A lower limb exoskeleton control method based on multi-modal data and variable impedance control

Abstract

The application discloses a lower limb exoskeleton control method based on multi-modal data and variable impedance control, and relates to the technical field of rehabilitation medical devices. The method comprises the following steps: determining a rehabilitation stage of a patient and generating a dynamic condition constraint vector based on multi-modal data of the patient; obtaining an ideal gait trajectory of a diseased side and an abnormal probability score of human-machine interaction according to a multi-joint kinematic state sequence of a healthy side of the patient and the dynamic condition constraint vector; estimating a human-machine interaction torque in real time based on the multi-joint kinematic state sequence, and determining a virtual impedance based on the rehabilitation stage and the abnormal probability score; the virtual impedance is used to constrain the interaction behavior between the lower limb exoskeleton and the human; combining the ideal gait trajectory, the human-machine interaction torque estimation value and the virtual impedance, a motor control instruction is determined, and the motor control instruction is used to drive the exoskeleton robot to assist the patient in moving. The method improves the rehabilitation efficiency of the patient.

Description

Technical Field

[0001] This application relates to the field of rehabilitation medical equipment technology, and in particular to a lower limb exoskeleton control method based on multimodal data and variable impedance control. Background Technology

[0002] In the clinical pathway for lower limb motor function reconstruction in patients with central nervous system injuries such as stroke and spinal cord injury, lower limb rehabilitation exoskeleton robots have become the core intelligent equipment in the field of neurorehabilitation. Through joint movement guidance and human-machine interaction force assistance, they can provide standardized gait training for patients with nerve injuries. This not only makes up for the shortcomings of traditional manual rehabilitation treatment, such as the high physical load on therapists, limited training time, and difficulty in consistently ensuring movement accuracy, but also promotes the remodeling of neural pathways and motor function compensation in the patient's central nervous system through repeated correct gait stimulation, helping patients gradually rebuild their ability to walk independently. It plays an irreplaceable key supporting role in the rehabilitation process.

[0003] However, existing exoskeleton control systems typically employ pre-set constant control parameters, such as fixed joint motion trajectories, stiffness, and damping coefficients. In actual rehabilitation applications, because systems using fixed control parameters cannot dynamically adjust assistive strategies based on the patient's actual rehabilitation progress and needs, this rigid control method prevents the exoskeleton robot from fully realizing its intended rehabilitation effectiveness. Fixed control parameters cannot dynamically adjust assistive strategies based on the patient's real-time physiological state and motor function recovery, resulting in patients not receiving the most suitable assistive support during rehabilitation training, thus leading to lower rehabilitation efficiency. Summary of the Invention

[0004] Therefore, it is necessary to provide a lower limb exoskeleton control method based on multimodal data and variable impedance control to address the aforementioned technical problems. This method improves the rehabilitation efficiency of patients.

[0005] The following technical solution is adopted in this specification: This specification provides a lower limb exoskeleton control method based on multimodal data and variable impedance control, including: Based on the patient's multimodal data, the patient's rehabilitation stage is determined and a dynamic condition constraint vector is generated. The multimodal data includes dynamic gait feature sequences and static physiological features. The rehabilitation stage includes early passive mode, mid-term mirror mode, and late-term active mode. The dynamic condition constraint vector includes gait speed limit and physical rigidity constraint. The physical rigidity constraint is the rigidity of the exoskeleton robot when the patient interacts with the exoskeleton robot. Based on the multi-joint kinematic state sequence and dynamic condition constraint vector of the patient's healthy limb, the ideal gait trajectory and abnormal probability score of human-computer interaction on the affected side are obtained. Based on multi-joint kinematic state sequences, the human-computer interaction torque is estimated in real time, and virtual impedance is determined based on the rehabilitation stage and abnormal probability score; virtual impedance is used to constrain the interaction behavior between the lower limb exoskeleton and the human. By combining the ideal gait trajectory, the estimated torque value of human-computer interaction, and the virtual impedance, the motor control commands are determined, and the exoskeleton robot is driven to assist the patient's movement through the motor control commands.

[0006] Optionally, based on the patient's multimodal data, the patient's rehabilitation stage is determined and a dynamic condition constraint vector is generated, specifically including: Static physiological features are mapped to the semantic space through a multilayer perceptron and then used as query guiding conditions. Based on query guidance conditions, a cross-attention mechanism is used to adaptively filter key segments in dynamic gait feature sequences and identify key segments as cross-modal fusion features. Based on cross-modal fusion features, predict core clinical indicators; core clinical indicators include lower limb motor function FMA score, functional walking ability classification FAC score, preset minute walking distance, and preset distance walking speed; The patient's rehabilitation stage is determined based on the lower limb motor function FMA score and the functional walking ability classification FAC score. Based on static physiological characteristics, preset walking distance per minute, preset walking speed, and rehabilitation stage, a dynamic condition constraint vector is generated, which includes walking speed limits and physical rigid constraints.

[0007] Optionally, based on cross-modal fusion features, predict core clinical indicators, specifically including: Cross-modal fusion features are input into a pre-trained multi-task prediction network for multi-branch regression processing to obtain the patient's FMA score, FAC grade, preset minute walking distance, and preset distance walking speed.

[0008] Optionally, the patient's rehabilitation stage can be determined based on the lower limb motor function FMA score and the functional walking ability classification FAC score, specifically including: When the FMA score is less than the first preset FMA score or the FAC grade is less than or equal to the first preset FAC grade, the rehabilitation stage is determined to be the early passive mode. When the FMA score is greater than or equal to the first preset FMA score and less than the second preset FMA score, and the FAC grade is the second preset FAC grade, the rehabilitation stage is determined to be the intermediate mirror mode. When the FMA score is greater than or equal to the second preset FMA score and the FAC grade is greater than or equal to the third preset FAC grade, the rehabilitation stage is determined to be the end-stage active mode.

[0009] Optionally, based on static physiological characteristics, preset walking distance per minute, preset walking speed over a distance, and rehabilitation stage, a dynamic condition constraint vector is generated, including walking speed limits and physical rigid constraints, specifically including: Convert the preset walking distance per minute into speed; The maximum value of the preset walking speed and the converted speed is determined as the patient's expected lower limit speed; Extract patient identification information and affected side indication markers from static physiological characteristics; The patient's identity information, affected side indicator, rehabilitation stage, expected lower limit speed, preset walking distance per minute, preset distance walking speed, and skeletal muscle strength level mapping value are concatenated to obtain a multidimensional dynamic condition constraint vector; the skeletal muscle strength level mapping value is a dynamic feature obtained based on the deconstruction of clinical core indicators, used to reflect the patient's current muscle control function.

