Neural state recognition method, device, storage medium, and program product based on EMG signals

Abstract

This invention discloses a neural state recognition method based on EMG signals, comprising: acquiring electromyography (EMG) detection signals; performing state recognition based on the EMG detection signals using a neural state recognition model to obtain state reference information; the state recognition of the neural state recognition model includes: preprocessing the acquired EMG detection signals to obtain standardized detection data; extracting detection feature information from the standardized detection data, establishing the time-series dependencies, focusing on changes in time segments through an attention mechanism, and matching the detection feature information with state reference information using data from a sample library. This invention discloses a neural state recognition method, device, storage medium, and program product based on EMG signals, constructing a neural state dataset, extracting key feature parameters, and using deep learning algorithms to establish a neural state discrimination model to achieve neural state recognition, which can help improve the stability, timeliness, and accuracy of neural state recognition.

Description

Technical Field

[0001] This invention relates to the field of electromyography (EMG) signal state recognition technology, specifically to a method, device, storage medium, and program product for neural state recognition based on EMG signals. Background Technology

[0002] In existing technologies, somatosensory evoked potentials (SEP) or motor evoked potentials (MEP) are commonly used to monitor neurological function. However, this method requires doctors to continuously monitor waveform changes, making the judgment process cumbersome and the workload heavy, with risks of missed diagnoses and delays.

[0003] Therefore, it is necessary to develop a neural state discrimination method that can automatically identify neural states by collecting and labeling multi-channel EMG signals, thereby improving the stability, timeliness, and accuracy of neural state identification, especially in the field of nerve injury. Summary of the Invention

[0004] This invention overcomes the shortcomings of existing technologies and provides a method, device, storage medium, and program product for neural state recognition based on EMG signals. It constructs a neural state dataset, extracts key feature parameters, and uses deep learning algorithms to establish a neural state discrimination model, thereby achieving automatic recognition of neural states and helping to improve the stability, timeliness, and accuracy of neural state recognition.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a neural state recognition method based on EMG signals, comprising:

[0006] Acquire detection data, which includes: electromyographic detection signals at the detection site;

[0007] Based on the electromyography (EMG) detection signal, a neural state recognition model is used to identify the state and obtain state reference information.

[0008] The state recognition of the neural state recognition model includes:

[0009] The acquired electromyography (EMG) detection signals are preprocessed using a CNN-BiLSTM deep learning model to obtain standardized detection data. Based on the CNN layer in the CNN-BiLSTM deep learning model, detection feature information is extracted from the standardized detection data. The time series dependencies are established according to the BiLSTM structure in the CNN-BiLSTM deep learning model. The changes in time segments are focused through an attention mechanism, and the detection feature information is matched with state reference information using data from the sample library.

[0010] In a preferred embodiment of the present invention, the acquisition of the sample library includes:

[0011] The acquired sample data is preprocessed to obtain standardized sample data; feature extraction is performed based on the standardized sample data to obtain label data; and a sample library is established based on the label data.

[0012] The sample data includes electromyographic signals at rest and electromyographic signals under traction at the corresponding locations in the case records.

[0013] In a preferred embodiment of the present invention, the sample data further includes electromyographic interference signals at corresponding locations in the case record.

[0014] In a preferred embodiment of the present invention, the CNN-BiLSTM deep learning model includes:

[0015] Step S1: The one-dimensional convolutional neural network module performs local feature extraction on the input 45-dimensional feature sequence to extract mutation points or local abnormal patterns in the electromyographic signal.

[0016] Step S2: The extracted temporal features are processed through a bidirectional long short-term memory network (BiLSTM) to enhance the model's ability to model temporal dependencies.

[0017] Step S3 involves weighting the sequence output by BiLSTM through an attention mechanism, enabling the model to autonomously focus on key feature segments and improve classification performance and stability.

[0018] In a preferred embodiment of the present invention, the preprocessing of the acquired electromyography (EMG) detection signal includes: performing sliding window normalization and segmentation on the EMG detection signal to obtain standardized detection data;

[0019] Alternatively, preprocessing of the acquired sample data may include: performing sliding window processing and segmentation on the electromyographic sample signals of the sample data, and standardizing the sample data.

