Multi-modal feedback lower limb rehabilitation robot cooperative control system

Abstract

The invention discloses a multi-modal feedback lower limb rehabilitation robot cooperative control system, which relates to the technical field of rehabilitation robots, and comprises a multi-source heterogeneous sensing layer, an intention and environment fusion understanding layer, a bilateral differentiation cooperative planning layer, a distributed cooperative execution layer, a central safety coordinator and a data management and remote service layer, the multi-source heterogeneous sensing layer collects multi-dimensional signals, the intention and environment fusion understanding layer generates a motion intention vector and a terrain disturbance expectation vector, and the double-side differential collaborative planning layer dynamically synthesizes a personalized gait track and deconstructs the personalized gait track into a double-side differential control target set. The distributed cooperative execution layer drives an actuator to move through a differential control algorithm, the central safety coordinator guarantees training safety, and the data management and remote service layer realizes control strategy iterative optimization. According to the invention, active fusion of intention and environment and dynamic generation of a cooperative motion strategy are realized, and adaptability and naturalness of rehabilitation training are improved.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation robot technology, specifically to a multimodal feedback collaborative control system for a lower limb rehabilitation robot. Background Technology

[0002] With the accelerating aging of the population and the rising incidence of cardiovascular and cerebrovascular diseases, the number of patients with lower limb motor dysfunction caused by stroke, spinal cord injury, and other conditions is increasing year by year, creating a demand for efficient and intelligent rehabilitation training methods. Traditional manual rehabilitation therapy has limitations such as low efficiency, difficulty in standardizing intensity, and heavy workload for therapists. In recent years, rehabilitation robot technology has become an important auxiliary tool for motor function rehabilitation by providing high-intensity, repetitive, and quantifiable training, and related research and applications are becoming increasingly widespread. Lower limb rehabilitation robots aim to help patients rebuild or improve their walking ability through mechanical support, guidance, and assistance. To achieve safe, effective, and humane rehabilitation training, its control system needs to be able to accurately sense the patient's state, understand the patient's intentions, and achieve compliant and coordinated human-computer interaction.

[0003] Currently, technological advancements in this field have yielded some progress, with various control schemes emerging to enhance system adaptability and interactivity. For example, patent CN114366556B discloses a multi-modal training control system and method for lower limb rehabilitation. This scheme employs a hierarchical control architecture, integrating multiple sensors to acquire multi-dimensional data, analyzing and processing it to predict the patient and robot states, and then adjusting the control output of the drive actuators online. Its core objective is to ensure that the lower limbs can move according to a preset desired trajectory or desired interactive force. Another patent, CN108392795B, discloses a multi-modal control method for rehabilitation robots based on multi-information fusion. This method uses multi-modal data such as the patient's surface electromyography, electroencephalography, and electrocardiography, along with joint motion information, as a foundation. It utilizes information fusion algorithms to dynamically adjust training parameters and control system parameters, and achieves automatic switching of training modes, aiming to improve the flexibility and intelligence of control. In addition, existing technologies focus on adaptively adjusting the parameters of the admittance controller through optimization algorithms to improve the compliance of human-computer interaction; other research is dedicated to enhancing the mechanical compliance of the system through a rigid-flexible hybrid drive module.

[0004] However, despite significant advancements in multimodal information acquisition, online parameter adjustment, and optimization of specific control algorithms, existing control systems often suffer from a limitation: they largely adhere to a servo control paradigm. In this paradigm, the system typically pre-sets or generates a fixed desired motion trajectory offline, with the fundamental task of control being to drive the patient's limbs to track this predetermined goal. Even with the introduction of multi-sensor information fusion, its primary function is often limited to fine-tuning control parameters or switching between several pre-set, discrete training modes, without altering the core logic that the control system needs to track a predefined reference target. This tracking-based control strategy becomes passive and rigid when facing dynamically changing training environments, such as transitioning from flat ground to a slope or stairs, or when dealing with the patient's complex and fluctuating active movement intentions. The system cannot proactively and collaboratively generate a personalized movement strategy that ensures both safety and efficient rehabilitation based on environmental changes and the patient's real-time intentions before movement occurs, thus limiting the naturalness, adaptability, and ultimate effectiveness of rehabilitation training.

[0005] Therefore, a pressing technical problem in the existing technology is how to transform the control system of lower limb rehabilitation robots from a passive trajectory tracking servo paradigm to a collaborative decision-making control paradigm that actively integrates environmental perception and real-time patient intent understanding, and dynamically generates and executes collaborative motion strategies accordingly. This invention is proposed against this backdrop. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multimodal feedback collaborative control system for lower limb rehabilitation robots. By collecting multi-dimensional signals through multi-source heterogeneous sensing, and understanding and analyzing movement intentions and terrain disturbances through the fusion of intention and environment, combined with differentiated control of bilateral differentiated collaborative planning and distributed collaborative execution, it can actively adapt to the patient's real-time intentions and dynamic environment, break through the passive trajectory tracking paradigm, realize safe and collaborative rehabilitation training, and improve the adaptability and effectiveness of training.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a multimodal feedback collaborative control system for a lower limb rehabilitation robot, the system comprising:

[0008] A multi-source heterogeneous sensing layer is used to collect in real time the patient's internal lower limb motion state signals, human bioelectric signals, and terrain visual signals of the robot's external environment;

[0009] The intent and environment fusion understanding layer, connected to the multi-source heterogeneous perception layer, is used to receive and fuse lower limb internal motion state signals and human bioelectric signals to generate a comprehensive patient motion intent vector, while processing the terrain visual signals to generate a terrain disturbance expectation vector.

[0010] A bilateral differentiated collaborative planning layer, connected to the intent and environment fusion understanding layer, is used to receive the motion intent vector and the terrain disturbance expectation vector, and combine them with pre-stored patient bilateral limb muscle strength assessment data to dynamically synthesize adaptive gait trajectory targets. The gait trajectory targets are deconstructed into a first control target set for the healthy limb and a second control target set for the affected limb. The first control target set is mainly composed of expected cross-impedance parameters and expected end force field parameters, while the second control target set is mainly composed of expected joint trajectory, variable joint impedance parameters, and feedforward auxiliary torque curve.

[0011] A distributed collaborative execution layer connects the bilateral differentiated collaborative planning layer, and the distributed collaborative execution layer includes a healthy side execution control unit and a diseased side execution control unit;

[0012] The healthy side execution control unit is used to generate a first motor control command based on the first control target set and a force-position hybrid control algorithm based on the admittance model, so as to drive the robot's healthy side joint actuator to achieve compliant assisted motion.

[0013] The affected side execution control unit is used to generate a second motor control command based on the second control target set using a composite control algorithm that integrates feedforward torque compensation and feedback impedance control, so as to drive the robot's affected side joint actuator to achieve active guidance and assisted movement.

[0014] Furthermore, the multi-source heterogeneous sensing layer includes an internal state sensing module and an external environment sensing module:

[0015] The internal state perception module includes a rotary encoder and a joint torque sensor installed at the robot's hip and knee joints for measuring joint angles, joint angular velocities, and human-machine interaction torques. The internal state perception module also includes surface electromyography (EMG) sensors arranged on the surface of the rectus femoris and biceps femoris muscle bellies of the patient's thigh for collecting EMG signals. The internal state perception module also includes a multidimensional force sensor installed on the robot's lower leg link or foot structure for measuring three-dimensional ground reaction forces and interaction forces.

[0016] The external environment perception module includes a depth vision camera fixed to the front of the robot's torso. The optical axis of the depth vision camera is oriented diagonally downwards, and its field of view covers the area in front of the robot. The external environment perception module has a built-in terrain parsing algorithm for processing depth image data in real time, generating a terrain map containing height information and terrain semantic classification, and extracting key terrain feature parameters.

[0017] Furthermore, the intent and environment fusion understanding layer includes a motion intent decoder and an environment terrain parser;

[0018] The motion intent decoder includes a frequency domain gait analysis submodule, an instantaneous interaction force analysis submodule, and a lightweight motion prediction submodule.

