Brachial plexus and exoskeleton cooperative control method and system based on brain-computer interface

Abstract

The invention provides a brachial plexus and exoskeleton cooperative control method and system based on a brain-computer interface, and relates to the field of intelligent control. The method comprises the steps that electroencephalogram signals and upper limb electromyographic signals of a target object during motor imagery are collected; predicting the motion intention of the target object through a neural network based on the electroencephalogram signals and the upper limb electromyographic signals; based on the motion intention, determining a target brachial plexus branch according to a preset motion intention and brachial plexus mapping rule; generating a nerve stimulation control instruction and an exoskeleton cooperative control instruction based on the target brachial plexus branch and the motion intention; and according to the nerve stimulation control instruction, performing electrical stimulation on the target brachial plexus branch to drive the corresponding muscle to contract, and meanwhile, according to the exoskeleton cooperative control instruction, controlling the upper limb exoskeleton to apply assistance to the corresponding joint. The method and the device are used in the cooperative control process of the brachial plexus and the exoskeleton, and the technical problem of inaccurate control in the existing limb control process based on neural signals is solved.

Description

Technical Field

[0001] This application relates to the field of intelligent control, and in particular to a method and system for coordinated control of the brachial plexus and exoskeleton based on a brain-computer interface. Background Technology

[0002] With the development of brain-computer interfaces and exoskeleton control technologies, limb control based on neural signals is gradually being applied. However, existing technologies suffer from inaccurate control in practical applications. Specifically, there is a discrepancy between the motor intention obtained from decoding electroencephalogram (EEG) and electromyographic (EMG) signals and the actual upper limb movements executed, resulting in the amplitude, direction, or degree of completion of the movement failing to match expectations. This is because neural signals are characterized by significant individual variability, insufficient stability, and susceptibility to interference, causing fluctuations in motor intention predictions across different objects or states. When these predictions are directly used to generate control commands, control accuracy is difficult to guarantee. Therefore, a new technical solution is urgently needed to address the inaccurate control issues in existing neural signal-based limb control processes. Summary of the Invention

[0003] This application provides a brain-computer interface-based method and system for coordinated control of the brachial plexus and exoskeleton, which solves the technical problem of inaccurate control in existing limb control processes based on neural signals.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a brain-computer interface-based method for coordinated control of the brachial plexus and exoskeleton is provided, comprising: collecting electroencephalogram (EEG) signals, upper limb electromyographic (EMG) signals, and joint motion state data of a target object during motor imagery; the joint motion state data includes joint angle information and joint motion trajectory information; predicting the target object's movement intention using a neural network based on the EEG signals and upper limb EMG signals; the neural network being a transfer learning model based on ICA (Independent Component Analysis); determining the target brachial plexus branch based on the movement intention and a preset mapping rule between the movement intention and the brachial plexus; generating neural stimulation control instructions and exoskeleton coordinated control instructions based on the target brachial plexus branch and the movement intention, wherein the neural stimulation control instructions include stimulation amplitude, stimulation frequency, and stimulation pulse width parameters; and the exoskeleton coordinated control instructions include target joint and assist output parameters; and applying electrical stimulation to the target brachial plexus branch to drive corresponding muscle contraction according to the neural stimulation control instructions, while simultaneously controlling the upper limb exoskeleton to apply assistance to the corresponding joint according to the exoskeleton coordinated control instructions.

[0005] Based on the above technical solution, in the brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application, this application collects EEG signals, upper limb electromyography signals, and joint motion state data during the target object's motor imagination process. It then uses a neural network to jointly analyze multi-source neural signals to predict motor intentions, and on this basis, collaboratively generates neural stimulation control commands and exoskeleton coordinated control commands to achieve coordinated driving of upper limb movements. Compared to control schemes that rely solely on a single neural signal or a single execution method, this application can achieve more comprehensive information representation at the level of motor intention, making the generated control commands closer to the target object's true motor intentions. This effectively reduces the deviation between motor intentions and actual executed actions, improving the accuracy and consistency of upper limb movements. Simultaneously, through the coordinated execution of neural stimulation and exoskeleton assistance, muscle contraction and joint movement can be coordinated, avoiding the action mismatch problem caused by a single control method, further improving the stability and reliability of the overall control effect.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, after controlling the upper limb exoskeleton to apply assistance to the corresponding joint according to the exoskeleton collaborative control command, the method further includes: acquiring joint motion state data and electromyographic feedback signals of the upper limb; and dynamically adjusting the neural stimulation control command and the exoskeleton collaborative control command through deep reinforcement learning based on the deviation between the joint motion state data and the electromyographic feedback signal and the preset target.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the target object's movement intention is predicted through a neural network based on electroencephalogram (EEG) signals and upper limb electromyography (EMG) signals. This includes: removing interference signals from the EEG and upper limb EMG signals using independent component analysis (ICA) to extract core EEG features and resting-state EMG features; fusing the core EEG features and resting-state EMG features; and then using a transfer learning model to predict and verify the probability distribution of the movement intention to obtain the target object's movement intention. The transfer learning model is a model constructed from a three-dimensional convolutional neural network and a long short-term memory network, initialized based on source domain movement intention samples.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves fusing EEG core features with EMG resting-state features, then using a transfer learning model to predict and verify the probability distribution of movement intentions to obtain the target object's movement intention. This includes: segmenting the EEG core features according to a preset time window and aligning them temporally with the EMG resting-state features within the same time window; concatenating the aligned features according to the channel dimension to construct a fused feature, which is a three-dimensional fused feature tensor containing temporal and spatial distribution information; inputting the fused feature into a pre-processed three-dimensional convolutional neural network for transfer learning to extract spatiotemporally relevant features to obtain high-level spatial features; inputting the high-level spatial features into a long short-term memory network to predict the predicted probability distribution of different movement intentions; and verifying the confidence of the predicted probability distribution. When the probability value corresponding to the target movement intention is greater than a preset confidence threshold, the target object's movement intention is determined.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the transfer learning 3D convolutional neural network is initialized by using a 3D convolutional neural network trained on multi-class motion intention samples in the source domain as the initial network. By freezing the first preset number of spatiotemporal feature extraction layers and fine-tuning the second preset number of feature representation layers, the motion intention features between different target objects are transferred to obtain the transfer learning initialized 3D convolutional neural network.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, based on the target brachial plexus branch and the movement intention, neural stimulation control instructions and exoskeleton collaborative control instructions are generated, including: determining the neural stimulation target marker corresponding to the movement intention based on the target brachial plexus branch, and matching the corresponding initial stimulation parameter template; generating neural stimulation control instructions containing stimulation amplitude, stimulation frequency, and stimulation pulse width based on the movement intention and the initial stimulation parameter template; determining the target joint that needs to be coordinated according to the movement intention, and generating exoskeleton collaborative control instructions containing assist output parameters; and applying temporal synchronization constraints to the neural stimulation control instructions and exoskeleton collaborative control instructions to ensure that neural stimulation and exoskeleton assistance are executed collaboratively within the same movement phase.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, based on the deviation between joint motion state data and electromyographic feedback signals and a preset target, deep reinforcement learning is used to dynamically adjust neural stimulation control commands and exoskeleton collaborative control commands. This includes: constructing a reinforcement learning strategy network based on joint motion state data and electromyographic feedback signals; the reinforcement learning strategy network includes: constructing a state space based on the motion stage, joint angle changes, and intensity of electromyographic feedback signals in the joint motion state data; constructing an action space based on the stimulation frequency, stimulation amplitude, stimulation pulse width of the neural stimulation control commands, and output parameters of the exoskeleton collaborative control commands; a reward function constrained by motion completion and abnormal muscle responses; determining the current motion stage and forming the current state based on the joint motion state data and electromyographic feedback signals; outputting corresponding action parameters through the reinforcement learning strategy network based on the current state, and updating the neural stimulation control commands and exoskeleton collaborative control commands; and calculating a reward value based on the motion effect after a single motion is completed, and updating the reinforcement learning strategy network accordingly.

