An Adaptive Control Method for Exoskeleton Based on a Two-Layer Computational Architecture and Bayesian Optimization

By employing a dual-layer computing architecture and a Bayesian-optimized exoskeleton adaptive control method, the lower-level and upper-level computers collaborate to achieve coordinated operation of online parameter optimization and underlying hardware real-time control. This solves the problems of exoskeleton parameter adaptability and safety, ensuring the stability and safety of the exoskeleton.

CN123060702APending Publication Date: 2026-11-10HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611489297.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-23
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

While existing technologies can improve the adaptability of lower limb exoskeleton parameters by using online parameter optimization methods with high computational requirements, they are difficult to maintain the deterministic real-time operation of the underlying dynamic control. Furthermore, communication processing, parameter mutations, and anomalies of nodes at different levels have a significant impact on the safety control of the exoskeleton.

Method used

An exoskeleton adaptive control method based on a two-layer computing architecture and Bayesian optimization is adopted. The lower-level machine is responsible for gait information acquisition and real-time motor control, while the upper-level machine performs Bayesian parameter optimization. Data is uploaded at high frequency through DMA, and a multi-dimensional weighted cost function is constructed for parameter evaluation. Physical safety boundary clamping and first-order low-pass filtering are performed on the lower-level machine to ensure parameter safety and stability.

Benefits of technology

It enables the coordinated operation of online parameter optimization and underlying hardware real-time control, reduces the obstruction of real-time control by communication and parameter parsing, limits the mechanical safety range of abnormal parameters, and improves the safety and stability of the exoskeleton.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123060702A_ABST
    Figure CN123060702A_ABST
Patent Text Reader

Abstract

This invention discloses an adaptive control method for exoskeletons based on a two-layer computing architecture and Bayesian optimization, belonging to the field of lower limb exoskeleton control technology. The lower-level computer collects gait data and motor feedback from the inertial measurement unit (IMU) in hard real-time cycles, determines the gait phase, and uploads the status data to the upper-level computer via direct memory access (DMA). Based on data from multiple gait cycles, the upper-level computer performs a weighted evaluation of safety, stability, and comfort, and uses a Gaussian process and upper confidence limit (UCB) strategy to optimize impedance and torque parameters online, sending them down at low frequency. The lower-level computer asynchronously receives and parses the parameters, which are then used for motor control after being clamped by physical safety boundaries and subjected to first-order low-pass filtering. Interrupt priority is used to ensure that the power control task is executed first, thereby reducing the impact of high-level computing, communication processing, and abnormal parameters on the underlying hard real-time control and power execution safety.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of wearable robot control and medical rehabilitation device technology, and in particular to an adaptive control method for exoskeleton based on a two-layer computing architecture and Bayesian optimization. Background Technology

[0002] Powered lower limb exoskeletons can provide motion assistance to the human lower limbs through joint actuators. In order to coordinate the exoskeleton's movement with the human gait, the control system usually needs to collect the human's motion status and actuator feedback, and adjust the exoskeleton's output according to the gait stage or feedback status.

[0003] For example, Chinese patent application CN106112985A (publication date: November 16, 2016) discloses a hybrid control scheme for an exoskeleton of a lower limb walking machine. This scheme uses plantar pressure information to determine whether the human body is in a support or swinging phase, and switches the corresponding control mode according to the different phases. In the support phase, high-impedance control is used for position tracking, while in the swinging phase, model reference adaptive control is used to adjust the exoskeleton movement according to the human's movement intention. This type of method can change the control state for different phases, but the acquisition and updating of control parameters mainly revolve around a predetermined controller.

[0004] For example, Chinese patent application CN118617388A (publication date: September 10, 2024) discloses a hierarchical control structure consisting of a host computer, a slave computer, and an exoskeleton execution end. The exoskeleton execution end is equipped with joint motors, torque sensors, and an inertial measurement unit (IMU). The slave computer calculates the control torque based on motor motion information, human posture information, and human-computer interaction forces. The host computer processes the desired gait trajectory and feedback information, and uses robust control to handle trajectory tracking errors and system uncertainties. It generates trajectory compensation based on human-computer interaction forces and corrects the desired trajectory based on the real-time human posture.

[0005] As the requirements for individual adaptability in lower limb exoskeletons increase, it becomes necessary to continuously adjust impedance and assist parameters based on the gait characteristics of different users and the actual human-computer interaction state. More complex online optimization algorithms typically involve significant computational loads. If such algorithms are directly deployed in the underlying controller responsible for motor closed-loop control, the algorithm calculations may consume underlying processing resources and affect the fixed control cycle. If parameter optimization is performed by the host computer, it further increases the processes of uploading status data, distributing optimized parameters, and communication between different computing nodes.

[0006] During collaborative operation between the upper and lower level machines, high-frequency status data and low-frequency control parameters have different data update rhythms. If communication reception, data parsing, or communication anomaly handling occupies the underlying processor for an extended period, it may also affect the continuous execution of the power control task. Furthermore, parameters obtained through online optimization may change significantly between adjacent update cycles, and communication anomalies may also generate anomalous parameters. If unconstrained parameters are directly used for the power actuator, control inputs that are incompatible with the exoskeleton's mechanical safety range may be generated.

[0007] Furthermore, the impact of failures in the host computer, the communication links between the upper and lower computers, and the motor communication links on the underlying power control is not the same. Using only a uniform communication anomaly handling method makes it difficult to balance the relationship between parameter optimization and power execution safety.

[0008] Therefore, how to improve the adaptability of lower limb exoskeleton parameters by using computationally intensive online parameter optimization methods, while maintaining the deterministic real-time operation of the underlying dynamic control and reducing the impact of communication processing, parameter mutations, and abnormalities of nodes at different levels on the safety control of the exoskeleton, has become an urgent technical problem to be solved. Summary of the Invention

[0009] The technical problem to be solved by this invention is to improve the adaptability of lower limb exoskeleton parameters by using online parameter optimization methods with high computational load, while maintaining the deterministic real-time operation of lower-level machine power control, and reducing the impact of communication processing, parameter mutations, and abnormalities of nodes at different levels on the safety control of the exoskeleton.

