A lower limb rehabilitation exoskeleton control method and system based on multi-joint torque synchronous prediction

Abstract

The application belongs to the technical field of exoskeleton control, and particularly discloses a lower limb rehabilitation exoskeleton control method and system based on multi-joint torque synchronous prediction, which comprises the following steps: collecting sEMG data and IMU data of a subject under different lower limb movement modes; calculating joint angles according to the IMU data; calculating joint torques of hip, knee and ankle joints according to the sEMG data combined with a Hill-type muscle-skeletal model; splicing the sEMG data and the joint angle data, and constructing a training set corresponding to the joint torques; training a deep learning model through the training set to obtain a lower limb multi-joint torque prediction model; and based on the lower limb multi-joint torque prediction model, real-time prediction of joint torques of the hip, knee and ankle joints is carried out, which is used as a core feedback signal of a user's movement intention for rehabilitation exoskeleton control. The application can realize individualized, compliant and safe rehabilitation exoskeleton control.

Description

Technical Field

[0001] This invention belongs to the field of exoskeleton control technology, and more specifically, relates to a method and system for controlling a lower limb rehabilitation exoskeleton based on synchronous prediction of multi-joint torques. Background Technology

[0002] Loss of motor function is the most common problem caused by neurological diseases such as stroke, Parkinson's disease, and cerebral palsy. Clinical studies have shown that continuous and effective rehabilitation therapy helps to restore neuroplasticity in the brain and regain functional abilities in daily living activities. In recent years, robot-assisted rehabilitation has provided a promising solution for the rehabilitation and recovery of people with disabilities.

[0003] When exoskeletons assist users in rehabilitation, accurately sensing the wearer's movement intentions is crucial for improving human-machine collaboration and enhancing user engagement. Related research mainly focuses on two directions: first, discrete motion pattern recognition, used to trigger different control modes of the exoskeleton, such as classifying movements like walking, standing, and sitting, or dividing the phase of the gait cycle; second, continuous motion intention prediction, such as real-time estimation of joint angles, angular velocities, or torques, to achieve more natural and smoother human-machine interaction. In particular, accurate joint torque estimation can avoid muscle compensation or insufficient force due to excessive exoskeleton assistance, thus affecting training effectiveness and ensuring the safety of the rehabilitation process. Joint torque estimation methods mainly include the following three categories: (1) Model-based estimation methods. These methods rely on pre-set physical or biomechanical models to calculate torque, specifically including: 1) Laboratory measurement method based on multibody inverse dynamics: This method requires the combination of three-dimensional kinematic data and ground reaction force data to calculate joint torque in a dedicated laboratory. Although considered the "gold standard," it cannot be widely applied to real-time estimation or non-laboratory environments due to its reliance on biomechanical models of marker points, specialized equipment, and laboratory environment; 2) Calculation method based on classical mechanical models: This method uses the Newton-Euler equation and Lagrange equation as its core, combined with joint angles / angular velocities measured by encoders, and parameters such as the mass and length of human limbs to calculate torque. No additional sensors are required, but the accuracy is highly dependent on the model's accuracy, and individual parameter differences or model simplification can easily lead to errors; 3) Model estimation method based on disturbance observers: This method treats joint torque as an external disturbance of the system and estimates it through a disturbance observer, without the need for additional force sensors. However, in practical applications, the output of the nonlinear observer will be superimposed with the unmodeled dynamic characteristics of the system, affecting the estimation accuracy. 4) Estimation based on the neuromuscular skeletal (NMS) model: Estimation is achieved by establishing a mapping relationship between electromyography (EMG) signals and joint torque using the NMS model. However, some model parameters are difficult to obtain, the calibration process is time-consuming and labor-intensive, and the signal quality of sEMG is easily affected by noise.

[0004] (2) Sensor-based direct / indirect measurement methods. Direct measurement methods use torque sensors, which are accurate, but increase the size and cost of the exoskeleton and are susceptible to interference from installation location and impact. Indirect measurement methods, such as calculating joint torques through human-computer interaction force sensors and kinematic data, are less expensive, but require overcoming the influence of sensor noise.

[0005] (3) Neural Network Mapping Method. This method does not require the construction of a complex human biomechanical model. It directly establishes a mapping relationship between wearable sensor inputs and joint torque outputs, and has become the mainstream technical path in the field of real-time joint torque prediction. Its core is to use signals from wearable sensors such as sEMG and IMU as model inputs. It can be mainly divided into shallow machine learning-based methods and deep learning-based methods. The former includes support vector machines, random forests, etc., but these methods require manual feature extraction and are limited by prior knowledge and expert experience. Deep learning can automatically extract key features from massive amounts of data, reducing the subjectivity and uncertainty brought about by manual operation.

[0006] However, existing methods for predicting joint torques based on neural network mapping still face the following challenges: On the one hand, joint torque, as a dynamic variable that is difficult to measure directly through wearable devices, is difficult to accurately label, which affects the quality of model training and prediction accuracy. On the other hand, surface electromyography signals can reflect the muscle activation state, and inertial measurement unit signals can reflect the limb kinematic state. The two have different data characteristics and temporal dynamic characteristics. How to achieve effective fusion of multimodal information and establish a high-precision mapping from fused features to joint torque is still an urgent problem to be solved.

[0007] Furthermore, exoskeleton control strategies directly impact patients' active participation levels and rehabilitation training effectiveness. Traditional passive control methods typically involve a robot guiding the patient through training movements along a pre-set trajectory, suitable for patients with severely impaired motor function or those in the early stages of rehabilitation. While this approach enables basic motor function training, it doesn't fully consider the patient's voluntary movement intentions. Long-term use may reduce the patient's active participation, even leading to muscle disuse atrophy or dependence on exoskeleton assistance, which is detrimental to stimulating neuroplasticity. In contrast, active intention-driven human-machine collaborative control strategies provide appropriate assistance based on the patient's residual motor ability and real-time exertion status, encouraging active exertion while ensuring exercise safety, thereby improving training participation and rehabilitation motivation.

