Control method of lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy

Abstract

The invention discloses a control method of a lower limb rehabilitation robot system based on personalized gait generation and an active compliance control strategy. The control method comprises the steps that S1, lower limb gait data of a plurality of volunteers are collected and preprocessed; s2, constructing a personalized gait generation model, and generating a personalized gait mode according to the preprocessed lower limb gait data; s3, based on a personalized gait mode, establishing an inner ring-position controller based on active disturbance compensation control (ADRC) to perform expected trajectory tracking; an outer ring-trajectory planner is established, and a preset trajectory is adjusted in real time according to the man-machine interaction torque and the predicted system state; and S4, realizing active gait rehabilitation training through a personalized gait generation model and a double-loop active compliance control strategy. The gait track designed and generated by the method is highly matched with a normal gait mode, the track loss is small, a safe and personalized rehabilitation training scheme is provided for a patient, and the applicability and effect of gait training are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of rehabilitation medicine and robotics technology, and in particular relates to a control method for a lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy. Background Art

[0002] In recent years, with the continuous development of rehabilitation medicine, the application of robotics in rehabilitation therapy has become increasingly widespread, especially in the rehabilitation training of patients with lower limb dysfunction caused by neurological diseases. The use of robot-assisted training has become an important method to improve rehabilitation effects and solve the problem of insufficient medical resources. Lower Limb Rehabilitation Robots (LLRR) help patients regain their ability to walk independently by providing safe and personalized training for them. However, current robot-assisted rehabilitation equipment has certain limitations, which are mainly reflected in the diversity and personalization of gait trajectories. Although existing rehabilitation robots are designed with physiologically standardized gait patterns and adjustable structures to adapt to the body shapes of different patients, due to the significant differences in individual gait, existing technologies lack flexibility in dealing with the gait characteristics of different patients, which can easily cause patients to feel uncomfortable during training and may induce muscle compensation.

[0003] Existing studies have shown that individual differences in gait can be quantified by measurable physical parameters, and neural networks provide an effective method for constructing a mapping between body parameters and gait. Some existing neural network-based gait generation models have provided valuable insights into personalized gait generation, but their application in rehabilitation training has not been fully explored. In addition, existing control methods often have difficulty maintaining the rhythmicity and coordination of gait in the process of real-time integration of patient movement intentions and preset training trajectories, which has an adverse effect on the effectiveness of rehabilitation training.

[0004] The rhythmicity and coordination of gait trajectory are the key to achieving efficient rehabilitation training. As an effective tool for fitting periodic data, Fourier series has been used to model gait trajectory. Its parameters can accurately describe the velocity and acceleration of gait trajectory, thereby maintaining the stability and consistency of trajectory during rehabilitation training. In addition, the regulation of gait trajectory also requires a high-performance controller to cope with environmental disturbances and modeling uncertainties. Active disturbance compensation control (ADRC) has been shown to provide superior control performance in complex uncertain environments.

[0005] Based on the analysis of the above existing technologies, there is an urgent need for a new personalized gait generation and control system that can generate customized gait trajectories according to the patient's physical parameters and combine with a dual-loop active compliant control structure to achieve safe and personalized rehabilitation training. Summary of the invention

[0006] In view of the shortcomings of the existing technology, the present invention proposes a control method for a lower limb rehabilitation robot system based on personalized gait generation and active compliant control strategy. The gait trajectory designed and generated by this method is highly consistent with the normal gait pattern, with small trajectory loss, providing patients with a safe and personalized rehabilitation training plan, and significantly improving the applicability and effectiveness of gait training.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A control method for a lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy comprises the following steps:

[0009] S1: Collect lower limb gait data of several volunteers and perform preprocessing;

[0010] S2: Build a personalized gait generation model to generate a personalized gait pattern based on the preprocessed lower limb gait data;

[0011] S3: Based on the personalized gait pattern, an inner loop-position controller based on active disturbance compensation control ADRC is established to track the desired trajectory; an outer loop-trajectory planner is established to adjust the preset trajectory in real time according to the human-machine interaction torque and the predicted system state;

[0012] S4: Active gait rehabilitation training is achieved through personalized gait generation model and dual-loop active compliance control strategy.

[0013] Preferably, in S1, the method of collecting lower limb gait data of several volunteers and preprocessing the data comprises:

[0014] The Nokov infrared optical motion capture system was used to collect the lower limb gait data of the volunteers and obtain the spatial coordinate data of the gait;

[0015] Filtering and detrending the spatial coordinate data, and obtaining the hip joint angle and the knee joint angle by inverse kinematics calculation;

[0016] Obtain the volunteers' physical parameters;

[0017] Based on the volunteers' physical parameters, the fixed-period Fourier fitting method was used to periodically decompose the hip joint angle and knee joint angle. The time interval between the maximum peak points of two consecutive knee joint cycles was taken as the step time, and the Fourier parameters were optimized according to the step time to approximate the original gait data.

[0018] Preferably, in S2, the personalized gait generation model is constructed based on a convolutional neural network (CNN), and includes a Hyper-search model and a gait generation model; wherein, except for the HS module and the RepOpt module, the Hyper-search model and the gait generation model have the same structure;

[0019] The RepOpt module consists of a convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU layer; the HS module is built based on a multi-branch structure with residual connections: when the number of input channels N i With the number of output channels N o At the same time, it contains two branches: one branch is a convolution layer with a kernel size of 3, and the other branch is a convolution layer with a kernel size of 1. In addition, each convolution layer is followed by a trainable scaling layer, which consists of a trainable matrix The trainable scaling layer uses a broadcast mechanism to convert S c Multiply with the corresponding input; if the number of input channels is N i Equal to the number of output channels N o , contains an additional branch consisting of a trainable scaling layer, similar to the RepOpt module, with batch normalization layers and ReLU activation layers added to enhance nonlinear modeling capabilities.

[0020] Preferably, in S3, the inner loop position controller based on active disturbance compensation control ADRC includes an extended state observer ESO and a nonlinear state error feedback controller NLSEF, and the angular velocity is calculated according to the Fourier series to simplify the system design; the specific formula is:

[0021]

[0022] Among them, ω, a n and b n is the Fourier fitting parameter, t is the walking time, N is the number of Fourier fitting series, is the joint angular velocity, and n represents the nth cycle.