[0010] To achieve the above objectives, this invention proposes an exoskeleton adaptive control method based on a two-layer computing architecture and Bayesian optimization, comprising: a lower-level computer acquiring lower limb IMU gait data and motor feedback at a preset hard real-time cycle, and determining the gait phase through a finite state machine; uploading sensor data carrying gait phase tags to an upper-level computer via direct memory access (DMA) at a frequency higher than the frequency of the optimization parameters; the upper-level computer accumulating data from multiple gait cycles, to... Here are the parameters to be optimized, where, and For impedance parameters, The torque parameter is evaluated based on a weighted cost function consisting of joint trajectory jerk, expected torque, feedback torque error, and an asymmetric penalty term when the interaction force exceeds a safety threshold. The optimized parameter is obtained through an iterative optimization using a Gaussian process surrogate model and a confidence upper bound (UCB) acquisition function. The host computer sends the optimized parameter to the slave computer at a frequency lower than the sensor data upload frequency. The slave computer uses a single-byte non-blocking receive interrupt to buffer the parameter and set the receive flag. During the main loop idle period, it parses the parameter, sequentially performs preset physical safety boundary clamping and first-order low-pass filtering smoothing on the parsed parameter, and uses the processed parameter for motor control. In the slave computer's interrupt priority, motor controller local area network (CAN) transmission and reception and emergency stop tasks are prioritized over IMU data acquisition, and IMU data acquisition is prioritized over host computer parameter reception.

[0011] Preferably, the comfort index in the weighted cost function is The stability index is The security index is triggered when the interaction force exceeds the maximum security threshold. An exponentially increasing asymmetric penalty is applied; the comfort, stability, and safety indices are each assigned normalized weights. , and ;in, For a single gait cycle, For joint angle, For time, and These are the desired torque and the feedback torque, respectively.

[0012] Preferably, the Gaussian process surrogate model adopts Maternity. Kernel function, the kernel function is ,in, The distance in the parameter space to be optimized. For kernel function amplitude parameters, is the kernel function scaling parameter.

[0013] Preferably, the exploration factor of the UCB acquisition function With control cycle according to Attenuation, where, These are the exploration factors corresponding to the initial exploration phase. The exploration factor corresponding to the convergence phase. This is the decay rate constant.

[0014] Preferably, the lower-level device captures the overflow error (ORE) flag and frame error (FE) flag in the status register in the Universal Asynchronous Receiver (UART) error interrupt callback function. After detecting the ORE flag or FE flag, it reads the data register to clear the error flag, resets the read pointer and write pointer of the global receive ring buffer, and restarts the single-byte non-blocking receive interrupt.

[0015] Preferably, the lower-level machine sets a communication timeout timer for the upper-level machine. If a valid impedance parameter frame sent by the host computer is not received within a preset timeout threshold, the control strategy will be switched to zero torque follow mode or fixed high damping safety mode.

[0016] Preferably, the lower-level computer detects the motor feedback packet in the CAN receive first-in-first-out (FIFO) buffer in each hard real-time control cycle. If no motor feedback packet is read for three consecutive control cycles, an emergency stop is triggered, and a power-off command is issued and the motor drive bridge enable signal is cut off.

[0017] Preferably, the obtained from the analysis , and Each with a preset mechanical safety extreme value range In comparison, among which, To ensure a safe lower limit, The upper limit is set as the safety limit, and parameters exceeding the mechanical safety limit range are truncated to the corresponding safety limit.

[0018] Preferably, the parameters after being clamped by the physical security boundary are according to Perform a first-order low-pass filter, where, For the new parameters obtained in the current period, The parameters are from the previous period. These are the parameters actually used for motor control. These are the filter coefficients.

[0019] Preferably, the lower-level machine sets the asymmetric interrupt preemption priority through the nested vector interrupt controller (NVIC), wherein the preemption priority of the motor CAN bus transmit / receive interrupt and the system emergency stop timer interrupt is 0, the preemption priority of the IMU data acquisition interrupt is 1, and the preemption priority of the upper-level machine parameter receiving UART interrupt is 2.

[0020] Preferably, the preset hard real-time control period is 10 ms.

[0021] The above technical solution has the following advantages: This application employs a two-layer computing architecture consisting of a host computer and a slave computer, allowing the computationally intensive Bayesian parameter optimization and the underlying hard real-time motor control to be executed separately. This reduces the direct impact of online optimization calculations on the underlying control resources. The slave computer uploads data containing gait phase information at high frequency via DMA. The host computer uses data from multiple gait cycles, along with safety, stability, and comfort evaluations, to iteratively optimize impedance and torque parameters, ensuring that parameter adjustments correspond to actual human gait and human-machine interaction states. Optimized parameters are returned to the slave computer at low frequency and parsed during the main loop's idle phase, minimizing the blocking of real-time control tasks during parameter communication and parsing. Before being used for motor control, optimized parameters undergo physical safety boundary clamping and first-order low-pass filtering, limiting abnormal parameters exceeding mechanical safety limits and reducing abrupt changes caused by adjacent parameter updates. By assigning different interrupt priorities to motor CAN transceiver and emergency stop tasks, IMU acquisition tasks, and host computer parameter reception tasks, power closed-loop and safety-related tasks can be prioritized over ordinary parameter communication, enabling online parameter optimization, data communication, and underlying real-time control to operate collaboratively at different time scales. Attached Figure Description

[0022] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a schematic diagram of the two-layer computing architecture and data interaction relationship provided in the embodiments of this application. Detailed Implementation

[0023] The technical solution of this application will be further described below with reference to specific embodiments. The described embodiments are used to illustrate the technical concept and implementation process of this application, and are not intended to limit the scope of protection of this application. Without changing the technical concept of this application, those skilled in the art can adopt corresponding conventional implementation conditions according to the specific exoskeleton structure, sensor configuration and control object.

[0024] Example 1 like Figure 1 As shown, this embodiment provides an adaptive control method for exoskeletons based on a dual-layer computing architecture and Bayesian optimization. This method is applied to a powered lower limb exoskeleton, forming a dual-layer computing platform consisting of an upper-level computer and a lower-level computer. The upper-level computer handles multi-gait cycle data processing and Bayesian parameter optimization, while the lower-level computer handles gait information acquisition, gait phase determination, real-time motor control, parameter safety processing, and fault safety handling.

[0025] In one specific implementation, the host computer employs a computing device capable of performing multi-step periodic data processing and Bayesian parameter optimization, the slave computer uses a microcontroller, and the exoskeleton's power execution part uses a motor. This configuration illustrates one implementation of a two-layer computing platform. The control logic of this embodiment does not rely on directly deploying the complex Bayesian optimization process to the slave computer. Instead, it utilizes the host computer to handle the computationally intensive online optimization process, allowing the slave computer to retain the computing resources required for performing hard real-time control tasks.

