Cross-user hand joint angle estimation method, system and computer device

Abstract

The application provides a cross-user hand joint angle estimation method, system and computer device. Hand activity signal data of multiple users is acquired, a multi-user model is constructed according to the hand activity signal data, the multi-user model is calibrated through training data of a new user, and the calibrated multi-user model estimates the finger joint angle of the new user. The method, system and computer device utilize surface electromyography signals and finger joint angle data of multiple users to establish a multi-user model, then adopt a newly proposed adversarial transfer learning strategy, use part of the training data of the new user to calibrate the multi-user model, and the calibrated model can estimate the joint angle of the hand joint in a continuous motion process of a person in real time. Meanwhile, since a data-driven algorithm is applied, the complex implementation process of establishing an electromyography-to-physiology model is avoided.

Description

Technical Field

[0001] This application relates to the field of medical robot control technology, and in particular to a method, system and computer device for estimating cross-user hand joint angles. Background Technology

[0002] For decades, robotic hands have been widely researched and applied in search and rescue, industry, and prosthetics. Surface electromyography (EMG) signals, with their characteristics of anticipating movement and being easy to collect, are ideal physiological signals for extracting human movement intentions. They have numerous applications in rehabilitation medicine and human-computer interaction, and are widely used worldwide for the dexterous control of robotic hands. Given the hand's high flexibility as one of the most distinctive human organs, multi-degree-of-freedom real-time continuous control methods will be the future direction of robotic hand development, as this provides more natural and intuitive control. Many studies have established algorithms to map EMG signals to finger joint angles to achieve continuous real-time control strategies. Furthermore, considering applications for amputees and the difficulty in collecting hand EMG signals, arm EMG is often used to establish regression algorithms to estimate joint angles during hand movements. Since the muscles in the arm are active during hand movements, estimating hand joint angles using arm EMG signals is feasible. However, due to physiological differences between individuals, estimation methods based on surface electromyography (SEMG) signals exhibit performance variations among different users. The most common solution is to retrain the model for each user and calibrate it using newly acquired data. However, this process requires additional data collection and model testing and validation, and its complexity and time consumption hinder real-world deployment and implementation. Research indicates that despite the individuality and specificity of different users, anatomical similarities and muscle synergy among users suggest that data from other users can be used to quickly construct recognition strategies that meet the requirements of new users. Therefore, research on cross-user intent recognition methods has gradually become a research hotspot. Based on current research, cross-user intent recognition methods in SEMG signal control can be divided into two categories: neural network-based model transfer methods and feature-based model transfer methods. In neural network-based model transfer methods, the neural network is first trained using sufficient labeled data. Then, a portion of the model's parameters are fixed by freezing a portion of the network weights, and the unfrozen model parameters are updated using labeled or unlabeled target data. In contrast, feature-based model transfer methods attempt to obtain feature-domain-invariant features through feature domain alignment. The neural network extracts features from the source and target domains to calculate the loss in the feature domain, with the aim of reducing the mismatch in feature distribution in the implicit space.

[0003] Currently, knowledge transfer methods based on neural network models are widely used due to their simplicity. However, due to physiological differences between individuals, the electromyographic signal features extracted by the model vary significantly among different users. The universality of these features across different users cannot be guaranteed. Therefore, the results of knowledge transfer methods based on neural network models are often unsatisfactory after model correction, and the corrected model is prone to losing its generalization ability to the original users. In contrast, feature-based transfer methods update the model parameters by simultaneously learning information from both the original and target users. This allows for the extraction of features applicable to different users and less dependent on individual differences. Therefore, the corrected model does not lose its robustness and generalization ability to other users. However, the correction process is time-consuming, and user acceptance is lower. Summary of the Invention

[0004] Therefore, it is necessary to provide a cross-user hand joint angle estimation method, system, and computer device that has a shorter correction time and higher user acceptance, in order to address the shortcomings of the existing technology.

