Anti-pose interference wrist electromyography pattern recognition calibration method based on transfer learning

Abstract

The present application belongs to the technical field of myoelectric pattern recognition, and particularly relates to a wrist myoelectric pattern recognition calibration method based on transfer learning and resisting gesture interference.The method comprises the following steps: collecting eight-channel myoelectric signals at the wrist by using a wrist myoelectric collection device, obtaining a characteristic matrix composed of four continuous characteristic values through window processing on the collected continuous sequence; using a prototype classifier for wrist myoelectric pattern recognition; using an unsupervised model of transfer learning for calibration, combining a target function for minimizing the characteristic distribution distance and a target function for minimizing the empirical error of the prototype classifier to obtain an optimization problem; and finally selecting pseudo-labels through the structural information of the source domain data with labels and the target domain data without labels themselves.The present application does not need to train the user in different postures in advance, thereby reducing the training burden of the user; the recognition accuracy is effectively improved without retraining the classifier; and the present application can be used for controlling a prosthetic hand in different postures.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of myoelectric pattern recognition, and particularly relates to a wrist myoelectric pattern recognition calibration method resisting gesture interference. BACKGROUND

[0002] Surface myoelectric signal is a superimposed electric signal formed by the electric signal generated by muscle contraction through muscle fiber to the skin surface, which reflects the contraction state of the moving muscle and can decode the individual's movement intention. The signal has good application prospect in human-computer interaction system and wearable device due to its simple and easy collection. In recent years, in the field of myoelectric signal intention recognition, most researches focus on improving the myoelectric recognition accuracy [2]-[4]. In practical application of myoelectric signal to wearable device, the user's movement intention is recognized through myoelectric signal analysis and modeling, so as to control the neural prosthesis, exoskeleton or other remote terminal [5], and many interference factors need to be considered. These interference factors will cause the characteristics of surface myoelectric signal to change, thereby affecting the use effect of wearable device [6].

[0003] The interference of the change of the user's limb posture in daily activities will cause the amplitude and frequency of the myoelectric signal collected by the wearable device to change, and the pattern recognition model trained by a small amount of data set will have limited generalization ability, and the performance will be significantly reduced when it is applied to new signals under different postures [7]. In actual use scene, the recognition accuracy is reduced due to the change of posture, and retraining the model by obtaining data under different postures of the user will increase the training cost, which is an important reason hindering the commercial application of myoelectric human-computer interaction system and wearable device.

[0004] The current solution is to first use the data under part of the posture scene as the training set to train the basic model, and then update the model through additional data for self-adaptive calibration. Compared with retraining the model under different scenes, the calibration through a small amount of additional data has small calculation cost, faster operation speed, more flexibility and is matched with the actual application scene. The calibration strategy is to find the relationship between the myoelectric signal characteristics under different postures of the same gesture, so the myoelectric data and gesture label under other postures are needed in the calibration process, that is, supervised calibration. Mainly includes:

[0005] Scheme one: an adaptive sparse representation classifier, each class is represented by a dictionary, for new data, the vector can be reconstructed by linear combination of the dictionary, and the solution of the optimization problem is found by sparse constraint, and finally the minimum residual of the vector and each class vector is calculated as the prediction result [8].

[0006] Scheme two: This scheme proposes that the motion intention of the new arm posture can be represented by the mixture of its corresponding similar postures. Therefore, this method uses ensemble learning to add the motion intentions of different postures with the likelihood of arm posture as the weight to obtain the prediction result[9].

[0007] Scheme three: This method proposes a linear supervised transfer learning by expectation maximization, which uses a small amount of labeled target domain data to learn the linear mapping matrix H from the target domain to the source domain, so that the distribution of the transformed target domain data in the labeled Gaussian mixture model in the source domain space has maximum likelihood. This algorithm can be combined with the learning vector quantization method, and only a small amount of new data is needed to significantly improve the performance of the classifier, and the performance is significantly improved compared to training a new classifier model

[10] .