[0019] The frequency domain gait analysis submodule performs time-frequency transformation on the joint angle sequence and extracts the dominant gait frequency and symmetry index as the first intention feature;

[0020] The instantaneous interactive force analysis submodule analyzes the force center trajectory and torque direction output by the multi-dimensional force sensor as a second intention feature;

[0021] The lightweight motion prediction submodule uses a recurrent neural network model, taking historical joint motion sequences as input, to predict joint motion trends within future time windows, which serve as third intent features.

[0022] The motion intention decoder also includes an intention fusion arbitrator, which takes the first intention feature, the second intention feature, the third intention feature, and the current gait phase signal as input, and outputs a comprehensive patient motion intention vector through a weighted fusion strategy.

[0023] The mathematical expression of the weighted fusion strategy is as follows:

[0024]

[0025] in, This represents the overall motion intent vector. Indicates the first The intent feature vector output by each submodule These correspond to the frequency domain gait analysis submodule, the instantaneous interaction force analysis submodule, and the lightweight motion prediction submodule, respectively. This indicates the current gait phase, determined in real time by the joint angle sequence. Indicates phase with gait The relevant adaptive weight coefficients satisfy And the weight coefficient mapping relationship Determined through a predefined fuzzy logic rule table.

[0026] Furthermore, the dual-sided differentiated collaborative planning layer includes a trajectory dynamic synthesis unit and a task deconstruction and allocation unit;

[0027] The trajectory dynamic synthesis unit uses the standard gait trajectory as a base and scales and deforms the base trajectory in real time and spatial amplitude according to the motion intention vector to generate a personalized gait trajectory.

[0028] The task deconstruction and allocation unit deconstructs the personalized gait trajectory into a healthy side support phase task and an affected side swing phase task based on the human biomechanics inverse dynamics model.

[0029] For the healthy side support phase task, the task deconstruction and allocation unit calculates the desired end force field parameters and desired interaction impedance parameters used to maintain balance and provide propulsion, which constitute the first control target set;

[0030] For the affected side swing phase task, the task deconstruction and allocation unit calculates the ideal joint moment curve in conjunction with the expected terrain disturbance vector, and scales the ideal joint moment curve according to the patient's bilateral limb muscle strength assessment data to generate the feedforward auxiliary moment curve. At the same time, the desired joint trajectory and variable joint impedance parameters are set to form the second control target set.

[0031] Furthermore, the health-side execution control unit is specifically used for:

[0032] A force-potential hybrid control law based on the admittance model is established. The input of the control law is the deviation between the desired end force field and the measured interaction force of the multidimensional force sensor.

[0033] The force deviation is converted into joint position adjustment using the admittance model.

[0034] The joint position adjustment amount is superimposed with a loose joint reference position command to generate the first motor control command, which is then sent to the healthy joint motor driver.

[0035] Furthermore, the affected-side execution control unit is specifically used for:

[0036] Establish a feedforward-feedback composite impedance control law;

[0037] The feedforward channel of the control law directly outputs the feedforward auxiliary torque curve as a torque command.

[0038] The feedback channel of the control law takes the positional deviation between the desired joint trajectory and the measured trajectory of the joint encoder as input, and calculates the feedback compensation torque through a variable impedance model. The stiffness and damping parameters of the variable impedance model are adjusted online in real time according to the signal amplitude of the surface electromyography sensor.

[0039] The torque command output from the feedforward channel is superimposed with the feedback compensation torque output from the feedback channel to generate the second motor control command, which is then sent to the motor driver of the affected joint.

[0040] Furthermore, the environmental terrain parser is specifically used for:

[0041] The system receives data from the depth vision camera and generates a terrain elevation map and a semantic map in the robot's body coordinate system through a synchronous localization and mapping algorithm and a semantic segmentation network.

[0042] Based on the current robot speed and the motion intention vector, predict the next expected landing point position;

[0043] Based on the terrain elevation map, the terrain height change value and slope angle at the expected landing point are calculated. Combined with the robot and human dynamics model, the additional joint torque requirement required to perform the next action is calculated, and the output is the expected terrain disturbance vector.

[0044] Furthermore, the system also includes a central security coordinator;

[0045] The central safety coordinator is connected to the distributed collaborative execution layer and the multi-source heterogeneous perception layer, and monitors all sensor signals, bilateral joint motion status, and the first motor control command and the second motor control command in real time.

[0046] The central safety coordinator is pre-set with joint motion angle limit thresholds, joint interaction torque safety thresholds, and bilateral motion coordination thresholds.

[0047] When any monitored quantity exceeds the corresponding threshold, the central safety coordinator sends a highest-priority interrupt signal to the distributed collaborative execution layer, controlling all joint actuators to enter a preset safety damping stop mode.

[0048] Furthermore, the system also includes a data management and remote service layer;

[0049] The data management and remote service layer is used to store the raw data collected by the multi-source heterogeneous perception layer, the historical records of the control target set generated by the dual-sided differentiated collaborative planning layer, and the execution logs of the distributed collaborative execution layer.

[0050] The data management and remote service layer constructs a digital twin model of the patient based on historical data, and uses a continuous learning algorithm to optimize and update the neural network model parameters of the lightweight motion prediction submodule and the weight coefficient mapping relationship of the intent fusion arbitrator offline.

[0051] The data management and remote service layer securely distributes the optimized model parameters to the intent and environment fusion understanding layer through an encrypted communication link, enabling remote iteration and personalized evolution of the control strategy.

[0052] Furthermore, the method for obtaining the patient's bilateral limb muscle strength assessment data includes:

[0053] During the initial system calibration phase, the control robot guides the patient's affected limb to perform a set of standardized isokinetic concentric and eccentric contraction movements.

[0054] The maximum active output torque of the affected joint at various motion angles is measured using the joint torque sensor.

[0055] The maximum active output torque of the healthy joint was measured in the same manner.

[0056] Calculate the ratio of the maximum active output torque of the corresponding joint groups on the affected side to that on the healthy side, and store this ratio as the muscle strength proportionality coefficient K in the patient's bilateral limb muscle strength assessment data.

[0057] Compared with existing technologies, this multimodal feedback collaborative control system for lower limb rehabilitation robots has the following advantages:

[0058] I. This invention uses a multi-source heterogeneous perception layer to collect real-time motion state signals of the patient's lower limbs, human bioelectric signals, and external environmental terrain visual signals. These signals are then fused and processed by an intent and environment fusion understanding layer to generate a comprehensive motion intent vector and a terrain disturbance expectation vector. A bilateral differential collaborative planning layer then dynamically synthesizes personalized gait trajectory targets by combining the patient's bilateral limb muscle strength assessment data and deconstructs them into a bilateral differential control target set. Finally, a distributed collaborative execution layer uses a differential control algorithm to drive the actuator movement. This achieves proactive fusion understanding of the patient's real-time motion intent and the dynamic environment, as well as the dynamic generation of collaborative motion strategies. This effectively improves the naturalness and adaptability of rehabilitation training and helps patients rebuild motor function more efficiently.

[0059] Second, this invention employs a bilateral differentiated control strategy, setting different control target sets and using adaptive control algorithms to address the functional differences between the healthy and affected limbs. This enables the healthy side to achieve compliant assisted movement while the affected side receives active guidance and assistance, aligning with the biomechanical characteristics of human movement and improving the coordination and comfort of human-computer interaction. Simultaneously, the central safety coordinator monitors various signals in real time and triggers a safety stop mode in case of abnormalities, ensuring the safety of the training process. The data management and remote service layer, through the construction of a digital twin model of the patient and continuous optimization of control parameters, achieves personalized evolution of the control strategy, adapting to the patient's rehabilitation progress and further enhancing the pertinence and effectiveness of rehabilitation training, providing patients with safe, personalized, and continuously optimized rehabilitation training support.