[0005] To solve the above problems, this application adopts the following technical solution:

[0006] One of the objectives of this application is to provide a method for estimating hand joint angles across users, comprising the following steps:

[0007] Acquire hand movement signal data from multiple users, including electromyographic signals and joint angle signals;

[0008] A multi-user model is constructed based on the hand movement signal data;

[0009] The multi-user model is calibrated using training data from new users;

[0010] The calibrated multi-user model estimates the finger joint angles of new users.

[0011] In some embodiments, in the step of acquiring hand activity signal data from multiple users, the hand activity signal data including electromyography (EMG) signals and joint angle signals, the EMG signals are from the extensor muscles of the fingers, flexor muscles of the fingers, biceps brachii, triceps brachii, and a ring of muscles in the forearm 2-6 cm from the elbow, and the EMG signals are sampled at a frequency of 2000 Hz; the joint angle signals are acquired through a data glove, and the sampling frequency is 20 Hz.

[0012] In some embodiments, the step of preprocessing the acquired hand movement signal data of multiple users is included after the step of acquiring hand movement signal data of multiple users and before proceeding to the next step.

[0013] In some embodiments, the preprocessing step of acquiring hand movement signal data from multiple users specifically includes the following steps:

[0014] A fourth-order Butterworth filter was used to bandpass filter the electromyography (EMG) signal, followed by baseline correction to remove noise. The EMG signal was then amplified using a u-law logarithmic scaling method. The joint angle signal was resampled to 2000Hz to synchronize the EMG and joint angle sequences in time. A low-pass filter was used to smooth the joint angle signal. The maximum and minimum values ​​of the acquired EMG and joint angle signals were recorded for normalization of training and testing data.

[0015] In some embodiments, the step of constructing a multi-user model based on the hand movement signal data specifically includes the following steps:

[0016] Training data on surface electromyography signals and joint angles from multiple users are acquired, and a supervised mean square loss function is used to train the model to obtain a multi-user neural network model. The multi-user neural network model includes a feature fusion module, two convolutional modules, and a multilayer perceptron module. The feature extraction module is composed of convolution or self-attention mechanism, and the feature fusion module is composed of convolution.

[0017] In some embodiments, the step of calibrating the multi-user model using training data from new users specifically includes the following steps:

[0018] Construct a new user neural network model with the same structure as the multi-user neural network model, and initialize it with the weights of the trained multi-user neural network model so that the initial parameters are the same as the parameters of the trained multi-user neural network model.

[0019] A feature region discriminator is set up, the parameters of the multi-user neural network model are frozen, and the parameters of the new user neural network model and the feature region discriminator are corrected and optimized to minimize the feature distance between the feature domains of the multi-user and the feature domains of the new user.

[0020] When the feature region discriminator has difficulty distinguishing two feature domains and finding the minimum distance between the two feature domains, the calibration of the multi-user model is completed.

[0021] In some embodiments, the feature region discriminator includes a structure of a densely connected residual module, which consists of a convolutional layer, a densely connected layer, a transition layer, and a residual connected layer. The preceding convolutional layers are used to extract features. The densely connected layer is directly connected to the convolutional layers. The convolutional layers are used to concatenate the extracted features as input to the subsequent convolutional layers. The transition layer uses 1×1 convolutions to compress the number of features and compress the network. The residual connected layer is used to give the network a deeper structure and richer feature representation.

[0022] In some embodiments, the step of estimating the finger joint angle of a new user in the calibrated multi-user model specifically includes the following steps: inputting a test dataset into the calibrated multi-user model, comparing the continuously estimated joint angle curve with the actual joint angle curve obtained by the joint angle sensor, and using three performance indicators—Pearson correlation coefficient, root mean square error, and coefficient of determination—as the evaluation criteria for the regression task.

[0023] A second objective of this application is to provide a cross-user hand joint angle estimation system, comprising:

[0024] The data acquisition unit is used to acquire hand activity signal data from multiple users, including electromyographic signals and joint angle signals.