[0008] Among all the model calibration methods, a representative method is an adaptive multi-modal control strategy proposed in 2021 to reduce the influence of limb position in myoelectric pattern recognition

[11] . This method first performs data fusion, fusing the myoelectric signal and IMU data to train a supervised LDA (linear discriminant analysis model) classifier as a reference model. The sensitivity and accuracy of the classifier under each combination of arm position and gesture are evaluated, and the representative combinations are found and expanded during daily use. Since the expanded data set does not contain labels, in order to reduce the spread of false labels, a majority voting strategy is used in the post-processing process, and the expanded database is used to retrain the classifier to obtain an adaptive model.

[0009] Based on the above technology, the main defects of existing myoelectric prostheses are as follows:

[0010] Defect 1: Requires labeled data of other limb postures for supervised calibration. The above scheme only needs a small amount of labeled data under new postures, which greatly reduces the training cost compared to retraining a new model, but still requires data and labels under other postures to be obtained in advance, increasing the user's additional training burden and the long configuration time before each use;

[0011] Defect 2: Limited to posture scenarios, several specific posture scenarios are set in advance, and the linear combination of several postures is not effective for completely new postures, lacking flexibility. The model does not update in real time with the user's experience during use;

[0012] Defect 3: High computational cost and high algorithm complexity. The ensemble learning method needs to train multiple models using multiple posture data and then derive the result based on the likelihood. SUMMARY

[0013] The present application aims to provide a gesture scene-adaptive, low-complexity algorithm, high-precision gesture interference-resistant wrist electromyography pattern recognition calibration method based on transfer learning.

[0014] The gesture interference-resistant wrist electromyography pattern recognition calibration method based on transfer learning provided by the present application is based on transfer learning, and specifically includes two parts: electromyography signal feature extraction and pattern recognition; and unsupervised calibration based on transfer learning.

[0015] Step (one) wrist electromyography signal feature extraction and wrist electromyography pattern recognition

[0016] The wrist electromyography signal feature extraction is performed by using a wrist electromyography acquisition device to collect eight-channel electromyography signals at the wrist at a certain frequency (equal to or greater than 1000 Hz, for example, 1000-1200 Hz, and the sampling rate is preferably 1000 Hz) to obtain continuous time series electromyography data of different gestures of the same subject in four states of sitting, standing (hand raised to shoulder level), walking and lying down, as shown in FIG. Figure 2 The electromyography data obtained in the sitting state is used as a training set, and the continuous time series are windowed with a certain window length (for example, 0.1s-1s, and the preferred window length is 0.5s) and a certain step (for example, 0.05s-0.2s, and the preferred step is 0.1s) from the training set electromyography data, and four feature values

[12] of the continuous samples in each window are calculated, which are the root mean square (RMS) of the electromyography signal, the variance, the absolute value of the summation of the exp power of the ASM (Absolute Value of Summation of exproot) signal, and the value of the Teager-Kaiser energy operator after the log of the Ltkeo (Log Teager Kaiser Energy Operator) signal

[13] , and the four feature values of the continuous samples form a feature matrix; further, the obtained feature matrix is subjected to feature Z-Score standardization and L2 norm normalization based on the samples, so as to convert the features of different magnitudes into the same magnitude, facilitating subsequent classification calculation and conversion matrix calculation of the transfer part.

[0017] The different gestures described herein are typically the following five gestures: wrist inversion, wrist eversion, index finger extension, OK, and thumb and index finger double-click, as shown in FIG. Figure 3

[0018] ​The wrist myoelectric pattern recognition (wrist myoelectric pattern refers to the myoelectric pattern generated by the wrist when making different gesture actions) adopts a prototype classifier

[14] , which is a parameter-free classifier. The prototype classifier mainly maintains a prototype for each category to represent the category, and then classifies through prototype matching. The classification result is determined according to the distance between the sample and the representative vector of each category prototype. Among them, the class center is used as the prototype of each category, and a training set is given The training set has |D| samples belonging to C different categories. The class center μ c of the c-th category is defined as:

[0019]

[0020] Among them, n c =|D c |, D c ={X i :X i ∈D,y(X i )=c}。x i represents the feature matrix of the i-th sample, y i represents the label value of the i-th sample; in the c-th category, c=1,2, …, C;

[0021] The conditional probability of a given sample x belonging to category y is expressed as follows:

[0022]

[0023] For a sample x i ∈c, the conditional distribution probability to different category prototypes is:

[0024]

[0025] The true distribution of the sample is a C-dimensional vector [0,0,…,1,…,0], where the c-th dimension of the category to which the sample belongs is 1, and the other dimensions are 0. In order to reduce the classification error, it is hoped that for a sample x i ∈c, it is close to the distance of the c-th class center while far away from the distance of other C-1 class centers. Therefore, the objective function of the classifier is obtained as:

[0026]

[0027] Among them, is the average of other class centers of x i , and is the total number of samples of other categories, and β is the regularization term coefficient.

[0028] Step (two) unsupervised calibration based on transfer learning

[0029] Unsupervised model calibration based on transfer learning is divided into three sub-steps: minimizing feature distribution distance, minimizing prototype classifier empirical error and pseudo-label selective propagation. The process is shown in Figure 1 .

[0030] (2.1) Minimizing feature distribution distance

[0031] The electromyographic data in a certain posture is taken as the source domain data (source domain) and marked as wherein represents the feature vector of a certain signal sample, i∈{1, 2,..., N S}, and its label is known as Ys. A small amount of electromyographic data in other postures is taken as the target domain (target domain) and marked as D T ={x i}, wherein N f =FxC is the dimension of the feature vector; N S and N T are the number of samples in the source domain and the target domain. The data matrices of the source domain and the target domain are marked as and Each row of the matrix represents the feature vector of a certain data sample. In order to reduce the shift of the feature distribution between the source domain and the target domain, the maximum mean discrepancy (MMD) is used as the distance measure between the two distributions

[15] . It is assumed that there is a feature conversion matrix P, which makes the MMD distance between the edge distribution and the conditional distribution of the mapped two domain data minimum. The edge distribution MMD distance between the source domain DS and the target domain DT is represented as:

[0032]

[0033] The above formula is equivalent to:

[0034] D(D S ,D T )=tr(P T XM0X T P), (6)

[0035] wherein, P is a feature conversion matrix, X is the data after splicing the source domain and the target domain, M0 is the MMD matrix:

[0036]

[0037] A prototype classifier is trained using the source domain data, and the pseudo-label of the target domain data is obtained using the classifier, which is then used to calculate the MMD distance of the conditional distribution, which is represented as:

[0038]

[0039] wherein,

[0040]

[0041] In formula (9), n c is the number of samples belonging to the c-th class in the source domain samples, n t is the number of samples belonging to the c-th class in the target domain samples.

[0042] Combining the two distribution distances shown in formula (5) and formula (8), the objective function for minimizing the feature distribution distance is obtained as:

[0043]

[0044] wherein,

[0045] (2.2) Minimizing prototype classifier empirical error

[0046] The pattern recognition classifier selected by the present application is a prototype classifier. Unlike some traditional classifiers which perform pattern recognition by calculating the division boundary. The classifier uses the distance of the sample as a classification means, without the need for additional parameter solving. For the source domain labeled data, the prototype of each class is represented by the class center, and the classification error is represented by the distance of the sample to the belonging class. The distance of the sample to the belonging class center is minimized, and the distance of the sample to the other class center is maximized.

[0047] For the target domain data, due to the lack of labels, the present application first generates the pseudo label of the target domain data by using the classifier of the source domain, and then calculates the class center of each class of the target domain data as the prototype representation of the target domain according to the pseudo label, to obtain the target domain classifier. Combining the source domain classification error and the target domain classification error, the error objective function is obtained as:

[0048]

[0049] After transformation, the above formula is equivalent to:

[0050]

[0051] wherein,

[0052]

[0053]

[0054] wherein, n s,c represents the number of samples belonging to the c-th class in the source domain, represents the number of samples not belonging to the c-th class in the source domain. nt,c The number of samples belonging to class c in the target domain. This represents the number of samples in the target domain that do not belong to class c. β is the regularization coefficient.