[0020] In a preferred embodiment of the present invention, the acquisition of electromyographic detection signals includes: placing the acquisition electrodes of the neurophysiological acquisition device on the surface of the acquisition muscle group of the subject; the acquisition muscle group includes: left abductor pollicis, left tibialis anterior, left adductor pollicis, right tibialis anterior, and right adductor pollicis.

[0021] In a preferred embodiment of the present invention, the extracted features include time-domain features, frequency-domain features, kurtosis features, and nonlinear features;

[0022] The time-domain features include maximum value, minimum value, and RMS; among them, the maximum value and minimum value are used to reflect the degree of muscle activation; RMS is the root mean square, used to measure the overall energy level.

[0023] Frequency domain characteristics include median frequency and average power frequency;

[0024] Kurtosis: Used to reflect the sharpness of a signal;

[0025] In kurtosis features, MP is the median frequency, and MPF is the average power frequency. MP and MPF are used to reflect the spectral distribution and reveal the changes in the neural drive frequency band.

[0026] Nonlinear features include sample entropy and permutation entropy: SampEn is the sample entropy and PE is the permutation entropy. SampEn and PE are used to describe the complexity and nonlinear features of a signal and to help identify abnormal activation patterns.

[0027] In a preferred embodiment of the present invention, a neural state recognition device based on EMG signals includes a memory for storing computer programs / instructions and a processor for executing the computer programs / instructions to implement the steps of the neural state recognition method based on EMG signals.

[0028] In a preferred embodiment of the present invention, a neural state recognition storage medium based on EMG signals is provided, on which a computer program / instruction is stored, which, when executed by a processor, implements the steps of a neural state recognition method based on EMG signals.

[0029] In a preferred embodiment of the present invention, a computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of a neural state recognition method based on EMG signals.

[0030] This invention addresses the deficiencies in the technical background, and the beneficial technical effects of this invention are:

[0031] A method, device, storage medium, and program product for neural state recognition based on EMG signals are disclosed. The method constructs a neural state dataset, extracts key feature parameters, and uses deep learning algorithms to establish a neural state discrimination model, thereby achieving automatic recognition of neural states and helping to improve the stability, timeliness, and accuracy of neural state recognition. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Figure 1 This is a comparison of time-domain, frequency-domain, and time-frequency data of electromyography (EMG) signals, EMG interference signals, and traction EMG signals in the sample data of a preferred embodiment of the present invention.

[0034] Figure 2 This is a flowchart of the neural state recognition model in a preferred embodiment of the present invention;

[0035] Figure 3 This is a structural diagram of the neural state recognition model in a preferred embodiment of the present invention;

[0036] Figure 4 This is a structure table of the neural state recognition model in a preferred embodiment of the present invention;

[0037] Figure 5 This is pseudocode data for a neural state recognition model in a preferred embodiment of the present invention;

[0038] Figure 6 The feature boundary of the t-SNE mapping in the preferred embodiment of the present invention Figure 1 ;

[0039] Figure 7 The feature boundary of the t-SNE mapping in the preferred embodiment of the present invention Figure 2 ;

[0040] Figure 8 This is a graph showing the model accuracy versus loss function in a preferred embodiment of the present invention.

[0041] Figure 9 This is a table of accuracy and recall parameters for three types of data in a preferred embodiment of the present invention;

[0042] Figure 10 This is a model comparison parameter table in a preferred embodiment of the present invention. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. These drawings are simplified schematic diagrams, which are only used to illustrate the basic structure of the present invention and therefore only show the components relevant to the present invention.

[0044] It should be noted that if directional indicators (such as up, down, bottom, top, etc.) are involved in the embodiments of the present invention, these directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of that feature. Unless otherwise explicitly specified and limited, the terms "set," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in the present invention can be understood according to the specific circumstances.

[0045] Example 1, as Figures 1-4As shown, a neural state recognition method based on EMG signals includes:

[0046] Acquire detection data, which includes: electromyographic detection signals at the detection site;

[0047] Based on the electromyography (EMG) detection signal, a neural state recognition model is used to identify the state and obtain state reference information.

