A human-computer interaction control method for a lower limb exoskeleton robot

Abstract

The application discloses a human-computer interaction control method of a lower limb exoskeleton robot. The method comprises the following steps: collecting surface electromyography data of a target; inputting the surface electromyography data into a trained motion intention decoding network to output a corresponding expected gait trajectory, wherein the motion intention decoding network comprises a generative adversarial network module, a graph neural network module and a Transformer module; and tracking the expected gait trajectory by using a designed robust following controller to ensure that the tracking error of the lower limb exoskeleton robot in the joint space is bounded. The application can generate a large amount of surface electromyography data under a small sample condition, and can obtain the features of local and global surface electromyography data through a closely coupled graph structure and a self-attention mechanism combining time domain and space domain, so that the complexity of sample collection is reduced, and the accuracy of human-computer interaction control is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robotics, and more particularly, to a human-robot interaction control method for a lower limb exoskeleton robot. BACKGROUND

[0002] Exoskeleton robots can be used in the field of medical rehabilitation, such as assisting the wearer to move. Taking lower limb exoskeletons (LLEs) as an example, they combine human mechanisms and mechanical properties closely to provide active balance and walking ability for the wearer. When assisting the wearer to walk, the lower limb exoskeleton robot needs to have the ability to perceive the human motion intention, improve the active participation of the wearer, and achieve excellent human-robot interaction performance. How to achieve natural, accurate and stable human-robot interaction control when the wearer wears the exoskeleton has not been completely solved at present. Therefore, it is of great significance to enhance the flexibility of LLE movement and improve the walking ability of LLE by realizing perfect human-robot interaction through advanced intelligent control.

[0003] Quick and accurate recognition of human motion intention is a prerequisite for efficient human-robot interaction control of LLEs. Using surface electromyography (sEMG) signals of the human body to detect human motion intention is an ideal signal source commonly used in the field of exoskeletons. sEMG signals contain rich human motion information corresponding to the motion state of the human body. sEMG has high accuracy and real-time performance, and the pre-position time of the human body motion is 30-150 milliseconds, which is conducive to integration into the real-time controller of the exoskeleton. Therefore, human-robot interaction recognition based on sEMG signals has gradually become a research hotspot in the field of exoskeletons.

[0004] In the human-robot interaction control of LLEs, the exoskeleton is endowed with the ability to decode human motion intention to generate corresponding motion trajectories. By designing a stable controller to control the LLE and achieve the required motion, safe interaction with the exoskeleton can be achieved.

[0005] There are currently various human-robot interaction control methods for lower limb exoskeleton robots. For example, there is a scheme that estimates motion intention and joint stiffness based on sEMG, and proposes an impedance-based adaptive controller for human-robot collaborative control of LLEs. There is a scheme that proposes a double-layer admittance control based on sEMG, which enables the wearer to actively adjust the gait of the LLE, and is verified by six subjects on a treadmill, and the subjects can continuously and stably adjust the gait trajectory.

[0006] Although existing research on exoskeleton human-computer interaction control has achieved human-computer interaction recognition and control based on sEMG, these methods are only applicable to unilateral LLE control and have not yet been validated for bilateral LLE. Furthermore, existing sEMG data acquisition methods are costly, hindering the widespread application of LLE, or lack robustness analysis for walking, failing to validate robustness against the subject and terrain, leading to unnatural human-computer interaction control. In summary, current research on exoskeleton human-computer interaction technology has limitations, including the need to collect large human-computer interaction datasets, poor versatility of human-computer interaction control, and a lack of high-precision human-computer interaction control technology. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a human-computer interaction control method for a lower limb exoskeleton robot. This method includes the following steps:

[0008] Collect surface electromyography data of the target;

[0009] The surface electromyography data is input into a trained motion intention decoding network, which outputs the corresponding desired gait trajectory.

[0010] The designed robust follower controller is used to track the desired gait trajectory to ensure that the tracking error of the lower limb exoskeleton robot in joint space is bounded.

