Exoskeleton auxiliary system and gait prediction control method thereof

Abstract

The invention discloses an exoskeleton auxiliary system and a gait prediction control method thereof. The system comprises a hardware architecture module and a controller, the hardware architecture module is used for providing exoskeleton auxiliary devices, and the exoskeleton auxiliary devices comprise a first exoskeleton auxiliary device arranged on the waist and a second exoskeleton auxiliary device arranged on the hip joint; the first exoskeleton auxiliary device is a waist clamp wrapping the waist; the second exoskeleton auxiliary device comprises a motor and a thigh clamp hinged to the motor, and the motor and the waist clamp are hinged to the hip joint. According to an existing exoskeleton system, torque is generally measured through a torque sensor, so that the motion intention of a user is judged. However, the method does not need the equipment, so that the hardware composition of the exoskeleton system is greatly simplified. And a torque sensor is omitted, so that the weight of the exoskeleton is reduced, potential fault points are reduced, and the reliability of the system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of exoskeleton, and in particular to an exoskeleton assistance method and system based on a neural network. Background Art

[0002] The control of lower-limb exoskeletons plays a crucial role in determining the effectiveness of walking assistance, but determining control objectives remains a major challenge. Although neural networks are commonly used to generate control objectives, many existing methods involve offline training and classifiers, or rely on pre-made models, lacking the adaptability required to support different users and real-time scenarios with different gait cycles. Meanwhile, balancing the intervention of the human limb between compliance and assistiveness during training remains an open question. Summary of the invention

[0003] In view of this, the purpose of the present invention is to propose an exoskeleton-assisted gait prediction control method that can automatically adapt to changes in movement patterns.

[0004] According to one aspect of the present invention, there is provided an exoskeleton assistive system, comprising a hardware architecture module, a dynamic model module and a controller;

[0005] A hardware architecture module is used to provide an exoskeleton assistive device, including a first exoskeleton assistive device arranged at the waist and a second exoskeleton assistive device arranged at the hip joint;

[0006] The first exoskeleton auxiliary device is a waist clamp wrapped around the waist;

[0007] The second exoskeleton auxiliary device includes a motor and a thigh clamp hinged to the motor, and the motor and the waist clamp are hinged at the hip joint;

[0008] The controller is used for controlling an exoskeleton assistive device based on a motion model and adopting a neural network method.

[0009] In the above technical solution, according to the described exoskeleton-assisted gait prediction control method and device, compared with the existing exoskeleton-assisted system, its advantages can be summarized as follows:

[0010] 1. No torque sensor required: Existing exoskeleton systems usually require torque sensors to measure torque and determine the user's movement intention. This method does not require this device, which greatly simplifies the hardware structure of the exoskeleton system. Omitting the torque sensor not only reduces the weight of the exoskeleton, but also reduces potential failure points and improves the reliability of the system.

[0011] 2. The product is lighter: Since there is no need for torque sensors, the overall weight of the exoskeleton system is reduced, making it more comfortable for users to wear and reducing their burden. The lightweight design also improves the portability and wearability of the exoskeleton, allowing users to use it in more scenarios.

[0012] 3. Better adaptability: This method uses a radial basis function neural network model and combines it with an online learning method of selective memory recursive least squares to acquire and process data samples of exoskeleton-assisted gait in real time and quickly adapt to the user's gait changes. This adaptability enables the exoskeleton system to better match the user's movement needs and provide more accurate and effective assistance. Specifically, the Recursive Least Squares (RLS) algorithm is an effective online learning algorithm that can be used to train neural network models. In the training of neural networks, the RLS algorithm can quickly adapt to new data and update network weights, thereby accelerating the convergence of the model.

[0013] 4. Lower cost: The omission of hardware equipment such as torque sensors and the adoption of efficient online learning methods and control strategies reduce the manufacturing cost of the exoskeleton system. The cost reduction makes the exoskeleton assistive system more affordable, which helps promote and popularize it, allowing more people to benefit from this technology.

[0014] In summary, the described exoskeleton-assisted gait prediction control method and device achieve real-time walking assistance without the need for a torque sensor, and have the advantages of being lightweight, adaptable, and low cost. These advantages make the method and device have broad application prospects in the fields of assisted walking, rehabilitation therapy, etc.

[0015] In some embodiments, the hardware architecture module further includes a wearable vest, which is connected to the first exoskeleton assistive device to share the weight to the shoulders;

[0016] The waist clamp is provided with a support portion extending upward to the back at the lumbar vertebrae, and the support member is connected to the wearable vest.

[0017] In the above technical solution, the structure is connected to a wearable vest, and part of the weight of the LLE is redistributed to the shoulders through the wearable clothing, reducing the weight burden on the waist and legs. Further, in order to enhance the stability of the structure, a support portion is added, extending from the waist to the back and integrated into the wearable vest.

[0018] According to another aspect of the present invention, there is provided an exoskeleton-assisted gait prediction control method, comprising:

[0019] Acquire data samples of exoskeleton-assisted gait in real time;

[0020] Input the data sample into a pre-trained radial basis function neural network model, which adopts the online learning method of selective memory recursive least squares;

[0021] Gait prediction is output based on the radial basis function neural network model, and the gait prediction is used for exoskeleton-assisted control;

[0022] The regression operator and model parameter solution formula of the selective memory recursive least squares are as follows:

[0023]

[0024] P -1 (k) = P -1 (k-1)+Φ(x(k))Φ T (x(k))-Φ(γ a (k-1))Φ T (γ a (k-1))

[0025] Where W(k) is the weight vector of the radial basis function neural network. P(k) is the covariance matrix at sampling time k, and λ∈(0,1] is the forgetting factor in the recursive process. x(k) and y(k) represent the input sample and output sample respectively. a (k-1) and Represent the final synthesized input sample and output sample in the normalized range respectively. Φ() is the vector composed of radial basis functions of the regressor.

[0026] In the above technical scheme, considering the scenario of real-time lower limb exoskeleton assistance, it is necessary to quickly and accurately obtain the expected trajectory prediction after tΔ time. However, due to the forgetting factor λ, the classic RLS will forget the previous training samples in real-time training. When λ=1, the convergence speed becomes slower as new samples are learned. Therefore, a derivative method of RLS, SMRLS, is proposed. Compared with RLS, it has a selective memory term as an additional data input, and has a faster convergence speed and generalization ability. The use of selective memory recursive least squares in neural networks can quickly learn and accurately predict the movement intentions of human users without prior training.

[0027] In some embodiments, obtaining data samples of exoskeleton-assisted gait in real time also includes:

[0028] The data samples are normalized into a unit hypercube of the input space, and new data samples are synthesized. In the above technical solution, the purpose of normalization is as follows:

[0029] 1. Improve model training efficiency: Normalization can unify the numerical range of data samples into a standard interval, avoiding the numerical instability problem during training caused by large differences in data ranges. This helps to speed up the convergence of the model and improve training efficiency.

