CN112043269B

CN112043269B - Muscle space activation mode extraction and recognition method in gesture motion process

Info

Publication number: CN112043269B
Application number: CN202011030863.5A
Authority: CN
Inventors: 陈香; 文庆庆; 张旭; 胡若晨
Original assignee: University of Science and Technology of China USTC
Current assignee: University of Science and Technology of China USTC
Priority date: 2020-09-27
Filing date: 2020-09-27
Publication date: 2021-10-19
Anticipated expiration: 2040-09-27
Also published as: CN112043269A

Abstract

The invention provides muscle space activation in the gesture action processThe pattern extraction and identification method comprises the following steps: step 1, collecting surface electromyographic signals generated when a user executes gesture actions by using an electrode array; step 2, preprocessing the collected surface electromyographic signal data set; step 3, extracting a muscle space activation mode matrix W and a recruitment mode matrix H by using a matrix decomposition algorithm, and determining the number of space activation modes according to a VAF threshold value; and 4, adjusting vectors in the W matrix by using the two-norm energy product of the W and the H, and rearranging to obtain an updated space activation mode matrix W_newThe alignment of the spatial activation modes of different motion samples is realized without supervision; step 5, adjusting the matrix W_newAs a gesture action sample, training and testing the designed network model; step 6, utilizing a space activation mode matrix W_newAnd the deep learning network completes the recognition of the gesture action.

Description

Muscle space activation mode extraction and recognition method in gesture motion process

Technical Field

The invention relates to the field of biological signal processing, in particular to a muscle space activation pattern extraction and identification method based on high-density surface myoelectricity.

Background

The surface electromyographic signals are superposition of bioelectricity signals generated by excitation of a plurality of movement units in muscles on time and space, and carry human emotion or intention information. The gesture recognition technology based on the surface myoelectricity can be used for realizing myoelectricity control, and simultaneously provides a novel human-computer interaction mode for the fields of virtual reality, consumer electronics and the like.

Because the high-density array electrode has the characteristic of capturing abundant muscle activation space-time distribution information, the gesture recognition problem based on the high-density surface myoelectricity is widely concerned. With the successful application and rapid development of the deep learning technology in the fields of video, image, voice recognition and the like, researchers try to introduce the deep learning technology into the field of gesture recognition based on high-density surface myoelectricity, and the deep learning technology has the performance which is remarkably superior to that of the traditional machine learning method in the aspect of recognition rate. However, noise caused by electrode offset, gesture movement and the like inevitably affects the quality of acquired signals, and when the deep learning technology is used for gesture recognition based on high-density array surface myoelectricity, sufficient samples are required to train a relatively robust recognition network, and heavy training burden is brought to a user by data acquisition of large samples. In addition, most research works input the raw electromyographic signals into a network as samples for pattern recognition, and the recognition of gesture actions often depends on the spatial distribution nonuniformity of the high-density array electromyographic signals. However, the presence of redundant information between channels in a high density array can obscure the spatial distribution non-uniformity, thereby adversely affecting the recognition performance. If effective gesture mode characteristics can be extracted from the high-density array surface myoelectricity, the influence caused by factors such as electrode offset noise, channel redundant information and the like is reduced while the dimension of network input data is reduced, and the method has important significance for promoting the practical application of the gesture recognition technology control based on the surface myoelectricity.

According to related researches, the gesture action high-density array surface electromyographic signal matrix X is decomposed by matrix decomposition algorithms such as principal component analysis, independent component analysis and non-negative matrix decomposition, and can be divided into a plurality of combinations of space activation distribution. In particular, a spatial activation pattern matrix W and a corresponding recruitment pattern matrix H containing a plurality of muscle relative activation state vectors may be extracted. Each vector in the W matrix can be considered as a spatial activation pattern of the target muscle group, and the corresponding vector of the H matrix is the recruitment pattern of the activation distribution. When the difference of the spatial activation modes exists in different gesture actions, the spatial activation distribution is utilized to possibly distinguish various gesture actions. Particularly, when the surface electromyographic signals generated by different execution actions of the same gesture are decomposed, the extracted space activation mode difference is small, and the influence of factors such as electrode offset is mainly reflected on the change of the recruitment mode. Therefore, the gesture action mode recognition is carried out by utilizing the space activation mode, and the robust recognition performance is expected to be obtained. Matrix decomposition algorithms such as principal component analysis, independent component analysis, non-negative matrix decomposition and the like are already the existing technologies, and most of the existing documents correspondingly apply the algorithms according to different research contents.

