A non-invasive closed-loop brain-computer interface system for precise recognition of left and right lower limb motor intentions from the contralateral upper limb.

By constructing a composite motor imagery paradigm and a dynamic ERD threshold optimization method, combined with audiovisual feedback, high-precision recognition of left and right lower limb motor intentions and closed-loop neural modulation were achieved. This solved the problems of insufficient recognition accuracy and stability in existing systems, and promoted the reconstruction of neural function and rehabilitation effects.

CN123116002APending Publication Date: 2026-11-17BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611261277.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing brain-computer interface systems have low accuracy and insufficient stability in recognizing the movement intentions of the left and right lower limbs, lack effective sensory feedback mechanisms, and traditional systems lack personalized control, making it difficult to meet the needs of different rehabilitation stages.

Method used

A non-invasive closed-loop brain-computer interface system for recognizing the left and right lower limb motor intentions of the contralateral upper limb was constructed. By introducing a composite motor imagery paradigm, a time-frequency cross-view cross-attention fusion algorithm, and a dynamic ERD threshold adaptive optimization method, combined with an audiovisual feedback module, high-precision motor intention recognition and closed-loop neural modulation were achieved.

Benefits of technology

It significantly improved the accuracy of motor intention recognition and system robustness, promoted neural function reconstruction, and enhanced rehabilitation effects and human-machine coordination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116002A_ABST
    Figure CN123116002A_ABST
Patent Text Reader

Abstract

This invention discloses a non-invasive closed-loop brain-computer interface system and method for accurate recognition of left and right lower limb motor intentions by combining contralateral upper limb movement. It belongs to the field of brain-computer interface and rehabilitation engineering technology. This invention enhances the distinguishability of left and right lower limb motor intentions at the neural level by constructing a composite motor paradigm that combines contralateral upper limb motor execution. Regarding motor intention recognition, differentiated technical solutions are proposed for different rehabilitation stages: in the early and middle stages of rehabilitation, a trained recognition method based on time-frequency cross-view cross-attention fusion is used to achieve high-precision, long-term stable decoding; in the middle and late stages of rehabilitation, a training-free technique based on dual-timescale dynamic ERD threshold adaptive optimization is used to achieve rapid response. Furthermore, this invention constructs a complete closed-loop control system of "active imagination—decoding command—motor execution—sensory feedback." This invention significantly improves the accuracy and robustness of left and right lower limb motor intention recognition, effectively promoting neural plasticity reconstruction and motor function recovery.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface and rehabilitation engineering technology, specifically relating to a non-invasive closed-loop brain-computer interface system and method for accurate recognition of left and right lower limb movement intentions in conjunction with contralateral upper limb movement. Background Technology

[0002] Brain-computer interface (BCI) is a technology that enables information exchange between the brain and external devices, and it has been widely used in the field of neurorehabilitation in recent years. For patients with central nervous system injuries such as spinal cord injury and stroke, the root cause of lower limb motor dysfunction is that the brain's motor commands cannot be effectively transmitted to peripheral actuators through damaged neural pathways. At the same time, the long-term lack of sensory-motor feedback further leads to weakened motor cortex representation and decreased neural plasticity, thus severely restricting the recovery of patients' motor function.

[0003] Existing brain-computer interface-based lower limb rehabilitation systems mostly use motor imagery EEG signals as control input, driving exoskeletons or rehabilitation devices to complete training movements by recognizing the user's motor intentions. However, many problems still exist in practical applications. On the one hand, the EEG signals corresponding to the left and right lower limb motor intentions have high similarity in spatial distribution and spectral characteristics, resulting in low classification accuracy and insufficient stability, making reliable control difficult. On the other hand, traditional motor intention decoding algorithms suffer from limitations in cross-subject generalization, poor robustness, long training time, and high response latency. In addition, existing systems generally lack effective sensory feedback mechanisms, making it difficult to form a stable neural modulation closed loop, which is not conducive to promoting neural pathway reconstruction. At the same time, existing exoskeleton control strategies are usually passive rehabilitation modes, with poor human-machine coordination, making it difficult to achieve effective rehabilitation. Finally, existing systems mostly adopt static and single technical strategies, failing to consider the dynamic evolution of patients' neuroplasticity and EEG characteristics in the early, middle, and late stages of rehabilitation, making it difficult to meet the actual needs of different rehabilitation stages.