[0055] In addition, considering that if the source domain data and the target domain data satisfy the same distribution, the prototype classifiers obtained from the source domain data and the target domain data should be as similar as possible, that is, the distance between the same class centers of the two domains should be minimized

[16] . Therefore, we also include the cross-domain classifier error as part of the objective function. That is, we put the source domain data into the target domain classifier and the target domain data into the source domain classifier to minimize the classification error. The objective function of the error is expressed as:

[0056]

[0057] Where, ε s and ε t These are the classification errors of the source domain classifier and the target domain classifier, respectively.

[0058]

[0059] Combining (2.1) and (2.2), the objective functions of minimizing the distribution distance and minimizing the classification error are combined to obtain the optimization problem:

[0060]

[0061] st P T XHX T P = I, (17)

[0062] in,

[0063] I is the central matrix.

[0064] The parameter λ is the feature distribution distance error term parameter. This parameter determines the importance of the feature distribution errors of the source and target domain data in the total error function. The larger the value of λ, the more important the feature distribution error term is. The default value is 1.

[0065] The parameter β is the regularization term of the classifier error. This parameter determines the importance of the distance from the sample to other class centers in the classifier error. The default value is 1.

[0066] The parameter γ is the regularization term parameter for the cross-domain classifier error. Its function is similar to that of the parameter β, determining the importance of the distance term between other class centers of cross-domain samples in the classifier error. The default value is 1.

[0067] The constrained optimization problem above is solved by the Lagrange multiplier method, and the characteristic transformation matrix P is obtained.

[0068] (2.3) Pseudo label selective propagation

[0069] Since pseudo labels are used in the operation of (2.1) and (2.2), and the wrong pseudo label will cause the accumulation of errors in the iteration process, therefore, when calculating with pseudo label, it is necessary to select the pseudo label with high confidence for propagation. In this application, the pseudo label is selected by the structured information of the labeled source domain data and unlabeled target domain data itself. The pseudo label is generated by the source domain classifier, but the pseudo label obtained by the source domain classifier is not necessarily reliable, therefore, the application combines the target domain data to select the pseudo label. First, the class center μ s,c of the source domain classifier is used to classify the target domain data, and the conditional probability of the sample x belonging to the y-th class is defined as:

[0070]

[0071] For the target domain data, the class center μ t,c of the target domain data is calculated by the unsupervised clustering algorithm K-Means, and the initialized value is the source domain class center. Similarly, the conditional probability of the sample x belonging to the y-th class in the target domain is defined as:

[0072]

[0073] According to the two conditional probabilities of formula (17) and (18), two pseudo labels are obtained. The selection strategy is to select the data whose two pseudo label prediction results are consistent first, and to sort them according to the prediction probability from large to small, and to select the pseudo label with high prediction probability preferentially. The final prediction label is the dynamic combination of the two conditional probabilities, and the weight factor is used to adjust the proportion of the two probabilities in the iteration process. For the sample x t , the final class conditional probability P(y|x t ) and the pseudo label are expressed as:

[0074]

[0075]

[0076] Where t is the current iteration number, and T is the total iteration number.

[0077] According to the anti-pose interference wrist electromyography pattern recognition calibration method based on the migration learning, the application further provides an anti-pose interference wrist electromyography pattern recognition calibration system based on the migration learning, which specifically comprises two modules: a wrist electromyography signal feature extraction module and a wrist electromyography pattern recognition module, and an unsupervised calibration module based on the migration learning; the two modules respectively perform the operation contents of the two steps of the wrist electromyography pattern recognition calibration method.

[0078] Characteristics of the application:

[0079] (1) The application first proposes to combine the migration learning technology, start from the correlation of the feature spaces of the electromyography signals under different postures, utilize the characteristics of the prototype classifier, minimize the distribution difference of the electromyography signal features under different postures, and minimize the difference of the class centers (i.e., the classification functions) of the source domain and the target domain, so as to calibrate the wrist electromyography pattern recognition under different postures;

[0080] (2) The application adopts selective propagation of pseudo-labels of the target domain data to obtain the conversion matrices of the electromyography signal features under two different postures, so as to calibrate the data under different postures, and greatly improve the recognition accuracy without retraining a new classifier.

