Upper limb motion angle estimation method based on local L2-constrained nonnegative matrix factorization

Abstract

This invention relates to a method for estimating upper limb motion angles based on local L2-constrained nonnegative matrix factorization (NMFLC). The proposed NMFLC algorithm addresses the problems of insufficient decoupling and unstable decoupling effects in estimating angles of continuous multi-degree-of-freedom motions. First, electromyographic (EMG) signals are preprocessed and features are extracted. Then, NMFLC decomposition is used to obtain the cofactor matrix. Activation coefficients are obtained using nonnegative least squares (NLS). Finally, the estimated angle for a given degree of freedom is obtained through a linear combination of the positive and negative activation coefficients. The proposed method achieves better suppression of inactive degrees of freedom, fully decouples single-degree-of-freedom movements, and provides a more stable decoupling effect, thus improving the accuracy of angle estimation for continuous multi-degree-of-freedom motions to some extent.

Description

Technical Field

[0001] This invention relates to a method for estimating multi-degree-of-freedom motion angles of the upper limbs based on local L2-constrained non-negative matrix decomposition (NMFLC). A NMFLC algorithm is proposed for estimating multi-degree-of-freedom motion angles of the human upper limbs. NMFLC is used to decompose the electromyographic signals of joint motion, extracting coercive elements and activation coefficients. Then, two activation coefficients related to the same degree of freedom are linearly combined to obtain the angle of that degree of freedom. This invention belongs to the field of pattern recognition model technology. Background Technology

[0002] Surface electromyography (sEMG) is generated by the electrical activity of active muscle fibers during muscle contraction. Control commands from the central nervous system are transmitted as electrical potentials within neurons, ultimately reaching motor neurons connected to the spinal cord. When limb control commands reach motor neurons, they continue to be transmitted as electrical potentials, eventually causing changes in the electrical potentials of muscle fibers in the motor unit, resulting in muscle contraction. Motor neurons and the muscle fibers they directly control constitute the smallest unit of motor control: the motor unit. When different motor units are activated sequentially, the potentials from multiple motor units are superimposed on the skin surface to form sEMG.

[0003] sEMG (sense electromyography) is a physiological electrical signal directly related to human movement. Changes in electromyographic signals can be detected by placing electrodes on the skin surface, a simple and convenient acquisition process. By analyzing sEMG, it is possible to estimate human movement intentions. Acquiring upper limb sEMG can help people with disabilities achieve flexible control of prostheses. Human-computer interaction platforms based on sEMG are also a major direction for next-generation human-computer interaction. By analyzing sEMG, human control intentions can be estimated, thereby enabling remote synchronous control of robots.

[0004] The earliest electromyographic (EMG) control of upper limb movements employed a switching control method based on sEMG amplitude. Movement was generated when the EMG amplitude exceeded a control threshold. While this discrete direct control system achieved a high success rate, it could only control one or two degrees of freedom at a time, failing to control smaller amplitudes. Furthermore, the EMG control threshold required multiple experiments to determine, resulting in poor real-time performance. Currently, upper limb movement angle estimation mainly falls into two categories: discrete action recognition and continuous motion estimation. Discrete action recognition uses supervised learning methods, extracting features from sEMG and training a classifier offline using samples. The model then identifies the action type of new samples. Although this method has high accuracy, it can only predict a few discrete actions and cannot achieve continuous, smooth, synchronous control. Therefore, using sEMG to achieve continuous proportional synchronous control of multi-degree-of-freedom movements has received widespread attention. Zhang et al. used a sparse pseudo-input Gaussian process regression method to map sEMG features and hand movements, achieving high estimation accuracy in both offline and online estimation of multi-degree-of-freedom angles. This is a supervised learning method that requires collecting joint angles to train the classifier offline. Although there was no significant difference in results between ipsilateral and contralateral training strategies, it is still not applicable to bilateral amputees. Jiang et al., based on muscle synergy theory, used Nonnegative Matrix Factorization (NMF) to achieve proportional synchronization control of two degrees of freedom movements in the wrist. They adopted a DOF-wise strategy, activating only one degree of freedom at a time and extracting its control information using the NMF algorithm. This is a quasi-unsupervised learning method, requiring only knowledge of which degree of freedom is activated in a single-degree-of-freedom movement. However, under single-degree-of-freedom movements, NMF's suppression effect on non-target degrees of freedom is weak, resulting in poor decoupling between different degrees of freedom and inaccurate estimation of non-target degree-of-freedom angles. To improve the suppression effect of NMF on non-target degrees of freedom when a single degree of freedom is activated, Lin et al. added an L1 norm regularization term for the coefficients to the NMF objective function. This sparsity constraint limited the space of possible solutions for NMF, improving the suppression effect on non-target degrees of freedom when a single degree of freedom is activated. However, this method adds constraints on the activation coefficients of all degrees of freedom, and also has a certain inhibitory effect on the target degree of freedom. In addition, when performing a single degree of freedom action, people may activate multiple degrees of freedom. Excessive inhibition of non-target degrees of freedom is not conducive to sufficient decoupling, and the decoupling robustness of this method is not strong.