[0030] 2. Improve model prediction accuracy: Normalization can eliminate the dimension and scale differences in the data, allowing the model to more accurately capture the intrinsic relationship between the data, thereby improving the prediction accuracy.

[0031] Furthermore, the normalized and segmented samples are replaced by The reason is due to the following considerations:

[0032] 1. Simplify the data processing process: by directly replacing The steps of normalizing and segmenting the original data samples can be omitted, thereby simplifying the data processing process and improving real-time performance.

[0033] 2. Adapt to specific application scenarios: In some specific application scenarios, such as real-time lower limb exoskeleton assistance, the data processing speed and real-time requirements are high. This alternative method may be able to better meet these requirements while maintaining or improving the prediction performance of the model.

[0034] 3. Improve model generalization ability: By introducing a selective memory term as an additional data input and combining it with this alternative normalization method, it may help improve the generalization ability of the model, enabling the model to better adapt to different users and gait changes.

[0035] In summary, normalization plays an important role in data preprocessing, which can improve model training efficiency and prediction accuracy. This may be to simplify the data processing process, adapt to specific application scenarios, and improve the generalization ability of the model. These measures together help to achieve fast and accurate trajectory prediction, thus meeting the needs of real-time lower limb exoskeleton assistance.

[0036] In some embodiments, the neural network uses a weighted moving filter and uses supervised learning of data at the current moment and the previous moment.

[0037] In the above technical solution, the use of weighted moving filters not only reduces fluctuations, but also reduces prediction errors, because these slight fluctuations lead to poor training samples, which affects the overall training effectiveness of the entire data set from the perspective of convergence and generalization. Furthermore, supervised learning of the current and previous moment data is used to achieve simultaneous learning and prediction.

[0038] In some embodiments, outputting gait prediction based on a radial basis function neural network model further comprises:

[0039] Based on the gait prediction results, the balance conditions are defined for balance torque control and adaptive admittance control.

[0040] In the above technical solution, the control law based on adaptive admittance control and equilibrium state ensures stable assistance during the training process and provides intuitive parameter adjustment.

[0041] In some embodiments, based on the gait prediction result, the balance torque control is performed by defining a control law, specifically:

[0042]

[0043] Where ee0, ei0, and τi0 are specific coefficients and torques under corresponding equilibrium conditions.

[0044] In the above technical solution, after gait prediction, the system will define the balance condition according to the prediction result and perform balance torque control. The control law here usually involves some specific coefficients and torques, such as ee0, ei0, etc., which are determined according to the balance condition. In the assistance phase, the exoskeleton system needs to use torque τi to assist human movement. In order to ensure that the control law of the middle-level controller can adapt to both the assistance phase and the user's autonomous movement phase, the system needs a mechanism to smoothly adjust the interactive torque. Although defining a torque threshold can simply distinguish the two phases, this method may cause discontinuous torque adjustment, thereby affecting the user's comfort and safety. Therefore, the system adopts a more continuous and smooth adjustment mechanism. Specifically, the system may dynamically adjust the torque output according to the gait prediction results and the user's actual movement state. For example, when it is predicted that the user is about to perform autonomous movement, the system will gradually reduce the output torque of the exoskeleton to avoid interfering with the user's movement. When the user needs assistance, the system will gradually increase the torque output to provide necessary support. This adjustment mechanism may be implemented based on an algorithm or control strategy, such as fuzzy logic, neural network or other advanced control methods. Through these methods, the system can dynamically adjust torque output based on real-time feedback and predicted results, ensuring smooth transitions between different stages.

[0045] In some embodiments, based on the gait prediction result, adaptive admittance control is performed by defining a control law, specifically:

[0046]

[0047] In the formula, K Pe , K De are the coefficients of the PD controller.

[0048] In the above technical solution, adaptive admittance control is a control method that automatically adjusts control parameters according to environmental changes and user needs. In exoskeleton-assisted gait prediction, adaptive admittance control can dynamically adjust the admittance parameters according to the gait prediction results and the actual motion state of the user to achieve more accurate and comfortable assistance. Specifically, the system may adjust the value of the admittance parameter in real time according to factors such as the user's gait characteristics, movement speed, and load conditions. For example, when the user's walking speed increases, the system may increase the value of the admittance parameter to increase the response speed and assistance of the exoskeleton. When the user's walking speed slows down or requires more support, the system will reduce the value of the admittance parameter to provide softer and more stable assistance. Through adaptive admittance control, the exoskeleton system can better adapt to the movement needs and gait characteristics of different users, thereby providing more accurate and comfortable assistance effects. At the same time, this control method also helps to improve the stability and reliability of the system, ensuring the safety and comfort of users during use. In summary, by combining balanced torque control and adaptive admittance control, the exoskeleton system can dynamically adjust control parameters and torque output according to the gait prediction results and the user's actual motion state to achieve a more precise, comfortable and safer assistance effect.

[0049] According to another aspect of the present invention, an exoskeleton-assisted gait prediction device is provided, based on the above method; comprising:

[0050] An acquisition module, used to acquire data samples of exoskeleton-assisted gait in real time;

[0051] A neural network module for inputting data samples into a pre-trained radial basis function neural network model using a selective memory recursive least squares online learning method;

[0052] The processing module is used to output gait prediction based on the radial basis function neural network model, and use the gait prediction to perform exoskeleton assisted control.

[0053] In the above technical scheme, in order to better use the above method, the present application proposes an exoskeleton-assisted gait prediction device, each module corresponds to each step of the above method, and its specific principles have been described above and will not be repeated here.

[0054] According to another aspect of the present invention, there is provided an exoskeleton-assisted gait prediction device, characterized in that it comprises:

[0055] at least one processor and a memory communicatively coupled to the at least one processor;

[0056] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method as described above.

[0057] In the above technical solution, in order to better run and process the method, the above method is stored in a memory, and the stored method is executed by a processor. It should be noted that the principle and effect of each step have been described above and will not be described in detail here.

[0058] According to another aspect of the present invention, a computer-readable storage medium is provided, storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0059] In the above technical solution, in order to better run and use the method, the above method is stored in a computer-readable storage medium, and the above method is implemented by a processor. It should be noted that the principle and effect of each step have been described above and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 work.