[0025] A multi-user model construction unit is used to construct a multi-user model based on the hand activity signal data;

[0026] A calibration unit is used to calibrate the multi-user model using training data from new users;

[0027] An estimation unit is used to estimate the finger joint angles of a new user based on the calibrated multi-user model.

[0028] The third objective of this application is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the cross-user hand joint angle estimation method.

[0029] The present application adopts the above technical solution, and its beneficial effects are as follows:

[0030] The cross-user hand joint angle estimation method, system, and computer device provided in this application acquire hand activity signal data from multiple users, construct a multi-user model based on the hand activity signal data, calibrate the multi-user model using training data from new users, and estimate the finger joint angles of new users using the calibrated multi-user model. The above method, system, and computer device use surface electromyography signals and finger joint angle data from multiple users to build a multi-user model, and then use a newly proposed adversarial transfer learning strategy to calibrate the multi-user model using partial training data from new users. The calibrated model can estimate the joint angles of the hand joints in real time during continuous human movement. At the same time, due to the application of data-driven algorithms, the complex implementation process of building a physiological model from electromyography is avoided. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart illustrating the steps of the cross-user hand joint angle estimation method provided in this application embodiment.

[0033] Figure 2 This is a schematic diagram illustrating the model correction and cross-user knowledge transfer strategy provided in this embodiment.

[0034] Figure 3 This is a schematic diagram of the structure of the densely connected residual module constructed in this embodiment.

[0035] Figure 4 This is a schematic diagram of the structure of a user hand joint angle estimation system provided in an embodiment of this application.

[0036] Figure 5 This is a schematic diagram of the structure of the computer device provided in this embodiment. Detailed Implementation

[0037] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0038] In the description of this application, it should be understood that the terms "upper", "lower", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0039] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0041] Example 1

[0042] Please see Figure 1 The flowchart of a cross-user hand joint angle estimation method provided in Embodiment 1 of this application includes the following steps S110 to S140. The implementation of each step is described in detail below.

[0043] Step S110: Acquire hand movement signal data from multiple users. The hand movement signal data includes electromyographic signals and joint angle signals.

[0044] In some embodiments, in the step of acquiring hand activity signal data from multiple users, the hand activity signal data includes electromyographic (EMG) signals and joint angle signals. The EMG signals are derived from the extensor muscles of the fingers, flexor muscles of the fingers, biceps brachii, triceps brachii, and a ring of muscles in the forearm 2-6 cm from the elbow. The EMG signals are sampled at a frequency of 2000 Hz. The joint angle signals are acquired using a data glove at a sampling frequency of 20 Hz.

[0045] Specifically, electromyographic signals acquired using differential electrodes in the Delsys electromyography acquisition system were used to acquire joint angle signals using the CyberGlove II data glove.

[0046] In some embodiments, a step of preprocessing the acquired hand movement signal data of multiple users is included after the step of acquiring hand movement signal data of multiple users and before proceeding to the next step.

[0047] Furthermore, the preprocessing step for acquiring hand movement signal data from multiple users specifically includes the following steps:

[0048] A fourth-order Butterworth filter was used to bandpass filter the electromyography (EMG) signal. Baseline correction was then applied to remove noise from the EMG signal. Because some channels of the surface EMG signal have small amplitudes, a u-law logarithmic scaling method was used to amplify the EMG signal. The joint angle signal was resampled to 2000Hz to synchronize the EMG and joint angle sequences in time. A low-pass filter was used to smooth the joint angle signal, avoiding step jitter and making it more like a normal human motion curve. The maximum and minimum values ​​of the acquired EMG and joint angle signals were recorded for normalization of training and testing data.

[0049] Step S120: Construct a multi-user model based on the hand movement signal data.

[0050] In this embodiment, the step of constructing a multi-user model based on the hand activity signal data specifically includes the following steps: acquiring training data of surface electromyography signals and joint angles of multiple users, using a supervised mean square loss function to train the model, and obtaining a multi-user neural network model.

