Variable stiffness and variable damping based wearable exoskeleton robot joint control method

Abstract

The application provides a wearable exoskeleton robot joint control method based on variable stiffness and variable damping, fuses a multi-modal perception output continuous gait phase and a terrain parameter, constructs a complete closed loop of "perception recognition-parameter mapping-bottom layer execution-online optimization-energy coordination", dynamically reproduces the stage variable impedance biomechanical characteristics of human muscle-tendon units, synchronously realizes global gait phase multi-window energy recovery, wide load adaptive DC-DC energy management, embedded lightweight perception model collaborative reasoning, and comprehensively solves the core pain points of the prior art, such as bulky structure, static discrete control, single perception, low energy efficiency and large human-computer interaction impact.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton robot control technology, and in particular to a wearable exoskeleton robot joint control method, electronic device, and computer-readable storage medium based on variable stiffness and variable damping. Background Technology

[0002] Wearable exoskeleton robots are becoming increasingly popular in the market, and their performance affects the user's comfort experience. The joint control method is the core control aspect of exoskeleton robots. Existing wearable exoskeleton robots employ multi-dimensional, single-variable control modes for joint control. For example, US20230025654A1 discloses an intelligent knee joint for human lower limb exoskeletons, prostheses, and orthotics. This intelligent knee joint reproduces part or all of the biomechanics of the human knee joint by using a motor drive unit based on a magnetorheological damper and a controllable elastic energy storage unit. Here, the motor drive unit can be replaced by a controllable damping unit. The core of this lower limb exoskeleton intelligent knee joint mechanism consists of an independent motor drive unit and an independent elastic energy storage unit assembled in series. Damping adjustment relies solely on a single set of magnetorheological dampers to achieve segmented mode switching control.

[0003] Although the aforementioned patent features a dual-mode motor drive unit capable of both driving and generating power, responsible for outputting assist torque or recovering negative energy, and includes an elastic energy storage unit comprising a fixed-stiffness spring assembly and a single MR damper, switching the working state of the elastic element (locked, free dissipation, energy storage) via a mechanical locking mechanism, its control logic employs a segmented threshold judgment mechanism. It relies solely on joint angle thresholds to classify discrete motion modes such as flat ground, climbing stairs, and squatting, pre-programming fixed stiffness and damping parameter tables, and directly calling preset parameters to adjust impedance during mode switching. Furthermore, the following joint control defects exist:

[0004] The adjustment link between spring stiffness and magnetorheological damping is mechanically coupled. Adjusting the spring preload will simultaneously change the load on the damper. It is impossible to independently decouple and control the two parameters K (stiffness) and B (damping). It can only achieve single-dimensional impedance adjustment, which makes it difficult to replicate the biomechanical characteristics of independent changes in stiffness and damping of human muscle-tendon units.

[0005] Using only fixed angle thresholds to classify movement modes is a "switch-type" discrete switching logic. It lacks a continuous gait phase φ fusion mechanism and cannot identify continuously changing working conditions such as gentle slopes, gradual steps, and user fatigue. The stiffness and damping parameters are fixed values ​​calibrated offline and lack an online closed-loop optimization link. The wearer's individual differences such as weight, gait symmetry, and fatigue level cannot be adapted in real time, resulting in a large impact on human-computer interaction and a significant increase in metabolic consumption during long-term walking.

[0006] Its MR damping only activates briefly at the moment of impact to buffer energy consumption, and only performs motor power generation to recover energy during a single window at the end of the swing, without a global gait phase coordinated recovery strategy; the two energy links of elastic energy storage and motor power generation lack hierarchical coordinated logic, the impact energy absorbed by the elastic element cannot be transferred to the motor power generation module, the pulse potential energy capture rate is less than 45%, and the improvement in battery range is low; at the same time, the DC-DC converter operates with a fixed number of phases and a fixed switching frequency, which cannot match gait pulse loads, and the conversion loss is high under light load conditions.

[0007] The motion pattern is determined solely by the joint encoder's single body sensor data, without integrating visual and laser ranging external terrain perception information, making it impossible to predict the slope and step height ahead; without an EKF extended Kalman filter continuous gait phase calculation module, it can only output off-walk gait events (ground contact, off-ground), lacking 0~100% continuous phase φ time sequence data, and cannot achieve staged fine impedance control.

[0008] Using a basic impedance control formula with a fixed ratio and differential coefficients, the parameters Kp and Kd are not dynamically updated with gait phase and terrain parameters. When the joint is disturbed by external force or gait oscillation, there is no decoupling control logic for instantaneous damping lift. It relies only on passive buffering by springs, which can easily lead to joint shaking and wearing imbalance. This is not safe enough for elderly rehabilitation and lower limb injury patients. Summary of the Invention

[0009] To address the technical problems existing in the prior art, the present invention provides the following technical solution:

[0010] On the one hand, a joint control method for a wearable exoskeleton robot based on variable stiffness and variable damping is provided, including the following steps:

[0011] S1. Multimodal cross-modal attention perception and continuous gait phase calculation: Raw data from the body sensor and external environment sensor are acquired and sensor error correction is performed. The calculated value range is determined using an extended Kalman filter (EKF). The continuous gait phase is used to output the terrain slope using a two-level cross-modal attention fusion network. Step height Online calculation of gait symmetry index User interaction torque Output structured feature vectors ;in Personalize quality factors for users;

[0012] S2. Gait phase coordinated stiffness-damping parameter hierarchical mapping: Using the structured feature vector as input, calculate the reference stiffness respectively. Reference damping The target stiffness is obtained by superimposing personalized quality factor scaling. Target damping Real-time detection of joint angular velocity oscillations; if oscillations are detected, only the damping coefficient is amplified while maintaining the target stiffness, and the final output is given. , With energy flow prediction tags;

[0013] S3, Stiffness / Damping Decoupling Servo Execution and Integrated Impedance Torque Synthesis: [This section appears to be incomplete and requires further context.] Mapped to the target displacement of the lever fulcrum The lever-type variable stiffness unit is driven by servo position closed-loop control to generate elastic torque. , This is the joint angle offset; Mapped to the target current of the magnetorheological fluid MRF damper excitation coil The damping torque is generated by driving the MRF damping unit through current closed-loop control. , The real-time angular velocity of the joint is used to obtain the total output impedance torque of the joint by superimposing these values. .

[0014] Preferably, the state vector of the Extended Kalman Filter (EKF) in step S1 is: The state prediction equation is The observation update equation is ;in Here is the gait state transition matrix. The process noise covariance matrix is... To observe the noise covariance matrix, The observation vector includes joint angles and corrected IMU angular velocities; the continuous gait phase is derived from the phase differential. The integral iteration yields a forced reset when a sudden change in heel impact is detected. Complete the gait cycle closed loop.