[0010] Optionally, based on the multi-joint kinematic state sequence and dynamic condition constraint vector of the patient's healthy limb, the ideal gait trajectory and abnormal probability score of the human-computer interaction on the affected side are obtained, specifically including: Obtain a cross-limb continuous gait generator; the cross-limb continuous gait generator includes a sequence embedding layer, a Transformer stack block modulated by adaptive layer normalization, a decoding generation branch and an anomaly detection branch; the Transformer stack block includes multiple Transformer layers connected in sequence; the anomaly detection branch includes a global average pooling layer and a classifier; The multi-joint kinematic state sequence of the patient's unaffected limb is input into the cross-limb continuous gait generator; In the sequence embedding layer, the multi-joint kinematic state sequence is patched to obtain non-overlapping time slice tokens, and learnable position codes are injected into the tokens to obtain the initial input tensor of the Transformer. The initial input tensor of the Transformer is input into the Transformer stack block, and the output of the last Transformer layer is determined as the deep hidden state; In the decoding generation branch, the deep hidden state is subjected to inverse patching and linear decoding to reconstruct the ideal gait trajectory of the affected side that is time-continuous. In the anomaly detection branch, global average pooling is performed on the deep hidden states to extract global features, and based on the global features, an anomaly probability score of human-computer interaction is calculated and output through a classifier.

[0011] Optionally, the method further includes: When the patient is in the end-stage active mode, a high-dimensional pressure manifold feature space is constructed. The plantar pressure manifold features acquired by the sensor are used as augmented observation vectors, and the augmented observation vectors are mapped to the high-dimensional pressure manifold feature space. The state variables of the exoskeleton system are defined as potential state vectors including continuous gait phase, angular frequency, and cross-impedance characteristics; continuous gait phase is used to characterize gait progress; angular frequency is used to characterize the dynamic rate of change of gait cycle; cross-impedance characteristics are used to reflect the mechanical characteristics of human-computer interaction. The adaptive filtering algorithm based on variational Bayesian inference estimates the observation noise covariance online during the joint estimation of gait phase and outputs continuous gait phase. The observation noise covariance is the noise generated by the physical environment interference during the actual measurement of the augmented observation vector. The continuous gait phase is used to synchronize the ideal gait trajectory so that the lower limb exoskeleton movement is synchronized with the patient's intention.

[0012] Optionally, the adaptive filtering algorithm based on variational Bayesian inference estimates the observation noise covariance online during the joint estimation of gait phase and outputs continuous gait phase, specifically including: In the variational prediction stage of the adaptive filtering algorithm, based on the posterior estimate of the previous time step, the state prior prediction is performed by linearizing the Jacobian matrix to obtain the mean of the state prior prediction and the prior prediction covariance at the current time step. In the variational update stage of the adaptive filtering algorithm, the expected value of the observation noise accuracy is estimated by Bayesian inference using the variational parameters at the current time, and the observation noise covariance is dynamically updated based on the expected value of the observation noise accuracy. The corrected Kalman gain is calculated based on the updated observation noise covariance. The posterior mean and covariance of the latent state vector are updated by combining the modified Kalman gain with the augmented observation vector at the current time step. The noise parameter estimation, Kalman gain correction, and state posterior update are performed iteratively until the iterative convergence condition is met. The iterative convergence condition includes variational free energy convergence or reaching a preset number of iterations. Extract the phase components from the latent state vector when the iteration converges, and determine the extracted phase components as continuous gait phases.

[0013] Optionally, the virtual impedance includes virtual stiffness and damping matrix; the virtual impedance is determined based on the recovery stage and anomaly probability score, specifically including: Convert the rehabilitation stage into the corresponding rehabilitation value; Based on the stage adaptive positive attenuation coefficient of stiffness and the safety adaptive positive sensitivity coefficient of stiffness, the rehabilitation value and the abnormal probability score are weighted and summed to obtain the stiffness-driven adaptive rehabilitation fusion score. The stiffness-driven adaptive rehabilitation fusion score is used as the input independent variable of the exponential function to obtain the dynamic weight scalar of virtual stiffness. Based on the damping stage adaptive positive attenuation coefficient and the damping safety adaptive positive sensitivity coefficient, the rehabilitation value and the abnormal probability score are weighted and summed to obtain the damping adaptive gait rehabilitation comprehensive index. The damping adaptive gait rehabilitation comprehensive index is used as the input independent variable of the exponential function to solve for the dynamic weight scalar of the damping matrix. Multiply the dynamic weight scalar of virtual stiffness by the virtual stiffness at the previous moment to obtain the virtual stiffness at the current moment. Multiply the dynamic weight scalar of the damping matrix by the damping matrix at the previous moment to obtain the damping matrix at the current moment. The virtual stiffness and damping matrix at the current moment are used to determine the virtual impedance at the current moment.

[0014] Optionally, the motor control commands are determined by combining the ideal gait trajectory, the estimated human-machine interaction torque, and the virtual impedance, specifically including: The error between the ideal gait trajectory and the actual joint position of the exoskeleton is defined as the trajectory error; Using the backstepping method, combined with the human-machine interaction torque estimation, trajectory error, and virtual impedance, the first-order reference velocity command is solved; The first-order reference speed command is subjected to high-frequency filtering, and based on the first-order reference speed after high-frequency filtering, a coupling compensation term is determined to counteract the interaction force between the exoskeleton link side and the motor side. The feedforward compensation term is obtained by performing inverse operations based on the rigid body dynamics model of the lower limb exoskeleton system. Based on Lyapunov stability theory, the feedback stabilization term is determined by utilizing the tracking error between the actual output motion parameters of the motor and the planned expected motion parameters. The motor control command is obtained by algebraically superimposing the feedforward compensation term, the feedback stabilization term, and the coupling compensation term.

[0015] This specification provides a lower limb exoskeleton control device based on multimodal data and variable impedance control, comprising: The determination module is used to determine the patient's rehabilitation stage and generate dynamic condition constraint vectors based on the patient's multimodal data. The multimodal data includes dynamic gait feature sequences and static physiological features. The rehabilitation stage includes early passive mode, intermediate mirror mode, and late active mode. The dynamic condition constraint vector includes gait speed limits and physical rigidity constraints. The physical rigidity constraints are the degree of rigidity of the exoskeleton robot when the patient interacts with the exoskeleton robot. The generation module is used to obtain the ideal gait trajectory and abnormal probability score of human-computer interaction on the affected side based on the multi-joint kinematic state sequence and dynamic condition constraint vector of the patient's healthy limb. The estimation module is used to estimate the human-computer interaction torque in real time based on the multi-joint kinematic state sequence, and to determine the virtual impedance based on the rehabilitation stage and abnormal probability score; the virtual impedance is used to constrain the interaction behavior between the lower limb exoskeleton and the human. The control module is used to combine the ideal gait trajectory, the estimated torque value of human-machine interaction, and the virtual impedance to determine the motor control commands, and drive the exoskeleton robot to assist the patient's movement through the motor control commands.

