A transfer learning method for myoelectric gesture recognition based on stability feature mining
Through the transfer learning method of stability feature discovery, the sample distribution change caused by electrode offset in the surface electromyography signal recognition system is solved, the recognition accuracy is improved, and the feature extraction is provided, which achieves the lack of strict requirements. High accuracy recognition in the case of wearing position.
Patent Information
- Application Number
- CN202410112859.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-26
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-01-26
AI Technical Summary
In daily use, the surface electromyography signal recognition system changes sample distribution due to electrode offset, and the traditional model has poor recognition effect and low accuracy.
Using the transfer learning method of stability feature discovery, the mapping matrix A, offset vector b and trust matrix F are optimized by joint iteratively to map the offset samples to the distribution of unoffset samples, and the knowledge training model is used for the unoffset samples. , identify the offset samples, and analyze the sample feature changes through stability feature discovery factors.
The accuracy of electromyography gesture recognition is improved, and guidance is provided on the impact of electrode offset on features, providing research directions for future feature extraction methods, and achieving high accuracy recognition without strict requirements on wearing positions.
Smart Images

Figure CN118034493B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of surface electromyographic signal (sEMG) recognition, and in particular relates to a transfer learning electromyographic gesture recognition method for exploring stability features. Background Art
[0002] Surface electromyography (sEMG) is a biological current generated by the contraction of surface muscles of the human body. Gesture recognition through electromyography plays an important role in the field of human-computer interaction. Typical applications of electromyography gesture recognition include software interaction, video games, prosthetic control, and sign language recognition. This ability to interpret gestures allows us to communicate with machines. In addition to surface electromyography, there are other methods for gesture recognition, such as visual sensors, invasive electromyography, motion sensors, etc., but they have some disadvantages. Visual sensors can capture the movement of the user's hands, but they have high requirements for the environment and light, and cannot function when blocked. Motion sensors cannot capture the details of muscle movement, affecting the accuracy of gesture recognition. Collecting invasive electromyography signals requires inserting the collection needle into the user's skin, which requires researchers to have the necessary medical skills or the help of professional doctors, increasing the difficulty of experiments and gesture recognition system development, and also making it difficult for some users to accept the system. Surface electromyography sensors only need to be attached to the user's skin to collect signals. Therefore, surface electromyography signals are favored by researchers in academic and medical fields.
[0003] However, surface electromyographic signals also have defects. In daily use of the system, user movement, sweating, and irregular wearing will cause electrode offset. Different electrode positions will also change the distribution of collected electromyographic signals. At present, most pattern recognition models are based on an assumption that the distribution of labeled samples and unlabeled samples is similar, and the data collected after electrode offset has limited similarity with the data distribution used in the training model, resulting in poor recognition effect and low precision of traditional models. Therefore, the present invention proposes a transfer learning electromyographic gesture recognition method for stability feature mining, establishes a mapping between offset samples and unoffset samples, re-establishes the distribution connection between offset samples and unoffset samples, and then uses the knowledge of unoffset samples to train a pattern recognition model to perform gesture recognition on offset samples, that is, transfer learning. Compared with directly using offset samples for recognition, the model obtains a higher recognition accuracy by using mapped samples for recognition. In addition, a stability feature mining factor is embedded in the method, through which the degree of change of sample features before and after mapping can be obtained, providing guidance for future research on feature extraction methods, such as paying more attention to features with high stability. Summary of the invention
[0004] The purpose of the present invention is to provide a transfer learning electromyographic gesture recognition method for stability feature mining; the method performs joint iterative optimization on the mapping matrix A, the offset vector b, the stability feature mining factor θ and the trust matrix F. First, as a semi-supervised method, in order to reduce the interference of unlabeled samples on the correctness of the method, we use the trust matrix to control the influence of each sample on the method; secondly, we can map the distribution of offset samples to non-offset samples according to the iterative optimization results to improve the accuracy of subsequent gesture recognition; finally, we can analyze the degree of change of sample features during the migration process through the stability feature mining factor θ, thereby analyzing the degree of influence of electrode offset on the features, and guiding the subsequent research on feature extraction of electromyographic gesture recognition.
[0005] In a first aspect, the present invention provides a method for myoelectric gesture recognition by transfer learning based on stability feature mining, which comprises the following steps:
[0006] Step 1: The subject wears the electromyographic acquisition device without deviation and makes a variety of different gestures. The collected electromyographic signals are preprocessed and feature extracted as non-deviation samples.
