Surface electromyogram signal multi-joint angle estimation method under symmetric time sequence Mangbar

Abstract

The invention discloses a surface electromyogram signal multi-joint angle estimation method under a symmetric time sequence Mangbar, belongs to the technical field of man-machine interaction, and solves the problems of poor signal quality, insufficient time sequence context information modeling capability and limited multi-joint angle prediction expandability of the existing joint angle estimation method. The multi-joint angle estimation method is a novel end-to-end frame, signal artifact suppression, cross-time sequence representation learning and multi-joint angle regression tasks are incorporated into a unified model architecture, and synchronous high-precision estimation of multi-joint motion angles is conveniently achieved. In the preprocessing of the sEMG signal, the noise modal component is accurately removed by combining the dual constraints of the peak frequency and the energy threshold after the sEMG signal is decomposed, and the quality of the surface electromyogram signal is improved. The symmetric time sequence Mangban module completes capture of a signal bidirectional time sequence context dependency relationship by aligning forward and reverse state space characterization, and the estimation precision of a joint angle is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology, specifically to a method for estimating the multi-joint angle of surface electromyography signals under symmetric temporal Mamba. Background Technology

[0002] Surface electromyography (sEMG) signals, as a non-invasive method for decoding neuromuscular activity, have been widely used and are applicable to a wide range of scenarios, including dexterity hand control and human-computer interaction.

[0003] In joint angle estimation research, many early studies employed regression-based methods, such as linear regression, support vector regression, or multilayer perceptrons, to map surface electromyography (EMG) signal features to joint kinematic information. With the development of deep learning technology, recent studies have introduced network architectures such as convolutional neural networks, long short-term memory networks, and attention mechanism models to enhance temporal feature extraction and spatial fusion capabilities between channels. While these methods have achieved acceptable performance, they generally suffer from the following problems: First, preprocessing involves simple filtering without fully utilizing the characteristics of EMG signals to improve signal quality; second, the use of shallow or basic network architectures limits their ability to capture the long-term dependencies and temporal structures required for modeling temporal biological signals; and third, most existing studies only estimate a small number of joint angles, making it difficult to verify the model's performance and scalability in complete multi-joint scenarios.

[0004] In summary, existing joint angle estimation methods based on sEMG signals suffer from poor signal quality, insufficient ability to model temporal context information, and limited scalability of multi-joint angle prediction. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a method for estimating the multi-joint angles of surface electromyography signals under symmetric temporal Mamba, which solves the problems of poor signal quality, insufficient ability to model temporal context information, and limited scalability of multi-joint angle prediction in existing joint angle estimation methods.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba is provided, including the following steps: S1. Acquire sEMG signals from multiple muscle channels as input sequences. ; S2. Using variational mode decomposition method to... The signal is decomposed into multiple modal components, and the peak frequency and energy ratio of each modal component are calculated. All modal components that fall within a preset interval corresponding to both peak frequency and energy ratio are selected, summed, and reconstructed to obtain the denoised signal. ; S3, will Input the symmetric timing Mamba module, and then use its forward Mamba module and reverse Mamba module to respectively... Perform forward and reverse timing processing to obtain the forward representation. and reverse representation ,Will and Perform feature fusion and combine the fused features with... By performing element-wise addition and layer normalization operations, enhanced temporal features are obtained. ; S4, will Input a pre-trained joint angle prediction model to predict the angles of joints corresponding to multiple muscle channels.

[0007] Furthermore, The method for decomposing into multiple modal components is as follows: S21. Construct a variational constraint model, the expression of which is: in, For the k-th decomposition, there is the modal component; Let be the modal component of the k-th decomposition at time t; for The center frequency; This represents the total number of modal components. This is for gradient calculation; Let be the impact function; It is a unit of virtual time; It is an L2 norm; These are constraints; S22, By introducing the augmented Lagrange function This transforms the constrained variational problem of the variational constraint model into an unconstrained variational problem. The expression is: in, Punishment factor; For Lagrange operators; This is an inner product operation; Let t be the Lagrange operator at time t; S23. The unconstrained variational problem is solved by using the alternating direction method of multipliers to obtain each modal component and its corresponding center frequency.

