Control method and device for wearable ankle exoskeleton and wearable ankle exoskeleton

Abstract

The present application relates to a control method and device for a wearable ankle exoskeleton and the wearable ankle exoskeleton, the control method comprising the following steps: calculating an estimated gait phase according to a step length estimation value and a timestamp of a heel strike event; generating a desired angle, a desired angular velocity and a desired angular acceleration of an ankle joint according to the step length estimation value and the estimated gait phase; calculating an angle tracking error and an angular velocity tracking error; obtaining a continuous impedance torque at a current sampling time according to the angle tracking error, the angular velocity tracking error and the desired angular acceleration, generating a pulse torque according to a pulse amplitude, a pulse start phase and a pulse width; synthesizing the continuous impedance torque and the pulse torque to obtain a desired assist torque; performing amplitude constraint and torque change rate constraint on the desired assist torque to obtain an actual assist torque. The present application can solve the problems of sharp peaks that cannot be executed by actuators, large torque control errors and poor phase adaptability.

Description

Technical Field

[0001] This invention belongs to the field of wearable ankle exoskeleton technology, and particularly relates to a control method, device and wearable ankle exoskeleton for wearable ankle exoskeleton. Background Technology

[0002] Wearable ankle assist devices are wearable devices that combine biomechanics and intelligent control technology, designed to reduce walking energy consumption, improve athletic performance, or aid in rehabilitation by assisting ankle movement. Existing wearable ankle assist devices primarily apply plantar flexion assistance to the ankle joint through motors, reels, or cable drives. The controller of a wearable ankle assist device typically generates preset phase-correlated torque, force, or displacement curves based on plantar pressure, inertial measurement units, or encoders to detect gait events, and these curves are tracked by the underlying motor control loop.

[0003] For example, the invention patent with application number CN202610433979.4 discloses "a control method for an ankle joint exoskeleton robot". This control method collects the error between the actual and expected angles of the joint to construct performance indicators; then, it calculates the gradient of the indicators with respect to the velocity factor based on the characteristics of the autocoupling PID controller; then, it inputs the gradient into the Adam optimizer to calculate the velocity factor update online; finally, it substitutes the gradient into the control law to obtain the output torque of the flexible exoskeleton.

[0004] While the above method can adaptively tune control parameters online and efficiently overcome nonlinear time-varying disturbances, it has the following problems: Although it can reduce the maximum dynamic peak deviation by about 56.4% under abrupt disturbances by utilizing the first and second moment estimation characteristics of the Adam optimizer, it may still generate an excessively large torque change rate by relying solely on the first and second moment estimation characteristics of the Adam optimizer, resulting in spikes or abrupt changes that the actuator cannot achieve.

[0005] Its torque control error and phase adaptability still need to be improved. Summary of the Invention

[0006] To address the problems of actuator failure, large torque control errors, and poor phase adaptability in existing technologies, this invention provides a control method, device, and wearable ankle exoskeleton for wearable ankle joints.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a control method for a wearable ankle exoskeleton, comprising the following steps: Based on stride duration estimates The most recent heel-to-toe incident The estimated gait phase at the current sampling time t is calculated. ; Based on the stride duration estimate Estimating gait phase Generate the desired angle of the ankle joint. Desired angular velocity and desired angular acceleration ; Obtain the real-time angle of the ankle joint. and real-time angular velocity And according to the desired angle and real-time angle Calculate angle tracking error According to the desired angular velocity and real-time angular velocity Calculate angular velocity tracking error ; The angle tracking error Angular velocity tracking error and desired angular acceleration By inputting the continuous impedance branch model, the continuous impedance torque at the current sampling time is obtained. The continuous impedance branch model is as follows: , In the formula, Equivalent stiffness gain, in units of ; Equivalent damping gain, in units of ; This is the inertial feedforward coefficient, in units of... Continuous impedance torque at the current sampling time The unit is Angle tracking error The unit is Angular velocity tracking error The unit is Desired angular acceleration The unit is ; Based on the estimated gait phase Pulse start phase and pulse width Determine whether the current sampling time is within the target assist window. If the current sampling time is within the target assist window, determine the pulse amplitude. Pulse start phase and pulse width Generate pulse torque ; The continuous impedance torque and the pulse torque Synthesized to obtain the desired assist torque ; For the desired assist torque By applying amplitude constraints and torque change rate constraints, the actual assist torque is obtained. .

[0008] A second aspect of the present invention provides a control device for a wearable ankle exoskeleton, comprising a memory and a controller connected in sequence, wherein the memory stores a computer program, and the controller is used to read the computer program and execute the control method for a wearable ankle exoskeleton described in the first aspect.