[0015] Preferably, the formula for calculating the attention weights of the two-level cross-modal attention fusion network in step S1 is as follows:

[0016] In the formula:

[0017] This is the Query vector for the ontology's gait temporal features. Let i be the i-th type of environmental perception feature key vector. The scaling factor for the feature vector dimension. The real-time attention weights for the i-th type of sensor are given, and the sum of the weights of all sensors is 1; the weighted fusion terrain parameters are output. , For single-sensor terrain parameter calculation; the swing phase enhances the attention weight of visual sensing, and the stable support phase enhances the attention weight of laser ranging and body sensing.

[0018] Preferably, the reference stiffness mapping formula in step S2 is: ;in The standard reference stiffness constant for flat ground. For the piecewise function of phase stiffness correction, This is the slope correction function. Here is the step height correction function; the reference damping mapping formula is: , The standard reference damping constant for flat ground. For the phase damping correction piecewise function, This is the slope damping correction function.

[0019] Preferably, the equivalent stiffness derivation formula for the lever-type variable stiffness unit in step S3 is as follows:

[0020] In the formula:

[0021] For the inherent linear stiffness of the preloaded disc spring assembly, The spring is fixed at its fixed end to the joint pivot, which is a long lever arm. The adjustable short lever arm of the servo screw drive fulcrum to the joint output arm; by adjusting Numerical implementation Continuous wide-range adjustment; stiffness servo position closed-loop control formula is: , This represents the actual position of the fulcrum measured by the lead screw and grating ruler. The drive voltage for the stiffness-adjustable servo motor.

[0022] Preferably, in step S3, the damping coefficient of the MRF damper and the excitation current have a linear mapping relationship as follows: , The damper's factory calibration constant; damping torque The negative sign indicates that the damping torque is opposite to the direction of the joint angular velocity. During the energy recovery phase, this damping torque drives the joint motor in the opposite direction to enter the power generation mode, converting the negative mechanical energy of the human body into electrical energy.

[0023] Preferably, the method further includes the following steps:

[0024] S4. Global Gait Multi-Mode Energy Coordinated Recovery and Adaptive DC-DC Management: Based on energy flow prediction tag matching, three preset gait energy recovery phase intervals are matched to execute a hierarchical energy recovery strategy of "elasticity priority, electricity recovery supplementation"; when the deformation of the elastic element exceeds the saturation threshold, the motor is activated to recover excess mechanical energy; a multi-phase interleaved bidirectional Buck-Boost DC-DC converter is used to dynamically adjust the number of activated phases according to gait energy flow prediction. When the energy recovery window is predicted to arrive, the Boost mode is switched in advance and the variable step size conductance incremental method MPPT algorithm is started to capture the pulse potential energy.

[0025] The three preset gait energy recovery phase intervals are: Mode 1, the end of the knee joint swing; Mode 2, the descent and ground contact phase; and Mode 3, the ankle joint rebound phase. When the phase falls into any interval, the high-damping parameter mapping and DC-DC pulse recovery mode in step S2 are triggered simultaneously.

[0026] Preferably, the adaptive DC-DC dynamic phase management (DPM) strategy in step S4 includes a three-level phase switching logic: all phases are enabled in high-power boost / pulse recovery operation. During stable operation and medium power consumption, half of the phases are activated; during light load standby conditions with oscillation, only single-phase operation is activated; the Global Efficiency Optimizer (GEO) runs in the background to monitor the converter's conversion efficiency in real time. With device junction temperature Dynamically adjust the phase number switching current threshold and self-calibrate the converter efficiency model online.

[0027] Preferably, the method further includes the following steps:

[0028] S5. Online Bayesian Adaptive Optimization Based on Implicit Performance Indicators: Construct a joint optimization loss function with minimizing the loss as the optimization objective. Apply small parameter perturbations to the baseline stiffness and baseline damping mapping coefficients in step S2, and observe N consecutive gait cycles. , As indicators change, the personalized mapping benchmark coefficients are updated iteratively.

[0029] Online Bayesian adaptive optimization serves as a low-frequency background auxiliary closed loop. The optimization iteration process includes: 1) Initializing the baseline mapping parameter set. ;2) Apply small parameter perturbations Obtain a new parameter set 3) Maintain Average data collected after running N consecutive gait cycles , Calculate the loss function ;4) Compare with the original ,like If the parameter set is not updated, the disturbance direction is changed and the test is repeated; the real-time control process of S1~S4 is not interrupted throughout the entire process.

[0030] 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.

[0031] 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.

[0032] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0033] 1. This invention can adjust the stiffness and damping of constant values ​​separately, and adjust the damping and stiffness of constant values ​​separately, achieving complete decoupled collaborative control; the axial dimension of the joint and the weight of the whole machine are greatly reduced, which can meet the needs of wearable lightweight. At the same time, the lever stiffness adjustment relies on the low power drive of the micro screw, and the MRF damping only consumes power from the excitation coil, reducing the static power consumption of the whole machine by 35%.

[0034] 2. Compared to traditional technical solutions such as US20230025654A1, this invention relies solely on joint angle thresholds to divide discrete motion modes. The parameters are fixed offline and cannot continuously adapt to gentle slopes, gradual steps, or user fatigue conditions. This invention, however, utilizes the EKF output of 0~100% continuous gait phase φ to construct a multivariate continuous impedance mapping function for phase, slope, and step height. Torque and impedance parameters change smoothly and continuously, eliminating mode switching impact during terrain changes. Coupled with online Bayesian closed-loop optimization, personalized parameters are autonomously iterated during daily walking, eliminating the need for laboratory calibration of a metabolic analyzer and adapting to real-time changes in the wearer's condition.

[0035] 3. Compared to existing technologies that rely solely on a single encoder for sensing, lacking external vision and laser terrain perception, and thus unable to predict steps or slopes ahead in advance, this invention constructs a two-level attention fusion perception network that integrates multi-source data from IMU, encoder, camera, and laser. The oscillating phase prioritizes visual anti-terrain sensing, while the supporting phase prioritizes laser body sensing. Simulation data incorporates sensor noise and drift errors, and hardware-in-the-loop distillation trains a lightweight embedded model. This improves the accuracy of complex terrain recognition by 24%, completes impedance parameter pre-adjustment 1-2 gait cycles in advance, and makes human-computer interaction smoother.

[0036] 4. This invention binds three gait negative power phase windows to simultaneously activate high-damping recovery, following the hierarchical energy management principle of "elasticity priority, electric recovery supplementation". The elastic element first absorbs low-frequency impact energy, and after saturation, the motor starts generating electricity to recover excess pulse potential energy. It is equipped with a predictive multiphase adaptive DC-DC converter, which switches the recovery window to Boost mode in advance for fast MPPT, increasing the pulse energy capture rate from 45% to 88% and extending battery life by more than 20%.