[0048] Specifically, state recognition in neural state recognition models includes:

[0049] The acquired electromyography (EMG) detection signals are preprocessed using a CNN-BiLSTM deep learning model to obtain standardized detection data. Based on the CNN layer in the CNN-BiLSTM deep learning model, detection feature information is extracted from the standardized detection data. The time series dependencies are established according to the BiLSTM structure in the CNN-BiLSTM deep learning model. The changes in time segments are focused through an attention mechanism, and the detection feature information is matched with state reference information using data from the sample library.

[0050] Specifically, obtaining the sample library includes:

[0051] The acquired sample data is preprocessed to obtain standardized sample data; feature extraction is performed based on the sample data to obtain label data, and a sample library is established based on the label data; the sample data includes electromyographic signals at rest and electromyographic signals under traction at corresponding locations in the case records.

[0052] Furthermore, the sample data also includes electromyographic interference signals at corresponding locations in the case records. The sample data was collected from existing case data.

[0053] Specifically, the CNN-BiLSTM deep learning model includes:

[0054] Step S1: The one-dimensional convolutional neural network module performs local feature extraction on the input multi-dimensional feature sequence (45-dimensional feature sequence in this embodiment) to extract mutation points or local abnormal patterns in the electromyographic signal.

[0055] Step S2: The extracted temporal features are processed through a bidirectional long short-term memory network (BiLSTM) to enhance the model's ability to model temporal dependencies.

[0056] Step S3 involves weighting the sequence output by BiLSTM through an attention mechanism, enabling the model to autonomously focus on key feature segments and improve classification performance and stability.

[0057] Among them, the CNN-BiLSTM deep learning model possesses strong sequence modeling and feature selection capabilities, making it particularly suitable for non-stationary and non-linear electromyography (EMG) signal processing tasks. It exhibits good classification performance within the feature boundaries of t-SNE. The convolutional layers (CNN) automatically extract local variation features, such as EMG mutations; the bidirectional LSTM (BiLSTM) considers the sequential dependencies of signals, enhancing temporal understanding; and the attention mechanism strengthens the model's focus on key segments. Furthermore, the dataset is divided into training and test sets at a ratio of 75% and 25% to ensure generalization ability. Cross-entropy is used as the loss function, the Adam optimizer is selected, the batch size is set to 32, and the number of training epochs is 100.

[0058] Specifically, preprocessing the acquired electromyography (EMG) signals includes: performing sliding window standardization and segmentation on the EMG signals to obtain standardized detection data; or, preprocessing the acquired sample data includes: performing sliding window processing and segmentation on the EMG sample signals of the sample data to standardize the sample data.

[0059] Specifically, the acquisition of electromyographic (EMG) signals involves placing the acquisition electrodes of a neurophysiological acquisition device on the surface of the target muscle groups. These muscle groups include the left abductor pollicis, left tibialis anterior, left adductor pollicis, right tibialis anterior, and right adductor pollicis. The acquisition locations cover typical and highly sensitive muscle groups, such as the adductor pollicis, tibialis anterior, and abductor pollicis, effectively reflecting the true state of nerve conduction. High-quality continuous EMG data provides a solid foundation for subsequent analysis, especially enabling continuous monitoring of neurological function during complex surgical procedures.

[0060] Specifically, the extracted features include time-domain features, frequency-domain features, kurtosis features, and nonlinear features;

[0061] The time-domain features include maximum value, minimum value, and RMS; among them, the maximum value and minimum value are used to reflect the degree of muscle activation; RMS is the root mean square, used to measure the overall energy level.

[0062] Frequency domain characteristics include median frequency and average power frequency;

[0063] Kurtosis: Used to reflect the sharpness of a signal;

[0064] In kurtosis features, MP is the median frequency, and MPF is the average power frequency. MP and MPF are used to reflect the spectral distribution and reveal the changes in the neural drive frequency band.

