Dynamic and personalized torque control method for exoskeleton based on multi-modal real-time perception

Abstract

The application provides a kind of exoskeleton dynamic and individualized torque control method based on multi-modal real-time perception, which regards terrain geometric parameters such as slope angle and step height as continuous variables, and realizes stepless continuous adjustment of baseline torque curve through biomechanics-verified scaling-deformation function.Torque output changes smoothly with terrain parameters, completely eliminating the step impact of mode switching and adapting to complex scenarios of any slope and any step height; performance indicators such as gait symmetry, step frequency stability and user effort level can be calculated only through IMU, encoder and motor current; Bayesian optimization is used to fine-tune control parameters online during user daily walking, quickly complete individualized adaptation, and can adapt to dynamic states such as user fatigue and weight change in real time.Torque execution adopts impedance control law dynamically adjusted with gait phase and terrain type, supports high stiffness to ensure stability, and is flexible in swing, further improving human-machine interaction comfort.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton robot technology, and in particular to an exoskeleton dynamic and personalized torque control method, electronic device, and computer-readable storage medium that dynamically senses terrain geometry parameters and user motion state by fusing information from the body sensor and environmental multimodal sensor in real time, and accordingly optimizes and adjusts the output torque of each joint of the exoskeleton online and adaptively. Background Technology

[0002] Lower limb exoskeleton robots can effectively reduce metabolic consumption during walking and improve motor skills by providing joint assistance to wearers. They have broad application prospects in fields such as rehabilitation medicine, industrial load-bearing, and elderly mobility assistance.

[0003] The core of high-performance exoskeleton control is applying the most precisely matched torque at the correct gait timing; its control precision directly determines the assist effect and wearing comfort. For example, Chinese invention patent CN113749956A, "An Adaptive Assist Control Method and System for Lower Limb Exoskeleton," represents a typical solution in the field of exoskeleton torque control: this technology pre-calibrates fixed torque-phase curves for typical terrains such as flat ground, uphill slopes, and stairs. After identifying the terrain type using a visual sensor, it switches to the corresponding preset torque mode. Personalized adaptation requires offline calibration of assist gain parameters in a laboratory environment, using a motion capture system and a metabolic analyzer to collect user motion data. However, this exoskeleton torque control solution has the following technical bottlenecks:

[0004] 1. Static control mode, coarse-grained terrain adaptation.

[0005] Existing solutions generally adopt a switch-type control logic of "terrain classification - mode switching", which can only adapt to a limited number of preset terrain types (such as 3° slope, 8° slope, 15cm step). In real-world scenarios, slope and step height are continuously changing, and the fixed torque curve cannot match the continuously changing terrain parameters, resulting in a large deviation between the assist torque and actual needs. Moreover, the torque will change abruptly when switching modes, causing human-computer interaction shock, reducing wearing comfort, and in severe cases, even interfering with the user's gait balance.

[0006] 2. Personalized calibration is cumbersome and cannot adapt to dynamic states.

[0007] Personalized parameter calibration relies on professional laboratory equipment (motion capture system, metabolic analyzer), and the calibration process takes 1-2 hours, is costly and has a high barrier to entry, making it impossible for ordinary users to complete independently. More importantly, the calibrated parameters are static values ​​and cannot adapt to dynamic states such as fatigue levels, muscle strength changes, and weight changes during the user's walking process, resulting in a significant decrease in the assistive effect after prolonged use.

[0008] 3. Lack of a refined dynamic adjustment mechanism

[0009] Even if some solutions achieve terrain recognition, their parameter adjustments are only made between gait cycles. They cannot achieve continuous and smooth torque modulation based on real-time phase and environmental parameters within a single gait cycle. The torque output is not well matched with human biomechanics, which limits the improvement of efficiency.

[0010] 4. Lacks closed-loop self-optimization capability

[0011] The control parameters of the existing system are fixed once calibrated, and cannot be automatically iterated and optimized based on the gait feedback of the user in actual use. The system performance will not improve with the length of use and it is difficult to adapt to changes in muscle strength and walking habits of users in long-term use.

[0012] Therefore, there is an urgent need in this field for an intelligent control scheme that can dynamically, precisely, and automatically optimize control torque based on real-time perceived environmental parameters and user status, and has continuous personalized learning capabilities. Summary of the Invention

[0013] To address the technical problems existing in the prior art, the present invention provides the following technical solution:

[0014] On the one hand, a method for controlling the dynamics and personalized torque of an exoskeleton based on real-time multimodal perception is provided, including the following steps:

[0015] S1. Multimodal real-time perception and state estimation: Fuse data from the body sensor and the environment sensor to calculate the continuously normalized gait phase, terrain type and geometric parameters, and calculate the user's implicit gait performance index.

[0016] S2. Parametric Dynamic Torque Generation: Based on continuous gait phase, real-time terrain geometry parameters are used as continuous variables. The standard baseline torque curve is continuously and steplessly adjusted through a preset scaling mapping function, and the initial target torque is generated by combining user-personalized parameters.

[0017] S3. Online Adaptive Parameter Optimization: Taking the user's implicit performance indicators as the optimization target, a lightweight Bayesian optimization algorithm is used to fine-tune the torque scaling coefficient and personalized parameters online during the user's walking process, continuously optimizing the assist effect;

[0018] S4, Bottom-level torque servo execution: Employs a dynamic impedance control law, combined with gait phase and terrain type, to adjust stiffness and damping parameters in real time, accurately outputting the target torque to the joint actuator.

[0019] As a preferred embodiment of this application, in step S1, the continuous gait phase is obtained by fusing IMU and joint encoder data through extended Kalman filtering, specifically including:

[0020] Constructing gait state vectors ,in Let be the normalized gait phase at time k, with a value range of . , The rate of change of phase;

[0021] Through the equation of state Perform state prediction, where Here is the state transition matrix. This is the process noise vector;

[0022] By observing the equation Observations are updated, among which The observation vector is composed of IMU acceleration and joint angle. For the observed noise vector;

[0023] The final output is a continuously changing gait phase.