[0037] 5. This invention monitors joint angular velocity fluctuations in real time, maintains constant stiffness when oscillations are detected, and rapidly suppresses vibrations with instantaneous lifting damping. The decoupled control logic does not damage the joint support stability, significantly reducing the impact upon landing. It is suitable for lower limb injury rehabilitation and for elderly people with mobility difficulties to wear, avoiding secondary injuries.

[0038] 6. Existing patented sensing, impedance control, and energy recovery subsystems are independent of each other and lack unified timing scheduling; this invention builds a three-level unified closed-loop architecture, with high-level sensing output characteristics synchronously driving mid-level impedance mapping and low-level energy flow prediction, and an external online optimization loop continuously correcting control reference parameters. The four modules are synchronized in timing, and the whole machine control logic is integrated, reducing communication delay between multiple modules and doubling the system response bandwidth. Attached Figure Description

[0039] 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.

[0040] Figure 1 This is a schematic diagram of control layer logic provided in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the control method provided in an embodiment of the present invention;

[0042] Figure 3 This is a block diagram of the hierarchical mapping of damping parameters in this invention;

[0043] Figure 4 This is a schematic diagram of the impedance torque synthesis mechanism of the present invention. Detailed Implementation

[0044] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0045] 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.

[0046] 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.

[0047] 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.

[0048] 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.

[0049] I. Core of the Invention

[0050] This invention designs a layered closed-loop control method for joints of a wearable exoskeleton robot, where stiffness and damping parameters can be independently / coupled and coordinated. It utilizes a lever principle combined with a magnetorheological fluid (MRF) parallel decoupled integrated joint mechanism as the execution carrier. The method uses a multimodal perception fusion network to output continuous gait phase φ, terrain geometry parameters, and user motion state as inputs, constructing a three-tiered control architecture: high-level motion intention recognition, mid-level impedance parameter mapping, and low-level decoupled servo execution. The mid-level introduces a gait phase-stiffness / damping dynamic mapping rule to achieve differentiated impedance configurations throughout the entire gait cycle, including early support, push-off, late swing, and downhill ground contact. The low-level employs a decoupled impedance control law to drive the lever stiffness adjustment unit and the rotary MRF damping mechanism. The exoskeleton unit independently controls the K and B parameters, and simultaneously integrates a multi-mode energy recovery strategy of "elasticity priority and energy recovery supplementation" to accurately match the high impedance damping stage with the gait negative energy recovery window. It is equipped with a cross-modal attention perception model trained by hardware-in-the-loop distillation, which can identify complex terrains such as flat ground, gentle slopes, and stairs in real time. A lightweight Bayesian online optimizer is introduced to continuously iterate personalized impedance parameters, dynamically mimicking the complex impedance characteristics of human muscle-tendon units that adapt to changes in motion. This eliminates the defects of traditional series variable impedance joints such as mechanical coupling, static parameter switching, single energy recovery window, and lack of perception dimension, and significantly improves the exoskeleton's human-machine compliance, energy utilization efficiency, wearability safety, and all-terrain adaptability.

[0051] II. Explanation of the Core Technology Architecture of the Invention

[0052] like Figure 1 As shown, the control method of this invention is divided into a high-level multimodal motion perception and recognition layer, a middle-level phase-coordinated stiffness-damping parameter mapping layer, and a bottom-level decoupled servo execution and energy coordination layer. The three layers form a complete closed loop of top-down command transmission and bottom-up state feedback. At the same time, an external online adaptive Bayesian optimization loop continuously corrects the basic parameters of the middle-level mapping.

[0053] (a) High layer: Multimodal motion sensing and recognition layer (input: all raw sensor data; output: structured sensing feature vector)

[0054] Input hardware signals: IMU three-axis acceleration / angular velocity, joint encoder angle data, laser ranging point cloud, visual RGB image, joint torque sensor current signal;

[0055] Internal processing flow: Sensor error simulation correction → EKF extended Kalman filter calculation of continuous gait phase φ ( EKF filtering calculates the temporal gait phase (0% = heel strike, 100% = end of the swing before the next heel strike) → cross-modal attention fusion network outputs terrain parameters. (Topographic slope angle) Laser + vision fusion perception output (uphill is positive, downhill is negative). (Step height) The vertical height of the steps is calculated by external sensors, and the input parameters for the staircase scene are as follows: motion mode labels (flat ground / uphill / downhill / upstairs / downstairs), and implicit performance indicators. (gait symmetry index) To optimize implicit metrics online, the closer the gait is to 1, the more symmetrical the gait becomes, and the optimization objective is to maximize the metric. ), (User interaction torque) The wearer's output torque is estimated using a motor current / torque sensor, and the target is minimized.

[0056] Output structured feature vectors , which serves as the sole input to the intermediate parameter mapping module.

[0057] (ii) Middle layer: Phase-coordinated stiffness-damping parameter mapping layer (Input: High-level sensing features; Output: Target impedance parameters) )

[0058] Internal processing flow: Gait phase segmentation determination → Scene-specific stiffness reference mapping and damping reference mapping → Personalized factor scaling correction → Oscillation disturbance detection and damping compensation.

[0059] Output two sets of target commands: stiffness commands Damping command The data is distributed to the underlying stiffness adjustment servo controller and the MRF damping current controller respectively; the energy flow prediction label (assist / energy recovery standby / pulse recovery) is synchronously output to the DC-DC adaptive management unit.

[0060] (III) Bottom Layer: Decoupling Servo Execution and Energy Coordination Layer (Input: Target Impedance Parameters of the Middle Layer; Output: Actual Joint Torque, Energy Storage Charging and Discharging Current)

[0061] Branch 1 (Stiffness Decoupling Execution): Query the fulcrum displacement calibration curve and output the target position of the lead screw. Servo motor closed-loop adjustment lever fulcrum, real-time generation ;

[0062] Branch 2 (Damping Decoupling Execution): Query the current-damping calibration curve and the target current of the output coil. MRF damper real-time generation The total impedance torque is obtained by superposition. Drive joint;

[0063] Branch 3 (Energy Coordination Management): Dynamically adjust the number of activated phases of DC-DC converters based on mid-level energy flow prediction tags. The recovery window initiates Boost mode and variable step size MPPT to capture pulse potential energy, and performs layered energy recovery with "elastic priority and electrical recovery supplementation";

[0064] Real-time feedback from the underlying layer: actual joint angle angular velocity Actual coil current, actual lead screw fulcrum position, DC-DC input and output power, device temperature, and data are fed back to the higher-level sensing layer and online optimization loop.