[0023] Preferably, the dynamic model of the man-machine system is: Where q=[q hip ,q knee ] T , and denote the angle, angular velocity and angular acceleration of the reference trajectory, respectively, q hip and q knee denote the hip joint angle and knee joint angle of the reference trajectory, and denote the hip joint angular velocity and knee joint angular velocity of the reference trajectory, and They represent the hip joint angular acceleration and knee joint angular acceleration of the reference trajectory, For parts that require precise modeling, and denote the inertia matrix, Coriolis matrix and gravity matrix respectively, d denotes the external disturbance, represents the control torque τ c and interaction torque The sum of the disturbances The expression is:

[0024] Preferably, the extended state observer ESO establishment method includes:

[0025] According to the total disturbance Angular acceleration Described as The system state is preset to x1=q, The disturbance variable is preset to Then the expanded state space equation is: where φ(x1,x2) represents the time derivative of x3, represents the angular velocity variable in the system state, represents the angular acceleration variable in the system state, and φ(x1,x2) represents the time derivative of x3 Used to estimate the disturbance change in the system, y is the system output, τ c represents the control input; the corresponding extended state observer is set as: Among them, z1, z2 and z3 represent the estimated values ​​of system state variables x1, x2 and x3 respectively. and The state variable derivatives of the extended state observer estimate the different state variables of the system and their rates of change. K1, K2 and K3 represent the gain coefficients in the extended state observer, which are used to adjust the dynamic response speed and stability of the observer. α1 and α2 represent the nonlinear coefficients of the fal function, which control the intensity of the nonlinear function. δ1 and δ2 represent the threshold parameters in the fal function, which define the linear interval of the fal function, so that it has a linear response in the small error area. is the difference between z1 and x1, that is The fal function is nonlinear and is expressed as: Among them, e represents the estimation error or input error, α represents the nonlinear coefficient, and δ represents the threshold parameter.

[0026] Preferably, the method for establishing the nonlinear state error feedback controller NLSEF includes:

[0027] Preset q d and They represent the desired joint angle and angular velocity respectively, and the control law is expressed as: Where e1 = q d -z1 represents the joint angle error, represents the angular velocity error, f3(e1) and f4(e2) represent fal(e1,α3,δ3) and fal(e2,α4,δ4) respectively, b1, b2 and b3 are gain coefficients;

[0028] Presets and Combined with the extended state space equation, the state space equation of the error is expressed as:

[0029] Preferably, in S3, the method of establishing an outer loop-trajectory planner and adjusting the preset trajectory in real time according to the human-machine interaction torque and the predicted system state includes:

[0030] Considering the human-robot system as a first-order spring-damper system, the expression is: Among them, D d and K d denote the desired damping and stiffness, ω d is based on τ t The expected gait period, ω r is the reference gait cycle designed by the therapist;

[0031] A first-order filter is applied to the interaction torque via a discrete time difference equation, which is expressed as Among them, τ tf represents the filtered moment, k represents the current time step, α f is the filter coefficient, α f ∈(0,1);

[0032] Using the filtered interaction torque τ tf and predicted system error, an adaptive damping method is preset; specifically: first, when the observer converges, the estimated error and Approaches zero, that is, the estimated states z1, z2 and z3 are approximately the system states. Represented as (z1,z2,z3) T , assuming that τ c Remains unchanged within the prediction range [k, k+N], where k represents the current time step, N represents the prediction duration, and Z(k+i|k) represents the system state at time k+i predicted at time k;

[0033] Set k+i =q d (k+i)-z1(k+ik) represents the prediction error in the prediction domain, qd (k+i) is calculated based on the Fourier series. The Fourier parameters remain unchanged in the interval [k, k+N]. The overall prediction error le is expressed as: le=l1|e k+1 |+l2|e k+2 |+l3|e k+3 |, where l1, l2 and l3 are constants. When le is greater than the preset requirement, it indicates that there is a deviation between the current state and the expected trajectory, and it is necessary to increase the virtual damping to adjust the trajectory. In addition, when the interaction torque is greater than the preset requirement, the damping coefficient D is appropriately reduced to make the period parameter ω d Ability to achieve the patient's desired exercise state in a timely manner;

[0034] Let ω max represents the maximum period parameter, ω min represents the minimum period parameter, and the final required period parameter is expressed as:

[0035] Preferably, the predicted system state is expressed as:

[0036] Z(k+ik)=A i Z(kk)+(A i-1 +A i-2 +...+I 6×6 )(Bτ+C)

[0037] Where i∈[0,N], the matrix Expressed as in is a diagonal matrix with all diagonal elements equal to the sampling period t. is represented as matrix is set to is the identity matrix.

[0038] Preferably, the method of increasing the virtual damping adjustment trajectory includes:

[0039] A Tanh-like function is preset to adjust the virtual damping d0 preset during training, expressed as: Where a, b, c and h are constants; the final human-robot interaction mapping is redefined as:

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention proposes a personalized gait generation method to provide a tailored trajectory, and introduces a dual-loop active compliance control strategy to integrate the patient's movement intention into the robot's control for robot-assisted training. Specifically, the method uses a gait generation model based on a convolutional neural network (CNN) to map the changes in gait patterns according to the differences in body parameters between individuals, and the model is trained through RepOptimizer. The outer loop-trajectory planner converts the wearer's movement intention into adjustment of the step time based on the interaction torque and the predicted system state. The inner loop-position controller is designed based on the active disturbance compensation control (ADRC) strategy. The gait trajectory designed and generated by this method is highly consistent with the normal gait pattern, with less trajectory loss, providing patients with a safe and personalized rehabilitation training program, and significantly improving the applicability and effectiveness of gait training. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 A schematic flow chart of a control method for a lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of the gait mobile phone experiment setup in an embodiment of the present invention; wherein (a) is a schematic diagram of the placement of an infrared camera; (b) is a schematic diagram of the locations of reflective markers attached to the outside of the hips, knees and ankles of volunteers;

[0045] Figure 3 Schematic diagrams of visualization of the angles of the hip joints and knee joints of three volunteers in the embodiments of the present invention; wherein, (a) is a schematic diagram of the hip joint angle value of Vol-1; (b) is a schematic diagram of the knee joint angle value of Vol-1; (c) is a schematic diagram of the hip joint angle value of Vol-2; (d) is a schematic diagram of the knee joint angle value of Vol-2; (e) is a schematic diagram of the hip joint angle value of Vol-3; (f) is a schematic diagram of the knee joint angle value of Vol-3;

[0046] Figure 4 An overview diagram of a personalized gait generation method in an embodiment of the present invention;

[0047] Figure 5 The overall design diagram of the outer loop-trajectory planner and the inner loop-position controller in the embodiment of the present invention;

[0048] Figure 6 Schematic diagram of the gait generated in the embodiment of the present invention and the real gait;

[0049] Figure 7 Schematic diagram of experimental results under constant torque and different k in the embodiment of the present invention; wherein (a) represents τ t1 and schematic diagram of hip joint trajectory under different k; (b) represents τ t1 and schematic diagram of knee joint trajectory under different k; (c) represents τ t2 and schematic diagram of hip joint trajectory under different k; (d) represents τ t2 and schematic diagram of knee joint trajectory under different k; (e) represents ω d With τ t1 Schematic diagram of changes; (f) represents ω d With τ t2 Schematic diagram of changes;