[0024] As a preferred embodiment of this application, in step S1, the terrain geometric parameters include the slope angle. with step height The solution is obtained by fusing semantic segmentation from a monocular camera with multi-point laser ranging:

[0025] The confidence distribution of terrain types is output through a lightweight semantic segmentation network. When the confidence of the terrain type is greater than 0.8, the terrain geometric parameters are calculated by combining the laser point cloud data.

[0026] Slope angle through Calculation, where The height difference of the laser point cloud. This corresponds to the horizontal distance;

[0027] The step height is calculated by taking the average of the height differences between the laser points on the edges of adjacent steps.

[0028] As a preferred embodiment of this application, in step S1, the user implicit performance index includes the gait symmetry index. Step frequency stability index Interaction torque with users Comprehensive performance indicators The calculation formula is:

[0029] in , , These are the weighting coefficients;

[0030] Gait symmetry index , , The duration of the left and right leg swing phases are respectively;

[0031] Step frequency stability index , The standard deviation of the gait frequency over five consecutive gait cycles. This represents the average step frequency.

[0032] User interaction torque It is the difference between the actual output torque and the target torque, used to characterize the degree of force exerted by the user.

[0033] As a preferred embodiment of this application, in step S2, the terrain parameterization moment scaling includes two models: slope scaling and step height scaling.

[0034] For uphill and downhill scenarios, the formula for calculating the baseline moment is: ,in Standard flat gait phase The corresponding reference torque, This is the slope scaling factor. The slope angle is real-time, with positive values ​​for uphill slopes and negative values ​​for downhill slopes;

[0035] For scenarios involving going up and down stairs, first calculate the peak torque. Then, the oscillating phase torque curve is scaled according to the peak value, where This is the step height coefficient. For real-time step height, This represents the peak torque corresponding to the standard step.

[0036] As a preferred embodiment of this application, in step S2, the user-personalized parameters include a boost intensity factor. With weight correction factor The formula for calculating the initial target torque is:

[0037] Among them, weight correction factor , For the user's actual weight, Standard body weight; strength factor The initial default value is 1.0, and the value range is 0.5~1.5. It can be dynamically adjusted through online optimization.

[0038] As a preferred embodiment of this application, the lightweight Bayesian optimization algorithm in step S3 uses a sparse Gaussian process as a surrogate model, specifically including:

[0039] With slope scaling factor Step height coefficient Assistance strength factor Optimization variables Limit the reasonable range of biomechanical values ​​for each parameter;

[0040] Using a sparse Gaussian process to fit and optimize variables With comprehensive performance indicators The mapping relationship between them is determined, the number of induced points is limited to 20, and stochastic variational inference is used to update the model hyperparameters;

[0041] The next test parameter point is selected using the expected improvement acquisition function. Each parameter adjustment does not exceed 10% of the current value. The surrogate model is updated after collecting performance data for 5 gait cycles.

[0042] As a preferred embodiment of this application, the optimization process in step S3 adopts an event-triggered and sparse update mechanism:

[0043] An optimization iteration is triggered when the performance index fluctuates by less than 5% for three consecutive gait cycles and the terrain condition is stable.

[0044] When the terrain type changes or the slope / step height fluctuates beyond the threshold, the optimization process is paused and restarted after the state stabilizes.

[0045] When the performance improvement is less than 1% after three consecutive iterations, the optimizer enters a sleep state and triggers a state check every 50 gait cycles.

[0046] As a preferred embodiment of this application, the dynamic impedance control law calculation formula in step S4 is as follows:

[0047] in:

[0048] This is the torque command output to the motor. For the target torque generated by the upper layer, This is the position stiffness coefficient. The damping coefficient is... , These are the target and actual joint angles, respectively. , These are the target and actual joint angular velocities, respectively.

[0049] and Dynamically adjusts according to gait phase and terrain type: the support phase uses high stiffness to ensure stability, the swing phase uses low stiffness to ensure flexibility, and the damping is increased at the end of the swing to assist deceleration; the overall stiffness coefficient is increased by 20% in uphill / downhill / staircase scenarios.

[0050] As a preferred embodiment of this application, the optimized control parameters are classified and stored in the user profile parameter set according to terrain type. The historical optimal parameters of the corresponding terrain are directly loaded as the initial values ​​when used next time. The system continuously records the optimal parameters under different terrains to form a user-specific torque mapping library, and the adaptation accuracy is gradually improved with the number of uses.

[0051] On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement the method described above.

[0052] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the above method.

[0053] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0054] 1. Breaking away from the traditional "mode switching" control approach, this system treats terrain geometry parameters such as slope angle and step height as continuous variables. Through a biomechanically validated scaling-deformation function, it achieves stepless continuous adjustment of the baseline torque curve. The torque output changes smoothly with the terrain parameters, completely eliminating the abrupt impact of mode switching and adapting to complex scenarios with arbitrary slopes and step heights.

[0055] 2. An online self-optimization method based on implicit indicators of the body sensor is proposed. It does not require additional metabolic instruments, motion capture equipment, etc. It can calculate performance indicators such as gait symmetry, gait frequency stability, and user force exertion level using only IMU, encoder, and motor current. A lightweight Bayesian optimization algorithm based on sparse Gaussian process is adopted to fine-tune control parameters online during the user's daily walking. Personalized adaptation can be completed within 10 minutes, and it can adapt to dynamic states such as user fatigue and load changes in real time.

[0056] 3. By integrating monocular visual semantic segmentation, multi-point laser ranging, and body sensor data, 0-100% continuous gait phase estimation is achieved through extended Kalman filtering, with a phase error of less than 2%. At the same time, joint calculation of terrain type and geometric parameters is achieved, with a slope estimation error of less than 1° and a step height error of less than 0.5cm, providing high-precision and high-real-time perception input for dynamic torque control.