[0065] (iv) External loop: Online Bayesian adaptive optimization loop input: Low-level feedback , Conversion efficiency Processing: Apply small parameter perturbations to the mid-level mapping reference coefficients, observe the changes in performance indicators over N gait cycles, and iteratively update the personalized scaling coefficients; Output: The corrected reference stiffness and damping mapping coefficients are written into the mid-level parameter mapping table to achieve long-term autonomous personalized adaptation.

[0066] IV. Control Method of the Invention

[0067] like Figure 2 As shown, this control method consists of the following steps:

[0068] Step 1: Multimodal cross-modal attention perception and continuous gait phase resolution

[0069] By uniformly collecting raw data from the main body and external multi-source sensors, sensor drift, noise, and distortion errors are corrected, and uninterrupted continuous gait phase is output through EKF filtering. It relies on a two-level attention fusion network to achieve adaptive weighted fusion of vision, laser, and proprioception to quantify and output terrain slope. Step height Gait symmetry User interaction torque This provides complete and accurate time-series input characteristics for mapping mid-level impedance parameters, addressing the shortcomings of existing technologies such as single sensing, off-state events, and lack of look-ahead terrain prediction. It includes the following sub-steps:

[0070] Sub-step 1.1: Multi-source sensor error modeling and raw data preprocessing

[0071] A unified error model is established to address IMU low-frequency drift, camera motion blur distortion, laser ranging distance noise, and encoder quantization error. Hardware defect compensation is performed at the data input end to reduce the data deviation between simulation and real domains and improve the generalization ability of the perception model.

[0072] Taking IMU angular velocity data correction as an example:

[0073] Raw angular velocity sampled value from the IMU (rad / s);

[0074] : IMU low-frequency drift bias, static offline calibration constant;

[0075] Gaussian white noise random disturbance term, with a mean of 0 and a variance given by the sensor manual;

[0076] The corrected true angular velocity is input into the EKF filter module. Laser ranging distance correction is as follows:

[0077] : The original laser ranging reading; Attenuation coefficient at the angle of incidence; : Angle offset compensation amount;

[0078] The embedded main control NPU performs multi-channel error correction in parallel. After eliminating the inherent bias of the sensor, the error in terrain parameter calculation can be reduced to less than 3%, and the drift in gait phase calculation can be reduced by 85%.

[0079] Sub-step 1.2: EKF Extended Kalman Filter Continuous Gait Phase Solution

[0080] Using the heel strike event as the phase zero point, the gait dynamics state equation is constructed by integrating joint angles and IMU angular velocity. The state variables are updated iteratively using EKF nonlinear filtering, and the output is a continuous phase of 0~100%, replacing the discrete gait state segmentation of the traditional finite state machine, and realizing fine impedance adjustment under arbitrary phase.

[0081] The EKF state equations and observation equations are expressed as follows:

[0082] State vector: State prediction equation: Observation update equation: ,

[0083] Gait state transition matrix, derived from the human lower limb kinematic model;

[0084] Process noise covariance matrix; : Observation noise covariance matrix;

[0085] : Observation vector, containing knee joint angle and IMU angular velocity data after correction; for example, the filtered output of normal walking on flat ground. After 1.5 seconds of integration, the result is obtained. It is in the early stage of support; when a sudden change in heel impact torque is detected, it immediately resets. Restart the periodic integration, without phase jumps or breaks.

[0086] Phase Iterative Update Core Derivation: Phase Differentiation of Filter Output Representing the rate of phase change, integration yields the continuous time sequence. ,

[0087] Force reset when a sudden change in heel strike observation is detected. This completes the cycle closed loop. It achieves uninterrupted phase output throughout the entire cycle, with a phase update frequency of 1kHz and a phase error of ≤0.5%, supporting fine-grained impedance control in the middle layer at different stages.

[0088] Sub-step 1.3: Terrain parameter calculation using a two-level cross-modal attention fusion network

[0089] Building a two-level attention deep learning network:

[0090] The first stage completes the fusion of visual and laser environmental perception, outputting terrain type and slope. Step height ;

[0091] The second-level cross-modal attention module uses the gait features of the body as the query and the terrain features as the key / value. It dynamically adjusts the weights of vision, laser, and IMU sensors, prioritizing vision for terrain in the swaying phase and prioritizing laser and body sensing in the stable support phase, thus adapting to complex scenarios.

[0092] For example, constructing separate student network models and teacher network models:

[0093] Lightweight student network (embedded deployment): 3-layer one-dimensional convolution to extract ontology temporal features, 2-layer convolution to extract visual image features, attention module hidden layer dimension 64, INT8 quantization, model parameter count ≤280k, RK3588 NPU inference latency ≤3ms.

[0094] Teachers use a network server for training, employing hardware-in-the-loop distillation training and randomly masking the visual / laser modalities to enhance robustness.

[0095] The attention weighting is set as follows:

[0096] ,

[0097] : Ontology gait temporal feature vector (Query);

[0098] : The i-th type of environmental perception feature vector (Key), where T represents the transpose of the vector matrix;

[0099] Feature vector dimension scaling factor;

[0100] : Real-time attention weights for the i-th type of sensor Weighted fusion terrain parameter output: , Calculate terrain parameters using a single sensor.

[0101] Sub-step 1.4: Online calculation of implicit performance indicators ( , )

[0102] (1) Gait symmetry index Calculation formula

[0103] ,

[0104] : Duration of the left / right leg swing phase; The closer to 1, the more symmetrical the gait, and the better the positive indicator.

[0105] (2) User interaction torque Estimation formula

[0106]

[0107] The total torque measured by the joint torque sensor; The output impedance torque of the exoskeleton joint; the difference is the torque exerted by the wearer's own force, and the optimization goal is to minimize its absolute value.

[0108] The final step is to output the overall structured feature vector. The data is then transmitted to the middle-layer parameter mapping module.

[0109] Step 2: Layered mapping of gait phase coordinated stiffness-damping parameters

[0110] like Figure 3 As shown, using the multi-dimensional perception features output from step one as input, a phase-separated and scenario-separated stiffness / damping benchmark mapping function is established based on the biomechanical laws of the human lower limb. This is then overlaid with a user-personalized mass factor scaling correction, real-time detection of joint oscillation disturbances, and instantaneous damping compensation. Finally, the decoupled target stiffness is output. Target damping This enables independent or coupled adjustment of stiffness and damping, overcoming the shortcomings of existing technologies such as fixed parameters, on / off switching, and lack of personalized adaptation. It includes the following sub-steps:

[0111] Sub-step 2.1: Gait phase mapping to scene segmented reference impedance

[0112] (1) Stiffness datum mapping, defined as:

[0113]

[0114] : Standard stiffness constant for flat ground;

[0115] The phase stiffness correction function needs to be defined piecewise, for example:

[0116] Support early ): (High rigidity, resistant to impact upon landing);

[0117] Pedaling period ): (Low stiffness, elastic energy release aids propulsion);

[0118] Swing period ): (Medium stiffness, reducing oscillation drag);

[0119] End of oscillation : (Medium stiffness, preparing to decelerate and touch down);

[0120] Slope correction function, uphill Linear lifting reference stiffness, downhill Slightly reduce stiffness;

[0121] Step height correction function: the higher the step, the linearly the reference stiffness increases.

