Exoskeleton robot adaptive control method and related devices

Abstract

The present disclosure provides an exoskeleton robot adaptive control method and related equipment, and relates to the technical field of exoskeleton robots. The method comprises the following steps: acquiring motion information and position information of each active joint in an exoskeleton robot; inputting the motion information and the position information into a pre-trained angle prediction model to obtain target angles of each active joint; taking a difference between the target angles and actual angles output by an exoskeleton controller as an angle error; constructing a virtual control variable based on the angle error; inputting the virtual control variable and motion parameters of the exoskeleton robot into a disturbance observer to generate an estimated disturbance; inputting the angle error, the virtual control variable and the estimated disturbance into an adaptive neural network controller to obtain control torques of each active joint; and generating control signals of each active joint based on the control torques of each active joint. The present disclosure has good control precision.

Description

Technical Field

[0001] The present disclosure relates to the technical field of exoskeleton robots, and in particular to an exoskeleton robot adaptive control method and related equipment. Background Art

[0002] Exoskeleton robots have important application value in areas such as assisted limb movement and rehabilitation training. The core challenge lies in accurately estimating human joint angles in real time and achieving adaptive control. Existing technologies have the following shortcomings:

[0003] 1. A single inertial measurement unit (IMU) suffers from signal drift and cumulative errors, making it difficult to meet high-precision requirements.

[0004] 2. Deep learning models (such as single LSTM) are not able to capture long-term dependencies when processing joint coupled motions;

[0005] 3. Traditional control methods are not capable of modeling dynamic joint coupling and cannot effectively handle time-varying interference in human-machine interaction;

[0006] 4. The deep learning-based angle estimation model lacks deep integration with the adaptive control strategy, making it difficult to simultaneously ensure tracking accuracy and constraint satisfaction.

[0007] 5. Existing systems have limited means of handling time-varying output constraints (such as the safety margin of joint movement). The traditional Barrier Lyapunov Function (BLF) has limited adaptability to asymmetric time-varying constraints, posing a safety hazard. Summary of the Invention

[0008] To overcome the problems existing in the related art, the present disclosure provides an exoskeleton robot adaptive control method and related equipment.

[0009] According to a first aspect of an embodiment of the present disclosure, there is provided an exoskeleton robot adaptive control method, characterized by comprising:

[0010] Obtain motion information and position information of each active joint in the exoskeleton robot;

[0011] Inputting the motion information and the position information into a pre-trained angle prediction model to obtain a target angle of each movable joint;

[0012] taking the difference between the target angle and the actual angle output by the exoskeleton controller as an angle error;

[0013] Based on the angle error, a virtual control quantity is constructed; wherein the virtual control quantity is used to indicate the ideal speed instruction required for each joint to achieve the target angle;

[0014] The virtual control quantity is combined with the motion parameters of the exoskeleton robot and inputted into a disturbance observer to generate an estimated disturbance; wherein the disturbance observer is used to quantify the error of the adaptive neural network controller and the external environment disturbance;

[0015] Inputting the angle error, the virtual control variable and the estimated disturbance into the adaptive neural network controller to obtain the control torque of each movable joint;

[0016] Based on the control torque of each active joint, a control signal of each active joint is generated.

[0017] In some embodiments, the method further comprises:

[0018] Constructing an adaptive law of the adaptive neural network controller;

[0019] Based on the adaptive law and in combination with the actual angle of each movable joint, generating the adaptive adjustment amount corresponding to each movable joint in the adaptive neural network controller;

[0020] Based on the adaptive adjustment amount, weight parameters in the adaptive neural network controller are updated.

[0021] In some embodiments, constructing an adaptive law of the adaptive neural network controller includes:

[0022] The adaptive law is formulated as:

[0023]

[0024] Where, Corresponding to the movable joints in the exoskeleton robot, , is the total number of movable joints in the exoskeleton robot; For the The estimated weights corresponding to the current active joints; for The derivative of It is Gaussian basis functions corresponding to active joints; is the global state input of the adaptive neural network controller; For the The normal number corresponding to the number of movable joints; For the Regularization coefficient corresponding to active joints; For the The velocity error corresponding to each active joint is ,in, For the The derivative of the actual angle of the active joint, For the The virtual control amount corresponding to the active joint, Used to indicate time.

[0025] In some embodiments, generating the adaptive adjustment value corresponding to each active joint in the adaptive neural network controller based on the adaptive law and in combination with the actual angle of each active joint includes:

[0026] Based on the actual angle of each active joint, a global state input of each joint in the adaptive neural network controller and a velocity error corresponding to each active joint are generated; wherein the global state input includes the actual angle, the actual angular velocity, the virtual control amount and the derivative of the virtual control amount;

[0027] The global state input and velocity error of each joint are respectively substituted into the corresponding adaptive law, and the adaptive law Integrate to obtain the adaptive adjustment amount corresponding to each active joint.

[0028] In some embodiments, constructing a virtual control variable based on the angle error includes:

[0029] Configuring the range of motion of each movable joint, wherein the range of motion is adapted to the physiological range of motion of the human joint;

[0030] Determining the time-varying boundaries of each movable joint based on the range of motion of each movable joint; wherein the time-varying boundaries are used to indicate the range of angular errors that can be generated by each movable joint within its range of motion;

[0031] Based on the time-varying boundaries of each movable joint and in combination with its angular error, the virtual control amount is constructed so that the ideal speed instruction shown by the virtual control amount does not break the constraint of the time-varying boundaries.

[0032] In some embodiments, constructing a virtual control quantity based on a time-varying boundary includes:

[0033] The virtual control quantity is constructed as:

[0034]

[0035] Where, Corresponding to the movable joints in the exoskeleton robot, , is the total number of movable joints in the exoskeleton robot; is the virtual control quantity; are the target angles of each active joint derivative of the target angle, i.e., target angular velocity; gain coefficients of respective active joints; angle errors of respective active joints;

[0036]

[0037] is a normal number, used to ensure derivative of the target angle is bounded; is an upper time-varying boundary corresponding to the active joint; is a lower time-varying boundary corresponding to the active joint; is a derivative of the target angle; is a derivative of the target angle. is a derivative of the target angle. In some embodiments, the disturbance observer is characterized by the following formula:

[0038]

[0039] wherein,

[0040] is an estimated disturbance; is an estimated value of an observation error, associated with a virtual control quantity and a motion parameter of the exoskeleton robot; is a function vector of a velocity error .