[0012] Secondly, a brain-computer interface-based brachial plexus and exoskeleton collaborative control system is provided to implement any of the methods in the first aspect. The system includes: a signal acquisition module for acquiring EEG signals, upper limb electromyography (EMG) signals, and joint motion state data of the target object during motor imagery, wherein the joint motion state data includes joint angle information and joint motion trajectory information; a motion intention prediction module for predicting the target object's motion intention based on EEG signals and upper limb EMG signals using a neural network; the neural network is a transfer learning model based on ICA (Independent Component Analysis) technology; a brachial plexus branch determination module for determining the target brachial plexus branch based on the motion intention and according to a preset mapping rule between the motion intention and the brachial plexus; and a collaborative control command generation module for generating commands based on the target brachial plexus branch. The system generates neural stimulation control commands and exoskeleton-assisted control commands based on neural branches and motor intentions. The neural stimulation control commands include parameters such as stimulation amplitude, stimulation frequency, and stimulation pulse width; the exoskeleton-assisted control commands include target joints and assist output parameters. An execution control module is used to apply electrical stimulation to the target brachial plexus branches to drive corresponding muscle contraction according to the neural stimulation control commands, and to control the upper limb exoskeleton to apply assistance to the corresponding joints according to the exoskeleton-assisted control commands. An adaptive adjustment module is used to acquire upper limb joint motion state data and electromyographic feedback signals after execution control, and to dynamically adjust the neural stimulation control commands and exoskeleton-assisted control commands through deep reinforcement learning based on the deviation between the joint motion state data and electromyographic feedback signals and the preset target.

[0013] In conjunction with the second aspect above, in one possible implementation, the system further includes: a hardware integration and timing synchronization module for synchronizing the interconnection and timing between the brain-computer interface, the brachial plexus stimulation module, the sensors, and the upper limb exoskeleton.

[0014] In conjunction with the second aspect above, in one possible implementation, the motion intention prediction module is further used to: remove interference signals from EEG signals and upper limb EMG signals using ICA independent component analysis technology, and extract EEG core features and EMG resting-state features; after fusing the EEG core features and EMG resting-state features, predict and verify the probability distribution of motion intention through a transfer learning model to obtain the motion intention of the target object; the transfer learning model includes: a transfer learning three-dimensional convolutional neural network and a long short-term memory network.

[0015] This application provides a brain-computer interface-based method and system for coordinated control of the brachial plexus and exoskeleton. By comprehensively utilizing electroencephalogram (EEG) signals, upper limb electromyographic (EMG) signals, and joint motion state data, it predicts movement intentions based on neural networks and generates corresponding neural stimulation control commands and exoskeleton coordination control commands, establishing a correspondence between the expression of movement intentions and the actual execution of upper limb movements. Simultaneously, by introducing joint motion state data and EMG feedback signals, the control commands are adaptively adjusted, reducing control errors caused by individual differences, signal fluctuations, or prediction biases. Furthermore, through feature processing, fusion modeling, and transfer learning, the stability and consistency of movement intention prediction are improved across different objects and states. This enhances overall control accuracy and reliability, solving the technical problem of inaccurate control in existing neural signal-based limb control systems.

[0016] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0017] Figure 1 A system architecture diagram of a brain-computer interface-based brachial plexus and exoskeleton collaborative control system is provided for embodiments of this application; Figure 2 A flowchart illustrating a brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application embodiment; Figure 3 A flowchart illustrating another brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application embodiment; Figure 4 A flowchart illustrating another brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application embodiment; Figure 5 A flowchart illustrating another brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application embodiment; Figure 6 This is a flowchart illustrating another brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in an embodiment of this application. Detailed Implementation

[0018] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0019] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] The brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application can be applied to, for example... Figure 1 In the brain-computer interface-based brachial plexus and exoskeleton coordinated control system shown, such as Figure 1 As shown, the system includes: a signal acquisition module 101, a motion intention prediction module 102, a brachial plexus branch determination module 103, a collaborative control command generation module 104, an execution control module 105, and an adaptive adjustment module 106.

[0021] Among them, the signal acquisition module 101 is used to acquire the electroencephalogram (EEG) signal, upper limb electromyogram (EMG) signal, and joint motion state data when the target object imagines motion. The joint motion state data includes joint angle information and joint motion trajectory information. The motion intention prediction module 102 is used to predict the motion intention of a target object based on electroencephalogram (EEG) signals and upper limb electromyography (EMG) signals through a neural network; the neural network is a transfer learning model based on independent component analysis (ICA) technology. The brachial plexus branch determination module 103 is used to determine the target brachial plexus branch based on the motor intention and according to the preset motor intention and brachial plexus mapping rules. The collaborative control command generation module 104 is used to generate neural stimulation control commands and exoskeleton collaborative control commands based on the target brachial plexus nerve branches and movement intentions. The neural stimulation control commands include stimulation amplitude, stimulation frequency and stimulation pulse width parameters; the exoskeleton collaborative control commands include target joint and assist output parameters. The execution control module 105 is used to apply electrical stimulation to the target brachial plexus branch according to the nerve stimulation control command to drive the corresponding muscle contraction, and to control the upper limb exoskeleton to apply assistance to the corresponding joint according to the exoskeleton coordination control command. The adaptive control module 106 is used to acquire upper limb joint motion state data and electromyographic feedback signals after execution of control, and dynamically adjust the neural stimulation control commands and exoskeleton collaborative control commands through deep reinforcement learning based on the deviation between the joint motion state data and electromyographic feedback signals and the preset target.

[0022] In one possible implementation, the system also includes a hardware integration and timing synchronization module for synchronizing the interconnection and timing between the brain-computer interface, the brachial plexus stimulation module, the sensors, and the upper limb exoskeleton.