[0050] Figure 8 Schematic diagram of experimental results under constant torque and different d in the embodiment of the present invention; wherein (a) represents τ t1 and schematic diagram of hip joint trajectory under different d; (b) represents τ t1 and schematic diagram of knee joint trajectory under different d; (c) represents τ t2 Schematic diagram of hip joint trajectory under different d; (d) represents τ t2 and schematic diagram of knee joint trajectory under different d; (e) represents ω d With τ t1 Schematic diagram of changes; (f) represents ω d With τ t2 Schematic diagram of changes;

[0051] Fig. 9 is the period parameter ω in the embodiment of the present invention d Schematic diagram of the change of; where (a) represents ω d With τ t1 Schematic diagram of changes; (b) represents ω d With τ t2 Schematic diagram of changes;

[0052] Fig.10 Schematic diagram of experimental results under different torque and d in the embodiment of the present invention; wherein (a) represents τ t1 and schematic diagram of hip joint trajectory under different d; (b) represents τ t1 and schematic diagram of knee joint trajectory under different d; (c) represents τ t2 Schematic diagram of hip joint trajectory under different d; (d) represents τ t2 and schematic diagram of knee joint trajectory under different d; (e) represents ω d With τ t1Schematic diagram of changes; (f) represents ω d With τ t2 Schematic diagram of changes;

[0053] Fig.11 Schematic diagrams of hip joint and knee joint angle tracking in an embodiment of the present invention; wherein, (a) is a schematic diagram of hip joint angle tracking under Case-A; (b) is a schematic diagram of knee joint angle tracking under Case-A; (c) is a schematic diagram of hip joint angle tracking under Case-B; (d) is a schematic diagram of knee joint angle tracking under Case-B;

[0054] Fig.12 Schematic diagram of tracking and observation errors in an embodiment of the present invention; wherein (a) represents a schematic diagram of tracking and observation errors under Case-A; and (b) represents a schematic diagram of tracking and observation errors under Case-B. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Embodiment 1

[0058] like Figure 1 As shown, it is a schematic diagram of the method flow of this embodiment. The personalized gait generation model of the present invention generates potential gait in a healthy state according to the physical parameters of the body and the expected gait parameters, and expresses the gait in a Fourier fitting way, retaining the main features of the generated data, and providing support for the trajectory planner based on step time. The ADRC controller has excellent anti-interference ability and tracking performance under uncertain dynamic modeling conditions. Its internal extended state observer (ESO) estimates the system state in real time, providing a basis for system state prediction. The designed variable damping trajectory planner adjusts the reference gait through the interaction torque, and incorporates the wearer's movement intention into the adjustment of gait and pace. The steps include:

[0059] S1. Collect lower limb gait data and perform preprocessing to obtain processed data.

[0060] In this embodiment, a gait collection experiment was constructed using the Nokov infrared optical motion capture system to collect gait data from 54 volunteers. Figure 2 As shown in (a), eight infrared cameras were placed near the wall and arranged in an ellipse around the measurement area. Reflective markers were attached to the outside of the volunteers’ hips, knees, and ankles to collect precise spatial coordinates, such as Figure 2 (b) as shown.

[0061] Joint coordinate data cannot be directly applied to controller design. Therefore, after filtering and detrending the collected data, we calculated the hip joint angle and knee joint angle (converting the raw data into hip and knee joint angles) through inverse kinematics (Ren J, Wang A, Li H, et al. A transformer-based neural network for gait prediction in lower limb exoskeleton robots using plantar force [J]. Sensors, 2023, 23 (14): 6547.). In addition to the gait trajectory, the gender of the volunteers was recorded and their physical parameters were measured, including height, weight, thigh length, calf length and hip width, all of which are highly correlated with the gait pattern. Table 1 lists the measured parameters and the maximum, minimum and average values ​​of the measured values ​​for three of the volunteers (Vol-1, Vol-2 and Vol-3). The data of the three volunteers' multiple walks were preprocessed and visualized, as shown in the figure. Figure 3 As shown, the trajectories of the hip and knee joint angles are intuitively displayed. Comparing the gait trajectories of different volunteers, it can be found that the maximum flexion angles of the hip and knee joints of Vol-1 are relatively large, which are 0.4199 rad and 1.1473 rad, respectively. In contrast, the maximum extension angle of the hip joint of Vol-2 is larger, which is -0.341 rad. For Vol-3, a clear peak is observed between 0.6s and 1.4s. In addition, the gait patterns of the same volunteer at different walking speeds show a high degree of similarity. Therefore, generating personalized gaits is necessary to provide appropriate training guidance for robots.

[0062] Table 1 Physical parameters of volunteers

[0063]

[0064] The Fourier fitting method decomposes the complex periodic sequence into the sum of multiple simple sine and cosine functions. In formula (1), q is the collected gait trajectory, t is the walking time, N is the series of Fourier fitting, a0, ω, a i and b j are the Fourier fitting parameters, i∈[0,n], j∈[1,n].

[0065]

[0066] Gait is usually approximately considered to be periodic, but in fact, the motion trajectory of each cycle is not exactly the same, and the starting and ending values ​​of the joint angles of each gait cycle will always change. For example, comparing the gait patterns of three volunteers across cycles, the maximum difference in hip flexion angle is 0.1293 rad, and the maximum difference in knee flexion angle is 0.1969 rad. Although these differences are small, Fourier fitting can only mechanically fit the data with a lower error without constraints, and cannot capture the inherent periodic characteristics of the original data.

[0067] To solve this problem, the fixed-period Fourier fitting method is used to periodically decompose the hip joint angle and knee joint angle, and the time interval between the maximum peak points of two consecutive knee joint cycles is used as the benchmark to determine the step time of the gait. The period parameters ω are calculated, where T represents the step time. Finally, the remaining finite Fourier parameters are iteratively optimized to approximate the original gait data.