[0051] Please see Figure 2 The multi-user neural network model includes a feature fusion module, two convolutional modules, and a multilayer perceptron module, wherein the feature extraction module is composed of convolution or self-attention mechanism, and the feature fusion module is composed of convolution.

[0052] Step S130: Calibrate the multi-user model using the training data of the new users.

[0053] Please refer to the following: Figure 2 This illustrates the process of model training and model correction using adversarial transfer learning (where dashed lines indicate the optimization direction of model parameters during knowledge transfer). This process consists of a multi-user neural network model (Mnet-s), a new-user neural network model (Nnet-t), and a feature region discriminator (DFD). The multi-user neural network model extracts features from training data from multiple users, the new-user neural network model extracts features from training data from new users, and the feature region discriminator minimizes the feature distance between multi-user features and new-user features during model correction.

[0054] In this embodiment, the step of calibrating the multi-user model using training data from new users specifically includes the following steps:

[0055] Step S131: Construct a new user neural network model with the same structure as the multi-user neural network model, and initialize it with the weights of the trained multi-user neural network model so that the initial parameters are the same as the parameters of the trained multi-user neural network model.

[0056] Specifically, prior to model correction, the multi-user neural network model was trained using surface electromyography signals and joint angle data from multiple subjects, employing a supervised mean square loss function.

[0057] Step S132: Set the feature region discriminator. The parameters of the multi-user neural network model are frozen. The parameters of the new user neural network model and the feature region discriminator are corrected and optimized to minimize the feature distance between the feature domains of the multi-user and the feature domains of the new user.

[0058] Based on the idea of ​​adversarial transfer learning, the cross-user model transfer learning process is trained using the following formulas (1)-(5), L total It is the overall loss function used to optimize the neural network model for new users and the feature region discriminator.

[0059] L total =L D +L N (1)

[0060] L D The loss function representing the feature region discriminator is used to indicate the relative distance between the feature domains of the multi-user neural network model and the new user neural network model.

[0061] L D =-log(DFD(F) s ))-log(1-DFD(F t (2)

[0062] Where F s F represents the features of a multi-user neural network model. t This represents the characteristics of the new user's neural network model. L N This represents the loss function of the neural network model for new users.

[0063] L N =L map +L subject (3)

[0064] Where L map This indicates the degree to which features from a multi-user neural network model are mapped to features from a new user neural network model.

[0065] L map =-log(DFD(F) t(4)

[0066] L subject This is used to improve the performance of the loss function, making the neural network model for new users more sensitive to the features of new users during the transfer learning process.

[0067]

[0068] Where Nnet(x) represents the finger joint angle estimated by the model. The labels represent the angles of the finger joints, where N represents the number of finger joints and α is a weighting coefficient.

[0069] Step S133: When the feature region discriminator has difficulty distinguishing two feature domains and finding the minimum distance between the two feature domains, the calibration of the multi-user model is completed.

[0070] Please see Figure 3 This demonstrates the structure of the main component of the feature region discriminator—the densely connected residual module. The densely connected residual module consists of convolutional layers, a densely connected layer, a transition layer, and residual connections. The convolutional layers extract features, while the densely connected layer provides direct connections between the convolutional layers. The features extracted by the preceding convolutional layers are concatenated as input to subsequent convolutional layers. This improves information propagation within the neural network, thus accelerating model convergence. The transition layer then uses 1×1 convolutions to compress the number of features and compress the network. Finally, the residual connection structure gives the network a deeper structure and a richer feature representation.

[0071] It is understandable that three representative features from a multi-user neural network model ( Figure 2 (represented by line A in the middle) and three representative features from the new user neural network model ( Figure 2 (represented by line B) are respectively fed into the densely connected residual module.

