Motion intention recognition method based on single-channel surface electromyogram signal decomposition

Abstract

The invention relates to the technical field of bio-electricity signal processing, in particular to a motion intention recognition method for single-channel surface electromyogram (sEMG) decomposition. Single-channel sEMG preprocessing: processing the acquired sEMG in sequence by adopting a mode of combining band-pass filtering and second-order differential filtering; single-channel sEMG signal decomposition: extracting a plurality of motion unit action potential sequences (MUAPT) from the mixed signal, then accurately classifying the sequences, and finally performing peak extraction; and motion intention recognition based on MUAPT: key features are extracted based on a motion unit action potential sequence (MUAPT), and after feature processing is performed by using an LSTM layer and a Softmax layer of an LSTM network, an accurate recognition result is output, so that recognition of different muscle actions is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of bioelectric signal processing, and in particular to a method for identifying movement intention by decomposing a single-channel surface electromyography (sEMG) signal. Background Art

[0002] Surface electromyography (sEMG) is a resource of the human body, containing rich information related to human movement. It has a natural advantage to use it as an interactive medium to build a human-computer interaction system. The key to achieving natural human-computer interaction through electromyography is to identify the human body's movement intention from the electromyography signal. The multi-channel surface electromyography (sEMG) movement intention recognition method based on time-frequency domain features is currently the mainstream method for analyzing movement intention using sEMG. The existing technology of this method mainly uses time-domain feature and frequency-domain feature extraction, multi-channel signal acquisition and traditional machine learning methods and feature dimension reduction method recognition algorithms. However, this method has some technical defects: (1) Feature extraction level: i. Time-frequency domain features fail to fully reflect the physiological mechanism of muscle activity; ii. Feature selection relies on experience; (2) Multi-channel signal acquisition and processing level: Electromyography signals detected by adjacent electrodes often come from similar muscle groups, resulting in a large redundancy in the observed signals; (3) Recognition algorithm level: Traditional statistical models have limited recognition capabilities for complex movement patterns. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for identifying movement intention based on single-channel surface electromyography signal decomposition to solve the problems existing in the above-mentioned background technology.

[0004] The specific method steps of the present invention include the following:

[0005] S1. Single-channel sEMG preprocessing: The collected sEMG data are processed sequentially using a combination of bandpass filtering and second-order differential filtering to obtain preprocessed data.

[0006] S2. Single-channel sEMG signal decomposition: extracting multiple motor unit action potential trains (MUAPTs) from the mixed signal, then accurately classifying these trains and finally performing spike extraction;

[0007] S3. Movement intention recognition based on motor unit action potential train (MUAPT): By extracting key features based on the motor unit action potential train (MUAPT), the key features are processed using the LSTM layer and Softmax layer of the LSTM network. At the same time, by calculating the target equation of the average error, the parameters of each gated unit in the LSTM unit in the LSTM layer are continuously updated through iterative gradients. Finally, accurate recognition results are output to achieve the recognition of different muscle movements.

[0008] Preferably, the extracting of multiple motor unit action potential sequences is specifically to identify and extract motor unit action potentials (MUAPs) by setting a threshold;

[0009] When the amplitude of the sEMG signal exceeds the preset threshold, the signal at that moment is considered a valid MUAP peak;

[0010] When the amplitude is lower than the threshold, it is considered as noise and its value is set to zero.

[0011] Furthermore, the threshold

[0012] in, is the threshold, 1.5 experience points, is the filtered sEMG signal, Indicates the index of the sEMG signal, is the total length of the filtered sEMG signal.

[0013] Preferably, the threshold is determined After that, two values ​​are drawn on the filtered EMG curve. And the baseline is parallel to the time axis. It will cross the positive baseline from bottom to top Or cross the negative baseline from top to bottom The horizontal coordinates of the points are marked as , the next one goes from top to bottom Or cross from bottom to top The horizontal coordinates of the points are marked as .exist There is a peak or valley between the two, and the horizontal coordinate of the peak or valley is marked as . Set the horizontal axis to The data point and its 11 adjacent data points are regarded as a MUAP, and its expression is:

[0014]

[0015] in, is the index of MUAP, After filtering in step S1, the total extracted signal MUAPs, and get the MUAP set MUAPs, that is .