[0008] Furthermore, the expression for calculating the energy percentage of each modal component is as follows: ; ; ; in, and These are the energy percentage and energy of the mode component in the k-th decomposition, respectively; for Total energy; The preset threshold; The time step length; It is an absolute value.

[0009] Furthermore, preset threshold The range of values ​​is .

[0010] Furthermore, the peak frequency range of each modal component is as follows: and .

[0011] Furthermore, positive representation The expression is: in, Forward Mamba module.

[0012] Furthermore, reverse representation The method for obtaining it is as follows: S31, will Perform reverse processing to obtain reverse sequence Its expression is: in, This is a reverse operation; S32, will Inputting into the reverse Mamba module yields the reverse encoded features. Its expression is: in, For reverse Mamba modules; S33, will Perform time-series realignment to obtain reverse representation .

[0013] Furthermore, enhance time-series features The expression is: ; ; in, For layer normalization operation; This is a residual enhancement feature; This is a feature of fusion.

[0014] Furthermore, the loss function of the pre-trained joint angle prediction model is: ; ; in, The loss function; and These represent the time step length and the total number of predicted joint angles, respectively. and The joints at time t are respectively The predicted angle mean and the actual angle; The joint at time t Variance of the predicted angle; The joint at time t Concentration parameters; The joint at time t Confidence level of the predicted angle; for degrees of freedom; This represents the weight of the regularization term.

[0015] Compared with the prior art, the present invention has the following significant advantages: 1. The multi-joint angle estimation method of this invention is a novel end-to-end framework that integrates signal artifact suppression, cross-temporal representation learning, and multi-joint angle regression tasks into a unified model architecture, facilitating synchronous high-precision estimation of multi-joint motion angles. Specifically, in the preprocessing of sEMG signals, variational mode decomposition (VMD) not only achieves intrinsic mode decomposition of the signal but is also commonly used for noise suppression. Based on this, this invention proposes a novel post-processing strategy, namely Variational Mode Decomposition-Energy Removal (VER). This strategy achieves precise removal of noise mode components by combining peak frequency and energy threshold constraints after sEMG signal decomposition, filtering out noise-dominated modes and improving the quality of surface electromyography signals. The Symmetric Temporal Mamba module (STMamba) captures the bidirectional temporal context dependencies of the signal by aligning the forward and reverse state space representations. This module can fully exploit the temporal dynamic features of multi-channel signals at each time step, thereby significantly improving the estimation accuracy of joint angles. This method addresses the problems of poor signal quality, insufficient ability to model temporal context information, and limited scalability of multi-joint angle prediction in existing joint angle estimation methods.

[0016] 2. This invention addresses the characteristics of surface electromyography (EMG) signal estimation of multi-joint angles, which is susceptible to noise / artifact interference and requires high-confidence prediction under smooth movements. It designs a deep evidence regression loss function for EMG adaptation: the model outputs the predicted joint angle value along with the corresponding confidence level (evidence). The loss function guides the model to adaptively adjust the confidence level according to the quality of the EMG signal, which not only improves the accuracy of angle estimation but also quantifies the uncertainty of prediction, thus meeting the safety requirements of rehabilitation medicine, exoskeleton control and other scenarios. Attached Figure Description

[0017] Figure 1 The flowchart shows a method for estimating the multi-joint angle of surface electromyography signals under symmetric temporal Mamba.

[0018] Figure 2 This is a block diagram of the symmetric timing Mamba module.

[0019] Figure 3 Comparison of the predicted trajectory and the actual trajectory for subject 1 with the five driving degrees of freedom. Detailed Implementation

[0020] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0021] Example 1 Joint angle estimation of the human hand / dexterous hand based on surface electromyography (sEMG) signals is a core technology for achieving refined and natural human-computer interaction, and also an indispensable foundational module for promoting the development and engineering implementation of embodied intelligence technology in robotic agents. However, existing methods generally suffer from problems such as poor signal quality, insufficient ability to model temporal context information, and limited scalability of multi-joint angle prediction. These shortcomings seriously hinder the implementation of high-precision motion decoding technology necessary for embodied intelligent interaction.