[0004] Therefore, how to construct a brain-computer interface system that can improve the accuracy of left and right lower limb movement intention recognition, enhance system robustness, and realize closed-loop neural modulation and personalized control has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a non-invasive closed-loop brain-computer interface system for accurate recognition of the left and right lower limb movement intentions of the contralateral upper limb. By constructing a brain-computer interface decoding method and closed-loop control system that conforms to the human motor neural mechanism, the system can achieve high-precision recognition and real-time control of the left and right lower limb movement intentions, thereby driving the lower limb exoskeleton to complete continuous, smooth and physiologically consistent movement output, and further promoting the reconstruction of neural function and the recovery of motor ability.

[0006] A non-invasive closed-loop brain-computer interface system for accurate recognition of left and right lower limb movement intentions from the contralateral upper limb includes a hardware device and a signal processing unit. The hardware device includes an electroencephalogram (EEG) signal acquisition module and an exoskeleton-assisted movement module. The signal processing terminal includes an EEG processing module, a module with a trained algorithm, and a module without a trained algorithm. The EEG signal acquisition module is used to acquire EEG signals generated by the user based on a compound motor imagery paradigm. The EEG processing module is used to preprocess the acquired EEG signals; The trained algorithm module is used to identify motor intentions from EEG signals in the early and middle stages of rehabilitation. The training-free algorithm module is used to identify motor intentions from EEG signals during the mid- and late-stages of rehabilitation. The exoskeleton-assisted motion module is used to receive the identified motion intention results and execute the corresponding motion actions.

[0007] Preferably, the composite motion imagination paradigm is as follows: While performing unilateral lower limb motor imagery, the contralateral upper limb is introduced to perform tasks in coordination, thereby activating multiple functional areas in the brain's motor cortex and enhancing the functional connections between different limb motor representations. When the system instructs the user to step with their left lower limb, the user's brain visualizes the movement of "stepping with the left leg," while the right upper limb performs the actual movement of "swinging the arm forward" or "clenching the fist." When the system instructs the user to step with their right lower limb, the user's brain visualizes the movement of "stepping with the right leg," while the left upper limb performs the actual movement of "swinging the arm forward" or "clenching the fist." Through the coordination of the actual physical swing of the arms and the mental stepping of the opposite leg, a cross-gait rhythm of "right arm and left leg" and "left arm and right leg" is formed.

[0008] Preferably, the preprocessing method of the EEG processing module is as follows: Adaptive notch filtering and independent component analysis are used together to remove power frequency interference and electromyography artifacts. The noise is then processed by co-space mode and continuous wavelet transform respectively before being input into the trained algorithm module and the untrained algorithm module.

[0009] Preferably, the trained algorithm module employs a trained recognition algorithm based on time-frequency cross-view cross-attention fusion, specifically including: A dual-branch network is constructed for feature extraction. The temporal branch uses multi-scale convolution to extract local temporal waveform features and captures the global long-distance dependencies of EEG sequences through a multi-head self-attention mechanism to obtain a temporal feature view. Frequency domain branching extracts the degree of decrease or increase of key frequency band power relative to the baseline state and the energy ratio of key frequency bands. Based on this, multi-scale convolution is used to further extract local spectral patterns at different frequency scales to form a frequency domain feature view. Based on an attention mechanism, the temporal feature view output from the temporal branch and the frequency feature view output from the frequency branch are bidirectionally complementary, and cross-view is introduced in this process. Figure 1 The consistency loss function forces the alignment of the feature distributions of the two views in the high-level semantic space. A dynamic reweighted fusion module is introduced to calculate weight values ​​in real time based on the classification contribution of the two views and to dynamically reweight the features after interaction. The fused features are fed into a classifier with a subject-specific adapter to complete motion intent recognition.