[0061] Figure 1 It is a structural schematic diagram of an embodiment of an exoskeleton-assisted gait system of the present invention;

[0062] Figure 2 1 is a schematic diagram of an embodiment of an exoskeleton-assisted gait system of the present invention, wherein A is an overview of components of a wearable device LLE; B is a schematic diagram of an exoskeleton-assisted structure;

[0063] Figure 3 It is a schematic diagram of electromechanical connections of an exoskeleton-assisted gait system according to an embodiment of the present invention;

[0064] Figure 4 Schematic diagram of different angles and errors of an embodiment of an exoskeleton-assisted gait system of the present invention;

[0065] Figure 5 It is a flow chart of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention;

[0066] Figure 6 It is a schematic diagram of an example of normalization and division of an input space of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention;

[0067] Figure 7 It is a block diagram of a control system of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention;

[0068] Figure 8 The simulation results of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention in an open-loop system. In order to make the advance prediction time 100ms, KF 1 reduces the sampling rate, and KF 2 performs iterative prediction. (A) Prediction results when the prediction time is 100ms. (B) Comparison of prediction results of SMRLS and RLS when the advance time is 400ms. (C) Prediction error A in the figure (D) Prediction error B in the figure;

[0069] Fig. 9 This is an open-loop system experimental comparison of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention. (A) Prediction angle when the lead time is 120ms. (B) Prediction error, the semi-transparent curve represents the original data, and the opaque curve represents the data processed by the filter;

[0070] Fig.10 These are two motion cycles in a stair climbing scenario of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention. The left figure shows the first stage of stair climbing, and the right figure shows another stage of flat ground walking, while the subject's motion switches between the two modes. The motion in the CS stage includes a step that simulates the action of climbing stairs by going up and down stairs.

[0071] Fig.11 The results of different control methods of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention, where the parameters of ETC τi0=4NM and τi0=1.7NM are marked with blue and green background areas respectively (the curves are horizontally scaled to match different periods). (A) Results of ETC and classic AC at a larger assist setting. (B) Results of ETC at different balance torque settings, AC and PD control at lower assist settings;

[0072] Fig.12 is a schematic diagram of poor and disordered assistance using PD control in a larger assistance setting according to an embodiment of an exoskeleton-assisted gait prediction control method of the present invention;

[0073] Fig.13The results of SMRLS, RLS and ETC of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention are shown in Figure 1. Each figure describes the results from the first transition to the FG stage and then to the second transition. (A) The results obtained using SMRLS. (B) The results obtained using RLS. (C) The prediction errors of SMRLS and RLS in the example segment, where the semi-transparent curve represents the original data and the opaque curve represents the data processed by the filter.

[0074] Fig.14 It is an ideal auxiliary tool for AR used in the present case of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention;

[0075] Fig.15 is the relationship between the human hip joint angle and torque in the FG phase gait cycle from the prior art in an embodiment of an exoskeleton-assisted gait prediction control method of the present invention, wherein the unit Nm / kg represents the applied torque divided by the body weight;

[0076] Fig.16 It is the assistance rate (AR) in different method experiments of an embodiment of an exoskeleton-assisted gait prediction control method of the present invention, where (high) and (low) represent experiments with different assistance requirements. ETC, AC, and PD all use SMRLS except the RLS bar. The data of PD (low) is not shown because it does not work. DETAILED DESCRIPTION

[0077] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present invention, but are not intended to limit the scope of the present invention. Similarly, the following examples are only partial embodiments of the present invention rather than all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0078] The present invention provides an exoskeleton-assisted gait prediction control method, which can automatically adapt to changes in movement patterns.

[0079] Embodiment 1

[0080] See also Figure 1 , an exoskeleton assistive system, including a hardware architecture module and a controller;

[0081] 1) Hardware architecture module, used to provide exoskeleton assistive device, see Figure 2 , including a first exoskeleton assistive device 1 arranged at the waist, and a second exoskeleton assistive device 2 arranged at the hip joint;

[0082] The first exoskeleton auxiliary device is a waist clamp 1 that wraps around the waist;

[0083] The second exoskeleton auxiliary device 2 includes a motor 21 and a thigh clamp 22 hinged to the motor, and the motor 21 and the waist clamp are hinged at the hip joint.

[0084] In this embodiment, the hardware architecture module further includes a wearable vest 3, which is connected to the first exoskeleton assistive device 1 to share the weight to the shoulders;

[0085] The waist clamp 1 is provided with a support portion 11 extending upward to the back at the lumbar vertebrae, and the support is connected to the wearable vest 3 .

[0086] Specifically: The first exoskeleton assistive device at the waist consists of a circular structure that can rotate between 0° and 225°, with an embedded fixation. In addition, the structure is connected to a wearable vest, redistributing part of the weight of the LLE to the shoulders through the wearable garment. The second exoskeleton assistive device with 2 degrees of freedom is set at the hip joint, using a hinge-like connection for abduction / adduction movement. This design allows flexion and extension of the thigh, while also providing the necessary degrees of freedom for the complex movement of the hip joint during walking. The total weight of the exoskeleton device is approximately 2.42 kg, with an additional weight of 1.94 kg for batteries and circuits (a total of 4.46 kg). The waist fixed structure extends downward and is connected to the hinge of the motor, ensuring that the center of the motor is fixed at a height corresponding to the hip joint. As a result, the pivot of the motor is aligned with the hip joint instead of the waist, causing the motor to shift when receiving significant torque. This shift may affect the assistive performance and affect data collection. In order to enhance the stability of the structure, additional components are added, extending from the waist to the back and integrated into the harness.

[0087] In this example, see Figure 2 The motor used in this embodiment is from the MIT Cheetah robot, which has a built-in gear ratio of 9:1 and requires a 24V power supply. The stall torque and rated torque are 22 Nm and 11 Nm respectively, and the maximum speed and rated speed are 230 rpm and 130 rpm respectively. The motor then transmits data to the STM32C8T6, which is regarded as a low-level controller, via CAN communication of the TJA1050 module. Subsequently, the processed data is transmitted to the JetsonNano B01 via serial communication for training. The user carries a 24V power battery in the backpack with a maximum current of 30A. The power is then provided directly or indirectly to the microcontroller unit (MCU) and the actuator through an overcurrent protection module and a voltage DC-DC converter. The corresponding circuit structure is shown as follows Figure 3 shown.

[0088] 2) A controller for controlling the exoskeleton assistive device based on a motion model using a neural network.

[0089] In this embodiment, it is assumed that there is a relative motion between the human leg and the LLE (see Figure 4 ). In the absence of external interactions, they all ideally follow a feed-forward dynamic model.

[0090]

[0091] Where M, C, and G are inertia matrix, Coriolis matrix, centrifugal matrix, and gravity matrix, respectively. h represents the rotation angle of the human limb. More specifically, the human dynamics model can be viewed as a feedforward model with a PD feedback term that causes the change of the human limb motion.

[0092]

[0093] In the formula, e h represents the tracking error between the actual human limb angle and the expected angle, τ h is the torque exerted by the body.

[0094]

[0095] e i =q e -q h (5)

[0096] In the formula, K Ph , K Dh , K Pi and K Di is the coefficient of the PD controller, τ i is the interaction torque between the LLE and the human limb (applied from the LLE to the human limb), q e represents the angle of LLE, τ l is the torque applied by humans with the help of LLE.