[0009] A third aspect of the present invention provides a wearable ankle exoskeleton, comprising a plantar pressure sensor, an inertial measurement unit, a lower leg wearable assembly, an ankle wearable assembly, an actuator assembly, a flexible transmission cable for transmitting the driving force generated by the actuator assembly to the lower leg wearable assembly and the ankle wearable assembly, and a controller. The plantar pressure sensor, the inertial measurement unit, and the actuator assembly are all signal-connected to the controller, which is a control device for a wearable ankle exoskeleton as described in the second aspect.

[0010] The beneficial effects of this invention are: This invention generates the desired assist torque based on the continuous impedance torque and pulse torque output by the continuous impedance branch model. It retains the full-cycle error adjustment and damping, and concentrates the positive power output during the target push-push phase, thereby improving the torque control accuracy and phase adaptability.

[0011] The pulse torque of this invention is only activated within the target assist window. When the estimated gait phase is outside this range, the pulse branch output is zero, thereby avoiding local assist from exceeding the target gait phase.

[0012] This invention improves the desired assist torque. After applying amplitude and torque change rate constraints, the output is finally satisfied with the amplitude and change rate boundaries, thus avoiding excessive torque change rate and spikes or abrupt changes that the actuator cannot achieve. Attached Figure Description

[0013] 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.

[0014] Figure 1 This is a schematic diagram illustrating the use of the wearable ankle exoskeleton of the present invention; Figure 2 The image shows a comparison between the ankle joint reference trajectory and the actual EC-TD3 trajectory under a representative parameter vector using the control method of this scheme. Figure 3 An ankle joint angle tracking error diagram under a representative parameter vector; Figure 4 A comparison chart of the expected assist torque, the torque after amplitude constraint, and the actual assist torque; Figure 5 This is a graph showing the actual rate of change response of the assist torque. Figure 6 A comparison chart of peak torque change rates under no execution constraints, only magnitude constraints, and combined magnitude-rate-of-change constraints; Figure 7 A comparison chart of constraint violation counts under conditions of no execution constraints, only magnitude constraints, and combined magnitude-rate of change constraints; Figure 8 Figures showing the structural ablation results for unassisted, continuous impedance branch, pulse branch, fixed parameter hybrid control and EC-TD3 hybrid control. Figure 9 A comparison chart of the ankle joint reference trajectory under the actual gait reference trajectory and the actual EC-TD3 trajectory; Figure 10 This is a graph showing the ankle joint angle tracking error under the actual gait reference trajectory.

[0015] The attached figure labels are named as follows: 1. User; 2. Controller; 3. Power supply assembly; 4. Actuator assembly; 5. Flexible transmission cable; 6. Lower leg wear assembly; 7. Ankle wear assembly; 8. Foot pressure sensor; 9. Inertial measurement unit. Detailed Implementation

[0016] This invention discloses a control method, device, and wearable ankle exoskeleton for wearable ankle joints, wherein, as Figure 1 As shown, the wearable ankle exoskeleton includes a plantar pressure sensor 8, an inertial measurement unit 9, a motion state acquisition unit for acquiring real-time ankle angle and real-time angular velocity, a lower leg wear assembly 6, an ankle wear assembly 7, an actuator assembly 4, a flexible transmission cable 5, and a controller 2. The plantar pressure sensor, inertial measurement unit, motion state acquisition unit, and actuator assembly are all signal-connected to the controller. The flexible transmission cable 5 is used to transmit the driving force generated by the actuator assembly to the lower leg wear assembly and the ankle wear assembly.

[0017] Of course, this wearable ankle exoskeleton also includes the necessary power supply component 3, such as Figure 1As shown, the controller 2 and power supply assembly 3 can be positioned on the user's back, waist, or other easily wearable locations; the actuator assembly 4 can be positioned on the ankle, lower leg, waist, back, or an external fixed platform. A flexible transmission cable 5 runs along the user's lower limbs and cooperates with the lower leg wear assembly 6 and ankle wear assembly 7 to transmit the driving force generated by the actuator assembly 4 to the vicinity of the ankle joint. The lower leg wear assembly 6 and ankle wear assembly 7 are used to convert the driving force into an assist torque acting on the ankle joint.

[0018] The plantar pressure sensor 8 is used to detect ipsilateral heel strike events and other gait events; the inertial measurement unit 9 is used to obtain the motion state of the foot, lower leg, or ankle joint. The controller 2 is the aforementioned control device for a wearable ankle exoskeleton. This device includes a memory and a controller connected in series. The memory stores a computer program, and the controller is used to read the computer program and execute any of the control methods for a wearable ankle exoskeleton described in this solution. Specifically, the control method for a wearable ankle exoskeleton includes steps S1 to S9.