[0011] The motion intent decoding network includes a generative adversarial network module, a graph neural network module, and a Transformer module. The Transformer module includes a temporal layer and a spatial layer. The temporal layer is used to process time-related features in surface electromyography (EMG) data and includes a temporal embedding layer, a temporal multi-head attention mechanism layer, and a first feedforward network. The spatial layer is used to capture the spatial distribution features of EMG data and includes a spatial embedding layer, a spatial multi-head attention mechanism layer, and a second feedforward network. In the temporal layer, the EMG data of each channel is used as a node in the graph structure of the graph neural network module. The data of each node is decomposed into a set of intrinsic mode functions as features of the corresponding node. The attention score obtained using the graph structure is passed to the temporal multi-head attention mechanism layer to learn the relationship between sequences.

[0012] Compared with existing technologies, the advantages of this invention lie in proposing a novel LLE (Low-Leg Exoskeleton) human-computer interaction control framework. This framework features a robust tracking controller that decodes the human motion intention network online, considers online gait generation and stable walking control, and enables the wearable exoskeleton to walk on various terrains. Furthermore, in the LLE human-computer interaction control framework proposed in this invention, a large amount of sEMG data can be generated using a transformer network based on a generative adversarial network-graph structure. Local and global sEMG features are obtained through a tightly coupled graph structure and a self-attention mechanism combining temporal and spatial domains, thereby reducing the amount of sample sEMG data required and improving the accuracy of human-computer interaction control.

[0013] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0014] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0015] Figure 1 This is a flowchart of a human-computer interaction control method for an exoskeleton robot according to an embodiment of the present invention;

[0016] Figure 2 This is a human-computer interaction control framework diagram for decoding human movement intentions according to an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the overall framework of a transformer network (GANGT) based on a generative adversarial network-graph structure according to an embodiment of the present invention;

[0018] Figure 4 This is a schematic diagram of a lower limb exoskeleton robot hardware system according to an embodiment of the present invention;

[0019] Figure 5 This is a schematic diagram of GAN-enhanced sEMG data for three test sets according to an embodiment of the present invention;

[0020] Figure 6 This is a schematic diagram of the motion trajectory and tracking results according to an embodiment of the present invention. Detailed Implementation

[0021] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0022] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0023] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0024] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0025] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0026] In summary, this invention designs a graph-based transformer network and introduces a generative adversarial network (GAN), or collectively referred to as a generative adversarial network-graph structure transformer (GANGT), for application under conditions of small sample surface electromyography (sEMG) datasets. Adversarial learning of GANs is used to generate feature-rich sEMG data, and a graph-based temporal and spatial self-attention mechanism is constructed to capture local and global features of sEMG signals during gait cycles. The designed GANGT achieves accurate end-to-end decoding of human movement intentions. Furthermore, a robust follower controller is designed to control tracking errors in predicted joint angles. Even with unmodeled errors and external disturbances in the exoskeleton dynamics system, the designed controller can keep the tracking error within a small range.

[0027] Specifically, see Figure 1 As shown, the human-computer interaction control method for the provided lower limb exoskeleton robot includes the following steps:

[0028] Step S110: Train the decoding motion intent network to obtain the corresponding expected gait trajectory based on the target-based sEMG data.

[0029] Figure 2 This is a schematic diagram of a human-computer interaction control framework for decoding human movement intentions, in which... Figure 2 (a) illustrates the process of decoding human movement intentions based on sEMG data. Figure 2(b) illustrates the human-machine interaction control process combined with a robust follower controller. The framework generally includes an offline training process and an online application process. During training, the training sEMG data (or sample sEMG data) is first preprocessed and its features are extracted, and then the GANGT network is trained. During application, new sEMG data is preprocessed and its features are extracted, and then the trained GANGT network is used to obtain the desired motion trajectory corresponding to the new sEMG data.