[0022] To address the aforementioned issues, this embodiment provides a method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba, referencing... Figure 1This method, with signal artifact suppression, multivariate temporal modeling, and end-to-end regression as its core design principles, constructs an integrated high-precision motion decoding system. The framework first designs a variational mode decomposition-energy removal (VER) filtering module. Relying on the dual criteria of peak frequency and energy threshold, it accurately filters and removes noisy modes, efficiently solving problems such as motion artifacts, baseline drift, and high-frequency interference in sEMG signals. This improves signal purity from the source, laying a solid data foundation for subsequent in-depth mining of temporal features. Simultaneously, it innovatively designs a symmetric temporal Mamba module. By aligning forward and reverse state space representations, it efficiently captures long-range dependencies in multivariate temporal data like sEMG, deeply mining dynamic temporal features across channels and time steps. This allows for accurate decoding of human motion intent without complex additional preprocessing. The framework adopts an end-to-end learning paradigm, connecting the signal preprocessing of the VER module and the temporal feature modeling of the STMamba module, two core components. It can directly and accurately estimate the synchronous angles of multiple joints from the raw sEMG signal. This overcomes the shortcomings of traditional methods in both signal quality and temporal modeling, and solves the scalability limitations of multi-joint prediction. It combines modeling accuracy with engineering deployment flexibility, better adapting to the complex real-world scenarios requiring embodied intelligent interaction. The specific steps include: S1. Acquire sEMG signals from multiple muscle channels as input sequences. ; S2. Using variational mode decomposition method to... The signal is decomposed into multiple modal components, and the peak frequency and energy ratio of each modal component are calculated. All modal components that fall within a preset range of both peak frequency and energy ratio are selected, summed, and reconstructed to obtain the denoised signal. ; S3, will Input the symmetric timing Mamba module, and then use its forward Mamba module and reverse Mamba module to respectively... Perform forward and reverse timing processing to obtain the forward representation. and reverse representation ,Will and Perform feature fusion and combine the fused features with... By performing element-wise addition and layer normalization operations, enhanced temporal features are obtained. ; S4, will Input a pre-trained joint angle prediction model to predict the angles of joints corresponding to multiple muscle channels.

[0023] Specifically, will The method for decomposing into multiple modal components is as follows: S21. Construct a variational constraint model, the expression of which is: in, For the k-th decomposition, there is the modal component; Let be the modal component of the k-th decomposition at time t; for The center frequency; This represents the total number of modal components. This is for gradient calculation; Let be the impact function; It is a unit of virtual time; It is an L2 norm; These are constraints; S22, By introducing the augmented Lagrange function This transforms the constrained variational problem of the variational constraint model into an unconstrained variational problem. The expression is: in, Punishment factor; For Lagrange operators; This is an inner product operation; Let t be the Lagrange operator at time t; S23. The constrained variational problem is solved iteratively using the alternating direction method of multipliers to obtain each modal component and its corresponding center frequency. , and The update expression and the expression for the iteration termination condition are: ; ; ; ; in, The center frequency is Fourier transform of the modal components of the k decompositions after the (n+1)th iteration; , and They are respectively The transformed signal after Fourier transform, the modal component of the i-th decomposition after Fourier transform, and the Lagrange operator after Fourier transform; The center frequency of the k-th modal component after the nth iteration; The center frequencies of the k modal components after the (n+1)th iteration; This represents the modal component of the k-th decomposition after the Fourier transform; and The center frequencies after the (n+1)th iteration and the nth iteration are respectively. The Lagrange operator after Fourier transform; The dual rate of ascent; and They are respectively and Fourier transform; and These are the modal components after the nth iteration and the (n-1)th iteration, respectively; The time step length, divided by It can normalize the error, eliminate the influence of signal length on convergence judgment, and make the convergence threshold... It has universality.