[0019] Step S1: Based on the stride duration estimate The most recent heel-to-toe incident The estimated gait phase at the current sampling time t is calculated. .

[0020] Two consecutive heel strikes constitute one gait cycle for that heel. Each gait cycle corresponds to multiple sampling times.

[0021] At the current sampling time t, the stride duration estimate is used. The estimated gait phase was calculated. , specifically , In the formula, This represents a phase normalization operator that maps the input to the interval [0,1), used to avoid the phase calculation result from exceeding the effective range of a complete gait cycle.

[0022] Stride duration estimate It can be the historical stride duration or an estimate obtained from a gait estimator, which is a constant value within a gait cycle.

[0023] The heel strike event can be determined by signals from plantar pressure sensors, inertial measurement units, joint encoders, electromyographic signals, vision, or a combination thereof.

[0024] Step S2: Based on the estimated stride duration value Estimating gait phase Generate the desired angle of the ankle joint. Desired angular velocity and desired angular acceleration .

[0025] Expectation angle Gait phase can be estimated The calculation yields the following results: In one implementation, the desired angle A piecewise cubic polynomial is used for generation. A complete gait cycle is divided into multiple phase intervals. For the i-th phase interval... Based on the desired angles corresponding to the start and end points of this phase interval. and desired angular velocity Determine the cubic polynomial. Wherein, Let be the starting phase node of the i-th phase interval. It is the terminating phase node of the i-th phase interval.

[0026] when At that time, for the estimated gait phase Normalization is performed to obtain the normalized gait phase. ,in, .

[0027] Then the expected angle for: , In the formula, , , and The coefficients are all based on the expected angles at both ends of the i-th phase interval. and desired angular velocity It has been confirmed.

[0028] Desired angular velocity and desired angular acceleration The desired angle can be obtained by time-scale conversion based on the expression for the estimated gait phase, specifically: , .

[0029] Step S3: Obtain the real-time angle of the ankle joint. and real-time angular velocity And according to the desired angle and real-time angle Calculate angle tracking error According to the desired angular velocity and real-time angular velocity Calculate angular velocity tracking error .

[0030] The real-time angle of the ankle joint during this step Real-time angular velocity can be obtained from sensor measurements or state estimators. It can be measured by an inertial measurement unit, a joint encoder, or obtained by a state estimator.

[0031] Angle tracking error for: , Angular velocity tracking error for: .

[0032] Step S4: Calculate the angle tracking error. Angular velocity tracking error and desired angular acceleration By inputting the continuous impedance branch model, the continuous impedance torque at the current sampling time is obtained. The continuous impedance branch model is as follows: , In the formula, The continuous impedance torque output by the continuous impedance branch model during the k-th gait cycle at the current sampling time t is given by the unit t. ; Equivalent stiffness gain, in units of ; This is the equivalent damping gain, in units of ; This is the inertial feedforward coefficient, in units of... In this embodiment, the control calculation of continuous impedance branches is uniformly performed using radians, and the angle tracking error is minimized. The unit is Angular velocity tracking error The unit is Desired angular acceleration The unit is This ensures that the dimensions on both sides of the equation of the continuous impedance branch model are consistent.

[0033] Equivalent stiffness gain Used to adjust the ankle joint angle error feedback intensity, affecting the reference trajectory tracking capability; equivalent damping gain Used to adjust the angular velocity error feedback strength and system damping, affecting motion smoothness and oscillation suppression capability; inertial feedforward coefficient This is used to adjust the feedforward compensation intensity corresponding to the desired angular acceleration, thereby improving the response capability to dynamic changes in gait.

[0034] The continuous impedance branch model operates throughout the gait cycle to provide error feedback, damping injection, and optional inertial feedforward. , , Within one gait cycle, , and The three parameters remain unchanged. This branch forms a stable, interpretable real-time feedback framework and continues to operate outside the pulse window.

[0035] Step S5: Based on the estimated gait phase Pulse start phase and pulse width Determine whether the current sampling time is within the target assist window. If the current sampling time is within the target assist window, determine the pulse amplitude corresponding to the gait cycle. Pulse start phase and pulse width Generating pulse torque ; when When the current sampling time is determined to be within the target assist window, the pulse branch model is activated to generate pulse torque. At this point, based on the pulse amplitude corresponding to the gait cycle... Pulse start phase and pulse width Generating pulse torque Among them, pulse amplitude Used to adjust the intensity of local assist torque and positive mechanical work output within the target assist window; pulse start phase Used to adjust the start time of localized assist, synchronizing the assist with the target pushing phase; pulse width Used to adjust the continuous phase range of local assistance, and together with the pulse amplitude, it affects the duration and energy distribution of the assist torque.