[0016] Preferably, before extracting the key features, the mean absolute value and the number of zero-crossing points of MUAPT are calculated;

[0017] Absolute mean: index is MUAPT data The MAV is calculated as:

[0018]

[0019] Zero crossing number: index is MUAPT data The calculation formula of ZC is:

[0020]

[0021] In the above two formulas, Indicates the MUAPT data length, Indicates the length is The sliding window, , Represents the MUAPT index, Indicates the starting index of the sliding window, Indicates the signal index in the current window, Indicates the of the MUAPT categories elements, Indicates the of the MUAPT categories elements, is a symbolic function defined as .

[0022] Preferably, the LSTM layer is composed of LSTM units;

[0023] Furthermore, The input of the LSTM unit at this moment includes the input at the current moment , the action input at the previous moment And unit input , output variables include action output and unit output ;

[0024] The output variable is calculated through three gate control units. The specific calculation formula is:

[0025]

[0026] in, represents the Sigmoid function, Indicates the unit status, represents the forget gate, represents the input gate, represents the output gate, express The hidden layer input at time t, express EMG input at all times, represents the LSTM unit weight of each gate, -Unit state weight, -Forget gate weight, -Input gate weight, - output gate weight, Represents the LSTM unit bias of each gate: -Cell state bias, -Forget gate bias, -Input gate bias, - output gate bias;

[0027] LSTM unit output at this moment Input from the previous unit and the current unit status Determine together, the final output vector can be obtained after the output gate transformation ,Right now:

[0028]

[0029] in, Represents element-by-element multiplication, and the output vector can be expressed as , 、 、 The feature representations correspond to the three actions of walking, climbing stairs, and descending stairs respectively.

[0030] Preferably, the input , superscript Represents the index of MUAPT, for Input at the moment;

[0031] Furthermore, the variable matrix input to the LSTM layer is represented as follows:

[0032] Among them, the superscript Represents the index of MUAPT ( ), subscript Represents the moment ( ).

[0033] Preferably, after the key features are processed by the Softmax layer, a probability distribution vector is obtained, and the calculation formula is:

[0034]

[0035] 、 、 Indicates The probability of the moment sample being predicted as walking, going up stairs, or going down stairs, is a natural constant, The exponential function with base is always greater than 0. Predicting results for the LSTM network In order to make the sum of all prediction results equal to 1, the prediction results are converted into probability form, that is, .

[0036] Compare the three probability values ​​to determine the recognition result of the current model.

[0037] Preferably, the average error target equation of the output vector is:

[0038]

[0039] For real action categories, is the error function, when Prediction results at the moment When the category is the same as the real action, ,on the contrary ; T represents the total time, It is the action category calculated by the LSTM network.

[0040] Preferably, the parameters of each gate unit in the LSTM unit are calculated using the back propagation algorithm based on the average error of the output vector. 、 The gradient of 、 ),in, include 、 、 、 , include 、 、 、 ;

[0041] Update, the update formula is as follows:

[0042]

[0043] in is the learning rate, represents the LSTM unit weight before each gate is updated, represents the LSTM unit weight after each gate is updated, represents the LSTM unit bias before each gate update, Represents the LSTM unit bias after each gate update.

[0044] Technical effects and advantages of the present invention:

[0045] 1. Reduced cost and complexity: The use of single-channel sEMG signal acquisition significantly reduces device cost and operational complexity, making the device easier to wear and use, and suitable for more scenarios, such as home rehabilitation training.

[0046] 2. Improved recognition accuracy: By deeply decomposing and comprehensively extracting features from single-channel sEMG signals, the system fully exploits the effective information in the signals. Combined with advanced recognition models, this significantly improves the accuracy of movement intention recognition, providing a more precise control basis for rehabilitation treatment and human-computer interaction.