[0024] In practical processing tasks of surface electromyography (EMG) signals, variational mode decomposition (VMD) can not only achieve intrinsic mode decomposition of the signal, but is also often used for noise suppression. Based on this, this embodiment proposes a novel post-processing strategy, namely variational mode decomposition-energy removal (VER). This strategy improves the quality of surface EMG signals by filtering out noise-dominant modes after signal decomposition. Specifically, this embodiment uses the following two criteria to identify and remove low-quality modes: Mode elimination criterion based on peak frequency: For each decomposition obtained mode component The frequency domain characterization is obtained by performing a Fourier transform, and the frequency corresponding to the maximum value of the modal component amplitude spectrum is defined as the peak frequency. If the peak frequencies of the modal components satisfy... or If so, the modal component is determined to be a noise mode and is removed. The aforementioned frequency bands typically correspond to motion artifacts and high-frequency interference components in surface electromyography (EMG) signals.

[0025] Mode rejection criterion based on energy proportion: The expression for calculating the energy proportion of each mode component is as follows: ; ; ; in, and These are the energy percentage and energy of the mode component in the k-th decomposition, respectively; for Total energy; The preset threshold; It is an absolute value.

[0026] In this embodiment, a preset threshold is used. The range of values ​​is This is used to balance the relationship between signal denoising effect and effective feature preservation. After removing high-noise and low-energy modes, the retained effective modes are summed and reconstructed to obtain the denoised signal. ,in, A set of indices representing all valid modal components. VER improves the interpretability and robustness of VMD-based signal decomposition results, ensuring that the reconstructed signal can fully retain physiologically relevant valid information, while effectively suppressing unwanted artifacts.

[0027] As a further solution in this embodiment, considering the long temporal characteristics of continuous multi-joint movements (such as continuous flexion and extension movements in rehabilitation training), traditional temporal models (such as LSTM and Transformer) suffer from gradient vanishing or low computational efficiency when processing long sequences, failing to simultaneously meet the requirements of long-term dependency capture and real-time estimation, and thus struggling to satisfy the requirements of real-time control scenarios such as exoskeleton robots. Furthermore, ordinary state-space models only model unidirectionally in the direction of increasing time steps, failing to utilize the relatively stable and symmetrical characteristics of electromyographic signals and joint angles when maintaining a fixed movement, thus failing to fully extract the intrinsic relationship of the entire signal features. To fully capture the temporal dependency features in surface electromyographic sequences, this embodiment designs a Symmetric Temporal Mamba (STMamba) module, referencing... Figure 2 This module consists of two parallel Mamba sub-modules: a forward Mamba module and a reverse Mamba module. Given the extracted denoised signal... ,in The number of time steps. For feature dimensions.

[0028] Forward Mamba module acquires positive representation The expression is: in, As a forward Mamba module, this module can model the temporal relationship between shallow features from front to back, directly encode the electromyographic features of the original time sequence, focus on capturing the "signal consistency in the stream" (which is the core identifier of the movement remaining unchanged), and extract the synchronization features of multi-channel electromyography to obtain the sequential deep feature representation, thereby enhancing the ability to represent signals.

[0029] Reverse representation The method for obtaining it is as follows: S31, will Perform reverse processing to obtain reverse sequence Its expression is: in, This is a reverse operation. Here, the superficial electromyographic features within the motion-invariant window are displayed. (For stationary features) reverse the time dimension, transforming the "early to late" sequence into "late to early". For stationary signals, this step can capture subtle fluctuations in muscle tension maintenance from a reverse perspective, avoiding the omission of weak features from a single forward perspective, and laying the groundwork for subsequent feature complementation.

[0030] S32, will Inputting into the reverse Mamba module yields the reverse encoded features. Its expression is: in, This is a reverse Mamba module; using the reverse branch of STMamba, the Mamba module, to analyze the stationary characteristics after reversal. Encoding is performed. It focuses on two points: first, capturing the consistency of signals across time steps within a window (corresponding to core features that remain unchanged in action); second, extracting the synergy of multi-channel electromyographic fluctuations, filtering single-channel noise, and improving the accuracy of joint angle estimation for static movements.

[0031] S33, will Perform time-series realignment to obtain reverse representation The features after reverse encoding The time sequence is then reversed again to restore the original features. This step ensures that the features of the reverse branch are precisely aligned with those of the forward branch. Subsequent fusion is equivalent to performing a "double check" on the stable signal, significantly reducing the interference of weak noise on the estimation of static motion angles.