[0065] Nonlinear features include sample entropy and permutation entropy: SampEn is the sample entropy and PE is the permutation entropy. SampEn and PE are used to describe the complexity and nonlinear features of a signal and to help identify abnormal activation patterns.

[0066] Working principle:

[0067] This invention provides a neural state recognition method based on EMG signals. By collecting and labeling multi-channel EMG signals, a neural state dataset is constructed, key feature parameters are extracted, and a neural state discrimination model is established using deep learning algorithms to achieve automatic recognition of neural states. This method can help improve the stability, timeliness, and accuracy of neural state recognition, especially in the field of nerve injury.

[0068] Example 2, based on Example 1, such as Figures 1-5 As shown, in this embodiment of a neural state recognition method based on EMG signals, the EMG signal at rest is the EMG signal without operation. EMG interference signals, such as changes in EMG signals caused by external devices (electrosurgical devices, electrocoagulation devices; the device model is recorded based on actual use and the real situation of the sample data, generally using commonly used models in the prior art; specific sample information and the selection of device models in actual operation will not be elaborated or listed here). Traction EMG signals are EMG responses caused by the plucked nerves marked in the case (plucking, for example, traction caused by the test subject's movement of the test position, thus plucking the nerve; but not limited to this, or it could also be the test robot automatically plucking the nerve according to a preset program). Attention mechanisms enhance the model's attention to key segments, such as subtle changes in EMG during nerve plucking.

[0069] Example 3, based on Example 1 or Example 2, such as Figures 1-5 As shown, this embodiment of a neural state recognition method based on EMG signals specifically includes:

[0070] The algorithm for the root mean square (RMS) is as follows: The root mean square (RMS) is used to reflect the average power of the signal; where, Let N be the i-th sampling point of the signal; N is the signal length.

[0071] The algorithm for integral electromyography (IEMG) values ​​is as follows: Integrated electromyography (EMG) values ​​are used to reflect the degree of muscle activation.

[0072] The algorithm for kurtosis (SK) is as follows: Kurtosis is used to reflect the sharpness of a signal; The mean of the signal within the window; the algorithms for the maximum value (Max) and minimum value (Min) are as follows: , This is used to reflect the amplitude range of the signal;

[0073] The algorithm for the average power frequency (MPF) is as follows: The average power frequency is used to reflect the median frequency of the power spectrum, where... For the k-th frequency component, for The power spectral density at point M, where M is the total number of frequency components;

[0074] The algorithm for the median frequency (MF) is as follows: Calculate the power spectral density using Fast Fourier Transform (FFT) and find the frequency that satisfies the following conditions. , The median frequency is used to reflect the average frequency of the power spectrum.

[0075] The algorithm for sample entropy SampEn is as follows: Embedding dimension m=2, matches is the number of matches in the dimension, and sample entropy is used to reflect the complexity of the signal; parameters of sample entropy (SampEn): tolerance r=0.2*std(window), the tolerance is 20% of the standard deviation of the current window signal, and time delay τ=1;

[0076] The algorithm for fractal dimension PFD is as follows: N: Signal length The number of directional changes in a signal (number of inflection points); fractal dimension reflects the degree of irregularity of the signal.

[0077] Example 4, based on Example 1 or Example 3, such as Figures 1-5 As shown, a neural state recognition method based on EMG signals in this embodiment includes:

[0078] Electromyography (EMG) data from case records were acquired as sample data. The entire experimental process, from signal acquisition and feature construction to model training and evaluation, was completed using this sample data. There was no specific limit to the number of cases used; a larger number generally resulted in better model stability and accuracy. The number of cases was adjusted according to actual training needs. The EMG acquisition system used had five channels, with a sampling frequency set to 1280Hz. The acquisition channels were positioned on the surface of muscles closely related to the patient's lower limb and spinal nerve pathways, using standard surface electrodes. The reference electrode was placed in a bony location to improve signal stability and reduce the impact of high-frequency interference during surgery.