[0030] In one embodiment, a wireless biofeedback system is used to collect sEMG data (or signals) of the muscles in the lower limbs. For example, muscle groups that have the greatest impact on the hip and knee joints are collected, including the rectus femoris (RF), vastus lateralis (VL), tibialis anterior (TA), biceps femoris (BF), semitendinosus, and gastrocnemius. Before acquisition, the surfaces at the measurement sites are wiped with alcohol to reduce external interference with the sEMG signals. Then, the raw sEMG data is preprocessed and features are extracted. For example, data normalization, sliding windowing, and low-pass filtering can be used for data preprocessing and feature extraction.

[0031] The goal of GANGT networks is to map human-computer interfaces to human joint angles. The detailed framework of GANGT is based on the core adversarial learning concepts of the generator-discriminator GAN architecture. See also... Figure 3 As shown, the GANGT network includes a GAN module, a graph module, and a Transformer module. Figure 3 (a) illustrates the input data, including random noise data, real sEMG data and real angle data from offline training, while the online application process uses the actual collected real sEMG data. Figure 3 (b) is a GAN module, which includes a discriminator (D) and a generator (G). Figure 3 (c) is the Transformer module, which includes graph structure, hyperempirical pattern decomposition, and self-attention mechanism.

[0032] The novel deep learning framework GANGT is provided to generate feature-rich sEMG data under small sample conditions. A graph structure is introduced into the transformer architecture to reconstruct the temporal and spatial characteristics of sEMG signals.

[0033] Combination Figure 3As shown in (b), GANGT is a framework that can generate sEMGs with features to increase the size of the network training set and change the network's generalization ability. First, a GAN module is trained, which contains a generator (G) and a discriminator (D). The input to the GAN module is random noise, real angle data, and real sEMG data. The generator's role is to generate fake data similar to real sEMG data, while the discriminator is responsible for determining the authenticity of the data. The generator and discriminator are trained adversarially in sync, dynamically generating new sEMG data with real features through self-supervised learning. During the offline training of the GAN module, the generator's goal is to generate sEMG data as realistic as possible to deceive the discriminator.

[0034] To ensure a mapping between the generated sEMG data and joint angles, actual angle data is incorporated into the generator's input. Unlike traditional GANs, this invention employs a four-layer network in both the generator and discriminator, each layer containing fractional convolutional layers, batch normalization layers, and ReLU activation functions. Specifically, fractional piecewise convolutions are introduced in the generator, and piecewise convolutions in the discriminator, allowing the GAN to learn its own spatial downsampling. Batch normalization addresses sample oscillations and model instability, improving the stability of GAN training. In the generator, the output layer uses the Tanh function, while other layers use the ReLU function. In the discriminator, each layer uses LeakyReLU activation, enabling the network model to learn more quickly. During the GAN's learning process, the batch size can be set to 64, and the number of iterations to 2000. The trained generator yields an enhanced sEMG dataset.

[0035] like Figure 3As shown in (c), after obtaining the enhanced sEMG dataset, a graph-based Transformer network was designed to decode the sEMG signal. This invention specifically discusses how to modify the original self-attention mechanism to capture the temporal and spatial features of the sEMG signal. Specifically, self-attention mechanisms in the temporal and spatial domains are constructed within the Transformer network. The temporal layer includes temporal embedding, temporal multi-head self-attention, and a feedforward network combined with graph-based time-series features. The spatial layer includes spatial embedding, spatial multi-head self-attention, and a feedforward network. When processing continuous sEMG signals, the temporal layer helps handle time-related features, while the spatial layer captures the spatial distribution features of the sEMG signal, thereby enhancing the network's versatility. In the temporal layer, the sEMG data of each channel is treated as a node in a graph structure. The data of each node is decomposed into a set of intrinsic mode functions (IMFs) using a hyperempirical mode decomposition method. The IMFs are used as features of the nodes to construct the graph structure and incorporated into the multi-head self-attention mechanism of the temporal layer to learn the relationships between sequences.

[0036] The output layers of the temporal and spatial domain layers use fully connected networks. The output of the spatial domain layer uses the sigmoid function as a mask. To ensure the regularization of the GANGT network, a global average pooling layer is added at the end of the network. In summary, to decode motion intent, the GANGT network generates feature-rich fake sEMG data during offline training to construct a data augmentation set, which is also the important role of the GAN module in offline training.

