, and near-infrared spectrum data; Extracting features of the electromyogram signal to obtain time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional space dynamic feature data of the electromyogram signal; Using the electromyogram signal intensity, the inertial measurement unit data, and the near-infrared spectrum data, and combining a multivariate linear regression algorithm, an electromyogram signal recognition classification result calibration model is constructed, and then corresponding electromyogram signal recognition classification result calibration coefficients are outputted; According to the principle of a self-attention mechanism algorithm, a Transformer model is constructed, the time-domain feature data, the frequency-domain feature data, and the time-frequency two-dimensional space dynamic feature data of the electromyogram signal are fused, and then correlation coefficients between various features are outputted; Combining the time-domain feature data, the frequency-domain feature data, the time-frequency two-dimensional space dynamic feature data of the electromyogram signal, and the correlation coefficients between various features, a convolutional neural network algorithm is used to construct an electromyogram signal recognition classification model, and corresponding electromyogram signal recognition classification coefficients are outputted; The electromyogram signal recognition classification coefficients are calibrated by the electromyogram signal recognition classification result calibration coefficients, the electromyogram signal recognition classification result is evaluated, and the electromyogram signal recognition classification result is outputted by issuing corresponding instructions.

2. The method of claim 1, wherein: The collection process of the electromyogram signal recognition classification data comprises: Different types of collection devices are deployed to collect the electromyogram signal data, the inertial measurement unit data, and the near-infrared spectrum data, wherein the collection devices include electromyogram sensors, electromyogram biofeedback instruments, accelerometers, gyroscopes, and Train.RED wireless muscle oxygen monitoring instruments; The electromyogram signal data includes electromyogram signals and their intensities; the inertial measurement unit data includes acceleration and angular velocity of limb movement; and the near-infrared spectrum data includes oxygenated hemoglobin concentration and deoxygenated hemoglobin concentration in muscle tissue; Specifically, electromyogram sensors are used to collect electromyogram signals at a sampling frequency of 500-2000 Hz, and electromyogram biofeedback instruments are used to collect electromyogram signal intensities; a Train.RED wireless muscle oxygen monitoring instrument is used to collect oxygenated hemoglobin concentration and deoxygenated hemoglobin concentration in muscle tissue based on near-infrared spectrum technology, by emitting near-infrared light to penetrate muscle tissue, combining a modified Beer-Lambert law with a spatially resolved spectroscopy method; The collected electromyogram signal intensities, the inertial measurement unit data, and the near-infrared spectrum data are subjected to data cleaning and data normalization processing, the electromyogram signal is subjected to denoising, rectification, and normalization processing, and the preprocessed electromyogram signal intensities, the inertial measurement unit data, and the near-infrared spectrum data are integrated to generate an electromyogram signal recognition classification data set.

3. The method of claim 2, wherein: The process of extracting features of the electromyogram signal to obtain time-domain feature data and frequency-domain feature data of the electromyogram signal comprises: The time-domain feature data of the electromyogram signal includes root mean square value, average absolute value, zero-crossing rate, waveform length, and variance; and the frequency-domain feature data of the electromyogram signal includes median frequency and average power frequency. The time domain feature extraction technology is used, the length of the segmentation window is set to 200 ms, the denoising, rectification and normalization processed electromyographic signals are segmented into a plurality of windows based on the set window length, the sampling rate is set, the number of electromyographic signal samples contained in each window is determined, the electromyographic signal samples of each window are obtained, the mean value of the electromyographic signal is calculated according to the electromyographic signal samples of each window and the number of electromyographic signal samples contained therein, the root mean square value, the average absolute value, the zero-crossing rate, the waveform length and the variance of the electromyographic signal are respectively calculated by combining the electromyographic signal samples of each window, the mean value of the electromyographic signal, the number of electromyographic signal samples contained therein and the zero-crossing rate indicator function; The power spectrum density curve of the electromyographic signal is obtained by the Fourier transform technology, the power spectrum density function of the electromyographic signal is obtained by combining the Welch method and the fast Fourier transform, and then the median frequency and the average power frequency of the electromyographic signal are respectively calculated. The obtained time domain feature data and frequency domain feature data of the electromyographic signal are integrated into the electromyographic signal recognition classification data set.

4. The method of claim 3, wherein: The process of extracting features from the electromyographic signal and obtaining time-frequency two-dimensional space dynamic feature data of the electromyographic signal includes: The time-frequency two-dimensional space dynamic feature data includes the center of gravity frequency, bandwidth and time-frequency entropy of the energy distribution on the two-dimensional time-frequency diagram of the electromyographic signal. The time-frequency analysis technology is used to convert the electromyographic signal into a two-dimensional time-frequency diagram, and the time, frequency and corresponding energy density thereof are extracted from the two-dimensional time-frequency diagram by combining the wavelet transform. The center of gravity frequency, bandwidth and time-frequency entropy of the energy distribution on the two-dimensional time-frequency diagram of the electromyographic signal are respectively calculated by using the time, frequency and corresponding energy density thereof extracted from the two-dimensional time-frequency diagram, and the center of gravity frequency, bandwidth and time-frequency entropy of the energy distribution on the two-dimensional time-frequency diagram of the electromyographic signal are integrated into the electromyographic signal recognition classification data set.

