A lower limb action accurate recognition method based on surface electromyogram signal decomposition and optimization processing

Abstract

The application discloses a lower limb action accurate recognition method based on surface electromyogram signal decomposition and optimization processing, and comprises the following steps: acquiring a surface electromyogram signal, decomposing the surface electromyogram signal to obtain a sub-signal; optimizing the sub-signal through a two-stage modal selection method, extracting and fusing features of the optimized sub-signal to obtain a feature vector, and reducing the dimension of the feature vector; wherein the two-stage modal selection method comprises an energy contribution selection method and a modal correlation selection method; and recognizing the reduced dimension feature vector through a classification algorithm to obtain a lower limb action recognition result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a lower limb action accurate recognition method based on surface electromyogram signal decomposition and optimization processing. BACKGROUND

[0002] With the rapid development of biomechanics, sports medicine and human-computer interaction technology, lower limb action recognition technology based on surface electromyogram signals is increasingly widely used in the fields of rehabilitation medicine, prosthesis control, sports training and the like. As a non-invasive physiological signal, surface electromyogram signals can reflect the muscle activity state and further reflect the dynamic process of human motion. For different lower limb action recognition tasks, how to extract accurate and effective features from surface electromyogram signals becomes a key problem of research.

[0003] Traditional surface electromyogram signal analysis methods mainly rely on time domain, frequency domain or time-frequency domain features of signals. However, due to the influence of noise, motion artifacts and other factors on surface electromyogram signals, direct processing of original surface electromyogram signals may lead to low recognition accuracy. In addition, surface electromyogram signals usually have high nonlinearity and time-varying characteristics, and it is particularly important to extract features that comprehensively and accurately represent the complexity of surface electromyogram signals.

[0004] At present, although there are some lower limb action recognition methods based on surface electromyogram signals, there are still problems such as low signal quality and incomplete feature extraction. Therefore, how to improve the processing accuracy of surface electromyogram signals and construct an efficient and accurate lower limb action recognition system is still a technical challenge to be solved. SUMMARY

[0005] To solve the above technical problems, the application provides a lower limb action accurate recognition method based on surface electromyogram signal decomposition and optimization processing to solve the problems existing in the prior art.

[0006] To achieve the above purpose, the application provides a lower limb action accurate recognition method based on surface electromyogram signal decomposition and optimization processing, comprising:

[0007] Obtaining surface electromyogram signals, decomposing the surface electromyogram signals to obtain sub-signals;

[0008] Optimizing the sub-signals by a two-stage modal selection method, extracting and fusing features from the optimized sub-signals to obtain a feature vector, and reducing the dimension of the feature vector; wherein the two-stage modal selection method comprises an energy contribution selection method and a modal correlation selection method;

[0009] Recognizing the reduced dimension feature vector by a classification algorithm to obtain a lower limb action recognition result.

[0010] Optionally, the surface electromyography signal acquisition process comprises:

[0011] An original surface electromyography signal is acquired, and the original surface electromyography signal is preprocessed to obtain a surface electromyography signal; wherein the preprocessing process comprises filtering, denoising and normalization processing; wherein the filtering adopts a 20-450 Hz band-pass filter and a 50 Hz notch filter, and the denoising adopts a multi-scale principal component analysis method.

[0012] Optionally, the surface electromyography signal is decomposed by an empirical mode decomposition method.

[0013] Optionally, the process of optimizing the sub-signals by a two-stage mode selection method comprises:

[0014] The energy of the sub-signals is calculated by an energy contribution method to obtain an energy contribution ratio of the sub-signals, and sub-signals with an energy contribution ratio lower than 2% are removed to obtain energy sub-signals;

[0015] The correlation between the energy sub-signals and the surface electromyography signal is calculated by a mode correlation selection method to obtain the correlation between the energy sub-signals and the correlation between the energy sub-signals and the surface electromyography signal, and energy sub-signals with a correlation between the energy sub-signals lower than 85% and a correlation between the energy sub-signals and the surface electromyography signal greater than 10% are retained to obtain the optimized sub-signals.

[0016] Optionally, the process of obtaining the energy contribution ratio of the sub-signals comprises:

[0017] The energy of the sub-signals is calculated:

[0018]

[0019] E k is the energy of the kth sub-signal sub, t i is the time sequence of the sub-signal, and N is the number of data points;

[0020] The energy contribution ratio of the sub-signals is calculated:

[0021]

[0022] wherein, E total is the total energy of the signal, and E k represents the energy contribution ratio of the kth sub-signal.