[0008] Therefore, how to achieve high-precision, personalized, compliant, and safe auxiliary control has become an important technical problem that urgently needs to be solved in the field of rehabilitation exoskeletons. Summary of the Invention

[0009] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a lower limb rehabilitation exoskeleton control method and system based on multi-joint torque synchronous prediction, the purpose of which is to achieve personalized, compliant and safe high-precision rehabilitation exoskeleton control.

[0010] To achieve the above objectives, according to a first aspect of the present invention, a method for controlling a lower limb rehabilitation exoskeleton based on synchronous prediction of multi-joint torques is proposed, comprising: Model building phase: sEMG and IMU data were collected from subjects under different lower limb movement patterns; The joint angles are calculated from IMU data; muscle activation is calculated from sEMG data, and then the joint torques of the hip, knee and ankle joints are calculated using the Hill musculoskeletal model. The sEMG data and joint angle data were concatenated and divided into sample data according to time. The joint torque corresponding to the time of the sample data was used as the corresponding label to construct the training set. The deep learning model is trained using a training set, and the trained deep learning model is the lower limb multi-joint torque prediction model. Model application phase: The system acquires sEMG and IMU data of the subject during exercise in real time, calculates the joint angles through the IMU data, and then inputs the sEMG data and joint angle data into the lower limb multi-joint torque prediction model to predict the joint torques of the hip, knee and ankle joints in real time. Rehabilitation exoskeleton control is performed based on real-time predicted joint torques of the hip, knee, and ankle joints.

[0011] As a further preferred embodiment, the sEMG data includes sEMG signals from the rectus femoris, vastus medialis, biceps femoris, and tibialis anterior muscles of the subject's leg.

[0012] As a further preferred option, the joint moments of the hip, knee, and ankle joints are calculated using a Hill-type musculoskeletal model, including: Based on the Hill-type musculoskeletal model, the active and passive muscle forces of the target muscle are calculated according to the muscle activation of the target muscle. Then, by combining the active and passive muscle forces, the muscle-tendon unit output force of the target muscle is obtained. The target muscle is the rectus femoris, vastus medialis, biceps femoris, or tibialis anterior. The joint torque is calculated based on the output force of the muscle-tendon unit:

[0013] in, Indicates the first j Joint torque of each joint j Indicates the hip, knee, or ankle joint; Indicates the first i The force output by the muscle-tendon unit of the target muscle; Indicates the first i Block target muscle relative to the first j The instantaneous lever arm of each joint; t Indicates time, To participate in the j The number of muscles calculated from the joint torque.

[0014] As a further preferred option, the muscles involved in calculating the joint torque of the hip joint are the rectus femoris and biceps femoris, the muscles involved in calculating the joint torque of the knee joint are the rectus femoris, biceps femoris and vastus medialis, and the muscle involved in calculating the joint torque of the ankle joint is the tibialis anterior.

[0015] As a further preferred option, the output force of the muscle-tendon unit of the target muscle. The formula for calculation is:

[0016] , , in, For active muscle strength, For passive muscle force, φ The pennate angle of the muscle fiber; f a To normalize the active force-length relationship function, f v To normalize the active force-velocity relationship function, l m , v m These are normalized muscle fiber length and normalized muscle fiber velocity, respectively. a ( t () represents muscle activation level. F max This represents the maximum isometric contractile force of the muscle. f p This is the normalized passive elastic force-length relationship function.

[0017] As a further preferred method, muscle activation is calculated based on sEMG data, including: Neural activation was calculated based on sEMG data. u ( t ):

[0018] in, express td sEMG data at any given time dFor time delay; u ( t ), u ( t -1) u ( t -2) respectively represent t , t -1、 t Neural activation at time -2; α , β 1. β 2 is the proportionality coefficient, which satisfies intermediate parameters | 1|<1、| 2|<1; Furthermore, based on neural activation level u ( t Muscle activation level was obtained. a ( t ):

[0019] in, A It is a non-linear shape factor. A The value ranges from 0 to 3.

[0020] As a further preferred embodiment, the deep learning model includes a convolutional neural network module, a spatial and channel reconstruction convolutional module, a gated recurrent unit module, and a multi-head attention mechanism module connected in sequence. The convolutional neural network module extracts local features of the spatial dimension from the input data to obtain features. Figure 1 The spatial and channel reconstruction convolutional module reconstructs features from both spatial and channel dimensions. Figure 1 Reconstruction is performed to obtain features. Figure 2 The gated loop unit module has specific features. Figure 2 Extract the dynamic dependencies between consecutive time steps to obtain features. Figure 3 Multi-head attention module based on features Figure 3 The temporal features are weighted and fused from multiple feature subspaces to obtain fused features; the predicted joint torque is output based on the fused features.

[0021] As a further preferred option, when controlling the rehabilitation exoskeleton, a dual-loop control structure including an outer loop and an inner loop is adopted. The outer loop converts the predicted joint torque into the desired motion trajectory increment and dynamically adjusts the target trajectory of the exoskeleton. The inner loop, based on the target trajectory, uses model-free adaptive terminal sliding mode control combined with time delay estimation to achieve rehabilitation exoskeleton control.

[0022] According to a second aspect of the present invention, a lower limb rehabilitation exoskeleton control system based on multi-joint torque synchronous prediction is provided, comprising a processor for executing the above-described lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction.

[0023] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the above-described lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction.