[0041] In some embodiments, the angle prediction model comprises: a feature extraction unit, a time series modeling unit, and a linear mapping unit connected in sequence.

[0042] The feature extraction unit comprises: a multi-layer time convolution network, configured to perform time series feature extraction on input motion information and position information to obtain first time series features.

[0043] The time series modeling unit comprises: a multi-layer long short-term memory network, configured to capture long-term dependencies of movements of respective active joints in the first time series features to obtain second time series features.

[0044] The linear mapping unit comprises: an attention mechanism layer, configured to dynamically allocate attention weights of each time step in the second time series features, and map the second time series features to a target angle.

[0045] In some embodiments, there is also a residual connection between the input and the output of the feature extraction unit.

[0046] According to a second aspect of the embodiments of the present disclosure, an exoskeleton robot adaptive control device is provided, comprising:​

[0047] An acquisition module is used to obtain motion information and position information of each movable joint in the exoskeleton robot;

[0048] A prediction module, configured to input the motion information and the position information into a pre-trained angle prediction model to obtain a target angle for each movable joint;

[0049] a calculation module, configured to use the difference between the target angle and the actual angle output by the exoskeleton controller as an angle error;

[0050] A construction module is used to construct a virtual control amount based on the angle error; wherein the virtual control amount is used to indicate the ideal speed instruction required for each joint to achieve the target angle;

[0051] A first generating module is configured to input the virtual control variable into a disturbance observer in combination with the motion parameters of the exoskeleton robot to generate an estimated disturbance; wherein the disturbance observer is configured to quantify the error of the adaptive neural network controller and the external environmental disturbance;

[0052] a second generating module, configured to input the angle error, the virtual control variable, and the estimated disturbance into the adaptive neural network controller to obtain a control torque for each movable joint;

[0053] The third generating module is used to generate a control signal for each movable joint based on the control torque of each movable joint.

[0054] The solution provided by the embodiment of the present disclosure can achieve better control effects through adaptive control. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart of an adaptive control method for an exoskeleton robot in an embodiment of the present disclosure is shown.

[0056] Figure 2 A schematic diagram of a control architecture in an embodiment of the present disclosure is shown.

[0057] Figure 3 A schematic diagram of the architecture of a TLA model in an embodiment of the present disclosure is shown.

[0058] Figure 4 A schematic diagram comparing TCN-LSTM-Attention angle prediction and actual value in an embodiment of the present disclosure is shown.

[0059] Figure 5 A schematic diagram comparing a wrist joint tracking trajectory and a planned trajectory in an embodiment of the present disclosure is shown.

[0060] Figure 6A wrist joint tracking error schematic diagram in the embodiment of the present disclosure is shown.

[0061] Figure 7 A comparison between an elbow joint tracking trajectory and a planned trajectory schematic diagram in the embodiment of the present disclosure is shown.

[0062] Figure 8 An elbow joint tracking error schematic diagram in the embodiment of the present disclosure is shown.

[0063] Figure 9 A comparison between a shoulder joint tracking trajectory and a planned trajectory schematic diagram in the embodiment of the present disclosure is shown.

[0064] Figure 10 A shoulder joint tracking error schematic diagram in the embodiment of the present disclosure is shown.

[0065] Figure 11 A structure schematic diagram of an electronic device in the embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0066] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers refer to the same or similar elements throughout the drawings. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present disclosure. Instead, they are merely examples of apparatuses and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0067] First, the present disclosure provides a kind of exoskeleton robot adaptive control method, which can be executed by any electronic equipment.

[0068] Figure 1 A flowchart of the exoskeleton robot adaptive control method in the embodiment of the present disclosure is shown, as Figure 1 The exoskeleton robot adaptive control method provided in the embodiment of the present disclosure includes the following steps.

[0069] S101, the motion information and position information of each active joint in the exoskeleton robot are obtained.

[0070] S102, the motion information and position information are input into a pre-trained angle prediction model to obtain the target angle of each active joint.

[0071] S103, the difference between the target angle and the actual angle output by the exoskeleton controller is taken as the angle error.

[0072] S104, a virtual control quantity is constructed based on the angle error.

[0073] The virtual control quantity is used to show the ideal speed instruction required by each joint to reach the target angle.

[0074] In S105, the virtual control quantity is combined with the motion parameters of the exoskeleton robot to input the disturbance observer to generate an estimated disturbance.

[0075] The disturbance observer is used to quantify the error of the adaptive neural network controller and the external environmental disturbance.

[0076] In S106, the angle error, the virtual control quantity, and the estimated disturbance are input into the adaptive neural network controller to obtain the control torque of each active joint.

[0077] In S107, the control signal of each active joint is generated based on the control torque of each active joint.

[0078] In some embodiments, the weight parameters of the adaptive neural network controller (Radial Basis Function, RBF) can be controlled by a pre-constructed adaptive law to enable the adaptive neural network controller to adapt to the dynamic changes and external disturbances of the exoskeleton robot system, thereby improving the control accuracy and robustness.

[0079] Exemplarily, the adaptive law of the adaptive neural network controller can also be pre-constructed. Based on the adaptive law, in combination with the actual angles of each active joint, the adaptive adjustment quantity corresponding to each active joint in the adaptive neural network controller can be generated. Based on the adaptive adjustment quantity, the weight parameters in the adaptive neural network controller can be updated.

[0080] For ease of understanding, the following will be combined with Figure 2 The implementation architecture of the embodiments of the present disclosure is fully described. Figure 2 An execution architecture schematic diagram of an adaptive control method of an exoskeleton robot in the embodiments of the present disclosure is shown.