[0016] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned lower limb exoskeleton control method based on multimodal data and variable impedance control.

[0017] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned lower limb exoskeleton control method based on multimodal data and variable impedance control.

[0018] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: In the lower limb exoskeleton control method based on multimodal data and variable impedance control provided in this specification, the method automatically determines the rehabilitation stage through multimodal data of dynamic gait feature sequences and static physiological features. Based on the rehabilitation stage and the probability of interaction abnormality, the virtual impedance is dynamically adjusted, which not only ensures the trajectory following accuracy in the early passive mode, but also releases the patient's autonomous movement space in the late active mode. The human-machine interaction torque is estimated in real time through multi-joint state sequences, which solves the pain point that traditional exoskeleton fixation control strategies cannot adapt to different rehabilitation processes. Based on the multi-joint kinematic sequence of the healthy limb, the ideal gait trajectory of the affected side is generated, which is completely in line with the patient's own physiological movement habits, thereby improving the patient's rehabilitation efficiency. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This specification provides a flowchart illustrating a lower limb exoskeleton control method based on multimodal data and variable impedance control. Figure 2 The overall flowchart provided for this invention; Figure 3A schematic diagram illustrating the rehabilitation assessment and strategy distribution method based on multimodal cross-attention provided by this invention; Figure 4 A schematic diagram of the cross-limb personalized gait generation and anomaly detection method based on JGT provided by the present invention; Figure 5 A schematic diagram of the adaptive variable impedance control method based on an extended state observer provided by the present invention; Figure 6 This specification provides a schematic diagram of a lower limb exoskeleton control device based on multimodal data and variable impedance control. Figure 7 This specification provides a schematic diagram of a computer device for implementing a lower limb exoskeleton control method based on multimodal data and variable impedance control. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0021] Devices such as desktop computers, servers, and laptops are capable of executing the solutions described in this manual. For ease of explanation, the following description will focus on servers as the primary execution method.

[0022] Deficiencies of existing technology: 1. Disconnect between macroscopic assessment and underlying control: Existing rehabilitation assessments (such as FMA and FAC scores) are mostly completed offline by doctors, and exoskeleton control parameters (such as stiffness, damping, and gait speed) rely on manual adjustment based on human experience. Current technology lacks a mechanism that can directly and autonomously map the patient's dynamic clinical indicators to the robot's underlying continuous control manifold.

[0023] 2. Gait trajectory generation lacks personalization and is difficult to ensure safety: Existing exoskeletons mostly use pre-programmed standard gait or fixed central pattern generators (CPGs). These methods cannot adapt to the patient's static biometrics such as height and weight, and when faced with non-stationary biometric constraints, it is difficult to reconstruct the trajectory of the affected side in real time based on the healthy limb.

[0024] 3. Gait phase detection suffers from discrete jumps and lags: Traditional finite state machines (FSMs) rely on discrete plantar pressure thresholds for gait switching. On noisy surfaces (such as carpets), misjudgments and state jumps are prone to occur, resulting in discontinuous control commands.

[0025] Human-computer interaction force detection is costly and has rigid impedance parameters: Traditional exoskeletons often use expensive and bulky multi-dimensional force sensors connected in series in various joints to achieve compliant control. At the same time, existing impedance control is mostly a constant parameter, which cannot adaptively adjust the impedance according to the patient's rehabilitation stage (passive, assisted, resistive) and transient abnormalities.

[0026] To address the shortcomings of existing technologies, this invention aims to provide a lower limb exoskeleton control method based on multimodal data and variable impedance control. Specific objectives include: 1. Construct a top-level perception brain (multimodal cross-attention gait assessment model) to achieve automated and rigorous mathematical mapping from multimodal clinical indicators to the robot's underlying control parameters, thus solving the problem of disconnect between assessment and control.

[0027] 2. A Joint-Graph Transformer (JGT) architecture is proposed, which integrates static biometrics and kinematic information from the healthy side to achieve real-time reconstruction of personalized gait on the affected side and high-frequency anomaly detection.

[0028] 3. Introduce a continuous phase estimator based on Variational Bayesian Inverse Model for Phase Estimation (VB-IMPE), which integrates pressure manifold and human-machine interaction torque to eliminate jumps caused by discrete state switching.

[0029] 4. Design a variable impedance control strategy based on Extended State Observer (ESO) to eliminate the hardware dependence on physical force sensors and dynamically adjust the system impedance in real time according to the recovery stage and abnormal score to ensure the safety of physical interaction throughout the entire cycle.

[0030] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0031] Figure 1 This is a schematic diagram of a lower limb exoskeleton control method based on multimodal data and variable impedance control, as described in this specification. The method specifically includes the following steps: S101: Based on the patient's multimodal data, determine the patient's rehabilitation stage and generate a dynamic condition constraint vector; the multimodal data includes dynamic gait feature sequences and static physiological features; the rehabilitation stage includes early passive mode, mid-term mirror mode and late-term active mode; the dynamic condition constraint vector includes gait speed limit and physical rigidity constraint; the physical rigidity constraint is the rigidity of the exoskeleton robot when the patient interacts with the exoskeleton robot.

[0032] In an exemplary embodiment, based on the patient's multimodal data, the rehabilitation stage of the patient is determined and a dynamic condition constraint vector is generated. Specifically, this includes: mapping static physiological features to a semantic space using a multilayer perceptron and using this as query guidance conditions; adaptively filtering key segments in the dynamic gait feature sequence using a cross-attention mechanism based on the query guidance conditions, and identifying these key segments as cross-modal fusion features; predicting core clinical indicators based on the cross-modal fusion features; the core clinical indicators include the lower limb motor function FMA score, the functional walking ability classification FAC score, a preset walking distance per minute, and a preset walking speed per distance; determining the patient's rehabilitation stage based on the lower limb motor function FMA score and the functional walking ability classification FAC score; and generating a dynamic condition constraint vector including gait speed limits and physical rigidity constraints based on static physiological features, the preset walking distance per minute, the preset walking speed per distance, and the rehabilitation stage.