[0010] Preferably, the classifier trains the main network using training data from multiple subjects during the pre-training phase, freezes the main parameters of the pre-trained main network during the new user phase, fine-tunes the subject-specific adapter and classification layer parameters using only a small amount of individual calibration data, and continuously iterates and optimizes the model parameters using an adaptive update strategy during the subsequent rehabilitation process, outputting gait intention recognition results.

[0011] Preferably, the training-free algorithm module employs a dynamic event correlation desynchronization threshold adaptive optimization method based on dual time scales, specifically including: Short-term optimization: After each trial, the ERD threshold is fine-tuned based on task success or failure, recent success rate trends, and historical adjustment directions to achieve a response to the user's real-time status. Long-term scale optimization: After each set number of valid trials, a periodic cosine annealing process is triggered, forcing the ERD threshold to be perturbed by a cosine curve within a preset reasonable range to actively explore potential better solutions, and then weighted and fused with the adaptive threshold.

[0012] Preferably, the training-free algorithm module employs a dynamic event correlation desynchronization threshold adaptive optimization method based on dual time scales, specifically executed according to the following steps: First, short-term optimization is triggered immediately after each single trial. The system comprehensively evaluates the success or failure of the current trial, the accuracy trend of recent consecutive trials, and the adjustment direction of historical thresholds. It then makes rapid fine-tuning of the ERD threshold in small steps to track and respond to the user's instantaneous EEG state fluctuations in real time.

[0013] in, The current ERD threshold, The ERD threshold is the one updated shortly. This indicates that the ERD determination result of the current trial is less than the ERD threshold. A value of 1; greater than the ERD threshold. The value is 0; This indicates the accuracy rate for consecutive recent trials that fall below the ERD threshold. This indicates the preset target accuracy rate. Adjust the step size based on the base. This is an adaptive adjustment coefficient. Through the above method, the threshold can be dynamically adjusted trial by trial based on the user's recent EEG response status.

[0014] Secondly, after each set number of valid experiments is completed, long-term scale optimization is triggered. The system starts a periodic cosine annealing process, which forcibly generates a perturbation threshold that fluctuates smoothly according to the cosine curve within a preset reasonable range. In this way, it actively jumps out of the local optimum solution trapped due to short-term fine-tuning and explores potential better threshold ranges.

[0015] in, Let be the perturbation amplitude, and k be a periodic variable for long-term optimization, so as to periodically explore other threshold regions near the current threshold and reduce the possibility that the threshold will fall into a local stable state for a long time.

[0016] Finally, the short-term adaptive threshold and the long-term cosine annealing threshold are weighted and fused to obtain the final dynamic ERD threshold, which is used to determine the motor intention of the EEG signal.

[0017] Preferably, the hardware device also includes an audiovisual feedback module, which is used to transmit the motion assessment information generated by the exoskeleton movement to the user in real time in the form of audiovisual feedback, forming a closed-loop neural modulation.

[0018] A method for accurately recognizing and controlling the movement intentions of the left and right lower limbs includes the following steps: S1: Guide users to perform a compound motor imagery paradigm, that is, while imagining motor movements of one lower limb, introduce the coordinated movement of the contralateral upper limb to perform the task. S2: The EEG signal acquisition module acquires the user's EEG signal during motor imagery, and the EEG processing module performs filtering and noise reduction preprocessing. S3: Determine the user's rehabilitation stage: If in the early or middle stage of rehabilitation, the trained algorithm module is invoked, and a time-frequency feature extraction algorithm based on time-frequency cross-view cross-attention fusion is used to extract time-frequency features and perform dynamic reweighting and individual adaptation, outputting the classification result; if in the middle or late stage of rehabilitation, the untrained algorithm module is invoked, and a motion intention recognition method based on dynamic event correlation desynchronization threshold adaptive optimization based on dual time scales is used, outputting the classification result. S4: Generate exoskeleton control commands based on the motion intention recognition results, and drive the exoskeleton assistive motion module to lead the user's limbs to perform continuous, smooth and physiologically consistent movements; S5: Transmits the exoskeleton's motion status and sensory feedback information to the user's brain in real time, forming a closed-loop neural regulation and promoting neural function reconstruction.