[0057] 4. The bottom-level torque execution adopts an impedance control law that dynamically adjusts according to the gait phase and terrain type. The high stiffness of the support phase ensures stability, the low stiffness of the swing phase ensures flexibility, and the high damping at the end of the swing phase assists in deceleration, achieving a balance between the smoothness and precision of torque output, and further improving the comfort of human-machine interaction. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is the control logic diagram for this method;

[0060] Figure 2 This is a schematic diagram of the dynamic torque generation mechanism of the present invention;

[0061] Figure 3 This is a block diagram of the torque servo execution logic of the present invention;

[0062] Figure 4 This is a schematic diagram of a lower limb assistive exoskeleton hardware system provided in an embodiment of the present invention. Detailed Implementation

[0063] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0064] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0065] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0066] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0067] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0068] like Figure 1 The control logic diagram of the present invention shown includes the following steps:

[0069] Step 1: Real-time multimodal sensing and multidimensional state estimation

[0070] This step serves as the sensing input for the entire system, responsible for fusing data from multiple sensors and outputting continuous gait phase, terrain parameters, and user status indicators, providing the data foundation for subsequent torque generation and optimization. It includes the following:

[0071] 1.1 Continuous gait phase estimation

[0072] First, based on the system sampling frequency (e.g., 100Hz, which is the standard configuration for exoskeleton systems, ensuring the time resolution of phase estimation without causing computational overload), the system needs to collect triaxial acceleration and triaxial angular velocity data from the lower limb IMU, and real-time joint angle data from the joint encoder, and then perform gait phase estimation by the system's main control platform.

[0073] Specifically, the Extended Kalman Filter (EKF) algorithm is used to model gait motion as a continuous state space, fusing observations from multiple sensors to output a 0-1 normalized continuous gait phase (corresponding to 0-100% of the gait cycle, including the continuous normalized gait phase). Phase change rate Compared to the discrete phase estimation of traditional state machines, the accuracy and smoothness are significantly improved.

[0074] The EKF algorithm operates as follows: (1) Define the gait state vector as: in:

[0075] : Normalized gait phase at time k, with a value range of 0 corresponds to the moment the heel strikes the ground, and 1 corresponds to the next heel strike, completing one complete gait cycle;

[0076] The rate of phase change at time k, in units of This reflects the pace of movement.

[0077] (2) State prediction equation (describing the evolution of gait phase): in:

[0078] Here is the state transition matrix. , The sampling time interval is 0.01s (0.01s in this scheme, corresponding to 100Hz sampling).

[0079] Let be the process noise vector, following a zero-mean Gaussian distribution, with a covariance matrix of... .

[0080] (3) Observation equation (establishing a mapping between state and sensor observations): in:

[0081] The observation vector contains two observations: IMU acceleration amplitude and knee joint angle.

[0082] This is a nonlinear observation function that maps the gait phase to the corresponding kinematic observations.

[0083] The observation noise vector follows a zero-mean Gaussian distribution, and the observation noise covariance is... .

[0084] The above values ​​for the covariance between process noise and observation noise are based on the actual accuracy calibration of the sensor. For example, the acceleration noise density of the BMI088 IMU is... The 16-bit encoder has an angle error of less than 0.005°, and the corresponding noise value matches the hardware precision.

[0085] (4) The EKF algorithm performs two iterations:

[0086] Prediction step: Estimate the prior state and prior covariance at the current time step based on the state at the previous time step;

[0087] Update step: Correct the prior state using the current sensor observations to obtain the posterior state estimate, and update the covariance matrix at the same time.

[0088] The final phase estimation error is less than 2%, which meets the phase accuracy requirements of torque control (when the phase error is greater than 5%, the timing deviation of the assist will significantly reduce the assist effect).

[0089] 1.2 Terrain Type Identification and Geometric Parameter Calculation

[0090] First, image data (30fps, 640×480 resolution) from a monocular camera and 8×8 dot matrix distance data (50Hz, laser ranging accuracy ±1cm, corresponding to a slope angle estimation error of less than 1°, a step height estimation error of less than 0.5cm, and a torque scaling error of less than 2% at this accuracy, which will not affect the assist effect) are collected and input. A two-level fusion scheme of "visual semantic classification + laser geometric calculation" is adopted. The vision is responsible for predicting the terrain type at a distance, while the laser is responsible for the accurate calculation of terrain parameters at close range. The fusion of the two achieves high-precision, forward-looking perception of terrain information. The implementation process is as follows:

[0091] (1) Visual semantic classification: A lightweight MobileNetV3+UNet semantic segmentation network is used to classify camera images at the pixel level and output the confidence distribution of five types of terrain: flat ground, uphill, downhill, upstairs stairs, and downstairs stairs. The network runs accelerated on the RK3588 NPU, with a single frame inference time of less than 10ms. It can achieve terrain prediction with a look-ahead of 2-3 meters (corresponding to a prediction time of 0.8-1.2 seconds when walking on flat ground, which is enough for the system to adjust the torque parameters in advance to achieve smooth terrain transition).

[0092] (2) Solving laser geometric parameters: Perform plane fitting and edge detection on the laser point cloud data, and calculate the terrain geometric parameters:

[0093] Slope angle calculation: Perform plane fitting on the laser point on the ground to obtain the plane normal vector, and calculate the angle between the normal vector and the vertical direction to obtain the slope angle. The formula is: in The height difference of the laser point cloud. This corresponds to the horizontal distance. When going uphill. Take positive values, and take negative values ​​when going downhill.

[0094] Step height calculation: The step height is obtained by identifying the edge points of the steps using an edge detection algorithm, calculating the height difference between adjacent step edges, and taking the average of multiple sets of data. The formula is:

[0095] ,in:

[0096] Let be the height value of the edge of the i-th step. The number of steps detected.

[0097] (3) Multi-source fusion decision: When the confidence level of visual terrain classification is greater than the system preset value (e.g., 0.8), the terrain type is adopted. And combined with laser parameters (terrain parameter vector) Output; when the confidence level is insufficient, it is downgraded to pure laser parameter estimation to ensure system robustness.