[0033] In an exemplary embodiment, predicting core clinical indicators based on cross-modal fusion features specifically includes: inputting the cross-modal fusion features into a pre-trained multi-task prediction network for multi-branch regression processing to obtain the patient's FMA score, FAC grade, preset minute walking distance, and preset distance walking speed.

[0034] In an exemplary embodiment, the rehabilitation stage of a patient is determined based on the lower limb motor function FMA score and the functional walking ability classification FAC score. Specifically, this includes: when the FMA score is less than a first preset FMA score or the FAC classification is less than or equal to a first preset FAC level, the rehabilitation stage is determined to be an early passive mode; when the FMA score is greater than or equal to the first preset FMA score and less than a second preset FMA score, and the FAC classification is a second preset FAC level, the rehabilitation stage is determined to be an intermediate mirror mode; when the FMA score is greater than or equal to the second preset FMA score and the FAC classification is greater than or equal to a third preset FAC level, the rehabilitation stage is determined to be a late active mode.

[0035] Specifically, the first preset FMA score, the second preset FMA score, the first preset FAC level, the second preset FAC level, and the third preset FAC level are set according to specific engineering practices.

[0036] In an exemplary embodiment, based on static physiological characteristics, a preset walking distance per minute, a preset walking speed over a distance, and a rehabilitation stage, a dynamic condition constraint vector including speed limits and physical rigid constraints is generated. Specifically, this includes: converting the preset walking distance per minute into speed; determining the maximum value of the preset walking speed over a distance and the converted speed as the patient's expected lower limit speed; extracting the patient's identity information and affected side indicator from the static physiological characteristics; and concatenating the patient's identity information, affected side indicator, rehabilitation stage, expected lower limit speed, preset walking distance per minute, preset walking speed over a distance, and skeletal muscle strength level mapping value to obtain a multidimensional dynamic condition constraint vector. The skeletal muscle strength level mapping value is a dynamic feature obtained based on the deconstruction of clinical core indicators and is used to reflect the patient's current muscle control function.

[0037] Specifically, the preset walking distance in minutes and the preset walking speed are set according to specific engineering practices. For example, the preset walking distance in minutes is 6 minutes of walking distance, and the preset walking speed is 10 meters of walking speed.

[0038] Based on static physiological characteristics, 6-minute walking distance, 10-meter walking speed, and rehabilitation stage level, a dynamic condition constraint vector is generated, including speed limits and physical rigid constraints. Specifically, this includes: converting the predicted 6-minute walking distance into speed; determining the maximum value of the predicted 10-meter walking speed and the converted speed as the patient's expected lower limit speed; extracting the patient's height, weight, age, and affected side indicator from the static physiological characteristics; and concatenating the height, weight, age, affected side indicator, rehabilitation stage level, expected lower limit speed, predicted 6-minute walking distance, and skeletal muscle strength level mapping values ​​to obtain a multi-dimensional dynamic condition constraint vector.

[0039] Specifically, the process involves acquiring a continuously sampled dynamic gait feature sequence matrix and static physiological features (such as height, weight, and age); mapping the static physiological features to a semantic space using a multilayer perceptron as a query guide; adaptively selecting key segments from the dynamic gait sequence using a cross-attention mechanism guided by the static physiological features to generate cross-modal fusion features; and predicting core clinical indicators (FMA score, FAC grade, 6-minute walking distance, etc.) based on these fusion features, and then applying them through a step-wise decision function. The system autonomously determines the patient's rehabilitation stage (e.g., 0-early passive mode, 1-mid-stage mirror mode, 2-late-stage active mode); based on the extracted features, it generates a dynamic condition constraint vector that includes gait speed limits and physical rigid constraints. , and then distribute it to the lower levels.

[0040] Figure 2 The overall flowchart provided for this invention is as follows: Figure 2As shown, the process begins: Acquire the patient's multimodal data (including dynamic gait feature sequences and static physiological features); input the multimodal data into the Multi-modal Cross-Attention Gait Evaluation Model (MC-GT) to predict core clinical indicators and output the patient's current rehabilitation stage and dynamic conditional constraint vector; Mid-level trajectory and phase planning: Branch 1: Input the healthy limb movement state and dynamic conditional constraint vector into the Cross-limb Continuous Gait Generator (JGT) to generate the ideal gait trajectory for the affected side and simultaneously output the abnormality probability score; Branch 2: Input the plantar pressure manifold features into the Velocity-Based Instantaneous Monotonic Gait Estimator (Velocity-Based Instantaneous Monotonic Gait Generator). PhaseEstimator (VB-IMPE) outputs continuous gait phase (primarily used during active rehabilitation); based on the extended state observer, it estimates the human-machine interaction torque in real time and adjusts the virtual impedance (stiffness and damping) in real time according to the rehabilitation stage and abnormal probability score; combining the ideal gait trajectory, continuous gait phase, and adjusted virtual impedance, it calculates the final motor control command to drive the exoskeleton robot to assist the patient's movement; end.

[0041] Figure 3 This is a schematic diagram of the rehabilitation assessment and strategy distribution method based on multimodal cross-attention provided by the present invention, as shown in the figure. Figure 3 As shown, we begin; two types of data are acquired in parallel: dynamic data: acquiring dynamic 28-dimensional temporal gait feature sequences, and static data: acquiring low-dimensional static patient physiological characteristics (such as height and weight). Feature Extraction: For dynamic data, high-dimensional dynamic gait representations are extracted through linear projection and absolute position encoding; for static data, high-dimensional static physiological representations are extracted through multilayer perceptron (MLP) mapping. Feature Fusion: The two types of high-dimensional representations are fused through a cross-attention mechanism (static physiological representation as Query, dynamic gait representation as Key and Value). Indicator Prediction: Four clinical indicators (FMA score, FAC grade, 6-minute walking distance, 10-meter walking speed) are output through a multi-task prediction network. Two types of results are output in parallel: Rehabilitation Stage Determination: Based on a step-wise determination function, the rehabilitation stage (passive / mirror / active) is output and passed to the underlying impedance controller. Dynamic Constraint Generation: An 8-dimensional dynamic condition constraint vector is generated by combining the predicted indicators and physiological features and passed to the gait generator. End.

[0042] S102: Based on the multi-joint kinematic state sequence and dynamic condition constraint vector of the patient's healthy limb, obtain the ideal gait trajectory and abnormal probability score of human-computer interaction on the affected side.