[0097] Although q e can be obtained accurately, but because the human body's limbs are soft, it is difficult for the gyroscope to adapt to the size of different people's limbs, so q h The measurement is inaccurate. In addition, e i is small, so the data accuracy is not enough to meet the subsequent calculation requirements. Therefore, only q e Can be used in controller design to be detected by a gyroscope or the motor's internal encoder.

[0098] The PD controller applied to the motor of this device is as follows:

[0099]

[0100] e e =qde -q e (7)

[0101] In the formula, K Pe and K De is the PD gain, e e is the tracking error of the motor, q de To control the desired angle of the target, τ E is the offset torque of LLE. Therefore, the control problem of LLE becomes how to design the reference trajectory q of LLE de A possible solution is to make q de The trajectory of precedes the expected trajectory of human lower limb movement:

[0102] q de (t) = q h (t+t Δ ) (8)

[0103] In the formula, t Δ It is the time to predict in advance.

[0104] During the training phase when the output of the neural network is far from the expected trajectory, humans need to guide the movement of the entire system. It is difficult to change the movement of the LLE based on the result calculated by (6). Therefore, AC is applied to the system as

[0105]

[0106] Among them, Δq d =q de -q dn is the angle changed by AC, q dn is the output of the neural network prediction angle. Considering the accuracy of the sensor and the derivation in the next section, the inertia term of AC is ignored in equation (9). The entire system can be implemented as two spring damping models with AC, as Figure 4 shown.

[0107] Embodiment 2

[0108] See also Figure 5 , an exoskeleton-assisted gait prediction control method, the method is based on the exoskeleton-assisted system described in one of the embodiments, and is used to implement auxiliary control of a controller, comprising:

[0109] S1, real-time acquisition of exoskeleton-assisted gait data samples;

[0110] S2, inputting the data sample into a pre-trained radial basis function neural network model, which adopts an online learning method of selective memory recursive least squares;

[0111] RBFNN is a lightweight structured neural network with one hidden layer, and has universal approximation capability. It has relatively low computational requirements and is therefore widely used in various applications. This embodiment uses a Gaussian function as a radial basis function.

[0112]

[0113] Where i = 1, 2, ..., N, N is the number of neurons, is the input vector, For the center of each radial basis function, To represent φ i (χ) Standard deviation of the receptive field width.

[0114] The output of RBFNN can be described as

[0115]

[0116] In the formula, is the output, is the weight vector, is the radial basis function φ i (χ) is the regression vector.

[0117] For the scenario in this case, considering the training from t0 to t1, the optimal weight vector is

[0118]

[0119] Commonly used online learning methods for RBFNN include stochastic gradient descent (SGD) and recursive least squares (RLS). Although they use forgetting factors to improve sensitivity to new samples, the phenomenon of forgetting past data is inherent.

[0120] The regression operator and model parameter solution formula of selective memory recursive least squares are as follows:

[0121] W(k)=W(k-1)+P(k)Φ(x(k))[y(k)-W T (k-1)Φ(x(k))],

[0122] P -1 (k) = λP -1 (k-1)+Φ(x(k))φ T (x(k)),

[0123] φ(x(k))=[φ1(x(k)),φ2(x(k)),...φ N (x(k))] (13)

[0124] Where W(k) is the weight vector of the radial basis function neural network. P(k) is the covariance matrix at sampling time k, λ∈(0,1] is the forgetting factor in the recursive process, and x(k) and y(k) represent the input sample and output sample respectively.

[0125] Considering the scenario of real-time LLE assistance, it is necessary to quickly and accurately obtain the expected trajectory prediction after tΔ time. However, due to the forgetting factor λ, the classic RLS in the previous section will forget the previous training samples in real-time training. When λ=1, the convergence speed will become slower as new samples are learned. Therefore, a derivative method of RLS, SMRLS, is proposed. Compared with RLS, it has a selective memory term as an additional data input, and has faster convergence speed and generalization ability.

[0126]

[0127] P -1 (k) = P -1 (k-1)+Φ(x(k))Φ T (x(k))-Φ(γ a (k-1))Φ T (γ a (k-1)) (14)

[0128] Where W(k) is the weight vector of the radial basis function neural network. P(k) is the covariance matrix at sampling time k, and λ∈(0,1] is the forgetting factor in the recursive process. -1 (k) is the regression operator. x(k) and y(k) represent the input sample and output sample respectively. a (k-1) and Represent the final synthesized input sample and output sample in the normalized range respectively. Φ() is the vector composed of radial basis functions of the regressor.

[0129] In this embodiment, real-time acquisition of exoskeleton-assisted gait data samples also includes:

[0130] The data samples are normalized to a unit hypercube of the input space and new data samples are synthesized.

[0131] The input range of SMRLS-RBFNN is normalized to a unit hypercube of the input space, e.g., [-1, 1] for a one-dimensional input space. Figure 6 An example of input synthesis is shown, where the trajectory is represented as angular velocity map phase (ASP). The current samples are then divided into different lattices and synthesized as new reference samples for the sample pairs in the corresponding lattices. The subscript a=1, 2, ..., NP represents the lattice corresponding to the current sample. The synthesis of can be done in many ways. This case adopts an alternative method, directly replacing the normalized and segmented samples in (x(k), y(k)) with

[0132] In this embodiment, the neural network uses a weighted moving filter and supervised learning of the data at the current moment and the previous moment. The use of a weighted moving filter not only reduces fluctuations, but also reduces prediction errors, because these slight fluctuations lead to poor training samples, thereby affecting the overall training effectiveness of the entire data set from the perspective of convergence and generalization. Further, supervised learning of the data at the current moment and the previous moment is used to achieve simultaneous learning and prediction. Specifically:

[0133] In order to generate the initial control target, supervised learning of the current and previous angles and angular velocities is performed based on SMRLS-RBFNN. a Defined as the number of samples to predict, so the lead time of the prediction is n a Δt. For the LLE system in this case, x(k) is expressed as q e (kn a )and Indicates that they are the previous n a The angular velocity and angular velocity of the moment, and y(k) is expressed by q e (k) and Therefore, at the current moment, the mapping between the current sample and the current sample is trained in the neural network as

[0134]

[0135] Wherein the initial state of the angle, q = 0, corresponds to the angle observed in the natural standing or hanging posture of the leg.