[0122] (2) Damping reference mapping

[0123]

[0124] : Standard reference damping constant for flat ground;

[0125] The phase damping correction function is also defined piecewise, as shown in the example below:

[0126] Supporting early, downslope contact: (High damping, absorbing impact potential energy);

[0127] Late swing phase and ankle rebound phase: (High damping, synergistically slowing down muscles and recovering energy);

[0128] Stable support during the push-off period: (Low damping, does not interfere with natural extension and force generation).

[0129] (3) Personalized factor scaling correction (core personalized adaptation)

[0130] , , To calibrate the user's weight scaling factor offline, the larger the weight, the better. A higher value automatically increases impedance under the same terrain phase to match the wearer's load requirements. For example, a wearer... Walking on flat ground (Supporting the early stages) Substituting, we get: , ; , This achieves high-rigidity, high-damping buffering upon landing.

[0131] Sub-step 2.2: Instantaneous compensation for oscillation disturbance damping (decoupling safety control)

[0132] (1) Disturbance determination logic

[0133] Real-time acquisition of joint angular velocity fluctuation variance at the bottom layer ,like (Oscillation threshold) determines that the joint is experiencing vibration disturbance, performs damping compensation, keeps the stiffness unchanged, and achieves complete decoupling control of K and B.

[0134] (2) The compensation mechanism is as follows:

[0135] , This is the oscillation damping amplification factor, after the disturbance disappears. Regression 1, Stiffness The entire process remains unchanged and does not affect the stability of the support.

[0136] Final target impedance parameters , Simultaneously output energy flow prediction labels (boost / recovery / standby).

[0137] Step 3: Stiffness / Damping Decoupling Servo Execution and Integrated Impedance Torque Synthesis

[0138] like Figure 4 As shown, this step involves outputting the middle layer... , These are converted into lever fulcrum displacement servo commands and MRF coil current drive commands, respectively. The two execution links are physically decoupled and controlled independently to generate elastic torques. Damping torque The total output impedance torque of the joint is obtained by vector superposition. It replicates the synergistic mechanism of elastic energy storage and damping dissipation in human muscles, solving the problems of K / B coupling and response lag in series mechanisms.

[0139] Sub-step 3.1: Decoupling servo control of lever variable stiffness unit

[0140] (1) Derivation of equivalent stiffness of lever (core mechanical principle)

[0141] Based on the principle of torque lever balance: Equivalent output stiffness of the joint: ,

[0142] : Inherent linear stiffness of the disc spring assembly;

[0143] Movable fulcrum short lever arm (servo screw adjustable variable) Unique correspondence Spring fulcrum to joint output arm lever arm Specifically, the short lever arm of the lever stiffness adjustment mechanism is changed by a servo motor driving the lead screw to move the fulcrum. ;

[0144] : Fixed long lever arm constant, which is the lever arm from the fixed end of the spring to the joint axis of rotation. With a fixed constant, a long lever arm, and a fixed mechanical structure, the logic is: decrease. , As the numerical value increases, the equivalent stiffness Linear increase; increase This reduces stiffness, allowing for a wide range of continuous adjustability.

[0145] (2) Servo position closed-loop control

[0146] , ,

[0147] Middle level The target displacement of the fulcrum is obtained from the table: mm. The target position of the stiffness adjustment servo screw is determined by the uniquely mapped equivalent stiffness. ;

[0148] : Actual position of the fulcrum measured by the lead screw and grating ruler;

[0149] Stiffness adjustment servo motor drive voltage; These are the PID coefficients for the position loop.

[0150] (3) Elastic torque output

[0151] , , The dynamic balance angle of the gait phase is used, and feedforward compensation eliminates static offset.

[0152] Sub-step 3.2: Decoupling current control of MRF variable damping unit

[0153] (1) Damping coefficient-excitation current calibration mapping relationship

[0154] The MRF damper calibration constant is defined as follows: the magnetic field strength is linearly related to the coil current, and the damping coefficient is continuously adjustable with the current; the target control current of the MRF excitation coil is defined as follows: A, the drive current output from the underlying controller to the damper coil is proportional to... .

[0155] (2) Damping torque output (coupled with gait phase and angular velocity)

[0156] The negative sign indicates that the damping torque is opposite to the direction of the joint angular velocity, and dissipates kinetic energy during movement; during the energy recovery stage, this torque drives the motor to generate electricity in the opposite direction, converting it into electrical energy for storage.

[0157] (3) Current closed-loop control logic

[0158] , The PWM output of the coil is adjusted in real time, with millisecond-level response to damping changes.

[0159] Sub-step 3.3: Integrated synthesis of total resistance and torque

[0160] .

[0161] For example, during the downhill landing phase of a wearer: , , , (The downward rotational angular velocity is negative) ; Total torque High elasticity support + high damping buffer absorbs the impact potential energy of downhill slopes and simultaneously enters energy recovery mode.

[0162] The output of this step includes the servo motor drive voltage, MRF coil PWM current, and total joint output torque. It will also feed back to the higher-level perception and optimization loop for further optimization and learning.

[0163] Step 4: Global Gait Multi-Mode Coordinated Energy Recovery and Adaptive DC-DC Management

[0164] The system implements hierarchical energy management with "elasticity priority and electrical recovery supplementation" by binding mid-level energy flow prediction tags and three gait energy recovery windows. Elastic elements prioritize absorbing impact negative work, and when the elastic load threshold is exceeded, the MRF high-damping + motor power generation is activated to recover excess mechanical energy. It is equipped with a 4-phase interleaved bidirectional Buck-Boost DC-DC converter, which dynamically adjusts the number of activated phases based on gait prediction. The recovery window initiates predictive fast MPPT, maximizing pulse potential energy capture efficiency and solving the problems of single recovery window and low efficiency of converter over wide load in existing technologies.

[0165] Sub-step 4.1: Three-window phase coordinated energy recovery determination logic

[0166] Three major recovery phase intervals (and high damping) (The stages completely overlap)

[0167] Mode 1: Late phase of knee swing (), slow down the lower leg and recover the kinetic energy of the swing;

[0168] Pattern 2: Descending phase (downhill, descending stairs and touching the ground) ), recover gravitational potential energy;

[0169] Mode 3: Ankle Rebound Phase ), recovering the elastic rebound kinetic energy of the ankle joint; judgment rule: middle layer output phase Falling into any range, synchronously triggering high-damping parameters The system is upgraded, and "pulse energy recovery" tags are sent to the DC-DC module.