[0019] The present invention has the following beneficial effects: (1) By constructing a composite motor imagery EEG paradigm, the distinguishability of left and right lower limb motor intentions at the neural level was significantly enhanced, and the decoding accuracy of EEG signals was improved.

[0020] (2) By introducing time-frequency cross-view cross-attention fusion, dynamic reweighting and subject-specific adaptation mechanisms, the complementary information of the time and frequency domains of EEG signals can be fully utilized to improve feature representation ability and adaptability among different subjects.

[0021] (3) By adopting a dual-timescale dynamic ERD threshold adaptive optimization method, combining short-term trial-by-trial adjustment and long-term periodic optimization, the judgment threshold can be dynamically adjusted according to the user's EEG state, thereby achieving fast and adaptive motor intention recognition without retraining the model.

[0022] (4) By constructing a closed-loop feedback mechanism, the reconstruction of neural plasticity is effectively promoted, thereby improving the overall stability, naturalness and rehabilitation effect of the system. Attached Figure Description

[0023] Figure 1 This is a structural diagram of a non-invasive closed-loop brain-computer interface system for accurately recognizing the left and right lower limb movement intentions of a composite upper limb movement according to the present invention.

[0024] Figure 2 This invention provides a motion imagination paradigm for recognizing the combined upper limb and lower limb movement intentions. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] This invention, starting from neural activation mechanisms, innovatively constructs a composite motor imagery paradigm. By simultaneously incorporating contralateral upper limb coordinated movement to perform a task during unilateral lower limb motor imagery, it activates multiple functional areas in the brain's motor cortex and enhances the functional connectivity between different limb motor representations. Compared to traditional single motor imagery paradigms, this composite paradigm significantly improves the spatial separation and expression intensity of neural signals, providing a more stable and discriminative EEG feature basis for subsequent motor intention recognition.

[0027] Regarding motor intention recognition, this invention proposes differentiated technical solutions for different rehabilitation stages. In the early and middle stages of rehabilitation, a trained recognition method based on time-frequency cross-view cross-attention fusion is used; in the middle and late stages of rehabilitation, a training-free technique based on dual-timescale dynamic ERD threshold adaptive optimization is used.

[0028] Because subjects exhibit weak motor imagery characteristics during the early and mid-stages of rehabilitation, accurate identification of left and right leg steps is crucial, and response times are slow. Therefore, this invention addresses this issue by providing a training algorithm for accurate identification of left and right leg motor intentions across subjects, and proposing a time-frequency cross-view cross-attention fusion algorithm based on specific subject adaptation. The core of this invention is a training mode designed to achieve high-precision, long-term stable decoding of left and right leg motor intentions. Through an integrated technical architecture encompassing "feature enhancement - dual-view extraction - cross-domain fusion - dynamic reweighting - individual adaptation - closed-loop optimization," it overcomes challenges such as the similar spatiotemporal distribution of lower limb motor intention signals, significant individual differences among subjects, and dynamic drift of neural signals during rehabilitation.

[0029] In the feature extraction stage, a dual-branch network is constructed. The temporal branch uses "multi-scale convolution + multi-head self-attention" to capture local temporal features such as potential deflection during the motion preparation period and amplitude changes during the execution period, as well as global dependencies across time points. The frequency domain branch focuses on specific features such as ERD / ERS intensity and frequency band energy ratio through "specific feature extraction + multi-scale convolution", highlighting the difference in frequency domain response of the brain's contralateral cortex when the left and right legs intend to move.