[0023] In one possible implementation, the motion intention prediction module is further used to: remove interference signals from EEG signals and upper limb electromyography (EMG) signals using ICA (Independent Component Analysis) technology, and extract EEG core features and EMG resting-state features; after fusing the EEG core features and EMG resting-state features, predict and verify the probability distribution of motion intention through a transfer learning model to obtain the motion intention of the target object; the transfer learning model includes: a transfer learning three-dimensional convolutional neural network and a long short-term memory network.

[0024] To address the technical problem of inaccurate control in existing limb control methods based on neural signals, this application provides a brain-computer interface-based brachial plexus and exoskeleton coordinated control method. The method includes: collecting electroencephalogram (EEG) signals, upper limb electromyographic (EMG) signals, and joint motion state data of a target object during motor imagery; the joint motion state data includes joint angle information and joint motion trajectory information; predicting the target object's movement intention using a neural network based on the EEG signals and upper limb EMG signals; the neural network is a transfer learning model based on ICA (Independent Component Analysis); determining the target brachial plexus branch based on the movement intention and a preset mapping rule between the movement intention and the brachial plexus; generating neural stimulation control commands and exoskeleton coordinated control commands based on the target brachial plexus branch and the movement intention, wherein the neural stimulation control commands include stimulation amplitude, stimulation frequency, and stimulation pulse width parameters; and the exoskeleton coordinated control commands include target joint and assist output parameters; and applying electrical stimulation to the target brachial plexus branch to drive corresponding muscle contraction according to the neural stimulation control commands, while simultaneously controlling the upper limb exoskeleton to apply assistance to the corresponding joint according to the exoskeleton coordinated control commands.

[0025] Figure 2 This is a flowchart illustrating the brain-computer interface-based brachial plexus and exoskeleton coordinated control method provided in this application embodiment. Figure 2 As shown, the method includes: S201. Collect EEG signals, upper limb electromyography signals, and joint movement data of the target object during motor imagery.

[0026] Among them, electroencephalogram (EEG) signals refer to the electrophysiological signals generated by the activity of the cerebral cortex during the target object's motor imagination; upper limb electromyography (EMG) signals refer to the surface or near-surface electrical signals generated by the relevant muscle groups of the upper limb under the drive of nerves; joint motion state data is a set of data used to characterize the motion state of the upper limb, including joint angle information and joint motion trajectory information.

[0027] In one possible implementation, EEG signals during the motor imagery phase are collected by an EEG acquisition device worn on the target's head, and corresponding upper limb electromyography (EMG) signals are collected by an EMG acquisition device placed on the surface of the upper limb muscle groups. At the same time, angle and displacement sensors installed on the upper limb exoskeleton or joints are used to acquire upper limb joint angle information and joint movement trajectory information in real time.

[0028] It should be noted that different sampling frequencies can be used for EEG signals, upper limb electromyography signals, and joint motion state data during the acquisition process, but timestamp marking or cache alignment are required before subsequent processing to ensure the correspondence between different modal data in the time dimension.

[0029] This step, through the synchronous acquisition of multi-source signals, enables motion intention recognition and control decisions to simultaneously combine central nervous system information, peripheral muscle information, and actual joint motion status, providing reliable data support for downstream tasks.

[0030] S202. Based on electroencephalogram (EEG) signals and upper limb electromyography (EMG) signals, predict the movement intention of the target object through a neural network.

[0031] Among them, motor intention refers to the type or direction of upper limb movement that the target object expects to complete during the motor imagination stage. In this application, it includes: elbow flexion, elbow extension, wrist flexion, wrist extension, grasping, and relaxation. The transfer learning model refers to a neural network model that is trained based on existing motor intention samples and can be adapted to different target objects. The neural network is a transfer learning model based on ICA independent component analysis technology.

[0032] In one possible implementation, the acquired EEG signals and upper limb EMG signals are input into a transfer learning model constructed based on ICA independent component analysis technology. After denoising and feature extraction of the signals, the model outputs prediction results corresponding to different movement intentions, thereby obtaining the movement intention of the target object.

[0033] Based on the above steps, this step uses the fusion of EEG signals and upper limb electromyography signals to predict motor intention, which helps to reduce the error caused by the instability of a single neural signal, thereby improving the accuracy of motor intention recognition.

[0034] S203. Based on the motor intention, determine the target brachial plexus branch according to the preset motor intention and brachial plexus mapping rules.

[0035] Among them, the brachial plexus mapping rule refers to the pre-established correspondence rules between motor intentions and brachial plexus branches, which are used to characterize the neural innervation pathways involved in different motor intentions.

[0036] In one possible implementation, based on the motor intention, a match is made between the preset motor intention and the brachial plexus mapping rules to determine the target brachial plexus branch corresponding to the motor intention, which is then used for subsequent neural stimulation control.

[0037] It should be noted that mapping rules can be set based on anatomical knowledge, experimental data, or empirical rules, and can be adjusted or expanded according to actual application needs. For example, the brachial plexus, composed of the anterior branches of the 5th to 8th cervical nerves (C5-C8) and the anterior branch of the 1st thoracic nerve (T1), is the main neural pathway for upper limb movement and sensation. Its structure differentiates hierarchically, with each level corresponding to a specific upper limb motor function: C5-C6 mainly innervates shoulder and elbow joint movements, C7 mainly participates in wrist and forearm movements, and C8-T1 mainly controls fine motor movements such as grasping and finger opposition. The nerve trunk further divides into anterior and posterior branches, corresponding to flexor and extensor muscle functions, respectively. Finally, it forms terminal branches such as the musculocutaneous nerve, radial nerve, median nerve, and ulnar nerve, which directly innervate specific muscles and complete corresponding movements. Based on the above clear and stable anatomical-functional correspondence, mapping rules between motor intentions and brachial plexus branches can be established, enabling accurate selection of nerve stimulation targets and improving the precision of upper limb control and functional recovery.

[0038] As an example, when the intention of movement is elbow flexion, the corresponding target brachial plexus branch can be the nerve branch that mainly innervates the biceps brachii.

[0039] Based on the above steps, this step clarifies the correspondence between motor intention and brachial plexus branches, enabling subsequent neural stimulation to act on neural pathways highly related to the target movement, thereby reducing control deviations caused by irrelevant stimuli.

[0040] S204. Based on the target brachial plexus branches and the movement intention, generate neural stimulation control commands and exoskeleton coordination control commands.

[0041] Among them, the neural stimulation control command is a set of commands used to control the electrical stimulation parameters of the brachial plexus, including stimulation amplitude, stimulation frequency and stimulation pulse width parameters; the exoskeleton coordination control command is a set of commands used to control the assist output of the upper limb exoskeleton, including target joint and assist output parameters.