[0024] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. This invention utilizes surface electromyography (sEMG) data, combined with a Hill-type neuromuscular skeletal model, to generate high-quality joint torque labels. This solves the problem of scarce training labels caused by the inability to directly measure dynamic parameters, thus improving the effectiveness and reliability of model training. Furthermore, through a deep learning model, using spliced ​​sEMG and joint angle data as input, it achieves synchronous and high-precision prediction of lower limb multi-joint torques. The real-time predicted joint torques are used as the core feedback signal for the user's movement intention, enabling personalized, compliant, and safe rehabilitation exoskeleton control.

[0025] 2. The deep learning model constructed in this invention can extract features in both spatial and temporal dimensions. It utilizes spatial correlation information and temporal dynamic information from electromyography (EMG) signals and joint angle data, and incorporates spatial and channel reconstruction convolution and multi-head attention mechanisms. This enhances feature representation capabilities while reducing model computational overhead and improving the model's ability to extract key features and long-distance correlation information. This deep learning model can fully integrate sEMG and kinematic multimodal information without complex manual feature extraction. With low computational resource consumption, it achieves synchronous and high-precision prediction of hip, knee, and ankle joint moments, overcoming the technical bottleneck of incomplete single-modal input information and limited prediction accuracy.

[0026] 3. This invention uses the predicted joint torque as the core feedback signal of the user's movement intention and embeds it into the exoskeleton control closed loop, realizing true intention-driven active-passive coordinated control, which significantly improves the patient's active participation and rehabilitation enthusiasm; the inner loop adopts a strategy that combines model-free adaptive terminal sliding mode control with time delay estimation, which has strong robustness and can maintain good trajectory tracking performance even in the presence of external disturbances and system uncertainties. Attached Figure Description

[0027] Figure 1 This is a flowchart of the lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction, according to an embodiment of the present invention.

[0028] Figure 2 This is a control framework diagram of a lower limb rehabilitation exoskeleton based on synchronous prediction of multi-joint torque, according to an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram illustrating the acquisition of sagittal lower limb joint torque data of a subject using a Hill-type neuromuscular skeletal model, as an embodiment of the present invention. Figure 4 This is a schematic diagram of the lower limb multi-joint torque prediction model structure according to an embodiment of the present invention, wherein (a) is the model structure, (b) is the convolutional neural network module, (c) is the gated recurrent unit module, (d) is the spatial and channel reconstruction convolutional module, and (e) is the multi-head attention mechanism module.

[0030] Figure 5 This is a schematic diagram of the detection of active segments of subjects in different lower limb movement modes according to an embodiment of the present invention, wherein (a) is the SitTS-S movement mode; and (b) is the LW mode.

[0031] Figure 6 The following are the raw data of the subjects in the SitTS-S exercise according to the embodiments of the present invention, wherein (a) is the raw sEMG data and (b) is the kinematic data.

[0032] Figure 7 This is a comparison chart of the actual values, model predictions, and filtered results of the torque at different joints in the flexion direction during the SitTS-S movement of the subjects in this embodiment of the invention.

[0033] Figure 8 This is a comparison chart of the prediction performance of different models in the SitTS-S motion according to an embodiment of the present invention.

[0034] Figure 9 This is a comparison chart of the prediction performance of the optimal model in the SitTS-S motion of this invention under different modes.

[0035] Figure 10 This is a comparison chart of the prediction performance of the CNN-SCRC-GRU-MHA model under different window lengths in the SitTS-S motion of this invention.

[0036] Figures 11 to 13 For each embodiment of the present invention, take k =0、 k =5、 k When the value is 10, the experimental results of the hip joint of the subject during the SitTS-S exercise are shown in the figure.

[0037] Figures 14 to 16 For each embodiment of the present invention, take k =0、 k =5、 k=10 results of knee joint experiment of subjects during SitTS-S exercise.

[0038] Figure 17 This is a comparison chart of the control performance of the control method of the present invention under different stiffness coefficients. Detailed Implementation

[0039] 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 and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] This invention provides a method for controlling a lower limb rehabilitation exoskeleton based on synchronous prediction of multi-joint torques, such as... Figure 1 and Figure 2 As shown, it includes the following steps: S1: Simultaneous acquisition of multimodal data. Using sEMG and IMU acquisition systems, sEMG (surface electromyography) and IMU data were simultaneously acquired from subjects during different lower limb movement modes, such as sitting, standing, walking, and leg raising and knee flexion.

[0041] In an optional embodiment, each subject independently completed five sit-stand-sit (SitTS-S) and level walk (LW) tasks. Surface electromyography (EMG) signals of four muscles in the right leg were acquired using a wireless sEMG acquisition system (RunE W2 EMG-Pro), including the rectus femoris (RF), vastus medialis (VM), biceps femoris (BF), and tibialis anterior (TA), at a sampling frequency of 500 Hz. An inertial measurement unit (IMU900) acquired kinematic data, including angles, velocities, and accelerations, in the pelvis, thigh, and lower leg, at a sampling frequency of 100 Hz.

[0042] S2: Joint torque calculation. The acquired sEMG and IMU data are aligned and preprocessed; then, based on the IMU data, the joint angles of the three joints are obtained through attitude calculation; muscle activation is obtained through sEMG data, and the joint torque data of the hip, knee, and ankle joints of the lower limb are calculated through the Hill musculoskeletal model.

[0043] Specifically, it includes: S201: Data alignment.

[0044] The sEMG and IMU data are aligned by timestamp and invalid data is removed. Then, the nearest neighbor principle is used to achieve asynchronous alignment between EMG signals, inertial measurement units (IMUs), and EMG-IMU data. The final output is time-synchronized and consistent-length EMG and IMU data frames for subsequent fusion analysis. The start and end timestamps of various movements are determined using IMU data, which enables the extraction of corresponding time periods based on time indices, thereby accurately extracting multi-channel data of seated extension-seat (SitTS-S) and standing walking (LW) activities.

