An unsupervised calibration method for myoelectric prosthetic control

通过在肌电假肢控制系统中采用无监督校准方法,利用肌电特征空间相关性进行数据加权,解决了肌电信号个体差异性导致的模型校准困难问题,实现了便捷的肌电假肢控制模型校准 and performance improvements have promoted its commercial production and clinical application.

CN113952093BActive Publication Date: 2024-11-22FUDAN UNIVERSITY
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202111336305.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-12
Publication Date
2024-11-22
Estimated Expiration
2041-11-12

AI Technical Summary

Technical Problem

Due to the individual differences in electromyography signal, existing electromyography prosthetic control systems require long-term model calibration and individual adaptation for each user, limiting their application in large-scale clinical use and commercial production.

Method used

Unsupervised calibration method is adopted based on the spatial correlation of electromyography characteristics. By comparing the electromyography signal characteristics of new users with existing databases, a model of electromyography signal and prosthetic motion control of new users is established without the need for new users. motion labels, reducing calibration steps and time.

Benefits of technology

The calibration of the unsupervised electromyography prosthesis control model is realized, which reduces the steps and time of calibration before using the prosthesis, improves the convenience of model calibration, and automatically updates model parameters during use, improving system performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113952093B_ABST
    Figure CN113952093B_ABST
Patent Text Reader

Abstract

The present invention discloses an unsupervised calibration method for myoelectric prosthesis control; the present invention focuses on a proportional control system based on regression, uses a data weighting criterion based on the spatial correlation of myoelectric features to perform unsupervised calibration on the prosthesis control model, and identifies the movement intention of the amputee through the myoelectric signals generated by muscle activity, thereby performing intention control on the prosthesis. The present invention does not need to collect the movement information (joint force or torque) of the target user, and can complete the user's myoelectric signal collection during the use of the prosthesis, greatly improving the convenience of the model calibration process; the present invention can continuously collect the user's myoelectric information during use, and repeat the calibration after a specific time interval, automatically update the model parameters, and regularly update the model, so that the performance of the myoelectric prosthesis control system is gradually improved during use, greatly reducing the cost of model calibration, and promoting its commercial production and wider clinical application.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence medical devices, and in particular, relates to an unsupervised calibration method for myoelectric prosthesis control. Background Art

[0002] Decoding the electromyographic signals of residual muscles to identify the amputee's movement intention and then intuitively control the prosthesis is an important technology to help amputees and other patients with movement disorders restore their motor functions [1]–[4]. Myoelectric prosthetic control can be divided into two categories: (1) control methods based on pattern recognition (such as gesture recognition) [5], and (2) proportional control methods based on regression [6]. Among them, proportional control based on regression can not only identify discrete movement categories, but also quantitatively estimate detailed information such as the force (torque) of a specific joint. Although myoelectric prosthetic control has great application potential, some key factors that need to be resolved hinder its practical application. One of the most critical factors is that electromyographic signals have individual differences [7]. That is, for each user, it is necessary to collect his own electromyographic signal training model. When the trained model is used for a new user, the performance drops significantly. The individual differences in electromyographic signals are an important factor that limits the generalization ability of myoelectric prosthetic control models. It requires long-term model calibration and individual adaptation for each user, which has also become an important reason hindering the large-scale clinical use and commercial production of myoelectric prosthetic control systems.

[0003] Existing solutions use models trained with other user data and further calibrate them with new subject data. Compared with retraining the model, calibrating an existing model for a new user often requires a smaller amount of data, so it is more convenient. The model calibration process is often supervised calibration, that is, calibration requires both the target user subject's electromyography data and the true value label of the movement intention (joint force, torque and other information). It mainly includes:

[0004] Solution 1: A covariate shift adaptation algorithm that only uses less model calibration data to complete the model calibration process [8]. Although the above existing solutions can greatly reduce the training cost, they still need to use less data for each user to calibrate the model, and this process is a supervised calibration, that is, both the user's electromyography and the true value of its movement (such as joint force) are required. However, collecting the force required by the joint (or other movement-related true values) is an additional burden for the user, and usually requires a long period of repeated experimental configuration.

