Method for executing output action by using electromyographic signal and action output system

Abstract

The invention belongs to the technical field of electromyographic signal detection, and particularly relates to a method for executing output actions by using electromyographic signals and an action output system. According to the method for executing the output action through the electromyographic signals, under the condition that the number of collection channels is limited, more recognition models are obtained through grouping, sorting and channel combination related to the control action, the number of the recognition models is increased, and more training normal forms can be met; therefore, the online detection result of the muscle activity is converted into the corresponding control instruction so as to output more control actions.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromyographic signal detection, and in particular relates to a method for executing output actions using electromyographic signals and an action output system. Background Art

[0002] Surface electromyography is a physiological signal that is closely related to human movement intention. Gesture recognition based on sEMG is of great significance to the research of human-computer interaction such as prosthetic control, industrial generation, virtual reality and rehabilitation medicine.

[0003] In order to effectively capture the characteristics of electromyographic signals under different gestures, electrodes are often placed near or around the forearm of a single arm, relying on the acquisition channel of the electromyographic signal and its corresponding acquisition position to complete the detection and gesture recognition, especially the output of complex gestures requires more channels in time sequence recognition results. In order to meet the output requirements of gesture decisions, the features of the signal itself are used in related technologies to increase the number of recognition results. For example, the patent CN116400812B is based on the surface electromyographic signal emergency rescue gesture recognition method and device as follows: the two-stage gesture recognition model includes a first feature extraction module, a gesture category decision module, a plurality of second feature extraction modules corresponding to different gesture categories, and a gesture action decision module connected after each second feature extraction module. Among them, the two-stage gesture recognition model is used to identify the gesture category of the input multi-channel surface electromyographic signal time-frequency diagram, and then recognize the gesture action according to the recognition result of the gesture category; the two-stage gesture recognition model is based on the sample multi-channel surface electromyographic signal time-frequency diagram with gesture type and gesture action labels. Summary of the invention

[0004] The present invention provides a method for executing output actions using electromyographic signals and an action output system, which meet the execution requirements of complex actions by increasing the number of recognition models.

[0005] In order to solve the above technical problems, the present invention provides a method for using electromyographic signals to perform output actions, including: setting channel combinations, that is, sorting and grouping acquisition channels according to control actions; setting model sets, that is, using offline electromyographic signals to perform model training according to different channel combinations to obtain recognition models and classify them into different model sets; loading models, that is, selecting corresponding model sets for loading according to control actions; online detection, traversing all loaded recognition models to detect online electromyographic signals to obtain different detection results; decision calling, smoothly outputting the detection results to call corresponding control instructions; control output, executing corresponding control actions based on control instructions.

[0006] Furthermore, the recognition model includes: a single-channel combination recognition model, a multi-channel combination fusion model, and a multi-channel combination arrangement model; wherein the single-channel combination recognition model is configured to be a recognition model obtained by model training a single channel combination; the multi-channel combination fusion model is configured to be a recognition model obtained by model training after fusing the signal features of multiple channel combinations; the multi-channel combination arrangement model is configured to be a recognition model obtained by model training after sorting the signal features of multiple channel combinations in time sequence.

[0007] Furthermore, the setting of the channel combination includes grouping and sorting the collection channels and collection parts according to the order corresponding to the execution of the control action.

[0008] Furthermore, the model training includes: detecting electromyographic burst activity, that is, detecting offline electromyographic signals to obtain the starting data segment of the electromyographic signal burst activity and its corresponding quantitative indicators or features; data slicing, that is, slicing the starting data segment, and marking the quantitative indicators or extracting the features of the data segment one by one to obtain model parameters; cross-validation to complete model training.

[0009] Furthermore, the detection of electromyographic burst activity includes: constructing a feature matrix based on characteristic parameters of offline electromyographic signals; extracting principal components in the projection space based on the feature matrix; performing classification tasks on the principal components based on a Gaussian mixture model; and obtaining the quantitative index or feature based on the classification results; wherein the characteristic parameters include: a combination of at least one of an autoregressive coefficient, a time domain feature, and a frequency domain feature.