[0043] In an exemplary embodiment, based on the multi-joint kinematic state sequence of the patient's healthy limb and the dynamic condition constraint vector, the ideal gait trajectory of the affected side and the abnormal probability score of human-computer interaction are obtained. Specifically, this includes: obtaining a cross-limb continuous gait generator; the cross-limb continuous gait generator includes a sequence embedding layer, a Transformer stack block modulated by adaptive layer normalization, a decoding generation branch, and an anomaly detection branch; the Transformer stack block includes multiple Transformer layers connected in sequence; the anomaly detection branch includes a global average pooling layer and a classifier; the multi-joint kinematic state sequence of the patient's healthy limb is input into the cross-limb continuous gait generator; in the sequence embedding layer, the multi-joint kinematic state sequence is patched to obtain non-overlapping time slice tokens, and learnable position codes are injected into the tokens to obtain the initial input tensor of the Transformer; the initial input tensor of the Transformer is input into the Transformer stack block, and the output of the last Transformer layer is determined as the deep hidden state; In the decoding and generation branch, the deep hidden states are subjected to inverse patching and linear decoding to reconstruct the ideal gait trajectory of the affected side with continuous time. In the anomaly detection branch, the deep hidden states are subjected to global average pooling to extract global features, and based on the global features, the anomaly probability score of human-computer interaction is calculated and output through a classifier.

[0044] Specifically, based on JGT, cross-limb personalized gait generation and anomaly detection are performed to obtain the multi-joint kinematic state sequence of the healthy limb; a patch-based token embedding strategy is used to convert the healthy limb state sequence into high-dimensional features and inject position encoding; an adaptive layer normalization (AdaLN) mechanism is introduced to transform the dynamic condition constraint vector output from step one into a higher-dimensional feature. As system adjustment parameters, the layer normalization scaling and translation factors inside the Transformer are modulated; the ideal gait trajectory of the affected limb is decoded and reconstructed. Simultaneously, global hidden layer features are extracted and fed into the anomaly detection head, outputting the anomaly probability score of human-computer interaction in real time. .

[0045] Figure 4 This is a schematic diagram of the cross-limb personalized gait generation and anomaly detection method based on JGT provided by the present invention, as shown below. Figure 4As shown, the process begins: 1. Obtain the multi-joint kinematic state sequence of the healthy limb. 2. Perform patching on the state sequence to obtain non-overlapping time slice tokens and inject learnable positional codes to generate the initial input tensor for the Transformer. 3. Conditional modulation and feature extraction (core recurrent layer) obtains the dynamic conditional constraint vectors from the upper layer. Through Adaptive Layer Normalization (AdaLN) mechanism, the conditional constraint vectors are transformed into scaling and translation factors to modulate the distribution of input features. 4. After processing by Multi-head Self-Attention (MSA) and Feed Forward Network (FFN), a deep hidden state incorporating patient features is obtained. 5. Dual-branch output: Anomaly detection branch: Global Average Pooling (GAP) is performed on the deep hidden state to extract global features. The classifier calculates and outputs the probability score of human-computer interaction anomalies. 6. Decoding and generation branch: Inverse patching and linear decoding are performed on the deep hidden state to retrieve the temporally continuous physiotherapy gait trajectory of the affected side. 7. End.

[0046] S103: Based on the multi-joint kinematic state sequence, the human-computer interaction torque is estimated in real time, and the virtual impedance is determined based on the rehabilitation stage and abnormal probability score; the virtual impedance is used to constrain the interaction behavior between the lower limb exoskeleton and the human.

[0047] In an exemplary embodiment, the virtual impedance includes virtual stiffness and a damping matrix. Determining the virtual impedance based on the rehabilitation stage and anomaly probability score specifically includes: converting the rehabilitation stage into a corresponding rehabilitation value; weighting and summing the rehabilitation value and anomaly probability score based on the stage adaptive positive attenuation coefficient of stiffness and the safety adaptive positive sensitivity coefficient of stiffness to obtain a stiffness-driven adaptive rehabilitation fusion score; solving for the stiffness-driven adaptive rehabilitation fusion score as the input independent variable of an exponential function to obtain a dynamic weight scalar of virtual stiffness; and determining the virtual impedance based on the stage adaptive positive attenuation coefficient of damping. The adaptive positive sensitivity coefficient of the damping and the rehabilitation value and the abnormal probability score are weighted and summed to obtain the damping adaptive gait rehabilitation comprehensive index. The damping adaptive gait rehabilitation comprehensive index is used as the input independent variable of the exponential function to obtain the dynamic weight scalar of the damping matrix. The dynamic weight scalar of the virtual stiffness is multiplied by the virtual stiffness at the previous moment to obtain the virtual stiffness at the current moment. The dynamic weight scalar of the damping matrix is ​​multiplied by the damping matrix at the previous moment to obtain the damping matrix at the current moment. The virtual stiffness and damping matrix at the current moment are used to determine the virtual impedance at the current moment.

[0048] The formula for calculating the dynamic weight scalar of virtual stiffness is formula (1): (1); in, For the dynamic weight scalar of virtual stiffness, For the recovery phase (e.g., early passive mode value is 0, mid-stage mirror mode value is 1, and late-stage active mode value is 2), The output of the cross-limb continuous gait generator is the probability score of human-computer interaction anomalies, with a value range of [value range missing]. , For the stage-adaptive positive attenuation coefficient of stiffness, The safety adaptive positive sensitivity coefficient is the stiffness coefficient.

[0049] The formula for calculating the dynamic weight scalar of the damping matrix is ​​formula (2): (2); in, The dynamic weight scalar of the damping matrix, For the damping stage, the adaptive positive attenuation coefficient is... The damped safety adaptive positive sensitivity coefficient.

[0050] The nonlinear design of formulas (1) and (2) makes the abnormal probability score When the exponential term suddenly increases, it can drive the weight scalar to drop sharply, instantly reducing the system stiffness of the exoskeleton and achieving a highly safe and compliant control response.

[0051] Specifically, based on sensorless variable impedance and underlying compliant servo control using ESO, 1. a second-order nonlinear system dynamic model of the exoskeleton and SEA actuator is constructed; 2. a linear high-gain extended state observer (ESO) is constructed to measure the human-machine interaction torque. The system's extended state is estimated in real time, eliminating the need for physical force sensors; 3. Construct the desired second-order virtual impedance model. The virtual stiffness matrix of this model... and virtual damping matrix By a dynamic weight function Modulation, which is further divided into: Stage Adaptive: based on the recovery stage output in Step 1. The higher the stage (the better the patient's function), the lower the impedance weight and the greater the compliance, thus stimulating the patient's active participation; safety adaptation: when the abnormal score output in step three... When the weight is increased, the weight function decreases rapidly, instantly reducing the system stiffness and avoiding secondary damage.