[0136] At the same time, real-time calculations using forward filters often have data lag problems. In the training of neural networks, even small fluctuations can lead to significant training differences because the entire training data set is affected. In addition, data lag undoubtedly has a considerable impact on the effect of the auxiliary. This case uses a weighted moving filter, and using a training method involving forward prediction, the data before and after the filter can be collected and used, resulting in a better filtering effect on the input sample. f Defined as a (2n a +1)×2 matrix, including 2n a +1 the angle and angular velocity of the previous sample. The final filtering result at the current time is W f x f , where W f is the weight matrix, and the weight of the center sample is

[0137] In order to ensure that the weight distribution of the filters used is linearly proportional from the edge to the center, and their sum is equal to 1, we set the weight matrix The first element in is defined as

[0138]

[0139] W f Is positive, indicating Therefore, W f Each element in is defined as

[0140]

[0141] In the formula, and W f Once the center weight is determined is determined. In addition, the weighted moving filter integrates the information of one trajectory period into each training sample.

[0142] S3, outputting gait prediction based on the radial basis function neural network model, and using the gait prediction to perform exoskeleton assisted control;

[0143] The trajectory generated based on SMRLS-RBFNN is then adjusted by ETC in the middle-level controller. The control target is then sent to the STM32 C8T6 as a low-level controller. The structural block diagram of the control system is shown in Figure 7 shown.

[0144] 1. Balance torque control

[0145] After obtaining the prediction results through RBFNN, the controller also needs to explain q dn and to achieve the desired assistance. In the early stages of training or when the human wants to change their current motion, the LLE should learn and predict both the angular velocity and the angular speed in the high-level controller. Then the control target is generated for the low-level PD controller through the ETC in the middle-level controller. Minimize excessive intervention on the human. In the assistance phase, the LLE will assist the human with τi. Therefore, it is necessary to ensure that the control law of the middle-level controller can adapt to both stages at the same time, see Figure 7 .

[0146] Although defining a torque threshold can solve this problem, the adjustment of the interaction torque needs to be continuous and smooth. According to the model (3)(6)(9), there exists a condition that satisfies τ i =τ e -τ E The equilibrium condition is that the torque provided by the person exactly matches the resistance of the LLE, resulting in static equilibrium.

[0147]

[0148] In the formula, e e0 、e i0 , τ i0 are the specific coefficients and torques under the corresponding equilibrium conditions. ne =q dn -q e is defined as the initial forward angle. i The concept of τ is different. i0 should be defined as a specific scalar quantity without direction. However, for the following expression, the reference torque direction is defined as sgn(Δq ne ), τ i0 The value of actually represents |τ in the following derivation i0 |sgn(Δq ne ). When the interaction torque exceeds the equilibrium torque, the direction of motion of the LLE changes, τ 00 is the torque threshold.

[0149] 2. Adaptive admittance control

[0150] Forward angle Δq ne The difference varies with speed in different torques. In addition, when the leg is to move in the opposite direction, this difference will inevitably decrease. Therefore, the training samples of the trajectory are constantly changing with the constant amount of communication and the relatively small coefficient, and humans need to be constantly retrained. In this case, the parameters of the communication variables based on (4), (6), and (9) are required.

[0151] In order to solve the interaction torque requirement problem in AC with uncertain forward angle, the interaction torque τ i0 For a certain equilibrium state, Δq d has the same effect. In this sense, τ i0 It can also be regarded as the reference tolerance of the interaction. Once the interaction torque q i Reach τ i0 , the limb and LLE will reach a state of equilibrium.

[0152] ρ e0 , D A0 , K A0 、e e0 , Δq d0 is defined as the corresponding parameter in the equilibrium state. If e is not detected i , once τ i0 is defined, we have:

[0153]

[0154] However, there are six unknown variables. de0 -q dn and q de0 -q e Replace e e0 and Δq d0 After that, there are still four unknown variables, such as q de0 , D A0 , K A0 For ease of calculation, the internal coefficients in (4), (6), and (9) can be kept in the same proportion as follows:

[0155]

[0156] In the formula, ρ a Human-LLE interaction model and ρ a K Pe =D A , a K De =K A Have the same proportions.

[0157] Therefore, there are three unknown variables ρ a , D A0 and K A0 , which can be solved as

[0158]

[0159] However, when τ i0 Exceed When a It may be equal to or even less than 0, resulting in D A0 and K A0 becomes infinite or negative. If is considered as targeting τ i0 The virtual torque τ n , then the boundary is defined as:

[0160]

[0161] The coefficient ρ e ∈(1,+∞) defines τ n and τ i0 In addition, in order to prevent equipment damage, it is necessary to determine ρ a The minimum value of .

[0162] In the above system, K Pe and K De It also limits D A and K A The higher the parameter, the greater the value of D A and KA From a physical point of view, it is more convenient to design a smaller motor control parameter K. Pe and K De aims to provide less help, which means that there is less interaction force in the corresponding scene. In this case, by τ i It is reasonable to have a smaller admittance parameter to facilitate easier trajectory changes. By introducing ρ a Instead of D A and K A To limit their variation range, it is directly ensured that the variation range of the admittance coefficient is adaptive without the need for manual adjustment by observing experimental data.

[0163] Based on the content of this embodiment, the description part provides the test results of the model simulation part and the experimental part, which are as follows.

[0164] 1. Simulation

[0165] In the simulation, the motion of the human body is then generated by (2). The two reference trajectories are defined as:

[0166]

[0167] Where T = 25 is the total period of the larger cycle, and n = 1, ..., n. The change in motion between the two trajectories in qr simulates the scene of continuously ascending stairs. According to the research of the prior art, the gait frequency range of humans during walking (per minute) is generally 52 to 82 Hz. The main frequency of both trajectories is 60 Hz, which is consistent with this. The weighted moving filter is not implemented in the simulation because the data is ideal.

[0168] Other parameters are determined as follows: the leg is the mass point, m h = 12kg, LLE mass is m e = 0.5 kg. The length of the leg l = 0.5 m (between the central pivot and the mass point).

[0169] Other control parameters are set as follows: K Ph =200, K Dh =50, K Pe =32, K De =8, K Pi =100 and K Di = 25. The angular acceleration is calculated from the angular velocity through the transfer function:

[0170] For the RBFNN setup, the number of nodes is 25 (5x5 for angle and velocity respectively). In SMRLS there are also 100 precision data points for angular velocity and angular speed respectively, which means it is divided into 10000 lattices for previous recording. In order to avoid the negative impact of the instability of the neural network data during the early training process, it only produces output to the system when t>10s. The forgetting factor λ of RLS is 1.

[0171] We considered the following trajectory prediction methods for comparison:

[0172] Method 1: The proposed method and SMRLS-RBFNN.

[0173] Method 2: RLS-RBFNN.

[0174] Method 3: Use Kalman filter (KF) for model prediction, where the measurement matrix is The transfer matrix is The state vector is

[0175] Method 4: Iterative Learning Control (ILC). Each cycle of ILC is set to 1s, and the sampling frequency is set to 50Hz (20ms).