[0081] Application scenarios of the application:

[0082] The application proposes the anti-pose interference wrist electromyography pattern recognition calibration method based on the migration learning, which can be applied to the control of a prosthetic hand under different postures, acquires gesture action labels through wrist electromyography pattern recognition to control the prosthetic hand to complete corresponding gesture actions, and improves the robustness of the control of the prosthetic hand.

[0083] Advantages of the application:

[0084] (1) The application solves the problem of the decrease of the accuracy of the pattern recognition based on electromyography signals caused by the changes of the electromyography signals under different postures by combining the migration learning method;

[0085] (2) The application is unsupervised calibration, and the calibration process does not need to acquire the real labels of the data under other postures, i.e., the user does not need to be trained in different postures in advance, the training burden of the user is reduced, and the model can be automatically calibrated in the background as long as it is detected that the user is in a different motion state;

[0086] (3) The method for calculating the general feature conversion matrix under different postures proposed by the application only needs a small amount of calculation when the posture scene is initialized, and has no limitation on the posture scene and can flexibly add posture scenes. In the use process, the conversion matrix can be updated in the background in a timely manner as the data is updated, the performance of the system is gradually improved in the use process, and the application is suitable for online recognition of real-time systems. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 Calibration flowchart of unsupervised model based on transfer learning of the present application.

[0088] Figure 2 Wrist electromyography signal collection experiment diagram for 60 healthy subjects.

[0089] Figure 3 Five gesture action diagrams in the experiment.

[0090] Figure 4 Influence of different schemes on gesture recognition accuracy of different postures. DETAILED DESCRIPTION

[0091] In the verification experiment, the wrist electromyography signals (sampling rate 1000 Hz) of 60 healthy subjects in four different postures (sitting, standing, walking, and lying down) were collected, and the experimental collection is shown in Figure 2 The experimental data contain five gestures in four postures, as shown in Figure 3 , which are wrist inversion, wrist eversion, finger extension, OK, and thumb and finger double-click. The experimental goal is to verify that the proposed calibration method can improve the accuracy of the model trained in a single posture on electromyography data in other postures, and to verify the effectiveness of the proposed method.

[0092] The electromyography signals in the sitting posture were used as the training set, and the window length was 0.5s and the step was 0.1s. The four feature values of the electromyography signals in each window were calculated, which were the root mean square value, the variance, the absolute value of the power sum, and the log value of the Teager-Kaiser energy. The four feature values of the continuous samples formed a feature matrix. Further, the obtained feature matrix was subjected to feature Z-Score standardization and L2 norm normalization based on the sample. Then, a traditional linear discriminant classification algorithm (LDA, Linear Discriminant Analysis) was used to train the classifier. Then, the data in the other three postures were used as the test set, and the normalized feature matrix was obtained according to the above process. Then, the classifier was tested respectively, and the recognition accuracy of the classifier in different postures was obtained as the baseline. In different posture scenarios, the number of samples is the same, and in this experiment, the sample number ratio of the training set and the test set is 1:1.

[0093] Then, the method proposed in the present application was used, and the electromyography features in the sitting posture were used as the source domain in the method. The electromyography features in the standing, walking, and lying down postures were used as the target domain in the method. The source domain prototype classifier was obtained by calculating the class center from the source domain data and the label. Then, the pseudo-label of the target domain data was obtained using the source domain prototype classifier, and the distribution alignment matrix M and the prototype classifier error minimization matrix Q Y , And matrix in cross-domain error minimization and Let the parameters λ, γ, β are all default values 1, get The eigenvalue decomposition of Ω gets the conversion matrix P; the conversion matrix P is used for calibrating the electromyographic features in different postures; the conditional probability is calculated for the calibrated data, and reliable pseudo labels are selected for the next iteration; after the iteration is completed, the final conversion matrix P and the predicted labels of the target domain are output. According to the predicted labels, the classification accuracy after calibration is calculated, and in this experiment, the number ratio of source domain and target domain samples is 1:1.

[0094] In addition, in order to further verify the universality of the conversion matrix, 40% of the target domain samples are selected as a verification set and do not participate in the calibration process, as new data, the conversion matrix of the calibration process is used for conversion and input into the classifier to obtain the test accuracy In this experiment, the number ratio of source domain samples, target domain calibration samples and target domain test samples is 5:3:2. The experimental results show that the overall average accuracy is improved by about 10% after the data in different postures are calibrated by the present application. Figure 4 The average result of the data of 60 healthy subjects.

