Control method and system of a rehabilitation training robot

Abstract

The present application relates to the technical field of medical rehabilitation robots, and particularly relates to a control method and system of a rehabilitation training robot. The method comprises: performing feature coding on neural signals and electromyographic signals through a multi-scale space-time encoder to obtain coded features; performing state compression on biomechanical signals through a biomechanical feature extractor to obtain biomechanical coded features; processing the coded features through a cross-modal attention fusion module to generate enhanced fusion features; outputting motion trajectory prediction, angular velocity prediction, interaction torque prediction and intention confidence through a multi-task prediction head module, and performing uncertainty quantification through Bayesian deep learning; and generating a control instruction based on a digital twin query and a variable parameter impedance controller module. The present application realizes low inference delay, significantly improves intention recognition accuracy and response speed, and enhances the safety of rehabilitation training and the naturalness of human-computer interaction.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation robot control technology, specifically to a control method and system for a rehabilitation training robot. Background Technology

[0002] With the aging population and rising incidence of neurological diseases, rehabilitation robots have become an important technological means to assist patients in restoring motor function. However, existing rehabilitation robot control systems face the following technical bottlenecks:

[0003] 1. Insufficient accuracy in intent recognition: Traditional methods rely on single biosignals (such as electromyography alone), making it difficult to comprehensively capture the user's motor intent, resulting in significant prediction errors in complex motor tasks. Template matching and pattern recognition methods have poor robustness under non-stationary physiological signals, leading to significant deviations between the auxiliary torque output and the user's true intent.

[0004] 2. High response latency: Traditional machine learning methods typically require a data processing window of over 500ms, and deep learning models, due to their high computational complexity, also struggle to meet real-time interaction requirements. This latency severely reduces the naturalness of human-computer interaction and user experience, and may even lead to incoordination and secondary injuries.

[0005] 3. Lack of personalized adaptation: Fixed-parameter models are difficult to adapt to the different physiological characteristics (such as muscle strength, nerve signal quality) and rehabilitation stages (from flaccid paralysis to recovery) of different patients. Insufficient assistance cannot achieve the training effect, while excessive assistance inhibits the patient's active participation, violating the principle of "assistance on demand".

[0006] 4. Weak security safeguards: Existing systems cannot quantify and predict uncertainties, and still output high-risk control commands even when model confidence is low. The lack of real-time monitoring of physiological constraints such as muscle fatigue and joint torque boundaries poses safety risks.

[0007] 5. Data privacy risks: Centralized model training requires uploading sensitive physiological data from patients, which does not comply with medical privacy protection requirements such as HIPAA and GDPR, and limits the application of multi-center clinical data.

[0008] In recent years, multimodal fusion technology has been introduced into the field of rehabilitation robots. However, existing methods mostly employ simple weighted fusion or early feature splicing, failing to fully explore the deep interactive relationships between modalities. At the same time, the "black box" nature of deep learning models makes it difficult for clinicians to understand the basis of control decisions, limiting their trustworthiness and interpretability in medical scenarios.

[0009] Therefore, there is an urgent need for a rehabilitation robot control solution that can identify users' movement intentions in real time, accurately and securely, and has adaptive learning and privacy protection capabilities. Summary of the Invention

[0010] The purpose of this invention is to provide a control method and system for a rehabilitation training robot. Through an innovative multimodal asynchronous fusion prediction network architecture and variable parameter impedance control strategy, it achieves millisecond-level intention recognition, uncertainty quantification, personalized adaptation, and safety-assisted control.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A control method for a rehabilitation training robot includes the following steps:

[0013] S100, acquire data collected by multimodal sensors and a trained multimodal asynchronous fusion prediction network; wherein, the multimodal asynchronous fusion prediction network includes a multi-scale spatiotemporal encoder module, a cross-modal attention fusion module, a multi-task prediction head module, and a variable parameter impedance controller module;

[0014] S200, the data collected by the multimodal sensor is input into the trained multimodal asynchronous fusion prediction network. The neural signals and electromyographic signals are processed by multi-scale temporal convolution, bidirectional gated recurrent unit processing and temporal self-attention calculation through the multi-scale spatiotemporal encoder module to obtain the encoded features. The biomechanical signals are compressed by the biomechanical feature extractor to obtain the biomechanical encoded features.

[0015] S300, the cross-modal attention fusion module processes the encoded features and biomechanical encoded features to generate enhanced fusion features; the cross-modal attention fusion module calculates the cross-attention weights between modalities through linear transformations of queries, keys, and values, and dynamically adjusts the contribution of each modality using a gating mechanism;

[0016] S400 processes the fused features through a multi-task prediction head module, outputting motion trajectory prediction, angular velocity prediction, interaction torque prediction, and intent confidence; the motion trajectory prediction uses Bayesian deep learning to quantify uncertainty and construct a diagonal Gaussian distribution.

[0017] The S500 obtains the user's maximum torque, muscle activation level, and fatigue index based on a digital twin query. It then calculates the adaptive stiffness and damping matrix through a variable parameter impedance controller module, generates control commands, and outputs them to the robot actuator.