[0079] The acquired EMG signals are continuous time-series signals. To facilitate subsequent processing and model input, they need to be segmented first. This embodiment uses a sliding window method for data segmentation, with a sliding window length of 128 points and a step size of 128 points, meaning there is no overlap between windows (overlap is 0), ensuring data processing efficiency and sample independence. Each window segment contains EMG data from 5 channels, and 9 types of features are extracted for each channel, including maximum value, minimum value, root mean square (RMS), integrated electromyography (IEMG) value, kurtosis, median frequency (MP), average power frequency (MPF), sample entropy, and permutation entropy (or fractal dimension). These features cover the signal's amplitude information, spectral structure, and nonlinear complexity, and can comprehensively reflect the degree of muscle nerve excitation and damage status.

[0080] After sliding window processing and feature extraction, 283 valid sample data segments were obtained from the sampling data of each case. Since each sample segment contains 5 channels and 9 features per channel, the feature fusion of the vectors uses simple concatenation, resulting in a feature vector dimension of 45, and therefore a feature matrix dimension of 283×45. The electromyographic signals of the sample data include: resting state, operation disturbance state (i.e., interference state, such as electrosurgery, electrocoagulation, etc.), and signal changes caused by nerve manipulation (the signal changes caused by nerve manipulation are defined as the "nerve injury state" label for model training). The 283 valid sample data segments in this embodiment include 100 resting state segments, 100 operation disturbance state segments, and 83 signal changes caused by nerve manipulation. However, this is not the only possibility. In other embodiments, the valid sample data can be adjusted or a specific number of samples can be set according to actual usage needs to obtain more comprehensive and accurate data analysis. It should be noted that the valid sample data of this invention are all legally acquired and used, and the data is anonymized using existing technologies during model training, testing, and use.

[0081] In the model construction phase, a CNN-BiLSTM deep neural network based on an attention mechanism is used for neural state discrimination modeling. This model first uses a one-dimensional convolutional neural network (CNN) module to extract local features from the input 45-dimensional feature sequence, extracting abrupt changes or local abnormal patterns in the electromyography (EMG) signal. Then, a bidirectional long short-term memory (BiLSTM) network is used to process the extracted temporal features, enhancing the model's ability to model temporal dependencies. Finally, an attention mechanism is added to weight the sequence output by the BiLSTM, enabling the model to autonomously focus on key feature segments, improving classification performance and stability. This model structure possesses strong sequence modeling and feature selection capabilities, making it particularly suitable for intraoperative non-stationary and nonlinear EMG signal processing tasks. It exhibits good classification performance within the feature boundaries of t-SNE. The t-SNE feature mapping is as follows: Figure 6 , Figure 7 As shown.

[0082] During model training, the dataset was divided into training and test sets at a ratio of 75% and 25% to ensure generalization ability. Cross-entropy was used as the loss function, Adam was selected as the optimizer, the batch size was set to 32, and the number of training epochs was 100. Model accuracy and loss function are as follows: Figure 8 As shown.

[0083] After training, the model performance was evaluated on the test set. The results showed that the model achieved a classification accuracy of 89.4% on the test set, indicating that the proposed method has high discrimination accuracy and robustness. The accuracy and recall rates for the three classes are shown in the table below. Figure 9 As shown.

[0084] This embodiment demonstrates the feasibility and effectiveness of the present invention, from EMG signal acquisition and feature extraction to deep learning-based neural state discrimination, and provides a comparison with other models. Figure 10 As shown, the performance of this model is superior to traditional machine learning and deep learning models. This verifies the practical application value of the invention; it not only improves the real-time performance and accuracy of neural state assessment but also provides an intelligent auxiliary tool for neuroprotection, demonstrating promising prospects for widespread adoption.

[0085] Working principle:

[0086] This invention demonstrates high reliability in data acquisition. The system utilizes CADwell neurophysiological equipment for multi-channel electromyography (EMG) signal acquisition, covering typical and highly sensitive muscle groups such as the adductor pollicis, tibialis anterior, and abductor pollicis, thus fully reflecting the true state of nerve conduction. High-quality continuous EMG data provides a solid foundation for subsequent analysis.