[0023] Optionally, the process of correlation calculation comprises:

[0024]

[0025] wherein, x i , y i are sample points of two signals, in the correlation calculation process between the energy sub-band signals and the surface myoelectric signals, x i represents the energy sub-band signals, y i represents the pure surface myoelectric signals, in the correlation calculation process between the energy sub-band signals, x i represents the energy sub-band signals, y i represents another energy sub-band signal. are the mean values of the two signals respectively.

[0026] Optionally, the process of feature extraction and fusion on the optimized sub-signals comprises:

[0027] The k-neighbor estimation entropy feature calculation method is used to extract features from the optimized sub-signals, and the extracted features are obtained, and the feature vectors are obtained by combining the extracted features.

[0028] Optionally, the process of obtaining the extracted features comprises:

[0029]

[0030] wherein, represents the k-neighbor estimation entropy feature, ψ is the digamma function, N is the number of samples, ε(i) is the distance between the ith sample and its kth neighbor sample, i represents the sample number, k represents the neighbor sample number, c d is the volume of a unit sphere with dimension d, d represents the corresponding dimension, and the Euclidean regularization calculation formula of c d is as follows:

[0031]

[0032] wherein, Γ(.) is the Gamma function.

[0033] Optionally, the feature vectors are reduced in dimension by a feature dimension reduction algorithm, and the feature dimension reduction algorithm adopts a diffusion mapping feature dimension reduction algorithm.

[0034] Optionally, the classification algorithm adopts a machine learning classification algorithm.

[0035] Compared with the prior art, the present application has the following advantages and technical effects:

[0036] The application provides a lower limb action accurate recognition method based on surface electromyogram signal decomposition and optimization processing.

[0037] The technical scheme provided by the application is suitable for various lower limb action recognition scenes, and has wide application potential in the fields of rehabilitation medicine, intelligent prosthesis control, motion monitoring and the like. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein for a purpose of explanations and are not intended as an undue limitation of the application. In the drawings:

[0039] Figure 1 It is a whole method flow diagram of the embodiment of the application;

[0040] Figure 2 It is a surface electromyogram signal processing flow chart of the embodiment of the application;

[0041] Figure 3 It is a surface electromyogram signal decomposition and optimization processing flow chart of the embodiment of the application;

[0042] Figure 4 It is a feature extraction flow chart of the embodiment of the application;

[0043] Figure 5 It is a feature dimension reduction flow chart of the embodiment of the application;

[0044] Figure 6 A lower limb action recognition flowchart of an embodiment of the present application. DETAILED DESCRIPTION

[0045] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0046] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0047] The present application provides a lower limb action accurate recognition method based on surface electromyography signal decomposition and optimization processing, comprising the following steps:

[0048] Surface electromyography sensors are installed on different muscle parts of the human lower limbs to collect original surface electromyography signals of multiple muscles during execution of different lower limb actions (such as walking on flat ground, climbing stairs, descending stairs, and jumping).

[0049] The collected original surface electromyography signals are preprocessed. Band-pass filtering is used to remove low-frequency noise and high-frequency noise, ensuring that the effective frequency band of the surface electromyography signals is retained; notch filtering is used to remove power frequency noise in the surface electromyography signals; multi-scale principal component analysis is used to remove random noise in the surface electromyography signals; amplitude normalization processing is performed on the surface electromyography signals to eliminate differences between different muscle parts or individuals, obtaining pure surface electromyography signals.

[0050] The pure surface electromyography signals after preprocessing are decomposed using a data-driven decomposition method, and the pure surface electromyography signals are decomposed into multiple sub-band signals; a two-stage modal selection method is used on the sub-band signals to obtain multiple sub-component signals. Through multi-level signal decomposition, more dynamic features hidden in the signals can be extracted, wherein the data-driven decomposition method uses an empirical mode decomposition method.

[0051] Entropy features are extracted from the multiple sub-components and combined into a feature vector. This feature can effectively reflect the complexity and non-linear characteristics of the signal, thereby providing more useful information for lower limb action recognition.

[0052] Based on a feature dimension reduction algorithm, the extracted entropy features are reduced in dimension to remove redundant information and improve computational efficiency and recognition accuracy.