[0052] Figure 5 A schematic diagram of the adaptive variable impedance control method based on an extended state observer provided by the present invention is shown below. Figure 5As shown, the process begins; Signal reception: Receives the ideal rehabilitation trajectory, actual joint kinematic state, rehabilitation stage indicators, and abnormal probability scores; Interactive torque observation: Inputs the actual joint kinematic state into the ESO to calculate the estimated human-machine interactive torque; Reference velocity solution: Using the backstepping method, combined with the estimated human-machine interactive torque, trajectory error, and the updated matrix, solves for the reference velocity command; Tracking differentiation and decoupling: Inputs the reference velocity command into the TD for filtering and calculates the coupling compensation term; Control law output: Combines feedforward, feedback, and coupling elimination terms to generate the motor control torque; Impedance weight modulation: Inputs the rehabilitation stage indicators and abnormal probability scores into the smoothing weight function to update the virtual stiffness / damping matrix; End.

[0053] S104: Combining the ideal gait trajectory, the estimated torque value of human-computer interaction, and the virtual impedance, determine the motor control command, and drive the exoskeleton robot to assist the patient's movement through the motor control command.

[0054] In an exemplary embodiment, the motor control command is determined by combining the ideal gait trajectory, the estimated human-machine interaction torque, and the virtual impedance. Specifically, this includes: determining the error between the ideal gait trajectory and the actual joint position of the exoskeleton as the trajectory error; using the backstepping method, combining the estimated human-machine interaction torque, the trajectory error, and the virtual impedance, to solve for the first-order reference velocity command; performing high-frequency filtering on the first-order reference velocity command, and determining the coupling compensation term to offset the interaction force between the exoskeleton link side and the motor side based on the high-frequency filtered first-order reference velocity; performing inverse operation based on the rigid body dynamics model of the lower limb exoskeleton system to obtain the feedforward compensation term; designing based on Lyapunov stability theory, using the tracking error between the actual output motion parameters of the motor and the planned desired motion parameters to determine the feedback stabilization term; and algebraically superimposing the feedforward compensation term, the feedback stabilization term, and the coupling compensation term to obtain the motor control command.

[0055] Specifically, by using backstepping combined with a nonlinear tracking differentiator (TD), the second-order impedance relationship is transformed into a first-order reference speed command, and the decoupled motor final control law is calculated, thereby achieving accurate and smooth tracking of the desired variable impedance trajectory.

[0056] Motor control commands (i.e., motor input torque) ) by feedforward compensation term Feedback and calming items Coupling compensation term Algebraic superposition constitutes, that is The specific derivation process is as follows: (1) Feedforward compensation term The formula is obtained by inverse operation based on the rigid body dynamics model of the lower limb exoskeleton system, and is used to counteract the system's own inertia, Coriolis force and gravitational torque, thereby improving the feedforward response speed. Its expression is formula (3): (3); in, This is the inertia matrix estimated in the exoskeleton dynamics model. The Coriolis force matrix estimated in the exoskeleton dynamics model. This is the gravity matrix estimated in the exoskeleton dynamics model. This represents the actual speed of the joint. The expected acceleration for the planned joints.

[0057] (2) Feedback calming items Based on Lyapunov stability theory, this design utilizes the tracking error (defined as position error) between the actual output motion parameters of the motor and the planned desired motion parameters. and speed error This causes the closed-loop system error to converge asymptotically. Its expression is given by formula (4): (4); in, The proportional feedback gain matrix is ​​positive definite. The differential feedback gain matrix is ​​positive definite. This is a robust compensation term used to suppress unmodeled dynamics and external bounded disturbances. It is a symbolic function.

[0058] In an exemplary embodiment, the method further includes: when the patient is in the terminal active mode, constructing a high-dimensional pressure manifold feature space, using the plantar pressure manifold features acquired by the sensor as an augmented observation vector, and mapping the augmented observation vector to the high-dimensional pressure manifold feature space; defining the state variables of the exoskeleton system as a latent state vector including continuous gait phase, angular frequency, and cross-impedance features; the continuous gait phase is used to characterize the gait progression; the angular frequency is used to characterize the dynamic rate of change of the gait cycle; the cross-impedance features are used to reflect the mechanical characteristics of human-computer interaction; based on an adaptive filtering algorithm of variational Bayesian inference, during the joint estimation of the gait phase, online estimation of the observation noise covariance is performed, and the continuous gait phase is output; the observation noise covariance is the noise generated by the physical environment interference during actual measurement of the augmented observation vector; the continuous gait phase is used to synchronize the time process of the ideal gait trajectory.

[0059] In an exemplary embodiment, the adaptive filtering algorithm based on variational Bayesian inference estimates the observation noise covariance online and outputs continuous gait phase during the joint estimation of gait phase. Specifically, this includes: in the variational prediction stage of the adaptive filtering algorithm, based on the posterior estimate of the previous time step, performing state prior prediction through Jacobian matrix linearization to obtain the mean and covariance of the state prior prediction at the current time step; in the variational update stage of the adaptive filtering algorithm, using the variational parameters at the current time step, estimating the expected value of the observation noise accuracy through Bayesian inference, and based on the observation... The expected value of the noise measurement accuracy is dynamically updated to update the observation noise covariance; based on the updated observation noise covariance, the corrected Kalman gain is calculated; using the corrected Kalman gain combined with the augmented observation vector at the current moment, the posterior mean and covariance of the latent state vector are updated; noise parameter estimation, Kalman gain correction, and state posterior update are iteratively performed until the iterative convergence condition is met; the iterative convergence condition includes variational free energy convergence or reaching a preset number of iterations; the phase component in the latent state vector when iterative convergence is met is extracted, and the extracted phase component is determined as the continuous gait phase.

[0060] Specifically, based on VB-IMPE-based continuous gait phase adaptive estimation (applied to active mode in the late stage of rehabilitation), a high-dimensional pressure manifold feature space is constructed, and the center of pressure (CoP) and its derivative are fused into an augmented observation vector. The exoskeleton state is defined as a latent state vector containing continuous gait phase, angular frequency, and cross-impedance features. An adaptive filtering algorithm based on variational Bayesian inference is introduced to estimate the observation noise covariance online during the joint estimation of gait phase. To adapt to environmental noise on different ground surfaces, it outputs smooth, continuous gait phase. Drive trajectory generation.