5. The method of claim 4, wherein: The process of constructing the electromyographic signal recognition classification result calibration model and then outputting the corresponding electromyographic signal recognition classification result calibration coefficient includes: The electromyographic signal intensity, the inertial measurement unit data and the near-infrared spectrum data in the electromyographic signal recognition classification data set are extracted, and the extracted data are converted into a first training set and a first test set; The first training set data and the multiple linear regression algorithm are combined, the first training set data are taken as input, the electromyographic signal recognition classification result calibration coefficient is taken as output, the non-linear relationship between the electromyographic signal intensity, the inertial measurement unit data, the near-infrared spectrum data and the electromyographic signal recognition classification result calibration coefficient is learned, and the electromyographic signal recognition classification result calibration model is trained; The first test set data are input into the electromyographic signal recognition classification result calibration model, the regression coefficient and the intercept term of the electromyographic signal recognition classification result calibration model are adjusted, the performance of the electromyographic signal recognition classification result calibration model is optimized, the final electromyographic signal recognition classification result calibration model is obtained, and the corresponding electromyographic signal recognition classification result calibration coefficient is output in combination with the current electromyographic signal intensity, the inertial measurement unit data and the near-infrared spectrum data.

6. The method of myoelectric signal recognition according to claim 5, wherein: The process of constructing the Transformer model, fusing the time domain feature data, the frequency domain feature data and the time-frequency two-dimensional space dynamic feature data of the electromyographic signal, and then outputting the correlation coefficients between the features comprises: The correlation coefficients between the features comprise correlation coefficients between the time domain feature data and the frequency domain feature data, correlation coefficients between the time domain feature data and the time-frequency two-dimensional space dynamic feature data, and correlation coefficients between the frequency domain feature data and the time-frequency two-dimensional space dynamic feature data; The time domain feature data, the frequency domain feature data and the time-frequency two-dimensional space dynamic feature data of the electromyographic signal are extracted from the electromyographic signal recognition classification data set, and the extracted data is mapped to a corresponding dimensional vector space through a fully connected layer to obtain time domain feature sequences, frequency domain feature sequences and time-frequency two-dimensional space dynamic feature sequences of the electromyographic signal; According to the self-attention mechanism, the time domain feature sequences, the frequency domain feature sequences and the time-frequency two-dimensional space dynamic feature sequences of the electromyographic signal are taken as inputs, and the correlation coefficients between the features are taken as outputs. Query matrix, Key matrix and Value matrix are generated through linear transformation to obtain attention weights. The self-attention mechanism is combined to learn the correlation between the time domain feature data and the frequency domain feature data, the correlation between the time domain feature data and the time-frequency two-dimensional space dynamic feature data, and the correlation between the frequency domain feature data and the time-frequency two-dimensional space dynamic feature data, respectively. The Transformer model is constructed. The correlation coefficients between the features are output in combination with the time domain feature sequences, the frequency domain feature sequences and the time-frequency two-dimensional space dynamic feature sequences of the current electromyographic signal, and the correlation coefficients between the features are integrated into the electromyographic signal recognition classification data set.

7. The method of myoelectric signal recognition according to claim 6, wherein: The process of constructing the electromyographic signal recognition classification model and outputting the corresponding electromyographic signal recognition classification coefficients comprises: The time domain feature data, the frequency domain feature data, the time-frequency two-dimensional space dynamic feature data of the electromyographic signal and the correlation coefficients between the features in the electromyographic signal recognition classification data set are extracted, and the extracted data is converted into a second training set and a second test set. The proportion of the second training set and the second test set is 7:3; A convolutional neural network architecture is constructed using a convolutional neural network algorithm. The convolutional neural network architecture comprises an input layer, a convolutional layer and a fully connected layer. The input layer receives the time domain feature data, the frequency domain feature data and the time-frequency two-dimensional space dynamic feature data of the electromyographic signal in the second training set data. The correlation coefficients between the features in the second training set data are analyzed through the convolutional layer to obtain the correlation between the time domain feature data and the frequency domain feature data, the correlation between the time domain feature data and the time-frequency two-dimensional space dynamic feature data, and the correlation between the frequency domain feature data and the time-frequency two-dimensional space dynamic feature data. The corresponding electromyographic signal recognition classification coefficients are output through the fully connected layer to train the electromyographic signal recognition classification model. The second test set data is input into the electromyogram recognition classification model, the performance of the electromyogram recognition classification model is evaluated, the electromyogram recognition classification model parameters are adjusted, the electromyogram recognition classification model is optimized, a final electromyogram recognition classification model is obtained, and corresponding electromyogram recognition classification coefficients are output in combination with the time domain feature data, the frequency domain feature data, the time-frequency two-dimensional space dynamic feature data, and the correlation coefficients between the features in the electromyogram recognition classification data set.

