Non-ideal myoelectricity gesture recognition method based on Riemannian manifold transfer learning

Abstract

The invention relates to a non-ideal myoelectricity gesture recognition method based on Riemannian manifold transfer learning, which is characterized by comprising four parts: 1) constructing a covariance matrix by using spatial features of a myoelectricity acquisition channel as a basis; 2) constructing a Riemannian geometric manifold structure in a Riemannian space by taking the covariance matrix as a positive definite matrix; 3) constructing a Riemannian geometric center of the data and aligning the geometric center in a manifold space to realize transfer learning of the data; and 4) classifying different gestures by using a minimum Riemannian distance classifier combined with Fischer linear discriminant filtering. According to the method, a Riemannian manifold alignment transfer learning method is applied to gesture recognition of electromyographic signals, the influence of non-ideal factors such as electrode displacement and muscle fatigue existing in electromyographic signal recognition is solved by aligning manifold structures of a source domain and a target domain, and an effective electromyographic gesture recognition model is established.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics, pattern recognition, and brain-computer interface technology, and in particular to a non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning. Background Technology

[0002] For individuals with movement disorders and stroke patients, hemiplegia is a common symptom, making basic hand movements such as grasping difficult. In clinical rehabilitation, rehabilitation gloves can effectively alleviate these symptoms and help patients gradually regain basic hand function. Recently, researchers combined rehabilitation gloves with surface electromyography (EMG) signals. By detecting and analyzing the electrical signals of muscles, they were able to capture the patient's hand movement intentions and thus control the rehabilitation gloves.

[0003] In traditional gesture recognition based on electromyography (EMG) signals, much attention has been paid to the time-domain and time-frequency domain features of EMG signals. Decoding these EMG signal features has played a crucial role in the development of gesture recognition. However, traditional EMG signal feature extraction processes extract features within a flat Euclidean geometric space, neglecting the spatial statistical characteristics of EMG signals as physiological features. Researchers have discovered that non-Euclidean representations of EMG signals in Riemannian space possess more comprehensive and usable statistical features, making them a more natural input for model data structures. In Riemannian manifold space, symmetric positive definite matrices are a commonly used Riemannian metric for studying nonlinear spaces. They utilize the covariance matrix as a symmetric positive definite matrix, leveraging its geometric properties to describe and analyze data structure, capturing the nonlinear relationships within the Riemannian manifold space. Representing and analyzing EMG signals within a Riemannian manifold space provides a new approach for more comprehensive extraction of their spatial features.

[0004] Although gesture recognition methods based on electromyography (EMG) signals have made significant progress, some non-ideal factors still exist during the acquisition of EMG signals, posing numerous challenges. How to effectively acquire high-quality EMG signals under these non-ideal conditions has become a key research focus for researchers. Summary of the Invention

[0005] To address the shortcomings or improvement needs of existing technologies, this invention applies the Riemann manifold alignment transfer learning method to gesture recognition of electromyographic signals. By aligning the manifold structures of the source and target domains, the influence of non-ideal factors such as electrode displacement and muscle fatigue in electromyographic signal recognition is resolved, and an effective electromyographic gesture recognition model is established.

[0006] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0007] A non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning includes the following steps:

[0008] 1) Collect electromyographic characteristic data and express them in Riemann space;

[0009] 2) Perform migration identification and classification on the processed data.

[0010] Step 1) includes the following steps:

[0011] 1.1) Obtain the electromyographic signal characteristic data of existing individuals as known data;

[0012] 1.2) Collect electromyographic signal characteristic data of new individuals or unknown characteristic data that need to be classified as unknown data;

[0013] 1.3) Bandpass filtering is performed on the known and unknown data to construct the covariance matrix, which is then expressed as a positive definite matrix in Riemann space.

[0014] The covariance matrix C is specifically:

[0015]

[0016] Where μ is the mean vector, T is the transpose of the matrix, and x i The data represents the electromyographic signal characteristics of the i-th channel, with the italicized T representing the time point.