[0098] 1.3 Calculation of User Implicit State Indicators

[0099] The system collects and inputs data from the user, including the angles and angular velocities of the left and right leg joints and the motor current, and directly calculates implicit indicators of the user's state. Without the need for additional sensors, it calculates these implicit indicators reflecting the assistive effect and the user's state solely through data from the system's own sensors, serving as the objective function for online optimization.

[0100] This scheme selects three core indicators—gait symmetry, gait frequency stability, and user interaction torque—to construct a comprehensive performance evaluation function.

[0101] (1) Gait symmetry index

[0102] This reflects the balance of gait between the left and right legs. Higher symmetry indicates more balanced assisted movement and greater comfort for the user. The formula is: in The duration of the left leg swing phase. The duration of the right leg swing phase is calculated from the ground contact event detected by gait phase detection. Range of values The closer it is to 1, the better the symmetry.

[0103] (2) Step frequency stability index

[0104] This reflects the degree of fluctuation in cadence. Higher stability indicates a smoother gait and a better assisted movement. The formula is:

[0105] in The standard deviation of gait frequency over 5 consecutive gait cycles (the statistical window is 5 gait cycles, which ensures the statistical significance of the indicator calculation and reflects changes in user status in a timely manner). This represents the average of the corresponding step frequencies. Range of values The closer it is to 1, the better the stability.

[0106] (3) Estimation of user interaction torque

[0107] This indirectly reflects the user's own exertion level; the less force the user exerts, the higher the proportion of exoskeleton assistance and the better the assistance effect. It is estimated by the difference between the actual output torque of the motor and the target torque.

[0108] in The actual output torque is calculated by inversely using the q-axis current of the motor. The target torque given to the system. The smaller the absolute value, the less effort the user needs to exert.

[0109] (4) Comprehensive performance indicators The three indicators are weighted and summed to form the objective function of the optimization algorithm: The weighting is based on human gait biomechanics research: gait symmetry is a core factor affecting wearing comfort and long-term safety, therefore it has the highest weight; gait frequency stability reflects gait smoothness, and user force application directly reflects assist efficiency, so both have roughly equal weight. Optimal weighting coefficients are selected. , , This is because the weighting ratio has been verified in multiple exoskeleton human-computer interaction experiments and conforms to the correspondence between user subjective feelings and objective performance.

[0110] Final output comprehensive performance indicators and gait symmetry Step frequency stability User interaction torque Three sub-indicators.

[0111] Step 2: Parametric Dynamic Torque Generation

[0112] This step is the core of the system's decision-making process. It receives gait phase, terrain parameters, and user-personalized parameters from the perception output, and generates dynamic target torque through a parameterized mapping function, thus achieving a leap from "fixed mode switching" to "continuous parameter adjustment".

[0113] like Figure 2 As shown, the torque construction process includes the following steps:

[0114] 1. Construction of Standard Baseline Torque Library

[0115] The system pre-stores the baseline moment curve of the standard flat gait. The gait phase was discretized into 100 sampling points, covering a complete gait cycle. This curve was calibrated based on biomechanical measurement data of normal adults walking on flat ground, conforming to the laws of human muscle exertion (calibrated using a typical adult weight of 70kg as an example, as shown below):

[0116] Knee joint: The stance phase provides knee extension assistance with a peak torque of 20 N·m; the swing phase provides knee flexion assistance with a peak torque of 12 N·m.

[0117] Ankle joint: Provides plantar flexion assistance during push-off, with a peak torque of 15 N·m.

[0118] The standard baseline torque serves as the benchmark for all terrain torque scaling, ensuring that the base assist conforms to human biomechanical characteristics.

[0119] 2. Terrain parameterization torque scaling

[0120] According to terrain type and parameters The standard baseline moment is continuously scaled steplessly, with different scaling models corresponding to different terrains.

[0121] (1) Slope scaling model (uphill and downhill scenarios)

[0122] When going uphill or downhill, the torque of human joints is approximately linearly related to the slope angle. Therefore, a linear scaling model is used to adjust the baseline torque as a whole. in:

[0123] Baseline moment after terrain adaptation, in units of ;

[0124] Phase under standard flat gait The corresponding reference torque, in units of ;

[0125] : Slope scaling factor, initial default value, e.g., uphill ,downhill ;

[0126] The real-time estimated slope angle, in units of Uphill values ​​are taken as positive, and downhill values ​​are taken as negative.

[0127] Operating mechanism: For every 1° change in slope, the baseline moment scales linearly. When going uphill... The torque increases proportionally, assisting the user in overcoming gravity; when going downhill... The torque is reduced proportionally to avoid excessive assistance interfering with the user's center of gravity control and to ensure stability when walking downhill.

[0128] Value Explanation: This value is based on biomechanical experimental data on human gait: the average knee extensor torque is approximately 15 N·m when walking on flat ground, and approximately 22.5 N·m when going uphill at a 5° angle, an increase of about 50%. The results are highly consistent with the measured data. When going downhill, the human body mainly controls speed through eccentric muscle contraction, and the need for assistance is lower than when going uphill. Therefore, the scaling factor for downhill is 0.08, which is slightly smaller than that for uphill, and is consistent with the human body's force exertion pattern.

[0129] (2) Stair height scaling model (going up and down stairs scene)

[0130] When going up or down stairs, the peak torque during the leg-lifting phase is linearly related to the step height. First, adjust the peak torque according to the step height, then scale the swing phase torque curve proportionally to ensure the curve shape conforms to biomechanical characteristics. First, calculate the target peak torque: Then, the oscillating phase torque curve is scaled as a whole: in:

[0131] Target peak assist torque during the leg lift and swing phase, in units of ;

[0132] The benchmark peak torque corresponding to a standard step (15cm) is expressed in units of... The default value is 25 N·m.

[0133] Step height coefficient, unit: Initial default value ;

[0134] Real-time detected step height, in units of .

