Surface electromyography signal action recognition method under multi-dimensional features

Abstract

The present application relates to the technical field of action recognition, and more particularly to a surface electromyogram signal action recognition method under multi-dimensional features, comprising the following steps: surface electromyogram signal acquisition and preprocessing, dividing the collected signals into a training set and a test set; extracting appropriate spatial, time and frequency domain features for the training set samples, and fusing the time domain features with the spatial and frequency domain features. The application trains a recognition model through the training set features, inputs the test set features into the recognition model, and outputs classification results. The spatial domain features are used as basic features, the time and frequency domain features are extracted for the top-ranked channels after channel sorting, and the features are used as auxiliary features and fused with the basic spatial domain features, so that the multi-dimensional signal feature representation is completed to classify the mode, and the operation speed, accuracy and stability are improved.

Description

Technical Field

[0001] This invention relates to the field of motion recognition technology, and in particular to a method for motion recognition based on surface electromyography signals with multidimensional features. Background Technology

[0002] Surface electromyography (EMG) signals are generated by the superposition of electrical signals sent from the brain onto muscle fibers. Surface EMG signals are the combined effect of electrical signals transmitted from nerves to muscles, reflecting muscle activity and possessing strong time-sensitivity. There are two methods for measuring EMG signals: invasive and surface-based. EMG signals obtained through surface measurement are called surface EMG signals, which have advantages such as being non-invasive and simple to operate. Surface EMG signals can reflect information such as the shape and position of limbs during the completion of movements, making them an important way to perceive human movement.

[0003] In the processing of surface electromyography (SEMG) signals, it is necessary to classify and process the processed signal parameters. SEMG signals contain a lot of mixed information and noise. One way to accurately obtain the muscle control signals is to apply the "feature" approach. Feature extraction uses different methods to extract features from the preprocessed SEMG signals from different angles and dimensions. These concise features can represent a long segment of the signal. Although theoretically, more features lead to more accurate signal classification, an increase in features also increases computational difficulty and processing time. Therefore, how to select an appropriate number and types of features to efficiently represent the signal and improve the accuracy of judgment is an urgent problem to be solved. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems existing in the prior art.

[0005] Therefore, the present invention provides a method for identifying surface electromyography signals based on multidimensional features.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A method for motion recognition of surface electromyography signals based on multidimensional features includes the following steps:

[0008] Surface electromyography signal acquisition and preprocessing, the acquired signals are divided into training set and test set;

[0009] Extract appropriate spatial, temporal, and frequency domain features from the training set samples, and then fuse the temporal features with the spatial and frequency domain features.

[0010] The identification model is trained using features from the training set, and the identification model is then input into features from the test set to output the classification result.

[0011] By adopting the above technical solution, the selected appropriate spatial, temporal, and frequency domain features are integrated to construct a new feature matrix, which represents the signal features from multiple dimensions to complete the pattern classification. Compared with the traditional surface electromyography signal pattern classification method, the present invention has significantly improved the calculation speed, accuracy, and stability.

[0012] Furthermore, temporal features are extracted for each type of action sample in the training set. The temporal features of different types are combined in pairs to obtain multiple sets of temporal feature groups for two types of actions. The channels are sorted by temporal feature groups according to the principle of "high weight first, low weight last".

[0013] Furthermore, the feature is that the features with the highest weights in the top 60% are extracted to determine which channel the action sample comes from.

[0014] Furthermore, the reliefF algorithm is used for channel selection. First, temporal features are extracted from the training set samples according to the principle of "per channel" and "per sample". The extracted features have a variance σ. 2 Standard deviation λ, root mean square X rms .

[0015] Furthermore, by calculating the variance, standard deviation, and root mean square, and generating six feature matrices, all matrices are combined in a "previous class, next class" manner. The combined matrices are then reorganized according to the principle of "placing matrices adjacent to each other in the same channel".

[0016] Furthermore, after concatenating the three feature sets, the concatenated feature set is used in the reliefF algorithm. The reliefF algorithm formula is as follows:

[0017]

[0018] Where diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A, k represents the number of nearest neighbor samples, and M j (C) indicates The j-th nearest neighbor sample.

[0019] Furthermore, the difference between samples R1 and R2 on feature A is as follows:

[0020]

[0021] Furthermore, spatial features are extracted from all channels of the training and test sets respectively, using a "one-to-one" strategy for spatial feature extraction.