[0026] The lower-level computer interacts with the lower limb IMU and exoskeleton motors. The IMU provides gait data related to the movement state of the human lower limbs, and the motors feed back their operating status to the lower-level computer. The lower-level computer and the motors exchange control and feedback information via a CAN bus. A data communication link is formed between the upper-level computer and the lower-level computer, which simultaneously handles the uplink of high-frequency status data and the downlink of low-frequency optimization parameters.

[0027] After system startup, the lower-level computer first enters a hard real-time control loop. In this embodiment, the hard real-time control cycle is 10ms. Upon entering each control cycle, the lower-level computer collects the current lower limb IMU gait data and motor feedback data, and inputs the collected results into a pre-configured finite state machine. The finite state machine performs gait state judgment based on the currently collected data and outputs the current gait phase of the human body. The gait phase includes at least the support phase and the swing phase.

[0028] After obtaining the gait phase, the lower-level computer associates the collected sensor data with the corresponding gait phase tags, enabling the upper-level computer to determine the gait stage corresponding to each set of sensor data during subsequent data processing. Thus, the data obtained within the same gait cycle reflects both the lower limb movement state and the motor operating state, and also possesses corresponding gait phase information.

[0029] For high-frequency status data sent from the lower-level device to the upper-level device, this embodiment uses DMA for transparent transmission. The lower-level device entrusts the sensor data and gait phase tags to be sent to the DMA transmission mechanism for processing, and the DMA completes the corresponding data transfer and transmission process. The lower-level device CPU does not need to participate in data transfer byte by byte during the entire uplink data transmission process, and can continue to execute the control tasks within the current hard real-time control cycle. Thus, the high-frequency sensor status feedback and the underlying hard real-time control loop are decoupled in terms of computing power consumption.

[0030] The host computer continuously receives data uploaded by the slave computer and organizes the data according to gait phase labels. After accumulating data from multiple gait cycles, the host computer evaluates the human-machine interaction status corresponding to the current impedance control parameters and assist torque parameters, and executes the next round of parameter optimization based on the evaluation results.

[0031] In this embodiment, the parameter to be optimized is represented as... .in, and The impedance parameters to be optimized are... These are the torque parameters to be optimized. Each set... This corresponds to the operational status of the exoskeleton under specific parameter conditions. The host computer uses operational data acquired over multiple gait cycles to calculate the evaluation result corresponding to this parameter combination.

[0032] To avoid using only a single trajectory error as the optimization objective, this embodiment simultaneously considers the safety, stability, and comfort during the interaction between the human body and the exoskeleton, and constructs a multi-dimensional weighted cost function. A single gait cycle is denoted as... The total cost function is expressed as:

[0033] in, For safety indicators, As a stability indicator, For comfort indicators. , and These are the corresponding normalized weight coefficients. Each weight is used to adjust the degree of influence of different evaluation indicators in the total cost.

[0034] Comfort indices are used to evaluate the smoothness of joint movement. This embodiment uses the acceleration of the joint trajectory as the evaluation basis. Let the joint angle be... , For time, the comfort index is expressed as:

[0035] When the joint trajectory undergoes a significant abrupt change in motion during the gait cycle, the third derivative of the joint angle increases accordingly. This increases accordingly. Therefore, this index can transform the smoothness of exoskeleton movement into an evaluation metric that can be used for parameter optimization.

[0036] Stability metrics are used to evaluate how closely the actual feedback torque of the motor approximates the desired torque. Let the desired torque corresponding to the upper-level control process be... The motor feedback torque is The stability index is then expressed as:

[0037] When the deviation between the motor feedback torque and the desired torque increases... The corresponding increase is made. By incorporating this index into the total cost function, the torque tracking status under the current impedance and torque parameter conditions can be evaluated during the optimization process.

[0038] Safety indicators The settings are based on the physical interaction boundaries between the human body and the exoskeleton. A maximum safety threshold is preset. When the interaction force exceeds When a state exceeds the safety boundary, an asymmetric penalty is imposed, and the corresponding cost increases exponentially with the degree of exceeding the boundary. Therefore, once the parameter combination causes the human-computer interaction state to fall outside the safety boundary, the total cost will increase accordingly, thus causing the subsequent optimization process to tend to avoid this parameter region.

[0039] The host computer, based on the obtained parameter samples and their corresponding... A Gaussian process surrogate model is constructed to fit the unknown mapping relationship between the parameters to be optimized and the human-computer interaction evaluation results. This embodiment uses Maternity... A kernel function describes the correlation between different parameter points in the parameter space. Its kernel function is:

[0040] in, This represents the distance between different combinations of parameters in the parameter space. For kernel function amplitude parameters, The kernel function scaling parameter is used. Using the evaluated parameter combinations and corresponding total costs, the Gaussian process surrogate model can be progressively updated to obtain the prediction mean and prediction uncertainty at different locations in the parameter space.

[0041] Based on the Gaussian process surrogate model, this embodiment employs the UCB strategy to select between the currently known optimal region and the parameter region that still has significant uncertainty. The next set of candidate parameters can be determined according to the following formula:

[0042] in, Indicates the next set of candidate parameters. This indicates the current proxy model's response to the parameters. The predicted mean corresponding to the objective function, This indicates the corresponding forecast uncertainty. An exploration factor to control the degree of exploration. This indicates the current control cycle.

[0043] To ensure sufficient parameter space exploration capability in the early stages of parameter optimization and to reduce parameter fluctuations caused by continuous large-scale exploration in the later stages of optimization, this embodiment uses an exploration factor... With control cycle Dynamic decay. The relationship is as follows:

[0044] in, Corresponding to the initial exploration level in the optimization process, The degree of exploration after the optimization gradually converges. is the decay rate constant. As... Increase The search progresses from a higher exploration level to a lower, more stable level. This allows for a wider search of the parameter space in the early stages of optimization, while mitigating the continuous changes caused by parameter searching in the later stages.

[0045] The host computer obtains new , and Subsequently, instead of directly participating in the hard real-time control of the underlying motor, the upper-level machine sends new parameters to the lower-level machine according to a parameter update rhythm lower than the state data upload frequency. The time required for the upper-level machine to perform Bayesian optimization is not used as a synchronization waiting condition for the lower-level machine's 10 ms hard real-time control cycle. While the upper-level machine has not yet generated new valid parameters, the lower-level machine continues to execute underlying control according to the currently valid control parameters.