[0135] Operating Mechanism: The system monitors the height of each step in real time and dynamically adjusts the peak leg-lifting torque; the higher the step, the greater the leg-lifting assistance. The supporting phase torque maintains a baseline value to ensure support stability. Compared to torque curves with fixed step heights, this model can adapt to steps of any height, significantly improving the accuracy of leg-lifting assistance.

[0136] Value Explanation: Biomechanical studies have shown that for every 1cm increase in step height, the peak torque of the quadriceps muscle increases by 1.1-1.3 N·m when climbing stairs. This scheme uses... It falls within the reasonable middle range, covering the common step height range of 12-18cm in civil buildings, making it the most adaptable.

[0137] 3. Personalized parameter mapping

[0138] Introducing the User Profile parameter set, the baseline torque is scaled a second time to adapt to individual user differences:

[0139] in:

[0140] Initial target torque value after personalized adaptation (target torque before online optimization and correction), unit: ;

[0141] : Assist strength factor, value range The initial default value is 1.0, which reflects the user's preference for the strength of the assist. It can be manually adjusted by the user or dynamically fine-tuned by the online optimizer.

[0142] Weight correction factor ,in For the user's actual weight, The ideal weight is 70kg.

[0143] When a user enters their weight upon first use, the system automatically calculates a weight correction factor to complete basic personalized calibration, ensuring a consistent assistive experience for users of different weights. The online optimizer can then dynamically adjust based on the user's status. It adapts to dynamic states such as user fatigue and changes in load.

[0144] Value explanation: The lower limb joint torque has an approximately linear relationship with human body weight. The greater the weight, the higher the joint torque required for walking. Therefore, the linear weight correction coefficient is in line with the principles of ergonomics. The value range of the assist intensity factor is 0.5-1.5, covering the demand range from light assistance to full assistance, meeting the usage needs of different users and different scenarios.

[0145] Step 3: Lightweight Bayesian Online Adaptive Optimization

[0146] This step is the intelligent core of the system. Using implicit performance metrics as the optimization target, it employs a lightweight Bayesian optimization algorithm to fine-tune control parameters online during the user's daily walking, continuously improving the assist effect without any additional calibration equipment. It includes the following:

[0147] 1. Optimization variables and objective function

[0148] (1) Optimization variables: Select the three parameters that have the most significant impact on the assist effect to form an optimization vector: These are the slope scaling factor, step height factor, and traction strength factor, respectively. At the same time, reasonable biomechanical value ranges are defined for each parameter, for example:

[0149] To avoid scaling ratios deviating from the natural force exertion patterns of the human body;

[0150] This covers a reasonable range of steps to assist with movement.

[0151] This ensures that the assist intensity is within an acceptable range for the user.

[0152] (2) Optimization objective: Maximize the overall performance index That is, finding the optimal parameters , so that: in For the feasible region of the parameters.

[0153] 2. Sparse Gaussian Process Surrogate Model

[0154] The core of Bayesian optimization is to fit the mapping relationship between parameters and performance using a surrogate model, avoiding a large amount of real sampling. To adapt to the computing power of embedded platforms, this scheme uses a Sparse Gaussian Process (SGP) as the surrogate model, significantly reducing the computational load by limiting the number of induced points. The Gaussian process assumes that the performance index follows a multivariate normal distribution, with the prior distribution as follows: in:

[0155] The mean function is used, and this scheme takes a constant mean of 0.6;

[0156] For the kernel function, a squared exponential kernel is used to describe the correlation between parameter points: in This represents the signal variance, with a default value of 0.04. This is the feature length scale, with a default value of 0.1.

[0157] Sparse Gaussian processes are achieved by selecting This scheme uses induced points to approximate the complete Gaussian process. At each induction point, stochastic variational inference (SVI) is used to update the model's hyperparameters. Compared to the standard Gaussian process, the computational complexity is reduced from... Down to This significantly improves computing speed.

[0158] Value explanation: In the 3D parameter space, 20 induced points are sufficient to fit a smooth performance surface, and the model fitting error is less than 1%, which fully meets the optimization accuracy requirements;

[0159] The squared exponential kernel is suitable for fitting smooth biomechanical performance surfaces and is the most commonly used kernel function in Bayesian optimization.

[0160] The model is updated every 10 gait cycles, with a single model update time of less than 10ms. It can run easily on the RK3588 platform without affecting the real-time performance of the control.

[0161] 3. Data Acquisition Function and Parameter Update Strategy

[0162] By employing Expected Improvement (EI) as the acquisition function, the exploration of the parameter space (exploring unknown regions) and utilization (optimizing known optimal regions) are balanced to efficiently find the optimal parameters. in:

[0163] ;

[0164] , The parameters predicted by the surrogate model are respectively The corresponding performance mean and standard deviation;

[0165] This represents the optimal performance value among the currently sampled points.

[0166] The cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

[0167] Operating mechanism: Find the maximum value of the EI function within the feasible region of the parameters, and select the point with the largest EI as the next test parameter point;

[0168] Apply parameter perturbation, update the control parameters to the test values, run continuously for 5 gait cycles, and collect the average performance index under these parameters;

[0169] Add new "parameter-performance" sample points to the sample set and update the sparse Gaussian process proxy model;

[0170] Repeat the above process to gradually approach the optimal parameters.

[0171] Parameter update constraint: The relative magnitude of each parameter adjustment shall not exceed 10% of the current value, i.e. This avoids sudden parameter changes that could cause large fluctuations in torque, ensuring a continuous user experience.

[0172] 4. Event Triggering and Sparse Update Mechanism

[0173] To reduce computational cost and avoid terrain changes interfering with optimization results, this solution adopts an event-triggered sparse update mechanism instead of continuous optimization:

[0174] Triggering conditions: When the performance index fluctuation is less than 5% for three consecutive gait cycles, and the terrain type and parameters are stable (slope fluctuation <1°, step height fluctuation <1cm), an optimization iteration is triggered.