[0010] Furthermore, the detection of myoelectric burst activity also includes: enhancing the characteristics of myoelectric burst, that is, extracting the first-order differential energy operator of the myoelectric signals of all acquisition channels in the channel combination and summing them before detecting the myoelectric burst activity.

[0011] Further, the decision call includes: setting the interval for obtaining the detection results to be a first time step, setting the interval for smoothing the output to be a second time step, and setting the interval for calling the control instruction to be a third time step, wherein the first time step ≤ the second time step ≤ the third time step; at the first time step, obtaining the real-time detection result each time; at the second time step, converting multiple real-time detection results into a decision result of a single smooth output according to the voting principle or prior probability; at the third time step, mapping the multiple decision results into corresponding control instructions.

[0012] In the second aspect, the present invention provides an action recognition system based on electromyographic signals, comprising: a processor unit, which runs the action recognition method to detect online electromyographic signals; a model database, which is connected to the processor unit and stores different model sets; a decision maker unit, which is connected to the processor unit and is used to output the decision results of the online electromyographic signals to call corresponding control instructions; and a controller unit, which is connected to the processor unit and executes corresponding control actions according to the control instructions.

[0013] In a third aspect, the present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0014] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method when executed by a processor.

[0015] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which implements the steps of the method when executed by a processor.

[0016] The beneficial effect of the present invention is that the method of using electromyographic signals to perform output actions of the present invention sets the channel combination according to the number, position, and order of the acquisition channels required for the control action, and then performs model training respectively to obtain multiple recognition models and classify them into different model sets. When detecting online electromyographic signals, the loaded recognition model is used for detection, and then the detection results are smoothly output into corresponding control instructions to execute corresponding control actions. In the case of a limited number of acquisition channels, more recognition models are obtained by grouping and sorting the channel combinations related to the control action, which increases the number of recognition models and can meet more training paradigms, thereby converting the online detection results of muscle activity into corresponding control instructions to output more control actions.

[0017] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a flow chart of using electromyographic signals to execute output actions.

[0021] Figure 2 It is a flowchart for obtaining a single-channel combination recognition model.

[0022] Figure 3 It is a flowchart for obtaining a multi-channel combination fusion model.

[0023] Figure 4 It is a flowchart for obtaining a multi-channel combination arrangement model.

[0024] Figure 5 It is the principle block diagram of the action recognition system.

[0025] Figure 6 is a schematic diagram of the five gesture categories.

[0026] Figure 7 This is the result of detecting the burst starting point of the electromyographic signal collected from the left forearm using the hard threshold method.

[0027] Figure 8 This is the result of detecting the burst starting point of the electromyographic signal collected from the right forearm using the hard threshold method.

[0028] Fig. 9 This is the detection result diagram of the left arm recognition model.

[0029] Fig.10 This is the detection result diagram of the right arm recognition model.

[0030] Fig.11 It is the detection result diagram of the multi-channel combination fusion model.

[0031] Fig.12 It is a graph of the detection results of the multi-channel combination arrangement model under the condition of timing synchronization.

[0032] Fig.13 The calculation results of the cumulative variance of the principal components are shown.

[0033] Fig.14 Gaussian classification results based on the characteristics of the autoregressive model parameters estimated by Burg's method are shown.

[0034] Fig.15The Gaussian classification results of the first two principal components extracted in the projected space are shown.

[0035] Fig.16 The results show the detection results of the starting and ending moments of the EMG signal of the flexor digitorum profundus when the left hand makes a fist. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Example 1.

[0038] In the related art, whether it is based on sparse surface electromyographic signals or high-density electromyographic signals, it is difficult to form more recognition models by only collecting the number of channels to meet the interpretation and execution of control actions for multi-category gesture tasks. Figure 1-Figure 5 , this embodiment 1 provides a method for executing output actions using electromyographic signals, including the following: setting channel combinations, that is, sorting and grouping acquisition channels according to control actions; setting model sets, that is, performing model training using offline electromyographic signals according to different channel combinations, obtaining recognition models and classifying them into different model sets; loading models, that is, selecting corresponding model sets for loading according to control actions; online detection, traversing all loaded recognition models to detect online electromyographic signals to obtain different detection results; decision calling, outputting the detection results smoothly to call corresponding control instructions; control output, executing corresponding control actions based on control instructions.