[0005] The above methods are ineffective at estimating non-target degrees of freedom when decomposing single-degree-of-freedom electromyographic signals, and their decoupling of single-degree-of-freedom movements is insufficient and unstable. To address these issues, this invention proposes a locally L2-constrained nonnegative matrix factorization algorithm to decompose electromyographic signals and then estimate joint motion angles. Summary of the Invention

[0006] The purpose of this invention is to address the unsatisfactory results of existing methods in estimating the angles of non-target degrees of freedom movements of users. This invention proposes a local L2-constrained nonnegative matrix factorization (NMFLC) method to decompose human electromyographic signals, achieving accurate estimation of human joint movement angles. This enables the estimation of angles for continuous multi-degree-of-freedom movements of the human body. While DOF-wise methods require the acquisition of electromyographic signals for each degree of freedom movement individually, non-target degrees of freedom also involve some movement during a user's single-degree-of-freedom movement. This can lead to insufficient decoupling between different degrees of freedom, resulting in inaccurate estimation of the angles of non-target degrees of freedom.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A method for estimating upper limb motion angles based on local L2-constrained nonnegative matrix factorization includes the following steps:

[0009] Step 1: Acquisition of electromyographic signals of the upper limb. The user completes single-degree-of-freedom and multi-degree-of-freedom movements of the hand in sequence. The Trigno electromyographic acquisition system is used to evenly place 7 wireless electrodes near the elbow joint at a frequency of 2kHz to acquire human electromyographic signals.

[0010] Step 2: Preprocessing and feature extraction of electromyography (EMG) signals. The acquired raw EMG signals are sequentially subjected to high-pass filtering, full-wave rectification, and low-pass filtering. Then, non-overlapping time windows are used to segment the EMG signals, and the temporal features of the EMG signals within the time window are extracted.

[0011] Step 3: Decompose the electromyographic feature signal to obtain the coercive matrix. Use NMFLC to decompose the electromyographic feature signal into the product of coercive elements and activation coefficients. Use the coercive matrix to estimate the angle of joint degrees of freedom in human movement.

[0012] Step 4: Obtain the activation coefficients of the test set and estimate the angles of different degrees of freedom in joint motion.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This invention proposes the NMFLC method to decompose the electromyographic features of human upper limb joint movements and obtain the joint motion angles through a linear combination of the decomposed activation coefficients. NMFLC, by adding appropriate L2 constraints to the objective function, better decouples single-degree-of-freedom movements, obtaining the cofactors corresponding to each degree of freedom, thereby achieving accurate estimation of multi-degree-of-freedom motion angles of the upper limbs. Furthermore, this method is more robust and provides more stable joint angle estimation results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the upper limb motion angle estimation method based on local L2-constrained nonnegative matrix decomposition according to the present invention;

[0017] Figure 2 This is a graph showing the average signal-to-noise ratio and significance analysis of the upper limb multi-degree-of-freedom based on local L2-constrained nonnegative matrix decomposition in this invention.

[0018] Figure 3 The figures show the estimation results of different methods of the upper limb motion angle estimation method based on local L2 constrained nonnegative matrix decomposition of the present invention.

[0019] Figure 4 This is a comparison chart of the standard deviations of ASNR for the upper limb motion angle estimation method based on local L2-constrained nonnegative matrix decomposition of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The following is a detailed explanation of each step.