[0047] 3. Promote the development of related fields: The application of this method will help promote technological progress in rehabilitation medicine, intelligent prostheses, human-computer interaction and other fields, improve the quality of life of patients and the level of intelligence of human-computer interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of the method for identifying movement intention based on single-channel surface electromyography signal decomposition of the present invention;

[0049] Figure 2 This is the filtering result diagram of the present invention; corresponding to (S1) and (S2) in single-channel preprocessing, the three columns of signals from left to right are the original signal (left), the signal after bandpass filtering (middle), and the signal after differential filtering (right). Each column of signals has eight columns, from top to bottom, EMG1 to EMG8, and the vertical axis of each column of signals represents the signal amplitude ( ), the horizontal axis represents time ( );

[0050] Figure 3 This is the peak detection result diagram of the present invention. The black signal in the figure is the single-channel EMG after preprocessing, and the dotted line represents the threshold The dots in the figure below represent the detected MUAP peaks. The vertical axes of the two figures above and below represent the amplitude ( ), the horizontal axis represents time ( );

[0051] Figure 4Schematic diagram of the identification model structure in the present invention;

[0052] Figure 5 Schematic diagram of the LSTM unit structure in the present invention;

[0053] Figure 6 Schematic diagram of the single-channel sEMG acquisition points in the present invention, wherein EMG1 corresponds to the rectus femoris, EMG2 corresponds to the medial broad muscle, EMG3 corresponds to the lateral broad muscle, EMG4 corresponds to the Achilles tendon muscle, EMG5 corresponds to the biceps femoris, EMG6 corresponds to the medial gastrocnemius, EMG7 corresponds to the lateral gastrocnemius, and EMG8 corresponds to the gastrocnemius; DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0055] The acquisition device in this embodiment adopts the Cometa PicoEMG wireless surface electromyography test system with a sampling frequency of 2000 Hz.

[0056] In this embodiment, the lower limb muscles are used as experimental objects to collect 8 channels of electromyographic data EMG1~EMG8 (such as Figure 6 The subjects were asked to collect signals from three movements: walking, climbing stairs, and descending stairs. Each movement was repeated four times, with a 2-minute interval between each movement to ensure that the collected signal quality was not affected by muscle fatigue. Single-channel decomposition and motion extraction were performed on each channel of the EMG for each movement.

[0057] The specific operations are as follows:

[0058] S1. Single-channel sEMG preprocessing

[0059] The sEMG signal collected from a single muscle mass is susceptible to the vibration of surrounding muscles, and sEMG itself has non-stationary characteristics. To eliminate interference from bioelectric signals and amplify the effective signal portion to address non-stationarity, this embodiment uses a combination of bandpass filtering and second-order differential filtering for preprocessing.

[0060] (1) Bandpass filtering: A 4th-order Butterworth bandpass filter (20 Hz to 450 Hz) is used on the sEMG signal to remove high-frequency noise interference. The filter transfer function is as follows:

[0061]

[0062] in, is a complex frequency variable, is the cutoff frequency of the filter, is the filter order.

[0063] (2) Second-order differential filtering: The second-order differential filtering is used to enhance the peak characteristics of the signal, which is convenient for the subsequent decomposition steps. The original observation signal after bandpass filtering is ,in is the collected sEMG signal, is the signal index, and the total length of the signal is The signal after second-order differential filtering is ,length , the expression is:

[0064]

[0065] Through this step, we can get Figure 2 The filtering result is shown in the figure.

[0066] S2. Single-channel sEMG signal decomposition

[0067] The core task of single-channel sEMG signal decomposition is to extract multiple motor unit action potential trains (MUAPTs) from the mixed signal and accurately classify these trains.

[0068] Based on the physiological generation mechanism of sEMG signals, the amplitude of the electrical signal generated by muscles in an active state is significantly higher than that in a resting state. Therefore, motor unit action potentials (MUAPs) can be identified and extracted by setting a threshold. Specifically, when the amplitude of the sEMG signal exceeds the preset threshold, the signal at that moment is considered a valid MUAP peak; the portion with an amplitude below the threshold is considered noise and its value is set to zero.

[0069] The threshold is calculated as follows:

[0070]

[0071] in, is the threshold; 1.5 is the empirical value; The sEMG signal after filtering in step S1 is: Indicates the index of the sEMG signal, is the total length of the filtered sEMG signal.