[0175] Pause conditions: When the terrain type changes, or when the slope / step height fluctuates beyond the threshold, the optimization process is immediately paused and restarted after the terrain condition stabilizes, in order to avoid performance fluctuations caused by terrain changes interfering with the optimization direction;

[0176] Convergence-Sleep: When the performance improvement is less than 1% after 3 consecutive optimization iterations, the parameters are considered to have converged, and the optimizer enters a sleep state. A state check is triggered every 50 gait cycles to confirm whether re-optimization is needed.

[0177] 5. Persistence of personalized parameters

[0178] The updated optimal control parameters are output. And a persistent set of user-personalized parameters, which is stored in the UserProfile parameter set after each optimization and categorized by terrain type.

[0179] The next time a user uses the service, the historical best parameters for the corresponding terrain are directly loaded as the initial values, greatly shortening the adaptation time. As the number of uses increases, the user-specific parameter library is continuously improved, helping to continuously improve accuracy.

[0180] Step 4: Low-level dynamic torque servo execution

[0181] like Figure 3 As shown, this step is the execution end of the system. It receives the target torque generated by the upper layer and uses a dynamic impedance control law to achieve precise and smooth torque output, ensuring torque tracking accuracy and human-machine interaction comfort.

[0182] 1. Dynamic impedance control law

[0183] Employing a position-velocity dual-feedback impedance control architecture, it provides smooth human-machine interaction while ensuring torque tracking accuracy. in:

[0184] The final torque command output to the motor, in units of... ;

[0185] The target torque generated by the upper layer, in units of ;

[0186] Position stiffness coefficient, unit: , with gait phase With terrain type Dynamic adjustment;

[0187] Damping coefficient, unit: , with gait phase With terrain type Dynamic adjustment;

[0188] , These represent the target joint angle and the actual joint angle, respectively, in units of... ;

[0189] , : These are the target joint angular velocity and the actual joint angular velocity, respectively, in units of . .

[0190] 2. Rules for dynamic adjustment of control parameters

[0191] Stiffness and damping parameters are dynamically matched based on gait phase and terrain type to simulate the natural impedance characteristics of human muscles and tendons.

[0192] Support phase ( ): Take 80 This ensures joint stiffness during the support phase and maintains body stability. Take 5 It buffers the impact force from the ground.

[0193] Oscillating phase ( ): Dropped to 40 This reduces joint resistance, ensures the flexibility of leg swing, and minimizes interference with the natural gait. Take 3 .

[0194] End of swing ( ): Increase to 10 It simulates the deceleration effect of the hamstring muscles, smoothly terminating the lower leg swing and preparing for heel strike.

[0195] Complex terrain (slopes, stairs): Overall improvement of 20%, enhancing position tracking accuracy and adapting to torque requirements in complex terrain; Simultaneously increase by 15% to enhance anti-disturbance capabilities.

[0196] Value Explanation: The above parameter values ​​are based on the measured impedance characteristics of human lower limb muscles and tendons: high muscle activation and high joint stiffness during the support phase; and low joint stiffness during the swing phase. This parameter configuration has been verified through human-computer interaction experiments. While ensuring assistive accuracy, it can reduce the impact force of human-computer interaction by more than 30%, significantly improving wearing comfort.

[0197] 3. Current loop torque servo

[0198] The underlying layer employs a field-oriented control (FOC) algorithm, which precisely controls the output torque by controlling the q-axis current of the motor. The current loop control frequency is 1kHz, and the torque response delay is less than 5ms, ensuring fast and accurate tracking of the target torque. Current feedforward compensation is also incorporated to further enhance dynamic response performance.

[0199] The application process of this method is described in detail below with reference to specific examples.

[0200] Example 1

[0201] like Figure 4 As shown, this embodiment is based on a prototype of a lower limb assistive exoskeleton, and the core hardware configuration is as follows:

[0202] Main control platform: The RK3588 embedded computing platform is used, which integrates a 4-core Cortex-A76 + 4-core Cortex-A55 CPU with a main frequency of 2.2GHz and a built-in 6TOPS computing power NPU, which is responsible for the operation of algorithms such as visual perception, Bayesian optimization, and torque control; the underlying torque servo is implemented by an STM32F407 microcontroller, which runs FOC and current loop control.

[0203] Sensor system:

[0204] Body sensors: BMI088 six-axis IMU (100Hz sampling) and 16-bit magnetoelectric encoder (1kHz sampling), respectively installed at the thigh, calf and joint axis;

[0205] Environmental sensors: OV5640 monocular camera (30fps, 640×480 resolution), VL53L5CX8×8 dot matrix laser rangefinder (50Hz sampling), mounted on the front of the waist of the exoskeleton;

[0206] Torque feedback: The joint has a built-in strain gauge torque sensor with an accuracy of ±0.5 N·m and a sampling frequency of 1 kHz.

[0207] Actuation system: permanent magnet synchronous motor + harmonic reducer, rated output torque 30 N·m, peak torque 50 N·m, meeting the assistance needs of flat ground and complex terrain.

[0208] Power system: 24V / 10Ah lithium-ion battery pack, with the adaptive DC-DC energy management system that is compatible with this solution.

[0209] I. Experimental Design

[0210] Experimental subjects: healthy adult males, weighing 70kg and 175cm tall, with no lower limb motor disorders and no prior experience wearing exoskeletons.

[0211] Experimental scenario: a continuous terrain consisting of a 5m flat ground → a 10m long 8° gentle slope → 12 steps of stairs with a height of 15cm, simulating terrain changes during daily walking.

[0212] Comparison group: The fixed mode switching scheme of the comparative file CN113749956A in the background technology was adopted. Three fixed torque curves were preset: flat ground, 8° uphill slope and 15cm stairs. The corresponding mode was switched after the terrain was visually identified (the experiment was conducted based on the same set of exoskeletons).

[0213] Evaluation metrics: terrain transition moment abrupt change rate, gait symmetry index, user metabolic consumption, and personalized adaptation time.

[0214] II. Implementation Process

[0215] Phase 1: Adaptive optimization for walking on flat ground (0-5m flat ground)

[0216] Perception layer output: The camera and laser fusion identifies the terrain as flat land, with slope... EKF continuously outputs gait phase with an average estimation error of 1.8%; initial gait symmetry index. Step frequency stability User interaction torque Overall performance .