When the matrix decomposition algorithm is used for gesture classification, besides principal component analysis, a plurality of space activation distribution vectors extracted by other matrix decomposition algorithms are randomly arranged; at this time, if the gesture classification is performed by using the spatial activation mode matrix extracted by the matrix decomposition algorithm, the spatial activation modes of different samples of the same type of gesture are blurred due to the random distribution, so that the classification accuracy is reduced.

Disclosure of Invention

In order to solve the technical problem, the invention provides a muscle space activation pattern extraction and recognition method in the gesture action process, which can realize the alignment of a space activation pattern matrix without supervision. Specifically, firstly, the gesture action high-density array surface electromyographic signals are decomposed by using a matrix decomposition algorithm to obtain a spatial activation mode matrix W and a recruitment mode matrix H. Then, two-norm energy products of corresponding components in the spatial activation mode matrix W and the recruitment mode matrix H are calculated, and all vectors in the spatial activation mode matrix W are reordered from large to small according to the energy products. And finally, inputting the reordered space activation pattern matrix serving as a gesture motion sample into a deep learning network for gesture recognition, so that the myoelectric pattern recognition effect can be improved.

The technical scheme of the invention is as follows: a muscle space activation mode extraction and recognition method in the gesture motion process comprises the following steps:

step 1, collecting surface electromyographic signals generated when a user executes gesture actions by using an electrode array;

step 2, preprocessing the collected surface electromyographic signal data set;

step 3, extracting a muscle space activation mode matrix W and a recruitment mode matrix H by using a matrix decomposition algorithm, and determining the number of space activation modes according to a VAF threshold value;

and 4, adjusting vectors in the W matrix by using the two-norm energy product of the W and the H, and rearranging to obtain an updated space activation mode matrix W_newThe alignment of the spatial activation modes of different motion samples is realized without supervision;

step 5, adjusting the matrix W_newAs a gesture action sample, the designed network model is processedTraining and testing;

step 6, utilizing a space activation mode matrix W_newAnd the deep learning network completes the recognition of the gesture action.

Further, gesture motions include combinations of multiple states that can involve the elbow, wrist, and joints of the fingers.

Furthermore, the number of channels of the electrode array is m, the number of channels of the row is n, and the distance between adjacent channels in the array is d.

Further, the electrode array collects surface electromyographic signals generated when the target muscle group executes gesture actions, and the method comprises the following steps: extensor muscles of forearm covering extensor total muscles of finger, extensor carpi radialis and extensor carpi ulnaris; flexor forearm, covering flexor muscles, flexor carpi radialis, flexor carpi ulnaris; the biceps brachii and triceps brachii.

Further, the myoelectric signals of the gesture action array in the step 2 are preprocessed as follows:

(2.1) dividing the active segment, namely dividing electromyographic data by adopting a moving average method based on a signal energy threshold, and judging the start and stop points of the active segment according to whether the energy average value of all channel signals is greater than a preset energy threshold;

(2.2) filtering and normalizing, namely firstly, carrying out high-pass filtering on the electromyographic signals of all channels to eliminate low-frequency noise possibly occurring in the data acquisition process; then, performing mean value removal and full-wave rectification on the filtered signal; finally, normalizing the rectified signals, wherein the normalization method is to divide each channel signal by the two-norm value of the channel signal;

and (2.3) sample expansion, wherein in view of the fact that the contraction state of each joint and muscle in the gesture action maintaining stage is kept unchanged, the sample expansion is carried out on the data in the stage by adopting a sliding window method, the window length is set to be L, the sliding step length is set to be S, and M sliding sample windows are obtained in each activity stage.