[0061] Beneficial effects: 1. Achieving a true closed loop of "perception-assessment-generation-control": MC-GT breaks through the isolated control and opens up the mathematical channel from clinical scales to the underlying motor torque of the robot, enabling the system to have brain-like intelligence to understand patients and autonomously switch rehabilitation strategies (passive / mirror / active).

[0062] 2. Extremely high individual adaptability and trajectory smoothness: JGT uses the AdaLN mechanism to deeply integrate the patient's static biometrics with dynamic intentions, solving the problem of rigid preset trajectories; VB-IMPE uses variational Bayesian evolution on Riemannian manifolds to completely eliminate the discrete jitter caused by traditional FSM.

[0063] 3. Low-cost, highly reliable, and inherently safe interaction: The ESO algorithm is used to reconstruct human-computer interaction force at high frequency, eliminating the dependence of each joint of the exoskeleton on expensive six-dimensional / single-axis force sensors, reducing hardware costs and failure rates; at the same time, the variable impedance controller, which is dually controlled by clinical stage and real-time abnormal score, ensures on-demand assistance and ultimate physical safety throughout the rehabilitation cycle.

[0064] When applying the lower limb exoskeleton control method based on multimodal data and variable impedance control provided in this manual, it is not necessary to... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0065] The above describes one or more embodiments of a lower limb exoskeleton control method based on multimodal data and variable impedance control provided in this specification. Based on the same concept, this specification also provides a corresponding lower limb exoskeleton control device based on multimodal data and variable impedance control, such as... Figure 6 As shown.

[0066] Figure 6 This specification provides a schematic diagram of a lower limb exoskeleton control device based on multimodal data and variable impedance control, including: The determination module 601 is used to determine the patient's rehabilitation stage and generate a dynamic condition constraint vector based on the patient's multimodal data; the multimodal data includes dynamic gait feature sequences and static physiological features; the rehabilitation stage includes early passive mode, mid-term mirror mode and late-term active mode; the dynamic condition constraint vector includes gait speed limit and physical rigidity constraint; the physical rigidity constraint is the rigidity of the exoskeleton robot when the patient interacts with the exoskeleton robot.

[0067] The generation module 602 is used to obtain the ideal gait trajectory and abnormal probability score of human-computer interaction of the affected side based on the multi-joint kinematic state sequence and dynamic condition constraint vector of the patient's healthy limb.

[0068] The estimation module 603 is used to estimate the human-computer interaction torque in real time based on the multi-joint kinematic state sequence, and to determine the virtual impedance based on the rehabilitation stage and the abnormal probability score; the virtual impedance is used to constrain the interaction behavior between the lower limb exoskeleton and the human.

[0069] The control module 604 is used to combine the ideal gait trajectory, the estimated human-machine interaction torque, and the virtual impedance to determine the motor control commands, and drive the exoskeleton robot to assist the patient's movement through the motor control commands.

[0070] Specific limitations regarding the lower limb exoskeleton control device based on multimodal data and variable impedance control can be found in the limitations of the lower limb exoskeleton control method based on multimodal data and variable impedance control described above, and will not be repeated here. Each module in the aforementioned lower limb exoskeleton control device based on multimodal data and variable impedance control can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0071] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A lower limb exoskeleton control method based on multimodal data and variable impedance control is provided.

[0072] This instruction manual also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 A lower limb exoskeleton control method based on multimodal data and variable impedance control is provided.

[0073] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A lower limb exoskeleton control method based on multimodal data and variable impedance control, characterized in that, include: Based on the patient's multimodal data, the patient's rehabilitation stage is determined and a dynamic condition constraint vector is generated; the multimodal data includes dynamic gait feature sequences and static physiological features; the rehabilitation stage includes early passive mode, mid-term mirror mode, and late-term active mode; the dynamic condition constraint vector includes gait speed limits and physical rigidity constraints; the physical rigidity constraints are the degree of rigidity of the exoskeleton robot when the patient interacts with the exoskeleton robot. Based on the multi-joint kinematic state sequence of the patient's healthy limb and the dynamic condition constraint vector, the ideal gait trajectory and abnormal probability score of human-computer interaction on the affected side are obtained. Based on the multi-joint kinematic state sequence, the human-computer interaction torque is estimated in real time, and the virtual impedance is determined based on the rehabilitation stage and abnormal probability score; the virtual impedance is used to constrain the interaction behavior between the lower limb exoskeleton and the human. By combining the ideal gait trajectory, the estimated human-computer interaction torque, and the virtual impedance, motor control commands are determined, and the exoskeleton robot is driven to assist the patient's movement through the motor control commands. The process of determining the patient's rehabilitation stage and generating a dynamic condition constraint vector based on the patient's multimodal data specifically includes: mapping the static physiological features to a semantic space using a multilayer perceptron and using this as a query guidance condition; adaptively filtering key segments in the dynamic gait feature sequence using a cross-attention mechanism based on the query guidance condition, and identifying these key segments as cross-modal fusion features; predicting core clinical indicators based on the cross-modal fusion features; the core clinical indicators include the lower limb motor function FMA score, the functional walking ability classification FAC score, a preset walking distance per minute, and a preset walking speed; determining the patient's rehabilitation stage based on the lower limb motor function FMA score and the functional walking ability classification FAC score; and generating a dynamic condition constraint vector including gait speed limits and physical rigid constraints based on the static physiological features, the preset walking distance per minute, the preset walking speed, and the rehabilitation stage.

2. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, The prediction of core clinical indicators based on the cross-modal fusion features specifically includes: The cross-modal fusion features are input into a pre-trained multi-task prediction network for multi-branch regression processing to obtain the patient's FMA score, FAC grade, preset minute walking distance, and preset distance walking speed.

3. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, The determination of the patient's rehabilitation stage based on the lower limb motor function FMA score and functional walking ability classification FAC score specifically includes: When the FMA score is less than the first preset FMA score or the FAC grade is less than or equal to the first preset FAC grade, the rehabilitation stage is determined to be the early passive mode. When the FMA score is greater than or equal to the first preset FMA score and less than the second preset FMA score, and the FAC grade is the second preset FAC grade, the rehabilitation stage is determined to be the intermediate mirror mode. When the FMA score is greater than or equal to the second preset FMA score and the FAC grade is greater than or equal to the third preset FAC grade, the rehabilitation stage is determined to be the end-stage active mode.

4. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, Based on the static physiological characteristics, the preset walking distance per minute, the preset walking speed, and the rehabilitation stage, a dynamic condition constraint vector is generated, including walking speed limits and physical rigid constraints, specifically including: Convert the preset walking distance per minute into speed; The maximum value of the preset walking speed and the converted speed is determined as the patient's expected lower limit speed; Extract the patient's identity information and affected side indicator from the static physiological characteristics; The patient's identity information, affected side indicator, rehabilitation stage, expected lower limit speed, preset minute walking distance, preset distance walking speed, and skeletal muscle strength level mapping value are concatenated to obtain a multidimensional dynamic condition constraint vector; the skeletal muscle strength level mapping value is a dynamic feature obtained based on the deconstruction of clinical core indicators, used to reflect the patient's current muscle control function.

5. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, The process of obtaining the ideal gait trajectory and abnormal probability score of human-computer interaction on the affected side based on the multi-joint kinematic state sequence of the patient's healthy limb and the dynamic condition constraint vector specifically includes: A cross-limb continuous gait generator is obtained; the cross-limb continuous gait generator includes a sequence embedding layer, a Transformer stack block modulated by adaptive layer normalization, a decoding generation branch and an anomaly detection branch; the Transformer stack block includes multiple Transformer layers connected in sequence; the anomaly detection branch includes a global average pooling layer and a classifier; The multi-joint kinematic state sequence of the patient's unaffected limb is input into the cross-limb continuous gait generator; In the sequence embedding layer, the multi-joint kinematic state sequence is patched to obtain non-overlapping time slice tokens, and learnable position codes are injected into the tokens to obtain the Transformer initial input tensor. The initial input tensor of the Transformer is input into the Transformer stack block, and the output of the last Transformer layer is determined as the deep hidden state; In the decoding generation branch, the deep hidden state is subjected to inverse patch operation and linear decoding to reconstruct the ideal gait trajectory of the affected side that is time-continuous. In the anomaly detection branch, global average pooling is performed on the deep hidden states to extract global features, and based on the global features, an anomaly probability score of human-computer interaction is calculated and output by a classifier.

6. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, The method further includes: When the patient is in the end-stage active mode, a high-dimensional pressure manifold feature space is constructed, the plantar pressure manifold features acquired by the sensor are used as augmented observation vectors, and the augmented observation vectors are mapped to the high-dimensional pressure manifold feature space. The state variables of the exoskeleton system are defined as potential state vectors including continuous gait phase, angular frequency, and interactive impedance characteristics; the continuous gait phase is used to characterize gait progress; the angular frequency is used to characterize the dynamic rate of change of the gait cycle; and the interactive impedance characteristics are used to reflect the mechanical characteristics of human-computer interaction. An adaptive filtering algorithm based on variational Bayesian inference estimates the observation noise covariance online during the joint estimation of gait phase and outputs a continuous gait phase. The observation noise covariance is the noise generated by physical environmental interference during the actual measurement of the augmented observation vector. The continuous gait phase is used to synchronize the ideal gait trajectory so that the lower limb exoskeleton movement is synchronized with the patient's intention.

7. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 6, characterized in that, The adaptive filtering algorithm based on variational Bayesian inference estimates the observation noise covariance online and outputs continuous gait phase during the joint estimation of gait phase, specifically including: In the variational prediction stage of the adaptive filtering algorithm, based on the posterior estimate of the previous time step, the state prior prediction is performed by linearizing the Jacobian matrix to obtain the mean of the state prior prediction and the prior prediction covariance at the current time step. In the variational update stage of the adaptive filtering algorithm, the expected value of the observation noise accuracy is estimated by Bayesian inference using the variational parameters at the current time, and the observation noise covariance is dynamically updated based on the expected value of the observation noise accuracy. The corrected Kalman gain is calculated based on the updated observation noise covariance. The posterior mean and covariance of the latent state vector are updated by combining the modified Kalman gain with the augmented observation vector at the current time step. The noise parameter estimation, Kalman gain correction, and state posterior update are performed iteratively until the iterative convergence condition is met; the iterative convergence condition includes variational free energy convergence or reaching a preset number of iterations; Extract the phase components from the latent state vector when the iteration converges, and determine the extracted phase components as continuous gait phases.

8. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, The virtual impedance includes a virtual stiffness and a damping matrix; the determination of the virtual impedance based on the rehabilitation stage and the abnormality probability score specifically includes: Convert the rehabilitation stage into the corresponding rehabilitation value; Based on the stage adaptive positive attenuation coefficient of stiffness and the safety adaptive positive sensitivity coefficient of stiffness, the rehabilitation value and the abnormal probability score are weighted and summed to obtain the stiffness-driven adaptive rehabilitation fusion score. The stiffness-driven adaptive rehabilitation fusion score is used as the input independent variable of the exponential function to obtain the dynamic weight scalar of virtual stiffness. Based on the damping stage adaptive positive attenuation coefficient and the damping safety adaptive positive sensitivity coefficient, the rehabilitation value and the abnormal probability score are weighted and summed to obtain the damping adaptive gait rehabilitation comprehensive index. The damping adaptive gait rehabilitation comprehensive index is used as the input independent variable of the exponential function to solve for the dynamic weight scalar of the damping matrix. Multiply the dynamic weight scalar of virtual stiffness by the virtual stiffness at the previous moment to obtain the virtual stiffness at the current moment. Multiply the dynamic weight scalar of the damping matrix by the damping matrix at the previous moment to obtain the damping matrix at the current moment. The virtual stiffness and damping matrix at the current moment are used to determine the virtual impedance at the current moment.

9. The lower limb exoskeleton control method based on multimodal data and variable impedance control as described in claim 1, characterized in that, The process of determining motor control commands by combining the ideal gait trajectory, the estimated human-machine interaction torque, and the virtual impedance specifically includes: The error between the ideal gait trajectory and the actual joint position of the exoskeleton is defined as the trajectory error; Using the backstepping method, combined with the estimated torque value of human-computer interaction, the trajectory error, and the virtual impedance, the first-order reference velocity command is solved; The first-order reference speed command is subjected to high-frequency filtering, and based on the first-order reference speed after high-frequency filtering, a coupling compensation term is determined to counteract the interaction force between the exoskeleton link side and the motor side. The feedforward compensation term is obtained by performing inverse operations based on the rigid body dynamics model of the lower limb exoskeleton system. Based on Lyapunov stability theory, the feedback stabilization term is determined by utilizing the tracking error between the actual output motion parameters of the motor and the planned expected motion parameters. The motor control command is obtained by algebraically superimposing the feedforward compensation term, the feedback stabilization term, and the coupling compensation term.

Images