[0045] In an optional embodiment, for seated extension-seated movements, the detection criteria for the start and end points of the active segment are as follows:

[0046]

[0047] in, θ τ Indicates the trunk angle in the sagittal plane; PP represents the peak-to-peak value; σ It is a coefficient ( σ >0), can be set to 0.05.

[0048] For standing walking, when the knee flexion angle of one lower limb reaches its peak value during the support phase (denoted as ( ) θ K When PP is reached, the starting and ending points are determined. Two consecutive such points mark a complete gait cycle. Figure 5 The diagram shows the activity segment detection of different lower limb movement patterns of the subjects.

[0049] S202: Data preprocessing.

[0050] The preprocessing of sEMG data includes: 1) resampling the sEMG data at 1000Hz; 2) using DC removal, stopband filter and bandpass filter to eliminate noise; 3) normalization using the positive and negative one method.

[0051] For IMU data, the joint angles are calculated using matrix transformation, and then filtered and normalized.

[0052] In an optional embodiment, the sEMG sampling frequency F SSet to 1000. The stopband filter specifically uses a notch filter to remove power frequency interference, with the following settings: 3rd order finite impulse response bandpass filter; centerline frequency: 50Hz; bandwidth: 2 Hz. The bandpass filter is used to filter out noise outside the sEMG spectrum range, with the following settings: 7th order Butterworth filter; low cutoff frequency: 20 Hz; high cutoff frequency: 450 Hz. For the filtered sEMG data... ,in M 1 indicates the number of sEMG sensors. L 1 represents the data length, using a positive and negative one method. x Normalize to the range [-1, 1]:

[0053] In an optional embodiment, the sagittal joint angles are solved by IMU attitude calculation, and the window size of the Savitzky-Golay filter in the angle data preprocessing (denoted as...) w l ) and polynomial order (represented as p o The values ​​are set to 7 and 3 respectively. Generally, w l It must be an odd number and must satisfy the following conditions: w l > p o .

[0054] S203: Muscle activation calculation.

[0055] The envelope of the preprocessed surface electromyography (sEMG) signal is obtained, and the neural activation level of the corresponding muscle is calculated based on the normalized sEMG signal. Specifically, considering the time delay between the sEMG signal and muscle contraction, a recursive filter is used to calculate the neural activation level. u (t):

[0056] in, d For time delay, α , β 1 and β 2 is the scaling factor of the model, and it satisfies the following constraints:

[0057] Among them, | 1|<1,| 2|<1.

[0058] In an optional embodiment, the following settings are provided: d =80ms β1 = -0.052, β 2 = 0.000627, α =0.9486.

[0059] Because there is a non-linear relationship between neural activation and muscle contractile force, neural activation and muscle activation... a ( t There is also a non-linear relationship between them:

[0060] in, A It is a non-linear shape factor that defines the curvature of a function, and its value is between 0 and... Between 3, A The smaller the value, the stronger the non-linear relationship between neural activation and muscle activation.

[0061] In an optional embodiment, the following settings are provided: A =-0.5.

[0062] S204: Calculation of output force of muscle-tendon unit.

[0063] Based on the Hill-type musculoskeletal model, the active muscle force of the target muscle is calculated according to muscle activation, muscle feather angle, muscle fiber length, muscle fiber velocity, and maximum isometric muscle force. F CE ( t Passive muscle strength F PE ( t This allows us to obtain the output force of the target muscle-tendon unit. F mu ( t ).

[0064]

[0065]

[0066]

[0067]

[0068] in, f a This is the normalized active force-length relationship function; f v This is the normalized active force-velocity relationship function; l , v These are normalized muscle fiber length and normalized muscle fiber velocity, respectively. l m This represents the current muscle fiber length. lm0 For optimal muscle fiber length; F max This represents the maximum isometric contractile force of the muscle. f p This is a normalized passive elastic force-length relationship function; φ It is the pennate angle of the muscle fiber.

[0069] In one embodiment, the effect of muscle fiber contraction velocity on active muscle force is ignored, and... f v ( v The normalized function of active muscle force-length and the normalized function of passive elastic force-length are expressed as follows:

[0070] Table 1. Parameters in the Hill-type musculoskeletal model

[0071] S205: Solving for joint torques.

[0072] For the hip, knee, and ankle joints, the output force and muscle lever arm of the target muscle-tendon unit acting on the corresponding joint are calculated respectively. R i Calculate the joint torques in the sagittal plane in the flexion and extension directions:

[0073] in, τ j Indicates the first j Joint torque of each joint j These can be used to represent the hip joint, knee joint, and ankle joint, respectively. Indicates the first i The force output by the muscle-tendon unit of the target muscle; Indicates the first i Block target muscle relative to the first j The instantaneous lever arm of each joint; To participate in the j The number of muscles calculated from the joint torque.

[0074] In this embodiment, the joint moments of the hip, knee, and ankle joints in the sagittal plane in the flexion and extension directions of the right leg are calculated as reference values ​​or labels for subsequent prediction models. The correspondence between each muscle and joint moment can be described as follows: hip joint (rectus femoris RF, biceps femoris BF), knee joint (rectus femoris RF, biceps femoris BF, vastus medialis VM); ankle joint (tibialis anterior TA).

[0075] The obtained joint torque data can be maximized (represented as...). max tNormalization is performed to avoid introducing normalization bias due to differences between the training and test sets; the processed multimodal signals are then spliced ​​along the channels. In an optional embodiment, max t Set it to 140.

[0076] S3: Offline model training.

[0077] The preprocessed sEMG data and joint angle data were spliced ​​together by channel and divided into equal-length samples. The corresponding joint torque data were used as labels to construct training and testing sets. A lower limb multi-joint torque prediction model was constructed based on a deep learning model, and the model was trained using the training set.