[0170] Sub-step 4.2: Elastic priority stratified energy recovery control law

[0171] Threshold determination: Monitoring spring deformation ,like (Elastic saturation threshold), start the motor to generate electricity to supplement recovery; energy stratification: ,

[0172] Elastic elements passively store mechanical energy (zero power consumption);

[0173] The motor generates electricity to recover excess mechanical energy, which is then stored in the battery via DC-DC converter.

[0174] Sub-step 4.3: Adaptive Multiphase DC-DC Dynamic Phase Management (DPM)

[0175] (1) Number of phases Switching rules

[0176] High-power boost / pulse recovery: It shares the current across all phases, reducing conduction losses;

[0177] Power output during smooth walking: Balance the switch and reduce conduction losses;

[0178] Standing / Swinging Light Load Standby: Minimize switching drive losses;

[0179] DC-DC converter enabling phase number It is positively correlated with the wearer's weight through offline calibration and used for basic scaling of impedance parameters.

[0180] (2) Predictive fast MPPT control (dedicated to the recovery window, combined with the Boost circuit of the exoskeleton robot controller for operation control)

[0181] : Initial output voltage setting value of the Boost converter circuit, in V; The boost reference voltage obtained by pre-calculation based on the gait phase is the target output voltage of the Boost converter at the moment of startup, used to replace the fixed initial value and realize phase prediction and pre-boost;

[0182] : Motor no-load / current instantaneous induced electromotive force reference voltage, unit V; that is, the original output voltage of the motor under gait generation conditions without phase compensation, which is directly determined by the gait speed and the back electromotive force constant of the motor windings, and is the basic basis of the boost voltage.

[0183] Gait phase compensation coefficient: a system offline calibration parameter, jointly tuned by motor internal resistance, Boost circuit gain range, and gait phase voltage fluctuation amplitude. It is used to quantify the phase correction weight on the output voltage and is fixedly written into the controller.

[0184] Predicting the gait advance phase value (phase normalization interval [-1,1] or [0,1]) is a way to predict the motion phase at the next moment in advance using gait sensors and motion timing models. It is an advance prediction quantity, not a real-time sampled phase, and is the core input for "pre-setting the initial duty cycle of Boost".

[0185] In the scenario of human gait-based power generation, the motor output voltage fluctuates periodically with the gait phase: the back electromotive force of the motor differs greatly during the leg lift, landing, and support phases. If the Boost circuit uses a fixed initial boost voltage, the voltage deviation is large at each phase switch, requiring the incremental conductance MPPT algorithm to undergo numerous iterations to match the optimal operating point, resulting in slow convergence speed and high pulse energy loss. By introducing advance prediction of gait phase for pre-compensation, the optimal initial boost voltage that fits the current gait is directly calculated at the moment of Boost startup / phase switching, significantly reducing the initial search range of the MPPT algorithm.

[0186] The control process is as follows:

[0187] The motion prediction module acquires gait acceleration / angular velocity signals in real time and outputs the predicted phase for the next cycle through a time-series prediction algorithm. ;

[0188] Read the current real-time base electromotive force of the motor ;

[0189] Substitute the fixed calibration compensation coefficient Calculate the phase correction term :

[0190] The phase is within the peak power generation range: If the value is positive, the correction term is positive, increasing the initial boost voltage;

[0191] The phase is in the low power generation range: If the value is negative, the correction term is negative, reducing the initial boost voltage;

[0192] The initial target voltage of the Boost converter is obtained by superimposing the phase correction gain on the base voltage. ;

[0193] The controller calculates the corresponding initial duty cycle of the Boost based on this voltage and applies it directly to the power switch transistor;

[0194] Starting with this pre-compensated duty cycle, the variable step size conductance incremental MPPT algorithm is initiated for fine optimization:

[0195] The initial point is close to the maximum power point, so a large-scale coarse search is not required;

[0196] The algorithm uses a dynamic step size. When the voltage deviation is large, the step size is large to quickly approximate the target. When the target is near the optimal point, the step size is automatically reduced to stabilize the output.

[0197] Therefore, phase prediction and pre-compensation can be achieved, directly outputting the boost initial voltage adapted to the gait, and calculating the initial duty cycle, thus solving the problem of large deviations in traditional fixed initial values. Traditional non-phase pre-compensation schemes typically have convergence times of tens of milliseconds. This pre-phase compensation formula combined with variable step-size MPPT sets the initial duty cycle of the boost voltage in advance based on the predicted gait phase, and uses the variable step-size conductance incremental method to quickly match the motor's internal resistance, compressing the pulse energy capture convergence time to less than 3ms, significantly improving the energy utilization rate of gait pulse power generation.

[0198] (3) Global Efficiency Optimizer (GEO) Background Iteration

[0199] Continuous monitoring of conversion efficiency (DC-DC conversion efficiency) (Monitoring real-time conversion efficiency and device junction temperature through a global efficiency optimizer) Dynamically adjust the phase number switching current threshold Online self-calibrating converter efficiency model, and scenarios such as adapter aging and temperature drift.

[0200] The Global Efficiency Optimizer (GEO) continuously iterates and calibrates asynchronously in the background, without interfering with the normal power conversion of the main power circuit. The system collects DC-DC input and output voltages and currents in real time and calculates the conversion efficiency in real time. The value range is limited to [0,1], and the junction temperature of the power transistor and inductor is sampled synchronously. As a constraint variable.

[0201] Iterative calibration is performed in three cyclical steps: The first step is data sampling and storage. The GEO caches multiple sets of operating condition data according to a fixed control cycle, including real-time efficiency, junction temperature, current load current, and current phase switching threshold. The first step involves removing instantaneous impact noise data and constructing a temperature-load-efficiency operating condition sample library. The second step is online model self-calibration, which calls the built-in converter theoretical efficiency model, compares the difference between the sampled measured efficiency and the model's predicted efficiency, and uses gradient descent iterative correction of the loss coefficients to update the corresponding model parameters for conduction loss, switching loss, and core loss. This eliminates model deviations caused by device aging, resistor temperature drift, and core performance degradation, completing real-time fitting of the efficiency model. The third step is threshold iterative optimization, using the maximization of the overall conversion efficiency as the objective function, with small perturbations. Two sets of thresholds were used to compare the results under the same temperature and load conditions before and after the disturbance. Numerical values; if efficiency improves, the new threshold is retained; if efficiency decreases, the original value is reverted, while a junction temperature constraint is added: T When the load exceeds the safe range, the switching current for light-load phases is automatically increased to reduce the number of conducting phases and lower heat generation. The entire iterative cycle runs at low frequency in the background, slowly fine-tuning threshold parameters to avoid output voltage fluctuations. Continuously adapting to long-term aging of adapter components and drift in high and low temperature environments, the system dynamically matches the optimal operating phase range of the multiphase converter, achieving real-time optimal global conversion efficiency across all operating conditions.