Further, in the step 4, for each signal array in the sliding analysis window, a matrix decomposition algorithm is used to extract a spatial activation mode matrix W and a recruitment mode matrix H, and reordering is performed on the matrices based on two-norm energy, which specifically includes:

decomposing a plurality of space activation mode vectors W by using a matrix decomposition algorithm_iWhere i 1.. N, and the corresponding recruitment mode vector H_iN, calculating the product of the two-norm energy of each space activation mode vector and the two-norm energy of the recruitment mode vector, rearranging the space activation mode vectors from large to small to obtain an updated space activation mode matrix W_new。

Further, in the step 5, the extracted spatial activation pattern matrix W is utilized_newAnd the deep learning network carries out gesture action pattern recognition.

Further, the space activation mode samples of different gestures are divided into a training set, a verification set and a test set, and the deep learning network model is trained by using training set data; adjusting network hyper-parameters according to the effect of the network on the verification set, and confirming the identification effect of the model on the test set, which is as follows:

(5.1) firstly, dividing a training set, a verification set and a test set, and generating corresponding data set labels;

(5.2) determining the number of network layers and the number of nodes of each hidden layer according to the convergence error and the recognition rate of the training set; illustratively, when the training error converges to a set lower limit value and the recognition rate also converges to approach 1, the number of network layers and the number of hidden layer units are optimal choices;

(5.3) adjusting relevant hyper-parameters such as regularization, neuron node random inactivation probability, learning rate and the like in the network according to the identification rate of the verification set, optimizing the network performance, and enabling the accuracy rate of the verification set to rise and converge to a set standard;

(5.4) validating the effect of the trained network model using the test set.

Has the advantages that:

the matrix reordering method based on the two-norm energy product can solve the problem of difference among a plurality of samples of the same action caused by the randomness of result arrangement when a space activation mode is extracted by using a matrix decomposition algorithm, namely, the matrix alignment of the space activation mode of the same gesture action and different samples is realized unsupervised.

An electromyographic pattern recognition experiment based on a space activation pattern and a deep learning network is carried out on the finger gesture actions, and feasibility of reducing training load of a user and reducing influence of electrode offset and channel redundant information by the method is verified.

The method has the advantages that the requirement on the diversity of training samples is reduced, a large-scale data set does not need to be collected to cover enough muscle change information, and a new thought is provided for solving the problems of electrode offset, channel redundancy and the like in gesture recognition based on high-density surface myoelectricity.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.

Fig. 1 is a flowchart of a muscle space activation pattern extraction and recognition method in a gesture action process according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of gesture actions according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a spatial activation mode of a representative gesture 5 according to the present embodiment;

FIG. 4 is a schematic diagram of a spatial activation pattern of the representative gesture 14 in the embodiment;

FIG. 5 is a schematic diagram of a gesture recognition network structure adopted by the high-density array electromyographic signals according to the embodiment of the invention;

FIG. 6 is a schematic diagram of a gesture recognition network structure used by the spatial activation distribution matrix according to the embodiment of the present invention;

FIG. 7 shows the average gesture recognition accuracy and the standard deviation probability obtained by the embodiment of the invention.

Detailed Description

The technical solutions in the embodiments of the present invention are clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.

The embodiment of the invention provides a muscle space activation mode extraction and recognition method in a gesture action process, which comprises the following main steps as shown in figure 1:

(1) collecting surface electromyographic signals X epsilon R generated when a user executes gesture actions by using a high-density electrode array^C×T(wherein C is the number of high-density array electrode channels, and T is the number of signal sampling points), and preprocessing is carried out;

(2) extracting a space activation mode matrix W epsilon R by adopting a matrix decomposition algorithm^N×C(where N is the number of active patterns) and a recruitment pattern matrix H ∈ R^T×NRearranging each vector line in the space activation mode matrix W based on the two-norm energy to obtain an adjusted matrix W_new∈R^N×CThe alignment of the spatial activation modes of different motion samples is realized without supervision;

(3) the adjusted matrix W_newAnd as a gesture action sample, training and testing the designed network model to realize gesture action recognition.

For convenience of understanding, a specific implementation process of the muscle space activation pattern extraction and gesture recognition method based on high-density surface myoelectricity is described in detail below by taking a non-Negative Matrix Factorization (NMF) algorithm and a Long Short Term Memory (LSTM) neural network as examples.