[0037] By using a trained Transformer network and taking real sEMG data as input, the corresponding predicted gait trajectory can be obtained.

[0038] Step S120: Construct a robust follower controller to track the joint angle error between the desired gait trajectory and the actual gait trajectory, so as to ensure that the tracking error of the exoskeleton robot in the joint space is bounded.

[0039] Combination Figure 2 As shown in (b), q d The smooth gait trajectory is generated by the output of the decoding motion intent network. The robust tracking controller takes the joint angle error e as input and outputs u. The control objective is to ensure that the tracking error of the lower limb exoskeleton robot in joint space is bounded, thus demonstrating robustness to modeling errors and external disturbances.

[0040] Considering model errors and external disturbances, the dynamic model of the lower limb exoskeleton robot can be expressed as:

[0041]

[0042] in, This represents the position coordinates and velocity matrix of the exoskeleton robot. In the developed LLE, q, It represents the angle and angular velocity of the hip and knee joints of the two legs in the sagittal plane. It is the total control input matrix of LLE. It is the inertia matrix. It is centripetal force and the Coriolis matrix. It is a gravity term. This indicates the modeling error of the exoskeleton. This indicates an external disturbance.

[0043] Before designing a robust tracking controller, it is essential to define the relevant characteristics and make reasonable assumptions, as this will be very helpful for the subsequent controller design.

[0044] The dynamic equations of LLE have the following properties and reasonable assumptions:

[0045] Feature 1: Matrix It is tilted symmetrical, and

[0046] Assumption 1: The desired trajectory and its first and second time derivatives are described by q d ,

[0047] Assumption 2: Modeling error and external disturbances Both are bounded. Positive definite functions exist. And δ, making and It always holds true.

[0048] Define the exoskeleton joint angle q and the desired exoskeleton joint angle q d The error between them is e = q d -q. And the virtual output is defined as:

[0049]

[0050] Where α > 0.

[0051] To ensure that the gait trajectory tracking error is bounded in the presence of modeling errors and external disturbances, in one embodiment, the robust following controller is designed as follows:

[0052]

[0053]

[0054] in, Represents positive integers. The matrix K is symmetric and positive definite.

[0055] In summary, the overall process of this invention is as follows: GAN is trained using a small sample dataset; for newly acquired sEMG data, approximate sEMG data is generated using a trained generator; an enhanced training set is constructed using this approximate sEMG data combined with the original sample data; a Transformer network is trained using the enhanced training set; real-time acquired sEMG data is input into the trained Transformer network to obtain the desired (or predicted) motion trajectory; and then a follower controller is used to track the desired motion trajectory, thereby ensuring that the tracking error of the exoskeleton robot in the joint space is bounded.

[0056] It should be noted that, without departing from the spirit and scope of this invention, those skilled in the art can make appropriate changes or modifications to the above embodiments. For example, the proposed human-computer interaction control framework is not limited to specific lower limb exoskeleton products, but is also applicable to other types of exoskeleton robots. The preprocessing and feature extraction of raw data are not limited to specific methods. The amount of data augmentation can be adjusted according to actual needs.

[0057] To further verify the effectiveness of this invention, it was tested on a self-developed lower limb exoskeleton robot platform. The lower limb exoskeleton robot, as shown in the figure... Figure 4 As shown. The LLE hardware system mainly includes a backpack, frame links, joint modules, straps, and foot pedals.

[0058] The developed LLE system provides six degrees of freedom for the lower limb, including four active degrees of freedom (hip and knee joints) and two passive degrees of freedom (ankle). The hip and knee joint modules consist of limit blocks, fixation plates, movement plates, actuator units, and protective covers. The actuator units include frameless torque motors, harmonic drivers, and encoders. The developed LLE system utilizes a Next-Level Computing Unit (NUC) to build a real-time operating system for controller development. The control frequency is set to 1000Hz. All drive units communicate with the real-time control cabinet via an EtherCAT bus. Furthermore, a compliant ankle joint with elastic deformation was designed to achieve passive dorsiflexion / plantar flexion. The designed ankle joint can withstand greater LLE deflection and reduces the impact of the ground on the LLE.