[0095] Figure 4 The influence of using the calibration algorithm on the gesture recognition accuracy in different postures. Among them, in the three different posture scenes of lying, standing and walking, for each posture scene, the leftmost column Baseline accuracy is the test accuracy obtained by taking the sitting posture scene as the training set, using the linear discriminant classification algorithm (LDA, Linear Discriminant Analysis) to train the classifier, and then taking the data in other posture scenes as the test set (the number ratio of training samples to test samples is 1:1); the middle column is the accuracy of the classifier after calibration using the calibration scheme proposed in the present application, taking the sitting posture data as the source domain and the other posture data as the target domain (the number ratio of source domain samples to target domain samples is 1:1); the rightmost column is the accuracy of the test obtained by applying the conversion matrix to the new target domain data after using the present application to obtain the conversion matrix with part of the other posture data as the target domain (the number ratio of source domain samples to target domain calibration samples to target domain test samples is 5:3:2). According to the accuracy results in the figure, the recognition accuracy of the wrist electromyographic data in standing, lying and walking postures is improved by more than 10% after calibration using the present application compared with the recognition accuracy using the LDA classifier. In addition, although the recognition accuracy decreases after reducing the calibration data by 40%, it is still about 10% higher than the accuracy of the LDA classifier.

[0096] REFERENCES

[0097] [1] Dai C, Hu X. Extracting and Classifying Spatial Muscle Activation Patterns in Forearm Flexor Muscles Using High-Density Electromyogram Recordings. Int J Neural Syst. 2019 Feb;29(1):1850025. doi: 10.1142 / S0129065718500259. Epub 2018 Jun 3. PMID: 29954235.

[0098] [2] He J, Jiang N. Biometric from surface electromyogram (sEMG): Feasibility of user verification and identification based on gesture recognition [J]. Frontiers in bioengineering and biotechnology, 2020, 8:58.

[0099] [3] Dai C, Hu X. Finger joint angle estimation based on motoneuron discharge activities [J]. IEEE journal of biomedical and health informatics, 2019, 24(3):760-767.

[0100] [4] Dai C, Zhu Z, Martinez-Luna C, et al. Two degrees of freedom, dynamic, hand-wrist EMG-force using a minimum number of electrodes [J]. Journal of Electromyography and Kinesiology, 2019, 47:10-18.

[0101] [5] Du Y, Jin W, Wei W, et al. Surface EMG-based inter-session gesture recognition enhanced by deep domain adaptation [J]. Sensors, 2017, 17(3): 458.

[0102] [6] Bi L, Guan C. A review on EMG-based motor intention prediction of continuous human upper limb motion for human-robot collaboration [J]. Biomedical Signal Processing and Control, 2019, 51: 113-127.

[0103] [7] Park K H, Suk H I, Lee S W. Position-independent decoding of movement intention for proportional myoelectric interfaces [J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2015, 24(9): 928-939.

[0104] [8] Betthauser J L, Hunt C L, Osborn L E, et al. Limb position tolerant pattern recognition for myoelectric prosthesis control with adaptive sparse representations from extreme learning [J]. IEEE Transactions on Biomedical Engineering, 2017, 65(4): 770-778.

[0105] [9] Park K H, Suk H I, Lee S W. Position-independent decoding of movement intention for proportional myoelectric interfaces [J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2015, 24(9): 928-939.

[0106]

[10] Paaβen B, Schulz A, Hahne J, et al. Expectation maximization transfer learning and its application for bionic hand prostheses [J]. Neurocomputing, 2018, 298: 122-133.

[0107]

[11] Spieker V, Ganguly A, Haddadin S, et al. An Adaptive Multi-Modal Control Strategy to Attenuate the Limb Position Effect in Myoelectric Pattern Recognition [J]. Sensors, 2021, 21(21): 7404.

[0108]

[12] Phinyomark A, Phukpattaranont P, Limsakul C. Feature reduction and selection for EMG signal classification [J]. Expert systems with applications, 2012, 39(8): 7420-7431.