[0039] As an optional implementation for setting the channel combination.

[0040] The channel combination is set up including grouping and sorting the acquisition channels and acquisition parts according to the order corresponding to the execution of the control action. Generally, the acquisition channel can be freely placed at the target muscle position to collect the myoelectric information of a specific muscle group. The specific muscle group is not limited to the forearm and hind arm, lower limbs and other muscle groups of the arm, and the corresponding channel combination is formed by the muscle group as a group. Taking the arm as an example, the acquisition channels of the left forearm (such as the ulnar flexor carpi, ulnar extensor carpi, radial extensor carpi brevis) are set as the left arm channel combination; the acquisition channels of the right forearm (such as the ulnar flexor carpi, ulnar extensor carpi, radial extensor carpi brevis) are set as the right arm channel combination.

[0041] As an optional implementation of setting the model collection.

[0042] First, we need to obtain the corresponding recognition model through model training according to different channel combinations. Figure 2 , Figure 3 , Figure 4 ,The basic process of model training includes: (1) Preprocessing: removing DC components and power frequency interference from offline electromyographic signals. For example, but not limited to, high-pass filter, 50 Hz notch filter, periodic difference, etc.

[0043] (2) Detecting the starting point of the EMG burst activity after envelope smoothing. The detection method of the starting point includes, for example but not limited to, hard threshold detection based on envelope amplitude or starting point detection based on unsupervised clustering under a feature matrix. The starting point detection based on unsupervised clustering under a feature matrix includes: constructing a feature matrix based on the feature parameters of the offline EMG signal; extracting the principal components in the projection space based on the feature matrix; performing classification tasks on the principal components based on a Gaussian mixture model; extracting the quantitative index based on the classification results; wherein the feature parameters include: a combination of at least one of: autoregressive coefficients, time domain features (such as statistics based on EMG amplitude), frequency domain features (such as band energy, median frequency, mean frequency based on wavelet transform or Fourier transform), etc. Specifically, the feature matrix is ​​an m*n matrix, where m represents the observed quantity and n represents the feature dimension. When the characteristic parameter is only the autoregressive coefficient, the characteristic matrix is ​​an m*1 matrix; when the characteristic parameter is the autoregressive coefficient and the time domain feature or the frequency domain feature, the characteristic matrix is ​​an m*2 matrix; when the characteristic parameter is the autoregressive coefficient and the time domain feature or the frequency domain feature, the characteristic matrix is ​​an m*3 matrix.

[0044] Optionally, the autoregressive coefficient is obtained based on the non-stationary characteristics of the electromyographic signal. , further strengthen the feature description of the electromyographic signal. For example, select the third-order autoregressive model and obtain the autoregressive coefficients a1, a2, and a3.

[0045] Optionally, the time domain features are extracted based on the amplitude characteristics of the electromyographic signal, such as absolute amplitude energy , waveform length , mean square root , Hilbert transform envelope energy, etc.

[0046] Optionally, the frequency domain features are extracted from multiple sub-band energies of electromyographic signals under spectrum estimation. Since electromyographic activity is a high-frequency electrical activity in a wide frequency band, the frequency domain features include M sub-band energy extractions under spectrum estimation (not limited to parametric and non-parametric spectrum estimation), such as after non-parametric fast Fourier transform or according to the estimated autocorrelation function of the time series to perform autoregressive spectrum estimation of the time series, to achieve the logarithmic energy of the sub-bands. , is the imaginary value after fast Fourier transform in each sub-band. Represents the frequency after fast Fourier transform. For example, select four sub-bands: 0-45.25hz, 45.25-90.5hz, 90.5hz-135.75hz, 135.75hz-181hz.