[0036] Preferably, the pulse branch model uses raised cosine finite-width pulses, specifically: , = , In the formula, , Let be a dimensionless raised cosine window function. The window function satisfies ... Therefore, the amplitude of the pulse branch does not exceed This raised cosine window function is continuous at both the start and end points and exhibits a smooth first-order change, thus avoiding abrupt changes caused by ideal impulses or rectangular pulses.

[0037] This is the pulse termination phase of the gait cycle.

[0038] when or The pulse branch model outputs zero, that is: 0.

[0039] This method ensures that both the pulse start phase and the pulse end phase are within the preset assist phase range, thereby preventing local assist from exceeding the target gait phase.

[0040] Step S6: The continuous impedance torque and the pulse torque Synthesized to obtain the desired assist torque ; Specifically, .

[0041] This step enables the continuous impedance branch model to be responsible for full-cycle tracking and damping, while the pulse branch model is responsible for concentrated push-pull work. Structurally, this avoids two types of problems: insufficient push-pull due to continuous control only and insufficient full-cycle regulation due to pulse control only.

[0042] Step S7: For the desired assist torque... By applying amplitude constraints and torque change rate constraints, the actual assist torque is obtained. .

[0043] This step first determines the desired assist torque. Amplitude constraints are applied to obtain the intermediate torque. Specifically: = In the formula, It is a symmetric saturation function. Limit the input to a range Inside, This represents the upper limit of the actual assist torque amplitude. ; The unit is This represents the torque reference value provided to the actuator's underlying control loop after torque amplitude constraints.

[0044] Furthermore, based on the aforementioned intermediate torque amount The actual assist torque at the previous sampling time t-1 The deviation between them is used to calculate the actual rate of change of assist torque at the current sampling time. Specifically: , In the formula, This represents the upper limit of the actual rate of change of assist torque. ; The time constant is used to realize the equivalent first-order torque of the actuator assembly and transmission mechanism. .

[0045] Finally, based on the actual rate of change of assist torque at the current sampling time The actual assist torque at the previous sampling time t-1 The actual assist torque at the current sampling time is calculated. ,in, = , In the formula, The sampling interval is the time interval between two adjacent sampling times.

[0046] The time rate of change of the actual assist torque is obtained by dividing the deviation between the expected assist torque after amplitude constraint and the current actual assist torque by the first-order torque realization time constant, and is constrained to an interval. Inside.

[0047] but ， .

[0048] Then execution deviation for: .

[0049] It is expected that the assist torque will not be directly applied to the human-exoskeleton coupled object, but will instead be applied to the human-exoskeleton coupled object after passing through amplitude constraints and torque change rate constraints in sequence to form the actual assist torque.

[0050] Parameter vector at multiple sampling times within a gait cycle The parameter vector remains unchanged. This includes relevant parameters in the continuous impedance branch model and the pulse branch model, namely, the equivalent stiffness gain. Equivalent damping gain Inertial feedforward coefficient Pulse amplitude Pulse start phase and pulse width All remain unchanged. Specifically, the parameter vector... for: .

[0051] After one gait cycle is completed, the above parameter vector can be... Each parameter is updated. Specifically, the update method is implemented using steps S8 to S9.

[0052] Step S8: After detecting the corresponding heel landing event, update the stride duration estimate.

[0053] Specifically, in response to the detection of a heel strike event, the time interval between the two most recent adjacent heel strike events is calculated; when the time interval is within a preset stride duration range, it is determined that a gait cycle has been completed, and the time interval between the two strike events corresponding to the completed gait cycle is updated to the stride duration estimate corresponding to the ongoing gait cycle.

[0054] The timestamps corresponding to two adjacent landing events are respectively , ,and For timestamp The time interval is calculated based on the timestamp of the most recent heel-to-toe landing event and the timestamps of the two adjacent landing events. ,in, = .

[0055] when If the stride duration is within a preset reasonable range, then the gait cycle is determined. Completed, and entering the next ongoing gait cycle. +1. At this point, the time interval of the completed gait cycle is... Update to the next ongoing gait cycle +1 is the estimated stride duration.