[0087] The system also demonstrates professionalism and specificity in signal preprocessing and feature extraction. Due to the strong anti-interference capabilities of the acquisition equipment itself, traditional filtering or notch filtering is unnecessary; input requirements can be met simply through sliding window partitioning and standardization. Feature selection covers the time domain (maximum, minimum, RMS), frequency domain (median frequency, average power frequency), kurtosis, and nonlinear features (sample entropy, permutation entropy). These features comprehensively characterize the time-frequency distribution and complexity of electromyographic signals from multiple perspectives, with nonlinear features being particularly effective in capturing sudden neural responses.

[0088] Furthermore, this method fully incorporates the actual surgical procedure in its label design. By using a deep learning model to learn the different characteristics of the original signals, the original signals are divided into three categories: electromyographic signals at rest, traction electromyographic signals, and electromyographic interference signals. This improves the accuracy and robustness of the model training, enabling the system to accurately distinguish between normal physiological changes and potential damage reactions caused by intraoperative procedures.

[0089] In terms of model structure, the adoption of a CNN-BiLSTM deep learning model with an attention mechanism is one of the core advantages of this method. CNN layers can extract local mutation features in the signal, while the BiLSTM structure can model the sequential dependencies of time series. In particular, the attention mechanism can focus on subtle changes in key time segments, improving the model's sensitivity and recognition accuracy for abnormal electromyographic activity. After network optimization, the model is capable of running on embedded devices, completing a discrimination with a response time of no more than 50ms, fully meeting the real-time requirements of intraoperative surgery.

[0090] Compared to traditional nerve monitoring methods that often rely on the surgeon's subjective judgment and experience, which suffer from problems such as delayed response, ambiguous judgment, and lack of objective quantification, making it easy for intraoperative injuries to go undetected and leading to irreversible consequences such as postoperative neurological dysfunction, the real-time nerve injury discrimination system proposed in this invention, combined with multi-channel acquisition of intraoperative electromyography (EMG) signals and deep learning technology, can achieve automated, objective, and rapid feedback of nerve status, demonstrating significant advantages in multiple aspects.

[0091] Example 5, as Figures 1-6 As shown, based on Embodiment 1 or Embodiment 3, a neural state recognition device based on EMG signals includes a memory for storing computer programs / instructions; and a processor for executing the computer programs / instructions to implement the steps of the neural state recognition method based on EMG signals.

[0092] Example 6, as Figures 1-6 As shown, based on Embodiment 1 or Embodiment 3, a neural state recognition storage medium based on EMG signals is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, it implements the steps of the neural state recognition method based on EMG signals.

[0093] Example 7, as Figures 1-6 As shown, based on Embodiment 1 or Embodiment 3, a computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of a neural state recognition method based on EMG signals.

[0094] The above specific embodiments are specific support for the concept proposed in this invention, and should not be used to limit the scope of protection of this invention. Any equivalent changes or modifications made on the basis of this technical solution in accordance with the technical concept proposed in this invention shall still fall within the scope of protection of this invention.

Claims

1. A neural state recognition method based on EMG signals, characterized in that, include: Acquire detection data, which includes: electromyographic detection signals at the detection site; Based on the electromyography (EMG) detection signal, a neural state recognition model is used to identify the state and obtain state reference information. The state recognition of the neural state recognition model includes: The acquired electromyography (EMG) detection signals are preprocessed using a CNN-BiLSTM deep learning model to obtain standardized detection data. Based on the CNN layer in the CNN-BiLSTM deep learning model, detection feature information is extracted from the standardized detection data. The time series dependencies are established according to the BiLSTM structure in the CNN-BiLSTM deep learning model. The changes in time segments are focused through an attention mechanism, and the detection feature information is matched with state reference information using data from the sample library. The CNN-BiLSTM deep learning model includes: Step S1: The one-dimensional convolutional neural network module performs local feature extraction on the input multi-dimensional feature sequence to extract abrupt change points or local abnormal patterns in the electromyographic signal. Step S2: The extracted temporal features are processed through a bidirectional long short-term memory network to enhance the model's ability to model temporal dependencies. Step S3: Through the attention mechanism, the sequence output by BiLSTM is weighted, enabling the model to autonomously focus on key feature segments and improve classification performance and stability. The acquisition of the sample library includes: The acquired sample data is preprocessed to obtain standardized sample data; feature extraction is performed based on the standardized sample data to obtain label data; and a sample library is established based on the label data. The sample data includes electromyographic signals at rest and electromyographic signals under traction at the corresponding locations in the case records.