[0202] Step 5: Online Bayesian Adaptive Optimization Based on Implicit Performance Metrics (Auxiliary)

[0203] With gait symmetry Maximize, user interaction torque Minimization is the joint optimization objective. A lightweight Bayesian optimization algorithm is used to apply small perturbations to the mid-layer stiffness and damping reference mapping coefficients. The performance index changes are observed over N gait cycles, and personalized mapping parameters are iteratively updated. No offline metabolic calibration in the laboratory is required, enabling continuous and autonomous adaptation to user fatigue and gait changes during daily walking.

[0204] Optimize the objective loss function

[0205] , Let be the weighting coefficients, and minimize the loss as the optimization objective.

[0206] Optimize the iteration process:

[0207] Initialize the current reference parameter set (Slope, step scaling factor);

[0208] Apply small parameter perturbation Generate a new parameter set ;

[0209] maintain Run N consecutive gait cycles and collect average values. , calculate ;

[0210] Compared to the original :like If necessary, permanently update the parameter set; otherwise, abandon the perturbation, change the perturbation direction, and retest.

[0211] The optimization iteration is repeated every M stable gait cycles, and the low-frequency background operation does not affect real-time control.

[0212] For example, when a user walks uphill at an 8° angle, the initial slope scaling factor... Optimize perturbation to After observing 50 gait cycles: Increase by 5%. Reduced by 3%, Loss decreased, system permanently updated. At the same slope, the assist is more in line with human biomechanics.

[0213] Example 1: Switching to descending stairs on level ground

[0214] I. Selection of Exoskeleton Robot Hardware Platform (Refer to commercially available exoskeleton robots; this section provides preferred hardware examples)

[0215] 1. Knee joint body with variable stiffness and variable damping

[0216] Variable stiffness unit: preloaded nonlinear disc spring assembly, precision ball screw, miniature servo adjustable motor, movable lever fulcrum; lever arm Fixed at 80mm, Adjustment range 20~70mm, equivalent stiffness adjustment range ;

[0217] Variable damping unit: Rotary MRF magnetohydrodynamic damper, excitation coil drive current 0~2A, damping range It is coaxially connected in parallel with the stiffness element for output, thus achieving physical decoupling;

[0218] 2. Sensing sensor kit

[0219] Body sensing: six-axis IMU, 16-bit high-precision joint encoder, joint torque sensor;

[0220] External environment sensing: monocular RGB camera, laser ranging radar;

[0221] 3. Embedded computing platform

[0222] High-level perceptual reasoning: RK3588 ARM embedded platform, NPU-accelerated cross-modal attention-based perceptual network reasoning;

[0223] The underlying servo control uses an ARM Cortex-M7 microcontroller with a 1kHz control frequency to execute a PID position / current closed-loop control.

[0224] DC-DC power control: TIC2000 digital power chip, 4-phase interleaved bidirectional GaNBuck-Boost converter;

[0225] 4. Energy storage system: 24V lithium battery pack, equipped with supercapacitor buffer pulse recovery energy.

[0226] II. Basic parameters of the wearer

[0227] The wearer weighs 75kg and has their personalized quality factor calibrated offline. Initial mid-layer reference stiffness Reference damping Factory calibration complete; initial slope scaling factor. .

[0228] III. Complete Scenario Implementation Process: Stable walking on flat ground → Stairs appearing ahead → Descending the stairs (entire process)

[0229] Phase 1: Stable walking on flat ground ( Cycle from 0 to 100%. (No steps)

[0230] Step 1: Sensing and Processing: Laser + Vision determines the terrain to be flat; EKF continuously outputs gait phase, at the end of the swing phase. Trigger energy recovery mode one; , Mean 6.2 N·m;

[0231] Step 2 Parameter Mapping: Supporting Early Stages , , ; Departure period Automatic switching low impedance , The elastic element releases stored energy to assist in the extension;

[0232] Step 3: Decoupling Execution: The servo motor adjusts the lever fulcrum to... match The MRF coil outputs a 0.7A current to match the damping; at the end of the oscillation, the damping is raised, generating a reverse damping torque to recover the oscillation kinetic energy;

[0233] Step Four: Energy Management - Smooth Walking During the final recovery window of the swing phase, Boost mode MPPT is activated to capture kinetic energy and store it in the battery.

[0234] Step 5: Online optimization: Backend Bayesian optimization is continuously monitored. The current loss is stable and there are no parameter updates.

[0235] Phase 2: Visual laser forward-looking recognition of the stairs ahead (transition phase) )

[0236] Cross-modal attention networks enhance laser ranging weights for calculating step height. The terrain label is switched to "going downstairs", and the energy flow prediction label is updated to "descent potential energy recovery"; the mid-level mapping function automatically raises the benchmark damping coefficient during the landing phase.

[0237] Phase 3: Descending the stairs and touching the ground ( (During the period of declining center of gravity, recovery mode two is activated)

[0238] Perception layer output (downhill), ;

[0239] Mid-level mapping: Extremely high damping compensation is triggered instantaneously upon landing. , Oscillation detection is free of jitter and requires no additional damping amplification.

[0240] Underlying execution: The MRF coil current is increased to 1.6A, the damping torque is significantly increased, and the lever fulcrum is slightly adjusted to maintain medium stiffness; Simultaneously, it achieves elastic support and high damping buffer to absorb the impact potential energy of the stairs falling down.

[0241] DC-DC management: Predicting pulse potential energy recovery and switching in advance. Full-phase operation, MPPT quickly matches the internal resistance of the generator motor to capture the potential energy of gravity during falling;

[0242] Online optimization: 30 consecutive downhill gait cycles were collected and calculated. , The loss decreased by 12% compared to the initial parameters, and the Bayesian optimization permanently updated the step height scaling factor. The impedance parameters of stairs at the same height will be further optimized in the future.

[0243] IV. Actual Measured Results

[0244] Control: No sudden impedance shock during transitions on flat ground or down stairs; smooth and continuous torque output; no switching or lag.