[0053] The feature set after dimension reduction is input into a classification algorithm to accurately recognize four lower limb actions (such as walking on flat ground, climbing stairs, descending stairs, and jumping).

[0054] This invention can be integrated into wearable devices, rehabilitation training systems, prosthetic control systems, and other applications. Based on the above technical solution, it can achieve the monitoring and accurate recognition of human lower limb movements, and is applicable to various scenarios such as sports training, rehabilitation therapy, and assisted prosthetic control.

[0055] This invention proposes a precise lower limb movement recognition method based on surface electromyography (SEMG) signal decomposition and optimization. It employs SEMG signal acquisition technology from multiple lower limb muscles to obtain richer and more comprehensive raw SEMG signals. Noise removal is performed on the raw SEMG signals through signal preprocessing to obtain purer and more reliable signals. The pure SEMG signals are then decomposed into multiple sub-band signals using signal decomposition technology, and a two-stage modality selection method is used to extract multi-level SEMG signal information, improving the expressive power and information richness of the SEMG signals. Entropy features are used to quantify the SEMG signals. Feature dimensionality reduction methods are employed to reduce redundant features and improve classification efficiency and accuracy. Finally, a classification algorithm is used to achieve efficient and accurate lower limb movement recognition.

[0056] The technical solutions of the present invention will be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, the present invention provides a method for accurate recognition of lower limb movements based on surface electromyography signal decomposition and optimization processing, which specifically includes the following:

[0057] Step 1: Acquisition of surface electromyography signals:

[0058] Surface electromyography (EMG) sensors are installed in multiple key muscle areas of the lower limbs of the human body, and surface EMG signals are collected synchronously at a sampling frequency of 1000Hz to obtain raw surface EMG signals with richer and more comprehensive motion information.

[0059] The muscles involved include: rectus femoris, vastus medialis, vastus lateralis, semitendinosus, tibialis anterior, lateral gastrocnemius, and medial gastrocnemius. These muscle groups play an important role in motor control during lower limb movements.

[0060] like Figure 2 As shown, step two: surface electromyography signal preprocessing:

[0061] The process of surface electromyography (EMG) signal processing includes filtering, denoising, and normalization of the EMG signal. This step ensures the purity of the EMG signal, laying the foundation for subsequent analysis. Specifically, it includes:

[0062] The acquired raw surface electromyography (EMG) signals were filtered using a 20-450Hz bandpass filter to remove low-frequency and high-frequency noise, while retaining the frequency bands related to the surface EMG signals of lower limb movement.

[0063] For the frequency band related to the surface electromyogram signal of lower limb movement, a 50 Hz notch filter is used to remove power frequency noise, and the filtered surface electromyogram signal is obtained.

[0064] For the filtered surface electromyogram signal, a multi-scale principal component analysis is used to remove random noise and artifacts in the surface electromyogram signal, and a denoised surface electromyogram signal is obtained.

[0065] The denoised surface electromyogram signal is normalized to keep the signal amplitude consistent between different electrodes. The normalization process selects the signal to be normalized to the range [0, 1]

[0066] The calculation formula of the normalization process is:

[0067]

[0068] Where x min is the minimum value of the surface electromyogram signal, and x max is the maximum value of the surface electromyogram signal.

[0069] As shown in Figure 3 Step 3: Surface electromyogram signal decomposition and optimization processing:

[0070] The pure surface electromyogram signal is decomposed into multiple sub-signals by a two-stage modal selection method. Specifically, it includes:

[0071] The preprocessed pure surface electromyogram signal is decomposed into multiple sub-band signals by the empirical mode decomposition method.

[0072] The two-stage modal selection method is applied to the multiple sub-band signals to select the sub-components that can effectively represent the lower limb movement.

[0073] The first stage of the two-stage modal selection method is the energy contribution selection method: the energy of each sub-band signal is calculated, and the proportion of the energy contribution of each sub-band signal is calculated. The sub-band signal with an energy contribution proportion less than 2% is discarded, and the energy sub-band signal is obtained;

[0074] The energy calculation formula of each sub-band signal is:

[0075]

[0076] Where E k is the energy of the kth sub-band signal sub, t i is the time series of the signal, N is the number of data points, and sub k () represents the kth sub-band signal sub in the time series t i of the signal.