[0007] Step 2: Use the unbiased samples as the training set to train the classifier. Cluster the samples of each class in the training set separately, and obtain p prototype points for each class.
[0008] Step 3: Establish an EMG offset mapping model embedded with the stability feature discovery factor θ; the objective function of constructing the EMG offset mapping model is as follows:
[0009]
[0010] Where F is the sample trust matrix; S, T are the original image and the image respectively; A, Θ are the mapping matrix and the stability feature discovery matrix respectively; I is the unit matrix; b is the offset vector; β, γ are adjustable parameters.
[0011] Step 4: Each time the subject needs to start myoelectric gesture recognition, he / she wears the myoelectric acquisition device (at this time, the wearing position is not strictly required, and it is considered that there is an offset in wearing), and performs initial offset calibration. The process of initial offset calibration is: the subject makes a variety of different gestures according to the instructions, and obtains the labeled sample X with offset. l Each labeled sample selects the prototype point with the closest Euclidean distance from all prototype points of the same category as the prototype point. l There will be labeled samples X l and Bishop T l As the original image S and image T respectively, the objective function is iteratively updated to obtain the mapping matrix A, offset vector b and stability feature discovery matrix Θ of the labeled samples.
[0012] Step 5: After the offset calibration is completed, the electromyographic gesture recognition is performed; the subject makes the gesture to be recognized, and the electromyographic acquisition device collects the electromyographic signal of the subject; the obtained electromyographic signal is preprocessed and feature extracted as the unlabeled sample X u .
[0013] Step 6: Use parameters A, b, Θ to map the labeled sample X l and unlabeled samples X u , and iteratively update the parameters A, b, Θ based on the mapping results.
[0014] Step 7: Use the parameters A, b, Θ updated in step 6 to map the unlabeled sample X u ; The mapped unlabeled sample X′ u Input the classifier trained in step 2 to obtain the electromyographic gesture recognition result.
[0015] Among the above steps, steps 1-4 are the preparation stage, which only need to be performed once for each subject; steps 5-7 are the actual detection stage, which need to be re-executed each time the electromyography acquisition device is re-donned or the electromyography acquisition device is offset; thus, accurate gesture recognition results can still be obtained without strict requirements on the wearing position during the actual detection process.
[0016] Preferably, the specific process of step 6 is as follows:
[0017] (1) The sample trust matrix F is a diagonal matrix, and its i-th diagonal element is f i is the trust level of the i-th sample. Set the first l diagonal elements of the sample trust matrix F to constants α is the EMG offset mapping model for labeled samples X l The confidence level of α∈[0,1] is an adjustable parameter. l is the number of labeled samples.
[0018] (2) Use parameters A, b, Θ to map labeled samples X l and unlabeled samples X u ; After mapping, the sample [X′ l ; X′ u ]=AΘ[X l ;X u ]+b1 T , X′ l and X′ u They are the labeled samples and unlabeled samples after mapping, respectively.
[0019] (3) The mapped unlabeled sample X′ u Input the classifier trained in step 3 to obtain the mapped unlabeled sample X′ uTemporary labeling. According to the unlabeled sample X′ after mapping u The temporary mark selects the prototype point with the closest Euclidean distance among the prototype points of the same category as the prototype point T u .
[0020] (4) According to the trust update strategy, update the last u elements of the sample trust matrix F. u is the number of labeled samples.
[0021] (5) The mapped sample [X′ l ; X′ u ] as the original image S, and replace the image [T l ; T u ] as the image T; iteratively update the objective function to obtain the updated mapping matrix A, offset vector b and stability feature discovery matrix Θ.
[0022] (6) Use the updated mapping matrix A, offset vector b and stability feature discovery matrix Θ to map the unlabeled samples again and update the mapped unlabeled samples X′ u =AΘX u +b1 T .
[0023] (7) Iterate steps (2)-(6) multiple times to obtain the final mapping matrix A, stability feature discovery factor θ, and offset vector b.
[0024] Preferably, the trust updating strategy is:
[0025] Construct the confidence level f of the unlabeled samples i The update expression is as follows:
[0026] f i =ψ(dist2-dist1) (29)
[0027] Among them, dist1 is s i and the corresponding prototype point d i The distance, s i The distance to the nearest prototype point in other categories is dist2.