Firstly, gesture action high-density array surface electromyogram data acquisition

In the embodiment of the invention, as shown in fig. 2, 30 gesture motions involving a combination of a plurality of states of the elbow, the wrist and the joints of the fingers are taken as recognition objects, and 10 healthy people of different sexes and different ages are recruited to participate in the data acquisition experiment. All subjects were asked to perform a gestural action with the right hand. Before starting the experiment, all participants were asked to practice the gesture action execution pattern until they were able to complete the task according to the experiment requirements.

The number of the row channels of the electrode array is m, the number of the column channels of the electrode array is n, the distance between adjacent channels in the array is d (the distance between the adjacent row channels and the adjacent column channels is equal), and two flexible high-density electrode arrays with different specifications are adopted in the embodiment of the invention to acquire the surface myoelectric signals of the gesture action. Illustratively, the parameters of the first electrode array are m-8, n-6, and d-14 mm, and the parameters of the second electrode array are m-4, n-4, and d-18 mm.

Before data acquisition, a subject sits on a chair in a comfortable state, the forearm is rested on a table top, the target muscle area is wiped with alcohol, conductive paste is coated on an electrode array to reduce the impedance of the skin to the electrode, and an experimenter fixes the electrode array on the surface of a main muscle group related to gesture actions by using a silica gel sheet with good adhesive force and elasticity. Illustratively, two electrode arrays of a first gauge are attached to the extensor forearm (extensor digitorum communis, extensor radialis, extensor ulnaris, etc.) and flexor forearm (flexor digitorum, flexor radialis, flexor ulnaris, etc.) and two electrode arrays of a second gauge are attached to the biceps brachii and triceps brachii, respectively.

During the data acquisition process, the subject performs the gesture with a comfortable force. The execution of the gesture motion is divided into three phases: the first phase is the initial execution phase, which includes the process from muscle contraction to elbow, finger and wrist movement, lasting about L1 seconds; the second phase is a gesture maintenance phase, in which all joints and muscle contraction states are kept unchanged for about L2 seconds; the third phase is the finishing phase, during which the joints return to a free state and the muscles relax, lasting approximately L3 seconds. Each gesture motion is repeated Nc times. To avoid muscle fatigue, a certain time L4 is left between each gesture action. The four high-density electrode arrays acquire 128-channel electromyogram signals in total, and exemplarily, the sampling rate of each channel is 1000Hz, and the four high-density electrode arrays are set as follows: l1-1 s, L2-2 s, L3-1 s, L4-7 s, Nc-8.

Secondly, surface electromyographic signal preprocessing

In the embodiment of the invention, the surface electromyographic signals generated by the gesture actions are preprocessed in the following way:

(1) and (5) dividing the active segment. And (3) adopting a moving average method based on a signal energy threshold value to segment the electromyographic data, and judging the start and stop points of the active segment according to whether the energy average value of all channel signals is greater than a preset energy threshold value.

(2) And (5) filtering and normalizing. Firstly, the electromyographic signals of all channels are subjected to high-pass filtering so as to eliminate low-frequency noise which possibly occurs in the data acquisition process. Illustratively, a FIR high-pass filter with a cutoff frequency of 20Hz of order 50 is used; then, performing mean value removal and full-wave rectification on the filtered signal; finally, the rectified signal is normalized, and in an exemplary method, each channel signal is divided by the two-norm value of the channel signal.

(3) And (5) sample expansion. In view of the fact that the state of contraction of each joint and muscle remains unchanged in the gesture motion maintenance phase, the data of this phase (about L2 seconds) is subjected to sample expansion by a sliding window method. Setting the window length as L and the sliding step length as S, and obtaining M sliding sample windows in each activity section. Illustratively, L2 is 2S, L is 100ms, S is 50ms, and M is 39.