[0018] Optionally, in S200, the multi-scale spatiotemporal encoder module includes a parallel multi-scale temporal convolution unit, a bidirectional gated recurrent unit, a temporal self-attention unit, and a feature fusion and dimensionality reduction unit;

[0019] The multi-scale temporal convolutional unit adopts 4-way parallel one-dimensional dilated convolution with dilation rates of 1, 4, 8 and 16, with each channel outputting 32 channels. After concatenation, it is fused by 1×1 convolution to obtain 128 channels of output.

[0020] The bidirectional gated loop unit contains two layers of Bi-GRU, with 128 hidden units. The output dimension after splicing the forward and backward hidden states is 256.

[0021] The temporal self-attention unit adopts an 8-head attention mechanism, with a head dimension of 16 and a key vector dimension of 64;

[0022] The feature fusion and dimensionality reduction unit reduces the feature dimensionality to 128 dimensions through residual connections, layer normalization, and adaptive max pooling.

[0023] Optionally, in S300, the cross-modal attention fusion module is implemented in the following way:

[0024] Neural coding features, electromyographic coding features, and biomechanical coding features are projected onto the query space, key space, and value space, respectively, with a projection dimension of 64 for each.

[0025] Calculate the cross-attention weights of mode i to mode j. ,in =64;

[0026] A four-head attention mechanism is used to compute modal relationships in different subspaces in parallel;

[0027] Through learnable gating vectors Enhancement features from various modalities are dynamically fused. The gating vector is generated using the sigmoid function, and the distribution smoothness is controlled using a temperature parameter of 0.5.

[0028] Optionally, in S400, the multi-task prediction head module includes a shared feature extraction layer and four parallel branches;

[0029] The shared feature extraction layer consists of two fully connected layers with dimensions of 256 and 128, respectively, and employs ReLU activation, layer normalization, and regularization with a Dropout rate of 0.3.

[0030] The motion trajectory prediction branch outputs the mean of joint angle increments. Sum of logarithmic variance Construct a diagonal Gaussian distribution ;

[0031] The angular velocity prediction branch outputs angular velocity. ;

[0032] The interactive torque prediction branch outputs torque. ;

[0033] The confidence branch of the intent outputs confidence ∈ [0,1].

[0034] Optionally, the multi-task prediction head module is trained using a composite loss function, which is:

[0035] ;

[0036] in, For mean square error loss, For negative log-likelihood loss, For variational attention KL divergence loss, For attention sparsity L1 regularization loss; .

[0037] Optionally, in S500, the variable parameter impedance controller module is implemented in the following way:

[0038] Calculate the auxiliary ratio ,in The maximum torque predicted for the digital twin;

[0039] Calculate the uncertainty factor ,in To predict the standard deviation of uncertainty;

[0040] Basic stiffness The MLP network is dynamically generated based on intent confidence and fatigue index using a 3-layer MLP network, which is based on historical states. For additional input;

[0041] Final joint stiffness Damping Where ξ = 0.7 is the critical damping ratio. This represents the joint inertia.

[0042] Optionally, the method further includes an online continuous learning step:

[0043] Maintain a circular buffer with a capacity of N=1000 samples to store the most recently trained data;

[0044] Perform a mini-batch update after every K=100 training cycles;

[0045] Elastic weights are used to prevent catastrophic forgetting; the loss function is... ,in For the diagonal elements of the Fisher information matrix, .

[0046] Optionally, the method further includes a federated learning aggregation step:

[0047] Aggregate model parameters from multiple clients in the cloud, based on the data volume of each client. Calculate initial weights ;

[0048] Clients are clustered by feature statistics using K-means clustering. Divide into K clusters and calculate the intra-cluster similarity. ;

[0049] Generate weighted weights Then perform intra-cluster model aggregation, and then sort by performance metrics. Perform global aggregation;

[0050] After receiving the global model, the client integrates it with the local model in a personalized manner, with an integration coefficient α=0.3.

[0051] On the other hand, embodiments of the present invention provide an environment-adaptive multi-channel feature fusion system, including:

[0052] At least one processor;

[0053] At least one memory for storing at least one program;

[0054] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0055] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the above-described method.

[0056] The beneficial effects of this invention include:

[0057] 1. Innovative Multimodal Asynchronous Fusion Predictive Network Architecture: A multi-scale spatiotemporal encoder module extracts deep temporal features from neural and electromyographic signals, combined with a biomechanical feature extractor to process biomechanical signals. A cross-modal attention fusion module then dynamically mines intermodal interactions, achieving effective fusion of data from different types of sensors with varying sampling rates. This enhances the comprehensiveness and robustness of motion intent recognition. The multi-task predictive head module not only outputs key control parameters such as motion trajectory, angular velocity, and interaction torque, but also quantifies uncertainty through Bayesian deep learning and outputs intent confidence scores, providing a quantitative basis for safety control decisions.