[0077] The proportion of the energy contribution of each sub-band signal is calculated as follows:

[0078]

[0079] where E total is the total energy of the signal, Z k represents the proportion of the kth sub-band signal energy contribution.

[0080] The second stage is the modal correlation selection method: the correlation between the energy sub-band signals and the correlation between the energy sub-band signals and the pure surface myoelectric signal are calculated, and the energy sub-band signals that satisfy the correlation between the energy sub-band signals less than 85% and the correlation between the energy sub-band signals and the pure surface myoelectric signal greater than 10% are selected, to obtain the energy sub-band signals that can effectively represent the lower limb action sub-component.

[0081] The correlation calculation formula is:

[0082]

[0083] where x i , y i are sample points of two signals, in the correlation calculation process between the energy sub-band signals and the surface myoelectric signal, x i represents the energy sub-band signal, y i represents the pure surface myoelectric signal, in the correlation calculation process between the energy sub-band signals, x i represents the energy sub-band signal, y i represents another energy sub-band signal; x, y are the mean values of the two signals respectively.

[0084] As shown in Figure 4 , step four: feature extraction:

[0085] Extract features from the multiple sub-signals obtained by decomposition and combine them into a feature vector. Specifically, it includes:

[0086] Extract the k-nearest neighbor estimated entropy feature from the decomposed sub-component which is helpful for lower limb action recognition. The calculation formula is as follows:

[0087]

[0088] where, represents the k-nearest neighbor estimated entropy feature, i.e. the extracted feature, ψ is the digamma function, N is the number of samples, ε(i) is the distance between the ith sample and its kth nearest neighbor sample, i represents the sample number, k represents the number of adjacent samples, c d is the volume of a unit sphere with dimension d, and the Euclidean regularization calculation formula is as follows:

[0089]

[0090] Where Γ(.) is the Gamma function.

[0091] Each sub-component is calculated with a k-nearest neighbor estimated entropy feature. The k-nearest neighbor estimated entropy features calculated from all sub-components are combined into a feature vector, which serves as the input feature for lower limb action recognition.

[0092] like Figure 5 As shown, step five: feature dimensionality reduction algorithm:

[0093] Feature reduction algorithms are used to reduce the dimensionality of feature vectors, thereby improving the efficiency and accuracy of lower limb movement recognition. Specifically, this includes:

[0094] The feature vectors are reduced in dimensionality using the diffusion mapping feature dimensionality reduction algorithm to remove redundant information and retain the principal components that are most discriminative for lower limb movement recognition, thus obtaining the dimensionality-reduced feature set.

[0095] The diffusion mapping feature dimensionality reduction algorithm includes the following steps:

[0096] (1) For the eigenvector F = {f1,...,f N}, where i represents the eigenvector label, i = 1, 2, ... N. A corresponding finite graph is constructed, and the weights of the edges in the graph are defined using a positive definite symmetric Gaussian kernel function, resulting in a weight matrix W, whose elements w(f i ,f j )for:

[0097]

[0098] Where σ is the scaling parameter of the Gaussian kernel.

[0099] (2) Construct a Markov weight matrix P, with elements p(f i ,f j )for:

[0100]

[0101] in Represented as data point f i The degree.

[0102] Each row in the matrix represents the transition probability of a one-step random move between any two data points, and P represents the transition probability matrix after t random moves. t (f i ,f j )for:

[0103] P t (f i ,f j )=(P(f i ,fj )) t

[0104] (3) Determining diffusion distance

[0105] Diffusion distance D of two points t (f i ,f j ) is defined as follows:

[0106]

[0107] wherein, represents the feature that the weight of the high-density area in the graph is greater.

[0108] (4) Under the premise of maintaining the diffusion distance between data points, the low-dimensional manifold Y can be obtained by the following formula:

[0109] P t Y = λY

[0110] wherein λ represents the eigenvalue of the matrix P t , the largest eigenvalue (trivial eigenvalue) λ0=1 and the corresponding eigenvector v0 are discarded, and the eigenvectors corresponding to the remaining D eigenvalues are taken as the low-dimensional embedding result, to obtain the low-dimensional manifold Y, and complete the dimension reduction of the eigenvector F. The low-dimensional manifold Y: Y = {λ1v1, λ2v2,..., λ D v D}, wherein v represents an eigenvector, the subscripts 0, 1, 2, and D represent the serial number of the corresponding eigenvector, and D represents the total number of eigenvectors extracted.