[0078] S301: Sample and dataset partitioning.

[0079] The processed sEMG data and joint angle data are concatenated column-wise, and the concatenated multimodal data is divided into equal-length samples, labeled with the preprocessed reference joint torque data for the corresponding time interval. Training and test sets are divided according to a preset ratio. When dividing the samples, the training set uses an overlapping window method (preferably a window length of 48ms and an overlap length of 12ms), while the test set uses a non-overlapping window method (window length of 48ms and an overlap length of 0ms), adapting to real-time control scenarios.

[0080] In an optional embodiment, firstly, for a specific target subject... S The sitting-standing-sitting method represents the preprocessed sEMG data as follows: ,in M 1 = 4 (meaning 4 channels of sEMG data are used). L Indicates the data length, representing the preprocessed joint angle data as... ,in M 2 = 3 (indicating the use of 3-channel joint angle data, i.e., the flexion angles of the hip, knee, and ankle joints in the sagittal plane), preprocessed sEMG and joint angle data are stitched together by channel (column) to obtain a single data set of size [size missing]. Two-dimensional matrix The preprocessed joint torque data is expressed as follows: , as a label for the prediction model, where M 3 = 3 (indicating the use of 3-channel joint moment data, i.e., the flexion moments of the hip, knee, and ankle joints in the lower limb in the sagittal plane). The resulting two-dimensional matrix is ​​divided into a sample set using a sliding windowing method, where the lengths of the analysis window and the overlapping window are denoted as... L S and L O Optionally, σSetting it to 0.8 means that the training set comprises 80% of the total samples. For the training sample set, L S and L O Setting the time to 48ms and 16ms respectively minimizes the similarity between samples while increasing the number of training samples, thus enhancing the reliability of the test results. For the test sample set, L S and L O The settings were set to 48ms and 0ms respectively, which is non-sliding windowing.

[0081] S302: Construction and offline training of a multi-joint torque prediction model for the lower limbs.

[0082] The lower limb multi-joint torque prediction model was trained offline using a training set. The input to the model was a stitched data set of sEMG and joint angles, and the output was the joint torques of the hip, knee, and ankle joints. Figure 4 As shown, the lower limb multi-joint torque prediction model is a CNN-SCRC-GRU-MHA model, which includes a convolutional neural network module, a spatial and channel reconstruction convolutional module, a gated recurrent unit module, and a multi-head attention mechanism module connected in sequence, wherein: The Convolutional Neural Network (CNN) module extracts local spatial features from the input data; the Spatial and Channel Reconstruction Convolutional (SCRC) module reconstructs features from both spatial and channel dimensions to suppress redundant features; the Gated Recurrent Unit (GRU) module models the dynamic dependencies between consecutive time steps; the Multi-Head Attention (MHA) module weightedly fuses temporal features from multiple feature subspaces to obtain more discriminative fused features; and finally, a fully connected layer obtains the predicted joint torque based on the fused features. Through this structure, the model can simultaneously utilize spatial correlation information and temporal dynamic information from electromyography (EMG) signals and joint angle data to achieve synchronous prediction of lower limb multi-joint torques.

[0083] Specifically, Spatial and Channel Reconstruction Convolution (SCRC) is an attention-enhanced convolutional module that combines spatial and channel information reconstruction. Its structure consists of spatial reconstruction units and channel reconstruction units connected in series, as shown in the diagram. Specifically, the spatial reconstruction unit first measures the effective information contribution of each feature map in the spatial dimension through feature separation and reduces redundant responses in the spatial dimension through the reconstruction process. The channel reconstruction unit, on the other hand, compresses and integrates redundant information between channels through operations such as channel grouping, feature transformation, and information fusion, thereby improving feature representation capabilities while reducing model computational overhead.

[0084] Multi-head attention (MHA) is an attention computation module used to enhance feature representation capabilities. It maps and models the relationships between different subspaces of input features using multiple parallel attention heads. Specifically, the input features are first transformed linearly to generate a query matrix, a key matrix, and a value matrix. Then, each attention head calculates an attention weight based on the similarity between the query and key matrices, and uses this weight to perform a weighted summation of the value matrix, thus obtaining contextual representations under different feature subspaces. The outputs of multiple attention heads are further concatenated and linearly mapped to obtain the final fused features. Through this approach, MHA can simultaneously capture the dependencies between features from multiple representation subspaces, enhancing the model's ability to extract key features and long-distance correlation information, thereby improving the model's prediction accuracy and generalization ability.

[0085] In an optional embodiment, for all four models, the input layer is... A two-dimensional matrix of size, with all output layers being [size missing]. A two-dimensional matrix of size represents the joint moments of the three joints of the lower limb. In an optional embodiment, M = M 1+ M 2 = 12 L S = 96. Specifically, two CNN blocks are used for spatial feature extraction, two GRU layers are used for temporal feature extraction, and three fully connected layers are used as predictors to perform dimensionality transformation on the last output neuron of the second GRU layer. A CNN block is defined as consisting of a one-dimensional convolutional layer, an activation function layer, a batch regularization layer, a pooling layer, and a dropout layer. Optionally, the invention uses the "PReLU" activation function. The invention designs a symmetric spatiotemporal attention mechanism, namely, connecting a plug-and-play SCRC attention module after the CNN, enabling the model to selectively focus on important information from different modalities and feature channels, and adapting to the fusion of arbitrary single-modal or multi-modal signals; a multi-head attention mechanism module is connected after the GRU to improve the GRU's performance in capturing long-range dependencies in sequence data processing.

[0086] The model training parameters are shown in Table 2. The training set is used to train the model, and the test set samples are input into the trained model for evaluation.