Thirdly, extracting the space activation mode based on the nonnegative matrix factorization algorithm

(1) non-Negative Matrix Factorization (NMF) algorithm

The NMF algorithm is decomposed under the constraint condition that all elements in the matrix are nonnegative numbers, and the core idea is to use a target matrix

Decomposition into a basis matrix

And a coefficient matrix

When applied to high-density array electromyography, W is called a spatial activation pattern matrix, and H is called a recruitment pattern matrix. Where '+' indicates no negative values in all three matrices, C indicates the number of channels, N indicates the number of spatial activation patterns extracted, and T is the sliding window length, which indicates the number of channels in the three matricesIllustratively, T is L.

The process of implementing the NMF algorithm can be considered as an optimization solution process. A decomposition error E is introduced, as in equation (1), such that the signal reconstructed from W and H is continuously approximated to X, with the optimization objective being to minimize the squared euclidean distance of the decomposition error.

X＝WH+E (1)

An objective function based on the sum of squared euclidean distances is shown in equation (2), where i is 1,2, … …, C; j ═ 1,2, … …, T; k is 1,2, … …, N:

minE(W,H)²＝∑_i,j||X_i,j-W_i,kH_k,j||² (2)

solving the optimization problem by using a gradient descent method, wherein the iteration rules of W and H are shown as formulas (3) and (4):

the alternating optimization of W and H according to the iteration rule can achieve a local optimal solution on the basis of keeping nonnegativity.

A spatial activation mode matrix W and a recruitment mode matrix H are extracted from a high-density electrode array signal X based on an NMF algorithm, and the method mainly comprises the following steps:

1) initializing the W and H matrixes randomly and ensuring that the parameters of the W and H are not negative;

2) updating and iterating W and H by using the iteration formulas of the formulas (3) and (4) until one of the termination conditions is met: iteration is carried out for enough times; secondly, setting a threshold value of reconstruction errors, stopping iteration when the threshold value is reached, and setting 10 in the embodiment^-5。

(2) Spatial activation pattern number determination

The more the number of the selected and extracted space activation modes in the NMF algorithm is, the higher the accuracy of the reconstructed electromyographic signals is; but it follows that interference signals are introduced during the extraction process, thereby affecting the validity of the data. In this embodiment, a method for determining the number of spatial activation modes based on a VAF threshold is adopted. Vaf (the variance computed) is a measure of the reconstruction error, which represents the variation of the reconstruction data. The specific calculation formula is shown as (5):

the VAF range is between 0 and 1, the larger the value is, the smaller the reconstruction error is, namely, the higher the reconstruction precision is. The appropriate number of synergies was selected by setting a threshold for VAF, which was set to 0.95 in this study.

(3) Matrix reordering based on two-norm energy

Spatial activation pattern matrix W^N×CEach row represents a spatial activation pattern, activation matrix H^T×NThe corresponding column of (a) represents the recruitment mode. When a plurality of spatial activation modes are extracted, each mode Wi (i epsilon [1, N) in a matrix W obtained by NMF algorithm]) Is arranged randomly. When a spatial activation mode of a certain action is specifically researched, a rule needs to be set for matrixes extracted from a plurality of samples of the same action, and the spatial activation mode is aligned.

In the embodiment of the invention, the vectors in W and H are rearranged according to the product of the two-norm energy of each row vector Wi in the space activation mode matrix W and the two-norm energy of the corresponding column vector Hi in the recruitment mode matrix H.

The method comprises the following specific steps:

firstly, calculating the two-norm energy product of each vector in the W and H matrixes according to a formula (6) to obtain an energy matrix e shown in a formula (7)_2×N。

And the second behavior in the matrix e is the original index of each cooperative vector in the cooperative structure matrix W.

Rearranging indexes of a second row in the matrix e from large to small according to values of the first row, and rearranging vectors of rows in the matrix W according to the adjusted values of the second row.