[0046] For low-frequency parameters sent by the host computer, the slave computer uses a single-byte non-blocking receive interrupt to receive them. After receiving each byte, the UART stores it in the receive buffer and sets the corresponding receive flag. After completing the above operations, it immediately exits the receive interrupt and does not perform time-consuming parameter string parsing in the UART receive interrupt.

[0047] After completing the current hard real-time task and entering the idle processing phase, the lower-level main loop checks the receive flag to determine if there is any parameter data that has already been received. If there is data to be processed, parameter string parsing is then performed. In one specific implementation, the atof function can be used to convert character-form parameters to numeric-form parameters. Because the parsing operation is moved out of the parameter receive interrupt, parameter communication and parameter parsing will not block the underlying control cycle by occupying the UART interrupt for a long time.

[0048] The parameters obtained from the analysis are first subjected to physical boundary clamping before entering the motor control process. For , and The analytical results were compared with the preset mechanical safety extreme value range. In comparison, among which, Indicates the lower limit of safety. This indicates the upper safety limit. When a parameter is below the corresponding lower safety limit, it is truncated to the lower safety limit; when a parameter is above the corresponding upper safety limit, it is truncated to the upper safety limit; when a parameter is within the safe extreme range, its analytical result is retained. This process prevents parameters that clearly exceed the mechanical safety range from directly entering the execution layer when the algorithm output jumps or abnormal data occurs during communication.

[0049] After completing the physical boundary clamping, the lower-level machine continues to perform first-order low-pass filtering smoothing on the old and new parameters. For any parameter to be updated... The smoothing relationship is:

[0050] in, This represents the new parameters after safety boundary processing for the current period. This indicates the parameter in the previous parameter update process. This indicates the parameters actually used for control execution in this instance. These are the filter coefficients. Through this process, the parameters do not jump directly from the old values ​​to the new optimization results, but rather transition according to a first-order low-pass filter relationship.

[0051] After completing safety boundary clamping and first-order low-pass filtering, the lower-level computer sends the processed parameters to the exoskeleton motor according to the control communication protocol used by the motor. Subsequent hard real-time control cycles participate in motor control with the updated actual parameters. This forms a closed-loop update process from human gait state acquisition, parameter performance evaluation, upper-level computer Bayesian optimization, low-frequency parameter feedback, low-level safety processing to motor execution.

[0052] To further prevent communication anomalies between the upper and lower level machines from affecting the underlying control process, this embodiment incorporates an error interrupt self-recovery mechanism during UART communication. When a serial port overflow error (ORE) or frame error (FE) occurs, the lower level machine reads the corresponding status register and captures the ORE and FE flags during the serial port error interrupt callback process. Upon detecting the anomaly, the corresponding error status is cleared by reading the data register, and the read and write pointers of the global receive ring buffer are reset. After anomaly cleanup is complete, the single-byte non-blocking receive interrupt is restarted. This process ensures that after the abnormal communication frame is discarded, the UART receive function can re-enter the receive state, preventing communication anomalies from continuously occupying or locking the parameter receive link.

[0053] In case of malfunctions in the host computer itself, this embodiment maintains the host computer communication timeout timer in the slave computer. Whenever the lower-level machine receives a valid impedance parameter frame, it updates the corresponding communication status. If a valid impedance parameter frame from the upper-level machine is not received within a preset timeout threshold, it is determined that the upper-level machine can no longer provide valid parameters. In one specific implementation, this timeout threshold is set to 2 seconds. After this occurs, the lower-level machine no longer relies on the missing dynamic parameters from the upper-level machine, but instead downgrades its control strategy to a zero-torque follower mode or a fixed high-damping safety mode.

[0054] For CAN communication between the lower-level machine and the motor, this embodiment employs a different safety handling method than that for upper-level machine disconnection. Every 10 ms hard real-time control cycle, the lower-level machine checks the motor feedback in the CAN receive FIFO. If no motor feedback packet is read for three consecutive control cycles, the corresponding time is 30 ms. At this point, the system determines this state as an actuator bus hardware failure and triggers an emergency stop. The lower-level machine sends a power-off control command to the actuator and simultaneously cuts off the motor drive bridge enable signal, causing the power actuator to exit normal drive mode.

[0055] Different safety procedures are used for host computer disconnection and motor CAN disconnection because they correspond to different fault levels. The host computer is responsible for parameter optimization; if it cannot provide new parameters for a short period, the slave computer still retains its real-time control capabilities and can therefore enter a preset safety degradation state. If the motor CAN fails to return feedback information for an extended period, the slave computer can no longer reliably obtain the operating status of the power actuator, therefore an emergency stop and drive enable shutdown are implemented.

[0056] This embodiment also sets the interrupt preemption priority for different tasks through the NVIC of the lower-level machine. The motor CAN bus transmit / receive interrupt and the system emergency stop timer interrupt, which are directly related to the power closed loop, are set to preemption priority 0. The IMU data acquisition interrupt used for gait motion state perception is set to preemption priority 1. The UART interrupt used to receive optimization parameters from the upper-level machine is set to preemption priority 2.

[0057] Based on the above priority settings, when host computer parameter communication occurs simultaneously with motor control, emergency stop, or IMU data acquisition, the task with higher safety criticality can be executed first. Host computer parameter communication can be preempted by higher priority tasks. Combining DMA high-frequency upload, UART single-byte fast reception, and idle-period parameter parsing, status communication and parameter communication can minimize the impact on the underlying hard real-time control cycle.

[0058] During normal operation, the lower-level machine continuously executes a 10 ms control cycle and sends high-frequency data with gait phase tags to the upper-level machine. The upper-level machine uses multiple gait cycles as evaluation units, updates the Gaussian process proxy model based on safety, stability, and comfort indicators under the current parameters, and then uses the UCB strategy to generate the next set of parameters to be verified. The new parameters are sent to the lower-level machine at a low frequency, and after asynchronous parsing, physical boundary clamping, and first-order low-pass filtering, they enter the execution layer. The new operational data is returned to the upper-level machine again to form the next round of evaluation data, thus performing iterative updates.

[0059] When using the above operating mode, the large amount of computation required for Bayesian optimization is mainly handled by the host computer, and the slave computer does not need to complete the Gaussian process model update and acquisition function solution in every hard real-time control cycle. On the other hand, the parameters generated by the host computer do not directly bypass the slave computer and act on the motor, but are received, parsed, safety boundary checked and parameter smoothed at the lower level before being put into execution.