[0042] In one possible implementation, based on a determined target brachial plexus branch and its corresponding movement intention, the movement intention is structurally parsed to extract the action type and corresponding joint information. According to a preset mapping rule between movement intention and brachial plexus stimulation parameters, the movement intention is converted into a set of neural stimulation control parameters for the target brachial plexus branch. These neural stimulation control parameters include at least stimulation amplitude, stimulation frequency, and stimulation pulse width. The mapping rule is used to limit the target brachial plexus branch and its stimulation parameter value range corresponding to different movement intentions, thereby generating corresponding neural stimulation control commands. Simultaneously, according to a preset movement intention and exoskeleton assistance configuration rule, the target joint corresponding to the movement intention and its assistance output parameters are determined. These assistance output parameters include at least assistance magnitude or assistance level, thereby generating exoskeleton collaborative control commands. Both types of control commands are generated within the same control cycle and sent to the neural stimulation module and exoskeleton control module, respectively.

[0043] It should be noted that both the neural stimulation control command and the exoskeleton collaborative control command are triggered and generated by the same motor intention, but their parameter configurations are independent of each other to avoid the impact of a single execution channel malfunction on the overall motion control.

[0044] As an example, when the movement intention is identified as an elbow flexion movement, the target brachial plexus branch is determined to be the musculocutaneous nerve according to the preset movement intention and brachial plexus stimulation parameter mapping rules. The stimulation amplitude, stimulation frequency and stimulation pulse width corresponding to the elbow flexion movement are selected from the preset parameter range to generate a neural stimulation control command for the musculocutaneous nerve. At the same time, according to the preset movement intention and exoskeleton assistance configuration rules, the target joint is determined to be the elbow joint, and an exoskeleton collaborative control command for controlling the upper limb exoskeleton to output corresponding assistance at the elbow joint is generated.

[0045] As an example, the decoded effective motor intentions pinpoint the target nerve branches, such as elbow flexion corresponding to the musculocutaneous nerve and grasping intention corresponding to the median and ulnar nerves; simultaneously, the joints requiring coordinated movement are identified, such as elbow flexion corresponding to the elbow joint; two types of core instructions are generated: the first is a stimulation instruction, containing the target nerve branch ID and initial stimulation parameters, and the second is an exoskeleton coordination instruction, containing the target joint and initial assist torque. The initial stimulation parameters are matched based on the target object's current functional level, for example, for severe paralysis, the default amplitude is 4mA, frequency is 60Hz, and pulse width is 200μs; the two types of instructions are transmitted to the neurostimulation module and exoskeleton control respectively via a high-speed synchronous bus, with an instruction transmission delay of ≤5ms, ensuring the consistency of movement coordination.

[0046] Based on the above steps, this step achieves coordinated output of neural activation and mechanical assistance by synchronously converting motor intentions into neural stimulation control commands and exoskeleton coordination control commands, thereby improving the stability and controllability of upper limb movement execution.

[0047] S205. According to the neural stimulation control command, electrical stimulation is applied to the target brachial plexus branch to drive the corresponding muscle contraction. At the same time, according to the exoskeleton coordination control command, the upper limb exoskeleton is controlled to apply assistance to the corresponding joint.

[0048] Among them, execution control refers to the process of applying electrical stimulation to the target brachial plexus nerve and providing assisted control to the upper limb exoskeleton based on the generated control instructions.

[0049] In one possible implementation, electrical stimulation is applied to the target brachial plexus branch according to the neural stimulation control command to drive the corresponding muscle to contract. At the same time, according to the exoskeleton coordination control command, the upper limb exoskeleton is controlled to apply assistance to the corresponding joint, thereby completing the expected upper limb movement.

[0050] Based on the above steps, this step, through the coordinated execution of neural stimulation and exoskeleton assistance, enables the predicted movement intention to be more accurately translated into actual movement output, thereby reducing control deviation and improving the degree of movement completion.

[0051] This application achieves multidimensional perception of the target object's movement intention and actual movement state by simultaneously acquiring EEG signals, EMG signals, and joint motion state data, reducing information bias caused by insufficient single signals. It predicts movement intention based on joint modeling of EEG and EMG signals, reducing uncertainty in intention recognition caused by neural signal fluctuations. According to preset movement intention and brachial plexus mapping rules, abstract movement intentions are transformed into specific target neural branches, improving the targeting of neural stimulation objects. Furthermore, movement intentions are mapped to specific neural stimulation control commands and exoskeleton-assisted control commands, ensuring the control parameters are executable and consistent, avoiding ambiguity or conflict in control strategies. Through the coordinated execution of neural stimulation and exoskeleton assistance, the synchronous effect of muscle activation and mechanical assistance is achieved, reducing the accumulation of movement deviations. By constructing a complete collaborative control process from neural signal perception and movement intention recognition to execution control, the accuracy and stability of brain-computer interface-based upper limb control are effectively improved, solving the technical problem of inaccurate control in existing neural signal-based limb control processes.

[0052] In one possible implementation, combining Figure 2 ,like Figure 3 As shown, following S205, the brachial plexus and exoskeleton coordinated control method based on brain-computer interface provided in this application embodiment further includes the following S301 and S302: S301. Acquire joint motion data and electromyographic feedback signals of the upper limb.

[0053] Among them, joint motion state data refers to data collected by motion state sensors that reflect the actual motion of the upper limb joints, including changes in joint angle, angular velocity, and motion trajectory deviation information; electromyographic feedback signal refers to the surface electromyographic signal generated by the target muscle or synergistic muscle group during the coordinated execution of nerve stimulation and exoskeleton, which is used to characterize the actual activation intensity and contraction state of the muscle.

[0054] In one possible implementation, sensors placed at key joints of the upper limb exoskeleton are used to collect joint angle and motion trajectory data in real time, and electromyography (EMG) sensors placed on the surface of the target muscle are used to collect EMG feedback signals synchronously. The collected joint motion state data and EMG feedback signals are time-aligned and preprocessed to obtain the actual motion performance data within the current action cycle, which serves as the input information for subsequent adaptive control.

[0055] It should be noted that the acquired data is objective feedback data after the execution of control. The acquisition process does not participate in motion intent prediction, but is only used to evaluate the execution effect of the current control command, thereby avoiding interference with the front-end motion intent decoding process.

[0056] As an example, during the elbow flexion movement, the actual elbow flexion angle curve is obtained through a sensor at the elbow joint position, and the electromyographic feedback signal of the biceps brachii is collected simultaneously. When the actual elbow flexion angle is detected to be less than the expected angle or the electromyographic activation intensity is insufficient, the corresponding data is recorded as the feedback result of this movement.

[0057] Based on the above steps, this step introduces the joint acquisition of joint motion state data and electromyographic feedback signals, enabling the system to objectively reflect the actual execution effect of control commands, providing a reliable basis for the adaptive adjustment of subsequent control parameters, thereby avoiding the accumulation of control deviations caused by relying solely on prediction results for open-loop control.

[0058] S302. Based on the deviation between joint motion state data and electromyographic feedback signals and preset targets, the neural stimulation control commands and exoskeleton collaborative control commands are dynamically adjusted through deep reinforcement learning.

[0059] Among them, the preset target refers to the expected motion state set for different motion intentions, including the target joint angle range, motion trajectory characteristics and corresponding electromyographic activation intensity range; the deviation refers to the difference between the actual collected joint motion state data and electromyographic feedback signals and the preset target.