Fig. 3 and 4 are schematic diagrams of the spatial activation pattern of 2 representative gestures obtained by the above steps. 6 space activation pattern vectors are extracted from each gesture action high-density array surface electromyography signal. Illustratively, the vectors of each row 1 x 128 in the spatial activation pattern matrix are arranged in rows and columns (here, the vectors of 1 x 128 are not arranged in rows or columns alone to 8 x 16, but channels belonging to the same electrode array are arranged in 16/row and then rows are juxtaposed together according to the sensor array used at the time of acquisition) to form an 8 x 16 matrix, with two electrodes belonging to format one being arranged at the first 6 x 16 position and two electrodes belonging to format two being arranged at the last 2 x 16 position. In the anterior 6 x 16 position, the anterior 3 lines are electrode array channels on the extensor group of the right arm, and the posterior 3 lines are electrode array channels on the flexor group of the right arm; in the posterior 2 x 16 position, anterior 1 is the electrode array channel on the right brachial biceps, and posterior 1 is the electrode array channel on the right brachial triceps. In each graph, the left 6 sub-graphs are respectively a space activation mode vector graph of the gesture action, and the right sub-graph is a binary image obtained from the corresponding activation mode vector through an Otsu method. As can be seen from fig. 3 and 4, after the matrix reordering, the difference of the main activated muscle regions of the spatial activated pattern vectors at the corresponding positions in the spatial activated pattern matrices of different actions is increased, so that it is possible to distinguish multiple gesture actions by using the spatial activated distribution

Fourth, construction and training of electromyographic pattern recognition network

Deep learning is a multi-layer representation learning method, and a deep learning network can extract multi-dimensional features of input data layer by layer, so that representation learning of complex functional relationships is completed. Among a plurality of deep learning network structures, a time sequence neural network represented by a Long Short Term Memory (LSTM) network has strong time sequence feature extraction capability on time sequence signals, and has a certain optimization effect on the problems of gradient explosion or gradient disappearance and the like. The surface electromyogram signal is a time series signal, which contains abundant time sequence information. Therefore, the LSTM is adopted as the electromyographic pattern recognition network in the embodiment of the present invention.

As shown in fig. 5 and fig. 6, the network structures of the embodiment when gesture classification is performed by using the high-density array electromyogram signal and the spatial activation pattern matrix are respectively shown. The long-short term memory neural network comprises LSTM hidden layers and batch normalization layers BN which are alternately arranged, and finally a softmax classification layer is added. The LSTM hidden layer is used for extracting the time sequence characteristics of sample values on each input sampling point, and the BN layer forcibly enables the distribution of internal input values of the network nodes to be in normal distribution with the mean value of 0 and the variance of 1 in a normalized mode, so that the offset influence of training on parameters is eliminated. In addition, to prevent overfitting of the network, a dropout layer is added after each hidden layer of the LSTM, and regularization coefficients of L1 and L2 are added to the final loss function. Each hidden layer adopts a tanh activation function, and the optimizer of the network is RMSprop. When the electromyographic signals corresponding to the gesture actions are subjected to pattern recognition, as shown in fig. 5, the number of cell units of the network is the step length L of the sliding window, that is, the LSTM layer comprehensively extracts time information contained in L time sampling points. Similarly, as shown in fig. 6, when the LSTM is used to identify the spatial activation pattern matrix W of the gesture motion, the number of cell units of the network is the number of spatial activation pattern vectors, that is, the LSTM network comprehensively considers a plurality of pieces of spatial activation pattern information.

Based on a cross-validation method, illustratively, the electromyography data obtained by repeatedly executing 8 times of Nc of each gesture action of the subject is taken as a test set, and the data acquired in the last 1 time is taken as the test set; of the remaining 7 data, data from 1 to 6 executions were sequentially selected as training set, and others were selected as validation set. The specific training process is as follows:

(1) firstly, a training set, a verification set and a test set are divided, and corresponding data set labels are generated.

(2) And determining the number of network layers and the number of nodes of each hidden layer according to the convergence error and the recognition rate of the training set. Illustratively, when the training error converges to a set lower limit value and the recognition rate also converges to approach 1, the number of network layers and the number of hidden layer units are the optimal choices.

(3) And adjusting relevant hyper-parameters such as regularization, neuron node random inactivation probability, learning rate and the like in the network according to the identification rate of the verification set, optimizing the network performance, and enabling the accuracy rate of the verification set to rise and converge to a set standard.

(4) The effect of the trained network model is confirmed using the test set.