[0060] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0062] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0063] Figure 2 This is a flowchart illustrating the process of the intent and environment integration understanding layer of this invention;

[0064] Figure 3 This is a flowchart illustrating the control command generation and execution process of the present invention. Detailed Implementation

[0065] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0066] Example 1

[0067] like Figures 1 to 3 As shown in the accompanying drawings, the specific embodiments of the present invention will be further described in detail below. This embodiment uses a rehabilitation training robot for patients with lower limb motor dysfunction as an application scenario. The implementation process of the multimodal feedback lower limb rehabilitation robot collaborative control system will be described in detail below with reference to the specific implementation content.

[0068] The multimodal feedback lower limb rehabilitation robot collaborative control system of this invention adopts a layered architecture design, with each layer cascaded sequentially and data transmission using a real-time communication protocol to ensure low latency and reliability of signal transmission. From top to bottom, the system includes a multi-source heterogeneous perception layer, an intent and environment fusion understanding layer, a dual-sided differentiated collaborative planning layer, and a distributed collaborative execution layer. It is also equipped with a central safety coordinator and a data management and remote service layer, respectively realizing safety monitoring and iterative optimization of control strategies. Data interaction between layers occurs through standardized data interfaces, with data format unified to a preset structured format, ensuring compatibility and collaboration between modules.

[0069] In this embodiment, the multi-source heterogeneous sensing layer is responsible for collecting in real time the patient's lower limb internal motion state signals, human bioelectric signals, and the terrain visual signals of the robot's external environment. Its specific implementation is as follows.

[0070] The internal state sensing module consists of a rotary encoder, a joint torque sensor, a surface electromyography sensor, and a multi-dimensional force sensor. The selection, installation method, and working principle of each sensor are as follows.

[0071] Both the rotary encoder and the joint torque sensor are mounted on the robot's hip and knee joints, coaxially arranged with the joint actuators. The rotary encoder is a high-precision incremental encoder with a measurement accuracy better than 0.1 degrees. It collects pulse signals during joint rotation, which are then converted into joint angle and angular velocity data by signal processing circuitry, providing basic data for subsequent gait analysis and motion intention recognition. The joint torque sensor is a strain gauge torque sensor with a range matched to the robot's maximum output torque. By sensing strain changes during joint movement, it converts these changes into human-machine interface torque data to determine the patient's active force exertion in the limbs.

[0072] The surface electromyography (EMG) sensors are placed on the surface of the rectus femoris and biceps femoris muscles in the patient's thigh. The sensor electrodes are medical-grade silver chloride electrodes to ensure good contact with the skin and biocompatibility. The sensors collect the weak bioelectrical signals generated during muscle contraction, which are then processed by a preamplifier and filter circuit to output a stable EMG signal. This signal reflects the muscle's activation state and degree of force exertion, providing a basis for decoding movement intentions and adjusting variable impedance parameters.

[0073] A multi-dimensional force sensor is installed on the robot's lower leg link near the sole of the foot. A three-dimensional force sensor is selected to measure the ground reaction force and human-robot interaction force in three-dimensional space. The sensor's measurement axis is aligned with the mechanical axis of the robot's lower leg link. By sensing the contact force between the sole of the foot and the ground, as well as the interaction force between the patient's limb and the robot, it outputs the magnitude and direction of the force data for analyzing the force center trajectory and torque direction.

[0074] The core component of the external environment perception module is a depth vision camera, which is fixed on a mounting bracket on the front side of the robot's torso with its optical axis pointing diagonally downwards. Its field of view can completely cover the area in front of the robot. The depth vision camera is based on the time-of-flight principle, and its resolution and frame rate meet the requirements for real-time terrain perception, enabling it to quickly acquire depth image data of the environment ahead.

[0075] The terrain parsing algorithm built into the external environment perception module is implemented as follows: First, the original depth image acquired by the depth vision camera is preprocessed, including denoising, image alignment, and coordinate transformation, converting the image pixel coordinates into three-dimensional coordinates in the robot's body coordinate system; then, the preprocessed depth image is semantically classified through a semantic segmentation network to identify different terrain types such as flat ground, slopes, and steps; finally, key feature parameters of the terrain are extracted, including terrain height, slope, and flatness, to generate a forward terrain map containing height information and terrain semantic classification, providing data support for the subsequent generation of terrain disturbance expectation vectors.

[0076] In this embodiment, the intent and environment fusion understanding layer is connected to the multi-source heterogeneous perception layer, receives and processes various signals collected by the multi-source heterogeneous perception layer, and generates a comprehensive patient motion intent vector and terrain disturbance expectation vector, which is specifically implemented as follows.

[0077] The motion intent decoder consists of a frequency domain gait analysis submodule, an instantaneous interaction force parsing submodule, a lightweight motion prediction submodule, and an intent fusion arbitrator. The implementation process and function of each submodule are as follows.

[0078] The specific implementation of the frequency domain gait analysis submodule is as follows: It receives the joint angle sequence acquired by the rotary encoder, performs time-frequency transformation on this sequence, selects the short-time Fourier transform as the time-frequency transformation method, and converts the time-domain joint angle sequence into time-frequency domain data by setting an appropriate time window and window function. From the time-frequency domain data, it extracts the dominant gait frequency and symmetry index as the first intentional features. The dominant gait frequency reflects the rhythmic characteristics of the patient's gait, and the symmetry index reflects the degree of symmetry in the movement of the patient's two limbs.

[0079] The instantaneous interactive force analysis submodule is implemented as follows: it receives three-dimensional ground reaction force and interactive force data output from a multi-dimensional force sensor, calculates the resultant force direction and point of application, obtains the force center trajectory and torque direction, and uses these as secondary intent features. The force center trajectory reflects the change in the patient's center of gravity during walking, and the torque direction reflects the patient's force application trend, providing supplementary information for motion intent recognition.

[0080] The lightweight motion prediction submodule employs a recurrent neural network model, using a long short-term memory network as its core structure. This network possesses excellent time-series data processing capabilities and long-term dependency learning abilities. The network input consists of historical joint motion sequences, selecting joint angle and angular velocity data from the most recent period as the input vector. The network output is a predicted value of joint motion trends within a future time window, which is used as the third intention feature. The network training process involves using gait data from healthy individuals and patients with lower limb motor dysfunction as training datasets. Backpropagation is used to iteratively optimize the network parameters until the prediction error meets preset requirements.

[0081] The core of the intention-to-merge arbitrator is the weighted merging strategy, whose mathematical expression is:

[0082]

[0083] in, The integrated motion intent vector is the final expression of intent after fusion, which can comprehensively reflect the patient's motion needs; This represents the intent feature vector output by the i-th submodule. When i=1, it corresponds to the first intent feature output by the frequency domain gait analysis submodule; when i=2, it corresponds to the second intent feature output by the instantaneous interaction force analysis submodule; and when i=3, it corresponds to the third intent feature output by the lightweight motion prediction submodule. The current gait phase is determined in real time by the joint angle sequence. The specific determination process is as follows: by analyzing the change pattern of joint angles and combining the preset gait phase division criteria, the gait cycle is divided into the support phase and the swing phase, and then further subdivided into multiple sub-phases. This represents the adaptive weighting coefficients related to gait phase, used to adjust the contribution of each intention feature under different gait phases.

[0084] The weighting coefficient mapping relationship is determined through a predefined fuzzy logic rule table. The process of establishing the fuzzy logic rule table is as follows: based on statistical analysis of a large amount of gait data, the importance of each intention feature under different gait phases is determined. For example, in the mid-stability phase, the second intention feature output by the instantaneous interaction force analysis submodule more accurately reflects the motion intention, and therefore has a larger corresponding weighting coefficient; in the mid-swing phase, the third intention feature output by the lightweight motion prediction submodule is more important, and its weighting coefficient increases accordingly. The weighting coefficients satisfy... This ensures the rationality and consistency of the fused motion intent vector.