[0022] A method for estimating upper limb motion angles based on local L2-constrained nonnegative matrix factorization includes the following steps:

[0023] Step 1: This invention mainly estimates the single-degree-of-freedom and multi-degree-of-freedom simultaneous movements of the wrist and metacarpophalangeal joints of the human upper limb. Single-degree-of-freedom movements include wrist flexion and extension movements and hand-holding and reaching movements, while multi-degree-of-freedom movements are the simultaneous movement of two degrees of freedom: wrist flexion and extension, and hand opening and closing.

[0024] Using the Delsys Trigno electromyography (EMG) acquisition system, seven wireless electrodes are evenly placed on the forearm near the elbow joint in the first third. This system can acquire signals from muscles such as the extensor carpi radialis longus, extensor carpi ulnaris, and flexor digitorum superficialis. The electrodes are parallel to the direction of the muscle fibers. The system has a sampling frequency of 2 kHz and outputs signals after passing through a 20–450 Hz filter.

[0025] Step 2: Perform high-pass filtering, rectification, and low-pass filtering on the original electromyography (EMG) signal in sequence to obtain the envelope signal of the original EMG signal. The high-pass filter and low-pass filter are Butterworth filters with cutoff frequencies of 25Hz and 4Hz, respectively.

[0026] Electromyography (EMG) signal characteristics can be categorized into time-domain characteristics, frequency-domain characteristics, and time-frequency-domain characteristics. Time-domain characteristic analysis treats the EMG signal as a signal with a mean of 0; the magnitude of the signal variance determines the size of the time-domain characteristic. Commonly used time-domain characteristics include root mean square (RMS), mean square value (MSV), and absolute mean (MAV). Frequency-domain characteristic analysis reflects the changes in the EMG signal with frequency; common frequency-domain characteristics include the number of zeros crossed (ZC) and median frequency (MPF). Time-frequency-domain characteristics can reflect both the changes in the EMG signal over time and reveal frequency information; common time-frequency-domain analysis methods include short-time Fourier transform and wavelet transform.

[0027] Temporal features are closely related to time, can reflect the real-time state of muscles, and are easy to extract; therefore, temporal features were chosen. A non-overlapping time window of 100ms was used to segment the envelope signal, and finally, the root mean square (RMS) features of the signal within the time window were extracted.

[0028] Step 3: Assume the electromyographic signal feature matrix is ​​Z M×T M and T represent the number of acquisition channels and the number of samples, respectively. The electromyographic state can be approximated by the co-element matrix W. M×2N and its activation coefficient matrix F 2N×T The product of:

[0029]

[0030] W i +(-) Let F be the i-th coordinator. The size of each element represents the contribution of different muscles to this coordinator. N is the number of degrees of freedom involved in the movement. In this method, N=2. The superscripts "+" and "-" represent the positive and negative directions under the same degree of freedom, respectively. i +(-) It is W i +(-) The activation vector represents the activation level of the i-th cooperation element;

[0031] This method estimates single-degree-of-freedom and multi-degree-of-freedom movements of the wrist and metacarpophalangeal joints. The electromyographic feature matrix can be represented as follows:

[0032]

[0033] In the formula, F 1j + ( -) (j=1,2,3,4) and F 2j+(-) (j=1,2,3,4) represent the activation coefficients for DOF1 and DOF2, respectively. In single-degree-of-freedom motion, F ij +(-) (i = j = 1, 2) are the activation coefficients of the target degrees of freedom, F ij +(-) (i≠j; i,j=1,2) are the activation coefficients of non-target degrees of freedom;

[0034] The NMFLC method is used to decompose the electromyographic feature matrix, and its objective function is:

[0035]

[0036] In the formula, λ is the regularization coefficient, which can be obtained through cross-validation; C is a constraint matrix of the same size as F. The size of C is determined before training begins. The elements in C corresponding to the activation coefficients of non-target degrees of freedom are 1, and the other elements are 0.

[0037] The expression can be solved using a multiplication iteration method.

[0038] W = W.*(ZF) T ). / (WFF T (4)

[0039] F = F.*(W) T Z). / (W T WF+2λC.*F) (5)

[0040] The collaborative matrix W will be used for angle estimation on the test set.

[0041] Step Four:

[0042] The activation coefficients of the electromyographic signals in the test set were obtained using the non-negative least squares method.

[0043]

[0044] In the formula, To test the electromyographic feature matrix, which includes the time-domain features of two single-degree-of-freedom movements and a multi-degree-of-freedom movement, W is the co-element matrix obtained by equation (5). is the activation coefficient of the test set.