[0017] Step 2) includes the following steps:

[0018] 2.1) Map the covariance matrix to the tangent space using a logarithmic mapping;

[0019] 2.2) By employing a Riemannian space data alignment strategy, the Riemannian geometric centers of known and unknown data are aligned in the manifold space;

[0020] 2.3) Perform Fisher linear discriminant filtering on the mapped data;

[0021] 2.4) Based on the filtered data, the gesture type of the unknown data is identified by the minimum Riemann distance classifier.

[0022] Step 2.1) specifically refers to:

[0023] Let the space of the set of tangent vectors of a point in the Riemannian manifold be called the tangent space. Points in this space are mapped onto the Riemannian manifold through an exponential mapping, i.e.:

[0024] P i =Exp p (S i ) = P 1 / 2exp(P -1 / 2 S i P -1 / 2 )P 1 / 2

[0025] The points on the Riemannian manifold are mapped to the tangent space by the inverse logarithmic mapping, i.e.:

[0026] S i =Log p (P i ) = P 1 / 2 exp(P -1 / 2 P i P -1 / 2 )P 1 / 2

[0027] Among them, P i For a point S in Riemannian space i Let P be the point i The tangent vectors, each tangent vector S i Starting from point P and heading to P i geodesic Γ i The derivative of (t) at t=0.

[0028] The Log-Euclidean mapping method is used to map each covariance matrix to a tangent space with its geometric mean as the reference point, where the Riemann geometric mean is... Specifically:

[0029]

[0030] Where p(C) is the geometry of the Riemannian manifold space formed by the covariance matrix, δ R (P,P i Let λ be the Riemann distance between the two points. i For P -1 P i The eigenvalues ​​are I, the total number of samples in the matrix set, and n is the dimension of the covariance matrix.

[0031] Step 2.2) specifically involves:

[0032] Use the known data as the reference matrix

[0033]

[0034] Assume the covariance matrix of the unknown data is The Riemann contrast method performs the following alignment transformation on each covariance matrix:

[0035]

[0036] in, It is the aligned covariance matrix.

[0037] Step 2.3) includes the following steps:

[0038] 2.3.1) Transforming the covariance matrix through logarithmic mapping Mapping onto the tangent space yields S i ;

[0039] 2.3.2) Using Fisher's linear discriminant analysis criterion to analyze S i Solving for the geodesic direction yields the discriminant vector.

[0040] 2.3.3) S i and Perform filtering:

[0041]

[0042] Where vec(·) is the vectorization operator;

[0043] 2.3.4) Through exponential mapping Mapping onto the Riemannian manifold space yields the covariance matrix.

[0044] Step 2.4) includes the following steps:

[0045] 2.4.1) Calculate the Riemann mean distance P for the classification task. φ ;

[0046] 2.4.2) Calculate the Riemann distance d from the test sample to each mean. k

[0047]

[0048] 2.4.3) By finding the minimum category index k of different tasks min The task corresponding to this index is taken as the final classification result, that is:

[0049]

[0050] The present invention has the following beneficial effects and advantages:

[0051] 1. This invention applies the Riemannian manifold alignment transfer learning method to solve the gesture recognition problem under non-ideal conditions, which can effectively capture the spatial feature information of electromyographic signals and reduce the adverse effects of non-influencing factors on the gesture recognition process.

[0052] 2. This invention employs the Riemann contrast method to center-align the covariance matrices of data from different sources, thereby achieving effective transfer of gesture recognition based on electromyographic signals under non-ideal conditions.

[0053] 3. This invention utilizes the Fisher-Riemann classifier method to classify on the tangent plane of geodesics, thus avoiding the problem of relevant information being ignored in the traditional classification process. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method of the present invention;

[0055] Figure 2 This is a visual comparison of the feature distributions of the source and target domains before and after alignment in the binary classification electrode offset experiment.

[0056] Figure 3 This is a visualization of the seven-category classification results of some experimental individuals in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0058] A non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning, including an EEG signal correction process and an integrated recognition process;

[0059] This includes the Riemannian spatial representation of electromyographic features and the transfer recognition and classification process;

[0060] The Riemann space expression process of the electromyographic features includes the following steps:

[0061] 1.1) Obtain the electromyographic signal features of existing individuals as known data (source feature data), and collect the electromyographic signal feature data of new individuals (or unknown feature data that needs to be classified) as unknown data;

[0062] 1.2) The multi-channel electromyography (EMG) signal features are processed, and the covariance matrix of the EMG features is expressed as a positive definite matrix in Riemann space;

[0063] The migration identification and classification process includes the following steps:

[0064] 2.1) Introduce a Riemann space data alignment strategy to align the Riemann geometric centers of the two sets of data in the manifold space to achieve data transfer learning;

[0065] 2.2) Combine Fisher linear discriminant filtering and minimum Riemann distance classifier to identify the gesture type of unknown data.