[0072] It is understood that the transfer method proposed in this embodiment corrects parameters by learning information from both the original user and the target user. It can extract features applicable to different users and largely independent of individual differences. Therefore, the corrected model can reuse more features from the original user, achieving higher estimation accuracy and exhibiting better robustness and generalization ability. Furthermore, the knowledge transfer algorithm proposed in this embodiment has a short correction time, which is beneficial for practical deployment. The algorithm proposed in this invention achieves model correction by minimizing the feature distribution differences between different users, eliminating the need for model retraining. Therefore, it has a short correction time and high user acceptability.

[0073] Step S140: The calibrated multi-user model estimates the finger joint angles of new users.

[0074] In some embodiments, the step of estimating the finger joint angle of a new user in the calibrated multi-user model specifically includes the following steps: inputting a test dataset into the calibrated multi-user model, comparing the continuously estimated joint angle curve with the actual joint angle curve obtained by the joint angle sensor, and using three performance indicators—Pearson correlation coefficient, root mean square error, and coefficient of determination—as the evaluation criteria for the regression task.

[0075] Specifically, the Pearson correlation coefficient (CC). CC measures the linear correlation between estimated finger joint angles and corresponding actual data. Its calculation formula is as follows.

[0076]

[0077] The closer the CC value is to 1, the closer the predicted finger movement trajectory is to the actual trajectory, and the higher the estimation accuracy of the method.

[0078] 2) Root Mean Square Error (RMSE) (°). RMSE is used to evaluate the deviation between the estimated and measured finger joint angles, and the unit is degrees (°). The method for calculating RMSE is as follows:

[0079]

[0080] 3) Coefficient of Determination (R²). As a comprehensive evaluation index for the overall accuracy of the model, R² ranges from 0 to 1. It is defined as the percentage change in the true value explained by the estimated value. The larger the R² value, the better the estimation performance. It is expressed as:

[0081]

[0082] In the above formula, N represents the sample size, p i The sample points representing the predicted finger joint angles, g i Sample points representing the actual angles of finger joints.

[0083] The cross-user hand joint angle estimation method provided in this application establishes a multi-user model using surface electromyography signals and finger joint angle data from multiple users. Then, a newly proposed adversarial transfer learning strategy is adopted to calibrate the multi-user model using partial training data from new users. The calibrated model can estimate the joint angles of the hand joints in real time during continuous human movement. At the same time, due to the application of data-driven algorithms, the complex implementation process of building a physiological model from electromyography is avoided.

[0084] The cross-user hand joint angle estimation method provided by this application can be used for the control and operation of industrial and aerospace robots. Based on the surface electromyography (sEMG) information of the human body, it manipulates a robotic arm to complete the grasping of objects of different sizes. The continuously estimated hand joint angles through sEMG signals can make the entire control process natural and dexterous, enabling efficient interaction between humans and the robotic arm. The cross-user model transfer technology can improve the generalization performance of the hand state estimation algorithm for new users who have never been seen before. At the same time, since there is no need to retrain the model, it can improve the rapid deployment ability of the model in the real world.

[0085] The cross-user hand joint angle estimation method provided by this application can provide a more intelligent and user-friendly rehabilitation training device for patients with motor dysfunction. By introducing the sEMG pattern recognition technology into the control link of the rehabilitation training system, the model is trained through the residual sEMG signals of the patients and the virtual hand joint angle information when the patients imagine completing the movements, helping the patients with motor dysfunction to complete the rehabilitation training more actively. At the same time, based on the transfer learning technology, the model pre-trained with the initial data is generalized to new users. While the generalization performance is improved, the time for recalibrating the model is greatly reduced, which greatly alleviates the burden on the users.

[0086] Please refer to Figure 4 , which provides a cross-user hand joint angle estimation system for Embodiment 2, including: a data acquisition unit 110 for acquiring hand activity signal data of multiple users, where the hand activity signal data includes sEMG signals and joint angle signals; a multi-user model construction unit 120 for constructing a multi-user model according to the hand activity signal data; a calibration unit 130 for calibrating the multi-user model with the training data of a new user; and an estimation unit 140 for estimating the finger joint angles of a new user with the calibrated multi-user model.