[0060] In one possible implementation, a reinforcement learning state input is constructed based on feedback data. The current motion stage, joint angle deviation, and electromyographic activation deviation are input as state variables into the deep reinforcement learning policy network. The stimulation frequency, stimulation amplitude, stimulation pulse width in the neural stimulation control command, and the assist output parameters in the exoskeleton collaborative control command are used as the action space. The reward value is calculated by comparing the deviation between the actual motion result and the preset target, and the policy network is updated according to the reward result, thereby outputting the adjusted neural stimulation control command and exoskeleton collaborative control command corresponding to the next action cycle.

[0061] It should be noted that the deep reinforcement learning process gradually updates the control strategy in the continuous action cycle, and the adjustment process is limited to the preset safety parameter range to avoid sudden changes in the stimulus parameters or assist parameters, thereby ensuring the stability and continuity of the control process.

[0062] Based on the above steps, this step introduces deep reinforcement learning to achieve online adaptive adjustment of neural stimulation control commands and exoskeleton collaborative control commands, enabling control parameters to be dynamically optimized according to actual execution results, thereby reducing the deviation between motor intention and actual action and improving the accuracy and robustness of the collaborative control process.

[0063] This application acquires upper limb joint motion state data and electromyographic feedback signals in real time after the execution of collaborative control, compares them with preset targets, and introduces deep reinforcement learning to dynamically adjust neural stimulation control commands and exoskeleton collaborative control commands. This makes the control strategy no longer dependent on the prediction results of a single movement intention, but can be continuously corrected based on the actual movement performance. This effectively reduces the impact of movement intention decoding errors, individual differences, and state fluctuations on control accuracy, avoids insufficient stimulation or excessive assistance, and improves the accuracy, consistency, and stability of upper limb movements under the collaborative action of neural stimulation and exoskeleton.

[0064] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S202 can be implemented through the following S401 and S402, which are explained in detail below: S401. The EEG signal and the upper limb electromyography signal are processed by ICA independent component analysis to remove interference signals and extract the core features of EEG and the resting state features of electromyography.

[0065] Among them, ICA (Independent Component Analysis) is a blind source separation method that decomposes multi-channel observation signals into several statistically independent source signals; EEG core features refer to the time-frequency or spatial features related to motor imagery and after artifact removal processing; EMG resting-state features refer to the amplitude, spectrum, or envelope features that reflect the level of neuromuscular excitation without obvious voluntary contraction.

[0066] In one possible implementation, the acquired raw EEG and upper limb EMG signals are first bandpass filtered and normalized. ICA decomposition is then performed on the EEG and EMG signals respectively to obtain multiple independent signal components. Interference components unrelated to electrooculography, power frequency interference, baseline drift, etc., are identified and removed based on the spectral characteristics, spatial distribution features, or energy distribution of the components. On this basis, power spectral density, time-frequency energy, or spatial projection coefficients are extracted from the retained independent EEG components as core EEG features. The root mean square value, spectral centroid, or envelope change of the EMG signal in the resting or low-activation state are extracted as resting-state EMG features, forming a feature set that can be used for subsequent fusion.

[0067] It should be noted that EEG signals and EMG signals can be processed by independent ICA processes, and the number of channels and component screening thresholds for ICA decomposition can be adaptively set according to the device channel size and signal-to-noise ratio to avoid excessive decomposition leading to the loss of effective information.

[0068] As an example, under the conditions of 8-channel EEG and 4-channel EMG acquisition, the signal can be decomposed into independent components with the same number of channels. By setting the energy proportion threshold of the 0.5–40Hz frequency band, the EEG components can be screened, while components with abnormal high-frequency noise proportion in the EMG can be removed, thus obtaining stable feature input.

[0069] Based on the above steps, this step effectively reduces the impact of noise and individual state fluctuations on feature quality by independently separating and screening the components of EEG and EMG signals, thereby improving the reliability and discriminativeness of input information in subsequent feature fusion and motion intention prediction processes.

[0070] S402. After fusing the core features of EEG with the resting state features of EMG, the probability distribution of the movement intention is predicted and verified through a transfer learning model to obtain the movement intention of the target object.

[0071] Among them, the transfer learning model is a model constructed from a three-dimensional convolutional neural network and a long short-term memory network, which is initialized based on source domain motion intention samples.

[0072] In one possible implementation, the core EEG features are segmented according to a preset time window and time-aligned with the resting-state EMG features within the same time window. The aligned features are then concatenated along the channel and time dimensions to construct a three-dimensional fusion feature tensor containing time, channel, and feature type. The fusion features are input into a three-dimensional convolutional neural network initialized based on transfer learning to extract multi-scale spatiotemporal features and output a high-level feature representation. This high-level feature sequence is then input into a long short-term memory network to output the predicted probability distributions corresponding to different motion intentions. A confidence verification process is performed on the probability distributions to determine whether the maximum predicted probability is greater than a preset confidence threshold and whether the difference between the maximum probability and the second-highest probability meets a preset discrimination threshold. When both conditions are met, the corresponding motion intention is determined to be a valid prediction result and output as the current motion intention of the target object. When the above conditions are not met, the motion intention is not output, and a new prediction is performed.

[0073] It should be noted that the transfer learning 3D convolutional neural network is initialized using a 3D convolutional neural network trained on multi-class motion intention samples in the source domain. By freezing the first preset number of spatiotemporal feature extraction layers and fine-tuning the second preset number of feature representation layers, the network learns the motion intention features between different target objects, resulting in the initial 3D convolutional neural network for transfer learning. For example, the network parameters of the first three 3D convolutional feature extraction layers are frozen, and the last two 3D convolutional feature extraction layers and two high-level feature representation layers are fine-tuned during training.

[0074] As an example, when the model outputs predicted probabilities of flexion, extension, and grasping motion intentions of 0.78, 0.12, and 0.10, respectively, and the preset confidence threshold is 0.7 and the discrimination threshold is 0.3, the prediction result corresponding to flexion passes the verification, is determined as the current motion intention, and is output; if the maximum probability is only 0.55, it is determined to fail the verification and the motion intention is not output.

[0075] As an example, during the training of a transfer learning 3D convolutional neural network, a gradient descent-based backpropagation algorithm is used to update the parameters of the trainable layers; the initial learning rate is set to 1×10⁻⁶. -4 The training process employs either exponential decay or adaptive adjustment strategies based on the number of training epochs; the batch size is set to 16, and the training epochs are set to 50. To avoid overfitting, L2 regularization constraints and a dropout mechanism are introduced into the trainable layers (the last four layers) during training, with the dropout ratio set to 0.2. The loss function is a cross-entropy loss function based on motion intent classification, used to improve the stability and generalization ability of motion intent prediction. The training parameters can be adaptively adjusted according to the data scale, signal-to-noise ratio, and number of motion intent categories for different target objects, and do not constitute a limitation on the technical solution of this application.