[0032] By performing a time reversal operation again on the output of the inverted Mamba module, the following can be achieved: and Timing alignment. The two outputs are added element-wise to obtain the fused bidirectional timing features: in, To integrate features, the positive branches The captured "forward time flow signal consistency" and the reverse branch The captured "reverse-view subtle fluctuations / multi-channel collaborative features" are fused. By complementing the information from the forward and reverse branches, the "signal consistency (the core identifier of movement stability)" and "subtle feature robustness (resistance to noise interference)" are combined to further enhance the representation ability of electromyographic signals, making subsequent multi-joint angle estimation more accurate and stable.

[0033] To enhance feature representation while fully preserving the stationary characteristics and basic temporal information of electromyographic signals, this embodiment introduces a residual connection mechanism to... Integration features with STMamba The residual characteristics are obtained by adding elements one by one. , This operation avoids losing the stable, fundamental information of the electromyographic signal itself during feature fusion, allowing the model to simultaneously capture both the enhanced features that depend on both positive and negative contexts and the core characteristics of the original signal.

[0034] Subsequently, layer normalization is performed on the residual fusion features to unify the mean and variance of the feature distribution, resulting in enhanced temporal features. : This step can eliminate the fluctuations in feature distribution caused by subtle fluctuations in electromyographic signals, thereby stabilizing and standardizing the output features; it can also alleviate gradient fluctuations during model training and improve training stability.

[0035] The temporal Mamba module is said to fully capture the forward and reverse temporal dependencies of electromyographic signals while retaining key information of the original input, providing a more robust feature basis for subsequent high-precision multi-joint angle estimation.

[0036] As a further solution in this embodiment, addressing the characteristics of surface electromyography (EMG) signal estimation of multi-joint angles—signals being susceptible to noise / artifact interference and requiring high-confidence predictions under smooth movements—a deep evidence regression loss function for EMG adaptation is designed. This function allows the model to output the predicted joint angle value along with the corresponding confidence level (evidence). The loss function guides the model to adaptively adjust the confidence level based on the quality of the EMG signal, improving angle estimation accuracy and quantifying prediction uncertainty, thus meeting the safety requirements of rehabilitation medicine, exoskeleton control, and other scenarios. The final output layer of the model needs to output two types of information simultaneously (corresponding to each time step). Dimensions of each joint Therefore, the loss function of the pre-trained joint angle prediction model is: ; ; in, The loss function; and These represent the time step length and the total number of predicted joint angles, respectively. and The joints at time t are respectively The predicted angle mean and the actual angle; The joint at time t Variance of the predicted angle; The joint at time t Concentration parameters; The joint at time t Confidence level of the predicted angle; for degrees of freedom; The weight for the regularization term is preferably 0.1, which is suitable for the high confidence requirement of stable signals.

[0037] In scenarios where electromyographic signals are stable (movement remains unchanged), the model will output a large value. ,correspond As the regularization term increases, its penalty effect weakens, and the model focuses more on optimizing the regression error term, thus outputting high-precision, high-confidence joint angle predictions. Conversely, when the electromyographic signal contains noise or artifacts, the model will proactively output smaller values. ,correspond Decreasing the regularization term strengthens its penalizing effect, causing the model to amplify the prediction variance. This helps to quantify uncertainty and avoid the adverse effects of incorrect predictions on scenarios such as rehabilitation assessment and real-time exoskeleton control.

[0038] The loss function in this embodiment is divided into three parts: a decomposition term of the negative log-likelihood of the normal distribution and an evidence regularization term for electromyography adaptation. The meaning and design rationale of each part are strongly related to the task characteristics of "estimating joint angles from electromyography signals". Part One: (Weighted squared error term) (1) Mathematical meaning: using the predicted variance Weight the squared errors of the predicted and actual angles—when The greater the uncertainty, the smaller the weight of this item; conversely, the smaller the uncertainty, the larger the weight.

[0039] (2) Reason for use: To adapt to the noise characteristics of electromyographic signals - when there are artifacts / noise in the electromyographic signals, the model will output large noise. (This indicates "this prediction is unreliable"), at which point the weight of this term is reduced to prevent the model from being "biased" by errors from noisy samples; while when the electromyographic signal is stable, The smaller the weight, the more the model focuses on optimizing accuracy.