[0056] If the time interval is not within the preset stride duration range, the stride duration estimate and parameter vector will not be updated, that is, the ongoing gait cycle will not be updated. The stride duration estimate and parameter vector are updated by +1, while the stride duration estimate and parameter vector are retained as the most recently verified safe and feasible parameters.

[0057] Step S9: Update the parameter vector. Specifically, this step includes steps 91 to 93.

[0058] Step 91: Based on the angle tracking error within the completed gait cycle. Angular velocity tracking error Pulse start phase Pulse width Actual assist torque The time variation rate of actual assist torque Expected assist torque and real-time angular velocity Gait diagnostic features were calculated.

[0059] Specifically, gait diagnostic features include the root mean square error of the angles corresponding to the completed gait cycle. Root mean square error of angular velocity Root mean square error of angles within the target assistance window Torque usage Torque change rate usage The squared mean of the execution deviation The actual assist torque is the positive mechanical work input to the ankle joint within the target assist window. Positive mechanical work generated by non-target windows Utilization rate of execution constraints ,in, , , , , , , , , , , In the formula, For timestamp The most recent time stamp of the incident where a foot followed the ground To prevent deviations from the expected outcome; Indicates the maximum value; This represents the utilization rate of the execution constraint, which is the maximum value between the utilization rate of the actual assist torque amplitude and the utilization rate of the actual assist torque change rate.

[0060] In this step, the constraint utilization is performed. This refers to maximizing the utilization of the actual assist torque amplitude and actual torque change rate relative to the corresponding execution constraint boundaries. This is achieved by improving... And suppress This allows the assist energy to be concentrated in the preset pushing phase.

[0061] Execution bias is used both for gait-level performance evaluation and as a bounded input in closed-loop analysis. Compared to existing schemes that only truncate peak values ​​at the output, this execution layer writes the execution bandwidth into the control channel as a dynamic constraint, ensuring that both training and deployment are based on achievable actual torque rather than ideal commands.

[0062] Step 92: After normalizing the gait diagnostic features and parameter vectors, calculate the diagnostic cost and learning state. The parameter vectors include the equivalent stiffness gain. Equivalent damping gain Inertial feedforward coefficient Pulse amplitude Pulse start phase and pulse width .

[0063] Since the parameters in the parameter vector have different physical dimensions and numerical ranges, they are normalized without dimensions before being input into the learning-based parameter optimization module. For the i-th parameter in the parameter vector... Let its preset physical lower bound and physical upper bound be respectively and The corresponding normalization parameter for: Therefore: The normalization parameters are combined into a normalization parameter vector. ,in, , In the formula, To normalize the equivalent stiffness gain, To normalize the equivalent damping gain, For normalized inertial feedforward coefficients, To normalize the pulse amplitude, The normalized pulse start phase, This is the normalized pulse width.

[0064] Through the above normalization process, parameters with different physical dimensions and numerical ranges are transformed to a unified dimensionless scale.

[0065] The diagnostic cost is calculated after normalizing the gait diagnostic features. ,in: , In the formula, All are non-negative weights. The normalized root mean square error of the angle , The normalized root mean square error of angular velocity , The root mean square error of the angle within the normalized target window. , Normalized torque usage , The amount used for the normalized torque change rate , The normalized mean square of the execution bias , The normalized actual assist torque is the positive mechanical work input to the ankle joint within the target assist window. , Positive mechanical work generated by the normalized non-target window , Normalized execution constraint utilization .

[0066] The above steps involve controlling the root mean square error of the angle. Root mean square error of angular velocity Root mean square error of angles within the target assistance window Torque usage Torque change rate usage The squared mean of the execution deviation Positive mechanical work generated by non-target windows and execution constraint utilization As a penalty, the positive mechanical work input to the ankle joint within the target assist window is the actual assist torque. As a reward.

[0067] Based on the normalized parameter vector Learning state is obtained from gait diagnostic features. ,in, , In the formula, T represents transpose.

[0068] Step 93: Based on the learning state and the diagnostic cost The parameter vector is updated. Specifically, this step includes steps 931 to 934.

[0069] Step 931, based on the diagnostic cost The network parameters of the learning parameter optimization module are updated, and the learning parameter optimization module is used to generate candidate parameter vectors based on the learning state. Specifically, the learning parameter optimization module can be EC-TD3. The learning parameter optimization module includes a policy network, two value evaluation networks, and a corresponding target network. The policy network is an actor network used to adjust the target network based on the current learning state. Output the candidate parameter vector for the next state period, i.e.: In the formula, Generate a dimensionless normalized candidate parameter vector for the actor network after adding exploration noise. This means that each component is restricted to the interval [-1, 1]. The parameter update strategy represented by the actor network. For actor network parameters; This represents the exploration noise that decays during training. All components in the normalized candidate parameter vector are confined to the interval [-1, 1].