[0060] High-frequency state streams, low-frequency parameter streams, and low-level power control are handled by different processing mechanisms. DMA is used to reduce the continuous CPU usage of high-frequency state uploads on the lower-level machine; single-byte interrupts and idle period parsing are used to reduce the blocking of low-frequency parameter communication on real-time control; and interrupt priority is used to prioritize power closed-loop and emergency stop tasks over ordinary parameter communication. These processes work together to enable the upper-level machine parameter optimization process to run in parallel with the low-level hard real-time control process.

[0061] Meanwhile, the multi-dimensional cost function incorporates safety, stability, and comfort into the same parameter evaluation process. Dynamic exploration factors allow parameter search to explore at different levels in different optimization stages. Physical boundary constraints limit abnormal parameters from entering the mechanical execution range, and first-order low-pass filtering reduces cross-cycle parameter mutations. Communication self-recovery, host computer disconnection degradation, and CAN disconnection emergency stop execute corresponding safety measures for different levels of anomalies, thus forming a hierarchical control process from algorithm output to power execution.

[0062] Example 2 This embodiment further explains the multi-objective Bayesian parameter optimization process executed by the host computer, based on Embodiment 1. Unless otherwise specified, the underlying data acquisition, gait phase determination, high-frequency data upload, parameter distribution, and motor execution processes are the same as in Embodiment 1.

[0063] The host computer uses data from multiple complete gait cycles as the data basis for a single parameter evaluation. The slave computer collects lower limb IMU gait data and motor feedback within each hard real-time control cycle, establishes a correspondence between the current gait phase and the collected data, and then uploads the data. The host computer organizes the received data according to gait phase and gait cycle, ensuring that the same set of parameters to be optimized... It can correspond to the human-computer interaction data generated during the actual operation of this parameter.

[0064] For the current parameter This embodiment generates evaluation results from three dimensions: safety, stability, and comfort, and combines the three types of evaluation indicators into a total cost function:

[0065] in, , and These are safety indicators, stability indicators, and comfort indicators. , and These are the normalized weighting coefficients corresponding to the safety, stability, and comfort indicators, respectively. By combining evaluation indicators of different dimensions using these weighting coefficients, each set of parameters to be optimized can form a corresponding comprehensive cost.

[0066] For comfort evaluation, this embodiment determines the comfort index based on the acceleration of the joint trajectory within the gait cycle. Let a single gait cycle be... The joint angle is , If the time factor is used, then the comfort index is:

[0067] When the joint trajectory changes rapidly over time The amplitude increases accordingly, formed by its square integral. This increases accordingly. Therefore, the degree of abrupt changes in the motion trajectory can be transformed into an evaluation quantity in the parameter optimization process, allowing parameter evaluation to consider not only whether motion control can be achieved, but also the smoothness of the motion process.

[0068] For stability evaluation, this embodiment compares the difference between the desired torque and the motor feedback torque. Let the desired torque be... The feedback torque is The stability index is:

[0069] Under the current parameter combination, if the motor feedback torque can follow the desired torque well, the error between the two is small, and the resulting stability cost is correspondingly small. Conversely, when the actual output torque deviates significantly from the desired torque, the stability cost increases accordingly.

[0070] For security evaluation, this embodiment sets a maximum security threshold corresponding to human-computer physical interaction. When the current interaction state does not exceed the maximum security threshold, a corresponding security evaluation result is generated based on the current state. When the interaction force exceeds... At that time, the security index is increased using an asymmetric penalty method, and the cost of exceeding the security boundary increases exponentially.

[0071] After adopting the above safety evaluation method, even if a certain parameter combination has good evaluation results in terms of trajectory smoothness or torque tracking, as long as the human-computer interaction state it causes exceeds the preset safety boundary, the overall cost of the parameter combination will still be significantly affected by the safety penalty. Therefore, safety constraints can directly participate in the parameter selection process of Bayesian optimization, rather than only making a post-hoc judgment on the results after the parameter optimization is completed.

[0072] The host computer will evaluate the parameters. and the corresponding total cost Known samples, used as surrogate models of Gaussian processes, are employed to progressively fit the relationship between the parameter space to be optimized and the objective function of human-computer interaction. This embodiment uses the Matern process... Kernel function:

[0073] in, This represents the distance between two parameter points in the parameter space. For kernel function amplitude parameters, The kernel function scaling parameters are used. After each set of parameters has been tested and evaluated, the new parameters and their corresponding total cost are added to the existing samples to update the Gaussian process surrogate model.

[0074] By using the updated Gaussian process surrogate model, the predicted mean of the candidate parameters in the parameter space can be obtained. and forecast uncertainty This embodiment determines the next set of parameters according to the UCB strategy:

[0075] in, Indicates the next set of candidate parameters. An exploration factor used to adjust the degree of exploration. This indicates the current control period. The predicted mean reflects the surrogate model's current estimate of the candidate parameter's performance based on existing samples, while the predicted uncertainty reflects the degree to which the region containing the candidate parameter has not been fully evaluated. By having both participate in the candidate parameter selection, dynamic selection can be made between parameter regions that have received relatively good evaluations and parameter regions that are yet to be explored.

[0076] This embodiment further enhances the exploration factor. With control cycle The exploration factor changes, rather than remaining constant throughout the optimization process. The exploration factor decays according to the following formula:

[0077] in, This is to ensure that there is sufficient exploratory capability in the early stages of optimization. This is used to maintain relatively stable parameter output after the optimization gradually converges. This is the decay rate constant.

[0078] In the initial stage of parameter optimization Relatively large, with high forecast uncertainty. It plays a relatively strong role in candidate parameter selection, thus enabling the exploration of parameter regions that have not yet been fully evaluated. With the control cycle... Increase, Gradually towards As things change, the extent to which large-scale exploration of unknown areas is gradually reduced, and parameter selection focuses more on areas that have already demonstrated a better overall cost.

[0079] Therefore, this embodiment does not employ a fixed search intensity throughout the entire rehabilitation process, but rather allows the parameter search process to vary with each iteration. This preserves the exploratory capabilities needed to find suitable parameter combinations in the early stages, while reducing the impact of continuous, large-scale parameter changes on the exoskeleton's dynamic output in the later stages.

[0080] Each new set Then, the host computer will send the corresponding , as well as The new parameters to be executed are sent to the lower-level machine at low frequency. The lower-level machine processes the parameters according to the non-blocking reception, idle period parsing, safety boundary clamping, and first-order low-pass smoothing process described in Embodiment 1. Only after processing are the parameters actually entered into the motor control process.