[0045] The motion angle estimate for a given degree of freedom is obtained by linearly combining two activation coefficients under the same degree of freedom.

[0046]

[0047] τ ij The value of (i,j=1,2) is adjusted according to the range of motion angles.

[0048] Four-fold cross-validation was performed on each collected data set to determine the regularization parameter λ. Coeromes were extracted using only EMG data from single-DOF movements in the training set, and these coeromes were used to predict angles for both single-DOF and multi-DOF movements in the test set. To validate the algorithm's performance, NMFLC was compared with existing methods NMF and SCNMF. For the regularization methods SCNMF and NMFLC, values ​​were taken in intervals from 0.001 to 10 (0.001, 0.01, 0.05, 0.1, 0.5, 1, 5, 10), and the average performance of different parameter values ​​across all test sets was compared to obtain the optimal regularization parameter value.

[0049] The degree of suppression of non-target degrees of freedom is an important indicator for evaluating the performance of algorithms in single-degree-of-freedom movements. The average signal-to-noise ratio (ASNR) is widely used for offline evaluation in single-degree-of-freedom movements. Since humans tend to grip each other during wrist flexion and extension, the actual angle is not zero. To represent the algorithm's ability to suppress noise from inactive coefficients, the denominator of the ASNR is changed to the absolute value of the difference between the predicted angle and the actual angle.

[0050]

[0051] In the formula, i represents the i-th single-degree-of-freedom action, (T i ,T i+1 Let F be the time of the i-th single-degree-of-freedom action. act (t), F inact (t) represent the estimated angles of the target degrees of freedom and the non-target degrees of freedom, respectively, H inact (t) represents the actual angle of the non-target degree of freedom.

[0052] ASNR was used to evaluate the decoupling effect of different methods. The experimental results are as follows: Figure 2 As shown. Comparing the average ASNR of the three algorithms, NMFLC performed best (2.57±0.43), followed by SCNMF (2.24±0.44), and NMF (2.23±0.46) performed worst. A two-way ANOVA was performed on the experimental results, showing a significant interaction between the two (p<0.001). Therefore, a one-way ANOVA was conducted for each factor. NMFLC and NMF showed significant differences in performance, with NMFLC outperforming NMF, and NMFLC also showing a significant improvement over SCNMF.

[0053] Figure 3 The decoupling effects of the three methods, NMF, SCNMF, and NMFLC, are shown. DOF1 represents the wrist flexion and extension degrees of freedom, and DOF2 represents the handshake and reach degrees of freedom. The average ASNRs of the three methods on the four test sets are 1.71, 1.95, and 2.38, respectively.

[0054] Since nonnegative matrix factorization algorithms are difficult to obtain the global optimal solution, three methods were used to repeat the training and prediction process 10 times for each test set. The ASNR of each test set was calculated and the standard deviation (STD) was calculated to evaluate the robustness of the algorithm.

[0055] Repeated training and prediction will cause fluctuations in ASNR results. Figure 4 The robustness of the three methods in decoupling during single-degree-of-freedom motion is shown. The mean standard deviations of NMF, SCNMF, and NMFLC across the 20 test sets are 0.22, 0.24, and 0.03, respectively. NMFLC exhibits the smallest mean standard deviation across the 20 test sets, demonstrating the best robustness in single-degree-of-freedom motion decoding.

[0056] In summary, the NMFLC method proposed in this invention has a better suppression effect on the activation coefficients of non-target degrees of freedom, thereby better decoupling the cooperating elements of different degrees of freedom and achieving accurate estimation of multi-degree-of-freedom motion angles. Furthermore, the NMFLC method exhibits the smallest standard deviation of ASNR results in different experiments, demonstrating more stable decoupling performance. The local L2-constrained nonnegative matrix factorization algorithm proposed in this invention solves, to some extent, the coupling problem of different degrees of freedom in multi-degree-of-freedom motion angle estimation, improving the accuracy of upper limb multi-degree-of-freedom motion angle estimation. The embodiments of this invention have been described in detail above with reference to the accompanying drawings, but this invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations of these embodiments without departing from the principles and spirit of this invention still fall within the protection scope of this invention.