[0081] As Figure 2 shown, the execution architecture can include a multi-IMU acquisition module, a TLA angle prediction module, a virtual control quantity construction module, a disturbance observer, an adaptive neural network controller, an exoskeleton controller, and an adaptive law.

[0082] In some embodiments, the multi-IMU (Inertial Measurement Unit) acquisition module is used to acquire the motion parameters of each active joint. For example, for an exoskeleton robot for the upper limb, a plurality of IMUs can be used to acquire the joint angles and angular velocities of the shoulder joint, elbow joint, and wrist joint of the upper limb. At the same time, Opencap is used to capture the motion of the upper limb and extract the joint flexion values of each joint.

[0083] In some embodiments, the TLA (TCN-LSTM-Attention) angle prediction module is configured to predict the target angle of each active joint of the exoskeleton robot based on the motion parameters collected by the IMU collection module. The module includes a feature extraction unit, a time series modeling unit, and a linear mapping unit connected in sequence.

[0084] The feature extraction unit includes a multi-layer temporal convolutional network (TCN) configured to perform time series feature extraction on the input motion information and position information to obtain first time series features.

[0085] The time series modeling unit includes a multi-layer long short-term memory network (LSTM) configured to capture long-term dependencies of the motion of each active joint in the first time series features to obtain second time series features.

[0086] The linear mapping unit includes an attention mechanism layer (Attention) configured to dynamically allocate attention weights for each time step in the second time series features and map the second time series features to the target angle.

[0087] By way of example, there is also a residual connection between the input and output of the feature extraction unit.

[0088] In some embodiments, the input of the virtual control quantity construction module is and which are the lower limit safety angle and the upper limit safety angle of the joint range of motion, respectively, reflecting the physiological or mechanical safety constraints of the human joint and the exoskeleton; and the angle error , is the actual angle output by the exoskeleton controller, is the target angle predicted by the TLA angle prediction module, is used to represent time. The output of the virtual control quantity construction module is the virtual control quantity which can be understood as an ideal speed command (see equation 16 below), which serves as an intermediate control variable for the subsequent adaptive neural network controller.

[0089] The virtual control quantity construction module can combine the safety constraints of joint motion and the current tracking deviation to construct auxiliary logic, convert the "safety boundary limit" and "tracking error feedback" into an intermediate guiding quantity for the control strategy, avoid the exoskeleton motion exceeding the safety range, and preliminarily adjust the control direction.

[0090] By incorporating safety constraints and error feedback in the control flow in advance, a constrained intermediate reference can be provided for adaptive control, ensuring that the exoskeleton movement not only tracks the desired angle but also does not break the joint safety activity boundary, thereby improving control safety and rationality.

[0091] In some embodiments, the inputs of the disturbance observer are the inertia matrix (reflecting inertia resistance), the Coriolis force and centrifugal force matrix (reflecting the coupling effect of motion), the virtual control variable , and the actual angle . The output is the estimated disturbance , including external environmental disturbance and error due to inaccurate modeling of the exoskeleton dynamics model.

[0092] The disturbance observer can observe and separate various disturbances received by the exoskeleton robot system in real time based on the exoskeleton dynamics model and motion state, decouple the exoskeleton mechanical properties, motion state, and disturbance influence, and provide "disturbance compensation basis" for the controller.

[0093] It can be understood that the exoskeleton robot may be subject to various unknown disturbances in actual operation, which will directly affect the control accuracy and stability. Through the disturbance observer, these unknown disturbances can be perceived and quantified in advance, facilitating subsequent controller to make targeted compensation, improving system robustness, and ensuring stable operation of the exoskeleton in complex environments or human-machine interaction.

[0094] In some embodiments, the inputs of the adaptive neural network controller are the virtual control variable output by the virtual control variable construction module, the estimated disturbance output by the disturbance observer, and the aforementioned angle error . The output is the control torque τ, which is used to drive the exoskeleton actuator (i.e., the motor).

[0095] The adaptive neural network controller can utilize the nonlinear approximation capability of neural networks to integrate the virtual control variable, estimated disturbance, and angle error, and adaptively adjust the control strategy to convert the multi-element input into torque instructions for driving the movement of each active joint of the exoskeleton robot, realizing the closed-loop logic of "error-driven-disturbance compensation-precision control".

[0096] That is, the adaptive neural network controller can cope with the nonlinearity, time-varying nature, and disturbance uncertainty of the exoskeleton system, adjust the control parameters through online learning of the adaptive neural network, ensure the joint angle tracking accuracy, and at the same time, combine the safety constraints and disturbance compensation provided by the aforementioned modules to make the exoskeleton movement accurate, stable, and safe.

[0097] In some embodiments, the inputs of the adaptive law are the actual trajectory The output is the adaptive adjustment amount, which is used to update the internal weight parameters of the adaptive neural network controller in real time.

[0098] The adaptive law can construct parameter adaptive update rules based on the operating errors and dynamic model characteristics of the exoskeleton robot system, so that the weight parameters of the adaptive neural network controller can be dynamically adjusted with the system state, always adapting to the time-varying dynamic characteristics and interference environment of the exoskeleton.

[0099] The adaptive law solves the problem of uncertainty in the exoskeleton model parameters. Through the adaptive law, the adaptive neural network controller can achieve "self-learning and self-adjustment", avoiding the failure of the fixed parameter controller when the system characteristics change, ensuring the long-term effectiveness of the control strategy, and improving the adaptability and robustness of the exoskeleton control.

[0100] In some embodiments, the input of the exoskeleton controller is the control torque τ output by the adaptive neural network controller, and the output is the control signal (e.g., motor drive voltage, current) actually acting on the exoskeleton joints, which is also used to feedback the actual angle of each active joint. , angular velocity and other motion parameters.