[0030] Cross-view interaction enables bidirectional information exchange between time-domain and frequency-domain features, combined with cross-view... Figure 1 The consistency loss function aligns the feature distribution, and then a dynamic reweighting fusion module is introduced. The weight values ​​are calculated in real time based on the classification contribution of the two views, and the features of the two views after interaction are dynamically reweighted to strengthen the key effective feature representation and weaken the interference of redundant information.

[0031] To adapt to individual differences and changes in neural plasticity, the algorithm embeds a subject-specific adapter. During the pre-training phase, it uses data from similar subjects to learn common feature patterns. New subjects only need a small amount of training data to complete the adapter fine-tuning. At the same time, an adaptive update strategy for the rehabilitation process is designed, which automatically collects a small amount of data every week to fine-tune the parameters and ensures long-term decoding stability.

[0032] Because the motor imagery signals of the subjects are stronger in the middle and late stages, the system response time can be accelerated by setting a threshold. Therefore, considering the user's training burden and practical application needs in the middle and late stages of rehabilitation, this invention further proposes a training-free dynamic threshold adaptive optimization method based on dual time scales. The training-free mode will deeply analyze the neurophysiological mechanism of motor intention EEG signals and propose a dynamic ERD threshold adaptive optimization technology based on dual time scales. This invention innovatively introduces a dynamic ERD threshold adaptive optimization mechanism, which integrates a dual time scale strategy: on a short-term scale, after each trial, the threshold is fine-tuned based on task success or failure, recent accuracy trends, and historical adjustment directions to achieve rapid response to the user's real-time state; on a long-term scale, a periodic cosine annealing process is triggered every 10 valid trials, forcing the threshold to be perturbed by a cosine curve within a preset reasonable range, actively exploring potential better solutions, and weighted and fused with the adaptive threshold to escape local optima. The short-term adaptive mechanism ensures the real-time response capability of the system. Immediate fine-tuning after each trial allows for rapid adaptation to immediate changes in participants' states, such as fatigue and fluctuations in attention, avoiding consecutive failures or successes due to fixed thresholds and maintaining the task's challenge within a reasonable range. The periodic annealing mechanism provides long-term optimization capabilities. Every N trials, a new threshold range is forcibly explored, effectively preventing the system from getting trapped in local optima.

[0033] Furthermore, this invention constructs a complete closed-loop brain-computer interface control system. In this system, EEG signals serve as control inputs, which are processed by a motor intention recognition module to generate control commands that drive the exoskeleton to perform actions. Simultaneously, the exoskeleton's motion state and sensory feedback information are transmitted to the user's brain in real time, thus forming a closed-loop control process of "active imagination—decoding commands—motor execution—sensory feedback." This closed-loop mechanism can continuously strengthen the functional connection between the brain and the peripheral system, promote the reconstruction of neural plasticity, gradually stabilize motor representation, and ultimately restore walking ability close to a natural gait.

[0034] By comprehensively applying the above technical solutions, this invention can not only significantly improve the accuracy and response speed of left and right lower limb movement intention recognition, but also achieve efficient collaborative control between brain-computer interface and exoskeleton system, and promote neural function reconstruction under the action of closed-loop feedback mechanism, thereby providing a highly efficient, safe rehabilitation solution with good clinical application prospects for patients with lower limb motor dysfunction.

[0035] Example 1: System Overall Structure like Figure 1As shown, a non-invasive closed-loop brain-computer interface system for accurate recognition of left and right lower limb movement intentions in conjunction with contralateral upper limb movement includes a hardware device and a signal processing unit. The hardware device includes an EEG signal acquisition module, an exoskeleton-assisted movement module, and an audiovisual feedback module; the signal processing unit includes an EEG processing module, a module with a trained algorithm, and a module without a trained algorithm.