[0066] like Figure 1 The diagram shown is a flowchart of the method of the present invention.

[0067] Source individual (known individual) electromyographic signal characteristics D m ∈R c×l m = 1, 2, ..., N, where N is the total number of existing individuals, and D is the electromyographic signal characteristic of the target individual (unknown individual). n ∈R c×l Here, c represents the number of EMG signal channels, and l represents the number of sampling points. The dataset used for testing consisted of 36 individuals, with 8 EMG signal channels and a sampling frequency of 200Hz. Before the classification task, the EMG signal features of known and unknown individuals were filtered using a bandpass filter, and the filtered EMG signal features from 10 to 100Hz were used as the model input. The network addresses non-ideal factors in EMG-based gesture recognition tasks, including electrode offset, individual differences, diurnal variations, muscle fatigue, and arm posture during the data acquisition process.

[0068] For each segment of preprocessed electromyography data Where C represents the number of channels and T represents the number of time points, its spatial covariance matrix is ​​constructed using the following formula:

[0069]

[0070] Where μ is the mean vector and T represents the transpose of the matrix, the resulting covariance matrix is ​​a symmetric positive definite matrix.

[0071] In Riemannian manifolds, the tangent space is often used to map data on the manifold. The tangent space is the space consisting of the set of tangent vectors at a given point, and points on it can be mapped to the Riemannian manifold through an exponential mapping:

[0072] P i =Exp p (S i ) = P 1 / 2 exp(P -1 / 2 S i P -1 / 2 )P 1 / 2

[0073] Similarly, an inverse (logarithmic) mapping can be used to map points on a Riemannian manifold to the tangent space:

[0074] S i =Log p (P i ) = P 1 / 2 exp(P -1 / 2 P i P -1 / 2 )P 1 / 2

[0075] Each tangent vector S iIt can be viewed as starting from point P and going to point P. i geodesic Γ i The derivative of (t) at t=0.

[0076] To achieve efficient Riemannian manifold mapping, the Log-Euclidean mapping method is used, mapping each covariance matrix to a tangent space with its geometric mean as the reference point. The Log-Euclidean mapping method requires defining the Riemann geometric mean, which can be calculated using the following formula:

[0077]

[0078] Where p(C) is the geometry of the Riemannian manifold space formed by the covariance matrix, δ R (P,P i ) represents the Riemann distance between two points, λ i For P -1 P i The eigenvalues ​​of the Riemannian manifold are shown above. This representation of the Riemannian manifold will be used in the subsequent filtering process.

[0079] After obtaining the covariance matrix, the centers of the covariance matrices from data from different sources are aligned using the Riemann contrast method. The subject data under known conditions is used as the reference matrix, which is defined as follows:

[0080]

[0081] Assume the covariance matrix of the unknown subjects is The Riemann contrast method performs the following alignment transformation on each covariance matrix:

[0082]

[0083] in It is the aligned covariance matrix. Due to the affine invariance of the Riemann distance, this transformation will not change the geometric relationship between the covariance matrices within the subjects, nor will it destroy the category information contained in the original pattern. Figure 2 This is a visualization comparison of the feature distributions of the source and target domains before and after alignment in a binary classification electrode offset experiment.

[0084] To achieve filtering, the covariance matrix is ​​transformed using a logarithmic mapping. Mapping onto the tangent space yields S i By solving the Fisher linear discriminant analysis criterion, the discriminant vector of the geodesic direction is obtained. S i and Perform filtering

[0085]

[0086] Where vec(·) is a vectorization operator that preserves the vector form of the upper triangular matrix. After the filtering operation, it is transformed through exponential mapping. Mapping onto the Riemannian manifold space yields the covariance matrix. This allows filtering operations in the manifold space to achieve linear processing similar to those in the tangent space while preserving the characteristic constraints of the space.