[0101] The exoskeleton controller, acting as the execution terminal for control commands and a source of motion state feedback, converts abstract control torques into actual control signals for the exoskeleton's movable joints. It also collects motion parameters in real time. This feedback allows higher-level control modules (such as the adaptive neural network controller) to perceive the control effects in real time, ensuring the exoskeleton's movements accurately follow the desired trajectory, forming a closed "control-feedback" loop.

[0102] The above describes the execution architecture of the exoskeleton robot adaptive control method provided in the embodiment of the present disclosure. Next, the implementation details of the above architecture will be explained with reference to a specific example, taking the control of an exoskeleton robot for an upper limb arm as an example.

[0103] 1. Data collection and preprocessing.

[0104] Subjects can wear three IMUs (located at the shoulder, elbow, and wrist) to collect 3D acceleration and angular velocity signals at each joint in real time at a 100Hz sampling rate. Simultaneously, an OpenCAP system (based on the iOS device camera) simultaneously captures the 3D coordinates of each joint's landmarks, providing highly accurate flexion and extension measurements.

[0105] Subsequently, the IMU signal can be denoised and filtered, and the Opencap data can be coordinate transformed. The sampling frequency of the data collected by the IMU and the data collected by Opencap can be unified to 100 Hz, and the data collected by the two can be aligned by timestamps to construct a multidimensional dataset containing joint motion information (collected by IMU) and position information (collected by Opencap).

[0106] 2. Implementation of angle prediction model.

[0107] like Figure 2 As shown in Figure 3, the angle prediction model can construct a three-level angle prediction model including a temporal convolutional network, a long short-term memory network, and an attention mechanism layer.

[0108] The temporal convolutional network can adopt a three-layer causal convolution structure, with the convolution kernel size of each layer being 15, the number of channels being 128, 256, and 512, and the dilation rates d being 1, 2, and 4, respectively. It can extract local temporal features from the multidimensional dataset constructed above, ensure real-time performance through causal convolution, and avoid gradient vanishing by connecting the input and output of the temporal convolutional network through residual connections.

[0109] For example, the output of the causal convolutional architecture at layer k at time t is Depends only on the data of the k-1th layer before time t , so future information will not affect the past. For a one-dimensional input sequence X=( , , ,…, ) and convolution filter F=( , , ,…, ), a larger filter is generated from the original filter by introducing zero values ​​into the original filter. Its mathematical expression is:

[0110]

[0111] in, is the convolution filter, is a one-dimensional input sequence. Indicates that the output corresponds to the input sequence at time location. is the length of the filter, It is a filter In the The weight coefficient of each position, is the expansion factor. Used to represent a time offset.

[0112] The LSTM network can be configured with a two-layer structure, with 64 hidden units in each layer, using Dropout (probability 0.2) to avoid overfitting, and using a gate structure (input gate, forget gate, output gate) to capture the long-term dependency of joint motion. j The update process can be described as:

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] in, represents the sigmoid function, represents the hyperbolic tangent function. [ , ] indicates hidden state With the current input The concatenation vector of Represent the activation vectors of the LSTM input gate, forget gate, cell candidate value, and output gate respectively. Represents element-wise multiplication. , , , The weight matrices corresponding to the input gate, forget gate, cell candidate value, and output gate are used to convert the current input With the previous hidden state Mapped to the activation value of the corresponding gate. is the bias vector of each gate, used to adjust the activation threshold.

[0120] The attention mechanism layer weights dynamic features. By calculating attention weights at each time step, the model focuses on key movement phases (such as gait transitions and peak force), enhancing the model's sensitivity to dynamic features. It also makes the model interpretable, visualizing weights to illustrate the basis for model decisions (e.g., high weights correspond to important events), suppresses noise interference, and enhances prediction of relevant features.

[0121] In order to evaluate the prediction performance of the angle prediction model, the root mean square error (RMSE) can be introduced. ) and the average coefficient of determination (Coefficient of Determination, ) to quantify the prediction accuracy. It indicates the degree of fit between the predicted value and the observed value curve. The closer its value is to 0, the better the prediction performance. The calculation formula is as follows:

[0122]

[0123] The closer it is to 1, the stronger the model's anti-interference ability is. A value greater than 0.8 indicates a good model fit. The calculation formula is as follows:

[0124]

[0125] in, is the total number of samples, is the sample index value, is the predicted value, is the true value, is the arithmetic mean of the true values.

[0126] For example, during the training of the angle prediction model, and In order to optimize the goal, the back propagation algorithm is used to train the angle prediction model.

[0127] 3. Implementation of adaptive control.

[0128] The dynamic model of the rigid exoskeleton system corresponding to the movable joints with 3 degrees of freedom (DOF) in space can be expressed as:

[0129]

[0130] in, is the joint variable vector, represents the control input torque, is a symmetric positive definite inertia matrix, Represents centrifugal force and Coriolis torque, reflecting the physical parameters of the exoskeleton manipulator's dynamic characteristics, which are determined by the exoskeleton structure and motion state. Represents gravity, are interference vectors from humans and the environment.

[0131] in, is obliquely symmetric, that is .

[0132] Consider a positive constant ,if ,and , then the following inequality holds:

[0133]

[0134] when ,and When the symmetric positive definite is true, the following inequality holds:

[0135]

[0136] in, express The minimum eigenvalue of express Based on the actual application environment, assuming that the interference imposed by humans and the environment is continuous and bounded. In order to design a control law for the exoskeleton system, the actual trajectory output by the control law Able to track target trajectory , while ensuring that the time-varying output constraint is never violated, i.e., if the time-varying bound is given as (lower limit of joint movement) and (the upper limit safety angle of joint movement), where Corresponding to the movable joints in the exoskeleton robot, , is the total number of movable joints in the exoskeleton robot.

[0137] The output of the control law is Should be kept in the following areas:

[0138]

[0139] In terms of control, an adaptive neural network controller can be used to compensate for the effects of system uncertainty, and an asymmetric time-varying barrier Lyapunov function can be used to avoid violating time-varying constraints.