[0109]

[13] Solnik S, Rider P, Steinweg K, et al. Teager–Kaiser energy operator signal conditioning improves EMG onset detection [J]. European journal of applied physiology, 2010, 110: 489-498.

[0110]

[14] Graf, A. B. A., Bousquet, O., Ratsch, G., Scholkopf, B.: Prototype classification: insights from machine learning. Neural computation 211, 272-300 (2009)

[0111]

[15] Long M, Wang J, Ding G, et al. Transfer feature learning with joint distribution adaptation [C] / / Proceedings of the IEEE international conference on computer vision. 2013:2200-2207.

[0112]

[16] Du Y, Chen Y, Cui F, et al. Cross-domain error minimization for unsupervised domain adaptation [C] / / Database Systems for Advanced Applications: 26th International Conference, DASFAA 2021, Taipei, Taiwan, April 11-14, 2021, Proceedings, Part II 26. Springer International Publishing, 2021:429-448.

Claims

1. A method for anti-pose interference wrist myoelectric pattern recognition calibration based on transfer learning, characterized in that, Specifically includes two steps: wrist electromyography signal feature extraction and wrist electromyography pattern recognition; Unsupervised calibration based on transfer learning; Wherein: Step one, the wrist electromyography signal feature extraction and wrist electromyography pattern recognition: The electromyography signal feature extraction is to use the wrist electromyography collection device to collect eight-channel electromyography signals at the wrist at a frequency of 1000-1200Hzd, to obtain continuous time series electromyography data of different gestures of the same subject in four states of sitting, standing, walking and lying down; The electromyography data obtained in the sitting state is used as the training set; In the training set electromyography data, the continuous time series is windowed with a window length of 0.1s-1s and a step length of 0.05s-0.2s, and four feature values of each window are calculated, which are: the root mean square value of the electromyography signal, the variance, the absolute value of the power sum, and the log value of the Teager-Kaiser energy; The four feature values of the continuous samples form a feature matrix; The obtained feature matrix is subjected to feature Z-Score standardization and L2 norm normalization based on the sample, so as to convert the feature data of different orders of magnitude into the same order of magnitude, facilitating subsequent classification calculation and conversion matrix calculation of the transfer part; Wrist myoelectric pattern recognition, using a prototype classifier, which is a parameter-free classifier, mainly to maintain a prototype for each category to represent the category, and then through the prototype matching classification, according to the distance between the sample to each class prototype representative vector to determine the classification result; wherein, the class center is used as the prototype of each category, and a training set is given The training set has |D| samples belonging to C different categories; the class center μ c is defined as: Where, n c =|D c |,D c ={X i :X i ∈D,y(X i )=c},x i Let y represent the feature matrix of the i-th sample. i Let represent the label value of the i-th sample, where c = 1, 2, ..., C in the c-th class. For a given sample x, the conditional probability that it belongs to class y is expressed as follows: For a sample x i ∈c, its conditional distribution probability to different class prototypes is: The true distribution of the sample is a C-dimensional vector [0, 0, …, 1, …, 0], where the cth dimension of the class to which the sample belongs is 1, and the other dimensions are 0; in order to reduce the classification error, it is hoped that for a sample x i ∈c, it is close to the center of the cth class and far from the centers of the other C-1 classes; thus, the objective function of the classifier is obtained as follows: wherein is x i the average of other class centers, is the total number of samples of other class, and β is a regularization term coefficient. Step two, unsupervised model calibration based on transfer learning, divided into three sub-steps: minimum feature distribution distance sub-step, minimum prototype classifier empirical error sub-step and pseudo-label selective propagation sub-step; (2.1) Minimum feature distribution distance The electromyographic data in a certain posture is taken as source domain data, and is marked as wherein represents a feature vector of a certain signal sample, i is in {1, 2,..., N S}, and the label thereof is known as Ys; a small amount of electromyographic signals in other postures are taken as target domains, and are marked as wherein N f =FxC is the feature vector dimension; N S and N T are the sample numbers of the source domain and the target domain; the data matrices of the source domain and the target domain are respectively marked as and each row of the matrix represents a feature vector of a certain data sample; in order to reduce the deviation of the feature distribution between the source domain and the target domain, the maximum mean difference MMD is taken as the distance measurement between the two distributions; it is assumed that there is a feature conversion matrix P, so that