[0047] Optionally, the projection space is composed of principal components whose cumulative variance energy is greater than a threshold. Principal component analysis (PCA) is a feature extraction and dimensionality reduction algorithm. The goal of PCA is to find a set of ordered orthogonal basis vectors that can capture the maximum variation in the data. The basis vectors are called principal components (PC). Specifically, the sample matrix X is set to m*n, indicating that there are m samples of n dimensions, and the sample matrix is ​​decentralized, X = X- , is the mean of the jth dimension of the sample. Calculate the covariance matrix C of X, C = . Perform SVD decomposition on the covariance matrix C to obtain the eigenvalues ​​λ and orthogonal eigenvectors P in the feature space, and project the sample matrix X into a new space based on the P matrix, Y = XP. Sort the eigenvalues ​​λ from large to small. The larger the λ, the greater the difference in the principal components in the projection space. Calculate the cumulative variance of each principal component ;See Figure 4 , when the sum of the contributions of N principal components is greater than the set threshold thd_c, the corresponding N principal components can be extracted based on the projection variance, eigenvalue, etc. of the basis vector. Fig.13 In the example, the threshold value thd_c is set to be, for example but not limited to, 95%.

[0048] Optionally, the task of classifying the principal components in the selected projection space is performed based on the Gaussian mixture model. The spatial clustering method may be a partition clustering method, a density clustering method, a hierarchical clustering method, etc. The Gaussian mixture model algorithm assumes that the sample points are generated by one or more Gaussian distributions, and estimates the parameters of the Gaussian distribution of each cluster by the maximum likelihood estimation method. The steps of solving the parameters of the Gaussian mixture model using the expectation maximization algorithm include: initializing the parameter model, such as the mean and covariance of each Gaussian distribution; in order to predict whether there is electromyographic burst activity, two mixed multi-dimensional Gaussian distributions can be designed. Calculate the responsiveness of all data points, that is, the probability that the data point belongs to each Gaussian distribution, and the data point is the principal component in the extracted projection space. Update the parameters of each Gaussian distribution. Calculate the likelihood function of the model. Iterate to determine whether the parameters of the model converge. If converged, end the iteration. According to the Gaussian mixture model, calculate the probability that the observed sample belongs to each cluster to complete the soft classification. Calculate the probability that the observed sample belongs to each cluster. According to the probability, determine the category of the observed sample and obtain the classification result of the principal component. Figure 5 As shown, in the original feature space, the first eigenvector is set as the autoregressive model parameter estimated based on the Burg method, expressed as Figure 5 In Figure 1, the second eigenvector is the autoregressive model parameter estimated based on the Burg method, expressed as Figure 5 Figure 2 in Figure 2, from Fig.14 It can be seen from the figure that in the original feature space, the distances between the two cluster centers of the binary Gaussian after Gaussian classification are relatively close. Fig.15 As shown, the Gaussian mixture model is unsupervisedly classified in the projection space to extract the corresponding principal components, namely the first principal component and the second principal component. Fig.15 It can be seen that the principal components extracted in the projection space effectively remove the redundancy of the variables in the feature space and improve the classification accuracy of the mixed Gaussian model. In summary: (1) For the principal components extracted in the projection space, the distance between each cluster center of the binary Gaussian after unsupervised training is farther and the probability of overlap is smaller; (2) In the original feature space, the distance between each cluster center of the binary Gaussian after unsupervised training is relatively close; (3) The principal components extracted in the projection space effectively remove the redundancy of the variables in the feature space and improve the classification accuracy of the mixed Gaussian model.