[0081] Once the new parameters are implemented, the lower-level computer continues to collect IMU data, motor feedback, and other data used to evaluate the current parameters within the corresponding gait cycle, and uploads it to the upper-level computer. The upper-level computer then recalculates. , and ,renew The Gaussian process surrogate model is then updated using the new samples, and the next set of parameters is selected. This forms an iterative process between parameter execution, state acquisition, performance evaluation, surrogate model update, and next parameter selection.

[0082] Example 3 This embodiment, based on Embodiment 1, further explains the communication anomaly recovery and node fault safety control in a two-layer computing architecture. This embodiment focuses on handling three different fault states: communication anomalies between the host computer and the slave computer, host computer disconnection, and CAN communication disconnection between the slave computer and the motor.

[0083] There is a high-frequency status data upload and a low-frequency parameter data download between the host computer and the slave computer. Since the data communication process is continuously accompanied by the underlying hardware real-time control operation, communication failures cannot occupy or lock the slave computer's main loop for an extended period of time.

[0084] During UART reception, if the data transmission rate of the host computer or the reception processing status prevents the slave computer's serial port receive register from being read in time, a serial port overflow error (ORE) may occur. Additionally, a frame error (FE) may also be detected during communication. For these errors, this embodiment performs exception recovery during the UART error interrupt callback process.

[0085] When the UART enters an error interrupt, the lower-level device reads the serial port status register and checks the ORE and FE flags. If ORE or FE is detected, the corresponding data register is read to clear the error status, and the global circular buffer used to receive parameter data is reset.

[0086] For the global circular buffer, the lower-level device resets its read and write pointers, thereby discarding the received data that has lost its integrity during this abnormal communication. After completing the error state clearing and buffer reset, the lower-level device restarts the single-byte non-blocking receive interrupt, causing the UART to re-enter the receive state.

[0087] The aforementioned exception recovery process only performs the operations required to restore communication within the interrupt service routine, without continuing parameter string parsing, control parameter calculation, or Bayesian optimization operations during the error interrupt. In one specific implementation, the time taken for this exception recovery process is less than [amount missing]. Therefore, while discarding abnormal data, the communication link can be restored, without allowing communication errors to occupy the main control process for an extended period.

[0088] In cases where the host computer itself stops working properly or the communication link between the host and slave computers is interrupted, this embodiment sets a function in the slave computer. A timer is used to determine whether the host computer is continuously providing valid impedance parameter frames.

[0089] During normal operation, once the lower-level machine receives a complete and valid parameter frame from the upper-level machine, it updates the corresponding communication status. As long as valid parameter frames are continuously obtained within the preset timeout threshold, the lower-level machine parses and executes the optimization parameters provided by the upper-level machine according to the normal parameter update process.

[0090] when If a preset timeout threshold is exceeded and no new valid impedance parameter frame is received, the lower-level machine determines that the upper-level machine can no longer provide parameters normally. In one specific implementation, the timeout threshold is set to 2 seconds.

[0091] In the aforementioned host computer disconnection state, the slave computer does not directly shut down its own control tasks, but instead switches from the normal online optimization control state to a preset safety degradation state. This safety degradation state can be either a zero-torque follow mode or a fixed high-damping safety mode.

[0092] When using the zero-torque follower mode, the lower-level computer no longer relies on the disconnected upper-level computer to continue providing dynamic impedance parameters, allowing the power output to enter the corresponding safe follower state. When using the fixed high-damping safety mode, the lower-level computer switches to a pre-determined fixed safety control state and no longer executes upper-level computer optimization parameters that are incomplete or cannot be updated.

[0093] Therefore, in this embodiment, the host computer performs parameter optimization but does not perform the irreplaceable and unique control function required for maintaining the underlying security state. Even if the host computer stops operating normally, the slave computer can still autonomously enter a preset security degradation state based on the communication timeout result.

[0094] For CAN bus failures between the lower-level machine and the motor, this embodiment employs a more direct and safer handling method than upper-level machine disconnection. The lower-level machine checks the motor feedback status in the CAN receive FIFO every 10 ms hard real-time control cycle.

[0095] If the motor feedback packet can be read normally in the current control cycle, the normal control process continues. If no motor feedback packet is read for three consecutive control cycles, corresponding to 30 ms without motor execution status feedback, the lower-level computer determines that there is a hardware fault in the CAN bus or power execution link.

[0096] Upon determining that the actuator is disconnected, the lower-level computer immediately triggers an emergency stop interrupt, sends a power-off command to the motor actuator, and disconnects the motor drive bridge enable signal. After completing the above processing, it stops driving the corresponding motor according to the normal Bayesian optimization parameters.

[0097] In this embodiment, the same fault handling methods are used for the host computer disconnection and the CAN actuator disconnection. When the host computer disconnects, the slave computer can still execute its own low-level control, thus entering zero torque following mode or fixed high damping safety mode. When the CAN continuously loses motor feedback, the slave computer cannot reliably obtain the current power status of the actuator, thus directly triggering an emergency stop.

[0098] This layered fault handling allows communication anomalies, computing node anomalies, and execution node anomalies to be handled according to their respective impacts on the power closed loop. When the UART itself experiences an ORE or FE (Extended Response) event, communication is restored first. When the host computer can no longer output valid parameters, a safety degradation is performed. When the motor feedback link fails, an emergency stop is performed. This avoids a single fault handling logic covering abnormal states of different safety levels.

[0099] Example 4 Based on Example 1, this embodiment further explains the safety handling before the upper computer's optimization parameters enter the exoskeleton's power execution layer and the lower computer's interrupt scheduling method.

[0100] The parameters output by the host computer include , and Since these parameters originate from a continuously iterative Bayesian optimization process, they may change relative to the previous set of parameters. Furthermore, when data anomalies occur during communication between the lower and upper control machines, the parsed results may deviate from the allowable mechanical operating range. Therefore, the lower control machine does not directly send the parsed parameters to the motor, but instead sequentially performs physical boundary clamping and parameter smoothing.

[0101] For any analytical parameter, the lower-level machine pre-stores the corresponding mechanical safety extreme value range. The mechanical safety extreme value range is represented as follows: ,in, Indicates the lower limit of safety. This indicates the safety upper limit. The lower-level machine respectively... , and Perform a safety boundary comparison.