the MMD distance of the edge distribution and the conditional distribution of the mapped two domain data is minimum; the MMD distance of the edge distribution between the source domain DS and the target domain DT is represented as: The above formula is equivalent to: D(D S ,D T )=tr(P T XM0X T P), (6) Wherein, P is a feature conversion matrix, X is the data after splicing the source domain and the target domain, M0is the MMD matrix, n s is the number of source domain samples, n t is the number of target domain samples: A prototype classifier is trained using source domain data, and pseudo-labels of target domain data are obtained using the classifier, which are then used to calculate the MMD distance of the conditional distribution, denoted as: Wherein, In formula (9), n s,c is the number of samples belonging to the cth class in the source domain samples, and n t,c is the number of samples belonging to the cth class in the target domain samples. The two distribution distances shown in formulas (5) and (8) are combined to obtain the objective function for minimizing the feature distribution distance as: wherein (2.2) Minimum prototype classifier empirical error For the labeled data of the source domain, the prototype of each class is represented by the class center, and the classification error is represented by the distance from the sample to the class to which it belongs. The distance from the sample to the class center is minimized while the distance from the sample to the class center is maximized. For target domain data, due to the lack of labels, first use the prototype classifier of the source domain to generate the pseudo-labels of the target domain data, and then calculate the class center of each class of the target domain data as the prototype representation of the target domain, to obtain the target domain prototype classifier. The source domain classification error and the target domain classification error are combined to obtain the error objective function as: After transformation, the above formula is equivalent to: Wherein, wherein n s,c represents the number of samples belonging to the c-th class in the source domain, represents the number of samples not belonging to the c-th class in the source domain; n t,c represents the number of samples belonging to the c-th class in the target domain, represents the number of samples not belonging to the c-th class in the target domain, and β is a regularization term coefficient. Considering that if the source domain data and the target domain data satisfy the same distribution, then the prototype classifiers obtained from the source domain and the target domain data should be as similar as possible, that is, the distance between the same class centers of the two domains should be minimized. Therefore, the cross-domain classifier error is also part of the objective function, that is, the source domain data is put into the target domain classifier, and the target domain data is put into the source domain classifier, so that the classification error is minimized, and the error objective function is represented as: where ε s and ε t are the classification errors of the source and target domain classifiers, respectively. Combining (2.1) and (2.2), the objective functions of minimizing distribution distance and minimizing classification error are combined to obtain an optimization problem: s.t. P T XHX T P = I, (17) wherein, I is a center matrix; The optimization problem with constraints is solved by Lagrange multiplier method to obtain a feature conversion matrix P. (2.3) Pseudo label selective propagation The pseudo labels are selected by the structured information of the labeled source domain data and the unlabeled target domain data itself; first, the class center μ s,c of the source domain classifier is used to classify the target domain data, and the conditional probability that a sample x belongs to the y-th class is defined as: For target domain data, use unsupervised clustering algorithm K-Means to iteratively calculate the class center μ of target domain data t,c initialized value is the source domain class center; similarly, the conditional probability that the sample x belongs to the y-th class in the target domain is defined as: According to the two conditional probabilities of formula (17), (18), two pseudo labels are obtained; the strategy is to first select the data whose two pseudo label prediction results are consistent, and then sort them according to the prediction probability from large to small, and preferentially select the pseudo label with high prediction probability; the final prediction label is the dynamic combination of the two conditional probabilities, and the proportion of the two probabilities is adjusted in the iteration process by using the weight factor. For sample x t , the final class conditional probability P(y|x t ) and pseudo label are represented as: Wherein, t is the current iteration number, and T is the total iteration number.

2. The method of claim 1, wherein, The different gestures are specifically the following five kinds: wrist varus, wrist valgus, stretch index finger, OK, and thumb and index finger double click.

3. The method of claim 1, wherein, The parameters λ, γ, β take value 1.

Images