[0068] Specifically, the fixed-period Fourier fitting method sets a fixed periodic parameter to fit the gait trajectory through the periodic characteristics of the gait data. In order to determine the step time, the following process is usually adopted: 1) Determine the periodic step time: First, by observing the original gait data (knee joint angle time series), find the maximum peak point of the knee joint in each gait cycle. Calculate the time interval between two adjacent peak points and take it as the step time of a cycle. Assume that the time of two adjacent peak points is t k and t k+1 , then the step time T can be expressed as T = t k+1 -t k , through this time interval, the gait cycle can be approximately defined to capture its main periodic characteristics. 2) Apply Fourier fitting: Use the step time T as the period parameter in Fourier fitting to fix the period of the fitting function. The basic form of the Fourier fitting function is shown in formula (1). 3) Iterative optimization of Fourier parameters: Under the condition of fixed period T, iteratively optimize the Fourier parameters a0, a i and b j , in order to minimize the fitting error and make the Fourier fitting curve as close to the actual gait data as possible. 4) Iterative optimization of Fourier parameters: The least squares method is used to fit the Fourier series model to approximate the original gait data. Specifically, the Fourier series model is shown in formula (1). The optimal Fourier parameters are obtained by minimizing the sum of squared errors between the original data and the fitting function. We first determine the step time T, and then use the angular frequency Fixed period, the Fourier parameters a0, a are obtained by least squares method nand b n , so that the fitted Fourier model can be as close to the original data as possible.

[0069] In this way, the fixed-period Fourier fitting can maintain the periodic description of the gait data with the lowest possible error, while avoiding the error accumulation caused by periodic changes.

[0070] S2. Construct a personalized gait generation model to generate personalized gait patterns based on body physical parameters, step time and stride length as a passive training trajectory in early rehabilitation.

[0071] Specifically, the personalized gait generation model is built based on a convolutional neural network (CNN). Figure 4 As shown in the figure, except for the HS module and the RepOpt module, the gait generation model and the corresponding Hyper-search model are constructed with the same structure. The main function of the Hyper-search model is to search for hyperparameters and generate optimization weights. It mainly includes the following parts: input layer, Gated module, HS module, RepOpt module, Pooling layer, Linear layer, and output layer.

[0072] Input layer: Both models take body parameters (height, weight, thigh length, calf length, hip width and gender) and gait parameters (step time and stride) as input. Output layer: The output of the model is the sixth-order Fourier fitting parameters of the left hip and knee joint angles (the hip and knee joint angles are predicted by the constructed personalized gait generation model, that is, the personalized gait pattern).

[0073] RepOpt module: The RepOpt module consists of a convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU layer, which is used to build the feature extraction part of the personalized gait generation model, with the aim of enhancing the nonlinear modeling capability of the model.

[0074] HS module: The HS module is a multi-branch structure based on residual connections. i With the number of output channels N o At the same time, it contains two branches: one branch is a convolution layer with a kernel size of 3, and the other branch is a convolution layer with a kernel size of 1. In addition, each convolution layer is followed by a trainable scaling layer, which consists of a trainable matrix The trainable scaling layer uses a broadcast mechanism to convert S c Multiply by its input. If N i Equal to N o, contains an additional branch consisting of a trainable scaling layer. Similar to the RepOpt module, a batch normalization layer and a ReLU activation layer are added to enhance the nonlinear modeling capability.

[0075] The Hyper-search model is trained on a small gait dataset using the AdamW optimizer. The training weights of the scaling layer are extracted and combined with prior knowledge to construct the gradient multiplication coefficients (Grad Mults) of the RepOptimizer to optimize the training process of the RepOpt module. In addition, during the training process of the Hyper-search model, we initialize the weights of the RepOpt module based on the weights of the HS module.

[0076] Gated module: In the Hyper-search model and the gait generation model, the Gated module is used to fine-tune the weights of the input parameters. In the Gated module, the Sigmoid layer acts as a "gate" to map the value to [0,1]. The closer the value is to 1, the more important the information is. The Tanh layer adjusts the feature distribution to be centered at 0 to enhance the nonlinearity of the model. The input and output of the Gated module are represented as In g and Out g , then its forward calculation process is shown in formula (2). Among them, w1, w2, b1 and b2 are the weights and biases of the two linear layers in the Gated module. Although all input elements are closely related to the gait pattern, their weights need to be further adjusted to enhance important features. To this end, Gated modules are added before and after the HS module and the RopOpt module. Pooling is a global average pooling layer. In order to reduce the impact of numerical size on network performance, the first linear layer is used to adjust the input weights. The last two linear layers and the ReLU layer are used to further integrate the learned features.

[0077] Out g =Sigmoid(w1·In g +b1)⊙Tanh(w2·In g +b2) (2)

[0078] S3. An outer loop trajectory planner is established to adjust the preset trajectory in real time according to the interaction torque and the predicted system state. At the same time, an inner loop position controller based on active disturbance compensation control (ADRC) is established to ensure accurate trajectory tracking.

[0079] The overall design of the outer loop - trajectory planner and the inner loop - position controller is as follows Figure 5As shown in the figure. The gait generation model is used to generate the possible gaits of patients in a healthy state. The generated gaits are represented by Fourier parameters and can be directly set as the desired trajectory of the controller for passive training, which is suitable for early rehabilitation. As the patient's autonomous movement ability improves and the motivation for active exercise increases, the generated gaits can be used as the input of the trajectory planner and adjusted in real time according to the human-computer interaction torque and system status for active training. In addition, a position controller based on active disturbance compensation control (ADRC) is designed to accurately track the desired trajectory.

[0080] Active disturbance compensation control is a model-free control strategy. Usually, ADRC consists of three main parts: transient process, extended state observer (ESO) and nonlinear state error feedback controller (NLSEF). However, in our method, the corresponding joint angular velocity is relatively smooth and can be directly calculated using formula (3), i.e., Fourier series, where is the joint angular velocity, and n represents the nth cycle. Therefore, for our specific application, the transient process is unnecessary.

[0081]

[0082] 1) Dynamic model: The dynamic model of the human-machine system is shown in formula (4), where q = [q hip ,q knee ] T , and and q knee Represent the hip joint angle and knee joint angle of the reference trajectory, and denote the hip joint angular velocity and knee joint angular velocity of the reference trajectory, and represent the hip joint angular acceleration and knee joint angular acceleration of the reference trajectory respectively. For parts that require precise modeling. and denote the inertia matrix, Coriolis matrix and gravity matrix respectively. d denotes external disturbance. represents the control torque τ c and interaction torque The sum of the total disturbance As shown in formula (5).

[0083]

[0084] Property 1: Matrix and its inverse matrix is symmetric and bounded.

[0085] Assumption 1: d and its first-order derivative is bounded.

[0086] 2) Extended State Observer (ESO): ESO is a tool for simultaneously estimating the system state and uncertainty. According to formula (5), the angular acceleration q can be described as The system state is defined as x1 = q, The disturbance variable is defined as Then the extended state space equation is shown in formula (6), where represents the angular velocity variable in the system state, represents the angular acceleration variable in the system state, and φ(x1,x2) represents the time derivative of x3 Used to estimate the disturbance change in the system, y is the system output, τ c Indicates control input.