[0022] Furthermore, by calling the corresponding spatial feature set and the corresponding channel ranking, the time-domain features of the top 50% of channels and the frequency-domain features of the top 25% of channels can be extracted based on the corresponding channel ranking group. The time-domain features are the variance σ. 2 Standard deviation λ, root mean square X rms The frequency domain characteristics are selected by the average frequency F. MF Center of gravity frequency F FC Frequency standard deviation F RVF The calculation formula for each feature is as follows:

[0023]

[0024]

[0025]

[0026] Where u(n) is the new data obtained after the Fast Fourier Transform algorithm, which transforms the time-domain signal into a frequency-domain signal to extract the frequency-domain signal.

[0027] Furthermore, the time-domain, spatial-domain, and frequency-domain features are normalized and then combined to fuse the spatial and frequency-domain features.

[0028] The beneficial effects of this invention are that it obtains the channel ranking under the "one-to-one" strategy by sorting the channels. Through this sorting, time-domain and frequency-domain features are extracted from the top-ranked channels, while the bottom-ranked channels are discarded. Spatial-domain features are extracted from all channels. The three types of features are normalized and then fused together to construct a new feature matrix, which represents the signal features from multiple dimensions. The "support vector machine" is used to complete the pattern classification. Compared with the traditional surface electromyography signal pattern classification method, this invention has significantly improved the computation speed, accuracy, and stability. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Figure 1 This is a schematic diagram illustrating the implementation process of the identification method in this invention.

[0031] Figure 2 This is a schematic diagram of the data acquisition action in this invention.

[0032] Figure 3 This is a schematic diagram of temporal feature reconstruction in this invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0034] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0035] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0036] A method for action recognition of surface electromyography signals based on multidimensional features includes the following steps:

[0037] Step 1: Data acquisition and preprocessing of surface electromyography signals.

[0038] S1.1 Acquire surface electromyography signals during a specific action. This embodiment takes hand movements as an example, referring to... Figure 2 Surface electromyography (EMG) signals were acquired during 10 different actions, including action A (see reference). Figure 2 (Image a) Show a thumbs-up; Action B (refer to...) Figure 2 (See small image b) All five fingers are extended; Action C (refer to...) Figure 2 (small image c) clench your fist; Action D (refer to...) Figure 2 (See small image d) Extend the middle and ring fingers; Action E (refer to...) Figure 2 (See small image e) Extend the thumb, index finger, and little finger; Action F (refer to...) Figure 2 (In the small image f) Only the thumb is bent; Action G (refer to...) Figure 2(In the small image g) Only the index finger is bent; Action H (refer to...) Figure 2 (In the small image) Only the middle finger is bent; Action I (refer to...) Figure 2 (See small image i) Only the little finger is bent; Action J (see reference) Figure 2 (The small image in the image shows only the ring finger bent.)

[0039] S1.2 Signal data preprocessing: Remove noise from the acquired surface electromyography (EMG) signals and perform baseline correction. Divide the baseline-corrected EMG signals into time windows to obtain multiple sample sets. Divide these sample sets into training set and test set, with a ratio of 8:2.

[0040] Step 2, Spatial feature extraction

[0041] Spatial features are extracted from all channels of the training and test sets respectively. During the extraction of spatial features, the extracted action signals are compared using a "one-to-one" strategy.

[0042] First, spatial domain feature extraction is performed on the test set and training set signals using the "co-space mode". The co-space mode essentially involves the interaction between the signal matrix and a spatial filter to extract the feature vector f. The spatial filter is calculated using the following formula:

[0043]

[0044] Where β is the obtained spatial filter when W is maximum, X1 and X2 represent the two categories respectively, and T represents the matrix transpose; the obtained spatial filter is interacted with the signal matrix to obtain the filtered signal: Z = βX, and the eigenvector calculation formula is as follows.

[0045]

[0046] Where var represents the calculated variance, i is the number of signal channels, and m is a parameter, typically 2m equal to the number of channels. Since co-space mode extraction of spatial features involves the cooperation of two specific types of actions—a "one-to-one" strategy—it will produce... There are n spatial feature sets, where n represents n types of actions.