8. The method of claim 7, wherein: The process of calibrating the electromyogram recognition classification coefficients by the electromyogram recognition classification result calibration coefficients and evaluating the electromyogram recognition classification result includes: when the myoelectric signal recognition classification result calibration coefficient is less than 0.3, the myoelectric signal recognition classification coefficient is not calibrated; when the myoelectric signal recognition classification result calibration coefficient is between 0.3 and 0.5, the myoelectric signal recognition classification coefficient is calibrated by the myoelectric signal recognition classification coefficient is calibrated, wherein, is the calibrated myoelectric signal recognition classification coefficient, QC and EP0 are the myoelectric signal recognition classification result calibration coefficient and the myoelectric signal recognition classification coefficient respectively, and the calibrated myoelectric signal recognition classification coefficient is obtained.​ When the calibrated electromyogram recognition classification coefficients are between 0-0.1, 0.1-0.2, 0.2-0.3, 0.3-0.4, 0.4-0.5, 0.5-0.6, 0.6-0.7, 0.7-0.8, and 0.8-1, they correspond to resting electromyogram, contraction electromyogram, voluntary electromyogram, involuntary electromyogram, normal electromyogram, abnormal electromyogram, mild fatigue electromyogram, moderate fatigue electromyogram, and severe fatigue electromyogram, respectively, and electromyogram recognition classification results are obtained.

9. The method of claim 8, wherein: The process of outputting the electromyogram recognition classification result by issuing corresponding instructions includes: The voice feedback instruction text is combined with the display instruction to control the voice broadcast and the terminal interface text display, respectively, so as to output the electromyogram recognition classification result; When the electromyogram recognition classification result is resting electromyogram, the voice broadcast is that the current electromyogram is resting, and the terminal interface text display is resting; When the electromyogram recognition classification result is voluntary electromyogram, the voice broadcast is that the current electromyogram is voluntary, and the terminal interface text display is voluntary; When the electromyogram recognition classification result is normal electromyogram, the voice broadcast is that the current electromyogram is normal, and the terminal interface text display is normal; When the electromyogram recognition classification result is mild fatigue electromyogram, the voice broadcast is that the current electromyogram is mild fatigue, and the terminal interface text display is mild fatigue; when the electromyogram recognition classification result is moderate fatigue electromyogram, the voice broadcast is that the current electromyogram is moderate fatigue, and the terminal interface text display is moderate fatigue; and when the electromyogram recognition classification result is severe fatigue electromyogram, the voice broadcast is that the current electromyogram is severe fatigue, and the terminal interface text display is severe fatigue.

10. An electromyographic signal recognition and classification system for implementing the method of any one of claims 1-9, comprising a recognition and classification data acquisition module, a feature extraction module, a feature fusion analysis module, an electromyographic signal recognition and classification module, and an instruction output module, wherein, The modules are in communication, and the system further includes: The recognition classification data collection module collects electromyographic signal recognition classification data, and the electromyographic signal recognition classification data includes electromyographic signal data, inertial measurement unit data, and near-infrared spectrum data. The feature extraction module extracts features of the electromyographic signal, and obtains time-domain feature data, frequency-domain feature data, and time-frequency two-dimensional space dynamic feature data of the electromyographic signal. The feature fusion analysis module uses a self-attention mechanism to construct a Transformer model, fuses the time-domain feature data, the frequency-domain feature data, and the time-frequency two-dimensional space dynamic feature data of the electromyographic signal, and then outputs correlation coefficients between various features. The electromyographic signal recognition classification module includes a recognition classification unit and a result calibration unit, and is configured to recognize and calibrate electromyographic signals. The recognition classification unit combines the time-domain feature data, the frequency-domain feature data, the time-frequency two-dimensional space dynamic feature data of the electromyographic signal, and the correlation coefficients between various features, uses a convolutional neural network algorithm to construct an electromyographic signal recognition classification model, and outputs corresponding electromyographic signal recognition classification coefficients. The result calibration unit combines the electromyographic signal intensity, the inertial measurement unit data, and the near-infrared spectrum data, uses a multiple linear regression algorithm to construct an electromyographic signal recognition classification result calibration model, outputs corresponding electromyographic signal recognition classification result calibration coefficients, calibrates the electromyographic signal recognition classification coefficients, and then evaluates electromyographic signal recognition classification results. The instruction output module outputs the electromyographic signal recognition classification results by issuing corresponding instructions.

Images