[0028] The function ψ(c) is defined as follows:
[0029]
[0030] Among them, θ, τ are adjustable parameters; c is the function independent variable.
[0031] Preferably, the stability feature discovery matrix θ is a diagonal matrix, and its i-th diagonal element is θ i , corresponding to the weight of the i-th feature of the sample;
[0032] Preferably, when the worn electromyographic acquisition device does not move relative to the human body, the measured electromyographic signals are mapped using the mapping matrix A, offset vector b and stability feature discovery matrix Θ obtained in step 6 and then input into the classifier trained in step 2 to continuously obtain new electromyographic gesture recognition results.
[0033] Preferably, when the worn electromyographic acquisition device does not move relative to the human body, each time a new electromyographic signal is measured, the new electromyographic signal is added to the unlabeled sample, and step 6 is repeated to further iteratively update the mapping matrix A, the stability feature discovery factor θ, and the offset vector b, and use the updated mapping matrix A, the offset vector b, and the stability feature discovery matrix Θ to map the newly measured electromyographic signal, and the mapping result is input into the classifier trained in step 2 to continuously obtain new electromyographic gesture recognition results.
[0034] Preferably, the feature extraction method in step 1 adopts time domain-autoregressive features, including root mean square, mean absolute value, zero crossing rate, slope change rate, waveform length, and autoregressive coefficient.
[0035] Preferably, in step 2, the trained classifier is a support vector machine.
[0036] Preferably, in step 2, K-means clustering method is used for clustering.
[0037] Preferably, in step 5, the method for iteratively updating the electromyographic offset mapping model is:
[0038] First, initialize the parameters, A = I, b = 1 d , θ=1 d / d, Θ=diag(θ), and then solve them separately. The specific process is as follows:
[0039] Fix other parameters and update the offset vector b. The objective function is equivalent to:
[0040]
[0041] Take the partial derivative of the formula with respect to b and set the derivative to 0, simplifying it to get:
[0042]
[0043] Where, M = (AθS-T)F, So the solution is:
[0044]
[0045] The specific process of iteratively solving the mapping matrix A is as follows:
[0046] Fix other parameters and update the mapping matrix A. The objective function is equivalent to:
[0047]
[0048] Taking the partial derivative of A and setting the derivative to 0, we can simplify to:
[0049] A=(βI-QP T )(PP T +βI) -1 (6)
[0050] Where P = ΘSF,
[0051] Fix other parameters and update the stability feature discovery factor θ. The objective function is equivalent to:
[0052]
[0053] The formula is simplified to:
[0054]
[0055] Let P = SFF T S T , R = SFF T (1 n b T -T T )A, the formula is equivalent to:
[0056]
[0057] Since Θ is a diagonal matrix, is the Hadamard product operator. Therefore, the formula can be transformed into:
[0058]
[0059] set up c = -2diag(R), there are:
[0060]
[0061] Introducing the auxiliary variable v, the formula becomes:
[0062]
[0063] The corresponding augmented Lagrangian method is:
[0064]
[0065] stθ T 1=1,θ≥0 (13)
[0066] Among them, λ is the Lagrange multiplier and μ is the augmentation parameter.
[0067] First, fix θ and update v. Take the partial derivative of v and set the derivative equal to 0 to get:
[0068]
[0069] Then fix v i Update f i , the formula is converted to:
[0070]
[0071] According to the perfect square formula of the matrix, it is transformed into:
[0072]
[0073] remember The above formula can be transformed into:
[0074]
[0075] The Lagrange equation is:
[0076]
[0077] where δ and v are Lagrange multipliers. Assume θ * is the optimal solution for θ, δ * and θ * is the corresponding multiplier. According to the KKT condition, for have
[0078]
[0079] The first equation in the formula can be rewritten as:
[0080] θ * -m-δ * 1-η * =0 (20)
[0081] Because θ T 1=1, so the above formula becomes:
[0082]
[0083] According to the formula and we can get:
[0084]
[0085] set up The formula can be rewritten as:
[0086] θ* =g+η * -η * 1 (23)
[0087] That is, for any t∈[1,d]:
[0088]
[0089] According to the formula and, we can know (a) + =max(a,0). So we can solve:
[0090]
[0091] If the optimal solution Determine, then the optimal solution θ * can also be determined by the above formula. The formula is rewritten as so So It can be calculated by the following formula:
[0092]
[0093] According to the restriction θ T 1=1 and, we define the function:
[0094]
[0095] when When It can be calculated by Newton's method
[0096]
[0097] In summary, the optimization results are as follows: θ=diag(θ).