[0087] For the detailed working mode of the estimation system provided in this embodiment, reference can be made to the cross-user hand joint angle estimation method provided in the embodiment, which will not be elaborated here.

[0088] The cross-user hand joint angle estimation system provided by this application uses the surface electromyography signals and finger joint angle data of multiple users to establish a multi-user model, and then adopts a newly proposed adversarial transfer learning strategy to calibrate the multi-user model with partial training data of a new user. The calibrated model can real-time estimate the joint angles of the hand joints during continuous human movement. At the same time, due to the application of a data-driven algorithm, it avoids the complex implementation process of establishing a model from sEMG to physiological model.

[0089] Please refer to Figure 5This is a schematic diagram of a computer device structure according to an embodiment of this application. The computer device 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0090] The memory 52 stores program instructions for implementing the above-described cross-user hand joint angle estimation method.

[0091] The processor 51 is used to execute program instructions stored in the memory 52 to implement the cross-user hand joint angle estimation method.

[0092] The processor 51 can also be referred to as a CPU (Central Processing Unit). The processor 51 may be an integrated circuit chip with signal processing capabilities. The processor 51 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0093] Example 2

[0094] In this embodiment, a sliding window was used to generate surface electromyography (EMG) signal sequences and joint angle sequences with a window length of 2000 sampling points. The sliding window step size was 100 sampling points. The joint angle and EMG signal data in each sliding window were used as one sample data. The dimension of the joint angle vector represented the estimated number of joint angles. Data from 35 users was collected in this embodiment. Data from 28 users was used to train the multi-user model, and the remaining 7 users were used as new users to correct and test the algorithm performance. For the new users' data, 80% of the data samples were used to correct the model, and 20% of the data samples were used as test data to test the algorithm performance.

[0095] This invention uses joint and electromyographic data from 35 subjects performing grasping actions on objects of different sizes to train, correct, and test the network model. Data from 28 of these subjects is used to build a multi-user model, while the remaining 7 subjects are used as new users for model correction and testing. 80% of the data from each new user is used for model transfer correction, and the other 20% is used for model performance testing. The CC, RMSE, and R2 regression model performance indicators for the proximal interphalangeal joints and metacarpophalangeal joints of the hand are tested, as shown in Table 1.

[0096] Table 1. Regression performance evaluation of three different model transfer algorithms on the same hand multi-joint angle estimation task.

[0097] Evaluation indicators This invention uses a network. Long Short-Term Memory Network Sparse pseudo-input Gaussian process CC 0.859±0.012 0.830±0.011 0.775±0.009 RMSE 7.344±0.242 7.973±0.219 8.953±0.267 R2 0.761±0.035 0.711±0.040 0.643±0.047

[0098] The results show that the performance metrics are superior to the two commonly used model transfer algorithms. This indicates that the transfer learning algorithm in this embodiment not only brings higher accuracy to the finger kinematics estimation for new users, but also has better generalization performance for new users.

[0099] It is understood that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above are merely preferred embodiments of this application, and only specifically describe the technical principles of this application. These descriptions are only for explaining the principles of this application and should not be construed as limiting the scope of protection of this application in any way. Based on this explanation, any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application, as well as other specific embodiments of this application that can be conceived by those skilled in the art without creative effort, should be included within the scope of protection of this application.