[0085] The specific implementation process of the environmental terrain parser is as follows: First, it receives depth image data collected by a depth vision camera, processes the depth image data through a synchronous localization and mapping algorithm, and generates a terrain elevation map in the robot's body coordinate system. This algorithm can update the three-dimensional terrain information of the robot's surrounding environment in real time to ensure the accuracy of terrain perception. At the same time, it performs semantic classification on the depth image through a semantic segmentation network to generate a semantic map and clarify the terrain type of different areas.

[0086] Then, based on the current robot speed and motion intention vector, the next expected landing point is predicted. The specific prediction method is as follows: combining the robot's current speed and motion direction information in the motion intention vector, the robot's travel distance and direction in the future are calculated based on kinematic principles, thereby determining the approximate location of the next expected landing point.

[0087] Finally, based on the terrain elevation map, the terrain height change and slope angle at the expected landing point are calculated. Combined with the robot and human dynamics models, the additional joint torque required to execute the next action is calculated through dynamic equations. This requirement is output as the terrain disturbance expectation vector, providing environmental reference information for the dynamic synthesis of the subsequent gait trajectory target.

[0088] In this embodiment, the bilateral differentiated collaborative planning layer connects to the intent and environment fusion understanding layer, receives the motion intent vector and the terrain disturbance expectation vector, and combines the pre-stored patient bilateral limb muscle strength assessment data to dynamically synthesize personalized gait trajectory targets and perform task deconstruction. Its specific implementation is as follows.

[0089] During the initial system calibration phase, the patient's bilateral limb muscle strength assessment process is executed. Specifically, the robot's affected-side joint actuator guides the patient's affected limb to perform a set of standardized isokinetic concentric and eccentric contraction movements according to preset motion parameters. During the movement, the joint movement speed remains constant, and the range of motion covers the normal range of motion of the lower limb joints.

[0090] The maximum active output torque of the affected joint at various motion angles is measured in real time using a joint torque sensor, and the peak torque corresponding to each angle is recorded. Using the same motion parameters and measurement methods, the robot is controlled to guide the patient's healthy limb to perform the same contraction movement, and the maximum active output torque of the healthy joint at various motion angles is measured.

[0091] The ratio of the maximum active output torque of the corresponding joint groups on the affected side to that on the healthy side is calculated. This ratio is stored as the muscle strength ratio coefficient K in the patient's bilateral limb muscle strength assessment data. This coefficient can reflect the recovery of muscle strength of the patient's affected limb relative to the healthy limb, and provides a basis for the subsequent generation of feedforward auxiliary torque curves.

[0092] The trajectory dynamic synthesis unit is based on the standard gait trajectory, which is derived from the gait database of healthy people. It is an average gait trajectory obtained through statistical analysis, including the angle change curves of the hip and knee joints during the gait cycle.

[0093] The gait trajectory is scaled and deformed in real time based on the motion intention vector. The specific implementation process is as follows: the motion intensity and rhythm information in the motion intention vector are extracted, and the time axis of the gait trajectory is scaled to match the patient's motion rhythm; at the same time, the spatial amplitude of the gait trajectory is deformed and adjusted according to the motion amplitude information in the motion intention vector and the environmental constraint information in the terrain disturbance expectation vector, so that the gait trajectory can adapt to the patient's motion ability and environmental changes, and finally generate a personalized gait trajectory that can fully fit the patient's individual characteristics and actual motion needs.

[0094] The task decomposition and allocation unit is based on the human biomechanical inverse dynamics model. This model is established based on the anatomical structure and biomechanical characteristics of the human lower limbs and is derived through the Lagrange equation. It can accurately reflect the mechanical relationship during the movement of the human lower limbs.

[0095] Personalized gait trajectories are deconstructed into a support phase task on the healthy side and a swing phase task on the affected side. The specific deconstruction method is as follows: based on gait phase information, the movement requirements of the healthy limb when it is in the support phase and the movement requirements of the affected limb when it is in the swing phase are clarified. Based on the human biomechanical inverse dynamics model, the movement task objectives of both limbs under the corresponding gait phase are calculated respectively.

[0096] For the healthy side support phase task, the task deconstruction and allocation unit calculates the desired end-effector force field parameters and desired cross-impedance parameters for maintaining balance and providing propulsion using a human biomechanical inverse dynamics model, forming the first control target set. The desired end-effector force field parameters define the force output range and distribution that the healthy foot needs to achieve, while the desired cross-impedance parameters are used to adjust the compliance of the interaction between the healthy limb and the robot, ensuring that the healthy limb can stably exert force and maintain body balance during the support process.

[0097] For the swing phase task on the affected side, the task deconstruction and allocation unit combines the expected terrain disturbance vector with the human biomechanical inverse dynamics model to calculate the ideal joint torque curve. This curve can meet the torque requirements of the affected limb to complete the swing movement under the current terrain conditions. Then, based on the muscle strength ratio coefficient K in the muscle strength assessment data of the patient's bilateral limbs, the ideal joint torque curve is scaled proportionally to generate a feedforward auxiliary torque curve, ensuring that the auxiliary torque matches the patient's muscle strength recovery and avoiding over- or under-assistance. At the same time, the desired joint trajectory is set according to the personalized gait trajectory, and the variable joint impedance parameter is set in combination with the patient's real-time movement state. The desired joint trajectory, variable joint impedance parameter and feedforward auxiliary torque curve together constitute the second control target set.

[0098] In this embodiment, the distributed collaborative execution layer connects the two-sided differentiated collaborative planning layer, including the healthy side execution control unit and the affected side execution control unit, which respectively generate corresponding motor control commands to drive the robot joint actuators to move. The specific implementation is as follows.

[0099] The control unit on the healthy side adopts a force-position hybrid control algorithm based on the admittance model. The specific implementation process is as follows: First, a force-position hybrid control law based on the admittance model is established. The mathematical expression of the admittance model is:

[0100]

[0101] Where M is the inertial parameter, B is the damping parameter, and K is the stiffness parameter, these parameters are preset according to the motion characteristics and interaction requirements of the healthy limb; Joint acceleration x represents the joint angular velocity and the joint position adjustment amount. This represents the deviation between the desired end force field and the actual interaction force measured by the multidimensional force sensor.

[0102] The input to the control law is the deviation between the desired end force field and the measured interaction force of the multi-dimensional force sensor. This force deviation is converted into a joint position adjustment amount through the admittance model. The specific conversion process is as follows: the force deviation is substituted into the admittance model, and the joint acceleration, joint angular velocity and joint position adjustment amount are obtained through numerical calculation.

[0103] The joint position adjustment amount is superimposed with a lenient joint reference position command to generate a first motor control command. This joint reference position command is based on a personalized gait trajectory setting and has a certain tolerance range, which can adapt to position adjustments caused by force deviations. The first motor control command is sent to the healthy side joint motor driver to drive the robot's healthy side joint actuator to achieve compliant assisted movement, ensuring that the healthy side limb can exert force flexibly during movement, while maintaining good interaction with the robot.

[0104] The affected side execution control unit adopts a composite control algorithm that integrates feedforward torque compensation and feedback impedance control. The specific implementation process is as follows: a feedforward-feedback composite impedance control law is established, which consists of a feedforward channel and a feedback channel.

[0105] The feedforward channel of the control law directly outputs the feedforward auxiliary torque curve as a torque command. This torque command can provide an active guiding torque to the affected limb, helping the affected limb to complete the preset gait trajectory.

[0106] The feedback channel of the control law takes the positional deviation between the desired joint trajectory and the measured trajectory of the joint encoder as input, and calculates the feedback compensation torque through a variable impedance model. The mathematical expression of the variable impedance model is as follows:

[0107]

[0108] in, For feedback compensation torque; Stiffness parameters of the variable impedance model These are the damping parameters for the variable impedance model; Positional deviation This refers to the speed deviation.