[0049] Optionally, based on whether there is a burst after classification by Gaussian mixture model, quantitative parameters such as the starting point, offset, and end point of the EMG burst activity are extracted. The quantitative indicators include but are not limited to: (1) the starting time Start_i of a single EMG burst activity; (2) the ending time Stop_i of a single EMG burst activity; (3) the number of EMG bursts of a certain duration, such as the number of bursts per minute; (4) the average EMG burst time and burst intensity under a certain duration; (5) statistical calculations of burst time and burst intensity, etc. For example, see Fig.16 , while making a fist with the left hand, Fig.16 The electromyographic signal in the figure is the electromyographic signal of the deep flexor muscle of the forearm. Fig.16 The Gaussian model classification in the figure is the result of detecting the EMG burst activity of the EMG signal based on the Gaussian mixture model, where a high level 1 represents the detection of muscle burst activity, and a low level 0 represents the absence of muscle activity. When it jumps from 0 to 1, it is recorded as the starting point of the burst 'start', and when it changes from 1 to 0, it is recorded as the end of the burst 'stop'. Calculate statistics, such as the number of bursts per minute, to extract quantitative parameters, such as the starting point, the length of a single burst, etc.

[0050] Optionally, before detecting the starting point of the myoelectric burst activity, the myoelectric burst feature may be enhanced, that is, the first-order differential energy operator is extracted from the myoelectric signals of all acquisition channels in the channel combination and the sum is calculated.

[0051] (3) Data slicing: slicing the starting point data segment of the electromyographic burst activity according to the time domain, marking the quantitative indicators or features one by one in the data segments, and obtaining the model parameters.

[0052] (4) Cross-validation: completing model training based on a machine learning algorithm or a deep learning algorithm. The machine learning algorithm is not limited to support vector machines, random forests, convolutional neural networks, etc.

[0053] Secondly, different recognition models are classified into recognition model sets with different attributes according to the control action or the location of the channel, so as to selectively load the corresponding recognition model during online detection.

[0054] Generally speaking, the recognition model of this case can be obtained by training the existing model using offline data, wherein the types of recognition models can be divided into: single-channel combination recognition model, multi-channel combination fusion model, and multi-channel combination arrangement model. The single-channel combination recognition model is configured as a recognition model obtained by model training on a single channel combination, such as the multi-channel combination fusion model is configured as a recognition model obtained by model training after fusing the signal features of multiple channel combinations. The multi-channel combination arrangement model is configured as a recognition model obtained by model training after sorting the signal features of multiple channel combinations in time sequence. This can greatly increase the number of recognition models in the recognition model set, including: n single-channel combination recognition models, 1 multi-channel combination fusion model, In addition, m of the n single-channel combination recognition models can be randomly selected for sorting to further expand the number of multi-channel combination recognition models. For example, the arm-related channel combinations can be classified into a gesture action recognition model set, such as single-channel combination recognition models such as the left arm recognition model and the right arm recognition model, a multi-channel combination fusion model obtained by combining the left arm and the right arm (only one in number), and a multi-channel combination arrangement model (multiple in number). In fact, the multi-channel combination arrangement model also contains multiple combinations. Taking the arm as an example, the left arm recognition model and the right arm recognition model form three multi-channel combination arrangement models in order, namely, left first, right first, and left and right synchronization, so as to realize the recognition of different actions. Therefore, the number of recognition models in the gesture action recognition model set is 2+1+3=6. Similarly, the leg action recognition model set or other action recognition model sets are classified.

[0055] In addition, the basic process of model training is different for different types of recognition models. The multi-channel combination arrangement model actually contains multiple single-channel combination recognition models sorted in time sequence. For example, when training a single-channel combination recognition model and a multi-channel combination arrangement model, the basic process of the two is the same. The difference is that the single-channel combination recognition model only needs to process the data of a single channel combination (such as Figure 2 As shown in Figure 2, the multi-channel combination arrangement model needs to process the data of different channel combinations independently and then sort them in time sequence (as shown in Figure 2). Figure 4 When training a multi-channel combination fusion model, it is necessary to perform fusion processing on the channel combination data (as shown in Figure 3 As shown), for example, when performing EMG burst feature enhancement, the channel energy operators of all single channel combinations involved in the multi-channel combination fusion model are summed.

[0056] As an optional implementation of decision calling.