[0058] 2. Variable parameter impedance control strategy to achieve personalized adaptation: Based on physiological parameters such as the user's maximum torque, muscle activation level and fatigue index provided by the digital twin, the assistance ratio and uncertainty factor are dynamically calculated. The basic stiffness is generated through the MLP network, and the joint stiffness and damping are finally determined by combining the joint characteristics. This allows the robot assistance force to accurately match the user's real-time physiological state and motor ability, avoiding both insufficient assistance affecting the training effect and excessive assistance inhibiting the user's active participation. This truly realizes personalized rehabilitation training with "assistance on demand".

[0059] 3. Online Continuous Learning and Federated Learning Ensure System Performance and Data Security: Online continuous learning maintains a circular buffer to store the latest data, performs regular small-batch updates, and employs elastic weight solidification technology to effectively prevent catastrophic forgetting. This allows the model to continuously adapt to changes in users' physiological characteristics during practical applications, continuously optimizing prediction accuracy. The federated learning aggregation step, while protecting the privacy of sensitive medical data on each client, achieves effective aggregation of multi-center model parameters, improves the model's generalization ability, and meets medical privacy protection requirements such as HIPAA and GDPR. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0061] Figure 1 This is a flowchart of a control method for a rehabilitation training robot according to an embodiment of the present invention;

[0062] Figure 2 This is a structural diagram of the control system of a rehabilitation training robot according to an embodiment of the present invention. Detailed Implementation

[0063] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0065] It is understood that the terms “first,” “second,” etc., used in this invention may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to determination” as used herein may be interpreted as “when…” or “when…” or “in response to determination.”

[0066] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0067] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0068] refer to Figure 1 This invention provides a control method for a rehabilitation training robot, the method comprising the following steps:

[0069] S100, acquire data collected by multimodal sensors and a trained multimodal asynchronous fusion prediction network; wherein, the multimodal asynchronous fusion prediction network includes a multi-scale spatiotemporal encoder module, a cross-modal attention fusion module, a multi-task prediction head module, and a variable parameter impedance controller module;

[0070] S200, the data collected by the multimodal sensor is input into the trained multimodal asynchronous fusion prediction network. The neural signals and electromyographic signals are processed by multi-scale temporal convolution, bidirectional gated recurrent unit processing and temporal self-attention calculation through the multi-scale spatiotemporal encoder module to obtain the encoded features. The biomechanical signals are compressed by the biomechanical feature extractor to obtain the biomechanical encoded features.

[0071] S300, the cross-modal attention fusion module processes the encoded features and biomechanical encoded features to generate enhanced fusion features; the cross-modal attention fusion module calculates the cross-attention weights between modalities through linear transformations of queries, keys, and values, and dynamically adjusts the contribution of each modality using a gating mechanism;

[0072] S400 processes the fused features through a multi-task prediction head module, outputting motion trajectory prediction, angular velocity prediction, interaction torque prediction, and intent confidence; the motion trajectory prediction uses Bayesian deep learning to quantify uncertainty and construct a diagonal Gaussian distribution.

[0073] The S500 obtains the user's maximum torque, muscle activation level, and fatigue index based on a digital twin query. It then calculates the adaptive stiffness and damping matrix through a variable parameter impedance controller module, generates control commands, and outputs them to the robot actuator.

[0074] In this embodiment, steps S100 to S500, by fusing multimodal sensor data with a deep learning model, achieve accurate prediction and personalized control of the user's motion intentions. The multi-scale spatiotemporal encoder module captures multi-scale temporal features of neural and electromyographic signals through four parallel one-dimensional dilated convolutions (dilation rates 1, 4, 8, and 16). Combined with a bidirectional gated recurrent unit (2-layer Bi-GRU, 128 hidden units) and an 8-head temporal self-attention mechanism (head dimension 16), it effectively extracts dynamic change patterns and key time node information from the signals. The biomechanical feature extractor performs state compression on force, position, and other signals, providing structured features for subsequent fusion. The cross-modal attention fusion module innovatively projects neural and electromyographic encoded features onto the query space, and biomechanical encoded features onto the key and value spaces. By calculating intermodal cross-attention weights (such as the attention of neural features to biomechanical features), and utilizing learnable gating vectors (g∈R), it achieves this. 3The contribution of each modality is dynamically adjusted by a temperature parameter of 0.5, solving the problems of asynchronicity and complementarity of multi-source heterogeneous data. The generated enhanced fusion features have both high discriminativeness and robustness. The multi-task prediction head module, based on the shared feature extraction layer (2 fully connected layers, 256 / 128 dimensions, ReLU activation, Dropout 0.3), outputs motion trajectory, angular velocity, interaction torque, and intent confidence in parallel. The composite loss function balances the requirements of prediction accuracy and uncertainty quantification. The variable parameter impedance controller module is deeply integrated with digital twin feedback. It reflects the user's active participation through the assistance ratio, quantifies the prediction risk through uncertainty factors, and then dynamically generates basic stiffness based on intent confidence and fatigue index through a 3-layer MLP network. Finally, joint stiffness and damping are determined to ensure that the robot's assistance force can effectively compensate for motion obstacles while avoiding dependence caused by over-assistance, achieving synergistic optimization of safety and training effect.