[0072] Determine the threshold After that, two values ​​are drawn on the filtered EMG curve. And the baseline is parallel to the time axis. It will cross the positive baseline from bottom to top Or cross the negative baseline from top to bottom The horizontal coordinates of the points are marked as , the next one goes from top to bottom Or cross from bottom to top The horizontal coordinates of the points are marked as .exist There is a peak or valley between the two, and the horizontal coordinate of the peak or valley is marked as . Set the horizontal axis to The data point and its 11 adjacent data points are regarded as a MUAP, and its expression is:

[0073]

[0074] in, is the index of MUAP, After filtering in step S1, the total extracted signal MUAPs, and get the MUAP set MUAPs, that is .

[0075] The k-means method was used to cluster MUAPs and we obtained Different categories ( ), No. The clustering results are Indicates that is the index of the clustering result, For the Class Results , the center of mass is defined as .

[0076] The specific steps are shown in Algorithm 1:

[0077] Algorithm 1: MUAP clustering algorithm based on K-means Input: MUAP set #timg#, number of clusters #timg#, convergence threshold #timg#, maximum number of iterations #timg# Output: clustering result #timg# Initialize cluster center: Randomly select #timg# #timg# from #timg# as the initial centroid #timg# Initialize the number of iterations #timg#while #timg# do Clear the results in all clusters #timg#for #timg#in#timg# do Calculate Euclidean distance: #timg# Traverse #timg# from #timg# to #timg# and find the centroid #timg# closest to #timg#, whose index is #timg# Add #timg# to the corresponding cluster #timg# Save the current centroid #timg#, #timg# for #timg#to #timg# do if #timg# is empty: Skip else: Initialize a new cluster center #timg# with a dimension of 12, the same as #timg#for #timg# in #timg# do for #timg# to #timg# do #timg# #timg#, #timg# represents the number of MUAPs contained in #timg# Calculate the change in cluster center #timg# if #timg#: Jump out of the loop and get the center of mass#timg# #timg#

[0078] According to Algorithm 1, we can get Different categories Each category contains several MUAPs. The MUAP samples in each category are sorted according to the sequence obtained after step S1. By rearranging the positions of MUAP trains (MUAP trains, MUAPT) are arranged in sequence. The expression of MUAPT set is: ,in It is MUAPT in the class.

[0079] After step S2, we can get Figure 3 The spike detection results are shown in Figure 4.

[0080] S3. Movement Intention Recognition Based on MUAPT

[0081] This example proposes an action recognition method based on motor unit action potential train (MUAPT) feature extraction and long short-term memory (LSTM) neural network. This method first extracts key features from MUAPT and then uses the LSTM network to process the feature samples to accurately identify different muscle movements.

[0082] MUAPT is characterized by nonstationarity and sparsity, while also containing rich physiological information about electromyographic signals. Feature extraction effectively captures stationary information related to movement, providing a reliable basis for movement recognition. Specifically, this example extracts the following features from MUAPT.

[0083] Before extracting features, for any channel of MUAPT data, the length of MUAPT data is known to be , set the length to The sliding window ( ).

[0084] (1) Absolute mean

[0085] Index is MUAPT data The MAV calculation formula is:

[0086]

[0087] (2) Zero-crossing points

[0088] Index is MUAPT data The ZC calculation formula is:

[0089]

[0090] In the above two formulas, Represents the MUAPT index, Indicates the starting index of the sliding window, Indicates the signal index in the current window, Indicates the of the MUAPT categories elements, Indicates the of the MUAPT categories elements, is the data length of MUAPT, Set the length to The sliding window ( ); is a symbolic function defined as .

[0091] This embodiment proposes a motion intention recognition model with LSTM network as the main unit. The basic structure of the network is as follows: Figure 4 Shown: LSTM network input matrix It consists of MAV and ZC. After being processed by the recognition model, the recognition result corresponding to the channel (walking, going up stairs, going down stairs) will be obtained.

[0092] The recognition model consists of an LSTM layer and a Softmax layer.

[0093] The input matrix expression is as follows:

[0094]

[0095] Among them, the superscript Represents the index of MUAPT ( ), subscript Represents the moment ( ). for LSTM network input at time t, .