[0076] Based on the above steps, this step achieves probabilistic prediction and verification of motor intention by spatiotemporally fusing EEG and EMG features and combining the three-dimensional convolutional network of transfer learning with temporal modeling capabilities. This reduces the impact of individual differences and state changes on the recognition results, thereby improving the stability and accuracy of motor intention prediction.

[0077] This application transforms the process of acquiring motor intention from direct judgment to a multi-level decision-making process by removing interference, extracting features, fusing and modeling, and performing probability verification on EEG and EMG signals. This effectively reduces the impact of noise interference and individual differences on the prediction results. At the same time, by leveraging the prior knowledge of the transfer learning model on the source domain samples, it achieves stable recognition of motor intention under limited sample conditions. Furthermore, confidence verification prevents unreliable prediction results from entering subsequent control links, thereby enhancing the reliability of overall collaborative control.

[0078] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S205 can be implemented by the following S501 to S504, which are explained in detail below: S501. Based on the target brachial plexus branches, determine the neural stimulation target markers corresponding to the motor intention and match the corresponding initial stimulation parameter templates.

[0079] Among them, the neural stimulation target identifier refers to the identification information used to uniquely characterize the terminal branches of the brachial plexus or their functional stimulation location; the initial stimulation parameter template refers to the set of parameter combinations pre-configured for different motor intentions and nerve branches, including the safe initial range of stimulation amplitude, stimulation frequency and stimulation pulse width.

[0080] In one possible implementation, the system receives the target brachial plexus branch identifier determined by the preceding steps and uses it together with the current motor intention as an index condition to query a preset motor intention, brachial plexus branch, and stimulation template mapping table. In the mapping table, each record corresponds to a recommended stimulation parameter range for a certain type of motor intention on a specific nerve branch, thereby outputting an initial stimulation parameter template that matches the target brachial plexus branch.

[0081] It should be noted that the initial stimulus parameter template is not the final execution parameter, but a starting configuration that meets the requirements of neural safety constraints and functional activation. It is used for subsequent instruction generation and adaptive regulation to avoid insufficient or excessive stimulation caused by directly using fixed or empirical parameters.

[0082] As an example, when the intended movement is elbow flexion and the target brachial plexus branch is the musculocutaneous nerve, the system can match stimulation parameter templates corresponding to elbow flexion, such as stimulation frequency 30–60Hz, stimulation amplitude 2–5mA, and stimulation pulse width 100–300μs.

[0083] Based on the above steps, this step introduces a parameter template matching mechanism based on motor intention and neural branches, which makes the selection of stimulation targets and the initial parameter configuration have clear functional orientation, reduces the uncertainty of parameter configuration, and improves the matching accuracy between neural stimulation and target action.

[0084] S502. Based on the motor intention and the initial stimulus parameter template, generate neural stimulation control instructions containing stimulus amplitude, stimulus frequency and stimulus pulse width.

[0085] Among them, the neural stimulation control command refers to the structured instruction data used to drive the neural stimulation module to execute, which includes at least the stimulation target identifier, stimulation amplitude, stimulation frequency and stimulation pulse width parameters.

[0086] In one possible implementation, the system generates specific combinations of stimulation amplitude, stimulation frequency, and stimulation pulse width based on the movement intention and initial stimulation parameter template, within the range of values ​​given in the parameter template, according to the current movement stage or preset strategy. The parameters are then encapsulated with the corresponding neural stimulation target identifiers to form neural stimulation control instructions that can be directly sent to the neural stimulation module.

[0087] It should be noted that the stimulation parameter generation process can adopt rule constraints or model output methods, but both must ensure that the parameter values ​​fall within the safe threshold and that the parameters satisfy the physical and physiological constraints of neural activation. The motor intention and initial stimulation parameter template refer to a pre-established set of neural stimulation parameters for different brachial plexus branches and their innervated muscle groups. These parameters describe the basic combination of stimulation amplitude, stimulation frequency, and stimulation pulse width that can safely and effectively induce contraction of the target muscle group under typical motor patterns. They can be configured according to actual needs and medical principles, and this application does not impose any limitations on this.

[0088] As an example, during the elbow flexion initiation phase, the system can select a stimulation frequency of 50Hz, a stimulation amplitude of 4mA, and a stimulation pulse width of 200μs from the initial template, and generate a neural stimulation control command containing the above parameters to drive the musculocutaneous nerve to produce elbow flexion-related muscle contraction.

[0089] Based on the above steps, this step achieves a clear mapping from abstract motor intentions to executable neural stimulation parameters, making the neural stimulation process controllable and repeatable, and improving the consistency and accuracy of stimulus execution.

[0090] S503. Determine the target joint that needs to be coordinated based on the motion intention, and generate exoskeleton coordination control instructions containing assist output parameters.

[0091] Among them, the exoskeleton collaborative control command refers to the control information used to control the upper limb exoskeleton to perform assisted movements, which includes at least the target joint identifier and the corresponding assisted output parameters.

[0092] In one possible implementation, the system determines the target joints that need to participate in the coordinated action based on the current motion intention, the preset motion intention, joint, and assist configuration rules, and generates corresponding assist output parameters in combination with the motion intention type to form exoskeleton coordinated control instructions.

[0093] It is important to note that the assist output parameters are used to assist, not replace, the voluntary movements generated by neural drives. Their output amplitude must be coordinated with the muscle contraction effect produced by neural stimulation to avoid over-assistance or movement conflicts. The preset movement intention, joint, and assist configuration rules are a structured mapping rule indexed by movement intention. This rule clarifies the target joints corresponding to the movement intention and the joint positions where the exoskeleton needs to provide assistance. Based on the human upper limb biomechanical model and exoskeleton structure, the rules pre-determine and store the joint combinations and assist parameter ranges corresponding to each movement intention in the offline phase, forming a rule base. During runtime, the system directly calls the corresponding rules based on the identified movement intention to generate exoskeleton assist control commands consistent with the actual movement needs of the human body.

[0094] As an example, when the intended movement is elbow flexion, the system identifies the elbow joint as the target joint and generates exoskeleton-assisted control commands containing assist torque or assist ratio to provide appropriate mechanical assistance during elbow flexion.

[0095] Based on the above steps, this step maps the motion intention to a specific exoskeleton assist configuration, so that the exoskeleton output is consistent with the human motion target, thereby improving the coordination and stability of human-machine collaborative movements.

[0096] S504. Apply timing synchronization constraints to neural stimulation control commands and exoskeleton coordination control commands to ensure that neural stimulation and exoskeleton assistance are executed in coordination within the same movement phase.

[0097] Among them, the temporal synchronization constraint refers to the consistency requirement of the execution time of the neural stimulation control command and the exoskeleton collaborative control command, which is used to limit the two to take effect in the same movement phase.

[0098] In one possible implementation, the system attaches a unified time stamp or motion phase stamp to the neural stimulation control command and the exoskeleton collaborative control command, respectively, and performs synchronous scheduling of the two types of commands in the execution control module according to the stamp, so as to ensure that the stimulation output and the exoskeleton assistance are executed simultaneously or in a predetermined order within the corresponding motion phase.