[0102] When the parsed parameter is less than When that happens, adjust the parameter to the corresponding value. When the parsed parameters are greater than When that happens, adjust the parameter to the corresponding value. When the parsed parameter is within the safe extreme value range, the parameter value is retained for subsequent processing.

[0103] Physical boundary clamping primarily addresses whether parameters fall within the mechanically permissible range. For parameters that are already within the safe range but have changed significantly compared to the previous cycle, this embodiment further employs a first-order low-pass filter for discontinuous smoothing.

[0104] Let the new parameters obtained through the current optimization and physical boundary constraints be: The parameters actually used in the previous parameter update phase were: After smoothing, the actual parameters output to the execution layer are: ,but:

[0105] in, These are the filter coefficients.

[0106] for , and All parameters can be updated according to the above relationship. Therefore, the new parameters output by the host computer are first constrained by mechanical safety boundaries, and then smoothed according to the relationship between the old and new parameters, rather than directly replacing the previous set of execution parameters in a step manner.

[0107] Physical boundary clamping and first-order low-pass filtering serve different functions. Physical boundary clamping is used to limit abnormal parameters that significantly exceed the preset mechanical safety boundary, while first-order low-pass filtering is used to suppress cross-cycle parameter abrupt changes that may still exist within the safety range. Both are processed in the order of clamping first, then smoothing, and then the lower-level computer processes the data. Used for corresponding motor control.

[0108] To ensure that the hard real-time tasks are not preempted by low-security-level communication processing while the above parameter communication, status acquisition and power control are running simultaneously, this embodiment uses the NVIC of the lower-level machine to set the asymmetric interrupt preemption priority.

[0109] The motor CAN bus transmit / receive interrupt, which is directly related to the power closed loop, is set to preemptive priority 0. The system emergency stop timer interrupt is also set to preemptive priority 0. Both are at the highest preemptive level in this embodiment.

[0110] The IMU data acquisition interrupt is set to preemption priority 1. IMU data is used for lower limb motion state acquisition and gait phase determination, so its priority is lower than tasks directly related to dynamic closed loop and emergency stop, but higher than host computer parameter communication tasks.

[0111] The host computer's UART parameter reception interrupt is set to preemptive priority 2. This interrupt is only responsible for receiving and buffering low-frequency optimization parameters and does not directly complete the power closed-loop control. When a higher-priority CAN transceiver, emergency stop, or IMU data acquisition task occurs, the host computer's UART parameter reception is allowed to be preempted by the corresponding task.

[0112] Through the above priority configuration, the underlying control process forms a scheduling relationship in which the power closed-loop task takes priority, motion state perception takes second, and upper computer parameter communication is subordinate.

[0113] This scheduling relationship also works in conjunction with the communication decoupling method in Example 1. When the lower-level machine uploads high-frequency gait data to the upper-level machine, DMA is used to reduce the CPU's data handling burden. When the lower-level machine receives low-frequency optimization parameters from the upper-level machine, it only performs single-byte buffering and flag setting in the UART interrupt, and parameter string parsing is performed during the main loop's idle phase.

[0114] Therefore, the parameter receiving task itself is at a lower interrupt priority, and the time-consuming parsing process in the parameters is not executed within the receive interrupt. Even if parameter communication continues between the host computer and the slave computer, the CAN power control and emergency stop tasks can still be executed with higher priority.

[0115] For emergency stop events, as long as the corresponding emergency stop conditions are met, the emergency stop timer interrupt can be executed at preemption priority 0. For normal power feedback, the CAN transmit / receive interrupt is also at the highest preemption level. Therefore, the host computer parameter optimization and parameter communication will not gain higher execution priority than the underlying power safety tasks due to the calculation or communication tasks themselves.

[0116] Example 5 This embodiment illustrates the parameter update process of the above adaptive control method for different users or different stages of rehabilitation. This embodiment does not change the two-layer computing architecture and control logic described in Embodiments 1 to 4, but instead uses online collected data to re-evaluate the impedance and torque parameters under the current usage state.

[0117] In actual rehabilitation training, the gait and human-machine interaction of dynamic lower limb exoskeletons may vary among different users. The movement patterns of the same user may also change at different rehabilitation stages. If fixed impedance and torque parameters are consistently used, a mismatch may occur between the current parameters and the actual movement patterns of the human body.

[0118] In this embodiment, after the corresponding training process begins, the lower-level computer still collects IMU gait data and motor feedback according to a 10 ms hard real-time cycle, identifies the support phase and swing phase through a finite state machine, and continuously uploads the data with gait phase tags to the upper-level computer through DMA.

[0119] The host computer does not directly use the evaluation results formed in other operating states as the sole control basis for the current state. Instead, it accumulates data from multiple gait cycles in the current operating state and uses the current parameters... The corresponding safety, stability, and comfort indicators reconstruct the total cost. .

[0120] As the actual operational data under the current usage state gradually increases, the Gaussian process surrogate model continuously updates its evaluation samples using new parameters. The UCB data acquisition process then selects the next set of samples based on the current surrogate model. Therefore, parameter optimization is based on data generated from the actual interaction between the exoskeleton and the human body, rather than simply using pre-fixed uniform parameters.

[0121] At the start of a new parameter optimization phase, a larger This is used to maintain the ability to explore the parameter space. As the optimization cycle increases, according to Gradual decay allows parameter updates to transition from early exploration to stable output in the later stages.

[0122] When the host computer receives a new combination of parameters, it still sends it to the slave computer as the optimization result. The slave computer first confirms the parameter reception status, and then parses the parameters during the idle phase of the main loop. The parsed result is then subjected to mechanical safety boundary clamping and first-order low-pass filtering before being used as the actual execution parameters in subsequent control cycles.

[0123] Therefore, even when personalized parameter searches are performed for different users or different stages of rehabilitation, the parameter search process cannot bypass the underlying security constraints of the lower-level machine. Bayesian optimization is responsible for finding a parameter range more suitable for the current interaction state based on actual operating data, while the lower-level machine is continuously responsible for hard real-time control, communication decoupling, parameter boundary constraints, and fault safety handling.

[0124] After several gait cycles of the current parameters, the host computer continues to obtain corresponding safety, stability, and comfort evaluation data. The new evaluation results are then used again to update the Gaussian process surrogate model and select the next set of parameters, thus forming a continuous parameter adaptive process.