[0109] The stiffness and damping parameters are adjusted online in real time based on the amplitude of the surface electromyography (SEMG) sensor signal. The specific adjustment logic is as follows: when the SEMG sensor signal amplitude is large, it indicates that the patient's active force exertion ability on the affected side is strong, so the stiffness and damping parameters are appropriately reduced to decrease the robot's assistance force and encourage the patient to move actively; when the SEMG sensor signal amplitude is small, it indicates that the patient's active force exertion ability on the affected side is weak, so the stiffness and damping parameters are appropriately increased to increase the robot's assistance force and ensure accurate tracking of the gait trajectory.

[0110] The torque command output from the feedforward channel is superimposed with the feedback compensation torque output from the feedback channel to generate a second motor control command, which is sent to the motor driver of the affected joint to drive the actuator of the robot's affected joint to achieve active guidance and assisted movement, helping the patient's affected limb to gradually recover motor function.

[0111] In this embodiment, the central security coordinator is connected to the distributed collaborative execution layer and the multi-source heterogeneous perception layer. Its core function is to realize the security monitoring and emergency protection of the system, which is specifically implemented as follows.

[0112] The central safety coordinator receives real-time sensor signals from all sources in the heterogeneous sensing layer, the bilateral joint motion status of the distributed collaborative execution layer, and the control commands of the first and second motors, and performs real-time analysis and monitoring of this data.

[0113] The central safety coordinator is pre-set with joint motion angle limit thresholds, joint interaction torque safety thresholds, and bilateral motion coordination thresholds. These thresholds are set based on the physiological limits of human lower limb joint movement, the safety limitations of the robot's mechanical structure, and the clinical safety standards for lower limb rehabilitation training, ensuring the rationality and safety of the thresholds.

[0114] When the central safety coordinator detects any monitored quantity exceeding the corresponding threshold, it immediately determines that the system is in an unsafe state. At this time, it sends a high-priority interrupt signal to the distributed collaborative execution layer. This interrupt signal can forcibly stop the execution of the current motor control commands. After receiving the interrupt signal, the distributed collaborative execution layer controls all joint actuators to enter a preset safety damping stop mode. By adjusting the damping parameters of the joint actuators, the joints gradually stop moving, avoiding impact injuries to the patient from sudden stops and ensuring the patient's safety during rehabilitation training.

[0115] In this embodiment, the data management and remote service layer is used to implement the data storage model construction and control strategy optimization, and its specific implementation is as follows.

[0116] The data management and remote service layer uses a distributed database to store the raw data collected by the multi-source heterogeneous perception layer, the historical records of the control target set generated by the dual-sided differentiated collaborative planning layer, and the execution logs of the distributed collaborative execution layer. The database supports high-concurrency data writing and fast querying, ensuring the stability and availability of data storage.

[0117] A digital twin model of the patient is constructed based on stored historical data. This model is based on a multibody dynamics model. By inputting the patient's physiological parameters, motion data and rehabilitation training records, the model parameters are calibrated and optimized, so that the digital twin model can accurately reflect the patient's actual motion state and rehabilitation progress.

[0118] The continuous learning algorithm is used to optimize and update the neural network model parameters and weight coefficient mapping relationship of the intent fusion arbitrator of the lightweight motion prediction submodule offline. The continuous learning algorithm can use the newly added rehabilitation training data to iteratively adjust the model parameters without forgetting the existing knowledge, so that the model can adapt to the changes in the patient's state during the rehabilitation process.

[0119] The data management and remote service layer securely distributes optimized model parameters to the intent and environment fusion understanding layer via an encrypted communication link. This encrypted communication link employs a mature encryption protocol to ensure the security and integrity of data transmission, preventing parameter tampering or leakage. This process enables remote iteration and personalized evolution of control strategies, allowing the control system to continuously adapt to the patient's rehabilitation progress and improve rehabilitation training effectiveness.

[0120] This embodiment achieves comprehensive acquisition of multi-dimensional signals through a multi-source heterogeneous sensing layer, providing reliable data support for understanding movement intentions and environmental perception. Through a weighted fusion strategy and terrain analysis algorithm in the intention and environment fusion understanding layer, it achieves accurate analysis of movement intentions and terrain disturbances, enabling control strategies to better align with the patient's actual needs and environmental changes. Through dynamic trajectory synthesis and task deconstruction in the bilateral differentiated collaborative planning layer, combined with the patient's bilateral limb muscle strength assessment data, it achieves personalized control target setting for the healthy and affected sides. Through differentiated control algorithms in the distributed collaborative execution layer, it respectively realizes compliant assisted movements of the healthy limbs and active guided and assisted movements of the affected limbs, improving the coordination and comfort of human-computer interaction. The central safety coordinator ensures the safety of the rehabilitation training process, and the continuous optimization capabilities of the data management and remote service layers enable the control system to adapt to the patient's rehabilitation progress, further improving the effectiveness and adaptability of rehabilitation training.

[0121] Example 2

[0122] like Figures 1 to 3 As shown, this embodiment takes home rehabilitation training for patients with lower limb motor dysfunction after stroke as the application scenario, and focuses on illustrating the complete workflow of the multimodal feedback lower limb rehabilitation robot collaborative control system, including the entire execution process of system startup, calibration, real-time operation, safety monitoring and strategy optimization. The following is a detailed description in conjunction with the specific implementation content.

[0123] The overall system architecture is as follows:

[0124] The system in this embodiment still adopts a layered architecture design. A multi-source heterogeneous perception layer, an intent and environment fusion understanding layer, a dual-sided differentiated collaborative planning layer, and a distributed collaborative execution layer are cascaded sequentially. The central security coordinator and the data management and remote service layer establish real-time data interaction with each core layer. Low-latency data transmission is achieved between layers via industrial Ethernet, and the data format adopts a standardized binary stream format to ensure data transmission stability and inter-module compatibility. After system startup, each layer completes initialization in a preset order. After the calibration process is completed, it enters real-time operation. The entire process is monitored by the central security coordinator, while the data management and remote service layer synchronously completes data storage and model optimization.

[0125] The system startup and initial calibration process is as follows:

[0126] After system startup, a self-test program is executed first, with each module sequentially completing hardware initialization and communication link verification. All sensors in the multi-source heterogeneous sensing layer initiate self-tests: the rotary encoder, joint torque sensor, surface electromyography sensor, multi-dimensional force sensor, and depth vision camera each output self-test signals. Once the sensors are confirmed to be functioning correctly, a ready signal is sent back to the central safety coordinator. The healthy and affected joint actuators in the distributed collaborative execution layer undergo no-load operation to verify the consistency of response between the motor driver and the actuator, ensuring no jamming or abnormal vibration.

[0127] After the self-test is completed, the initial calibration phase begins. The core task is to obtain muscle strength assessment data for both sides of the patient's limbs. First, the rehabilitation therapist assists the patient in wearing the robotic device, adjusting the tightness of the straps to ensure good contact between the sensors and the body. The electrode pads of the surface electromyography sensor are attached to the rectus femoris and biceps femoris muscle bellies, and the multidimensional force sensor is attached tightly to the foot or calf linkage without any looseness.

[0128] The calibration process officially begins. The system controls the actuator on the affected side to operate according to preset motion parameters, guiding the patient's affected limb to complete standardized isokinetic concentric and eccentric contraction movements. During the movement, the joint movement speed remains constant, covering the normal range of motion of the hip and knee joints. The joint torque sensor collects the maximum active output torque of the affected joint at each movement angle in real time, and simultaneously records the joint angle data from the rotary encoder to ensure that the timestamps of the torque data and angle data are consistent.

[0129] After data acquisition on the affected side is completed, the actuator on the healthy side is controlled using the same motion parameters and procedures to guide the patient's healthy limb to complete the same contraction movement. The joint torque sensor also collects the maximum active output torque at various angles of the healthy joint. The system automatically calculates the ratio of the maximum active output torque of the corresponding joint groups on the affected and healthy sides, storing this ratio as a muscle strength proportionality coefficient in the system's local database to provide foundational data for subsequent bilateral differentiated collaborative planning.