[0140] Here we define a new combined disturbance consisting of the estimation error of the adaptive neural network controller and the unknown external disturbance from humans and the environment, which is estimated by the disturbance observer (DO).

[0141] The angle error between the actual angle output by the exoskeleton controller and the target angle output by the angle prediction model can be expressed as , which can be used as error feedback of the control law to drive the neural network controller to adjust the output.

[0142] It is understandable that Here we define a new error variable, namely the speed error ,in is a virtual control quantity, which is used to represent the ideal speed instruction required to achieve the target angle. for The derivative of The time derivative of is as follows:

[0143]

[0144] in, is the inertia matrix The inverse matrix of .

[0145] The time-varying boundary of is set as:

[0146]

[0147]

[0148] in, and .

[0149] Considering the asymmetric time barrier Lyapunov function is:

[0150]

[0151] It is understandable that when the exoskeleton robot is used for rehabilitation, the range of joint motion is subject to human physiological limitations (for example, the maximum flexion / extension angle of the elbow joint is asymmetric), and the target angle is time-varying (for example, the range of joint motion is gradually expanded during rehabilitation training). In this case, the traditional symmetric / steady-state constrained Lyapunov function cannot adapt, so it is necessary to construct an asymmetric time-varying obstacle Lyapunov function. Equations (9) and (10) define the time-varying boundary and meet the requirement of "the range of joint motion can be adjusted in real time". When → or → hour, →∞, forcing the controller to apply a strong correction torque when the error approaches the boundary to avoid crossing the boundary.

[0152] Among them, formula (11) is expressed in the form of function , mandatory constraint error Do not break the time-varying boundaries, Defined as:

[0153]

[0154] in, is the output angle error No. A quantity, .

[0155] It is understandable that the asymmetric structure The symbol selection function is adapted to the physiological characteristics of human joints with different maximum flexion / extension angles. When approaching the upper boundary, =1; when approaching the lower boundary, =0, thus ensuring that the error is far away from the boundary.

[0156] Through the coordinate transformation of the error, we can get:

[0157]

[0158] Therefore, Equation (11) can be rewritten into a new form:

[0159]

[0160] By changing the formula (11) The derivative is:

[0161]

[0162] Equation (15) can relate the joint angle error, the time-varying boundary change rate and the system dynamics, providing a mathematical description of the error evolution for subsequent stability analysis. and is the rate of change of the time-varying boundary (corresponding to the dynamic adjustment of the range of joint motion during rehabilitation training, such as when the patient changes from passive training to active training, the boundary Expanded). Subsequent integration with adaptive laws and interference compensation demonstrates that these error terms ultimately converge exponentially to a small neighborhood. This derivation rigorously demonstrates the system's semiglobally uniformly bounded stability (i.e., all signals ultimately converge to a small range and do not diverge), guaranteeing the safety and effectiveness of the exoskeleton in rehabilitation scenarios (preferably without damaging the patient's joints while providing precise training assistance).

[0163] Please also note that due to space constraints, the above formula (15) is broken down into two lines, and the two lines arranged above and below in the outermost brackets are in a sequential addition relationship.

[0164] Based on this, the virtual control quantity shown in the following formula (16) can be constructed: Its physical meaning is the intermediate control instruction that integrates error feedback and time-varying constraints. It can also be understood as an ideal speed instruction, thereby providing a constrained reference signal for the adaptive neural network controller.

[0165]

[0166] where, is given by

[0167]

[0168] is a small positive constant, ensuring that even and are both zero, is also bounded, the gain matrix . Here, diag[⋅] is a mathematical notation to construct a diagonal matrix. The role is to match the dimensions, for example, the exoskeleton has n active joints, each active joint corresponds to a n is arranged in a diagonal matrix , so that the gain coefficient and the joint dimension one by one.

[0169] Substitute equation (16) and equation (17) into equation (14) to get

[0170]

[0171] In equation 18, the part after ≤ is the scaling derivation. By inequality scaling, the complex constraint term and error term are transformed into the "negative definite + positive definite" form that can be analyzed for stability. It provides the basis for subsequent design of control law and adaptive law, and proves that "by reasonably choosing the gain and , the system error can converge" (due to space constraints, equation 18 is split into three lines, and the above three lines are sequentially linked).

[0172] where, is set to:

[0173]

[0174] In the above equation, ， is the "gain coefficient" of the error cross term, by adjust the "coupling strength" of the two errors, so that the angle error and the dynamic error constrain and converge each other. For example, when the angle error is large, the feedback on is enhanced, and the trajectory is quickly pulled back.

[0175] Exemplarily, To compensate for the disturbance, a nonlinear disturbance observer can be used to estimate it. Define the following auxiliary variables:

[0176]

[0177] where, yes The function vector of (20) is used to find the time derivative, that is,

[0178]

[0179] Please note that due to space constraints, the above formula (21) is broken down into two lines, and the above two lines are sequentially connected.

[0180] in, .Will Set to A linear function of is a constant matrix, and DO can be easily designed and implemented. can be obtained through state feedback, so by estimating As an approximation of the interference To obtain an approximate value of The approximate value of The expression is as follows:

[0181]

[0182] in, yes From Equation (20), we can get the estimated value of interference Estimated value of:

[0183]

[0184] Thus we get

[0185]

[0186] Taking the time derivative, we can get

[0187]

[0188]

[0189] in, is an unknown constant, which is a conservative assumption for the “range of variation of the interference derivative”. In actual engineering implementation, it can be given by combining offline experiments with online adjustments (for example, using historical interference data to calculate the maximum derivative). A feasible estimate.

[0190] In the above equations, equation (21) is the disturbance observer, and equation (22) is the engineered disturbance observer. , but in formula (21) Unknown time-varying interference derivatives (for example, when a person exerts force suddenly, the interference change rate cannot be directly measured) cannot be used directly in engineering. In formula (22), Replace real interference The problem of “unknown interference derivative” is transformed into “based on observation error” through formula (23). Adaptive adjustment of Approach Formulas (24) to (26) are proofs. It will eventually converge to near 0.