[0111] As shown in FIG. 6, step six is lower limb action recognition. Figure 6

[0112] According to the reduced feature set, the accurate recognition of the lower limb action is completed in combination with a classification algorithm. Specifically, the reduced feature set is input into a support vector machine classification algorithm for lower limb action recognition, to realize the accurate recognition of four lower limb actions. It should be noted that the classification algorithm can also use a linear discriminant analysis classification algorithm, an integrated bagging classification algorithm, a K-nearest neighbor classification algorithm, a naive Bayes classification algorithm, and other machine learning classifiers.

[0113] The four lower limb actions include walking on flat ground, climbing stairs, descending stairs, and jumping.

[0114] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements easily thought of by those skilled in the art within the technical range disclosed in the present application should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.​

Claims

1. A method for precise recognition of lower limb movement based on decomposition and optimization processing of surface electromyography signals, characterized in that, The method comprises the following steps: obtaining a surface electromyogram signal, decomposing the surface electromyogram signal to obtain a sub-signal; optimizing the sub-signal through a two-stage modal selection method, extracting and fusing features of the optimized sub-signal to obtain a feature vector, and reducing the dimension of the feature vector; wherein the two-stage modal selection method comprises an energy contribution selection method and a modal correlation selection method; identifying the reduced feature vector through a classification algorithm to obtain a lower limb action recognition result; the process of optimizing the sub-signal through the two-stage modal selection method comprises: calculating the energy of the sub-signal through the energy contribution method to obtain an energy contribution ratio of the sub-signal, removing the sub-signal with an energy contribution ratio lower than 2% to obtain an energy sub-signal; calculating the correlation between the energy sub-signals and the surface electromyogram signal through the modal correlation selection method to obtain the correlation between the energy sub-signals and the correlation between the energy sub-signals and the surface electromyogram signal, and retaining the energy sub-signal with a correlation between the energy sub-signals lower than 85% and a correlation between the energy sub-signals and the surface electromyogram signal greater than 10% to obtain the optimized sub-signal; the process of obtaining the energy contribution ratio of the sub-signal comprises: calculating the energy of the sub-signal: ； the energy of the kth sub-signal sub, is the time series of the sub-signals, and N is the number of data points. calculating the energy contribution ratio of the sub-signal: ； wherein, is the total energy of the signal, which is the sum of the energy of the sub-signals, represents the proportion of the energy contribution of the kth sub-signal. the process of correlation calculation comprises: ； wherein, , are sample points of two signals, in the correlation calculation process between the energy sub-band signals and the surface myoelectric signals, denotes an energy sub-band signal, denotes a pure surface myoelectric signal, in the correlation calculation process between the energy sub-band signals, denotes an energy sub-band signal, denotes another energy sub-band signal; , are mean values of two signals respectively; the process of obtaining the extracted features comprises: ； wherein, denotes the k-nearest neighbor estimated entropy feature, i.e., the extracted feature, is the digamma function, and N is the number of samples, is the distance between the ith sample and its kth nearest neighbor sample, i denotes the sample index, and k denotes the nearest neighbor index, is the volume of a unit sphere of dimension d, d denotes the corresponding dimension, and The Euclidean regularization formula of is: wherein is the Gamma function.

2. The method of claim 1, wherein the process of obtaining the surface electromyogram signal comprises: obtaining an original surface electromyogram signal, and obtaining the surface electromyogram signal through preprocessing of the original surface electromyogram signal; wherein the preprocessing process comprises filtering, denoising, and normalization processing; wherein the filtering adopts a 20-450 Hz band-pass filter and a 50 Hz notch filter, and the denoising adopts a multi-scale principal component analysis method.

3. The method of claim 1, wherein the surface electromyogram signal is decomposed through an empirical mode decomposition method.

4. The method of claim 1, wherein the process of extracting and fusing features of the optimized sub-signal comprises: extracting features of the optimized sub-signal through a k-nearest neighbor estimated entropy feature calculation method to obtain extracted features, and combining the extracted features to obtain a feature vector.

5. The method of claim 1, wherein the feature vector is reduced in dimension through a feature dimension reduction algorithm, wherein the feature dimension reduction algorithm adopts a diffusion mapping feature dimension reduction algorithm.

6. The method of claim 1, wherein the classification algorithm adopts a machine learning classification algorithm.

Images