[0176] 1. Open-loop system

[0177] This subsection shows the results of one-way training without interaction feedback to the control system. The interaction torque gradually converges, but the trajectory enters the next sub-period before fully convergence and cannot work successfully. Figure 8 As shown in Figure A, ILC fails to properly converge to the nested periodic variations, even though such variations are very common in normal motion. Due to the relatively high learning gain, the fluctuations can be clearly observed even with a second-order moving average filter.

[0178] For the fixed model using KF, we tried two different approaches. The first approach was to keep the same sampling period as the other methods (20ms), but iterate the prediction 5 times to get the same lead time (100ms). The second approach was to set the sampling period exactly equal to the lead time. It is clear that neither approach can predict the motion trajectory very well. This is partly due to the lag problem caused by the Kalman filter itself. In addition, due to the limitations of the fixed linear model, it lags at fast speeds and overestimates at slow speeds. The tracking error is as follows Figure 8 As shown in C, the errors of the remaining methods are larger than those of RLS and SMRLS.

[0179] Although SMRLS outperforms RLS at a lead time of 100 ms, the difference is not significant at a lower lead time (100 ms). Figure 8 Figure B shows the difference at a higher lead time (400ms). In the first transition (13.5-17s), SMRLS converges faster, even though the sampled motion is new to both neural training methods, suggesting better generalization of our proposed method. In the subsequent first motion return (26-29.5s), although the RLS method converges slightly better at the first peak, it never fully converges to the reference trajectory and converges significantly slower in subsequent iterations. The tracking errors of the two methods are shown in Figure 2. Figure 8 As shown in d, it shows that SRMLS increases in advance and performs better than RLS.

[0180] 2. Experiment

[0181] In this section, only RLS is used for comparison, as other prediction methods clearly show poor availability in simulations and may harm the LLE device. Similar to the simulation, the neural network provides output after 10 seconds of training in the early training, while the actuator does not provide assistance until 12 seconds. The following sampling, communication, and calculation frequencies are all set to 66.7Hz (Δt=15ms). Similar to the simulation, the number of nodes in the RBFNN is 25 (5x5), and the precision used in SMRLS is 100 data points. In order to demonstrate the adaptability of the method, all datasets in the experiment are not pre-trained, that is, the initial weight vector W of the RBFNN is set to zero. The forgetting factor λ of RLS is 1. The normalized range of angles is set to [-0.2,1.2], while the normalized range of angular velocity is set to [-2,2].

[0182] During the initialization of the program, the subjects were asked to maintain a natural standing posture with their legs perpendicular to the ground. The angle of the motor drive was then calibrated to 0.

[0183] 1. Scene setup:

[0184] This experiment focuses on the scene of climbing stairs. Normally, humans do not maintain a constant state of climbing stairs, but alternate between climbing stairs (CS phase) and walking on flat ground (FG phase) after climbing a flight of stairs. This movement involves a nested periodic motion that requires a continuous transition of lower limb movements.

[0185] For evaluation purposes, the staircase in this experiment was simulated by one step and had a height of 17.5 cm, as Figure 8As shown. The experiment focuses on the movement of one leg, the subjects will repeatedly lift this step using their left leg, mimicking the downward movement of one leg during the CS phase, and then transfer the center of gravity to the right leg while bending the left leg to simulate the lifting movement when ascending the stairs. Throughout the process, the contact position between the left leg and the stairs will remain unchanged. After completing several cycles of these movements, the subjects will leave the steps, walk a short distance on a flat surface, and then return to the steps again to start the next cycle.

[0186] 2. Open-loop system

[0187] This section shows the offline training results of SMRLS and RLS, where the training samples are collected from a closed-loop experiment. Compared with simulation, even in the same stage, the samples are perturbed and noisy, which requires better generalization and faster adaptation. However, the normalization range of RBFNN is set to the same as the input sample range as offline training, which is different from the closed-loop experiment. The number of samples for advance training na is 8, that is, the advance prediction time is 120ms.

[0188] Comparison between forward filter and weighted moving filter is shown in Fig. 9 In this experiment, Fig.10 As shown in the figure, the subject first climbed the stairs until the system reached a relatively stable state, and then transitioned to another movement. The approximate time intervals of the CS phase were: 0-40s, 58-75s, and 87-103s, and the time intervals of the FG phase were: 40-58s, 75-87s, and 103-120s. Fig. 9 As can be seen in Figure A, while the convergence of all methods gradually improves with training after switching motion, RLS performs worse. In the initial transition, the superior prediction accuracy of SMRLS can be attributed to its better generalization. However, at the first switch back, a clear gap between the two methods is observed, indicating that our method benefits from its selective memory ability. For the derivative method, the use of a forward filter produces a significant phase difference. The forward filter introduces data lag, and when used for training input samples, it effectively increases the time interval between input and output samples. Therefore, when the trained mapping is used, it results in a more advanced output stage from the corresponding input. After a motion transition, the prediction error may increase suddenly, but SMRLS can quickly reduce the prediction error to a lower magnitude, see Fig. 9 B. The average error after 20 seconds of program execution is shown in the table. I.SMRLS method with forward filter is even worse than RLS, mainly due to its phase difference. The results show that the weighted moving filter not only reduces fluctuations but also reduces the prediction error, because these slight fluctuations lead to poor training samples, thus affecting the overall training effectiveness of the whole dataset from the perspective of convergence and generalization.

[0189] 2. Assisted stair climbing (closed loop system)

[0190] This subsection will demonstrate the overall assistive performance of the entire system. The strengths and weaknesses of the approach are primarily demonstrated through the interaction torque. The derived ETC will also be compared with fixed parameter AC and PD control. Experiments are performed using different PD control parameters, including larger and smaller settings. This is used to elucidate the consequences of different magnitude requirements for assist torque, thus providing insights into the performance of the system under different conditions.

[0191] 1) Setting: The number of samples for advance training na is 12, that is, the advance prediction time is 180ms. As described in the open-loop experiment description, the first switch of the movement mode occurs when the subject perceives that the movement mode is relatively stable. Due to the limitations of ETC in all experiments, ρe=1.01 and ρmin=0.4.

[0192] When lowering the legs at a slow speed, the proposed method will not provide assistance but introduce resistance. However, this problem is often alleviated in stair climbing scenarios. Therefore, in this experiment, the distance when lowering the legs to contact the steps is ignored, and it is assumed that a person immediately contacts the steps when lowering the legs. For ideal assistance in ordinary walking (FG phase), the interaction force can only move backward when in contact with the ground and provide forward assistance in other phases. But the contact with the ground in the FG phase is relatively small, which means that the movement of the suspended leg cannot be ignored.