[0087] Table 2 Training parameters of the lower limb multi-joint torque prediction model

[0088] Optionally, the root mean square error (RMSE) and the coefficient of determination (CQD) can be used. R 2 The predictive performance of the model is evaluated using RMSE. RMSE reflects the predicted joint moments. Compared with actual joint torque labels The smaller the RMSE value, the better the prediction performance.

[0089] R 2 This reflects the proportion of all variables in the dependent variable that can be explained by the independent variables through the regression relationship, among which... R 2 The closer the value is to 1, the better the prediction performance.

[0090] S4: Online model prediction.

[0091] Real-time acquisition and preprocessing of sEMG and IMU data from subjects were performed. The processed sEMG data and joint angles were then stitched together and input into a trained lower limb multi-joint torque prediction model to obtain three-joint torque prediction results, represented as follows: ,in L ' represents the length of the predicted joint torque, and the result is smoothed by filtering.

[0092] In an optional embodiment, max t Set to 140. Therefore, the actual predicted joint torque value should be expressed as: Optionally, for Perform moving average filtering to smooth the signal, and obtain The filter window size is set to... .

[0093] S5: Exoskeleton Cooperative Control.

[0094] A collaborative intention control method based on joint torque prediction, ICC-EJTP, is proposed. This method integrates the smoothed predicted joint torques output by the model into a dual-loop control structure of the exoskeleton, representing the user's motion intention. The outer loop employs a reduced-order admittance algorithm to map the predicted joint torques into motion trajectory increments, which are then incorporated into the control loop to reflect the user's active motion intention. The inner loop uses model-free adaptive terminal sliding mode control combined with time delay estimation technology to compensate for system dynamic uncertainties and achieve precise and compliant control of the exoskeleton.

[0095] Specifically, for a 2-DOF exoskeleton with hip and knee joints, the dynamic characteristics of the model-free system are as follows:

[0096] in q , , ( t )∈ 2 The angular position coordinates, angular velocity, and angular acceleration of the exoskeleton are not represented separately. τ ( t ) and τ h ( t ) ∈ 2 These represent the control input and the joint torque vector, respectively. d ( t ) ∈ 2 This represents the total disturbance affecting the system, including uncertain external disturbances and joint torque estimation errors; α ∈ 2×2 This represents a diagonal matrix.

[0097] The lumped uncertainty disturbances of the system can be described as follows: . accurate estimation d ( t This is crucial for implementing model-free controllers. Therefore, TDE (Transform-by-Determination) techniques are employed to estimate the changes in the system caused by lumped uncertainties. , , , in, ( t ) indicates concentrated interference in the assessment; ε Indicates the data sampling period; τ ( t - ε )and ( t - ε These represent the input torque and acceleration at the previous time step, respectively. In practical applications, d ( t It changes continuously over time. When the delay interval... ε Sufficient hours, through use d ( t-ε It can be effectively estimated d (t ).in, ( t The ) represents the error in estimating the disturbance. The limit of the disturbance estimation error is defined as where η >0 limits the maximum value of the estimated interference error.

[0098] Joint torque decoding using the CNN-SCRC-GRU-MHA model represents the human's active movement intention. Therefore, a reduced-order admittance algorithm is employed to convert the estimated joint torque into joint angle increments to ensure the exoskeleton exhibits appropriate compliant behavior. Where Δ q ( t This represents the angular change caused by the predicted lower limb joint torque. k It is a constant representing the stiffness coefficient. The new expected trajectory, incorporating the subject's perceived motion intention, is defined as follows: , , in q ref(t) , q des(t) These represent the predetermined reference trajectory and the generated desired trajectory, respectively.

[0099] The relevant error variables are defined as follows:

[0100] Design an adaptive terminal sliding mode control algorithm, with its sliding surface s The definition is as follows:

[0101]

[0102] in p and q All values ​​are positive constants and are odd numbers to ensure convergence within a finite time. The adaptive update law is designed as follows: , Where sign(•) represents the sign function, and ξ(t) represents the gain adjusted online.

[0103] The following are specific examples: To verify the effectiveness of the rehabilitation exoskeleton coordinated control method based on synchronous prediction of lower limb multi-joint torque proposed in this invention, this embodiment selected 7 subjects (denoted as S1-S7; 28.43±3.17 years old, 172.43±8.37 cm, 60.59±5.71 kg) to conduct experimental verification in the sit-to-stand-to-sit (SitTS-S) movement of the lower limbs. The data acquisition system adopted: 1) The surface electromyographic signals of four muscles of the right leg were collected using a wireless sEMG acquisition system (RunE W2 EMG-Pro), including: rectus femoris (RF), vastus medialis (VM), biceps femoris (BF), and tibialis anterior (TA), with the sampling frequency set to 500 Hz; 2) Kinematic data of the pelvis, thigh, and calf were collected using IMUs, with the sampling frequency set to 100 Hz.

[0104] The specific implementation steps are as follows: (1) Data acquisition and preprocessing Raw sEMG and kinematic data of 7 subjects were obtained during the sit-stand-sit exercise, such as Figure 6 As shown. Joint torque data of the lower limbs of each subject were obtained.

[0105] (2) Experimental results The actual values, model predictions, and filtered results of the torques at different joints of the lower limb in the flexion direction during the SitTS-S movement of subject S1 were visualized, such as... Figure 7 As shown. By Figure 7 It can be seen that the method provided by this invention can predict joint torques relatively accurately.

[0106] (3) Ablation test First, taking the SitTS-S exercise as an example, to illustrate the superiority of the deep learning model CNN-SCRC-GRU-MHA of this invention, other deep learning models and traditional machine learning models, including Support Vector Machines (SVM), Random Forest (RF), and Gaussian Process Regression (GPR), were compared. The test results of different models on 7 subjects under the SitTS-S exercise are as follows: Figure 8 As shown.