[0191] Since the exoskeleton is susceptible to external interference, it is assumed that the actual value Can be divided into certain parts and the uncertain part ,in ,and is a symmetric positive definite matrix. Similarly, ,and is skew-symmetric. We use an adaptive neural network to estimate uncertainty. The actual output of the neural network is and approximate.

[0192]

[0193] in, is the input variable of RBF, is the estimation error, satisfying ,in yes The adaptive law is:

[0194]

[0195] In the above formula, It is the "global state input" of the neural network offline training / online learning, including the actual angle , actual angular velocity , virtual control quantity and its derivatives , which is used to approximate the unknown dynamics of the system. Figure 3 In the neural network controller, 、 and d is the “local error input” of the controller closed-loop feedback, since, and Directly related, belong The derived error, yes The basic variables of Figure 2 So there is no conflict.

[0196] Subscript Used to indicate the Joints (exoskeletons have multiple joints, such as shoulders, elbows, and wrists, which are used to distinguish the independent control of each joint in a multi-joint system. Indicates the joints error.

[0197] The purpose of the derivative of the above formula (28) (the core is to "let the neural network learn the unknown dynamics of the system online") is to Drive weight update, let Approximating ideal weights online , and finally Accurately offset the unknown parts of the system to ensure control accuracy.

[0198] Represents the ideal weight used to estimate NNS The actual weight of . is a positive constant, is the basis Gaussian function, is the number of nodes of NN, satisfying , It is a small positive number, which is the regularization coefficient. It avoids excessive weight updates and improves stability.

[0199] Ideal weight is the “neural network weight that can theoretically perfectly compensate for system uncertainty”, but it cannot be directly obtained in practice (because the system uncertainty is unknown). The role of the adaptive law: through the online update of formula (28), the actual weight Approximating ideal weights online , rather than being directly equal to . This is a gradual approximation process that allows the neural network to gradually adjust from "initial random weights" to "weights that can effectively compensate for uncertainty."

[0200] The global state input and velocity error of each joint are respectively substituted into the corresponding adaptive law, and the adaptive law Integrate to obtain the adaptive adjustment amount corresponding to each active joint, that is, the new weight parameter.

[0201] Next, the combined interference can be defined as:

[0202]

[0203] in, For external environmental interference, is the internal error of the model that is not accurately approximated.

[0204] The control law of the adaptive neural network controller is obtained as follows:

[0205]

[0206] This is the "iterative convergence" logic of closed-loop feedback control, not a true circular dependency. Its core lies in: 1. Initial moment: When control starts, Based on the initial observation (such as assuming that the initial interference is 0, or using the initial sensor data for rough estimation), an "initial compensation value" can be given first, and then the initial Drive the exoskeleton to move. 2. Iteration process: After the movement, the system state ( , , etc.), the disturbance observer is updated in real time according to the new state. , while the control law Also according to the new Adjust the output. This is a closed-loop iteration of "observe → control → feedback → re-observe," essentially a dynamic convergence process.

[0207] , is a positive-definite gain matrix, defined in Equation 19, and associated with the joint constraints. Its purpose: Directly feeds back the error, allowing the control law to prioritize "pullback" of angular deviations, ensuring accurate tracking of joint trajectories during rehabilitation training (for example, quickly correcting a patient's arm if it deviates from the planned trajectory). is the positive definite gain matrix at the velocity level, The action mode suppresses system oscillation and processes the dynamic fluctuations caused by "interference compensation and neural network learning" (for example, when the interference observer estimates a delay, Feedback quickly damps oscillations.

[0208] By updating formula (8), we can obtain:

[0209]

[0210] in, for The inverse matrix of . Since is the actual inertia matrix of the exoskeleton robot. Since the power system of the exoskeleton robot is relatively complex, it is difficult to accurately model it in actual engineering and there is uncertainty. Therefore, the ideal state of the exoskeleton controller is adopted. , replace the formula (8) Approximation, here It is the “nominal inertia matrix”, which can also be understood as the “ideal model matrix”.

[0211] Auxiliary variables can be expressed as:

[0212]

[0213] Taking the derivative we get:

[0214]

[0215]

[0216] Finally, the estimated interference is obtained:

[0217]

[0218] In the above equations, equation (32) takes the interference and system dynamics Bind, construct an auxiliary variable containing the interference.

[0219] Formula (33) is derived to establish the interference change rate The connection with the system dynamics is the "theoretical dynamic equation" of the disturbance observer.

[0220] Equation (34) uses a measurable control input , , Substitute unknowns (such as real interference ), transforming the theoretical dynamic equation into an observer update law that can be implemented in engineering.

[0221] Formula (35), from the estimated value of the auxiliary variable (Equation 34 can be integrated to solve) The interference estimate is calculated .

[0222]

[0223]

[0224]

[0225] right Derivation, we get

[0226]

[0227] Due to space constraints, the above formula (39) is broken down into three lines, and the above three lines are connected sequentially.

[0228] in, is the Lyapunov function The time derivative of is used to analyze the stability of the system. When it is less than or equal to 0, the system is asymptotically stable. and See eq. 18, 19. To constrain the transformation error, see eq. 13. is the disturbance observation error. is the ideal weight (optimal parameter of the neural network approximating the unknown dynamics), , is the actual online adjusted weight.

[0229] Substitute eq. (30) into eq. (39)

[0230]

[0231] Due to space constraints, the above eq. (40) is split into two lines, and the above two lines are sequentially connected.

[0232] According to the mean inequality, we get

[0233]

[0234]

[0235]

[0236]

[0237] Due to space constraints, the above eq. (44) is split into two lines, and the above two lines are sequentially connected.

[0238] where, is a positive number used to adjust the scaling degree.

[0239]

[0240] Simplify

[0241]

[0242] Considering the basis function of the Gaussian RBF NN, we have .

[0243]

[0244] Due to space constraints, the above eq. (47) is split into three lines, and the above three lines are sequentially connected.