[0193] The magnitude of the auxiliary torque is mainly affected by K Pe and K De In this experiment, the performance of ETC, AC and PD control strategies is studied under relatively high auxiliary parameters, where K Pe =45, K De =5 and τ i0 =4.0Nm. For AC strategy, D A =20 and K A = 2.5, and the PD control strategy can be assumed to be an AC strategy with infinite parameters. For relatively small auxiliary programs, K Pe =20, K De =3. The parameters of AC power and AC power under high auxiliary conditions (D A =20 and K A =2.5) remains the same. In both cases, the balancing torque of ETC is set to τ i0 =4.0Nm and τ i0 =1.7Nm, indicating that the maximum expected auxiliary torque is 4.0Nm and 1.7Nm respectively.

[0194] 2) Comparison of interactive torque and control law: In order to highlight the contribution of this case to the control law, the results of using ETC, AC and PD control in the CS stage are shown in Fig.11 For a clear presentation, the torque curves are scaled in the time domain to normalize the different cycles. Fig.11 Figure A shows the results of using ETC and AC under higher assistance conditions. It is obvious that ETC provides smoother results, which indicates that the assistance is more stable. This is also related to its better adaptability. For example, the fixed parameters of AC used in this case are relatively low, which requires greater assistance, and smaller forces will lead to larger Δqd (change in qde), which is prone to more fluctuations. Due to the learning of RBFNN, those samples with poor fluctuations will be sent back to learning, becoming a vicious cycle. In addition, the current trajectory q e , RBFNNq dn Predicted trajectory and interaction torque τ using classical PD control i The results are as follows Fig.12 As shown, it is shown that PD control cannot even work successfully in this case. In PD control, a higher parameter represents a larger upper limit of the assist torque, but this requires the motion compliance of the AC. Although the training of the RBFNN is generally completed within 12 seconds and provides acceptable assistance within two initial steps (12-14 seconds), the actions may be different even in the same movement phase, which leads to poor assistance and stiffness.

[0195] However, these fixed parameters and PD control work better at lower assist demands, e.g. Fig.11 As shown in B, this also shows that τ i0 =4.0Nm and τ i0 =1.7Nm ETC result. Without adjusting other parameters, the magnitude of torque interaction decreases as the PD parameter decreases, while τ i0 = 4.0Nm ETC still maintains 4.0Nm of interaction torque as before. Although PD control is Fig.11 The value of A is smaller than that of A, but it also limits the torque to 4.0Nm as the upper limit. i0 = 1.7Nm ETC generally limits the interactive torque to 1.7Nm, indicating that ETC is easier and more intuitive to adjust the auxiliary torque and can better adapt to changes caused by other settings.

[0196] 3) Comparison between SMRLS and RLS: The results of SMRLS and RLS tend to reach a relatively stable state after a period of training, such as Figure 8 As shown in the simulation of Figure A, the convergence speed is fast. Although the prediction error of RLS is larger after convergence, this slight difference is not obvious in the experiment. Therefore, the difference between SMRLS and RLS is mainly reflected in the early stage when learning is not complete.

[0197] Fig.13 An example of the transition from the first to the second transition of the SMRLS and RLS segments is depicted. The corresponding prediction error obtained by processing with a mean moving filter with a window size of 150 is shown in Fig.13 As shown in C. The prediction error of RLS in the experiment is significantly greater than the simulation result. In addition, the prediction error of RLS increases significantly during the transition process of different stages, and it is even indistinguishable from the results of SMRLS. The average prediction error after the second cycle remains within 0.1rad (6°).

[0198] SMRLS resolves assistance after the first 6 steps of the first transition (after 33 seconds), with the interaction torque starting to provide assistance, while the results of RLS still fail to resolve even when approaching the first return to the initial movement mode. When the prediction still reaches higher angles, the overall direction of the assistance tends to be in the positive direction, similar to the movement of CS rather than FG. Similarly, it fails to converge effectively during the first transition to the CS phase, while SMRLS usually converges after 3 steps (after 44 seconds). Similar results can be found during the second transition to the FG phase.

[0199] Furthermore, when the second transition to the CS phase was performed using RLS, a significant decrease in the maximum amplitude of the subjects’ leg lifting could be observed compared to the results using SMRLS, which means that the method affected the subjects’ native leg movements instead of providing assistance in reverse. In this case, according to our experiment, the subjects subconsciously tilted their trunks backwards while walking. It is also obvious that the subjects tried to resist this poor assistance and took extra efforts to correct it, preventing it from continuing to affect subsequent training as a defective sample. However, it resulted in an overall bias of the assistance backwards, causing the assistance to act as resistance with negative torque during the leg lifting phase, similar to the shortcomings of AC. If the subjects tended to fix it, it would have produced worse assistance. In other words, the subjects had to either tolerate the inappropriate changes imposed by LLE or exert extra efforts to correct this problem.

[0200] 3. Assistance rate

[0201] This experiment introduces the assistance rate (AR) to quantify the proportion of correct assistance actions under different control methods. If the direction of the interaction torque is consistent with the corresponding theoretical direction mentioned above, it is considered to be a beneficial assistance effect. Therefore, in the context of the CS phase, the correct assistance can be defined as follows: 0-100% (positive) in the leg lifting phase, 0-0100% (negative) in the leg lowering phase. FG phase: 0-20% and 85-100% in the back swing phase, 0-100% (positive) in the front swing phase, and 20-100% (negative) in the back swing phase. Therefore, 20% is used as the standard in this case. However, a certain 10% phase difference is allowed between the interaction torque and the ideal point. When the phase is leading, it can be regarded as an exoskeleton that provides torque to resist the inertial effect and change the current state. When the phase lags, it can be regarded as an exoskeleton that follows the user's intention and provides assistance. Both forms successfully contribute to providing assistance. Due to the significant uncertainty and interference in the data, the segmentation of each phase will be manually labeled. The stance phase (foot contact) accounts for 62% (time) of a single gait cycle on average. Therefore, during normal walking, more torque should be directed toward forward movement. In this case, the ideal assistance should be as follows Fig.14 shown.

[0202] Fig.15 The gait analysis diagram from the prior art shows an example of data of hip joint angle and torque in a gait cycle of normal human walking (FG phase) based on clinical gait analysis. Although the definition of AR is slightly different from the example provided, the curve trend of AR defined below is basically consistent with the curve of the experimental results, which confirms the rationality of the comparison in this case.

[0203] Fig.16 An example of how AR is calculated in a weaker assistance condition is shown, where the reversal torque of the interaction torque has been marked on the respective trajectories. It is clear that according to AR, ETC provides a more appropriate portion of assistance (green part) compared to AC, where the interaction torque with AC is reduced earlier while the leg is still lifting. Control target q de and It is easy to be changed by a smaller torque because the predicted front angle and angular velocity are reduced, although this torque is necessary to help. Although PD control has better AR, it cannot even provide assistance under larger assistance conditions, such as Fig.12 As shown, this means that the good performance of PD control is accidental. In addition, the low PD parameter leads to a low upper limit of the assist torque, while other control methods usually require a larger margin for the assist torque. For example, in this case, it has limited the maximum assist torque of ETC, although ETC is still good.