[0070] The normalized candidate parameter vector output by the actor network is denormalized. For the i-th component in the normalized candidate parameter vector... According to the preset physical lower bound of the corresponding control parameters and the physical upper limit Candidate parameters with actual physical dimensions are obtained. : A candidate parameter vector with actual physical dimensions is formed by the inverse normalized candidate parameters. Specifically, the equivalent stiffness gain, equivalent damping gain, inertial feedforward coefficient, and pulse amplitude are restored to their corresponding physical dimensions, while the pulse start phase and pulse width are restored to their respective preset phase ranges. The inverse-normalized candidate parameter vector... Then, adjacent gait variation constraints, feasible region projection, and safety threshold checks are performed.

[0071] Diagnostic costs Used to characterize the current parameter vector within the k-th gait cycle. The overall execution effect is used as a feedback signal for the learning-based parameter optimization module to update the network parameters of the value evaluation network and the policy network. When using the reward-maximization-based EC-TD3 implementation, the diagnostic cost can be converted into a reward signal, and the value evaluation network and policy network can be trained and updated based on this reward signal.

[0072] Step 932: Perform adjacent gait variation constraints, feasible region projection, and safety threshold checks on the candidate parameter vector.

[0073] Specifically, the candidate parameter vectors are first subjected to adjacent gait variation constraints and feasible region projection, i.e., candidate parameter vectors that pass the safety threshold check. for: , In the formula, This is the parameter vector actually used in the k-th gait cycle; Used to limit the amount of parameter variation between adjacent gait cycles; Used to project the parameter vector onto a preset feasible region. Specifically: Equivalent stiffness gain in candidate parameter vector Equivalent damping gain The inertial feedforward coefficient remains positive and within preset upper and lower bounds. Pulse amplitude Within the allowable range, and ensuring that both the pulse start phase and pulse end phase are within the preset target assist phase range, that is: , , In the formula, and These are the lower and upper bounds of the preset target assist phase range, respectively.

[0074] Pulse width It should also be no less than the preset minimum width to avoid generating pulses that are too narrow and difficult to execute.

[0075] Next, a safety threshold check is performed on the candidate parameter vector.

[0076] Step 933: In response to the candidate parameter vector passing the safety threshold check, update the candidate parameter vector to a parameter vector, that is, use it as the parameter vector for the next state cycle. .

[0077] Step 934: In response to the candidate parameter vector failing the security threshold check, reject the candidate parameter vector and update the preset security parameter vector to the parameter vector.

[0078] To further illustrate the superiority of this solution, a simulation case and a real-world case will be used for explanation.

[0079] The single ankle joint model was solved using a fourth-order Runge-Kutta method with a fixed step length of 1 ms, and a stride duration of 1.08 s; the execution layer time constant was also used. The upper limit of torque amplitude is 15 ms. The upper limit of the torque change rate is 5 N·m. It is 100 N·m / s.

[0080] The learning-based parameter optimization module uses EC-TD3, which always generates a six-dimensional candidate parameter vector. After adjacent gait change constraints, feasible region projection, and safety threshold checks, the candidate parameter vector generates actual assist torque through the same execution layer.

[0081] The most representative parameter vector is: like Figure 2 and Figure 3 As shown, under the above parameter vector, the actual trajectory of the ankle joint can follow the reference trajectory well. The root mean square error of the unidirectional period angle is 0.23417°, the root mean square error of the angle within the target assistance window is 0.21144°, the root mean square error of the angular velocity is 2.9134° / s, and the positive mechanical work within the target assistance window is... The energy level was 196.13 mJ, and the number of constraint violations was 0. This result indicates that the continuous impedance branch model can maintain full-cycle error regulation, while the phase local pulse branch model does not degrade the overall tracking performance.

[0082] like Figure 4 and Figure 5 As shown, the expected assist torque, after being constrained by amplitude and torque change rate, becomes the actual assist torque. The actual peak torque is 2.940 N·m, and the peak torque change rate is 80.53 N·m / s, both within the preset boundaries. This result demonstrates that the execution of the constrained torque layer can avoid unattainable spike torques generated by learning or pulse assist.