[0130] The initial calibration phase also includes sensor parameter calibration, zero-point calibration and range calibration of the multi-dimensional force sensor using a standard force source to ensure force measurement accuracy; internal and external parameter calibration of the depth vision camera to correct imaging distortion and ensure the accuracy of terrain detection; and baseline calibration of the surface electromyography sensor by collecting signals from the patient at rest to eliminate signal drift caused by environmental interference.

[0131] The real-time operation process is as follows:

[0132] After the initial calibration is completed, the system enters real-time operation, and each layer works collaboratively according to the following process.

[0133] Data acquisition process of the multi-source heterogeneous sensing layer: The internal state sensing module and the external environment sensing module of the multi-source heterogeneous sensing layer simultaneously start data acquisition. In the internal state sensing module, the rotary encoder continuously collects joint angle data of the hip and knee joints, generates a continuous joint angle sequence by counting pulse signals in real time, and then calculates the joint angular velocity; the joint torque sensor simultaneously collects the human-machine interaction torque, capturing the force changes between the patient's limb and the robot actuator; the surface electromyography sensor continuously collects the electromyographic signals of the rectus femoris and biceps femoris muscles, and outputs a stable signal sequence after filtering and amplification; the multi-dimensional force sensor measures the three-dimensional ground reaction force and interaction force in real time, capturing the force changes of the foot in contact with the ground and the force transmission of human-machine interaction.

[0134] The depth vision camera in the external environment perception module continuously captures depth images of the area in front of the robot, with the optical axis pointing diagonally downwards to ensure that the field of view covers the area where the robot might land next. The raw image data acquired by the depth vision camera is transmitted in real time to the built-in terrain analysis module. After denoising, image alignment, and coordinate transformation, 3D point cloud data in the robot's body coordinate system is generated. Then, semantic segmentation is used to identify the terrain type and extract key parameters such as terrain height and slope to form a complete terrain map of the area ahead.

[0135] All collected internal motion state signals, human bioelectric signals, and terrain visual signals are packaged and transmitted to the intent and environment fusion understanding layer at fixed time intervals. During the transmission process, a data verification mechanism is used to ensure signal integrity.

[0136] The signal processing flow of the intent and environment fusion understanding layer is as follows:

[0137] After receiving the signals transmitted by the multi-source heterogeneous perception layer, the intent and environment fusion understanding layer processes the motion intent signal and the terrain visual signal in parallel in two paths.

[0138] In the motion intent decoding process, the frequency domain gait analysis submodule receives joint angle sequences, performs time-frequency transformation processing on them, and extracts the dominant gait frequency and symmetry index to reflect the rhythmic characteristics of the patient's gait and the degree of bilateral motion symmetry; the instantaneous interactive force analysis submodule receives force data from multi-dimensional force sensors, analyzes the force center trajectory and torque direction, and captures the patient's force application trend and center of gravity changes; the lightweight motion prediction submodule receives historical joint motion sequences and outputs the joint motion trend within the future time window through a preset prediction model.

[0139] The intent fusion arbitrator receives the output features from the three sub-modules mentioned above, and simultaneously acquires the current gait phase determined in real time by the joint angle sequence. It then fuses the three features according to a preset weighted fusion strategy. The weight coefficients of each feature under different gait phases are adjusted according to predefined rules, with higher weights for instantaneous interaction force features during the support phase and higher weights for motion prediction features during the swing phase. After fusion, a comprehensive patient motion intent vector is generated, fully reflecting the patient's motion needs and trends.

[0140] In the environmental terrain analysis process, the environmental terrain analyzer receives terrain map data processed by the depth vision camera, combines it with the current robot movement speed and movement intention vector, and predicts the next expected landing point location. Based on the terrain map, it calculates the terrain height change and slope angle at the landing point, and combines the robot and human dynamics characteristics to calculate the additional joint torque required to complete the next action, generating a terrain disturbance prediction vector, and clarifying the constraints and requirements of the environment on gait movement.

[0141] After the motion intention vector and the terrain disturbance expectation vector are generated, they are synchronously transmitted to the bilateral differentiated collaborative planning layer.

[0142] The trajectory planning and task deconstruction process of the two-sided differentiated collaborative planning layer is as follows:

[0143] After receiving the motion intention vector and the expected terrain disturbance vector, the bilateral differential collaborative planning layer first retrieves the patient's bilateral limb muscle strength assessment data stored in the initial calibration stage, and then performs personalized gait trajectory synthesis in combination with the standard gait trajectory base.

[0144] The trajectory dynamic synthesis unit is based on the standard gait trajectory of healthy individuals, adjusting the time axis and spatial amplitude of the trajectory according to the movement intention vector. When the movement rhythm is fast, the trajectory time axis is shortened; when the movement amplitude is large, the trajectory spatial range is expanded. At the same time, the spatial shape of the trajectory is adjusted in combination with the expected vector of terrain disturbance. For example, when encountering a slope, the angle change range of the hip and knee joints is adjusted to ensure that the trajectory adapts to the terrain undulations, ultimately generating a personalized gait trajectory that fits the patient's movement ability and environmental conditions.

[0145] Based on the human biomechanical inverse dynamics model, the task decomposition and allocation unit decomposes the personalized gait trajectory into a healthy side support phase task and an affected side swing phase task. For the healthy side support phase task, the desired end-effector force field parameters and desired cross-impedance parameters required to maintain body balance and provide propulsion are calculated to form the first control target set, ensuring that the healthy limb can exert force stably and smoothly during the support process. For the affected side swing phase task, the ideal joint moment curve is calculated by combining the expected terrain disturbance vector. The curve is scaled according to the muscle force proportionality coefficient to obtain the feedforward auxiliary moment curve. At the same time, the desired joint trajectory and variable joint impedance parameters are set to form the second control target set, ensuring that the affected limb accurately completes the swinging movement under the guidance of the robot.

[0146] The first control target set and the second control target set are respectively transmitted to the healthy side execution control unit and the affected side execution control unit of the distributed collaborative execution layer.

[0147] The control instruction generation and execution process of the distributed collaborative execution layer is as follows:

[0148] After receiving the control target set transmitted by the two sides of the differentiated collaborative planning layer, the two execution control units respectively carry out control calculations.

[0149] The healthy side actuator control unit adopts a force-position hybrid control strategy based on an admittance model. It receives the desired end force field parameters and desired cross-impedance parameters from the first control target set, and calculates the force deviation by combining the measured cross-impedance forces collected in real time by multi-dimensional force sensors. The force deviation is converted into a joint position adjustment amount through the admittance model, which is then superimposed with a preset relaxed joint reference position command to generate a first motor control command. This command is sent to the healthy side joint motor driver to drive the healthy side joint actuator to perform compliant assisted movement. The patient's healthy limb can exert force flexibly with robot assistance to maintain motor balance.

[0150] The affected-side actuator control unit employs a composite control strategy integrating feedforward torque compensation and feedback impedance control. It receives the desired joint trajectory variable joint impedance parameters and feedforward auxiliary torque curve from the second control target set. The feedforward channel directly outputs the feedforward auxiliary torque curve as the basic torque command, while the feedback channel receives the measured trajectory acquired by the joint encoder, calculates the positional deviation from the desired joint trajectory, and calculates the feedback compensation torque through a variable impedance model. The stiffness and damping parameters of the variable impedance model are adjusted in real time based on the signal amplitude of the surface electromyography sensor. When the patient's active force is strong, the parameter values ​​are reduced to decrease assistance; when the active force is weak, the parameter values ​​are increased to enhance guidance. The feedforward torque command and the feedback compensation torque are superimposed to generate a second motor control command, which is sent to the affected-side joint motor driver to drive the affected-side joint actuator to achieve active guidance and assisted movement, helping the patient's affected limb follow a personalized gait trajectory and gradually restore motor function.

[0151] The real-time monitoring process of the Central Security Coordinator is as follows:

[0152] The central safety coordinator works continuously throughout the entire system operation process, receiving in real time all sensor signals from the multi-source heterogeneous perception layer, the joint motion status of the distributed collaborative execution layer, and two motor control commands.