[0204] Finally, the ARs for the different experiments above are Fig.16The results are marked in , starting from 30 seconds after operation. The results show that ETC produces more appropriate assistance compared to AC. Although the PD controller performs well at lower assistance, it does not work at higher assistance and is therefore not included in Fig.16 For ordinary walking, the movement of the human legs is largely influenced by inertia and gravity. Therefore, a large part of the human effort is devoted to maintaining standing balance. This can also be achieved by the knee and ankle joints. According to the subjects, they experienced significant assistance when support was provided during the middle phase of the forward swing and the initial moment of ground contact. In other phases, even if the direction of assistance differed from the above-mentioned AR definition, the subjects did not experience noticeable effects as long as the resistance was not too great.

[0205] The AR in this case only considers the direction of the torque, not the magnitude. Fig.11 In A, the interaction with AC is clearly unstable, but there is no great difference in AR. Therefore, AR generally proves that ETC and SMRLS are better than other methods, but there are still many advantages that cannot be intuitively observed through the current results.

[0206] In this paper, a real-time learning-based walking assistance strategy based on SMRLS-RBFNN is proposed. A control strategy using trajectory prediction is used to balance the interactive torque to achieve more stable and controllable assistance. Compared with many existing exoskeleton control strategies, the proposed strategy does not use classifiers for advanced control. Instead, it relies on enhanced adaptability achieved by optimizing the neural network training method. In addition, compared with many complex real-time training methods, the settings and neural network structures involved in this strategy are relatively simple. This simplicity greatly reduces the data processing time, thereby greatly improving the control and communication frequency, which is a key factor in real-time training and assistance. According to further testing, SMRLSRBFNN can not only operate in the standard 5V / 4A power mode of Jetson Nano, but also successfully operate in the low power mode of 5V / 2A.

[0207] In simulation, we compare with other trajectory prediction methods that do not rely on neural networks. These methods include methods based on fixed models (KF) with external optimization, and methods involving continuous training with iterative and recursive properties (ILC). Compared with other methods, the strategy proposed in this case has superior adaptability in handling specific scenarios. Comparisons are made with the classic RLS method in both simulation and experiments. The results show that SMRLS-RBFNN can better adapt to previously trained data and show better generalization. Therefore, it converges faster when changing the motion. Although the scenario described in this case is ascending stairs to emphasize the adaptability of nested loops and faster and more efficient convergence of real-time training, the proposed strategy is not limited to this scenario. Although the gait advance prediction method may not be applicable to situations that require individuals to perform negative work, such as descending stairs, where resisting leg weight or maintaining balance is necessary, it can still promote the increase of human step frequency. The definition of AR is based on the assumption of ideal situations, but the experimental results are generally consistent with the gait analysis results of the literature.

[0208] Embodiment 3

[0209] An exoskeleton-assisted gait prediction device, comprising:

[0210] at least one processor and a memory communicatively coupled to the at least one processor;

[0211] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method described in the second embodiment.

[0212] In the above technical solution, in order to better run and process the method described in the second embodiment, the above method is stored in a memory, and a processor is used to execute the stored method. It should be noted that the principle and effect of each step have been described above and will not be described in detail here.

[0213] Embodiment 4

[0214] A computer-readable storage medium stores a computer program, which implements the method described in the second embodiment when executed by a processor.

[0215] In the above technical solution, in order to better run and use the method described in the second embodiment, the above method is stored in a computer-readable storage medium, and the above method is implemented using a processor. It should be noted that the principle and effect of each step have been described above and will not be described in detail here.

[0216] The above descriptions are only some embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent device or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An exoskeleton-assisted gait prediction control method, characterized in that: The method comprises: Acquire data samples of exoskeleton-assisted gait in real time; Input the data sample into a pre-trained radial basis function neural network model, which adopts the online learning method of selective memory recursive least squares; Gait prediction is output based on the radial basis function neural network model, and the gait prediction is used for exoskeleton-assisted control; The regression operator and model parameter solution formula of the selective memory recursive least squares are as follows: P -1 (k) = P -1 (k-1)+Φ(x(k))Φ T (x(k))-Φ(γ a (k-1))Φ T (γ a (k-1)) where W(k) is the weight vector of the radial basis function neural network; P(k) is the covariance matrix at sampling time k, λ∈(0,1] is the forgetting factor in the recursive process; P -1 (k) is the regression operator; x(k) and y(k) represent the input sample and output sample respectively; a (k-1) and Represent the last synthesized input sample and output sample within the normalized range respectively; Φ() is the vector composed of radial basis functions of the regressor.

2. An exoskeleton-assisted gait prediction control method as claimed in claim 1, characterized in that: Real-time acquisition of exoskeleton-assisted gait data samples, including: The data samples are normalized to a unit hypercube of the input space and new data samples are synthesized.

3. An exoskeleton-assisted gait prediction control method as claimed in claim 1, characterized in that: The neural network adopts a weighted moving filter and uses supervised learning of data at the current moment and the previous moment.

4. The exoskeleton-assisted gait prediction control method according to claim 1, characterized in that: Based on the radial basis function neural network model output gait prediction, it also includes: Based on the gait prediction results, the balance conditions are defined for balance torque control and adaptive admittance control.

5. The exoskeleton-assisted gait prediction control method according to claim 1, characterized in that: Based on the gait prediction results, the balance torque control is performed by defining the control law, specifically: Where ee0, ei0, and τi0 are specific coefficients and torques under corresponding equilibrium conditions.

6. The exoskeleton-assisted gait prediction control method according to claim 1, characterized in that: Based on the gait prediction results, adaptive admittance control is performed by defining the control law. Specifically: In the formula, K Pe , K De are the coefficients of the PD controller.

7. An exoskeleton assistive system, characterized in that: An exoskeleton-assisted gait prediction control method according to any one of claims 1 to 6, comprising a hardware architecture module and a controller; A hardware architecture module is used to provide an exoskeleton assistive device, including a first exoskeleton assistive device arranged at the waist and a second exoskeleton assistive device arranged at the hip joint; The first exoskeleton auxiliary device is a waist clamp wrapped around the waist; The second exoskeleton auxiliary device includes a motor and a thigh clamp hinged to the motor, and the motor and the waist clamp are hinged at the hip joint; The controller is used for assisting the control of an exoskeleton assistive device based on a motion model and adopting a neural network method.

8. An exoskeleton assist system as claimed in claim 7, characterized in that: The hardware architecture module also includes a wearable vest, which is connected to the first exoskeleton assistive device to share the weight to the shoulders; The waist clamp is provided with a support portion extending upward to the back at the lumbar vertebrae, and the support member is connected to the wearable vest.

9. An exoskeleton-assisted gait prediction device, characterized in that: include: at least one processor and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method as claimed in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images