[0083] like Figure 6 and Figure 7 The diagram shows a comparison of the results when the method in step S7 is not used (i.e., no execution layer is set), only the amplitude constraint method in step S7 is used, and the method in step S7 is used. When no execution layer is set and only amplitude constraints are set, the peak torque change rate reaches 1044.05 N·m / s, with 6 violations of the change rate constraint in each case. After setting both amplitude and change rate constraints, the peak change rate drops to 80.53 N·m / s, a reduction of 92.3%, and the number of violations is zero. This result demonstrates that amplitude limiting cannot replace dynamic change rate constraints, and the execution constraint torque layer of this patent has an independent technical function.

[0084] like Figure 8The control structure ablation results shown demonstrate that, compared to fixed-parameter hybrid control, EC-TD3 parameter optimization reduced the full-cycle root mean square error from 0.3108° to 0.2342°, and the root mean square error of the assist window from 0.2379° to 0.2114°. The positive mechanical work within the target assist window... The efficiency was improved from 148.77 mJ to 196.13 mJ while maintaining zero execution violations. This result demonstrates that the continuous impedance branch model, the pulse branch model, and the execution constraint learning optimization work synergistically to further improve performance.

[0085] When combined perturbations occur in model parameters, stride duration, and execution constraints, a full-cycle root mean square error of 0.2749°, positive work of 221.29 mJ, and zero execution violations can still be achieved without retraining. Even with large phase and stride estimation errors, the execution layer can still maintain torque boundaries, but local impulses may miss the optimal push-off timing. Therefore, practical systems should incorporate protection logic such as gait event confidence detection, abnormal phase freezing, or degradation to a continuous impedance branch.

[0086] In another practical implementation, the ankle joint angle trajectory of a real human walking gait is used as a reference input. This reference trajectory contains more pronounced changes in the support-push-off phase, thus allowing for further testing of the phase adaptation, assist timing, and execution constraint capabilities of this patent under a real gait reference. Figure 9 and Figure 10 As shown, under a real gait reference trajectory, this patent can still achieve relatively small gait cycle tracking errors and assist window tracking errors, while keeping the actual torque and torque change rate within the set boundaries. This result demonstrates that this patent is not only applicable to nominal spline trajectories but can also be adapted to parameters for real human gait trajectories.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A control method for wearable ankle exoskeletons, characterized in that, Includes the following steps: Based on stride duration estimates The most recent heel-to-toe incident The estimated gait phase at the current sampling time t is calculated. ; Based on the stride duration estimate Estimating gait phase Generate the desired angle of the ankle joint. Desired angular velocity and desired angular acceleration ; Obtain the real-time angle of the ankle joint. and real-time angular velocity And according to the desired angle and real-time angle Calculate angle tracking error According to the desired angular velocity and real-time angular velocity Calculate angular velocity tracking error ; The angle tracking error Angular velocity tracking error and desired angular acceleration By inputting the continuous impedance branch model, the continuous impedance torque at the current sampling time is obtained. The continuous impedance branch model is as follows: ， In the formula, Equivalent stiffness gain, in units of ; Equivalent damping gain, in units of ; This is the inertial feedforward coefficient, in units of... Continuous impedance torque at the current sampling time The unit is Angle tracking error The unit is Angular velocity tracking error The unit is Desired angular acceleration The unit is ; Based on the estimated gait phase Pulse start phase and pulse width Determine whether the current sampling time is within the target assist window. If the current sampling time is within the target assist window, determine the pulse amplitude. Pulse start phase and pulse width Generating pulse torque ; The continuous impedance torque and the pulse torque Synthesized to obtain the desired assist torque ; For the desired assist torque By applying amplitude constraints and torque change rate constraints, the actual assist torque is obtained. .

2. The control method for a wearable ankle exoskeleton according to claim 1, characterized in that, The estimation of gait phase Pulse start phase and pulse width Determine whether the current sampling time is within the target assist window, and then include: In response to the current sampling time not being within the target assist window, the pulse torque will be... Set to 0.

3. The control method for a wearable ankle exoskeleton according to claim 1, characterized in that, According to pulse amplitude Pulse start phase and pulse width Generating pulse torque for: ， = ， In the formula, It is a dimensionless raised cosine window function.

4. The control method for a wearable ankle exoskeleton according to claim 1, characterized in that, The desired assist torque Apply amplitude constraints and torque change rate constraints, including: For the desired assist torque Amplitude constraints are applied to obtain the intermediate torque. ,in, = In the formula, It is a symmetric saturation function. Limit the input to a range Inside, This represents the upper limit of the actual assist torque amplitude. According to the intermediate torque amount The actual assist torque at the previous sampling time t-1 The deviation between them is used to calculate the actual rate of change of assist torque at the current sampling time. ,in, ， In the formula, This represents the upper limit of the actual rate of change of assist torque. The time constant is the equivalent first-order torque of the actuator assembly and transmission mechanism. Based on the actual rate of change of assist torque at the current sampling time The actual assist torque at the previous sampling time t-1 The actual assist torque at the current sampling time is calculated. ,in, = ， In the formula, The sampling interval is the time interval between two adjacent sampling times.