[0153] The central safety coordinator incorporates limit thresholds for joint motion angles, safety thresholds for joint interaction torques, and thresholds for bilateral motor coordination. These thresholds are set based on human physiological limits, robot mechanical safety constraints, and clinical rehabilitation standards. Monitoring data is compared to these thresholds in real time. When a joint motion angle exceeds the limit threshold, it is determined that the limb movement is out of range; when the joint interaction torque exceeds the safety threshold, it is determined that the interaction force is too large and may injure the patient; when bilateral limb motor coordination is below the threshold, it is determined that there is a risk of imbalance and fall.

[0154] Once any monitored quantity is detected exceeding the corresponding threshold, the central safety coordinator immediately sends a high-priority interrupt signal to the distributed collaborative execution layer, forcibly suspending the execution of the current motor control command. Upon receiving the interrupt signal, the distributed collaborative execution layer controls all joint actuators to switch to a safe damping stop mode. By adjusting the actuator damping parameters, the joints are slowly decelerated to a stop, preventing secondary injury to the patient from sudden stop impact. Simultaneously, the system issues audible and visual alarm signals to alert the rehabilitation therapist or caregiver to take timely action.

[0155] The optimization process for the data management and remote service layer is as follows:

[0156] The data management and remote service layer synchronously stores various types of data during system operation, including the raw acquisition data of the multi-source heterogeneous perception layer, the historical records of the control target set of the bilateral differentiated collaborative planning layer, and the execution logs of the distributed collaborative execution layer. All data is stored in categories according to timestamps and supports historical data backtracking and querying.

[0157] A digital twin model of the patient is built based on stored historical data. This model simulates the patient's limb movement characteristics, muscle strength, and rehabilitation progress. The model's accuracy is continuously calibrated by receiving real-time operational data, enabling the model to accurately reflect the patient's actual rehabilitation status.

[0158] A continuous learning algorithm is used to optimize the model parameters and weight coefficient mapping relationship of the lightweight motion prediction submodule and the intent fusion arbitrator offline. The system periodically summarizes rehabilitation training data over a period of time and inputs it into the continuous learning algorithm. While retaining the existing effective parameters, the model parameters and weight coefficient mapping relationship are adjusted according to the patient's rehabilitation progress, making the motion intent decoding more accurate and adapting to the changes in the patient's muscle strength improvement and motor ability recovery.

[0159] The optimized model parameters are sent to the intent and environment fusion understanding layer via an encrypted communication link, updating the corresponding module parameters and enabling remote iteration and personalized evolution of the control strategy. The entire parameter update process is carried out during non-training periods to avoid affecting real-time rehabilitation training, while encrypted transmission ensures parameter security and prevents data leakage or tampering.

[0160] The following are adaptation and optimization schemes for different terrain scenarios, such as the adaptation scheme for slope terrain:

[0161] When the external environment perception module detects a sloping terrain ahead, the environmental terrain parser calculates the slope angle and height changes of the slope, generates the corresponding terrain disturbance prediction vector, and determines the additional joint torque required when the affected limb swings to adapt to the needs of going up or down the slope.

[0162] When synthesizing personalized gait trajectories, the bilateral differentiated collaborative planning layer increases the trajectory amplitude of the hip flexion angle and knee extension angle on the affected side for uphill scenarios, while also increasing the amplitude of the feedforward assist torque curve to provide stronger uphill assistance to the affected limb. For downhill scenarios, the time axis of the gait trajectory is adjusted to prolong the swing time of the affected limb, reduce the amplitude of the feedforward assist torque, and increase the damping value of the variable joint impedance parameter to prevent the patient's limb from descending too quickly due to gravity.

[0163] The affected side execution control unit of the distributed collaborative execution layer optimizes the feedback compensation torque in real time based on the adjusted second control target set, ensuring the movement stability and trajectory accuracy of the affected limb on sloping terrain.

[0164] The terrain adaptation scheme for stepped structures is as follows:

[0165] When the external environment perception module detects that there are steps in front, the environmental terrain parser identifies the height, width and number of steps, predicts the foot placement position, generates a terrain disturbance prediction vector, and determines the height and forward reach of the affected limb.

[0166] When synthesizing personalized gait trajectories, the bilateral differentiated collaborative planning layer adjusts the swing height of the affected limb according to the step height and the limb extension amplitude according to the step width, while extending the swing phase time of the affected limb to ensure that the limb lands smoothly on the step. When generating the second control target set, the task deconstruction and allocation unit increases the torque value of the feedforward auxiliary torque curve during hip flexion and knee extension phases to provide sufficient assistance for limb elevation and extension. Simultaneously, it optimizes the variable joint impedance parameters to increase the stiffness value during the swing process, thereby enhancing trajectory guidance accuracy.

[0167] The healthy side execution control unit adjusts the desired end force field parameters of the first control target set to increase the supporting force of the healthy side limb, maintain the patient's body balance, and avoid the center of gravity shifting due to the elevation of the affected side limb.

[0168] In some alternative implementations, the internal state perception module of the multi-source heterogeneous perception layer can employ alternative sensor combinations. In addition to the surface electromyography (EMG) sensors positioned on the rectus femoris and biceps femoris muscles of the thigh, surface EMG sensors can be added to the gastrocnemius and tibialis anterior muscles of the calf, collecting electrical signals from more muscles and enabling more comprehensive decoding of movement intentions. Multi-dimensional force sensors can be optionally installed on the foot structure instead of the calf link, directly collecting contact force data between the foot and the ground, reducing signal loss during force transmission and improving the accuracy of force data measurement. The external environment perception module can add ultrasonic sensors to the depth vision camera. These ultrasonic sensors are positioned on both sides of the robot's torso to assist in detecting side obstacles, complementing the frontal terrain detection of the depth vision camera, expanding the environmental perception range, and improving the accuracy of generating terrain disturbance prediction vectors.

[0169] In some alternative implementations, the lightweight motion prediction submodule can use a gradient boosting tree model instead of a recurrent neural network model. The gradient boosting tree model predicts future joint motion trends by integrating the prediction results of multiple decision trees. This model has fast training speed, low inference latency, and is more suitable for resource-constrained robot control systems. During training, historical joint angles, joint angular velocities, and electromyographic signal sequences are used as input features, and joint angle sequences within future time windows are used as output labels. The model parameters are optimized using a gradient descent algorithm to ensure that the prediction accuracy meets the requirements of motion intent decoding. In actual operation, the model receives real-time collected historical motion data, quickly outputs motion trend prediction results, and fuses them with the other two intent features to generate a comprehensive motion intent vector.

[0170] In some optional implementations, the central safety coordinator can use tiered triggering safety logic instead of a single interrupt triggering logic. Safety thresholds are divided into warning thresholds and emergency thresholds. When the monitored quantity reaches the warning threshold but does not exceed the emergency threshold, the central safety coordinator does not send an interrupt signal. Instead, it sends a warning signal to the bilateral differentiated collaborative planning layers, adjusting the control target set, reducing joint movement speed, or decreasing interaction torque to prevent risk escalation. When the monitored quantity exceeds the emergency threshold, the interrupt signal is then sent and the safety damping stop procedure is executed. This tiered triggering logic reduces unnecessary training interruptions, improves the continuity of rehabilitation training, and ensures patient safety. The warning and emergency thresholds are set based on clinical rehabilitation data and robot performance test results to ensure the rationality and safety of the tiered logic.