[0217] Torque generation layer: initial parameters , The baseline torque is equal to the standard flat ground torque, and the peak assist torque of the knee joint is 20 N·m.

[0218] Online optimization process: After walking approximately 20 steps (15 seconds), the terrain and gait stabilize, triggering the optimizer to start. The optimizer then performs tests sequentially. and Two sets of parameter perturbations were used, and performance data for each set was collected over five gait cycles.

[0219] test Average performance ;

[0220] test Average performance ;

[0221] Prediction after fitting the surrogate model The optimal performance is achieved when the slope scaling factor is updated to 0.095.

[0222] Optimization results: Overall performance after the update Compared to the initial value, it improved by 1.26%, the gait symmetry improved to 0.93, and the user interaction torque decreased to 3.9 N·m.

[0223] Execution layer: Support phase , oscillating phase , oscillation end The torque tracking error is less than 2%.

[0224] Phase 2: Smooth transition adjustment of gentle slope (starting point of 5-7m ramp)

[0225] Perception layer output: The camera identifies uphill terrain 2m ahead, and the laser rangefinder calculates the slope angle in real time, continuously increasing from 0° to 8°. The entire transition process lasts about 1.2 seconds (2 gait cycles).

[0226] Moment generation layer: The baseline moment increases continuously and linearly with the slope angle, smoothly transitioning from 1.0 times the standard moment to... The peak torque at the knee joint increased continuously from 20 N·m to 35.2 N·m, with no step jumps throughout the entire process.

[0227] Performance of the comparison group: The comparison scheme requires waiting for the terrain recognition confidence level to reach the standard before switching modes. The torque jumps from 20 N·m to 36 N·m within one control cycle, with a torque mutation rate of 80%, and users can clearly feel the assist impact.

[0228] Optimization layer status: The terrain is in a transition phase, and the slope fluctuation is greater than the threshold. The optimizer will automatically pause and restart after the slope stabilizes.

[0229] Phase 3: Continuous uphill fatigue adaptation (7-15m stable 8° slope)

[0230] Initial state: After entering a stable uphill phase, the slope stabilizes at 8°±0.5°, and the optimizer restarts. User performance is good in the initial phase, and overall performance is good. .

[0231] Changes in fatigue status: After walking uphill for 3 minutes, the user experienced mild fatigue, and gait symmetry was observed. The step frequency stability decreased to 0.87. The user interaction torque dropped to 0.89. Increased to 5.8 N·m, overall performance The value dropped to 0.571, indicating a decrease in the effectiveness of the boost.

[0232] Online optimization process: The optimizer detects a performance degradation and triggers multi-parameter joint optimization. and Synchronous optimization was performed. After 6 iterations (approximately 60 gait cycles, 2 minutes), the surrogate model converged, and the optimal parameters were updated to... , .

[0233] Optimized result: Gait symmetry It rebounded to 0.91, and the step frequency stability... Rebounded to 0.92, user interaction torque Reduced to 4.5 N·m, overall performance Compared to before the optimization, it improved by 8.2%, effectively offsetting the decline in assistive effect caused by user fatigue.

[0234] Performance of the comparison group: The comparison scheme used fixed uphill parameters and had no adaptive adjustment capability. After the user became fatigued, the gait symmetry continued to drop to 0.83, the user's force continued to increase, and the assist effect continued to decline.

[0235] Phase 4: Dynamic adaptation for climbing stairs (15cm step after 15m)

[0236] Perception layer output: Laser ranging identifies the edge of the step and calculates the step height. The visual classification was "going up stairs," with a confidence level of 0.92.

[0237] Torque generation layer: peak torque The swing phase torque is scaled proportionally as a whole, and the leg lift assistance for each step is finely adjusted according to the real-time step height.

[0238] Optimization process: After the staircase scene stabilizes, the optimizer adjusts the step height coefficient. Optimize, initial After 4 iterations, the optimal value is Overall performance improved by 3.7%.

[0239] III. Data Analysis (See Table 1 below)

[0240] Terrain transition moment abrupt change rate 4.2% 32.7% Reduced by 87.2% Mutation rate = torque jump value / target torque value; the comparison scheme is the torque step at the moment of mode switching. gait symmetry index on flat ground 0.94 0.87 Increased by 8.0% Symmetry of the swing phase duration of the left and right legs; the closer the value is to 1, the better the symmetry. Gait symmetry after walking uphill for 3 minutes 0.91 0.83 An increase of 9.6% Symmetry maintenance capability under simulated user fatigue state Relative metabolic consumption of users 0.78 0.95 Decrease of 17.9% Using the metabolic energy expenditure of unassisted walking as a baseline of 1.0, it was indirectly estimated through user interaction torque. Personalized adaptation time 8min 120 min (offline calibration) Shortened by 93.3% The time required to achieve optimal assist effect needs to be calibrated using a laboratory motion capture system for comparison. Slope adaptability range 0-15° continuous stepless 3° / 8° / 12° three fixed positions Full range continuous coverage This solution can adapt to any slope and requires no preset gears.

[0241] Table 1

[0242] IV. Summary of Technical Advantages

[0243] Compared with the comparative solutions in the background art, the core advantages of the present invention can be summarized in four points:

[0244] Significantly improved control smoothness: Through a parameterized continuous scaling mechanism, the torque step impact caused by mode switching is completely eliminated, resulting in natural and smooth terrain transitions and significantly improved wearing comfort and gait safety.

[0245] Enhanced terrain adaptability: Breaking through the limitations of preset terrain settings, it can adapt to continuous terrain with any slope and step height, resulting in a higher match between the assist torque and actual needs, and a significant improvement in assist efficiency.

[0246] Personalized adaptation costs are extremely low: No professional laboratory equipment or offline calibration process is required. The system automatically completes personalized optimization during the user's daily life, reducing the adaptation time from 2 hours to 8 minutes, which greatly reduces the threshold for use.