[0245] where is greater than 0, and is the minimum convergence speed coefficient of the system. is greater than 0, and is a bounded constant determined by the system disturbance and the neural network approximation error. The system state (tracking error, constraint error, disturbance observation error, etc.) will converge to at a rate of In the neighborhood of the boundary, the asymptotic stability and constraint satisfaction of the control scheme are finally proved.

[0246]

[0247] is the convergence rate of the constraint error . The larger is, the more stringent the constraint is. is the convergence rate of the velocity error . is the control gain matrix. is the convergence rate of the neural network weight error . is the convergence rate of the disturbance observation error . Guarantees that the negative feedback dominates, and the system state will continue to converge.

[0248]

[0249] is the upper bound of the system steady-state error , is the boundedness of the derivative of the disturbance observation error , is the upper bound of . is the boundedness of the coupling between the ideal output of the neural network and the disturbance is the scaling parameter, is the number of basis functions. is the boundedness of the neural network weight error , is the upper bound of the weight error. Describes the "steady-state accuracy" of the system affected by disturbances and model uncertainties, and proves that the error will not grow unboundedly.

[0250] To ensure , select the constant , the gain matrix , and the auxiliary variable to satisfy the following two equations.

[0251]

[0252]

[0253] When , . The time-varying output constraint is never violated, i.e. . That is, the joint angle asymptotically converges to the desired angle , ensuring the tracking accuracy of the control system. At the same time, the active joint angle always strictly satisfies the time-varying safety boundary, , which guarantees the safety constraint of the control system.

[0254] Next, please refer to the following Tables 1 to 3, which respectively verify the effect of the TLA model. Table 1 shows the RMSE results of predicting the joint angle of the upper limb of the subject using the TLA model. Table 2 shows the results of predicting the joint angle of the upper limb of the subject using the TLA model.

[0255] Table 1 RMSE results of predicting the joint angle of the upper limb of the subject using the TLA model

[0256]

[0257] Table 2 results of predicting the joint angle of the upper limb of the subject using the TLA model

[0258]

[0259] Table 3 overall performance of five models in continuous motion estimation of arm joints

[0260]

[0261] It can be seen that the TLA model constructed by the embodiments of the present disclosure has high accuracy in angle prediction.

[0262] Next, please refer to the following Figures 4 to 10 , Figure 4 and Figure 10 respectively, which show the effect of the adaptive control method provided by the embodiments of the present disclosure. Through Figure 4 and Figure 10 it can be found that for the control of each joint, the accuracy of adaptive control gradually converges with the feedback of the controller, and has good effect.

[0263] ​This paper addresses the control problem of a 3-DOF exoskeleton system subject to time-varying output constraints. The exoskeleton system is described as a multi-input and multi-output nonlinear system. First, three IMUs are installed on the shoulder, elbow, and wrist joints of the exoskeleton, respectively. The exoskeleton's motion signals are collected through Opencap and transmitted to a TCN-LSTM neural network (TCN-LSTM-Attention, TLA) based on the attention mechanism for prediction. The network estimates the unknown dynamic characteristics of the exoskeleton system to address its uncertainty problem. Secondly, a disturbance observer is designed to estimate the unknown disturbances from the user and the environment. In the control design process, an asymmetric time-varying barrier Lyapunov function is used to avoid violating the time-varying output constraints. Experiments show that the system control scheme can achieve high-precision joint angle estimation, time-varying disturbance compensation, and constraint assurance, thereby improving the real-time, robustness, and safety of the exoskeleton system.

[0264] Based on the same inventive concept, the present disclosure also provides an exoskeleton robot adaptive control device, such as the following embodiment. Figure 1 The method embodiment shown is similar to that shown in FIG. 1 , so the implementation of the exoskeleton robot adaptive control device embodiment can refer to the above Figure 1 The implementation of the illustrated method embodiment will not be repeated any more.

[0265] Figure 11 FIG. 1 is a schematic diagram showing the structure of an exoskeleton robot adaptive control device according to an embodiment of the present disclosure. Figure 11 As shown, the exoskeleton robot adaptive control device 1100 includes: an acquisition module 1101, a prediction module 1102, a calculation module 1103, a construction module 1104, a first generation module 1105, a second generation module 1106, and a third generation module 1107.

[0266] An acquisition module 1101 is used to acquire motion information and position information of each movable joint in the exoskeleton robot;

[0267] Prediction module 1102, for inputting motion information and position information into a pre-trained angle prediction model to obtain target angles for each movable joint;

[0268] a calculation module 1103 for calculating the difference between the target angle and the actual angle output by the exoskeleton controller as an angle error;

[0269] A construction module 1104 is configured to construct a virtual control variable based on the angle error; wherein the virtual control variable is used to indicate an ideal velocity instruction required for each joint to achieve a target angle;

[0270] The first generation module 1105 is configured to input the virtual control variable into a disturbance observer in combination with the motion parameters of the exoskeleton robot to generate an estimated disturbance; wherein the disturbance observer is configured to quantify the error of the adaptive neural network controller and the external environment disturbance;

[0271] The second generating module 1106 is used to input the angle error, the virtual control amount and the estimated disturbance into the adaptive neural network controller to obtain the control torque of each movable joint;

[0272] The third generating module 1107 is used to generate a control signal for each movable joint based on the control torque of each movable joint.

[0273] Through the description of the above embodiments, it will be readily understood by those skilled in the art that the example embodiments described herein can be implemented via software or via a combination of software and necessary hardware. Therefore, the technical solutions according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or mobile hard drive) or on a network and includes several instructions for enabling a computing device (such as a personal computer, server, terminal device, or network device) to execute the methods according to the embodiments of the present disclosure.

[0274] In an exemplary embodiment of the present disclosure, a computer-readable storage medium is also provided, which may be a readable signal medium or a readable storage medium. A program product capable of implementing the above-mentioned method of the present disclosure is stored thereon. In some possible implementations, various aspects of the present disclosure may also be implemented in the form of a program product, which includes program code. When the program product is run on a terminal device, the program code is used to cause the terminal device to execute the steps according to various exemplary embodiments of the present disclosure described in the above-mentioned "Exemplary Method" section of the present disclosure.