[0047] Step 3, Temporal Feature Extraction

[0048] Temporal features are extracted for each class of samples in the training set. Temporal features of different classes are combined in pairs to obtain multiple sets of temporal feature groups for two-class action combinations. The channels are sorted using the temporal feature groups. The multiple feature groups are placed in the reliefF algorithm to obtain the feature weights of multiple sets of two-class actions. According to the principle of "high weight first, low weight last", the features with the top 60% weights are extracted and their respective channels are determined, thus obtaining the channel sorting of multiple different sets of two-class action combinations.

[0049] S3.1, the channel selection uses the reliefF algorithm. First, temporal features are extracted from the training set samples according to the principle of "each channel" and "each sample". The extracted features have variance σ. 2 Standard deviation λ, root mean square X rms The calculation formula for each feature is as follows:

[0050]

[0051]

[0052]

[0053] Where N is the sample length or the number of sample points, x i For sample point values, s is the sample mean, and s is the sample standard deviation.

[0054] By calculating the three features mentioned above, six feature matrices are generated. All matrices are combined in a "previous category, next category" manner. Then, the combined matrices are reorganized according to the principle of "placing matrices adjacent to each other on the same channel". The reorganization method is as follows. Figure 3 As shown.

[0055] S3.2 Since each feature generates a feature set, the three feature sets are concatenated and then used in the reliefF algorithm to calculate the weights of the correlation between features among their nearest neighbors of the same class and among their nearest neighbors of different classes. The reliefF algorithm formula is as follows:

[0056]

[0057] Where P(C) is the proportion of that class, diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A, P(Class(R)) is the proportion of a randomly selected sample class, m is the number of samplings, k represents the number of nearest neighbor samples, and M j (C) indicates The j-th nearest neighbor sample. As shown in the following formula:

[0058]

[0059] Obtain the feature weight array W for all features. Sort the features in W according to their weights, from largest to smallest. Use the top 60% of features as a selected interval, and determine which channels the features within that interval originate from through a voting process, thus completing the channel sorting. Note that channel sorting is specific to two classes of actions; if there are n classes of actions, it will generate... Each channel is sorted into groups.

[0060] Step 4, frequency domain feature extraction.

[0061] Because both spatial feature extraction and channel ranking are "one-to-one" strategies, when combining features for specific two types of actions, it is only necessary to call the corresponding spatial feature set and the corresponding channel ranking. Based on the channel ranking obtained in step 3, time-domain features are extracted from the top 50% of the channels and frequency-domain features are extracted from the top 25% of the channels. No features are extracted from the lower-ranked matrices. The time-domain features are selected in the same way as those in step 3.1, and the frequency-domain features are selected from the average frequency F. MF Center of gravity frequency F GC Frequency standard deviation F RVF The calculation formula for each feature is as follows:

[0062]

[0063]

[0064]

[0065] Where u(n) is the new data obtained after the Fast Fourier Transform algorithm, and u(n) is the mean of the new data obtained after the Fast Fourier Transform algorithm. The above process transforms the data from a time domain signal to a frequency domain signal, which facilitates the extraction of frequency domain features.

[0066] Step 5, Feature Fusion.

[0067] The time-domain, frequency-domain, and spatial-domain features of the obtained test set and training set are normalized separately, and then fused. The fused features are used as sample features. Each feature matrix is ​​normalized because different features have different orders of magnitude. To fuse them together to represent a single signal, all features must be normalized to ensure they are on the same order of magnitude. The normalization formula is as follows:

[0068]

[0069] Where lower represents the set lower limit value, upper represents the set lower limit value, max represents the maximum value among signal elements, min represents the minimum value among signal elements, and y is the selected data value. The normalized time-domain features, frequency-domain features, and spatial-domain features are sorted according to their normalized values, and the sorted time-domain features, frequency-domain features, and spatial-domain features are then merged together to complete feature fusion.

[0070] Step 6: Model Training and Result Output

[0071] The feature samples from the training set are input into a support vector machine to train the recognition model; the feature samples from the test set are used as input to achieve hand action classification, recognition and prediction using the obtained support vector machine model.

[0072] In summary, the present invention proposes a method for optimal classification and recognition of surface electromyography signals based on time-domain, frequency-domain, and spatial-domain features. The method uses spatial features as basic features, extracts time-domain and frequency-domain features from the top-ranked channels after channel sorting, and uses these features as auxiliary features to fuse with the basic spatial features.