[0125] The above process assigns parameter adaptation, algorithm computation, and dynamic safety to corresponding levels. Parameter adaptation is based on actual data generated by the human body and exoskeleton. The host computer handles the computationally complex Bayesian optimization, while the slave computer maintains deterministic control timing and parameter safety boundaries. Each part operates collaboratively according to different time scales, thus avoiding the need to directly place complex algorithms into the underlying hard real-time control loop to achieve personalized parameter searches.

[0126] The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit the scope of protection of this application. For those skilled in the art, without departing from the technical concept of this application, the specific implementation of the above embodiments can be adjusted according to the specific exoskeleton structure, the controlled object, and the actual operating conditions. Equivalent substitutions or conventional modifications that do not change the technical relationships such as dual-layer computing power collaboration, upper computer parameter optimization, low-level hard real-time control, parameter security processing, and abnormal state security handling can all be accomplished using the usual implementation methods in the art.

[0127] The technical features in the various embodiments of this application can be combined with each other where there is no technical conflict. Specific descriptions of the same technical step, parameter processing procedure, or control logic in different embodiments can be considered as further implementations of the corresponding technical solutions. For the optional control methods explicitly described in this application, implementation can be selected according to specific application conditions, but the data flow, control relationship, and security constraint relationship between the related technical features should not be changed.

[0128] For general power supply, conventional communication connections, data storage, program invocation, and mechanical installation methods not specifically described in this application, conventional technical means that can be directly implemented by those skilled in the art can be adopted. The above-mentioned conventional implementation conditions are only used to ensure that the technical solution of this application can operate normally and do not constitute a further limitation on the core technical solution of this application.

[0129] The parameter optimization process in this application uses actual operating data as the evaluation basis. The host computer is used to perform computationally intensive parameter optimization, while the slave computer is used to maintain the hard real-time control cycle and handle parameter safety. After the optimized parameters are generated by the host computer, they do not directly bypass the slave computer and act on the power actuator. Instead, they are received, parsed, constrained by safety boundaries, and smoothed before being used for motor control. Host computer anomalies, communication anomalies, and actuator bus anomalies are handled according to the corresponding level of safety policies. The above technical relationships together constitute a complete exoskeleton adaptive control process in the foregoing embodiments.

Claims

1. An adaptive control method for exoskeletons based on a two-layer computing architecture and Bayesian optimization, characterized in that, include: The lower-level computer collects lower limb IMU gait data and motor feedback according to a preset hard real-time cycle, and determines the gait phase through a finite state machine; Sensor data carrying gait phase tags is uploaded to the host computer at a frequency higher than the frequency of the optimization parameters via DMA. The host computer accumulates data from multiple gait cycles, in order to Here are the parameters to be optimized, where, and For impedance parameters, The torque parameter is evaluated based on a weighted cost function consisting of joint trajectory jerk, expected torque, feedback torque error, and asymmetric penalty term when the interaction force exceeds the safety threshold. The parameters after iterative optimization are obtained through Gaussian process surrogate model and UCB acquisition function. The host computer sends the optimized parameters to the slave computer at a frequency lower than the sensor data upload frequency. The lower-level machine uses a single-byte non-blocking receive interrupt buffer parameter and sets the receive flag. During the idle period of the main loop, it parses the parameter, performs preset physical safety boundary clamping and first-order low-pass filtering smoothing on the parsed parameter in sequence, and uses the processed parameter for motor control. In the interrupt priority of the lower-level machine, the motor CAN transceiver and emergency stop task takes precedence over IMU data acquisition, and IMU data acquisition takes precedence over upper-level machine parameter reception.

2. The exoskeleton adaptive control method according to claim 1, characterized in that, The comfort index in the weighted cost function is: The stability index is The security index is triggered when the interaction force exceeds the maximum security threshold. An exponentially increasing asymmetric penalty is applied; the comfort, stability, and safety indices are each assigned normalized weights. , and ;in, For a single gait cycle, For joint angle, For time, and These are the desired torque and the feedback torque, respectively.

3. The exoskeleton adaptive control method according to claim 2, characterized in that, The Gaussian process proxy model uses Matern Kernel function, the kernel function is: in, The distance in the parameter space to be optimized. For kernel function amplitude parameters, is the kernel function scaling parameter.

4. The exoskeleton adaptive control method according to claim 3, characterized in that, The exploration factor of the UCB acquisition function With control cycle according to Attenuation, where, These are the exploration factors corresponding to the initial exploration phase. The exploration factor corresponding to the convergence phase. This is the decay rate constant.

5. The exoskeleton adaptive control method according to claim 1, characterized in that, In the UART error interrupt callback function, the lower-level device captures the ORE flag and FE flag in the status register. After detecting the ORE flag or FE flag, it reads the data register to clear the error flag, resets the read pointer and write pointer of the global receive ring buffer, and restarts the single-byte non-blocking receive interrupt.

6. The exoskeleton adaptive control method according to claim 1, characterized in that, Lower-level machine sets the upper-level machine communication timeout timer If a valid impedance parameter frame sent by the host computer is not received within a preset timeout threshold, the control strategy will be switched to zero torque follow mode or fixed high damping safety mode.

7. The exoskeleton adaptive control method according to claim 1, characterized in that, The lower-level computer detects the motor feedback packet in the CAN receive FIFO in each hard real-time control cycle. If no motor feedback packet is read for three consecutive control cycles, it triggers an emergency stop, issues a power-off command, and cuts off the motor drive bridge enable signal.

8. The exoskeleton adaptive control method according to claim 1, characterized in that, The parsed , and Each with a preset mechanical safety extreme value range In comparison, among which, To ensure a safe lower limit, The upper limit is set as the safety limit, and parameters exceeding the mechanical safety limit range are truncated to the corresponding safety limit.

9. The exoskeleton adaptive control method according to claim 8, characterized in that, Parameters after being clamped by physical security boundaries according to Perform a first-order low-pass filter, where, For the new parameters obtained in the current period, The parameters are from the previous period. These are the parameters actually used for motor control. These are the filter coefficients.

10. The exoskeleton adaptive control method according to claim 1, characterized in that, The lower-level computer sets the asymmetric interrupt preemption priority through NVIC. The preemption priority of the motor CAN bus transmit / receive interrupt and the system emergency stop timer interrupt is 0, the preemption priority of the IMU data acquisition interrupt is 1, and the preemption priority of the upper-level computer parameter reception UART interrupt is 2.

Citation Information

Patent Citations

  • Exoskeleton mixed control system and method for lower limb walking assisting machine

    CN106112985A

  • Lower limb rehabilitation exoskeleton robot control method and system, electronic equipment and medium

    CN118617388A