[0057] The decision call includes: setting the interval for obtaining the detection result as the first time step, setting the interval for smooth output as the second time step, and setting the interval for calling the control instruction as the third time step, wherein the first time step < the second time step < the third time step; obtaining each real-time detection result at the first time step; converting multiple real-time detection results into a single smooth output decision result according to the voting principle or prior probability at the second time step; and mapping multiple decision results into corresponding control instructions at the third time step. Smoothly outputting the decision result can correct the gesture recognition result, eliminate the error of gesture class confusion due to the unstable state of the gesture switching process or the interference of background noise, and reduce the output of error information.

[0058] Optionally, the first time step is determined according to the muscle reaction time and can be set to 0.1 second; the second time step can be balanced according to the accuracy of the detection model and can be set to 1 second; the third time step needs to be determined according to the difficulty of the external control device or the control action, and can adopt the method of continuous output of control instructions or one-time output of control instructions, and the external control device can further output the control action according to the control instruction. The external control device can be, for example, a rehabilitation prosthesis, an unmanned aerial vehicle, a smart home, a voice broadcast, etc.

[0059] Example 2.

[0060] Based on Example 1, see Figure 5 , this embodiment 2 provides an action recognition system based on electromyographic signals, including: a processor unit, running the action recognition method to detect online electromyographic signals; a recognition model, connected to the processor unit, executing a training paradigm based on offline electromyographic signals to complete model training; a decision maker unit, connected to the processor unit, for outputting a decision result of the online electromyographic signal to call a corresponding control instruction; a controller unit, connected to the processor unit, executing a corresponding control action according to the control instruction.

[0061] Specifically, the motion recognition system further includes a data acquisition unit, which forwards the captured electromyographic signals in real time using TCPIP / LSL (connecting the upper and lower parts); aligns and packages the data from the sensor, and forwards the data from the lower computer where the data acquisition unit is located to the upper computer where the processor unit is located according to wireless network communication. The wireless network communication can be based on LSL forwarding of Lab Streaming Layer or data distribution of TCPIP protocol.

[0062] Example 3.

[0063] Based on the first embodiment, the third embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0064] Example 4.

[0065] Based on Example 1, this Example 4 provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method are implemented.

[0066] Example 5.

[0067] Based on Embodiment 1, this Embodiment 5 provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0068] Test example.

[0069] Test conditions.

[0070] Take the gesture recognition model set as an example, Figure 6 As shown in the figure, five gesture categories are selected to verify the accuracy of the recognition model, where 1, 2, 3, 4, and 5 represent five different gesture categories. Each acquisition channel is placed at the acquisition site of the left forearm and the right forearm (such as the ulnar flexor carpi, ulnar extensor carpi, and radial extensor carpi brevis of the forearm) to collect electromyographic signals. The electromyographic signals collected from the left forearm and the detection results of the starting point of the burst activity are shown in the figure. Figure 7 As shown in the figure, the electromyographic signals collected from the right forearm and the detection results of the starting point of the burst activity are as follows Figure 8 As shown. The hard threshold is selected as the amplitude average of the envelope. The moment when the envelope hard threshold after smoothing is greater than the hard threshold is recorded as the starting point of the EMG burst activity. The first-order difference energy operator is calculated to enhance the amplitude characteristics of the EMG burst activity. The left arm recognition model corresponding to experimental condition 1 (i.e., the first model), the right arm recognition model corresponding to experimental condition 2 (i.e., the second model), the multi-channel combination fusion model corresponding to experimental condition 3, and the multi-channel combination arrangement model corresponding to experimental condition 4 (sorted in the time sequence of left and right synchronization) are used to detect gesture recognition. The detection results are shown as follows: Fig. 9 , Fig.10 , Fig.11 , Fig.12 As shown.

[0071] Test results.

[0072] The accuracy of each recognition model, the number of classification categories and the number of control instructions mapped to them are shown in Table 1 below.

[0073] Table 1 The accuracy of each recognition model.