[0005] Solution 2: Use transfer learning algorithm to achieve fast myoelectric control model transfer [9]. This method uses a model calibration scheme based on least squares support vector machine. For amputees, due to incomplete limbs, some specific joints that need to estimate the applied force (torque) may not exist and therefore cannot be obtained.

[0006] Scheme 3: Among all the model calibration methods, the most representative method is a multi-user universal generalization model based on efficient model training proposed in 2020 [7]. This method first uses the EMG and joint force labels of multiple subjects to train a subject-specific model (a time-delay linear model) for the EMG data of each channel of each subject. The model fitting parameters corresponding to the EMG features in multiple time windows of a single EMG channel can form a finite impulse response (FIR) filter. These filters naturally have the frequency domain response shape of a low-pass filter

[10] –

[13] . These trained model filters are all normalized in the frequency domain response so that the response gain at 0 Hz is 1. By averaging the model filters corresponding to all subjects and EMG channels, a model filter with extremely generalized performance can be obtained, which can provide the time domain dynamic characteristics of the EMG features for as many subjects as possible. Therefore, when training a new model, it is not necessary to train the time domain dynamic characteristics within each channel, but only the gain of each channel needs to be trained (because the gain of the model filter of each channel is normalized). Training channel gains is a much easier task than training the time-domain dynamics of each channel, so it requires less data to complete. However, the model will not be updated in real time with the user's experience during use. Summary of the invention

[0007] In view of the shortcomings of the above-mentioned existing model calibration schemes, the present invention focuses on a regression-based proportional control system and provides an unsupervised calibration method for myoelectric prosthesis control, which uses a data weighting criterion based on the spatial correlation of myoelectric features to perform unsupervised calibration of the prosthesis control model. It can identify the amputee's movement intention through the myoelectric signals generated by muscle activity, thereby performing intentional control of the prosthesis.

[0008] The technical solution of the present invention is specifically described as follows.

[0009] The present invention discloses an unsupervised calibration method for myoelectric prosthesis control, which compares the myoelectric signal characteristics of a new user with those of users in an existing database, and establishes a model of the new user's myoelectric signal and prosthesis motion control without requiring a motion label of the new user, thereby greatly reducing the steps and time of calibration before using the prosthesis; the method comprises the following steps:

[0010] Step 1: Extraction of myoelectric features and construction of myoelectric prosthesis control model

[0011] First, the EMG signal in each channel is normalized to a time series with a mean of zero and a standard deviation of 1. The training data segment is normalized using its own mean and STD, and the test signal segment is normalized using the mean and standard deviation calculated from the corresponding channel of the EMG data obtained from the same subject during the test.

[0012] From the training EMG data, calculate the root mean square RMS value, zero crossing ZS and slope sign change SSC features of every 100 consecutive samples;

[0013] The extracted features are used to train a polynomial regression model with time lag that provides intra-channel temporal dynamic information. The model is described by the following formula (1):

[0014]

[0015] Where i is the sample index of the sampled data, V represents the polynomial model order, Q represents the maximum time delay, F represents the number of extracted features, C represents the number of channels, θ is the fitting parameter, and θ offset is a constant term and x is the eigenvalue; the optimal model parameters are estimated using linear least squares and rank-explicit QR factorization is used to remove feature correlation;

[0016] Step 2: Unsupervised Model Calibration

[0017] 1) Domain shift estimation

[0018] The training set MYO is regarded as the source domain data and marked as in Represents the feature vector of a signal sample, i∈{1, 2, ..., N S}, the target user's electromyographic data collected during the use of the prosthesis is regarded as the target domain and marked as in N f =F×C is the feature vector dimension, N S and N T is 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 eigenvector of a data sample;

[0019] The main goal of the domain shift estimation process is to estimate the linear transformation matrix By description and The offset between in is the source domain after linear transformation;