[0171] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multimodal feedback collaborative control system for a lower limb rehabilitation robot, characterized in that, The system consists of: A multi-source heterogeneous sensing layer is used to collect in real time the patient's internal lower limb motion state signals, human bioelectric signals, and terrain visual signals of the robot's external environment; The intent and environment fusion understanding layer, connected to the multi-source heterogeneous perception layer, is used to receive and fuse lower limb internal motion state signals and human bioelectric signals to generate a comprehensive patient motion intent vector, while processing the terrain visual signals to generate a terrain disturbance expectation vector. A bilateral differentiated collaborative planning layer, connected to the intent and environment fusion understanding layer, is used to receive the motion intent vector and the terrain disturbance expectation vector, and combine them with pre-stored patient bilateral limb muscle strength assessment data to dynamically synthesize adaptive gait trajectory targets. The gait trajectory targets are deconstructed into a first control target set for the healthy limb and a second control target set for the affected limb. The first control target set is mainly composed of expected cross-impedance parameters and expected end force field parameters, while the second control target set is mainly composed of expected joint trajectory, variable joint impedance parameters, and feedforward auxiliary torque curve. A distributed collaborative execution layer connects the bilateral differentiated collaborative planning layer, and the distributed collaborative execution layer includes a healthy side execution control unit and a diseased side execution control unit; The healthy side execution control unit is used to generate a first motor control command based on the first control target set and a force-position hybrid control algorithm based on the admittance model, so as to drive the robot's healthy side joint actuator to achieve compliant assisted motion. The affected side execution control unit is used to generate a second motor control command based on the second control target set using a composite control algorithm that integrates feedforward torque compensation and feedback impedance control, so as to drive the robot's affected side joint actuator to achieve active guidance and assisted movement.

2. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The multi-source heterogeneous sensing layer includes an internal state sensing module and an external environment sensing module: The internal state perception module includes a rotary encoder and a joint torque sensor installed at the robot's hip and knee joints for measuring joint angles, joint angular velocities, and human-machine interaction torques. The internal state perception module also includes surface electromyography (EMG) sensors arranged on the surface of the rectus femoris and biceps femoris muscle bellies of the patient's thigh for collecting EMG signals. The internal state perception module also includes a multidimensional force sensor installed on the robot's lower leg link or foot structure for measuring three-dimensional ground reaction forces and interaction forces. The external environment perception module includes a depth vision camera fixed to the front of the robot's torso. The optical axis of the depth vision camera is oriented diagonally downwards, and its field of view covers the area in front of the robot. The external environment perception module has a built-in terrain parsing algorithm for processing depth image data in real time, generating a terrain map containing height information and terrain semantic classification, and extracting key terrain feature parameters.

3. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The intent and environment fusion understanding layer includes a motion intent decoder and an environment terrain parser; The motion intent decoder includes a frequency domain gait analysis submodule, an instantaneous interaction force analysis submodule, and a lightweight motion prediction submodule. The frequency domain gait analysis submodule performs time-frequency transformation on the joint angle sequence and extracts the dominant gait frequency and symmetry index as the first intention feature; The instantaneous interactive force analysis submodule analyzes the force center trajectory and torque direction output by the multi-dimensional force sensor as a second intention feature; The lightweight motion prediction submodule uses a recurrent neural network model, taking historical joint motion sequences as input, to predict joint motion trends within future time windows, which serve as third intent features. The motion intention decoder also includes an intention fusion arbitrator, which takes the first intention feature, the second intention feature, the third intention feature, and the current gait phase signal as input, and outputs a comprehensive patient motion intention vector through a weighted fusion strategy. The mathematical expression of the weighted fusion strategy is as follows: in, This represents the overall motion intent vector. Indicates the first The intent feature vector output by each submodule These correspond to the frequency domain gait analysis submodule, the instantaneous interaction force analysis submodule, and the lightweight motion prediction submodule, respectively. This indicates the current gait phase, determined in real time by the joint angle sequence. Indicates phase with gait The relevant adaptive weight coefficients satisfy And the weight coefficient mapping relationship Determined through a predefined fuzzy logic rule table.

4. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The dual-sided differentiated collaborative planning layer includes a trajectory dynamic synthesis unit and a task deconstruction and allocation unit; The trajectory dynamic synthesis unit uses the standard gait trajectory as a base and scales and deforms the base trajectory in real time and spatial amplitude according to the motion intention vector to generate a personalized gait trajectory. The task deconstruction and allocation unit deconstructs the personalized gait trajectory into a healthy side support phase task and an affected side swing phase task based on the human biomechanics inverse dynamics model. For the healthy side support phase task, the task deconstruction and allocation unit calculates the desired end force field parameters and desired interaction impedance parameters used to maintain balance and provide propulsion, which constitute the first control target set; For the affected side swing phase task, the task deconstruction and allocation unit calculates the ideal joint moment curve in conjunction with the expected terrain disturbance vector, and scales the ideal joint moment curve according to the patient's bilateral limb muscle strength assessment data to generate the feedforward auxiliary moment curve. At the same time, the desired joint trajectory and variable joint impedance parameters are set to form the second control target set.

5. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The health side execution control unit is specifically used for: A force-potential hybrid control law based on the admittance model is established. The input of the control law is the deviation between the desired end force field and the measured interaction force of the multidimensional force sensor. The force deviation is converted into joint position adjustment using the admittance model. The joint position adjustment amount is superimposed with a loose joint reference position command to generate the first motor control command, which is then sent to the healthy joint motor driver.

6. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The affected-side execution control unit is specifically used for: Establish a feedforward-feedback composite impedance control law; The feedforward channel of the control law directly outputs the feedforward auxiliary torque curve as a torque command. The feedback channel of the control law takes the positional deviation between the desired joint trajectory and the measured trajectory of the joint encoder as input, and calculates the feedback compensation torque through a variable impedance model. The stiffness and damping parameters of the variable impedance model are adjusted online in real time according to the signal amplitude of the surface electromyography sensor. The torque command output from the feedforward channel is superimposed with the feedback compensation torque output from the feedback channel to generate the second motor control command, which is then sent to the motor driver of the affected joint.

7. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 3, characterized in that, The environmental terrain parser is specifically used for: The system receives data from the depth vision camera and generates a terrain elevation map and a semantic map in the robot's body coordinate system through a synchronous localization and mapping algorithm and a semantic segmentation network. Based on the current robot speed and the motion intention vector, predict the next expected landing point position; Based on the terrain elevation map, the terrain height change value and slope angle at the expected landing point are calculated. Combined with the robot and human dynamics model, the additional joint torque requirement required to perform the next action is calculated, and the output is the expected terrain disturbance vector.

8. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The system also includes a central security coordinator; The central safety coordinator is connected to the distributed collaborative execution layer and the multi-source heterogeneous perception layer, and monitors all sensor signals, bilateral joint motion status, and the first motor control command and the second motor control command in real time. The central safety coordinator is pre-set with joint motion angle limit thresholds, joint interaction torque safety thresholds, and bilateral motion coordination thresholds. When any monitored quantity exceeds the corresponding threshold, the central safety coordinator sends a highest-priority interrupt signal to the distributed collaborative execution layer, controlling all joint actuators to enter a preset safety damping stop mode.

9. The multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 1, characterized in that, The system also includes a data management and remote service layer; The data management and remote service layer is used to store the raw data collected by the multi-source heterogeneous perception layer, the historical records of the control target set generated by the dual-sided differentiated collaborative planning layer, and the execution logs of the distributed collaborative execution layer. The data management and remote service layer constructs a digital twin model of the patient based on historical data, and uses a continuous learning algorithm to optimize and update the neural network model parameters of the lightweight motion prediction submodule and the weight coefficient mapping relationship of the intent fusion arbitrator offline. The data management and remote service layer securely distributes the optimized model parameters to the intent and environment fusion understanding layer through an encrypted communication link, enabling remote iteration and personalized evolution of the control strategy.

10. A multimodal feedback lower limb rehabilitation robot collaborative control system according to claim 4, characterized in that, The methods for obtaining the patient's bilateral limb muscle strength assessment data include: During the initial system calibration phase, the control robot guides the patient's affected limb to perform a set of standardized isokinetic concentric and eccentric contraction movements. The maximum active output torque of the affected joint at various motion angles is measured using the joint torque sensor. The maximum active output torque of the healthy joint was measured in the same manner. Calculate the ratio of the maximum active output torque of the corresponding joint groups on the affected side to that on the healthy side, and store this ratio as the muscle strength proportionality coefficient K in the patient's bilateral limb muscle strength assessment data.

Images