[0040] Part Two: (Logarithmic variance constraint) (1) Mathematical meaning: This is an inherent component of the negative log-likelihood when the prediction error follows a normal distribution (this term will be decomposed after taking the negative log of the probability density function of the normal distribution).

[0041] (2) Reason for use: to prevent the model from "lazy" - if there is only the first part, the model may be scaled up indefinitely. To reduce the weight of the squared error; this term will affect the weight of "excessively large" squared errors. "Impose penalties to force the model to learn 'reasonable uncertainty' (neither too small (ignore noise) nor too large (avoid optimization)."

[0042] Part Three: (Evidence regularity) (1) Mathematical meaning: Regularization terms designed based on evidence (confidence level) — (Degrees of freedom) are based on evidence Decide( The larger, The larger the value, the more this option will penalize situations where "prediction accuracy is low but confidence is high".

[0043] (2) Reasons for use: This is the core of Deep Evidence Regression (DER), which adapts to the confidence requirements of electromyography tasks—when the electromyography signal is stable and the prediction accuracy is high, When the signal is large, the penalty for this term is weak, and the model can maintain high confidence; however, when the EMG signal noise is high and the prediction accuracy is low, The smaller the value, the stronger the penalty, forcing the model to lower its confidence level (smaller output). This allows "confidence level" to be matched with "confidence level".

[0044] When these three parts are combined, the loss function can both optimize the model's estimation accuracy of joint angles and adaptively quantify prediction uncertainty (which is crucial for tasks involving easily disturbed physiological signals such as electromyography).

[0045] In summary, the beneficial effects of this plan are as follows: We propose an end-to-end integrated framework to jointly solve the temporal dependency modeling of surface electromyography signals and the estimation of multi-joint angles, aiming to improve the estimation accuracy, robustness and scalability of human-computer interaction systems in the field of embodied intelligence.

[0046] A variational mode decomposition-energy rejection (VER) filtering method is proposed. This method relies on variational mode decomposition and energy-based mode selection strategy to effectively suppress artifact interference in surface electromyography signals and improve signal quality before feature learning.

[0047] A Symmetric Temporal Mamba (STMamba) module is constructed, which captures the forward and reverse temporal dependencies of signals based on a bidirectional state-space architecture, and achieves deep temporal modeling of signals with extremely low computational overhead.

[0048] Example 2 This embodiment is a further limitation based on Embodiment 1. Its purpose is to provide an experimental method for estimating the multi-joint angle of surface electromyography signals under symmetric temporal Mamba. Other parts not mentioned refer to Embodiment 1 or the prior art.

[0049] This embodiment uses the publicly available NinaPro DB8 dataset for experiments. This dataset contains multimodal data from 12 subjects, including 10 healthy subjects and 2 subjects with radial amputations. All subjects completed 9 types of bimanual mirror hand movements. Sixteen wireless electrodes were deployed on the right forearm of the subjects to collect surface electromyography (EMG) signals. The original sampling frequency of the signals was 1111 Hz, which was subsequently upsampled to 2000 Hz and synchronized with the kinematic data. Each subject's dataset contains three experimental phases. This embodiment follows the data partitioning scheme in the literature, merging the data from phases 1 and 2 for model training and validation, while the data from phase 3 is used as a separate test set.

[0050] The experiment used the CyberGlove II data glove worn by the subject on the contralateral left hand to collect corresponding 18-dimensional joint angle data. The original sampling frequency of this data was 100 Hz, which was also upsampled to 2000 Hz to achieve time-series matching with surface electromyography signals.

[0051] This embodiment implements a simplified data augmentation strategy based on the method described in the literature: Overlapping sliding windows are used to segment continuous surface electromyography (EMG) signals, with a window length of 2000 time steps (corresponding to 1 second) and a sliding step size of 100 time steps. Each segmented signal window is paired with the joint angle vector at the corresponding time point, serving as the input sample and label for the model, respectively. This processing method effectively improves the sample diversity of the dataset while preserving the inherent temporal correlation of EMG signals.