[0099] It should be noted that timing synchronization constraints do not require absolute simultaneous triggering, but rather ensure that the two are functionally within the same motion phase range, in order to avoid discontinuous actions caused by execution misalignment.

[0100] Based on the above steps, this step applies a unified temporal constraint to neural stimulation and exoskeleton assistance, enabling the two to form a synergistic relationship at the execution level, reducing the occurrence of movement lag or conflict, and improving the coherence and controllability of the overall movement output.

[0101] In one possible implementation of the embodiments of this application, combined with Figure 3 ,like Figure 6 As shown, the above S302 can be specifically implemented through the following S601 to S604, which are explained in detail below: S601. Construct a reinforcement learning strategy network based on joint motion state data and electromyographic feedback signals.

[0102] Among them, reinforcement learning strategy network refers to a decision model used to adaptively output control parameters based on human motion feedback during continuous motion. It includes: constructing a state space based on the motion stage, joint angle changes and electromyographic feedback signal intensity in joint motion state data; constructing an action space based on the stimulation frequency, stimulation amplitude, stimulation pulse width of neural stimulation control commands and the output parameters of exoskeleton collaborative control commands; and a reward function constrained by motion completion degree and abnormal muscle response.

[0103] In one possible implementation, the system first analyzes the joint motion state data, extracting the current motion phase, such as the initiation phase, maintenance phase, termination phase, joint angle change rate, and trajectory deviation. Simultaneously, it performs amplitude normalization and anomaly detection processing on the electromyographic feedback signal. These features collectively constitute the state vector for reinforcement learning. Furthermore, parameters such as stimulation frequency, stimulation amplitude, and stimulation pulse width from neural stimulation control commands, as well as assist output magnitude and assist change rate from exoskeleton collaborative control commands, are used as output dimensions in the continuous motion space to construct a differentiable policy network.

[0104] It should be noted that the reward function incorporates both positive incentives and constraint penalties in its design. The degree of motor completion, such as whether the joint reaches the target angle and whether the trajectory is smooth, is used as a positive reward, while abnormal muscle responses, such as sudden increases in electromyography and enhanced cocontraction, are used as penalties to avoid unsafe stimuli caused by simply pursuing joint displacement.

[0105] As an example, in an elbow flexion training scenario, the state space may include the current elbow angle, angular velocity, and mean electromyography value of the biceps brachii, while the motion space includes the amplitude of musculocutaneous nerve stimulation, stimulation frequency, and exoskeleton elbow joint assist torque. The reward function is calculated based on whether the flexion angle stably reaches the target range.

[0106] Based on the above steps, this step systematically introduces human feedback information into the reinforcement learning decision model, enabling neural stimulation and exoskeleton control to have online adaptive adjustment capabilities, thereby significantly improving the control strategy's responsiveness to individual differences and dynamic changes in motion.

[0107] S602. Based on joint motion state data and electromyographic feedback signals, determine the current motion stage and form the current state.

[0108] The current state refers to the reinforcement learning state vector determined by joint motion state data and electromyographic feedback signals within a certain control cycle, which reflects the actual motion execution of the target object at that moment.

[0109] In one possible implementation, joint angles, joint trajectory points, and corresponding electromyographic feedback signals are collected in real time during each control cycle. These are then compared with the standard motion model corresponding to the motion intention to determine the current motion stage. The determined motion stage identifier, joint angle deviation, and electromyographic activation level are integrated into a unified format of state input and fed into the reinforcement learning policy network.

[0110] Based on the above steps, this step accurately characterizes the current motion state, enabling the reinforcement learning strategy to adopt differentiated control strategies for different motion stages, thus avoiding the problem of control mismatch caused by uniform parameters in different stages.

[0111] S603. Based on the current state, output the corresponding action parameters through the reinforcement learning policy network, and update the neural stimulation control instructions and exoskeleton collaborative control instructions.

[0112] Among them, action parameters refer to the continuous set of parameters output by the reinforcement learning policy network, which are used to regulate neural stimulation control commands and exoskeleton coordination control commands.

[0113] In one possible implementation, after receiving the current state, the reinforcement learning policy network calculates and outputs the corresponding action vector, which is directly mapped to the new stimulation frequency, stimulation amplitude, stimulation pulse width, and exoskeleton assistance output parameters. The system then updates the original control commands and executes the updated neural stimulation and exoskeleton assistance in the next control cycle.

[0114] Based on the above steps, this step achieves continuous and adaptive updates of neural stimulation and exoskeleton control parameters, enabling the collaborative control strategy to be dynamically adjusted according to human feedback, thereby improving the stability and consistency of the training process.

[0115] S604. After a single exercise is completed, calculate the reward value based on the exercise effect and update the reinforcement learning policy network.

[0116] The reward value refers to the scalar feedback calculated based on the exercise effect and muscle response after a single exercise, which is used to guide the parameter updates of the reinforcement learning strategy network.

[0117] In one possible implementation, the reward value is obtained by weighting a motion completion index and a muscle response safety index. The motion completion index characterizes the consistency between the actual joint movement and the target movement, while the muscle response safety index characterizes whether muscle activation is within a reasonable range during neural stimulation. First, the motion completion index is calculated based on joint movement data, including the error between the final angle reached by the joint and the target angle, the average deviation between the joint movement trajectory and a preset reference trajectory, and a smoothness index of the movement process. When the joint reaches the target angle and the trajectory deviation is small, the motion completion index is assigned a higher value. The system calculates the muscle response safety index based on electromyographic feedback signals. By analyzing whether the amplitude of the electromyographic signal exceeds a preset threshold, or whether there are abnormal spikes or sustained high activation states, abnormal muscle responses are penalized and quantified.

[0118] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0119] Based on the above steps, this step continuously modifies the control strategy through a closed-loop reward mechanism, enabling the system to gradually approach the optimal combination of stimulus and assistance during multiple training sessions, thereby achieving personalized and progressive neural and exoskeleton synergistic control effects.

[0120] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0121] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for controlling brachial plexus and exoskeleton based on brain-computer interface, characterized in that, The method comprises the following steps: Collecting electroencephalogram signals and upper limb electromyogram signals during motor imagery of a target object; Predicting the motor intention of the target object through a neural network based on the electroencephalogram signals and the upper limb electromyogram signals; wherein the neural network is a transfer learning model based on ICA independent component analysis technology; Based on the motor intention, determining a target brachial plexus branch according to a preset mapping rule of motor intention and brachial plexus; Based on the target brachial plexus branch and the motor intention, generating a neural stimulation control instruction and an exoskeleton cooperative control instruction; wherein the neural stimulation control instruction comprises stimulation amplitude, stimulation frequency and stimulation pulse width parameters; the exoskeleton cooperative control instruction comprises target joint and assistance output parameters; According to the neural stimulation control instruction, electrically stimulating the target brachial plexus branch to drive the corresponding muscle to contract, and at the same time, controlling the upper limb exoskeleton to apply assistance to the corresponding joint according to the exoskeleton cooperative control instruction.