[0087]

[0088] The corresponding extended state observer is designed as shown in formula (7):

[0089]

[0090] Among them, z1, z2 and z3 represent the estimated values ​​of the system state variables x1, x2 and x3, respectively, which are used to estimate the state and disturbance of the system. and The state variable derivatives of the extended state observer estimate the different state variables of the system and their rates of change. K1, K2 and K3 represent the gain coefficients in the extended state observer, which are used to adjust the dynamic response speed and stability of the observer. α1 and α2 represent the nonlinear coefficients of the fal function, which control the strength of the nonlinear function. δ1 and δ2 represent the threshold parameters in the fal function, which define the linear interval of the fal function, so that it has a linear response in the small error area. is the difference between z1 and x1, that is The fal function is nonlinear and can be expressed as formula (8):

[0091]

[0092] Where e represents the estimation error or input error, which generally represents the deviation between the observed value and its estimated value. α represents the nonlinear coefficient, which usually takes a value between 0 and 1 and is used to control the intensity of the nonlinear function. When α is large, the nonlinearity is high and the response is fast; when α is small, the system response is smoother. δ represents the threshold parameter, which is used to define the segmentation points of the nonlinear function. Usually, when the error e is less than δ, the function is in a linear form; when e is greater than δ, the function is in a nonlinear form. This parameter can prevent the system from producing drastic nonlinear responses when the error is small, thereby improving the stability of the observer.

[0093] make and The corresponding Lyapunov equation can be defined as The derivative of V is derived as follows:

[0094]

[0095] in, and Abbreviated as and The existence of K1, K2 and K3 makes In addition, when e→δ, the left and right derivatives of V exist and are bounded. That is, when e=δ, the convergence still holds. Based on the above analysis, the convergence of the designed ESO is proved.

[0096] 3) Nonlinear state error feedback controller NLSEF: The NLSEF module calculates the system state variable error based on the estimated system state variables, performs feedback control, and compensates for disturbances to reduce their impact on the control variables. d and q d denote the desired joint angle and angular velocity respectively, and the control law can be written as formula (10):

[0097]

[0098] Where e1 = q d -z1 represents the joint angle error, represents the angular velocity error. f3(e1) and f4(e2) represent fal(e1,α3,δ3) and fal(e2,α4,δ4) respectively. b1, b2 and b3 are gain coefficients, and b3 is set to 1 in this paper.

[0099] set up and Combined with formula (6), the state space equation of the error is expressed as follows:

[0100]

[0101] The stability of the designed controller is also proved by Lyapunov stability theory. By designing the Lyapunov function as The derivative of V is calculated as follows:

[0102]

[0103] Considering the stability of ESO, e1, e2, and is bounded. In addition, fal(·) is an odd function, monotonically increasing. Therefore, f3(e1) and f4(e2) are also bounded. Let y(e1,e2) = (e1+e2)(b1f3(e1)+b2f4(e2)), we can see that y(e1,e2) is even symmetric about the origin. Therefore, there exist suitable b1, b2 and fal function parameters, so that when e1+e2<0, b1f3(e1)+b2f4(e2)<0, and when e1+e2>0, b1f3(e1)+b2f4(e2)>0. In addition, when b1 and b2 are large enough, It shows the stability of the controller.

[0104] In the early stage of rehabilitation, the personalized gait trajectory is used as the rehabilitation training path to meet the training needs. However, as the patient's mobility improves, the movement trajectory needs to be further adjusted according to the patient's exercise willingness.

[0105] We designed a trajectory planner to adjust the preset step time of the original trajectory by modifying the cycle parameter ω during walking training. Specifically, considering the coupling of lower limb joint motion, the hip and knee joints should share a unified step time. In addition, the hip joint is closer to the nerve center, and its motion trajectory largely reflects the patient's movement intention. The movement patterns of the left and right lower limbs are symmetrical. The movement of the thigh is directly driven by the hip joint, so only the interaction torque at the unilateral thigh is considered. When the interaction torque is positive, it indicates that the patient's intended movement is ahead of the robot's current movement, indicating that the patient can train at a faster speed. At this time, the step time should be shortened and the cycle parameter should be increased accordingly; conversely, when the interaction torque is negative, it means that the patient's thigh movement lags behind the robot's current movement, and the cycle parameter should be reduced.

[0106] To describe the interaction torque τ at the thigh t and the expected period parameter ω d The mapping relationship between them is used to regard the human-robot system as a first-order spring-damper system, as shown in formula (13). d and K d denote the desired damping and stiffness respectively. ω d is based on τ t The expected gait period is obtained.r is the reference gait cycle designed by the therapist. d >0 and K d > 0, if the interaction torque τ t = 0, the periodic parameter will gradually converge to ω r , indicating that the system has internal stability.

[0107]

[0108] During the training process, the sudden changes in the interaction forces may cause the gait trajectory to be non-smooth. To solve this problem, a first-order filter is applied to the interaction torque through a discrete time difference equation to ensure smoothness. The process can be expressed as Among them, τ tf Represents the filtered moment. k represents the current time step. α f is the filter coefficient, α f ∈(0,1).

[0109] In addition, the admittance parameters of the human body vary with the patient’s physiological state, so it is challenging to design appropriate damping and stiffness parameters. In order to better align the gait trajectory with the wearer’s motion intention, the filtered interaction torque τ is used to tf and the predicted system error, we further designed an adaptive damping method. First, when the observer converges, the estimated error and approaches zero, that is, the estimated states z1, z2 and z3 can be approximated as the system state. Represented as (z1,z2,z3) T Assume τ c It remains unchanged within the prediction range [k, k+N], where k represents the current time step and N represents the prediction duration. Z(k+i|k) represents the system state at time k+i predicted at time k. The predicted system state is shown in formula (14).

[0110] Z(k+i|k)=A i Z(k|k)+(A i-1 +A i-2 +…+I 6×6 )(Bτ+C) (14)

[0111] In formula (14), i∈[0,N]. The matrix As shown in formula (15), is a diagonal matrix with all diagonal elements equal to the sampling period t. is represented as matrix is set to is the identity matrix.

[0112]

[0113] Set k+i =q d (k+i)-z1(k+ik) represents the prediction error in the prediction domain. d (k+i) is calculated based on the Fourier series, and the Fourier parameters remain unchanged in the interval [k, k+N]. N is 3. The overall prediction error le is shown in formula (16), where l1, l2 and l3 are constants. When le is large, it indicates that there is a large deviation between the current state and the expected trajectory, and it is necessary to increase virtual damping to quickly adjust the trajectory. In addition, when the interaction torque is large, the damping coefficient D can be appropriately reduced to make the period parameter ω d The patient's desired motion state can be achieved in time. Therefore, a Tanh-like function is designed to adjust the virtual damping d0 preset during training, as shown in formula (17). a, b, c and h are constants. The final human-robot interaction mapping is redefined as shown in formula (18).