[0075] Optionally, in S200, the multi-scale spatiotemporal encoder module includes a parallel multi-scale temporal convolution unit, a bidirectional gated recurrent unit, a temporal self-attention unit, and a feature fusion and dimensionality reduction unit;

[0076] The multi-scale temporal convolutional unit adopts 4-way parallel one-dimensional dilated convolution with dilation rates of 1, 4, 8 and 16, with each channel outputting 32 channels. After concatenation, it is fused by 1×1 convolution to obtain 128 channels of output.

[0077] The bidirectional gated loop unit contains two layers of Bi-GRU, with 128 hidden units. The output dimension after splicing the forward and backward hidden states is 256.

[0078] The temporal self-attention unit adopts an 8-head attention mechanism, with a head dimension of 16 and a key vector dimension of 64;

[0079] The feature fusion and dimensionality reduction unit reduces the feature dimensionality to 128 dimensions through residual connections, layer normalization, and adaptive max pooling.

[0080] This embodiment achieves simultaneous capture of features at different time scales in neural and electromyographic signals through parallel multi-scale temporal convolutional units. Convolutions with a dilation rate of 1 focus on high-frequency detail fluctuations in the signal, while convolutions with a dilation rate of 16 capture long-term trend changes. The four-way parallel design enables the model to simultaneously process multimodal dynamic information ranging from instantaneous spikes to continuous activity. Bidirectional gated recurrent units effectively model the contextual dependencies of the signal through forward and backward propagation. The two-layer Bi-GRU structure enhances the learning ability for complex temporal patterns, and the 128 hidden units ensure sufficient expression of key features. The concatenated 256-dimensional output contains comprehensive information about the signal's past and future. The 8-head mechanism of the temporal self-attention unit allows the model to compute the correlation weights between features in parallel from different subspaces. The configuration of a 16-head dimension and a 64-key vector dimension balances computational complexity and attention resolution, enabling precise location of key timestamps and feature combinations in the signal. The feature fusion and dimensionality reduction unit alleviates the vanishing gradient problem in deep networks through residual connections, layer normalization stabilizes the training process, and adaptive max pooling compresses the dimension to 128 while preserving key features, providing a compact and efficient feature representation for subsequent cross-modal fusion. This modular design enables the multi-scale spatiotemporal encoder to robustly handle the non-stationarity and individual variability of neuromuscular signals, providing high-quality encoded features for intent prediction.

[0081] Optionally, in S300, the cross-modal attention fusion module is implemented in the following way:

[0082] Neural coding features, electromyographic coding features, and biomechanical coding features are projected onto the query space, key space, and value space, respectively, with a projection dimension of 64 for each.

[0083] Calculate the cross-attention weights of mode i to mode j. ,in =64;

[0084] A four-head attention mechanism is used to compute modal relationships in different subspaces in parallel;

[0085] Through learnable gating vectors Enhancement features from various modalities are dynamically fused. The gating vector is generated using the sigmoid function, and the distribution smoothness is controlled using a temperature parameter of 0.5.

[0086] This embodiment achieves deep collaboration and dynamic weighting of different modal features. Specifically, neural coding features are projected onto the query space Q via a linear transformation matrix WQ, electromyographic coding features are projected onto the bond space K1 via WK1, and biomechanical coding features are projected onto the bond space K2 via WK2. The projection dimensions of the three are unified to 64 to ensure dimensionality matching for attention calculation. During modal cross-attention weight calculation, the similarity between two modal features is quantified through a scaled dot product attention mechanism. The four-head parallel computation mines modal associations from different semantic subspaces. For example, head 1 focuses on feature coupling in the initiation stage of movement, head 2 focuses on collaborative patterns in the steady stage of movement, and the multi-head results are concatenated and output as modal enhancement features through a linear transformation. The gating vector g generates modal weights for each modality through a 3D sigmoid function. For example, g1 controls the contribution of neural features, g2 controls the contribution of electromyographic features, and g3 controls the contribution of biomechanical features. The temperature parameter 0.5 adjusts the slope of the sigmoid function to make the weight distribution smoother and avoid excessive influence of extreme values ​​on the fusion result. This design not only achieves information complementarity between modalities, but also dynamically adjusts the importance of each modality according to different movement stages and user states. For example, in the rapid movement intention initiation stage, the weight of neural and electromyographic features will be significantly increased, while in the movement maintenance stage, the weight of biomechanical features will be increased accordingly, thereby significantly improving the ability of fused features to represent the user's true movement intention.

[0087] Optionally, in S400, the multi-task prediction head module includes a shared feature extraction layer and four parallel branches;

[0088] The shared feature extraction layer consists of two fully connected layers with dimensions of 256 and 128, respectively, and employs ReLU activation, layer normalization, and regularization with a Dropout rate of 0.3.