2. The method of claim 1, wherein, After controlling the upper limb exoskeleton to apply assistance to the corresponding joint according to the exoskeleton cooperative control instruction, the method further comprises: Obtaining joint motion state data and electromyogram feedback signals of the upper limb; wherein the joint motion state data comprises joint angle information and joint motion trajectory information; According to the deviation between the joint motion state data and the electromyogram feedback signals and a preset target, dynamically adjusting the neural stimulation control instruction and the exoskeleton cooperative control instruction through deep reinforcement learning.

3. The method of claim 1, wherein, The method of predicting the motor intention of the target object through a neural network based on the electroencephalogram signals and the upper limb electromyogram signals comprises the following steps: Removing interference signals from the electroencephalogram signals and the upper limb electromyogram signals through ICA independent component analysis technology to extract electroencephalogram core features and electromyogram resting state features; After fusing the electroencephalogram core features and the electromyogram resting state features, predicting the probability distribution of the motor intention through a transfer learning model and verifying it to obtain the motor intention of the target object; wherein the transfer learning model is a model constructed by a three-dimensional convolutional neural network and a long short-term memory network based on source domain motor intention samples.

4. The method of claim 3, wherein, The method of predicting the probability distribution of the motor intention through a transfer learning model after fusing the electroencephalogram core features and the electromyogram resting state features to obtain the motor intention of the target object comprises the following steps: Segmenting the electroencephalogram core features according to a preset time window, time-aligning them with the electromyogram resting state features in the same time window, and concatenating the aligned features according to the channel dimension to construct fusion features, wherein the fusion features are three-dimensional fusion feature tensors containing time sequence information and spatial distribution information; Inputting the fusion features into a pre-transfer three-dimensional convolutional neural network to extract the spatio-temporal correlation features in the fusion features and obtain high-level spatial features; Inputting the high-level spatial features into a long short-term memory network to predict the prediction probability distribution of different motor intentions; Verifying the prediction probability distribution for confidence, and determining the motor intention of the target object when the probability value corresponding to the target motor intention is greater than a preset confidence threshold.

5. The method of claim 4, wherein, The migration learning three-dimensional convolutional neural network is initialized by a three-dimensional convolutional neural network trained on a source domain multi-class motion intention sample, and migration learning motion intention features between different target objects is performed by freezing a first preset number of spatial and temporal feature extraction layers and fine-tuning a second preset number of feature representation layers, so as to obtain the migration learning initialized three-dimensional convolutional neural network.

6. The method of claim 1, wherein, The generation of the neural stimulation control instruction and the exoskeleton cooperative control instruction based on the target brachial plexus branch and the motion intention comprises: According to the target brachial plexus branch, a neural stimulation target point identifier corresponding to the motion intention is determined, and a corresponding initial stimulation parameter template is matched; Based on the motion intention and the initial stimulation parameter template, a neural stimulation control instruction containing a stimulation amplitude, a stimulation frequency and a stimulation pulse width is generated; According to the motion intention, a target joint requiring cooperative action is determined, and an exoskeleton cooperative control instruction containing a power assistance output parameter is generated; A timing synchronization constraint is imposed on the neural stimulation control instruction and the exoskeleton cooperative control instruction, so as to ensure that the neural stimulation and the exoskeleton power assistance are cooperatively executed in the same motion phase.

7. The method of claim 2, wherein, The dynamic adjustment of the neural stimulation control instruction and the exoskeleton cooperative control instruction according to the deviation between the joint motion state data and the preset target and the electromyographic feedback signal through deep reinforcement learning comprises: A reinforcement learning strategy network is constructed according to the joint motion state data and the electromyographic feedback signal; wherein the reinforcement learning strategy network comprises: a state space constructed based on the motion phase, the joint angle change in the joint motion state data and the intensity of the electromyographic feedback signal; an action space constructed with the stimulation frequency, the stimulation amplitude, the stimulation pulse width of the neural stimulation control instruction and the output parameter of the exoskeleton cooperative control instruction; a reward function with the motion completion degree and the abnormal muscle reaction as constraints; Based on the joint motion state data and the electromyographic feedback signal, the current motion phase is determined and the current state is formed; Based on the current state, the corresponding action parameters are output through the reinforcement learning strategy network, and the neural stimulation control instruction and the exoskeleton cooperative control instruction are updated; After a single motion is completed, the reward value is calculated according to the motion effect, and the reinforcement learning strategy network is updated.

8. A system for coordinated control of brachial plexus and exoskeleton based on brain-computer interface, for implementing any of the methods of claims 1-7, characterized in that, The system comprises: a signal acquisition module, a motion intention prediction module, a brachial plexus branch determination module, a cooperative control instruction generation module, an execution control module and an adaptive regulation and control module; The signal acquisition module is configured to acquire electroencephalogram signals, upper limb electromyographic signals and joint motion state data during the target object's motor imagery, and the joint motion state data comprises joint angle information and joint motion trajectory information; The motion intention prediction module is configured to predict the target object's motion intention based on the electroencephalogram signals and the upper limb electromyographic signals through a neural network; the neural network is a migration learning model based on ICA independent component analysis technology; The brachial plexus branch determination module is configured to determine the target brachial plexus branch according to a preset motion intention and brachial plexus mapping rule based on the motion intention. The synergic control instruction generation module is configured to generate a neural stimulation control instruction and an exoskeleton synergic control instruction based on the target brachial plexus branch and the motion intention, wherein the neural stimulation control instruction comprises stimulation amplitude, stimulation frequency, and stimulation pulse width parameters; and the exoskeleton synergic control instruction comprises target joints and assistance output parameters; The execution control module is configured to perform electrical stimulation on the target brachial plexus branch to drive corresponding muscle contraction according to the neural stimulation control instruction, and control the upper-limb exoskeleton to apply assistance to corresponding joints according to the exoskeleton synergic control instruction. The adaptive regulation module is configured to acquire upper-limb joint motion state data and electromyographic feedback signals after execution control, and dynamically adjust the neural stimulation control instruction and the exoskeleton synergic control instruction through deep reinforcement learning according to deviations between the joint motion state data and the electromyographic feedback signals and a preset target.

9. The system of claim 8, wherein, The system further comprises: A hardware integration and timing synchronization module configured to synchronize interconnection and timing among the brain-computer interface, the brachial plexus stimulation module, the sensor, and the upper-limb exoskeleton.

10. The system of claim 8, wherein, The motion intention prediction module is further configured to: remove interference signals from electroencephalogram signals and upper-limb electromyographic signals through ICA independent component analysis technology, extract electroencephalogram core features and electromyographic resting-state features, and fuse the electroencephalogram core features and the electromyographic resting-state features to predict a probability distribution of the motion intention through a transfer learning model and verify the probability distribution to obtain the target object motion intention. The transfer learning model comprises a transfer learning three-dimensional convolutional neural network and a long short-term memory network. ​

Images