[0074] Test conditions Identification Model Accuracy Number of classification categories Number of control instructions Test condition 1 Left arm recognition model 93.31% 5 5 Test condition 2 Right arm recognition model 91.06% 5 5 Test condition 3 Multi-channel combination fusion model 95.34% 5 5 Test condition 4 Multi-channel combination arrangement model 84.97% 5 25 It can be seen that the recognition accuracy of the multi-channel combination fusion model and the single-channel combination recognition model is relatively high. By arranging and combining the left and right gesture recognition results (which can be time-synchronized or time-unsynchronized), that is, the multi-channel combination arrangement model, although a certain accuracy is sacrificed, the number of control instructions is greatly increased. For example, m left-hand actions are matched with n right-hand actions, which are mapped into at least m*n control instructions. Therefore, when performing action recognition, a set of highly correlated recognition models can be selected and loaded according to the relevance of the executed actions, and then a small number of channels are combined in the recognition model set to improve the accuracy of gesture recognition output instructions and the number of control instructions.

[0075] In the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0076] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0078] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for executing an output action using an electromyographic signal, characterized in that: include: Set channel combinations, that is, sort and group acquisition channels according to different control actions; Setting a model set, that is, using offline electromyographic signals to train models according to different channel combinations, obtaining multiple recognition models and classifying them into different model sets; Load the model, that is, select the corresponding model set to load according to the control action; Online detection, traverse all loaded recognition models to detect online electromyographic signals to obtain different detection results; Decision calling, smoothly output the detection results to call the corresponding control instructions; Control output, execute corresponding control actions based on control instructions.

2. The method according to claim 1, characterized in that The recognition model includes: a single-channel combination recognition model, a multi-channel combination fusion model, and a multi-channel combination arrangement model; wherein The single channel combination recognition model is configured as a recognition model obtained by performing model training on a single channel combination; The multi-channel combination fusion model is configured to fuse the signal features of multiple channel combinations and then perform model training to obtain a recognition model; The multi-channel combination arrangement model is configured to sort the signal features of multiple channel combinations in time sequence and then perform model training to obtain a recognition model.

3. The method according to claim 1, characterized in that The setting of the channel combination includes grouping and sorting the collection channels and collection parts according to the order corresponding to the execution of the control action.

4. The method according to claim 1, characterized in that: The model training includes: Detecting myoelectric burst activity, that is, detecting offline myoelectric signals to obtain starting point data segments of myoelectric signal burst activity and corresponding quantitative indicators or features; Data slicing, that is, slicing the starting data segment and marking the quantitative indicators or features one by one on each data segment to obtain model parameters; Cross validation is performed to complete model training.

5. The method according to claim 4, characterized in that The detecting of myoelectric burst activity comprises: Construct a feature matrix based on the feature parameters of the offline electromyographic signal; Extract the principal components in the projection space based on the feature matrix; Classification task of principal components based on Gaussian mixture model; The quantitative index or feature is obtained based on the classification result; wherein The characteristic parameters include: a combination of at least one of an autoregressive coefficient, a time domain feature, and a frequency domain feature.

6. The method according to claim 5, characterized in that The detecting of myoelectric burst activity also includes: enhancing the feature of myoelectric burst, that is, extracting the first-order differential energy operator of the myoelectric signals of all the acquisition channels in the channel combination and summing them before detecting the myoelectric burst activity.

7. The method according to claim 1, characterized in that The decision call includes: Set the interval for obtaining the detection results to the first time step, set the interval for smoothing the output to the second time step, and set the interval for calling the control instruction to the third time step, wherein the first time step ≤ the second time step ≤ the third time step; At the first time step, obtain the real-time detection result of the online signal each time; At the second time step, multiple real-time detection results are converted into a single smooth output decision result according to the voting principle or prior probability; At the third time step, multiple decision results are mapped into corresponding control instructions.

8. A motion output system based on electromyographic signals, characterized in that: include: A processor unit, executing the method according to any one of claims 1 to 7, to load the model to detect online electromyographic signals; A model database, connected to the processor unit, storing different model sets; A decision-maker unit, connected to the processor unit, is used to output the decision result of the online electromyographic signal to call the corresponding control instruction; The controller unit is connected to the processor unit and executes corresponding control actions according to the control instructions.

9. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images