2. The neural state recognition method based on EMG signals according to claim 1, characterized in that: The sample data also includes electromyographic interference signals at corresponding locations in the case records.

3. The neural state recognition method based on EMG signals according to claim 2, characterized in that: Preprocessing of the acquired electromyography (EMG) signals includes: performing sliding window standardization and segmentation on the EMG signals to obtain standardized detection data; Alternatively, preprocessing of the acquired sample data may include: performing sliding window processing and segmentation on the electromyographic sample signals of the sample data, and standardizing the sample data.

4. The neural state recognition method based on EMG signals according to claim 3, characterized in that: The acquisition of electromyographic detection signals includes: placing the acquisition electrodes of the neurophysiological acquisition device on the surface of the acquisition muscle group of the subject; the acquisition muscle group includes: left abductor pollicis, left tibialis anterior, left adductor pollicis, right tibialis anterior, and right adductor pollicis.

5. The neural state recognition method based on EMG signals according to claim 4, characterized in that: Extracted features include time-domain features, frequency-domain features, kurtosis features, and nonlinear features; The time-domain features include maximum value, minimum value, and RMS; among them, the maximum value and minimum value are used to reflect the degree of muscle activation; RMS is the root mean square, used to measure the overall energy level. Frequency domain characteristics include median frequency and average power frequency; Kurtosis: Used to reflect the sharpness of a signal; In kurtosis features, MP is the median frequency, and MPF is the average power frequency. MP and MPF are used to reflect the spectral distribution and reveal the changes in the neural drive frequency band. Nonlinear features include sample entropy and permutation entropy: SampEn is the sample entropy and PE is the permutation entropy. SampEn and PE are used to describe the complexity and nonlinear features of a signal and to help identify abnormal activation patterns.

6. The neural state recognition method based on EMG signals according to claim 5, characterized in that: The algorithm for Root Mean Square (RMS) is as follows: The root mean square (RMS) is used to reflect the average power of the signal; where, Let N be the i-th sampling point of the signal; N is the signal length. The algorithm for integral electromyography (IEMG) values ​​is as follows: Integrated electromyography (EMG) values ​​are used to reflect the degree of muscle activation. The algorithm for kurtosis (SK) is as follows: Kurtosis is used to reflect the sharpness of a signal; The mean of the signal within the window; the algorithms for the maximum value (Max) and minimum value (Min) are as follows: , This is used to reflect the amplitude range of the signal; The algorithm for the average power frequency (MPF) is as follows: The average power frequency is used to reflect the median frequency of the power spectrum, where... For the k-th frequency component, for The power spectral density at point M, where M is the total number of frequency components; The algorithm for the median frequency (MF) is as follows: Calculate the power spectral density using Fast Fourier Transform (FFT) and find the frequency that satisfies the following conditions. , The median frequency is used to reflect the average frequency of the power spectrum. The algorithm for sample entropy SampEn is as follows: Embedding dimension m=2, matches is the number of matches in the dimension, and sample entropy is used to reflect the complexity of the signal; parameters of sample entropy SampEn: tolerance r=0.2*std(window), the tolerance is 20% of the standard deviation of the current window signal, and time delay τ=1; The algorithm for fractal dimension PFD is as follows: N: Signal length The number of directional changes in a signal, i.e., the number of inflection points; fractal dimension reflects the degree of irregularity of the signal.

7. A neural state recognition device based on EMG signals, characterized in that: Includes memory used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the neural state recognition method based on EMG signals as described in any one of claims 1-5.

8. A neural state recognition storage medium based on EMG signals, wherein a computer program / instruction is stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the neural state recognition method based on EMG signals as described in any one of claims 1-6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the neural state recognition method based on EMG signals as described in any one of claims 1-6.

Images