[0059] In the experimental verification, the performance of the GANGT network was mainly analyzed, and a walking experiment was conducted on a wearer wearing an exoskeleton using the designed human-computer interaction control framework.

[0060] In the performance analysis experiments of the GANGT network, in order to lower the threshold of deep learning embedded exoskeleton control and reduce the cost of data sample acquisition, a GANGT network trained on small sample data was adopted. GANGT can generate fake sEMG data with features. The GAN module generates fake sEMG data through adversarial training to counter the side of labeled real data, thereby establishing a large number of real features with new sEMG data.

[0061] Figure 5 This is a diagram illustrating GAN-enhanced sEMG data from three test sets. The first three columns show the sEMG data at different time epochs, and the last column shows the real sEMG data. Figure 5 It is evident that after 2000 epochs, new data highly similar to real sEMG features is generated. Through offline training, the amount of sEMG data with 10,000 features can be increased to 100,000 samples, significantly reducing the interference from manual feature extraction. Notably, this method offers sufficient flexibility to generate sEMG data of any scale during adversarial learning.

[0062] In the experiment of walking with the exoskeleton, sEMG sensors were installed at the corresponding muscle positions in the right leg of all subjects to collect signals in real time. During the experiment, six subjects walked freely within a limited area. Figure 6 The trajectory and tracking results were generated based on the designed framework, including the results of four subjects walking on flat ground and two subjects walking uphill on three slopes. Figure 6 This image shows joint tracking results for the right hip and knee joints of six subjects wearing exoskeletons. The red dashed line represents the actual joint angle, and the blue solid line represents the predicted joint angle. Figure 6 It can be seen that, whether walking on flat ground or uphill on slopes of varying gradients, the developed framework can accurately decode the subject's movement intentions and generate gait trajectories in real time. Thanks to the robust follower controller, the actual joint movement trajectories are largely consistent with the predicted trajectories, indicating highly accurate tracking control.

[0063] Table 1 shows the mean average error (MAE) of the tracking error. It can be observed that there is no significant difference in tracking error between walking on flat ground and walking uphill, which verifies the robustness and generalization ability of the robust following controller designed in this invention. The mean following errors for all subjects were: left hip 0.1207±0.0074°, left knee 0.0851±0.0043°, right hip 0.1212±0.0081°, and right knee 0.0841±0.0043°. This demonstrates that the robust following controller has good control accuracy, is unaffected by terrain and subject variations, and exhibits good robustness.

[0064] Table 1: MAE of Tracking Error

[0065]

[0066] The results of the walking experiment demonstrate that the proposed controller generalizes to solving human-computer interaction problems across different subjects and terrains. Under the robust tracking controller, the mean absolute tracking error for the hip joint was 0.0846° and the mean absolute tracking error for the knee joint was 0.1210° for all subjects. This result indicates that all tracking errors are concentrated within a bounded region, thus validating the robustness and effectiveness of the proposed robust tracking controller. In summary, these validation results provide important insights into how to accurately decode human-computer interfaces using sEMG and achieve robust tracking control for LLE (Low-Least-Effect) systems.

[0067] In summary, compared with the prior art, the present invention has the following advantages:

[0068] 1) In gait prediction for LLE (Low-Leg Levitation), accurate prediction models can improve walking stability and achieve a natural and reasonable human-computer interaction effect. This invention combines adversarial learning with a transformer to construct a data-augmented decoding network for human motion intent, achieving accurate online prediction of the human-computer interface. Furthermore, under conditions of small sample sEMG data, the GANGT network utilizes adversarial learning to generate feature sEMGs to expand the dataset. It improves the network's generalization ability by constructing temporal and spatial self-attention mechanisms, and uses the sEMG data of each channel in the temporal domain as nodes in the graph structure to capture data features between local regions. In summary, GANGT can generate a large amount of sEMG data and obtain local and global sEMG features through a tightly coupled graph structure and self-attention mechanisms in the temporal and spatial domains.