[0245] Human-machine performance: The wearer's average interaction force torque is reduced by 28%, gait symmetry is improved by 6%, and the peak impact upon landing is reduced by 41%;

[0246] In terms of energy efficiency: the full-area three-window energy recovery system increases the total recovered energy by 2.7 times compared to the comparative patent, and extends the battery's single-charge range by 22%;

[0247] Adaptive aspect: No secondary laboratory calibration is required; personalized parameters are automatically iterated during the descent, adapting to fatigue levels during long-term walking.

[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] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0257] 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.

[0258] 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 joint control method for a wearable exoskeleton robot based on variable stiffness and variable damping, characterized in that, Includes the following steps: S1. Multimodal cross-modal attention perception and continuous gait phase calculation: Raw data from the body sensor and external environment sensor are acquired and sensor error correction is performed. The calculated value range is determined using an extended Kalman filter (EKF). The continuous gait phase is used to output the terrain slope using a two-level cross-modal attention fusion network. Step height Online calculation of gait symmetry index User interaction torque Output structured feature vectors ;in Personalize quality factors for users; S2. Gait phase coordinated stiffness-damping parameter hierarchical mapping: Using the structured feature vector as input, calculate the reference stiffness respectively. Reference damping The target stiffness is obtained by superimposing personalized quality factor scaling. Target damping Real-time detection of joint angular velocity oscillations; if oscillations are detected, only the damping coefficient is amplified while maintaining the target stiffness, and the final output is given. , With energy flow prediction tags; S3, Stiffness / Damping Decoupling Servo Execution and Integrated Impedance Torque Synthesis: [This section appears to be incomplete and requires further context.] Mapped to the target displacement of the lever fulcrum The lever-type variable stiffness unit is driven by servo position closed-loop control to generate elastic torque. , This is the joint angle offset; Mapped to the target current of the magnetorheological fluid MRF damper excitation coil The damping torque is generated by driving the MRF damping unit through current closed-loop control. , This refers to the real-time angular velocity of the joint. The total output impedance torque of the joint is obtained by superposition. .

2. The control method according to claim 1, characterized in that, In step S1, the state vector of the Extended Kalman Filter (EKF) is: The state prediction equation is The observation update equation is ;in Here is the gait state transition matrix. The process noise covariance matrix is... To observe the noise covariance matrix, The observation vector includes joint angles and corrected IMU angular velocities; the continuous gait phase is derived from the phase differential. The integral iteration yields a forced reset when a sudden change in heel impact is detected. Complete the gait cycle closed loop.

3. The control method according to claim 1, characterized in that, The formula for calculating the attention weights of the two-level cross-modal attention fusion network in step S1 is as follows: In the formula: This is the Query vector for the ontology's gait temporal features. Let i be the i-th type of environmental perception feature key vector. The scaling factor for the feature vector dimension. Let be the real-time attention weight for the i-th type of sensor, and the sum of the weights of all sensors is 1; Weighted fusion terrain parameter output , For single-sensor terrain parameter calculation; the swing phase enhances the attention weight of visual sensing, and the stable support phase enhances the attention weight of laser ranging and body sensing.

4. The control method according to claim 1, characterized in that, The formula for the reference stiffness mapping in step S2 is: ;in The standard reference stiffness constant for flat ground. For the piecewise function of phase stiffness correction, This is the slope correction function. Here is the step height correction function; the reference damping mapping formula is: , The standard reference damping constant for flat ground. For the phase damping correction piecewise function, This is the slope damping correction function.

5. The control method according to claim 1, characterized in that, The equivalent stiffness derivation formula for the lever-type variable stiffness element in step S3 is as follows: In the formula: For the inherent linear stiffness of the preloaded disc spring assembly, The spring is fixed at its fixed end to the joint pivot, which is a long lever arm. The adjustable short lever arm of the servo screw drive fulcrum to the joint output arm; by adjusting Numerical implementation Continuous wide-range adjustment; stiffness servo position closed-loop control formula is: , This represents the actual position of the fulcrum measured by the lead screw and grating ruler. The drive voltage for the stiffness-adjustable servo motor.

6. The control method according to claim 1, characterized in that, In step S3, the linear mapping relationship between the damping coefficient of the MRF damper and the excitation current is as follows: , The damper's factory calibration constant; damping torque The negative sign indicates that the damping torque is opposite to the direction of the joint angular velocity. During the energy recovery phase, this damping torque drives the joint motor in the opposite direction to enter the power generation mode, converting the negative mechanical energy of the human body into electrical energy.

7. The control method according to claim 1, characterized in that, It also includes the following steps: S4. Global Gait Multi-Mode Energy Coordinated Recovery and Adaptive DC-DC Management: Based on energy flow prediction tag matching, three preset gait energy recovery phase intervals are matched to execute a hierarchical energy recovery strategy of "elasticity priority, electricity recovery supplementation"; when the deformation of the elastic element exceeds the saturation threshold, the motor is activated to recover excess mechanical energy; a multi-phase interleaved bidirectional Buck-Boost DC-DC converter is used to dynamically adjust the number of activated phases according to gait energy flow prediction. When the energy recovery window is predicted to arrive, the Boost mode is switched in advance and the variable step size conductance incremental method MPPT algorithm is started to capture the pulse potential energy. The three preset gait energy recovery phase intervals are: Mode 1, the end of the knee joint swing; Mode 2, the descent and ground contact phase; and Mode 3, the ankle joint rebound phase. When the phase falls into any interval, the high-damping parameter mapping and DC-DC pulse recovery mode in step S2 are triggered simultaneously.

8. The control method according to claim 7, characterized in that, Step S4's adaptive DC-DC dynamic phase management (DPM) strategy includes a three-level phase switching logic: all phases are enabled in high-power boost / pulse recovery operation. During stable operation and medium power consumption, half of the phases are activated; during light load standby conditions with oscillation, only single-phase operation is activated; the Global Efficiency Optimizer (GEO) runs in the background to monitor the converter's conversion efficiency in real time. With device junction temperature Dynamically adjust the phase number switching current threshold and self-calibrate the converter efficiency model online.

9. The control method according to claim 1, characterized in that, It also includes the following steps: S5. Online Bayesian Adaptive Optimization Based on Implicit Performance Indicators: Construct a joint optimization loss function with minimizing the loss as the optimization objective. Apply small parameter perturbations to the baseline stiffness and baseline damping mapping coefficients in step S2, and observe N consecutive gait cycles. , As indicators change, the personalized mapping benchmark coefficients are updated iteratively. Online Bayesian adaptive optimization serves as a low-frequency background auxiliary closed loop. The optimization iteration process includes: 1) Initializing the baseline mapping parameter set. ;2) Apply small parameter perturbations Obtain a new parameter set 3) Maintain Average data collected after running N consecutive gait cycles , Calculate the loss function ;4) Compare with the original ,like If necessary, the parameter set will be permanently updated; otherwise, the perturbation direction will be changed and the test will be repeated.

Images