[0089] The motion trajectory prediction branch outputs the mean of joint angle increments. Sum of logarithmic variance Construct a diagonal Gaussian distribution ;

[0090] The angular velocity prediction branch outputs angular velocity. ;

[0091] The interactive torque prediction branch outputs torque. ;

[0092] The confidence branch of the intent outputs confidence ∈ [0,1].

[0093] This embodiment achieves joint learning of multiple tasks and uncertainty quantification. The shared feature extraction layer performs deep abstraction of fused features through two fully connected layers. The dimensionality reduction process from 256 to 128 dimensions, combined with the ReLU activation function, enhances the nonlinear expressive power. Layer normalization stabilizes the training distribution, and a dropout rate of 0.3 effectively prevents overfitting, providing robust basic features for each prediction branch. The motion trajectory prediction branch innovatively introduces Bayesian deep learning. By outputting the mean and logarithmic variance of the joint angle increment, a diagonal Gaussian distribution model is constructed, so that the prediction result not only includes the expected value of the trajectory, but also quantifies the uncertainty of the prediction through variance. For example, when the user's motion intention is ambiguous or the sensor signal is interfered with by noise, the model will output a large variance value, providing a risk assessment basis for subsequent control decisions. The angular velocity prediction branch and the interaction torque prediction branch output the velocity of joint motion and the expected interaction force through independent fully connected layers, respectively, to achieve a multi-dimensional characterization of the motion state. The intent confidence branch outputs a confidence value in the [0,1] interval through the sigmoid activation function, comprehensively reflecting the reliability of the model's prediction of the current movement intent. This value can be directly used to adjust the parameters of the variable-parameter impedance controller. When the confidence level is lower than a set threshold (e.g., 0.6), the system automatically reduces the assistance level and enters a safety monitoring mode. This multi-task parallel architecture, through the combination of shared feature extraction and task-specific parameters, reduces the total number of parameters while improving the overall prediction performance by utilizing the correlation information between tasks. In particular, the introduction of the uncertainty quantification mechanism provides key technical support for the safety and personalized adjustment of rehabilitation training.

[0094] Optionally, the multi-task prediction head module is trained using a composite loss function, which is:

[0095] ;

[0096] in, For mean square error loss, For negative log-likelihood loss, For variational attention KL divergence loss, For attention sparsity L1 regularization loss; .

[0097] This embodiment achieves precise control and multi-objective optimization of the multi-task learning process. Mean squared error loss (Lmse) is used to optimize the prediction accuracy of motion trajectory, angular velocity, and interaction torque. Negative log-likelihood loss (Lnll) targets the diagonal Gaussian distribution of motion trajectory prediction, combining the predicted mean and log-variance with the true angle increment to guide the model to simultaneously learn the trajectory expectation and uncertainty distribution. Variational attention KL divergence loss constrains the distribution of cross-modal attention weights, minimizing the KL divergence between the model's attention distribution and a uniform distribution, preventing excessive concentration of attention on a single modality or feature, and promoting balanced collaboration between modalities. Attention sparsity L1 regularization loss (Lsparse) applies an L1 penalty to the cross-modal attention weights, prompting the model to focus on key modal associations, filtering redundant information, and improving the discriminative power of fused features. The weighting coefficients of each loss term are set according to the task priority, with trajectory prediction accuracy as the core optimization objective, uncertainty quantification as the secondary objective, and modal fusion quality as an auxiliary constraint. Through this weighted combination, the model can ensure prediction accuracy while possessing reliable uncertainty assessment capabilities and robust cross-modal fusion performance, laying a high-quality prediction foundation for subsequent control command generation.

[0098] Optionally, in S500, the variable parameter impedance controller module is implemented in the following way:

[0099] Calculate the auxiliary ratio ,in The maximum torque predicted for the digital twin;

[0100] Calculate the uncertainty factor ,in To predict the standard deviation of uncertainty;

[0101] Basic stiffness The MLP network is dynamically generated based on intent confidence and fatigue index using a 3-layer MLP network, which is based on historical states. For additional input;

[0102] Final joint stiffness Damping Where ξ = 0.7 is the critical damping ratio. This represents the joint inertia.