[0275] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the appended claims.

Claims

1. An adaptive control method for an exoskeleton robot, characterized in that: include: Obtain motion information and position information of each active joint in the exoskeleton robot; Inputting the motion information and the position information into a pre-trained angle prediction model to obtain a target angle of each movable joint; The difference between the target angle and the actual angle output by the exoskeleton controller is used as the angle error; Based on the angle error, a virtual control quantity is constructed; wherein the virtual control quantity is used to indicate the ideal speed instruction required for each joint to achieve the target angle; The virtual control quantity is combined with the motion parameters of the exoskeleton robot and inputted into a disturbance observer to generate an estimated disturbance; wherein the disturbance observer is used to quantify the error of the adaptive neural network controller and the external environment disturbance; Inputting the angle error, the virtual control variable and the estimated disturbance into the adaptive neural network controller to obtain the control torque of each movable joint; Generate control signals for each active joint based on the control torque of each active joint; The method further comprises: Constructing an adaptive law of the adaptive neural network controller; Based on the adaptive law and in combination with the actual angle of each movable joint, generating the adaptive adjustment amount corresponding to each movable joint in the adaptive neural network controller; Based on the adaptive adjustment amount, weight parameters in the adaptive neural network controller are updated.

2. The method according to claim 1, characterized in that The adaptive law of the adaptive neural network controller is constructed, comprising: The adaptive law is formulated as: Where, Corresponding to the movable joints in the exoskeleton robot, , is the total number of movable joints in the exoskeleton robot; For the The estimated weights corresponding to the current active joints; for The derivative of It is Gaussian basis functions corresponding to active joints; is the global state input of the adaptive neural network controller; For the The normal number corresponding to the number of movable joints; For the Regularization coefficient corresponding to active joints; For the The velocity error corresponding to each active joint is ,in, For the The derivative of the actual angle of the active joint, For the The virtual control amount corresponding to the active joint, Used to indicate time.

3. The method according to claim 2, characterized in that The method of generating the adaptive adjustment amount corresponding to each active joint in the adaptive neural network controller based on the adaptive law and in combination with the actual angle of each active joint includes: Based on the actual angle of each active joint, a global state input of each joint in the adaptive neural network controller and a velocity error corresponding to each active joint are generated; wherein the global state input includes the actual angle, the actual angular velocity, the virtual control amount and the derivative of the virtual control amount; The global state input and velocity error of each joint are respectively substituted into the corresponding adaptive law, and the adaptive law Integrate to obtain the adaptive adjustment amount corresponding to each active joint.

4. The method according to claim 1, wherein The step of constructing a virtual control variable based on the angle error includes: Configuring the range of motion of each movable joint, wherein the range of motion is adapted to the physiological range of motion of the human joint; Determining the time-varying boundaries of each movable joint based on the range of motion of each movable joint; wherein the time-varying boundaries are used to indicate the range of angular errors that can be generated by each movable joint within its range of motion; Based on the time-varying boundaries of each movable joint and in combination with its angular error, the virtual control amount is constructed so that the ideal speed instruction shown by the virtual control amount does not break the constraint of the time-varying boundaries.

5. The method according to claim 4, characterized in that The virtual control amount is constructed based on the time-varying boundaries of each movable joint and its angle error, including: The virtual control quantity is constructed as: Where, Corresponding to the movable joints in the exoskeleton robot, , is the total number of movable joints in the exoskeleton robot; is the virtual control quantity; are the target angles of each active joint The derivative of , that is, the target angular velocity; are the gain coefficients of each active joint respectively; are the angle errors of each active joint; is a positive constant, used to ensure The derivative of It is bounded; For the The upper time-varying boundary corresponding to the active joint; For the The lower time-varying boundary corresponding to the active joint; for The derivative of for The derivative of .

6. The method according to claim 1, characterized in that The disturbance observer is characterized by the following equation: in, To estimate interference; is the estimated value of the observation error, which is associated with the virtual control quantity and the motion parameters of the exoskeleton robot; is the speed error The function vector of .

7. The method according to claim 1, characterized in that The angle prediction model includes: a feature extraction unit, a time series modeling unit and a linear mapping unit connected in sequence; The feature extraction unit includes: a multi-layer temporal convolutional network, configured to extract temporal features from the input motion information and position information to obtain a first temporal feature; The time series modeling unit includes: a multi-layer long short-term memory network, which is used to capture the long-term dependency relationship of the movements of each active joint in the first time series feature to obtain a second time series feature; The linear mapping unit includes: an attention mechanism layer, which is used to dynamically allocate the attention weight of each time step in the second time series feature and map the second time series feature to a target angle.

8. The method according to claim 7, characterized in that There is also a residual connection between the input and output of the feature extraction unit.

9. An exoskeleton robot adaptive control device, characterized in that: The device is used to perform the method according to any one of claims 1 to 8, and the device includes: An acquisition module is used to obtain motion information and position information of each movable joint in the exoskeleton robot; A prediction module, configured to input the motion information and the position information into a pre-trained angle prediction model to obtain a target angle for each movable joint; a calculation module, configured to use the difference between the target angle and the actual angle output by the exoskeleton controller as an angle error; A construction module is used to construct a virtual control amount based on the angle error; wherein the virtual control amount is used to indicate the ideal speed instruction required for each joint to achieve the target angle; A first generating module is configured to input the virtual control variable into a disturbance observer in combination with the motion parameters of the exoskeleton robot to generate an estimated disturbance; wherein the disturbance observer is configured to quantify the error of the adaptive neural network controller and the external environmental disturbance; a second generating module, configured to input the angle error, the virtual control variable, and the estimated disturbance into the adaptive neural network controller to obtain a control torque for each movable joint; The third generating module is used to generate a control signal for each movable joint based on the control torque of each movable joint.

Images