[0098] In a second aspect, the present invention provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the aforementioned transfer learning electromyographic gesture recognition method for stability feature mining.
[0099] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned transfer learning electromyographic gesture recognition method for stability feature discovery.
[0100] The present invention has the following beneficial effects:
[0101] 1. The existing electromyographic gesture recognition method ignores the problem that the electrode wearing position is not standard in daily use, and is easily offset during the recognition process. The offset samples are inconsistent with the sample classification used by the training model, resulting in a decrease in the recognition accuracy of the model. The present invention can map the offset samples to a distribution similar to the training samples through a simple correction process, so that the recognition accuracy of the model is maintained within a reliable range.
[0102] 2. The transfer learning electromyographic gesture recognition method for stability feature mining proposed in the present invention provides a method for improving recognition accuracy for surface electromyographic gesture recognition and human-computer interaction research, and the stability feature mining factor θ that can be obtained by the iterative update method can be used to obtain the degree of change of sample features before and after mapping, providing guidance for future research on feature extraction methods, such as paying more attention to features with high stability.
[0103] 3. The present invention can be used as a preprocessing method for other methods, allowing other methods to obtain information about the electrode in an offset position, and to achieve adaptive correction of the electrode without changing other models. It has many practical scenarios and high practicability. BRIEF DESCRIPTION OF THE DRAWINGS
[0104] Figure 1 This is a flow chart of the surface electromyography gesture recognition method provided by the present invention.
[0105] Figure 2 This is a framework diagram of the surface electromyography gesture recognition method provided by the present invention. DETAILED DESCRIPTION
[0106] The present invention will be further described below in conjunction with the accompanying drawings.
[0107] like Figure 1 and Figure 2 As shown in FIG. 1 , the surface electromyography gesture recognition method combining semi-supervised regression and graph label propagation specifically includes the following steps:
[0108] Step 1: The subject wears the electromyography acquisition device correctly and makes a variety of different gestures without electrode deviation. The staff records the electromyography signals and classifies the signals according to gestures. There are a total of c types of gestures that are recognized as non-deviation samples.
[0109] In this embodiment, the myoelectric acquisition device uses a gesture-controlled armband (Myo Armband). In this embodiment, c=7; the myoelectric data used in this embodiment comes from the SeNic (sEMG in non-ideal conditions) data set.
[0110] Step 2: Preprocess and extract features of the EMG data set, and obtain one or more EMG data matrices for each set of EMG signals. The feature extraction method uses time domain-autoregressive features. Time domain-autoregressive features include root mean square, mean absolute value, zero crossing rate, slope change rate, waveform length, and autoregressive coefficient.
[0111] Step 3: Use the unbiased samples as training sets to train a classifier C Sp At the same time, the K-means clustering method is used to cluster the samples of each class in the training set, and p prototype points are obtained for each class.
[0112] In this embodiment, the trained classifier is a support vector machine.
[0113] Step 4: The objective function of constructing the electromyographic offset mapping model is as follows:
[0114]
[0115] In the formula, is the sample trust matrix, and is a diagonal matrix, its i-th element is f i ∈[0, 1] is the confidence level of the i-th sample, and f is the confidence level of all f i The column vector composed of are the original image and the image respectively, d is the feature dimension, and n is the total number of samples; are the mapping matrix and the stability feature discovery matrix respectively, where Θ is a diagonal matrix and its i-th element is θ i , corresponding to the weight of the i-th feature of the sample; is the offset vector.
[0116] Step 5: When the subject needs to start myoelectric gesture recognition, he wears the myoelectric acquisition device (at this time, the wearing position is not strictly required, and it is considered that there is an offset in wearing), and makes a variety of different gestures to obtain l labeled samples X with offset. l There are labeled samples X l Each sample in the same category selects the prototype point with the closest Euclidean distance as the prototype point. l There will be labeled samples X l and Bishop T l Substitute into the EMG offset mapping model, let the sample trust matrix F = I, the original image S = X l , like T = T l , the objective function is iteratively updated to obtain the mapping matrix A, offset vector b and stability feature discovery matrix Θ of the labeled samples.