[0103] This embodiment achieves dynamic adaptive adjustment of the robot's assistive force, ensuring the safety and effectiveness of the training process. The assist ratio α is calculated as the ratio of the user's predicted interactive torque to the maximum torque provided by the digital twin, and is limited to the [0,1] interval by the clip function, intuitively reflecting the user's active participation level. When the user's active force approaches their maximum capacity (α→1), the assist ratio decreases, and the robot provides a smaller assistive force to encourage active movement; when the user's force is weak (α→0), the assist ratio increases, and the robot increases the assistive force to compensate for movement obstacles. The uncertainty factor β is calculated based on the standard deviation of the motion trajectory prediction and mapped to the [1,2] interval by the tanh function to quantify the prediction risk. When the prediction uncertainty is high (standard deviation increases), the β value increases, and the system enhances motion stability by increasing joint stiffness, reducing the safety risk caused by prediction errors. The basic stiffness is dynamically generated by a 3-layer MLP network. This network not only inputs the current intention confidence and fatigue index but also incorporates historical states (including kinematic data from the past 500ms), enabling the stiffness adjustment to have temporal continuity and memory. For example, when an increase in a user's muscle fatigue index is detected, the network proactively reduces the baseline stiffness to prevent excessive resistance from exacerbating fatigue; conversely, when the confidence level of intent increases, the baseline stiffness increases accordingly to provide more precise motion guidance. Ultimately, joint stiffness is adjusted synergistically by integrating baseline stiffness, the assist ratio, and uncertainty factors. This ensures that when user engagement is high and predictions are reliable, stiffness is reduced to preserve degrees of freedom; and when engagement is low or predictions are uncertain, stiffness is increased to ensure motion safety. Damping is calculated based on a critical damping ratio ξ=0.7 and joint inertia to maintain the stability of the system's dynamic response and prevent oscillations or overshoot during movement. This variable-parameter impedance control mechanism deeply integrates the user's physiological state, prediction uncertainty, and motion intent, realizing a three-tiered control strategy of "safety safeguard - active guidance - personalized adaptation." This effectively compensates for motor dysfunction and promotes the recovery of active movement ability, significantly improving the personalization and intelligence of rehabilitation training.

[0104] Optionally, the method further includes an online continuous learning step:

[0105] Maintain a circular buffer with a capacity of N=1000 samples to store the most recently trained data;

[0106] Perform a mini-batch update after every K=100 training cycles;

[0107] Elastic weights are used to prevent catastrophic forgetting; the loss function is... ,in For the diagonal elements of the Fisher information matrix, .

[0108] This embodiment achieves dynamic adaptation and continuous performance optimization of the control system during long-term use. A circular buffer maintains the most recent 1000 training samples using a first-in, first-out (FIFO) principle, ensuring the timeliness and representativeness of the stored data, covering recent changes in the user's movement patterns, such as changes in force exertion characteristics due to muscle strength improvement during rehabilitation or movement patterns introduced by new training movements. After every 100 training cycles, the system randomly selects a small batch of samples (e.g., 32 samples / batch) from the buffer for parameter updates, achieving incremental learning and avoiding the waste of computational resources caused by processing large amounts of historical data at once. The Elastic Weight Fixation (EWC) mechanism effectively mitigates the catastrophic forgetting problem of neural networks by introducing a regularization term into the loss function. Specifically, the diagonal elements of the Fisher information matrix... Quantify the importance of each parameter to the historical task, for key parameters of motion intention prediction (such as attention weights of temporal self-attention units and gating vectors of cross-modal fusion). For high values, the regularization term imposes a greater constraint on the parameter updates, ensuring that these parameters do not fluctuate drastically when learning new data; while for minor parameters (such as bias terms in some fully connected layers). The lower value allows it to adapt more flexibly to new data. The weighting coefficients balance the priority between learning new tasks and retaining old knowledge. For example, when a user transitions from initial passive assisted training to active resistance training, the system continuously learns online, retaining its basic motor intention prediction capabilities while gradually optimizing the recognition accuracy of high-intensity active force exertion patterns, without causing a decline in predictive performance for early low-intensity movement patterns due to learning new data. This online learning mechanism enables the rehabilitation trainer robot to dynamically evolve alongside the user's rehabilitation process, continuously matching changes in the user's motor abilities. This avoids the performance degradation problem that occurs in traditional fixed models over long-term use, significantly extending the system's clinical applicability and personalized rehabilitation effects.

[0109] Optionally, the method further includes a federated learning aggregation step:

[0110] Aggregate model parameters from multiple clients in the cloud, based on the data volume of each client. Calculate initial weights ;

[0111] Clients are clustered by feature statistics using K-means clustering. Divide into K clusters and calculate the intra-cluster similarity. ;

[0112] Generate weighted weights Then perform intra-cluster model aggregation, and then sort by performance metrics. Perform global aggregation;

[0113] After receiving the global model, the client integrates it with the local model in a personalized manner, with an integration coefficient α=0.3.