[0073] This paper obtains the channel ranking under the "one-to-one" strategy by channel sorting. The time-domain and frequency-domain features are extracted from the top-ranked channels, while the bottom-ranked channels are discarded. Spatial domain features are extracted from all channels. The three types of features are normalized and then fused together to construct a new feature matrix, which represents the signal features from multiple dimensions. The "support vector machine" is used to complete the pattern classification. Compared with the traditional surface electromyography signal pattern classification method, this invention has significantly improved the computation speed, accuracy and stability.

[0074] Experimental results showed that the accuracy rate of motion intention pattern classification for 1860 samples with 12-dimensional features and 10 categories was 91.27%.

[0075] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. A method for surface electromyography (sEMG) action recognition under multi-dimensional features, characterized in that, Includes the following steps: Surface electromyography signal acquisition and preprocessing, the acquired signals are divided into training set and test set; Extract appropriate spatial, temporal, and frequency domain features from the training set samples, and then fuse the temporal features with the spatial and frequency domain features. Spatial features are extracted from all channels of the training and test sets respectively. During the extraction of spatial features, the extracted action signals are compared using a "one-to-one" strategy. Temporal features are extracted for each class of samples in the training set. Temporal features of different classes are combined in pairs. When combining features of two specific action classes, the corresponding spatial feature set and channel ranking are called. Multiple feature sets are placed in the reliefF algorithm to obtain feature weights for multiple two-class actions. The top 60% of the features by weight are extracted and ranked by channel. Temporal features are extracted for the top 50% of the channels and frequency features are extracted for the top 25% of the channels. The identification model is trained using features from the training set, and the identification model is then input into features from the test set to output the classification result.

2. The sEMG action recognition method under multi-dimensional features according to claim 1, characterized in that, Temporal features are extracted for each type of action sample in the training set. Temporal features of different types are combined in pairs to obtain multiple sets of temporal feature groups for two types of actions. These multiple feature groups are placed in the reliefF algorithm to obtain the feature weights for multiple sets of two types of actions. The temporal feature groups are used according to the principle of "high weight first, low weight last". 3.The sEMG action recognition method under multi-dimensional features according to claim 1, characterized in that, Extract the features with the highest weights in the top 60% to determine which channel the action sample comes from, and infer the channel ranking based on the proportion of features in the top 60%.

4. The method for recognizing surface electromyography signals under multi-dimensional features according to claim 1, characterized in that, The channel selection uses the reliefF algorithm. First, temporal features are extracted from the training set samples according to the principle of "per channel" and "per sample." The extracted features have variance. Standard deviation Root mean square .

5. The method for recognizing surface electromyography signals under multi-dimensional features according to claim 4, characterized in that, By calculating the variance, standard deviation, and root mean square, and generating six feature matrices, all matrices are combined in a "previous class, next class" manner. The combined matrices are then reorganized according to the principle of "placing matrices adjacent to each other in the same channel".

6. The method for recognizing surface electromyography signals under multi-dimensional features according to claim 1, characterized in that, After concatenating the three feature sets, the concatenated feature set is used in the reliefF algorithm. The reliefF algorithm formula is as follows: in Indicates sample , The difference on feature A, where k represents the number of nearest neighbor samples. express The j-th nearest neighbor sample.

7. The method for recognizing surface electromyography signals under multi-dimensional features according to claim 1, characterized in that, sample , The difference on feature A is as follows: 。 8. The method for recognizing surface electromyography signals under multi-dimensional features according to claim 1, characterized in that, Simply call the corresponding spatial feature set and the corresponding channel sorting. Based on the corresponding channel sorting group, extract the time-domain features of the top 50% of channels and the frequency-domain features of the top 25%. The time-domain features are the variance... Standard deviation Root mean square The frequency domain characteristics are selected by the average frequency. Center of gravity frequency Frequency standard deviation The calculation formula for each feature is as follows: in, It is new data obtained after passing through the Fast Fourier Transform algorithm, which transforms the time-domain signal into a frequency-domain signal to extract the frequency-domain signal.

9. The method for recognizing surface electromyography signals under multi-dimensional features according to claim 1, characterized in that, The time-domain, spatial-domain, and frequency-domain features are normalized and then combined to fuse the spatial and frequency-domain features.

Images