[0020] Source domain data and target domain data Calibrate to obtain and X T The optimal linear transformation matrix A with the highest similarity * , that is, the domain offset matrix A * ;

[0021] 2) Data weight calculation

[0022] Target domain and each single source domain The degree of domain shift between them varies with different data of different subjects. The domain shift matrix A * It is used to generate the best data weight for each non-user test data in the original source domain. In model training, the electromyographic signals that are more similar to the target domain data are assigned higher weights;

[0023] The difference between a specific source domain and a target domain is calculated based on the matrix in formula (5) To describe:

[0024]

[0025] matrix The maximum singular value λ of The 2-norm of is used to characterize the degree of domain shift between a specific source domain and a target domain. The data weight in Obtained by the following formula:

[0026]

[0027] Finally, the parameter model values ​​of each non-user subject in the original source domain trained according to formula (1) are multiplied by their respective weight values ​​obtained by (6). The models with different weights are superimposed and normalized to obtain the myoelectric prosthesis control model of the final user. The larger the λ, the greater the feature gap between the user subject and a non-user subject. Therefore, the model of the non-user subject in the source domain is not suitable for application to the user subject in the target domain. Therefore, the reciprocal of λ is taken. The larger the λ, the smaller the weight will be given to the non-user subject model, and vice versa. Among them, m is a preset hyperparameter. The best value of m is found in the study to be 3 or 4. Since m is the exponent of λ, the role of m is to exponentially amplify λ, thereby further amplifying the multiplier of the entire weight coefficient.

[0028] Preferably, in step one, the number of channels C for collecting electromyographic signals is 8, the sampling rate is 4096 Hz, the number of features F extracted is 3, and if the signal sign and the slope sign change at the sampling index n, respectively, ZC[n] and SSC[n] are assigned the value 1, otherwise they are assigned to 0, where n represents the sampling point index, and ZC[n], SSC[n] and the true torque signal are low-pass filtered to below 16 Hz.

[0029] Preferably, in step 2) by minimizing and X T The covariance matrix of and C T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

[0030]

[0031] Preferably, in step 2) by minimizing and X T The probability density function of and P T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

[0032]

[0033] Preferably, in step 2) by minimizing and X T The characteristic distance and d T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

[0034]

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] (1) The myoelectric prosthesis control model calibration scheme of the present invention is an unsupervised calibration. It does not need to collect the real motion labels (joint forces or torques) of the target user and does not need the cooperation of the user. It can complete the user's electromyographic signal collection, complete the model calibration, and update the model parameters during the use of the prosthesis, which greatly improves the convenience of the model calibration process.

[0037] (2) The present invention transfers the model trained with other user data to the new target user. The adopted model unsupervised calibration scheme adopts the idea of ​​data weighting, and gives higher weight to the training data that is more similar to the target subject's electromyographic data.

[0038] (3) Based on the advantage of unsupervised calibration, the present invention can continuously collect the user's electromyographic information during use, repeat the calibration after a specific period of time, automatically update the model parameters, and regularly update the model, so that the performance of the myoelectric prosthesis control system is gradually improved during use, which greatly reduces the cost of model calibration and can promote its commercial production and wider clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 : Flowchart of the unsupervised calibration system for myoelectric prosthetic control models.

[0040] Figure 2 : Experimental data from 65 healthy subjects, with electrodes covering the entire forearm.

[0041] Figure 3 :Amputee controls prosthesis to open door.

[0042] Figure 4 :The influence of whether to use the calibration algorithm on the control accuracy of the prosthesis motion trajectory. The blue is the target trajectory, the red and purple are the empty trajectories without the calibration algorithm, and the yellow is the trajectory with the calibration algorithm. The yellow control trajectory is obviously closer to the target trajectory and has a better effect. DETAILED DESCRIPTION

[0043] The following will introduce the unsupervised myoelectric prosthesis control model calibration scheme proposed in this patent from two aspects in combination with the accompanying drawings:

[0044] Myoelectric feature extraction and myoelectric prosthesis control model description; unsupervised model calibration method.