[0117] The specific process of iterative solution is as follows: First, initialize the parameters, A = I, b = 1 d , θ=1d / d, Θ=diag(θ), and then solve them separately.
[0118] The specific process of iterative solution of b is as follows:
[0119] Fix other parameters and update b. The formula with respect to b is equivalent to:
[0120]
[0121] Take the partial derivative of the formula with respect to b and set the derivative to 0, simplifying it to get:
[0122]
[0123] Where M = (AθS-T)F, So the solution is:
[0124]
[0125] The specific process of iterative solution of A is as follows:
[0126] Fix other parameters and update A. The formula about A is equivalent to:
[0127]
[0128] Take the partial derivative of A and set the derivative to 0, and simplify it to get:
[0129] A=(βI-QP T )(PP T +βI) -1 (36)
[0130] Where P = ΘSF,
[0131] The specific process of iteratively solving Θ in the formula is as follows:
[0132] Fix other parameters and update θ. The formula about θ is equivalent to:
[0133]
[0134] The formula is simplified to:
[0135]
[0136] Let P = SFF T S T , R = SFF T (1 n b T -T T )A, the formula is equivalent to:
[0137]
[0138] Since Θ is a diagonal matrix, then is the Hadamard product operator. Therefore, the formula can be transformed into:
[0139]
[0140] set up c=-2diag, then:
[0141]
[0142] Introducing the auxiliary variable v, the formula becomes:
[0143]
[0144] The corresponding augmented Lagrangian method is:
[0145]
[0146] stθ T =1,θ≥0 (43)
[0147] λ is the Lagrange multiplier and μ is the augmentation parameter.
[0148] First, fix θ and update v. Take the partial derivative of v and set the derivative equal to 0 to get:
[0149]
[0150] Then fix v i Update f i , the formula is converted to:
[0151]
[0152] According to the perfect square formula of the matrix, it is transformed into:
[0153]
[0154] remember The above formula can be transformed into:
[0155]
[0156] The Lagrange equation is:
[0157]
[0158] δ and η are Lagrange multipliers. Assume θ * is the optimal solution for θ, δ * and η * is the corresponding multiplier. According to the KKT condition,
[0159] for have
[0160]
[0161] The first equation in the formula can be rewritten as:
[0162] θ * -m-δ * 1-η * =0 (50)
[0163] Because θ T 1=1, so the above formula becomes:
[0164]
[0165] According to the formula and we can get:
[0166]
[0167] set up The formula can be rewritten as:
[0168] θ * =g+η * -η * 1 (53)
[0169] That is, any t∈[1,d] has:
[0170]
[0171] According to the formula and, we can know (a) + =max(a,0). So we can solve:
[0172]
[0173] If the optimal solution Determine, then the optimal solution θ * can also be determined by the above formula. Similarly, the formula can be rewritten as so So It can be calculated by the following formula:
[0174]
[0175] According to the restriction θ T 1=1 and, we define the function:
[0176]
[0177] when When It can be calculated by Newton's method
[0178]
[0179] In summary, θ can be optimized to be θ=diag(θ).
[0180] Step 6: Perform electromyographic gesture recognition. The electromyographic acquisition device collects the electromyographic signal of the subject and processes it in the same way as in step 2 as the unmarked sample X with offset. u , the number of samples is u.
[0181] Step 7: Set the first l elements of the sample trust matrix F to constants α is the EMG offset mapping model for labeled samples X l The trust degree of , α∈[0,1] is an adjustable parameter.
[0182] Step 8: Use the parameters A, b, Θ obtained in step 5 to map the labeled sample X l and unlabeled samples X u ; The mapping strategy is X′ l and X′ u They are the labeled samples and unlabeled samples after mapping, respectively.
[0183] Step 9: Map the unlabeled sample X′ u Input the classifier C trained in step 3 Sp , get the unlabeled sample X u Temporary mark.
[0184] Step 10: According to the mapped unlabeled sample X′ u The temporary mark selects the prototype point with the closest Euclidean distance among the prototype points of the same category as the prototype point T u .