[0096] The LSTM layer is composed of LSTM units;

[0097] Furthermore, The input of the LSTM unit at this moment includes the input at the current moment , the action input at the previous moment And unit input , output variables include action output and unit output ;

[0098] The output variable is calculated through three gate control units. The specific calculation formula is:

[0099]

[0100] in, represents the Sigmoid function, Indicates the unit status, represents the forget gate, represents the input gate, represents the output gate, express The hidden layer input at time t, express EMG input at all times, represents the LSTM unit weight of each gate, -Unit state weight, -Forget gate weight, -Input gate weight, - output gate weight, Represents the LSTM unit bias of each gate: -Cell state bias, -Forget gate bias, -Input gate bias, - output gate bias;

[0101] LSTM unit output at this moment Input from the previous unit and the current unit status Determine together, the final output vector can be obtained after the output gate transformation ,Right now:

[0102]

[0103] in, Represents element-by-element multiplication, and the output vector can be expressed as , 、 、 The feature representations correspond to three actions respectively. In this embodiment, the three actions are {walking, climbing stairs, and descending stairs}.

[0104] After the key features are processed by the Softmax layer, the probability distribution vector is obtained. The calculation formula is:

[0105]

[0106] 、 、 Indicates The probability of the moment sample being predicted as walking, going up stairs, or going down stairs, is a natural constant, The exponential function with base is always greater than 0. Predicting results for the LSTM network In order to make the sum of all prediction results equal to 1, the prediction results are converted into probability form, that is, .

[0107] Compare the three probability values ​​to determine the recognition result of the current model.

[0108] For example, if , , , then the model believes that the action at the current moment is most likely to be "walking", that is, the LSTM output result is "walking".

[0109] In this embodiment, For real action categories, is the action category calculated by the LSTM network, and the average error target equation of the output vector is:

[0110]

[0111] For real action categories, is the error function, when Prediction results at the moment When the category is the same as the real action, ,on the contrary ; T represents the total time, It is the action category calculated by the LSTM network.

[0112] While processing data for the LSTM network, in order to reduce the error, the embodiment uses the back propagation algorithm to calculate the parameters of each gate unit in the LSTM unit according to the average error of the output vector. 、 The gradient of 、 ),in, include 、 、 、 , include 、 、 、 ;

[0113] Update, the update formula is as follows:

[0114]

[0115] in is the learning rate, represents the LSTM unit weight before each gate is updated, represents the LSTM unit weight after each gate is updated, represents the LSTM unit bias before each gate update, Indicates the LSTM unit bias after each gate is updated. In this embodiment, the learning rate is 0.01, and the gate unit parameters are updated by gradient so that the error is continuously Reduce it to get more accurate recognition results.

[0116] The role of gradient update is to optimize the LSTM network. In the initial state, the weights in the LSTM network and bias is formed randomly, which will lead to a large error between the predicted result and the real action category. By updating the weights and biases, the error can be As the learning ability of the LSTM network for the complex mapping relationship between EMG signals and action categories is gradually enhanced, the high-precision recognition of the actions corresponding to EMG signals can be achieved.

[0117] Finally, the recognition results are as follows (taking the EMG3 signal as an example):

[0118] Action Category Training accuracy Test accuracy walk 0.985 0.990 Go up the stairs 0.901 0.892 Go down the stairs 0.897 0.878

[0119] Table 1

[0120] Note: All experimental data are divided into training set and test set in a ratio of 7:3. The training set data is used to update the parameters of the LSTM network, and the trained network is used to directly perform action recognition on the test set data.

[0121] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A motion intention recognition method based on single-channel surface electromyography signal decomposition, characterized in that: The steps include: S1. Single-channel sEMG preprocessing: The collected sEMG is processed sequentially using a combination of bandpass filtering and second-order differential filtering. S2. Single-channel sEMG signal decomposition: Extract multiple motor unit action potential trains (MUAPs) from the mixed signal, then accurately classify the MUAPTs and perform spike extraction. S3. Movement intention recognition based on motor unit action potential train (MUAPT): By extracting key features based on the motor unit action potential train (MUAPT), the key features are processed using the LSTM layer and Softmax layer of the LSTM network. At the same time, by calculating the target equation of the average error, the parameters of each gated unit in the LSTM unit in the LSTM layer are continuously updated through iterative gradients. Finally, accurate recognition results are output to achieve the recognition of different muscle movements.

2. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 1, characterized in that: The extracting of multiple motor unit action potential sequences specifically includes identifying and extracting motor unit action potentials (MUAPs) by setting a threshold; When the amplitude of the sEMG signal exceeds the preset threshold, the signal at that moment is considered a valid MUAP peak; When the amplitude is lower than the threshold, it is considered as noise and its value is set to zero.

3. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 2, characterized in that: The threshold in, is the threshold, 1.5 experience points, is the filtered sEMG signal, Indicates the index of the sEMG signal, is the total length of the filtered sEMG signal.

4. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 3, characterized in that: Determine the threshold After that, two values ​​are drawn on the filtered EMG curve. And the baseline is parallel to the time axis; it will cross the positive baseline from bottom to top Or cross the negative baseline from top to bottom The horizontal coordinates of the points are marked as , the next one goes from top to bottom Or cross from bottom to top The horizontal coordinates of the points are marked as ;exist There is a peak or valley between the two, and the horizontal coordinate of the peak or valley is marked as ; Set the horizontal axis to The data point and its 11 adjacent data points are regarded as a MUAP, and its expression is: in, is the index of MUAP, After filtering in step S1, the total extracted signal MUAPs, and get the MUAP set MUAPs, that is .

5. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 1, characterized in that: Before extracting the key features, the absolute value mean and the number of zero-crossing points of MUAPT are calculated; Absolute mean: index is MUAPT data The MAV is calculated as: Zero crossing number: index is MUAPT data The calculation formula of ZC is: In the above two formulas, Indicates the MUAPT data length, Indicates the length is The sliding window, , Represents the MUAPT index, Indicates the starting index of the sliding window, Indicates the signal index in the current window, Indicates the of the MUAPT categories elements, Indicates the of the MUAPT categories elements, is a symbolic function defined as .

6. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 5, characterized in that: The LSTM layer is composed of LSTM units; Furthermore, The input of the LSTM unit at this moment includes the input at the current moment , the action input at the previous moment And unit input , output variables include action output and unit output ; The output variable is calculated through three gate control units. The specific calculation formula is: in, represents the Sigmoid function, Indicates the unit status, represents the forget gate, represents the input gate, represents the output gate, express The hidden layer input at time t, express EMG input at all times, represents the LSTM unit weight of each gate, -Unit state weight, -Forget gate weight, -Input gate weight, - output gate weight, Represents the LSTM unit bias of each gate: -Cell state bias, -Forget gate bias, -Input gate bias, - output gate bias; LSTM unit output at this moment Input from the previous unit and the current unit status Determine together, the final output vector can be obtained after the output gate transformation ,Right now: in, Represents element-by-element multiplication, and the output vector can be expressed as , 、 、 The feature representations correspond to the three actions of walking, climbing stairs, and descending stairs respectively.

7. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 6, characterized in that: The input , superscript Represents the index of MUAPT, for Input at the moment; Furthermore, the variable matrix input to the LSTM layer is represented as follows: Among them, the superscript Represents the index of MUAPT ( ), subscript Represents the moment ( ).

8. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 6, characterized in that: After the key features are processed by the Softmax layer, a probability distribution vector is obtained, and the calculation formula is: 、 、 Indicates The probability of the moment sample being predicted as walking, going up stairs, or going down stairs, is a natural constant, The exponential function with base is always greater than 0. Predicting results for the LSTM network In order to make the sum of all prediction results equal to 1, the prediction results are converted into probability form, that is, ; Compare the three probability values ​​to determine the recognition result of the current model.

9. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 8, characterized in that: The average error objective equation of the output vector is: For real action categories, is the error function, when Prediction results at the moment When the category is the same as the real action, ,on the contrary ; T represents the total time, It is the action category calculated by the LSTM network.

10. The method for identifying movement intention based on single-channel surface electromyography signal decomposition according to claim 9, characterized in that: According to the average error of the output vector, the back propagation algorithm is used to calculate the parameters of each gate unit in the LSTM unit. 、 The gradient of 、 ),in, include 、 、 、 , include 、 、 、 ; Update, the update formula is as follows: in is the learning rate, represents the LSTM unit weight before each gate is updated, represents the LSTM unit weight after each gate is updated, represents the LSTM unit bias before each gate update, Represents the LSTM unit bias after each gate update.

Images