Claims

1. A method for estimating hand joint angles across users, characterized in that, Includes the following steps: Acquire hand movement signal data from multiple users, including electromyographic signals and joint angle signals; A multi-user model is constructed based on the hand movement signal data; The multi-user model is calibrated using training data from new users; The calibrated multi-user model estimates the finger joint angles of new users; The step of constructing a multi-user model based on the hand movement signal data specifically includes the following steps: Training data of surface electromyography signals and joint angles from multiple users are obtained, and a supervised mean square loss function is used to train the model to obtain a multi-user neural network model. The multi-user neural network model includes a feature fusion module, two convolutional modules and a multilayer perceptron module, wherein: the feature extraction module is composed of convolution or self-attention mechanism, and the feature fusion module is composed of convolution. The step of calibrating the multi-user model using training data from new users specifically includes the following steps: Construct a new user neural network model with the same structure as the multi-user neural network model, and initialize it with the weights of the trained multi-user neural network model so that the initial parameters are the same as the parameters of the trained multi-user neural network model. A feature region discriminator is set up, the parameters of the multi-user neural network model are frozen, and the parameters of the new user neural network model and the feature region discriminator are corrected and optimized to minimize the feature distance between the feature domains of the multi-user and the feature domains of the new user. When the feature region discriminator has difficulty distinguishing two feature domains and finding the minimum distance between the two feature domains, the calibration of the multi-user model is completed.

2. The cross-user hand joint angle estimation method of claim 1, wherein, In the step of acquiring hand activity signal data from multiple users, including electromyography (EMG) signals and joint angle signals, the EMG signals are derived from the extensor muscles of the fingers, flexor muscles of the fingers, biceps brachii, triceps brachii, and a ring of muscles in the forearm 2-6 cm from the elbow. The EMG signals are sampled at a frequency of 2000 Hz. The joint angle signals are acquired using a data glove at a sampling frequency of 20 Hz.

3. The cross-user hand joint angle estimation method according to claim 1 or 2, wherein The process includes a preprocessing step for acquiring hand movement signal data from multiple users, following the step of acquiring hand movement signal data from multiple users and before proceeding to the next step.

4. The cross-user hand joint angle estimation method of claim 3, wherein, The preprocessing steps for acquiring hand movement signal data from multiple users specifically include the following steps: A fourth-order Butterworth filter was used to bandpass filter the electromyography (EMG) signal, followed by baseline correction to remove noise. The EMG signal was then amplified using a u-law logarithmic scaling method. The joint angle signal was resampled to 2000Hz to synchronize the EMG and joint angle sequences in time. A low-pass filter was used to smooth the joint angle signal. The maximum and minimum values ​​of the acquired EMG and joint angle signals were recorded for normalization of training and testing data.

5. The cross-user hand joint angle estimation method of claim 1, wherein, The feature region discriminator includes a structure of a densely connected residual module, which consists of a convolutional layer, a densely connected layer, a transformation layer, and a residual connection layer. The preceding convolutional layers are used to extract features. The densely connected layer is directly connected to the convolutional layers. The convolutional layers are used to concatenate the extracted features as input to the subsequent convolutional layers. The transformation layer uses 1×1 convolutions to compress the number of features and compress the network. The residual connection layer is used to give the network a deeper structure and richer feature representation.

6. The cross-user hand joint angle estimation method as described in claim 1, characterized in that, The step of estimating the finger joint angles of new users in the calibrated multi-user model specifically includes the following steps: inputting the test dataset into the calibrated multi-user model, comparing the continuously estimated joint angle curves with the actual joint angle curves obtained by the joint angle sensors, and using three performance indicators—Pearson correlation coefficient, root mean square error, and coefficient of determination—as the evaluation criteria for this regression task.

7. A cross-user hand joint angle estimation system, characterized in that, The system employs the cross-user hand joint angle estimation method as described in any one of claims 1-6; the system comprises: The data acquisition unit is used to acquire hand activity signal data from multiple users, including electromyographic signals and joint angle signals. A multi-user model construction unit is used to construct a multi-user model based on the hand activity signal data; A calibration unit is used to calibrate the multi-user model using training data from new users; An estimation unit is used to estimate the finger joint angles of a new user based on the calibrated multi-user model.

8. A computer device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the cross-user hand joint angle estimation method as described in any one of claims 1-6.

Images