[0185] Step 11: Update the last u elements of the sample trust matrix F according to the trust update strategy; the last u elements of the updated sample trust matrix F are used as the unlabeled sample X u degree of trust.
[0186] The strategy for updating F is as follows:
[0187] f i is the i-th element of F, representing the credibility of the i-th sample label, and the calculation formula is as follows:
[0188] f i =ψ(dist2-dist1) (59)
[0189] Where dist1 is si With the corresponding prototype d i The distance, s i The distance to the nearest prototype in other categories is dist2, and ψ(c) is defined as follows:
[0190]
[0191] θ, τ are adjustable parameters.
[0192] Step 12: Map the labeled samples and unlabeled samples [X′ l ; X′ u ] as the original image S, and the set of prototype points in the prototype point set that is taken as the image by any labeled sample or unlabeled sample is taken as the image T = [T l ; T u ], and use the same method as in step 5 to solve the equation to obtain the new mapping matrix A, offset vector b, and stability feature discovery matrix Θ.
[0193] Step 13: Use the mapping matrix A, the offset vector b and the stability feature discovery matrix Θ to map the unlabeled samples again and update the mapped unlabeled samples
[0194] Step 14: Repeat steps 9-13 until the number of iterations reaches the upper limit, and obtain the final mapping matrix A, stability feature discovery factor θ, offset vector b and mapped unlabeled sample X′ u .
[0195] Step 15: The unlabeled sample X′ obtained in step 14 u Input the classifier C trained in step 3 Sp , and obtain more accurate offset sample recognition results.
[0196] After the first round of unlabeled samples are identified, follow-up identification is performed in either of the following two ways:
[0197] Method 1: When the worn electromyographic acquisition device does not move relative to the human body, the measured electromyographic signals are mapped using the mapping matrix A, offset vector b and stability feature discovery matrix Θ obtained in step 14 and then input into the classifier trained in step 3 to continuously obtain new electromyographic gesture recognition results.
[0198] Method 1 helps to reduce the amount of calculation for subsequent detection and achieve rapid recognition of gestures.
[0199] Method 2: When the worn electromyographic acquisition device does not move relative to the human body, each time a new electromyographic signal is measured, the new electromyographic signal is added to the unlabeled sample, and steps 8-13 are repeated to further iteratively update the mapping matrix A, the stability feature discovery factor θ, and the offset vector b. The newly measured electromyographic signal is mapped using the updated mapping matrix A, the offset vector b, and the stability feature discovery matrix Θ. The mapping result is input into the classifier trained in step 3 to continuously obtain new electromyographic gesture recognition results.
[0200] Method 2 continuously optimizes the mapping matrix A, the offset vector b and the stability feature discovery matrix Θ according to the newly measured electromyographic data, which helps to continuously improve the accuracy of gesture recognition.
[0201] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.
Claims
1. A transfer learning method for myoelectric gesture recognition based on stability feature mining, Features: The following steps are involved: Step 1: The subject wears the electromyographic acquisition device without deviation and makes a variety of different gestures. The collected electromyographic signals are preprocessed and feature extracted as non-deviation samples; Step 2: Use the unshifted samples as the training set to train the classifier; cluster the samples of each class in the training set separately, and obtain p prototype points for each class; Step 3: Establish an EMG offset mapping model embedded with the stability feature discovery factor θ; the objective function of constructing the EMG offset mapping model is as follows: In formula (1), F is the sample trust matrix; S, T are the original image and the image respectively; A, Θ are the mapping matrix and the stability feature discovery matrix respectively; I is the identity matrix; b is the offset vector; β and γ are adjustable parameters; Step 4: The subject wears the electromyography acquisition device and performs initial offset calibration. The process of initial offset calibration is as follows: the subject makes a variety of different gestures to obtain labeled samples X l ; Each labeled sample selects the prototype point with the closest Euclidean distance among the prototype points of the same category as the prototype point T l ; There will be labeled samples X l and Elephant T l As the original image S and image T respectively, the objective function is iteratively updated to obtain the mapping matrix A, offset vector b and stability feature discovery matrix Θ of the labeled samples; Step 5: The subject makes a gesture to be recognized, and the electromyographic acquisition device acquires the electromyographic signal of the subject; The obtained electromyographic signal is used as the unlabeled sample X after preprocessing and feature extraction. u ; Step 6: Use