Claims

1. A method for estimating upper limb motion angles based on local L2-constrained nonnegative matrix decomposition, characterized in that: Includes the following steps: Step 1: Acquisition of electromyographic signals of the upper limbs. The user completes single-degree-of-freedom and multi-degree-of-freedom movements of the hand in sequence. The Trigno electromyographic acquisition system is used to collect human electromyographic signals by evenly placing multiple wireless electrodes near the elbow joint. Step 2: Preprocess and extract features from the raw electromyography (EMG) signals. The raw EMG signals are subjected to high-pass filtering, full-wave rectification, and low-pass filtering in sequence. Then, the EMG signals are segmented using non-overlapping time windows, and the temporal features of the EMG signals within the time window are extracted. Step 3: Decompose the electromyographic feature signal to obtain the coercive matrix. Use NMFLC to decompose the electromyographic feature signal into the product of coercive elements and activation coefficients. Use the coercive matrix to estimate the angle of joint degrees of freedom in human movement. Step three specifically includes: Let the electromyographic signal feature matrix be Z. M×T M and T represent the number of acquisition channels and the number of samples, respectively. The electromyographic state is approximated by a co-element matrix W. M×2N and its activation coefficient matrix F 2N×T The product of: In equation (1), W i +(-) Let F be the i-th coordinator. The size of each element represents the contribution of different muscles to this coordinator. N is the number of degrees of freedom involved in the movement. In this method, N=2. The superscripts "+" and "-" represent the positive and negative directions under the same degree of freedom, respectively. i +(-) It is W i +(-) The activation vector represents the activation level of the i-th cooperation element; The electromyographic feature matrix is ​​represented as follows In equation (2), F 1j +(-) (j=1,2,3,4) and F 2j +(-) (j=1,2,3,4) are the activation coefficients corresponding to DOF1 and DOF2, respectively. In single-degree-of-freedom motion, F ij +(-) (i = j = 1, 2) are the activation coefficients of the target degrees of freedom, F ij +(-) (i≠j; i,j=1,2) are the activation coefficients of non-target degrees of freedom; The NMFLC method is used to decompose the electromyographic feature matrix, and its objective function is: s.t.W M×2N ≥0,F 2N×T ≥0 (3)(3) In equation (3), λ is the regularization coefficient, which is obtained through cross-validation; C is a constraint matrix of the same size as F. The size of C is determined before training begins. The elements in C corresponding to the activation coefficients of non-target degrees of freedom are 1, and the other elements are 0. Solve using a multiplication iteration method: W=W.*(ZF T ). / (WFF T ) (4) F=F.*(W T Z). / (W T WF+2λC.*F) (5) The collaborative matrix W will be used for angle estimation on the test set; Step 4: Obtain the activation coefficients of the test set and estimate the angles of different degrees of freedom during joint movement; Step four specifically includes: The activation coefficients of the electromyographic signals in the test set were obtained using the non-negative least squares method. In equation (6), To test the electromyographic feature matrix, which includes the time-domain features of two single-degree-of-freedom movements and a multi-degree-of-freedom movement, W is the co-element matrix obtained by equation (5). The activation coefficients for the test set. The motion angle estimate for a given degree of freedom is obtained by linearly combining two activation coefficients under the same degree of freedom. τ ij The value of (i,j=1,2) is adjusted according to the range of motion angles; Change the denominator of the average signal-to-noise ratio to the absolute value of the difference between the predicted angle and the actual angle; In equation (8), i represents the i-th single-degree-of-freedom action, (T i ,T i+1 Let F be the time of the i-th single-degree-of-freedom action. act (t), F inact (t) represent the estimated angles of the target degrees of freedom and the non-target degrees of freedom, respectively, H inact (t) represents the actual angle of the non-target degree of freedom.

2. The method for estimating upper limb motion angles based on local L2-constrained nonnegative matrix decomposition according to claim 1, characterized in that: The single-degree-of-freedom movements in step one include wrist flexion and extension movements and hand-shaking and reaching movements, while the multi-degree-of-freedom movements include simultaneous movements of two degrees of freedom: wrist flexion and extension, and hand opening and closing.

3. The method for estimating upper limb motion angles based on local L2-constrained nonnegative matrix decomposition according to claim 2, characterized in that: In step one, seven wireless electrodes are evenly placed on the forearm near the elbow joint in the first third to collect data from the extensor carpi radialis longus, extensor carpi ulnaris, and flexor digitorum superficialis muscles. The electrode direction is parallel to the direction of the muscle fibers. The system sampling frequency is 2kHz, and the output signal is filtered through a 20-450Hz filter.

Images