[0114] This embodiment achieves secure and efficient utilization of multi-center data and collective intelligence enhancement of model performance. Under the federated learning architecture, each hospital or rehabilitation institution acts as a client, uploading only local model parameters rather than raw training data, fundamentally protecting patient privacy (such as sensitive physiological data like electromyography signals and movement trajectories) and avoiding compliance risks associated with cross-institutional data transmission. The cloud server first calculates initial weights based on the amount of data from each client, ensuring a positive correlation between data contribution and model influence weights, preventing small-sample clients from being marginalized. K-means clustering is used to group clients according to feature statistics (mean, standard deviation), for example, clients with similar types of motor dysfunction (e.g., stroke patients with upper limb spasticity compared to spinal cord injury patients) or similar sensor distribution characteristics are grouped into the same cluster, with intra-cluster similarity... Quantization is performed using a Gaussian kernel function; a higher value indicates a greater similarity in the data distributions between the two clients. Weighted weights are then generated based on this similarity. This allows similar client parameters within a cluster to receive higher weights during aggregation, reducing aggregation bias caused by data heterogeneity. After intra-cluster aggregation, the performance metrics of each cluster model on the independent validation set (such as intent prediction accuracy) are then used to determine the aggregation bias. A global secondary aggregation is performed to ensure that the high-quality parameters of the high-performance cluster dominate the global model. After receiving the global model, the client performs personalized fusion with the local model using a fusion coefficient of α=0.3. This preserves the adaptability of the local model to specific user groups while absorbing the generalization ability of the global model. For example, rehabilitation center A focuses on early-stage rehabilitation patients, and its local model excels at recognizing subtle movement intentions; center B focuses on mid-to-late-stage rehabilitation patients, and its model predicts high-intensity movement patterns more accurately. Through federated learning, the parameters of the two center models are clustered and aggregated. The global model retains its sensitivity to capturing subtle intentions while improving the robustness of high-intensity movement prediction. After fusion by the client, it can simultaneously serve patients at different stages of rehabilitation within the same center. This distributed learning mechanism breaks down data silos, enabling the model to learn universal rehabilitation patterns from multi-source heterogeneous data. At the same time, by balancing global performance and local adaptability through hierarchical aggregation and personalized fusion, it provides a safe and efficient solution for performance optimization of rehabilitation training robots in multi-center clinical applications.

[0115] Experimental verification:

[0116] A. Experimental Setup

[0117] 1) Dataset: The evaluation was conducted on 10 stroke patients and 5 spinal cord injury patients. Multimodal data were collected, including 16-channel EEG (250Hz), 8-channel sEMG (1000Hz), 6-dimensional force / torque, and 7-dimensional joint angle (100Hz). The dataset was divided into 70% training, 15% validation, and 15% testing.

[0118] 2) Implementation details: Implemented using PyTorch 1.12, with a training batch size of 32, an initial learning rate of 1e-3, and weight decay of 1e-4. Deployed on the NVIDIA Jetson AGX Orin platform, using TensorRT for INT8 quantization and layer fusion optimization.

[0119] 3) Evaluation indicators:

[0120] Prediction accuracy: RMSE, MAE, R²

[0121] Timing alignment: Dynamic Time Warping Distance

[0122] Control performance: task completion rate, active contribution rate, motion smoothness

[0123] Real-time performance: inference latency, end-to-end latency

[0124] B. Performance Comparison

[0125] Table I: Comparison of Intent Prediction Accuracy

[0126] method RMSE↓ MAE↓ R²↑(%) DTW↓ Traditional EMG proportional control 0.287 0.198 71.2 0.156 LSTM-Single Mode 0.198 0.142 81.5 0.124 CNN+ Early Fusion 0.167 0.118 85.3 0.108 Transformer 0.145 0.105 88.7 0.095 This invention, MAFP-Net 0.089 0.067 93.2 0.068

[0127] Table II: Real-time Performance Comparison

[0128] index Traditional methods Method of the present invention Inference delay (ms) 45.3 4.8 Control cycle (ms) 50 9.6 Model size (MB) 128.5 18.3 Power consumption (W) 35.2 15.7

[0129] C. Clinical validation results

[0130] A 4-week training trial was conducted on 10 stroke patients to compare traditional EMG control with the method of this invention:

[0131] Improvement in motor function: The Fugl-Meyer score improved by 12.3 ± 2.1 points (p < 0.001), significantly higher than the 7.5 ± 1.8 points achieved by the traditional method.

[0132] Active participation: The assistance ratio α gradually decreased from an initial 0.32 to 0.68, indicating an increase in patients' active contribution rate.

[0133] Safety: Zero serious adverse events; false trigger rate reduced by 62% compared to traditional methods.

[0134] Interpretability: Visualization of attention weights showed that the contribution of neural signals in the early stages of rehabilitation was g_n = 0.52 ± 0.08. As training progressed, the contribution of electromyographic signals g_e increased to 0.41 ± 0.06, consistent with the clinical pattern of motor function recovery.

[0135] refer to Figure 2 This invention also provides a control system for a rehabilitation training robot, comprising:

[0136] At least one processor;

[0137] At least one memory for storing at least one program;

[0138] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0139] The content of the above method embodiments is applicable to this embodiment. The specific functions implemented in this embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. Therefore, they will not be repeated here.

[0140] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0141] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0142] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0143] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0144] This invention also provides a computer program product, including a computer program or computer instructions, which are stored in a memory. A processor of a computer device reads the computer program or computer instructions from the memory and executes the computer program or computer instructions, causing the computer device to perform the above-described method.