[0107] Depend on Figure 8It can be seen that: 1) Among different deep learning models, the CNN-SCRC-GRU-MHA model of this invention achieved the best results; 2) Among the three MLP models, the RFR model performed better.

[0108] Then, using the CNN-SCRC-GRU-MHA and RF models, and again taking the SitTS-S motion as an example, we tested their prediction performance under different types of input data. The results are as follows: Figure 9 As shown. By Figure 9 It can be seen that both models perform best under fused signal input, followed by joint angle signal input, while the performance is worst when using only sEMG signals. This is because joint torque is mainly determined by the joint kinematic information of the base IMU, while the relationship between sEMG and joint torque is more of an indirect physiological correlation. The fused signal contains both kinematic and muscle activation features, thus exhibiting a higher correlation with the target torque and achieving better estimation results.

[0109] Finally, using the CNN-SCRC-GRU-MHA model, and again taking SitTS-S motion as an example, we tested its prediction performance under different window lengths. The results are as follows: Figure 10 As shown. By Figure 10 It can be seen that: 1) In L O = 0.25 L S Under these conditions, with L S As the value increases, the model's predictive performance first increases and then decreases. L S = 48 and L O = 12, prediction performance is optimal; 2) in L S = 48, with L O As the value decreases, the model's predictive performance gradually declines; 3) Smaller values L S The value implies a shorter sampling interval and processing delay, therefore... L O Set as L S A 25% ratio is more reasonable, as it balances model performance with training efficiency and testing reliability.

[0110] (4) Intention-driven rehabilitation exoskeleton collaborative control For patients who have partially regained motor function, incorporating voluntary effort into rehabilitation training helps encourage their active participation. Therefore, an exoskeleton trajectory tracking experiment was conducted using the proposed intention-based collaborative control method.

[0111] Specifically, using Matlab's curve fitting tool, curves are extracted for the reference trajectories of the three joints based on multiple SitTS-S motions.

[0112]

[0113] The following metrics were established: the mean absolute error between the reference trajectory and the actual trajectory (MAE-RA), the mean absolute error between the expected trajectory and the actual trajectory (MAE-DA), and the auxiliary level (AL), which is the ratio of MAE-RA to the root mean square value of the predicted torque:

[0114]

[0115]

[0116] in, q i , ref_j , q i,des_j and q i,j Representing the first i Reference angle, expected angle, and actual angle of each joint; τ i,h_j Indicates the first i Predicted human torque for each joint; j , N Representing the first of each angle or torque j Individual data points and total number of data points.

[0117] The joint torques predicted by CNN-SCRC-GRU-MHA represent the active motion intention of the human body, which is used to regulate the desired trajectory via a reduced-order admittance controller, where the stiffness parameter... k Set to 0, 5, and 10. Note that when... k When = 0, the controller is in passive training mode; when k When the value is greater than 0, it switches to active assisted training mode.

[0118] Figures 11 to 13 The experimental results on the S1 hip joint of the subjects were presented. Figures 14 to 16 The experimental results on the S1 knee joint of the subjects are presented. Specifically, Figure 11 and Figure 14 Corresponding to k The stiffness value is 0. Figure 12 and Figure 15 Corresponding to k =5 stiffness value, Figure 13 and Figure 16 Corresponding to k =10 stiffness value; Figures 11 to 16 The first line shows the difference between the reference trajectory, the expected trajectory updated according to the human's movement intention, and the actual feedback trajectory. The second line compares the deviation between the expected trajectory and the actual trajectory. The third line shows the estimated joint torque, and the last line depicts the control signal sent to the motor. It can be seen that: 1) In passive control mode, the active torque of the hip and knee joints is relatively high, reaching maximum values ​​of 22 N·m and 120 N·m, respectively. In contrast, in active control mode, the joint torques of the hip and knee joints are limited to the ranges of [-20, 25] N·m and [-15, 23] N·m, respectively. This difference arises because in passive mode, since the human's movement intention is not included in the control system, this inconsistency is amplified. Conversely, in active mode, the interaction between the human and the exoskeleton helps reduce the difference between the two. 2) The tracking error between the expected trajectory and the feedback trajectory varies slightly with the increase of the stiffness coefficient k. This is because the magnitude of the tracking error is mainly determined by the estimation and compensation of disturbances in the inner loop. 3) A larger stiffness coefficient... k This results in a higher admittance value, allowing the exoskeleton to adapt more flexibly to human movement while requiring less active joint torque.

[0119] For the proposed ICC-EJTP control strategy, the control performance of all subjects under different stiffness coefficients was compared, and the results are as follows: Figure 17 As shown. It can be seen that: 1) Regarding the difference in the total trajectory between the hip and knee joints, when k When = 0, the average MAE-RA is 0.643° and 0.589°, significantly lower than the values ​​in active mode (when = 0). k When = 5, they are 1.691° and 1.018° respectively. k = 10, which are 3.534° and 2.098° respectively. This clearly shows that MAE-RA increases with increasing stiffness value. 2) Regarding the inner ring tracking accuracy, the MAE of the joint shows a slight upward trend with increasing stiffness, when k When = 0, the average MAE-DA values ​​are 0.630° and 0.582°. k = 5, which are 0.646° and 0.610°, when k =10, the values ​​are 0.791° and 0.731°. 3) Regarding AL, when k When = 0, the average values ​​for the hip and knee joints reach their lowest values ​​of 0.133° / N·m and 0.114° / N·m, respectively; when kWhen the value increased to 5, AL rose to 0.172 ° / N·m and 0.166 ° / N·m, respectively, while for k When = 5, the AL value reaches its peak, at 0.434 ° / N·m and 0.284 ° / N·m, respectively. In summary, the experimental results verify the potential of applying the method provided in this invention to the collaborative control of exoskeleton robots.