[0052] To ensure the stability of numerical calculations, this embodiment refers to the processing method in the prior art (Zanghieri M, Benatti S, Benini L, et al. Event-based Low-Power and Low-Latency Regression Method for Hand Kinematics from Surface EMG[C]∥2023 9th International Workshop on Advances in Sensors and Interfaces (IWASI), 2023: 293-298), and amplifies the original surface electromyography signal. The optimal result of each model on the test set is selected as its final performance level during the model evaluation phase.

[0053] In this embodiment, a penalty factor for variational mode decomposition (VMD) is set. Noise tolerance The total number of modal components During the decomposition process, the DC component is removed, and the center frequency is uniformly initialized. When the relative change in iteration is lower than... When the algorithm converges, it is determined that the algorithm has converged.

[0054] Due to individual differences among subjects, the energy parameters corresponding to different subjects vary. The optimal values ​​vary for each subject. To address this issue, this embodiment uses a grid search to determine the optimal value for each subject. Value. Subject-specific values ​​used during model training and testing phases. The possible values ​​are as follows: To verify the effectiveness of the proposed method, this embodiment selects several representative pose estimation models based on surface electromyography signals for comparative experiments. Referring to the experimental settings and model design principles of existing models, this embodiment adapts all comparison methods to conform to the characteristics of the NinaPro DB8 dataset, ensuring fair and consistent performance comparisons of all models under the direct regression task setting.

[0055] To focus on the relevant degrees of freedom required for dexterity hand control, Krasoulis et al. transformed the 18-DOF joint data collected from the data glove into 5 driving degrees of freedom through linear transformation. In this embodiment, these driving degrees of freedom are used as the prediction targets of the neural network. Detailed information about this dataset and the matrix parameters used in the linear transformation can be found in relevant literature. The OnlineTransformer model incorporates a sliding window-based attention mechanism, supports element-wise processing of sequences, and integrates spiking neurons to improve network sparsity, making it suitable for online processing of surface electromyography signals in dexterity hand control scenarios. The TEMPONet model is a convolutional neural network built on a temporal convolutional network, containing 3 convolutional modules. By searching for optimal parameters such as convolutional kernel size and stride, a balance between model performance and computational cost is achieved.

[0056] Unless otherwise specified, all experimental results in this embodiment use the mean absolute error (MAE) as the core evaluation index, and the calculation method corresponds to formula (1). The unit of all result values ​​is angle (°).

[0057] To ensure the fairness of the comparative experiments, this embodiment evaluates the model under both the standard 5-class action setting and the complete 10-class hand action setting of the dataset, thereby verifying the robustness and generalization ability of the proposed method. Referring to Table 1, the experimental results show that the mean absolute error of the method in this embodiment reaches 6.02°, which is better than all the results of existing studies and refreshes the current state-of-the-art performance for this task; in particular, compared with the second-best method, Online Transformer, the mean absolute error of the method in this embodiment is reduced by 0.06°.

[0058] Table 1. Comparison of mean absolute error and 10° / 15° threshold accuracy of different models To verify the effectiveness of the proposed method, this embodiment conducted ablation experiments and set up three model configurations for comparison: the original baseline model without any post-processing strategy, the improved baseline model with the VER module embedded, and the complete model integrating the VER module and the STMamba module. The experimental results are shown in Table 2. Compared with the original baseline model, the introduction of the VER module reduced the mean absolute error (MAE) for almost all subjects, fully verifying the module's excellent noise suppression capability. Further integration of the STMamba module on the basis of the VER module further improved the model performance, achieving the best mean absolute error across all dimensions. This embodiment also included an additional control experiment, validating the original baseline model without any improvement strategy and the model integrating only the STMamba module at all 18 joint angles. The results showed that the STMamba module reduced the mean absolute error for all subjects. These stepwise performance improvements fully validate the effectiveness of the proposed VER module and STMamba module. The complementary gain effect formed by the two significantly improves the accuracy of joint angle estimation, providing dual core support for model performance optimization.