In order to verify the effectiveness and superiority of the proposed matrix decomposition algorithm combined with the electromyographic pattern recognition method of the deep learning network, gesture recognition experiments are carried out by directly utilizing the original electromyographic signals and the process.

Fig. 7 shows the accuracy statistics obtained by performing gesture recognition with the original electromyographic signal and the spatial activation pattern matrix extracted by non-negative matrix factorization, respectively. It can be found that: under the condition that the training set is of a certain size, compared with gesture recognition directly using high-density surface muscle electrical signals, the gesture recognition method based on the space activation mode has higher robustness; and under the condition of less training set samples, the method has higher recognition accuracy. Therefore, the method can provide a beneficial solution for the problems of heavy training burden of a user, poor algorithm robustness and the like in the gesture recognition technology based on the high-density array surface myoelectricity.

Through the above description of the embodiments, it is clear to those skilled in the art that the above embodiments can be implemented by software, and can also be implemented by software plus a necessary general hardware platform. With this understanding, the technical solutions of the embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a usb disk, a removable hard disk, etc.), and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods according to the embodiments of the present invention.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A muscle space activation mode extraction and recognition method in a gesture motion process is characterized by comprising the following steps:

step 2, preprocessing the collected surface electromyographic signal data set;

in the step 3, VAF is used to measure the amount of reconstruction error, which represents the variation of reconstruction data, and the specific calculation formula is shown in (1):

in the step 4, for each signal array in the sliding analysis window, a matrix decomposition algorithm is used to extract a spatial activation mode matrix W and a recruitment mode matrix H, and reordering of the matrices based on two-norm energy specifically includes:

decomposing a plurality of space activation mode vectors W by using a matrix decomposition algorithm_iWhere i 1.. N, and the corresponding recruitment mode vector H_iN, each of which is calculated as 1Rearranging the space activation mode vectors from large to small to obtain an updated space activation mode matrix W by multiplying the two-norm energy of the space activation mode vectors and the two-norm energy of the recruitment mode vectors_new；

Step 5, adjusting the matrix W_newAs a gesture action sample, training and testing the designed network model;

2. The method of claim 1, wherein the gesture comprises a combination of states that can involve the elbow, wrist, and joints of the fingers.

3. The method for extracting and recognizing muscle space activation pattern in gesture action process as claimed in claim 1, wherein the number of row channels of said electrode array is m, the number of column channels is n, and the distance between adjacent channels in array is d.

4. The method for extracting and identifying the muscle space activation pattern in the gesture action process according to claim 2, wherein the electrode array collects the surface electromyographic signals generated when the target muscle group performs the gesture action, and the method comprises the following steps: extensor muscles of forearm covering extensor total muscles of finger, extensor carpi radialis and extensor carpi ulnaris; flexor forearm, covering flexor muscles, flexor carpi radialis, flexor carpi ulnaris; the biceps brachii and triceps brachii.

5. The method for extracting and recognizing the muscle space activation pattern in the gesture action process according to claim 1, wherein the myoelectric signals of the gesture action array are preprocessed in the step 2 as follows:

6. The method for extracting and recognizing muscle space activation pattern in gesture action process as claimed in claim 1, wherein in step 5, the extracted space activation pattern matrix W is used_newAnd the deep learning network carries out gesture action pattern recognition.

7. The muscle space activation pattern extraction and recognition method in the gesture action process according to any one of claims 1 to 6, characterized by dividing space activation pattern samples of different gestures into a training set, a verification set and a test set, and training a deep learning network model by using training set data; adjusting network hyper-parameters according to the effect of the network on the verification set, and confirming the identification effect of the model on the test set, which is as follows:

(5.2) determining the number of network layers and the number of nodes of each hidden layer according to the convergence error and the recognition rate of the training set; when the training error converges to a set lower limit value and the recognition rate also converges to approach 1, the number of network layers and the number of hidden layer units are optimal choices;

(5.3) adjusting relevant hyper-parameters of regularization, neuron node random inactivation probability and learning rate in the network according to the identification rate of the verification set, optimizing the network performance, and enabling the accuracy rate of the verification set to rise and converge to a set standard;

(5.4) validating the effect of the trained network model using the test set.