[0114] le=l1|e k+1 |+l2|e k+2 |+l3|e k+3 | (16)

[0115]

[0116] To ensure the safety of the patient during training, the adjusted cycle parameters must be further restricted to a specific range. max represents the maximum period parameter, ω min represents the minimum period parameter. The final required period parameter is expressed as shown in formula (19):

[0117]

[0118] S4. Use the constructed personalized gait generation model and dual-loop active compliance control strategy to perform active gait rehabilitation training. Among them, the outer loop-trajectory planner and the inner loop-position controller are collectively referred to as the dual-loop active compliance control strategy.

[0119] Specifically, the formula and principle of the personalized gait generation model and the dual-loop active compliance control strategy are in steps S2 and S3, which are experimental verifications of the personalized gait generation model and the dual-loop active compliance control strategy.

[0120] In this embodiment, three experiments are mainly carried out, including a personalized gait generation experiment, a trajectory planner experiment, and a position tracking simulation experiment.

[0121] (1) Personalized gait generation experiment

[0122] Dataset setting: The data of 54 volunteers were randomly divided into 6 sets, each set contained the data of 9 volunteers, there was no overlap between the sets, the largest set contained 277 instances, and the smallest set contained 245 instances. Then one set was randomly selected to train the Hyper-search model, and the remaining 5 sets were used to train the gait generation model. The 5-fold cross-validation method was used, that is, 5 iterative experiments were performed, and one set was selected as the test set in each iteration, and the remaining sets were used as the training set.

[0123] Evaluation criteria: Mean absolute error (MAE) and dynamic time warping (DTW) are used as evaluation indicators. MAE quantifies the error between the generated value and the actual value at each time point, while DTW effectively handles the time axis offset, scaling, and nonlinear deformation in time series data. Assuming that the true fitting parameter is x and the output of the model is y, the corresponding fitting curves are represented as X and Y respectively. The error evaluation expression is shown in formula (20).

[0124]

[0125] Experimental setup: The Hyper-search model is trained using the regular optimizer AdamW to obtain appropriate RepOptimizer constant multipliers and initialization parameters for the gait generation model. Figure 4 N in is set to 9, and the number of output channels is set to 32, 32, 32, 64, 64, 128, 128, 256, 256. Subsequently, the gait generation model is further trained using the corresponding structural settings and RepOptimizer.

[0126] In addition, five neural network models, G1, G2, RV, TE and VGG-19, were applied to compare with the proposed gait generation model to verify the performance of the model. Model G1 is the same network as the proposed model, but the Gated module is deleted. Model G2 adds an ECA block after each RepOpt module. The EAC module is shown in the literature (Wang Q, Wu B, Zhu P, et al. ECA-Net: Efficient channel attention for deep convolutional neural networks [C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2020: 11534-11542.). Model RV is built based on the architecture and structure reparameterization method in the literature (Ding X, Zhang X, Ma N, et al. Repvgg: Making vgg-style convnets great again [C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2021: 13733-13742.), and the two-dimensional operations are adjusted to one-dimensional operations to adapt to gait data. Five stages are stacked, and the number of layers is set to 1, 2, 2, 2, and 2, respectively, while the number of output channels of each stage is 32, 32, 64, 128, and 256, respectively. Model TE adopts the encoder part of Transformer. The input in our experiment has no temporal relationship, so the position encoding module is deleted from the original structure. In addition, since the input dimension is fixed and the input data is globally visible, the padding and masking mechanisms in the original Transformer structure are not used. The number of attention heads in the model is set to 8. Before passing the input to the encoder, three convolutional layers (each followed by a batch normalization layer and a ReLU activation layer) are used to increase the data dimension to 64. VGG-19 uses the specific parameter configuration in the literature (Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition [J]. arXiv preprintarXiv: 1409.1556, 2014.). Similarly, in order to adapt to the data characteristics of this experiment, one-dimensional convolution is used to replace the two-dimensional convolution operation in the original structure.In order to ensure the effectiveness of the comparative experiment, the optimizer used in the training of the five models is also AdamW, the learning rate is set to 0.05, and the training generation number is 50.

[0127] Experimental results and analysis: Table 2 shows the loss of each model on the test set. Fold-i represents the i-th iteration experiment. The smaller the value of MAE and DTW, the more similar the generated gait trajectory is to the real gait trajectory. The results in the table are rounded to five decimal places, and the minimum loss loss2 in each round of iterative experiments is shown in bold. Table 3 shows the average time consumption corresponding to running one generation on the test set. The time unit is seconds.

[0128] Table 2. Loss of gait generation on the test set

[0129]

[0130] Table 3 Average time consumption of the previous generation of the test set

[0131] Proposed method G1 G2 RV TE VGG-19 time 3.32512 3.32119 3.33584 3.76525 9.24791 3.92139

[0132] The difference between the proposed method and model G1 is whether the Gated module is added. In the third iteration experiment Fold-3, the loss2 value of G1 is 23.26% higher than that of our model. This shows that the Gated module enhances the generalization ability of the model when processing new or unfamiliar data by further filtering out unimportant information, thereby improving the performance and stability of the model. Compared with our proposed model, G2 includes an ECA module, which can adjust channel weights to strengthen the important features learned. However, the performance of the G2 model is not significant in both MAE and DTW results. This can be attributed to two factors. On the one hand, the results of the proposed model are already close to the optimal, and further improvement is challenging. On the other hand, the Gated module is somewhat similar to the function of the attention mechanism.

[0133] The model RV has the smallest loss2 and relatively low time consumption on the test set, but in the training stage, the time consumed by re-parameterizing the multi-branch structure used before increased by 36.34%; the time consumption of the model AE based on the attention mechanism is 2.78 times that of our model. At the same time, compared with our model and the RV model, the model VGG-19 with more convolutional layers stacked also takes more time on the test set. Nevertheless, judging from the values ​​of loss1 and loss2, the experimental results of the VGG-19 model have not been significantly improved. The gait generation model we proposed achieved the lowest error in the second generation Fold-2 and the fifth generation Fold-5. The loss in the other three generations is also relatively small. In summary, considering the time cost consumed and the generation error of the model, our proposed model shows the best overall performance.

[0134] To further demonstrate the generated gait, the output of the proposed model is visualized, e.g. Figure 7 As shown in the figure, it can be observed that the generated data conforms to the human walking pattern. The error of the hip joint near the limbs is slightly larger, but does not exceed 0.1 radians. In addition, the generated data is relatively smooth and very consistent with the human walking pattern, making it suitable as a reference trajectory for the lower limb rehabilitation robot.

[0135] (2) Trajectory planner experiment based on step time

[0136] To verify the performance of the step-time based trajectory planner, simulation experiments with fixed and variable damping parameters were performed.