[0087] Finally, the Riemann mean distance P for the classification task is calculated. φ The Riemann distance from the test sample to each mean is calculated using the formula.

[0088]

[0089] Then, the classification task is completed by finding the smallest category index for different tasks.

[0090]

[0091] Figure 3 This is a visualization of the seven-class classification results for some of the experimental individuals in this experiment. The network proposed in this experiment can achieve good classification results.

Claims

1. A non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning, characterized in that, Includes the following steps: 1) Collect electromyographic characteristic data and express them in Riemann space; 2) Perform migration identification and classification on the processed data.

2. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 1, characterized in that, Step 1) Includes the following steps: 1.1) Obtain the electromyographic signal characteristic data of existing individuals as known data; 1.2) Collect electromyographic signal characteristic data of new individuals or unknown characteristic data that need to be classified as unknown data; 1.3) Bandpass filtering is performed on the known and unknown data to construct the covariance matrix, which is then expressed as a positive definite matrix in Riemann space.

3. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 2, characterized in that, The covariance matrix C is specifically: Where μ is the mean vector, T is the transpose of the matrix, and x i The data represents the electromyographic signal characteristics of the i-th channel, with the italicized T representing the time point.

4. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 1, characterized in that, Step 2) includes the following steps: 2.1) Map the covariance matrix to the tangent space using a logarithmic mapping; 2.2) By employing a Riemannian space data alignment strategy, the Riemannian geometric centers of known and unknown data are aligned in the manifold space; 2.3) Perform Fisher linear discriminant filtering on the mapped data; 2.4) Based on the filtered data, the gesture type of the unknown data is identified by the minimum Riemann distance classifier.

5. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 4, characterized in that, Step 2.1) specifically refers to: Let the space of the set of tangent vectors of a point in the Riemannian manifold be called the tangent space. Points in this space are mapped onto the Riemannian manifold through an exponential mapping, i.e.: P i =Exp p (S i )=P 1 / 2 exp(P -1 / 2 S i P -1 / 2 )P 1 / 2 The points on the Riemannian manifold are mapped to the tangent space by the inverse logarithmic mapping, i.e.: S i =Log p (P i )=P 1 / 2 exp(P -1 / 2 P i P -1 / 2 )P 1 / 2 Among them, P i For a point S in Riemannian space i Let P be the point i The tangent vectors, each tangent vector S i Starting from point P and heading to P i geodesic Γ i The derivative of (t) at t=0.

6. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 5, characterized in that, The Log-Euclidean mapping method is used to map each covariance matrix to a tangent space with its geometric mean as the reference point, where the Riemann geometric mean is... Specifically: Where p(C) is the geometry of the Riemannian manifold space formed by the covariance matrix, δ R (P,P i Let λ be the Riemann distance between the two points. i For P -1 P i The eigenvalues ​​are I, the total number of samples in the matrix set, and n is the dimension of the covariance matrix.

7. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 4, characterized in that, Step 2.2) specifically refers to: Use the known data as the reference matrix Assume the covariance matrix of the unknown data is The Riemann contrast method performs the following alignment transformation on each covariance matrix: in, It is the aligned covariance matrix.

8. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 4, characterized in that, Step 2.3) includes the following steps: 2.3.1) Transforming the covariance matrix through logarithmic mapping Mapping onto the tangent space yields S i ; 2.3.2) Using Fisher's linear discriminant analysis criterion to analyze S i Solving for the geodesic direction yields the discriminant vector. 2.3.3) S i and Perform filtering: Where vec(·) is the vectorization operator; 2.3.4) Through exponential mapping Mapping onto the Riemannian manifold space yields the covariance matrix.

9. The non-ideal electromyographic gesture recognition method based on Riemannian manifold transfer learning according to claim 4, characterized in that, Step 2.4) includes the following steps: 2.4.1) Calculate the Riemann mean distance P for the classification task. φ ; 2.4.2) Calculate the Riemann distance d from the test sample to each mean. k : 2.4.3) By finding the minimum category index k of different tasks min The task corresponding to this index is taken as the final classification result, that is:

Images