[0247] Continuous self-optimization capability: A closed loop of "performance awareness - parameter optimization - effect improvement" has been built. The system can continuously iterate and optimize as users use it, automatically adapting to dynamic states such as user fatigue and changes in load, and the long-term user experience is constantly improved.

[0248] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0249] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0250] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0251] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0252] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0253] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0254] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0255] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0256] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0257] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamic and personalized torque control of an exoskeleton based on multimodal real-time perception, characterized in that, Includes the following steps: S1. Multimodal real-time perception and state estimation: Fuse data from the body sensor and the environment sensor to calculate the continuously normalized gait phase, terrain type and geometric parameters, and calculate the user's implicit gait performance index. S2. Parametric Dynamic Torque Generation: Based on continuous gait phase, real-time terrain geometry parameters are used as continuous variables. The standard baseline torque curve is continuously and steplessly adjusted through a preset scaling mapping function, and the initial target torque is generated by combining user-personalized parameters. S3. Online Adaptive Parameter Optimization: Taking the user's implicit performance indicators as the optimization target, a lightweight Bayesian optimization algorithm is used to fine-tune the torque scaling coefficient and personalized parameters online during the user's walking process, continuously optimizing the assist effect; S4, Bottom-level torque servo execution: Adopting a dynamic impedance control law, combined with gait phase and terrain type, the stiffness and damping parameters are adjusted in real time to accurately output the target torque to the joint actuator.

2. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, In step S1, the continuous gait phase is obtained by fusing IMU and joint encoder data through extended Kalman filtering, specifically including: Constructing gait state vectors ,in Let be the normalized gait phase at time k, with a value range of . , The rate of change of phase; Through the equation of state Perform state prediction, where Here is the state transition matrix. This is the process noise vector; By observing the equation Observations are updated, among which The observation vector is composed of IMU acceleration and joint angle. For the observed noise vector; The final output is a continuously changing gait phase.

3. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, In step S1, the terrain geometric parameters include the slope angle. with step height The solution is obtained by fusing semantic segmentation from a monocular camera with multi-point laser ranging: The confidence distribution of terrain types is output through a lightweight semantic segmentation network. When the confidence of a terrain type is greater than 0.8, the terrain geometric parameters are calculated by combining the laser point cloud data. Slope angle through Calculation, where The height difference of the laser point cloud. This corresponds to the horizontal distance; The step height is calculated by taking the average of the height differences between the laser points on the edges of adjacent steps.

4. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, In step S1, the user's implicit performance metrics include the gait symmetry index. Step frequency stability index Interaction torque with users Comprehensive performance indicators The calculation formula is: ; in , , These are the weighting coefficients; Gait symmetry index , , The duration of the left and right leg swing phases are respectively; Step frequency stability index , The standard deviation of the gait frequency over five consecutive gait cycles. This represents the average step frequency. User interaction torque It is the difference between the actual output torque and the target torque, used to characterize the degree of force exerted by the user.

5. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, In step S2, the terrain parameterization moment scaling includes two models: slope scaling and step height scaling. For uphill and downhill scenarios, the formula for calculating the baseline moment is: ,in Standard flat gait phase The corresponding reference torque, This is the slope scaling factor. The slope angle is real-time, with positive values ​​for uphill slopes and negative values ​​for downhill slopes; For scenarios involving going up and down stairs, first calculate the peak torque. Then, the oscillating phase torque curve is scaled according to the peak value, where This is the step height coefficient. For real-time step height, This represents the peak torque corresponding to the standard step.

6. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 5, characterized in that, In step S2, the user-personalized parameters include the assist intensity factor. With weight correction factor The formula for calculating the initial target torque is: ; Among them, weight correction factor , For the user's actual weight, Standard body weight; strength factor The initial default value is 1.0, and the value range is 0.5~1.

5. It can be dynamically adjusted through online optimization.

7. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, The lightweight Bayesian optimization algorithm in step S3 uses a sparse Gaussian process as a surrogate model, specifically including: With slope scaling factor Step height coefficient Assistance strength factor Optimization variables Limit the reasonable range of biomechanical values ​​for each parameter; Using a sparse Gaussian process to fit and optimize variables With comprehensive performance indicators The mapping relationship between them is determined, the number of induced points is limited to 20, and stochastic variational inference is used to update the model hyperparameters; The next test parameter point is selected using the expected improvement acquisition function. Each parameter adjustment does not exceed 10% of the current value. The surrogate model is updated after collecting performance data for 5 gait cycles.

8. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 7, characterized in that, The optimization process in step S3 adopts an event-triggered and sparse update mechanism: An optimization iteration is triggered when the performance index fluctuates by less than 5% for three consecutive gait cycles and the terrain condition is stable. When the terrain type changes or the slope / step height fluctuates beyond the threshold, the optimization process is paused and restarted after the situation stabilizes. When the performance improvement is less than 1% after three consecutive iterations, the optimizer enters a sleep state and triggers a state check every 50 gait cycles.

9. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, The formula for calculating the dynamic impedance control law in step S4 is as follows: ; in: This is the torque command output to the motor. For the target torque generated by the upper layer, This is the position stiffness coefficient. The damping coefficient is... , These are the target and actual joint angles, respectively. , These are the target and actual joint angular velocities, respectively. and Dynamically adjusts according to gait phase and terrain type: the support phase uses high stiffness to ensure stability, the swing phase uses low stiffness to ensure flexibility, and the damping is increased at the end of the swing to assist deceleration; the overall stiffness coefficient is increased by 20% in uphill / downhill / staircase scenarios.

10. The method for dynamic and personalized torque control of exoskeleton based on multimodal real-time perception according to claim 1, characterized in that, The optimized control parameters are categorized by terrain type and stored in the user profile parameter set. The corresponding historical optimal parameters for the next use are directly loaded as the initial values. The system continuously records the optimal parameters under different terrains to form a user-specific torque mapping library, and the adaptation accuracy is gradually improved with the number of uses.

Images