[0045] 1. Myoelectric feature extraction and myoelectric prosthesis control model description

[0046] The solution proposed in the invention first normalizes the electromyographic signal (EMG) in each channel (a total of 8 channels) to a time series with a mean of zero and a standard deviation (STD) of 1.

[0047] The training data segment is normalized using its own mean and STD.

[0048] The test signal segments were normalized using the mean and standard deviation calculated from the corresponding channels of the EMG data obtained from the same subject during the test.

[0049] From the training electromyographic data (sampling rate is 4096Hz), the root mean square (RMS) value, zero crossing (ZS) and slope sign change (SSC) features are calculated for every 100 consecutive samples (24.4ms). Specifically, if the signal sign and slope sign change at sampling index n, respectively, ZC[n] and SSC[n] (where n represents the sampling point index) are assigned a value of 1, otherwise they are assigned a value of 0. ZC[n], SSC[n] and true value torque signals are low-pass filtered to below 16Hz. The addition of low-pass filtering is based on the fact that joint torque varies in a low and narrow frequency band (maximum frequency is below 16Hz). The high-frequency components of the measured torque signal and electromyographic features can be regarded as noise. The present invention downsamples the torque signal and features after low-pass filtering to 40.96Hz so that one value is retained for every original 100 samples (24.4ms). The selected sampling rate of 40.96 Hz can construct a data set with a suitable bandwidth (40.96 Hz × 0.5 = 20.48 Hz), which is about 10 times that of the synchronous torque signal. In addition, the selected sampling rate is high enough to capture the system dynamics and low enough to eliminate high-frequency noise outside the torque signal spectrum.

[0050] The present invention uses the above extracted features to train a polynomial regression model with time lag (providing time domain dynamics information within the channel). The model can be described by the following formula (1):

[0051]

[0052] Where i is the sample index of the 40.96Hz downsampled data; V represents the polynomial model order (optional parameter, default V = 1, linear model); Q = 15 (366ms) represents the maximum time delay; F = 3 and C = 8 represent the number of extracted features and channels respectively; θ is the fitting parameter (θ offset is a constant term); x is the eigenvalue. The optimal model parameters are estimated using linear least squares, and rank-explicit QR factorization is used to remove feature correlation.

[0053] 2. Unsupervised Model Calibration Methods

[0054] Unsupervised model calibration can be divided into two steps: domain offset estimation and data weight calculation. The flowchart is as follows Figure 1 As shown;

[0055] 2.1 Domain Shift Estimation

[0056] The training set muscle TV is regarded as the source domain data (source domain), marked as in Represents the feature vector of a signal sample, i∈{1, 2, ..., N S The target user’s electromyographic data collected during the use of the prosthesis is regarded as the target domain and marked as in N f =F×C is the feature vector dimension; N S and N T 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 eigenvector of a data sample.

[0057] The main goal of the domain shift estimation process is to estimate the linear transformation By description and The offset between in is the source domain after linear transformation. In order to estimate the matrix A, a comparison is required and X T Since this scheme aims at unsupervised calibration of the model, the only available target domain information is the EMG signal or the extracted EMG features (without the true torque label at each moment).

[0058] Therefore, this patent proposes the following three solutions for source domain data: and target domain data To perform calibration, any of the following three options can achieve the calibration purpose:

[0059] Solution 1: Calibrate the covariance matrix

[0060] if With X T Similar, we can conclude that their covariance matrix and C T Similarly, we can minimize the covariance matrix of the two and C T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

[0061]

[0062] Option 2: Calibrate the probability density distribution

[0063] if With X T Similar, we can draw the following conclusion, that is, their probability density distribution and PT Similarly, we can minimize the probability density functions of the two and P T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

[0064]

[0065] Solution 3: Calibration feature distance

[0066] if With X T If they are similar, we can conclude that their corresponding features are similar, that is, the numerical distances of the features are close. and d T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

[0067]

[0068] After obtaining the linear transformation matrix A * Afterwards, the data weighted calibration method proposed in this patent is implemented by using the domain offset matrix A * Get the weight of each set of training data.