[0059] Table 2 MAE (°) of ablation experiments at all joint angles As a further solution in this embodiment, refer to Figure 3 , Figure 3 The regression results of the driving degrees of freedom for subject number one are visualized. Figure 3It can be seen that the model's estimation accuracy varies across different driving degrees of freedom, reflecting the correlation between the difficulty of this regression task and the specific action type. Notably, the model in this embodiment consistently maintains high accuracy in predicting the latter three driving degrees of freedom—index finger flexion, middle finger flexion, and ring / little finger flexion—with good angle estimation results, and the estimated trajectories closely match the actual trajectories. This result demonstrates that even in motion scenarios involving multiple joint linkages and joint coupling, the proposed model still possesses excellent robustness and can accurately capture the motion pattern characteristics of the fingers.

[0060] In summary, this embodiment fully validates the proposed framework on the public dataset NinaPro DB8. Experimental results show that, compared with the current mainstream baseline models, this method exhibits significantly better overall performance: it effectively reduces the mean absolute error (MAE) in both full-subject and full-joint estimation tasks, and the model maintains stable and excellent performance when expanded from a subset of local hand movements to a full-dimensional movement set, fully verifying the generalization ability and scalability of the proposed method.

Claims

1. A method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba, characterized in that, Including the following steps: S1. Acquire sEMG signals from multiple muscle channels as input sequences. ; S2. Using variational mode decomposition method to... The signal is decomposed into multiple modal components, and the peak frequency and energy ratio of each modal component are calculated. All modal components that fall within a preset interval corresponding to both peak frequency and energy ratio are selected, summed, and reconstructed to obtain the denoised signal. ; S3, will Input the symmetric timing Mamba module, and then use its forward Mamba module and reverse Mamba module to respectively... Perform forward and reverse timing processing to obtain the forward representation. and reverse representation ,Will and Perform feature fusion and combine the fused features with... By performing element-wise addition and layer normalization operations, enhanced temporal features are obtained. ; S4, will Input a pre-trained joint angle prediction model to predict the angles of joints corresponding to multiple muscle channels.

2. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 1, characterized in that, Will The method for decomposing into multiple modal components is as follows: S21. Constructing a variational constraint model: in, For the k-th decomposition, there is the modal component; Let be the modal component of the k-th decomposition at time t; for The center frequency; This represents the total number of modal components. This is for gradient calculation; Let be the impact function; It is a unit of virtual time; It is an L2 norm; These are constraints; S22, By introducing the augmented Lagrange function This transforms the constrained variational problem of the variational constraint model into an unconstrained variational problem. The expression is: in, Punishment factor; For Lagrange operators; This is an inner product operation; Let t be the Lagrange operator at time t; S23. The unconstrained variational problem is solved by using the alternating direction method of multipliers to obtain each modal component and its corresponding center frequency.

3. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 2, characterized in that, The expression for calculating the energy percentage of each modal component is as follows: ； ； ； in, and These represent the energy percentage and energy of the mode component in the k-th decomposition, respectively. for Total energy; The preset threshold; The time step length; It is an absolute value.

4. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 3, characterized in that, Preset threshold The range of values ​​is .

5. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 1, characterized in that, The peak frequency range of each modal component is: and .

6. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 1, characterized in that, positive representation The expression is: in, Forward Mamba module.

7. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 6, characterized in that, Reverse representation The method for obtaining it is as follows: S31, will Perform reverse processing to obtain reverse sequence Its expression is: in, This is a reverse operation; S32, will Inputting into the reverse Mamba module yields the reverse encoded features. Its expression is: in, For reverse Mamba modules; S33, will Perform time-series realignment to obtain reverse representation .

8. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 7, characterized in that, Enhanced temporal features The expression is: ； ； in, For layer normalization operation; This is a residual enhancement feature; This is a feature of fusion.

9. The method for estimating multi-joint angles of surface electromyography signals under symmetric temporal Mamba according to claim 1, characterized in that, The loss function of the pre-trained joint angle prediction model is: ； ； in, The loss function; and These represent the time step length and the total number of predicted joint angles, respectively. and The joints at time t are respectively The predicted angle mean and the actual angle; The joint at time t Variance of the predicted angle; The joint at time t Concentration parameters; The joint at time t Confidence level of the predicted angle; for degrees of freedom; This represents the weight of the regularization term.

Images