[0120] In summary, this invention effectively avoids the complex manual feature extraction process of traditional solutions, achieving synchronous prediction of torques in the hip, knee, and ankle joints of the lower limbs with lower computational resource consumption. Addressing the issue of inability to directly measure joint torques, this invention utilizes a neuromuscular model to accurately calculate the autonomous joint torques. To achieve full integration of surface electromyography signals and kinematic data, this solution constructs and trains an end-to-end deep learning model, and selects the optimal model through multiple rounds of performance evaluation. This technical solution adopts a "dual-layer control architecture": the outer layer adjusts the motion trajectory increment by predicting joint torques, thereby deeply integrating the human body's active movement intentions into the control loop and ensuring the compliance of human-computer interaction; the inner layer employs model-free adaptive terminal sliding mode control and time delay estimation technology, which can effectively compensate for system dynamic uncertainties and achieve precise, compliant, and coordinated control of the exoskeleton.

[0121] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for controlling a lower limb rehabilitation exoskeleton based on synchronous prediction of multi-joint torques, characterized in that, include: Model building phase: sEMG and IMU data were collected from subjects under different lower limb movement patterns; The joint angles are calculated from the IMU data. Muscle activation is calculated based on sEMG data, and then the joint torques of the hip, knee, and ankle joints are calculated by combining the Hill-type musculoskeletal model. The sEMG data and joint angle data were concatenated and divided into sample data according to time. The joint torques corresponding to the sample data at the corresponding time are used as the corresponding labels to construct a training set; The deep learning model is trained using a training set, and the trained deep learning model is the lower limb multi-joint torque prediction model. Model application phase: The system acquires sEMG and IMU data of the subject during exercise in real time, calculates the joint angles through the IMU data, and then inputs the sEMG data and joint angle data into the lower limb multi-joint torque prediction model to predict the joint torques of the hip, knee and ankle joints in real time. Rehabilitation exoskeleton control is performed based on real-time predicted joint torques of the hip, knee, and ankle joints.

2. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in claim 1, characterized in that, The sEMG data includes sEMG signals from the rectus femoris, vastus medialis, biceps femoris, and tibialis anterior muscles of the subject's leg.

3. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in claim 2, characterized in that, The joint moments of the hip, knee, and ankle joints were calculated using the Hill-type musculoskeletal model, including: Based on the Hill-type musculoskeletal model, the active and passive muscle forces of the target muscle are calculated according to the muscle activation of the target muscle. Then, by combining the active and passive muscle forces, the muscle-tendon unit output force of the target muscle is obtained. The target muscle is the rectus femoris, vastus medialis, biceps femoris, or tibialis anterior. The joint torque is calculated based on the output force of the muscle-tendon unit: in, Indicates the first j Joint torque of each joint j Indicates the hip, knee, or ankle joint; Indicates the first i The force output by the muscle-tendon unit of the target muscle; Indicates the first i Block target muscle relative to the first j The instantaneous lever arm of each joint; t Indicates time, To participate in the j The number of muscles calculated from the joint torque.

4. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in claim 3, characterized in that, The muscles involved in calculating the joint torque of the hip joint are the rectus femoris and biceps femoris; the muscles involved in calculating the joint torque of the knee joint are the rectus femoris, biceps femoris, and vastus medialis; and the muscle involved in calculating the joint torque of the ankle joint is the tibialis anterior.

5. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in claim 3, characterized in that, The output force of the muscle-tendon unit of the target muscle The formula for calculation is: , , in, For active muscle strength, For passive muscle force, φ The pennate angle of the muscle fiber; f a To normalize the active force-length relationship function, f v To normalize the active force-velocity relationship function, l m , v m These are normalized muscle fiber length and normalized muscle fiber velocity, respectively. a ( t () represents muscle activation level. F max This represents the maximum isometric contractile force of the muscle. f p This is the normalized passive elastic force-length relationship function.

6. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in claim 1, characterized in that, Muscle activation was calculated based on sEMG data, including: Neural activation was calculated based on sEMG data. u ( t ): in, express td sEMG data at any given time d For time delay; u ( t ), u ( t -1) u ( t -2) respectively represent t , t -1、 t Neural activation at time -2; α , β 1. β 2 is the proportionality coefficient, which satisfies intermediate parameters | 1|<1、| 2|<1; Furthermore, based on neural activation level u ( t Muscle activation level was obtained. a ( t ): in, A It is a non-linear shape factor. A The value ranges from 0 to 3.

7. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in claim 1, characterized in that, The deep learning model comprises a convolutional neural network module, a spatial and channel reconstruction convolutional module, a gated recurrent unit module, and a multi-head attention mechanism module connected in sequence. The convolutional neural network module extracts local features in the spatial dimension of the input data to obtain feature map 1. The spatial and channel reconstruction convolutional module reconstructs feature map 1 from both spatial and channel dimensions to obtain feature map 2. The gated recurrent unit module extracts the dynamic dependencies between consecutive time steps from feature map 2 to obtain feature map 3. Based on feature map 3, the multi-head attention module performs weighted fusion of temporal features from multiple feature subspaces to obtain fused features. Based on these fused features, a predicted joint torque is output.

8. The lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in any one of claims 1-7, characterized in that, When controlling the rehabilitation exoskeleton, a dual-loop control structure consisting of an outer loop and an inner loop is adopted. The outer loop converts the predicted joint torque into the desired motion trajectory increment and dynamically adjusts the target trajectory of the exoskeleton. The inner loop, based on the target trajectory, uses model-free adaptive terminal sliding mode control combined with time delay estimation to achieve rehabilitation exoskeleton control.

9. A lower limb rehabilitation exoskeleton control system based on multi-joint torque synchronous prediction, characterized in that, Includes a processor for executing the lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the lower limb rehabilitation exoskeleton control method based on multi-joint torque synchronous prediction as described in any one of claims 1-8.

Images