[0137] 1) Fixed damping simulation: The interaction torque τ in the time interval from 0.5s to 3.5s is t1 Expressed as 10Nm. Filter coefficient α f Set to 0.4.

[0138] Figure 7 (a), 7(b) and 7(e) show the τ t1 The trajectories of the hip and knee joints and the expected cycle parameters under different k. At 0.5s, the original trajectory is subject to τ t1 The expected period parameter ω gradually increases. After 3.5s, the interaction torque τ t1 becomes 0Nm, ω d It gradually converges to the original expected value of 3.75 within 0.45s. On the contrary, suppose that in the time interval from 0.5s to 3.5s, the interaction torque τ t2 is -10Nm. Figure 7 As shown in (c), 7(d), and 7(f), when interaction occurs, the period parameter ω gradually decreases. After 3.5s, τ t2 = 0Nm, ω gradually converges to the original expected value. When the period parameter ω reaches its corresponding limit, it will remain unchanged despite the interaction torque to ensure the safety and effectiveness of the training. Figure 8 A similar trend can be found in .

[0139] Remarkably, despite the abrupt change in the interaction torque, the reformed trajectory remains smooth and retains the basic shape of the normal gait. Fig. 9 (a) and 9(b) show the trajectory changes without and with filtering, respectively. At 3.5 seconds, the interaction torque suddenly changes from 10Nm to 0Nm. Without filtering, the slope of the angular trajectory changes greatly, while with filtering, the transition is smoother. Filtering the interaction torque plays a buffering role and ensures the smoothness of the trajectory to a certain extent.

[0140] The values ​​of k and d will affect the change of periodic parameters. Figure 7 and Figure 8 , when d is constant and k increases, the rising and falling speeds of the periodic parameters also increase. When k is constant and d increases, the changing speed of the periodic parameters decreases accordingly. In further applications, considering the patient's physical condition, we do not recommend that k be too large and d be too small.

[0141] 2) Multi-parameter simulation: The simulation of variable interaction torque and variable damping was further performed. Specifically, τ t3 It is expressed as follows:

[0142]

[0143] Wherein, t=0.02s, k represents the current sampling time.

[0144] Fig.10 (a), (b) and (e) are the trajectories obtained with fixed damping parameters and the corresponding ω changes. Fig.10 (c), (d) and (f) are the trajectories obtained by using variable damping parameters and the corresponding ω changes. In equation (16), the parameters l1, l2, and l3 are set to 0.7, 0.2, and 0.1, respectively. In equation (17), d0 is set to 30, c is set to 100, and b is set to 0.015. Fig.10 As shown, in the variable τ t3 Under the interaction of , both methods can ensure the smoothness of the reorganized trajectory. The period parameter ω gradually increases in the first 1.8s, drops to the lowest point at 2.6s, then continues to increase for 1.4s and converges to 3.75 at about 13s. The variable damping method adjusts the damping parameter d in real time according to the actual interaction torque and system tracking error, and adjusts the corresponding period parameters more flexibly while ensuring safety. In the fixed damping experiment, the range of ω is 3.695-3.938, while the maximum value of ω in the variable damping method is 3.980 and the minimum value is 3.502, which can respond to the patient's movement intention more quickly.

[0145] (3) Position tracking simulation

[0146] The model parameter settings in formula (4) are as follows:

[0147]

[0148] In formula (7), the parameters of ESO, K1, K2, and K3 are set to [140, 145] respectively. T , [350,720] T and [2200,1500] TIn addition, α1=0.5, α2=0.25, δ1=δ2=0.05. For the control law in formula (10), b1, b2, and b3 are set to [550,750] respectively. T , [28,10] T and 1. α3 is set to 0.7, α4 is set to 1.4, δ3 and δ4 are set to 0.01 and 0.1 respectively. The stiffness k is set to 20, and the sampling period is also set to 0.02s.

[0149] The same experiments with constant interaction torque and variable interaction torque as in the step-time-based trajectory planner experiments were performed. The estimation performance of the ESO and the tracking performance of the controller were evaluated using MAE. In 10 experiments under constant interaction torque, including the cases of k = [10, 13, 16, 19, 22], d = 10, τ = [10, -10], the MAE for the hip angle did not exceed 0.00285 rad, and the maximum MAE for the knee angle was 0.00726 rad. For the tracking error, the MAE for the cases of k = 22, τ = 10, τ = 2 ... t = -10, the MAE value is slightly larger, the error of the hip joint trajectory is 0.0112rad, and the error of the knee joint is 0.0300rad. This case is denoted as Case A. In five experiments under variable interaction torque, including the cases of k = [10, 13, 16, 19, 22] and variable damping, the estimated error of the hip joint is in the range of 0.00306rad to 0.00307rad, and the estimated error of the knee joint is in the range of 0.00739rad to 0.00748rad. When k = 10, the tracking MAE of the hip and knee joints is the largest, which is 0.01150rad and 0.03107rad, respectively. We call this case Case B.

[0150] To further analyze the performance of the designed controller, we visualize the trajectory tracking process of Case A and Case B. Fig.11 As shown in Figure 2, in both cases, the observation error of the state observer converges to within 0.02 rad within 0.4 s. In addition, the controller has a fast response time and is able to track the trajectory and reach a stable tracking state within 0.15 s. Although the initial tracking trajectory has slight oscillations, the oscillation range does not exceed 0.05 rad. Fig.11 As shown in (b), from 2.75s to 3.5s and around 3.8s, the knee joint angle tracking error is slightly larger, but the maximum error does not exceed 0.16rad. Fig.12 As shown in (b), the maximum steady-state tracking error of the knee joint angle is 0.094 rad, and the maximum steady-state tracking error of the hip joint does not exceed 0.027 rad. In general, the controller has a fast response speed, stable tracking, and has a good ability to overcome environmental disturbances and modeling errors.

[0151] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A control method for a lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy, characterized in that: The following steps are involved: S1: Collect lower limb gait data of several volunteers and perform preprocessing; S2: Build a personalized gait generation model to generate a personalized gait pattern based on the preprocessed lower limb gait data; S3: Based on the personalized gait pattern, an inner loop position controller based on active disturbance compensation control (ADRC) is established to track the desired trajectory; Establish an outer loop - trajectory planner to adjust the preset trajectory in real time according to the human-machine interaction torque and predicted system status; S4: Active gait rehabilitation training is achieved through personalized gait generation model and dual-loop active compliance control strategy.

2. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 1 is characterized in that: In S1, the method of collecting lower limb gait data of several volunteers and performing preprocessing includes: The Nokov infrared optical motion capture system was used to collect the lower limb gait data of the volunteers and obtain the spatial coordinate data of the gait; Filtering and detrending the spatial coordinate data, and obtaining the hip joint angle and the knee joint angle by inverse kinematics calculation; Obtain the volunteers' physical parameters; Based on the volunteers' physical parameters, the fixed-period Fourier fitting method was used to periodically decompose the hip joint angle and knee joint angle. The time interval between the maximum peak points of two consecutive knee joint cycles was taken as the step time, and the Fourier parameters were optimized according to the step time to approximate the original gait data.

3. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 1 is characterized in that: In S2, the personalized gait generation model is constructed based on a convolutional neural network (CNN), and includes a Hyper-search model and a gait generation model; wherein, except for the HS module and the RepOpt module, the Hyper-search model and the gait generation model have the same structure; The RepOpt module consists of a convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU layer; the HS module is built based on a multi-branch structure with residual connections: when the number of input channels N i With the number of output channels N o At the same time, it contains two branches: one branch is a convolution layer with a kernel size of 3, and the other branch is a convolution layer with a kernel size of 1. In addition, each convolution layer is followed by a trainable scaling layer, which consists of a trainable matrix The trainable scaling layer uses a broadcast mechanism to convert S c Multiply with the corresponding input; if the number of input channels is N i Equal to the number of output channels N o , contains an additional branch consisting of a trainable scaling layer, similar to the RepOpt module, with batch normalization layers and ReLU activation layers added to enhance nonlinear modeling capabilities.

4. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 1 is characterized in that: In S3, the inner loop position controller based on active disturbance compensation control ADRC includes an extended state observer ESO and a nonlinear state error feedback controller NLSEF, and the angular velocity is calculated according to the Fourier series to simplify the system design; the specific formula is: Among them, ω, a n and b n is the Fourier fitting parameter, t is the walking time, N is the number of Fourier fitting series, is the joint angular velocity, and n represents the nth cycle.

5. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 4 is characterized in that: The dynamic model of the human-machine system is: Where q=[q hip ,q knee ] T , and denote the angle, angular velocity and angular acceleration of the reference trajectory, respectively, q hip and q knee denote the hip joint angle and knee joint angle of the reference trajectory, and denote the hip joint angular velocity and knee joint angular velocity of the reference trajectory, and They represent the hip joint angular acceleration and knee joint angular acceleration of the reference trajectory, For parts that require precise modeling, and denote the inertia matrix, Coriolis matrix and gravity matrix respectively, d denotes the external disturbance, represents the control torque τ c and interaction torque The sum of the disturbances The expression is:

6. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 5 is characterized in that: The extended state observer ESO establishment method includes: according to the total disturbance Angular acceleration Described as The system state is preset to x1=q, The disturbance variable is preset to Then the expanded state space equation is: where φ(x1,x2) represents the time derivative of x3, represents the angular velocity variable in the system state, represents the angular acceleration variable in the system state, and φ(x1,x2) represents the time derivative of x3 Used to estimate the disturbance change in the system, y is the system output, τ c represents the control input; the corresponding extended state observer is set as: Among them, z1, z2 and z3 represent the estimated values ​​of system state variables x1, x2 and x3 respectively. and The state variable derivatives of the extended state observer estimate the different state variables of the system and their rates of change. K1, K2 and K3 represent the gain coefficients in the extended state observer, which are used to adjust the dynamic response speed and stability of the observer. α1 and α2 represent the nonlinear coefficients of the fal function, which control the intensity of the nonlinear function. δ1 and δ2 represent the threshold parameters in the fal function, which define the linear interval of the fal function, so that it has a linear response in the small error area. is the difference between z1 and x1, that is The fal function is nonlinear and is expressed as: Among them, e represents the estimation error or input error, α represents the nonlinear coefficient, and δ represents the threshold parameter.

7. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 6 is characterized in that: The method for establishing the nonlinear state error feedback controller NLSEF includes: Preset q d and They represent the desired joint angle and angular velocity respectively, and the control law is expressed as: Where e1 = q d -z1 represents the joint angle error, represents the angular velocity error, f3(e1) and f4(e2) represent fal(e1,α3,δ3) and fal(e2,α4,δ4) respectively, b1, b2 and b3 are gain coefficients; Presets and Combined with the extended state space equation, the state space equation of the error is expressed as:

8. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 7 is characterized in that: In S3, the method of establishing an outer loop-trajectory planner and adjusting the preset trajectory in real time according to the human-machine interaction torque and the predicted system state includes: Considering the human-robot system as a first-order spring-damper system, the expression is: Among them, D d and K d denote the desired damping and stiffness, ω d is based on τ t The expected gait period, ω r is the reference gait cycle designed by the therapist; A first-order filter is applied to the interaction torque via a discrete time difference equation, which is expressed as Among them, τ tf represents the filtered moment, k represents the current time step, α f is the filter coefficient, α f ∈(0,1); Using the filtered interaction torque τ tf and predicted system error, an adaptive damping method is preset; specifically: first, when the observer converges, the estimated error and Approaches zero, that is, the estimated states z1, z2 and z3 are approximately the system states. Represented as (z1,z2,z3) T , assuming that τ c Remains unchanged within the prediction range [k, k+N], where k represents the current time step, N represents the prediction duration, and Z(k+i|k) represents the system state at time k+i predicted at time k; Set k+i =q d (k+i)-z1(k+ik) represents the prediction error in the prediction domain, q d (k+i) is calculated based on the Fourier series. The Fourier parameters remain unchanged in the interval [k, k+N]. The overall prediction error le is expressed as: le=l1|e k+1 |+l2|e k+2 |+l3|e k+3 |, where l1, l2 and l3 are constants. When le is greater than the preset requirement, it indicates that there is a deviation between the current state and the expected trajectory, and it is necessary to increase the virtual damping to adjust the trajectory. In addition, when the interaction torque is greater than the preset requirement, the damping coefficient D is appropriately reduced to make the period parameter ω d Ability to achieve the patient's desired exercise state in a timely manner; Let ω max represents the maximum period parameter, ω min represents the minimum period parameter, and the final required period parameter is expressed as:

9. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 8, characterized in that: The predicted system state is expressed as: Z(k+ik)=A i Z(kk)+(A i-1 +A i-2 +...+I 6×6 )(Bτ+C) Where i∈[0,N], the matrix Expressed as in is a diagonal matrix with all diagonal elements equal to the sampling period t. is represented as matrix is set to is the identity matrix.

10. The control method of the lower limb rehabilitation robot system based on personalized gait generation and active compliance control strategy according to claim 8, characterized in that: Methods for increasing virtual damping and adjusting trajectory include: A Tanh-like function is preset to adjust the virtual damping d0 preset during training, expressed as: Where a, b, c and h are constants; the final human-robot interaction mapping is redefined as:

Images