[0145] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0146] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0147] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0148] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A control method for a rehabilitation training robot, characterized in that, The method includes the following steps: S100, acquire data collected by multimodal sensors and a trained multimodal asynchronous fusion prediction network; wherein, the multimodal asynchronous fusion prediction network includes a multi-scale spatiotemporal encoder module, a cross-modal attention fusion module, a multi-task prediction head module, and a variable parameter impedance controller module; S200, the data collected by the multimodal sensor is input into the trained multimodal asynchronous fusion prediction network. The neural signals and electromyographic signals are processed by multi-scale temporal convolution, bidirectional gated recurrent unit processing and temporal self-attention calculation through the multi-scale spatiotemporal encoder module to obtain the encoded features. The biomechanical signals are compressed by the biomechanical feature extractor to obtain the biomechanical encoded features. S300, the cross-modal attention fusion module processes the encoded features and biomechanical encoded features to generate enhanced fusion features; the cross-modal attention fusion module calculates the cross-attention weights between modalities through linear transformations of queries, keys, and values, and dynamically adjusts the contribution of each modality using a gating mechanism; S400 processes the fused features through a multi-task prediction head module, outputting motion trajectory prediction, angular velocity prediction, interaction torque prediction, and intent confidence; the motion trajectory prediction uses Bayesian deep learning to quantify uncertainty and construct a diagonal Gaussian distribution. The S500 obtains the user's maximum torque, muscle activation level, and fatigue index based on a digital twin query. It then calculates the adaptive stiffness and damping matrix through a variable parameter impedance controller module, generates control commands, and outputs them to the robot actuator.

2. The method according to claim 1, characterized in that, In S200, the multi-scale spatiotemporal encoder module includes a parallel multi-scale temporal convolution unit, a bidirectional gated recurrent unit, a temporal self-attention unit, and a feature fusion and dimensionality reduction unit; The multi-scale temporal convolutional unit adopts 4-way parallel one-dimensional dilated convolution with dilation rates of 1, 4, 8 and 16, with each channel outputting 32 channels. After concatenation, it is fused by 1×1 convolution to obtain 128 channels of output. The bidirectional gated loop unit contains two layers of Bi-GRU, with 128 hidden units. The output dimension after splicing the forward and backward hidden states is 256. The temporal self-attention unit adopts an 8-head attention mechanism, with a head dimension of 16 and a key vector dimension of 64; The feature fusion and dimensionality reduction unit reduces the feature dimensionality to 128 dimensions through residual connections, layer normalization, and adaptive max pooling.

3. The method according to claim 1, characterized in that, In S300, the cross-modal attention fusion module is implemented in the following way: Neural coding features, electromyographic coding features, and biomechanical coding features are projected onto the query space, key space, and value space, respectively, with a projection dimension of 64 for each. Calculate the cross-attention weights of mode i to mode j. ,in =64; A four-head attention mechanism is used to compute modal relationships in different subspaces in parallel; Through learnable gating vectors Enhancement features from various modalities are dynamically fused. The gating vector is generated using the sigmoid function, and the distribution smoothness is controlled using a temperature parameter of 0.

5.

4. The method according to claim 1, characterized in that, In S400, the multi-task prediction head module includes a shared feature extraction layer and four parallel branches; The shared feature extraction layer consists of two fully connected layers with dimensions of 256 and 128, respectively, and employs ReLU activation, layer normalization, and regularization with a Dropout rate of 0.

3. The motion trajectory prediction branch outputs the mean of joint angle increments. Sum of logarithmic variance Construct a diagonal Gaussian distribution ; The angular velocity prediction branch outputs angular velocity. ; The interactive torque prediction branch outputs torque. ; The confidence branch of the intent outputs confidence ∈ [0,1].

5. The method according to claim 4, characterized in that, The multi-task prediction head module is trained using a composite loss function, which is: ； in, For mean square error loss, For negative log-likelihood loss, For variational attention KL divergence loss, For attention sparsity L1 regularization loss; .

6. The method according to claim 1, characterized in that, In S500, the variable parameter impedance controller module is implemented in the following way: Calculate the auxiliary ratio ,in The maximum torque predicted for the digital twin; Calculate the uncertainty factor ,in To predict the standard deviation of uncertainty; Basic stiffness The MLP network is dynamically generated based on intent confidence and fatigue index using a 3-layer MLP network, which is based on historical states. For additional input; Final joint stiffness Damping Where ξ = 0.7 is the critical damping ratio. This represents the joint inertia.

7. The method according to claim 1, characterized in that, The method also includes an online continuous learning step: Maintain a circular buffer with a capacity of N=1000 samples to store the most recently trained data; Perform a mini-batch update after every K=100 training cycles; Elastic weights are used to prevent catastrophic forgetting; the loss function is... ,in For the diagonal elements of the Fisher information matrix, .

8. The method according to claim 1, characterized in that, The method also includes a federated learning aggregation step: Aggregate model parameters from multiple clients in the cloud, based on the data volume of each client. Calculate initial weights ; Clients are clustered by feature statistics using K-means clustering. Divide into K clusters and calculate the intra-cluster similarity. ; Generate weighted weights Then perform intra-cluster model aggregation, and then sort by performance metrics. Perform global aggregation; After receiving the global model, the client integrates it with the local model in a personalized manner, with an integration coefficient α=0.

3.

9. A control system for a rehabilitation training robot, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

Images