[0069] 2.2 Data weight calculation

[0070] Target domain and each single source domain The degree of domain shift between varies with different data from different subjects. In this scheme, the matrix A * It is used to generate the best data weight for each non-user subject data in the original source domain. In model training, EMG signals that are more similar to target domain data should be assigned higher weights. For example, if the domain offset matrix A * =I, where If is the identity matrix, then the source domain and the target domain match perfectly, and the corresponding source domain data should be assigned the highest weight. The difference between the specific source domain and the target domain can be obtained by using the matrix in formula (5) To describe:

[0071]

[0072] matrix The maximum singular value λ of The 2-norm of can be used to characterize the degree of domain shift between a specific source domain and a target domain. The data weight in It can be obtained by the following formula:

[0073]

[0074] Finally, the parameter model values ​​of each non-user subject in the original source domain trained according to formula (1) are multiplied by their respective weight values ​​obtained by (6). The models with different weights are superimposed and normalized to obtain the final user model.

[0075] In the verification experiment, the upper limb muscle electrical signals of 65 healthy subjects were collected (see the experimental figure Figure 2 ), the experimental goal is to verify that the proposed algorithm can transfer the model of 65 healthy subjects to amputees.

[0076] First, we calculated the models of arm muscle electrical signals and movements of 65 healthy subjects, with one adapted model for each subject;

[0077] Secondly, by comparing the muscle electrical signals generated by 65 healthy subjects and the residual limbs of amputees, a calibration algorithm based on feature space correlation was used (three schemes can be selected to obtain similar results), and different weight coefficients were assigned to the models of the 65 healthy subjects. 65 models with different weights were then used as the prosthetic control model for amputees.

[0078] Finally, the amputee achieved fine motor control of the upper limbs using the trained model (see the experimental figure). Figure 3 ), and the experimental results were compared with those without using the proposed unsupervised calibration model (directly using the average model of 65 subjects for amputees) (e.g. Figure 4 ), the overall effect is improved by about 10%.

[0079] The present invention proposes for the first time that starting from the correlation of the feature space of the electromyographic signal, three schemes are proposed to calculate the similarity between the data of different users and the target user data, and then obtain the data weights and perform unsupervised calibration of the model through data weighting. The scheme for calculating the data weights through the correlation of the feature space of the electromyographic signal can be various variations and fine-tuning of these three methods. For example, alternative schemes can be: directly calculating the difference of the covariance matrix of the feature set, calculating the 2-norm of the difference matrix to obtain the weights, etc. All unsupervised machine learning algorithms that use the data weighting mechanism to perform model calibration to achieve prosthetic control are within the scope of protection of this patent.

[0080] References:

[0081] [1]C.Dai,Z.Zhu,C.Martinez-Luna,T.R.Hunt,T.R.Farrell,and E.A.Clancy,“Two degrees of freedom,dynamic,hand-wrist EMG-force using a minimum numberof electrodes,”Journal of Electromyography and Kinesiology,vol.47,pp.10–18,Aug.2019,doi:10.1016 / j.jelekin.2019.04.003.

[0082] [2]E.A.Clancy,C.Martinez-Luna,M.Wartenberg,C.Dai,and T.R.Farrell,“Twodegrees of freedom quasi-static EMG-force at the wrist using a minimum numberof electrodes,”Journal of Electromyography and Kinesiology,vol.34,pp.24–36,Jun.2017,doi:10.1016 / j.jelekin.2017.03.004.

[0083] [3]W.Wei,Q.Dai,Y.Wong,Y.Hu,M.Kankanhalli,and W.Geng,“Surface-Electromyography-Based Gesture Recognition by Multi-View Deep Learning,”IEEETransactions on Biomedical Engineering,vol.66,no.10,pp.2964–2973,Oct.2019,doi:10.1109 / TBME.2019.2899222.