parameters A, b, Θ to map the labeled sample X l and unlabeled samples X u , and iteratively update the parameters A, b, Θ based on the mapping results; The specific process is as follows: (1) The sample trust matrix F is a diagonal matrix, and its i-th diagonal element is f i is the trust level of the i-th sample; the first l diagonal elements of the sample trust matrix F are set to constants α is the EMG offset mapping model for labeled samples X l The confidence level, α is an adjustable parameter; l is the number of labeled samples; (2) Use parameters A, b, Θ to map labeled samples X l and unlabeled samples X u ; After mapping, the sample [X′ l ; X′ u ]=AΘ[X l ;X u ]+b1 T , X′ l and X′ u They are the labeled samples and unlabeled samples after mapping respectively; (3) The mapped unlabeled sample X′ u Input the classifier trained in step 3 to obtain the mapped unlabeled sample X′ u Temporary labeling; According to the unlabeled sample X′ after mapping u The temporary mark selects the prototype point with the closest Euclidean distance among the prototype points of the same category as the prototype point T u ; (4) According to the trust update strategy, update the last u elements of the sample trust matrix F; u is the number of labeled samples; The trust updating strategy is: Construct the confidence level f of the unlabeled samples i The update expression is as follows: f i =ψ(dist 2 -dist 1 ) (29) Among them, dist 1 For i and the corresponding prototype point d i The distance, s i The distance to the nearest prototype point in other categories is dist 2 ; The function ψ(c) is defined as follows: Among them, θ, τ are adjustable parameters; c is the function independent variable; (5) The mapped sample [X′ l ; X′ u ] as the original image S, and replace the image [T l ; T u ] as the image T; iteratively update the objective function to obtain the updated mapping matrix A, offset vector b and stability feature discovery matrix Θ; (6) Use the updated mapping matrix A, offset vector b and stability feature discovery matrix Θ to map the unlabeled samples again and update the mapped unlabeled samples X′ u =AΘX u +b1 T ; (7) Iterate steps (2)-(6) multiple times to obtain the final mapping matrix A, stability feature discovery factor θ, and offset vector b; Step 7: Use the parameters A, b, Θ updated in step 6 to map the unlabeled sample X u ; The mapped unlabeled sample X′ u Input the classifier trained in step 2 to obtain the electromyographic gesture recognition result.
2. According to the transfer learning method for myoelectric gesture recognition based on stability feature mining according to claim 1, Features: The stability feature discovery matrix θ is a diagonal matrix, and its i-th diagonal element is θ i , corresponding to the weight of the i-th feature of the sample.
3. According to the method for transfer learning of stability feature mining in claim 1, Features: When the worn electromyographic acquisition device does not move relative to the human body, the measured electromyographic signals are mapped using the mapping matrix A, offset vector b and stability feature discovery matrix Θ obtained in step 6 and then input into the classifier trained in step 2 to continuously obtain new electromyographic gesture recognition results.
4. According to claim 1, a transfer learning electromyographic gesture recognition method for discovering stability features, Features: When the worn electromyographic acquisition device does not move relative to the human body, each time a new electromyographic signal is measured, the new electromyographic signal is added to the unlabeled sample, and step 6 is repeated to further iteratively update the mapping matrix A, the stability feature discovery factor θ, and the offset vector b. The newly measured electromyographic signal is mapped using the updated mapping matrix A, the offset vector b, and the stability feature discovery matrix Θ. The mapping result is input into the classifier trained in step 2 to continuously obtain new electromyographic gesture recognition results.
5. According to claim 1, a transfer learning electromyographic gesture recognition method for stability feature mining, Features: The feature extraction method in step 1 uses time domain-autoregressive features, including root mean square, mean absolute value, zero crossing rate, slope change rate, waveform length, and autoregressive coefficient.
6. According to claim 1, a transfer learning electromyographic gesture recognition method for discovering stability features, Features: In step 2, the trained classifier uses a support vector machine; at the same time, the K-means clustering method is used to cluster the samples of each class in the training set.
7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, Features: The memory stores a computer program; the processor executes the transfer learning electromyographic gesture recognition method for stability feature mining as described in any one of claims 1 to 6.
8. A readable storage medium storing a computer program; Features: When the computer program is executed by a processor, it is used to implement the transfer learning electromyographic gesture recognition method for stability feature mining as described in any one of claims 1 to 6.