[0069] 2) This invention proposes a robust follower controller to enable the exoskeleton to accurately follow predicted joint angles. Even with unmodeled errors and external disturbances in the exoskeleton dynamics system, the controller can keep the tracking error within a small range. Its adaptability to different wearers and terrains was verified in wearable exoskeleton experiments, effectively realizing human-machine interaction control of the lower limb exoskeleton.

[0070] 3) This invention has strong robustness and is applicable to any exoskeleton robot, not limited to lower limb exoskeleton robot models.

[0071] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0072] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0073] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0074] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0075] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0076] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0077] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0078] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0079] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A human-robot interaction control method of a lower extremity exoskeleton robot, comprising the following steps: collecting surface electromyography data of a target; inputting the surface electromyography data into a trained motion intention decoding network to output a corresponding desired gait trajectory; tracking the desired gait trajectory using a designed robust following controller to ensure that the tracking error of the lower extremity exoskeleton robot in the joint space is bounded; wherein the motion intention decoding network comprises a generative adversarial network module, a graph neural network module and a Transformer module, the Transformer module comprises a time layer and a space layer, the time layer is used to process time-related features in the surface electromyography data, and comprises a time domain embedding layer, a time domain multi-head attention mechanism layer and a first feedforward network, the space layer is used to capture the spatial distribution features of the surface electromyography data, and comprises a spatial domain embedding layer, a spatial domain multi-head attention mechanism layer and a second feedforward network; in the time layer, the surface electromyography data of each channel is taken as a node of a graph structure in the graph neural network module, the data of each node is decomposed into a group of eigenmodes as the features of the corresponding node, and the attention scores obtained by using the graph structure are transmitted to the time domain multi-head attention mechanism layer to learn the relationship between sequences.

2. The method of claim 1, wherein, The motion intention decoding network is trained according to the following steps: training a generative adversarial network using random noise data, sample angle data and sample surface electromyography data, the generative adversarial network comprising a generator and a discriminator, the generator being used to generate false data similar to the sample surface electromyography data, and the discriminator being used to determine the authenticity of the data generated by the generator; for the actually collected surface electromyography data, generating new surface electromyography data using the trained generator; constructing an enhanced data set as a training set using the new surface electromyography data and the sample surface electromyography data, the training set reflecting the corresponding relationship between the surface electromyography data and the desired motion trajectory; training the Transformer module using the training set to meet a set loss standard.

3. The method of claim 2, wherein, The generator adopts a fractional form convolution layer, and the discriminator adopts a step-by-step convolution layer.

4. The method of claim 1, wherein, The exoskeleton robot is a lower extremity exoskeleton robot, and a dynamics model of the lower extremity exoskeleton robot is represented as: wherein, represents the position coordinates and velocity matrix of the lower extremity exoskeleton robot, is the total control input matrix of the lower extremity exoskeleton robot, is the inertia matrix, is the centripetal and Coriolis matrix, is the gravity term, represents the modeling error of the lower extremity exoskeleton robot, represents the external disturbance, and the dynamics model of the lower extremity exoskeleton robot has the following properties: matrix is skew-symmetric, and The desired trajectory and its first and second time derivatives are described as q d , modeling errors and external disturbances are bounded, and there exist positive definite functions and δ, such that and hold simultaneously.

5. The method of claim 4, wherein, the robust following controller is represented as: wherein: e = q d - q wherein denotes a positive integer, the matrix K is symmetric positive definite, and a >

0.

6. The method of claim 1, wherein, the output layer of the time layer and the space layer uses a fully connected network, and the output of the space layer uses a sigmoid function as a mask.

7. The method of claim 1, wherein, The eigenmodes are obtained by a hyper empirical mode decomposition method.

8. A computer readable storage medium having stored thereon a computer program, wherein, The computer program is executed by a processor to implement the steps of the method according to any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, having stored on the memory a computer program capable of running on the processor, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

Images