[0084] [4]D.C.Toledo-Pérez,J.Rodríguez-Reséndiz,R.A.Gómez-Loenzo,andJ.C.Jauregui-Correa,“Support Vector Machine-Based EMG Signal ClassificationTechniques:A Review,”Applied Sciences,vol.9,no.20,Art.no.20,Jan.2019,doi:10.3390 / app9204402.

[0085] [5]Y.Du,W.Jin,W.Wei,Y.Hu,and W.Geng,“Surface EMG-Based Inter-SessionGesture Recognition Enhanced by Deep Domain Adaptation,”Sensors,vol.17,no.3,Art.no.3,Mar.2017,doi:10.3390 / s17030458.

[0086] [6]N.Jiang,H.Rehbaum,I.Vujaklija,B.Graimann,and D.Farina,“Intuitive,Online,Simultaneous,and Proportional Myoelectric Control Over Two Degrees-of-Freedom in Upper Limb Amputees,”IEEE Transactions on Neural Systems andRehabilitation Engineering,vol.22,no.3,pp.501–510,May 2014,doi:10.1109 / TNSRE.2013.2278411.

[0087] [7]B.Bardizbanian et al.,“Efficiently Training Two-DoF Hand-WristEMG-Force Models,”in 2020 42nd Annual International Conference of the IEEEEngineering in Medicine Biology Society(EMBC),Jul.2020,pp.369–373.doi:10.1109 / EMBC44109.2020.9175675.

[0088] [8]M.M.-C.Vidovic,H.-J.Hwang,S.Amsuss,J.M.Hahne,D.Farina,and K.-R.Muller,“Improving the Robustness of Myoelectric Pattern Recognition forUpper Limb Prostheses by Covariate Shift Adaptation,”IEEE Trans.NeuralSyst.Rehabil.Eng.,vol.24,no.9,pp.961–970,Sep.2016,doi:10.1109 / TNSRE.2015.2492619.

[0089] [9]C.Prahm,B.Paassen,A.Schulz,B.Hammer,and O.Aszmann,“TransferLearning for Rapid Re-calibration of a Myoelectric Prosthesis After ElectrodeShift,”in Converging Clinical and Engineering Research on NeurorehabilitationII,Cham,2017,pp.153–157.doi:10.1007 / 978-3-319-46669-9_28.

[0090]

[10] K.Koirala,M.Dasog,P.Liu,and E.A.Clancy,“Using the Electromyogramto Anticipate Torques About the Elbow,”IEEE Transactions on Neural Systemsand Rehabilitation Engineering,vol.23,no.3,pp.396–402,May 2015,doi:10.1109 / TNSRE.2014.2331686.

[0091]

[11] V.T.Inman,H.J.Ralston,J.B.De C.M.Saunders,M.B.Bertram Feinstein,and E.W.Wright,“Relation of human electromyogram to muscular tension,”Electroencephalography and Clinical Neurophysiology,vol.4,no.2,pp.187–194,May1952,doi:10.1016 / 0013-4694(52)90008-4.

[0092]

[12] A.L.Hof and Jw.Van den Berg,“EMG to force processing I:Anelectrical analogue of the hill muscle model,”Journal of Biomechanics,vol.14,no.11,pp.747–758,Jan.1981,doi:10.1016 / 0021-9290(81)90031-2.

[0093]

[13] D.A.Winter and H.J.Yack,“EMG profiles during normal humanwalking:stride-to-stride and inter-subject variability,”Electroencephalography and Clinical Neurophysiology,vol.67,no.5,pp.402–411,Nov.1987,doi:10.1016 / 0013-4694(87)90003-4.