5. The control method for a wearable ankle exoskeleton according to claim 4, characterized in that, The actual assist torque is obtained And then it includes: In response to the detection of a corresponding heel strike event, the time interval between the two most recent adjacent heel strike events is calculated; In response to the time interval being within the preset stride duration range, it is determined that a gait cycle has been completed, and the time interval between the two grounding events corresponding to the completed gait cycle is updated to the estimated stride duration corresponding to the ongoing gait cycle. Based on the angle tracking error within the completed gait cycle Angular velocity tracking error Pulse start phase Pulse width Actual assist torque Actual assist torque variation rate Expected assist torque and real-time angular velocity Gait diagnostic features were calculated; After normalizing the gait diagnostic features and parameter vectors, the diagnostic cost and learning state are calculated. The parameter vectors include the equivalent stiffness gain. Equivalent damping gain Inertial feedforward coefficient Pulse amplitude Pulse start phase and pulse width ; The parameter vector is updated based on the learning state and diagnostic cost.

6. The control method for a wearable ankle exoskeleton according to claim 5, characterized in that, The angle tracking error based on the completed gait cycle Angular velocity tracking error Pulse start phase Pulse width Actual assist torque Actual assist torque variation rate Expected assist torque and real-time angular velocity The calculated gait diagnostic features include: Calculate the root mean square error of the angle corresponding to the completed gait cycle. Root mean square error of angular velocity Root mean square error of angles within the target assistance window Torque usage Torque change rate usage The squared mean of the execution deviation The actual assist torque is the positive mechanical work input to the ankle joint within the target assist window. Positive mechanical work generated by non-target windows Utilization rate of execution constraints ,in, ， ， ， ， ， ， ， ， ， ， In the formula, For timestamp The most recent time stamp of the incident where a foot followed the ground To prevent deviations from the expected outcome; This indicates taking the maximum value.

7. The control method for a wearable ankle exoskeleton according to claim 6, characterized in that, The step of updating the parameter vector based on the learning state and diagnostic cost includes: Based on the diagnostic cost The network parameters of the learning-based parameter optimization module are updated, and a candidate parameter vector is generated based on the learning state using the learning-based parameter optimization module. The candidate parameter vectors are subjected to adjacent gait variation constraints, feasible region projection, and safety threshold checks. In response to the candidate parameter vector passing the safety threshold check, the candidate parameter vector is updated to a parameter vector; In response to the candidate parameter vector failing the security threshold check, the preset security parameter vector is updated to the parameter vector.

8. The control method for a wearable ankle exoskeleton according to claim 7, characterized in that, The process of normalizing the gait diagnostic features and parameter vectors to calculate the diagnostic cost and learning state includes: The diagnostic cost is calculated after normalizing the gait diagnostic features. ,in: ， In the formula, All are non-negative weights. The normalized root mean square error of the angle , The normalized root mean square error of angular velocity , The root mean square error of the angle within the normalized target window. , Normalized torque usage , The amount used for the normalized torque change rate , The normalized mean square of the execution bias , The normalized actual assist torque is the positive mechanical work input to the ankle joint within the target assist window. , Positive mechanical work generated by the normalized non-target window , Normalized execution constraint utilization ; The parameter vector is normalized to obtain a normalized parameter vector. ; According to the normalized parameter vector Learning state is obtained from gait diagnostic features. ,in, ， In the formula, T represents transpose.

9. A control device for a wearable ankle exoskeleton, comprising a memory and a controller connected in sequence, wherein the memory stores a computer program, characterized in that: The controller is used to read the computer program and execute the control method for wearable ankle exoskeleton as described in any one of claims 1-8.

10. A wearable ankle exoskeleton, comprising a plantar pressure sensor, an inertial measurement unit, a motion state acquisition unit for acquiring real-time ankle joint angle and real-time angular velocity, a lower leg wear assembly, an ankle wear assembly, an actuator assembly, a flexible transmission cable for transmitting the driving force generated by the actuator assembly to the lower leg wear assembly and the ankle wear assembly, and a controller, wherein the plantar pressure sensor, the inertial measurement unit, the motion state acquisition unit, and the actuator assembly are all signal-connected to the controller, characterized in that... The controller is a control device for wearable ankle exoskeleton as described in claim 9.

Images