[0094]

[14] B.Sun,J.Feng,and K.Saenko,“Return of Frustratingly Easy DomainAdaptation,”presented at the Thirtieth AAAI Conference on ArtificialIntelligence,Mar.2016.Accessed:Aug.19,2019.[Online].Available:https: / / www.aaai.org / ocs / index.php / AAAI / AAAI16 / paper / view / 12443。

Claims

1. An unsupervised calibration method for myoelectric prosthesis control, characterized in that: By comparing the EMG signal characteristics of a new user with those of users in the existing database, a model of the new user's EMG signal and prosthetic motion control is established without the need for the new user's motion labels, thereby reducing the steps and time for calibration before using the prosthesis; The steps include: Step 1: Extraction of myoelectric features and construction of myoelectric prosthesis control model First, the EMG signal in each channel is normalized to a time series with a mean of zero and a standard deviation of 1. The training data segment is normalized using its own mean and STD, and the test signal segment is normalized using the mean and standard deviation calculated from the corresponding channel of the EMG data obtained from the same subject during the test. From the training EMG data, calculate the root mean square RMS value, zero crossing ZS and slope sign change SSC features of every 100 consecutive samples; The extracted features are used to train a polynomial regression model with time lag that provides intra-channel temporal dynamic information. The model is described by the following formula (1): Where i is the sample index of the sampled data, V represents the polynomial model order, Q represents the maximum time delay, F represents the number of extracted features, C represents the number of channels, θ is the fitting parameter, and θ offset is a constant term, and x is an eigenvalue; The optimal model parameters were estimated using linear least squares, and rank-explicit QR factorization was used to remove feature correlations; Step 2: Unsupervised Model Calibration 1) Domain Shift Estimation The training set muscle TV is regarded as the source domain data and marked as in Represents the feature vector of a signal sample, i∈{1, 2, ..., N S }, the target user's electromyographic data collected during the use of the prosthesis is regarded as the target domain and marked as in N f =F×C is the feature vector dimension, N S and N T is 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 eigenvector of a data sample; The main goal of the domain shift estimation process is to estimate the linear transformation matrix By description and The offset between in is the source domain after linear transformation; Source domain data and target domain data Calibrate to obtain and X T The best linear transformation matrix A with the highest similarity; 2) Data weight calculation Target domain and each single source domain The degree of domain shift between them varies with different data of different subjects. The domain shift matrix A * , that is, the optimal linear transformation matrix A is used to generate the optimal data weight for each non-user subject data in the original source domain. In model training, a higher weight is assigned to the electromyographic signal that is more similar to the target domain data; The difference between a specific source domain and a target domain is calculated based on the matrix in formula (5) To describe: matrix The maximum singular value λ of The 2-norm of is used to characterize the degree of domain shift between a specific source domain and a target domain. The data weight in Obtained by the following formula: Among them, m is a hyperparameter, and its value is 3 or 4; Finally, the parameter model values ​​of each non-user subject in the original source domain trained according to formula (1) are multiplied by their respective weight values ​​obtained by (6). The models with different weights are superimposed and normalized to obtain the myoelectric prosthesis control model of the final user.

2. The method according to claim 1, characterized in that: In step 1, the number of channels C for collecting electromyographic signals is 8, the sampling rate is 4096 Hz, the number of features F extracted is 3, and if the signal sign and slope sign change at sampling index n, ZC[n] and SSC[n] are assigned the value 1, otherwise they are assigned to 0, where n represents the sampling point index. ZC[n], SSC[n] and true torque signals are low-pass filtered to below 16 Hz.

3. The method according to claim 1, characterized in that In step 2, by minimizing and X T The covariance matrix of and C T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

4. The method according to claim 1, characterized in that: In step 2, by minimizing and X T The probability density function of and P T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

5. The method according to claim 1, characterized in that In step 2, by minimizing and X T The characteristic distance and d T The difference is calibrated and the optimal linear transformation matrix A is estimated. The objective function of the algorithm is as follows:

Citation Information

Patent Citations

  • Artificial limb control method and system thereof

    CN101987048A

  • Myoelectric prosthesis control source lead optimization method based on correlation coefficients

    CN102764167A

  • Upper limb prosthesis control system and method based on muscle coordination and human-machine voice interaction

    CN111616847A

  • Multifunction control of a prosthetic limb using syntactic analysis of the dynamic myoelectric signal patterns associated with the onset of muscle contraction

    EP0421780A2

  • Method and device for deducting pinch force, method and device for discriminating muscle to deduct pinch force

    KR1020100056254A