[0036] The hardware includes an EEG signal acquisition module and an exoskeleton or other motion-assisted module; the signal processing module includes an EEG processing module, a trained algorithm module, and a non-trained algorithm module; the EEG signal acquisition module is used to acquire EEG signals generated by the user based on a compound motor imagery paradigm; the EEG processing module is used to preprocess the acquired EEG signals; the trained algorithm module is used to recognize motor intentions from EEG signals in the early and middle stages of rehabilitation; the non-trained algorithm module is used to recognize motor intentions from EEG signals in the middle and late stages of rehabilitation; and the exoskeleton-assisted motion module is used to receive the recognized motor intention results and execute the corresponding motor actions.

[0037] Example 2: Composite Motion Imagination Paradigm like Figure 2 The diagram shows the EEG paradigm of the composite multiple motor imagery of the present invention. It constructs an EEG coding mechanism for composite multiple motor imagery actions, integrates typical motor imagery tasks such as lower limb and contralateral upper limb movements, enhances the spatial specificity of neural activation and enriches the category expression of motor intentions, and achieves accurate and rapid recognition of the motor intentions of the left and right legs of the lower limbs.

[0038] While performing unilateral lower limb motor imagery, the contralateral upper limb is introduced to perform coordinated motor tasks, thereby activating multiple functional areas in the brain's motor cortex, enhancing the functional connections between different limb motor representations, improving the spatial separation and expression intensity of neural signals, and driving the limb to step by decoding the exoskeleton on the same side.

[0039] When the system instructs the subject to step with their left lower limb, the subject's brain visualizes "stepping with the left leg" (MI); simultaneously, the subject's right upper limb performs actual motor execution (ME) tasks such as "swinging the arm forward" or "clenching the fist." When the system instructs the subject to step with their right lower limb, the subject's brain visualizes "stepping with the right leg" (MI); simultaneously, the subject's left upper limb performs actual motor execution (ME) tasks such as "swinging the arm forward" or "clenching the fist." Through the actual physical swing of the arm, coordinated with the mental stepping of the opposite leg, a cross-gait rhythm of "right arm, left leg" and "left arm, right leg" is formed.

[0040] Example 3: Preprocessing of EEG Signals First, adaptive notch filtering and independent component analysis (ICA) are used together to denoise, accurately removing noise components such as power frequency interference and electromyography artifacts. Then, after processing by common spatial pattern (CSP) and continuous wavelet transform (CWT), the noise is input into the feature extraction module.

[0041] Example 4: Includes a training algorithm module. The feature extraction process is divided into parallel time-domain and frequency-domain branches.

[0042] The temporal branch uses multi-scale convolution to extract local temporal waveform features and captures the global long-range dependencies of EEG sequences through a multi-head self-attention mechanism; The frequency domain branch extracts two types of frequency domain features related to motion intent: one is the ERD / ERS response feature, which is used to characterize the degree to which the power of key frequency bands decreases or increases relative to the baseline state during the motion task; the other is the key frequency band energy proportion feature, which is used to characterize the relative proportion of the energy of each key frequency band in the total energy of the target frequency range. Based on this, multi-scale convolution is used to further extract local spectral patterns at different frequency scales to form a frequency domain feature view. Next, based on the attention mechanism, the temporal feature view output from the temporal branch and the frequency feature view output from the frequency branch are bidirectionally complementary, and cross-view is introduced in this process. Figure 1 The consistency loss function forces the alignment of the feature distributions of the two views in the high-level semantic space. Subsequently, the temporal and frequency domain features, after cross-view interaction, are input into the dynamic reweighted fusion module. The system calculates the weight values ​​in real time based on the actual classification contribution of the temporal and frequency domain features in the current sample, thus completing the dynamic weighted fusion of features. Finally, the fused features are fed into a classifier with a subject-specific adapter to complete gait intention recognition.

[0043] The training strategy consists of two phases: During the pre-training phase, the classifier's main network is trained using training data from multiple subjects, enabling it to learn common motion intention features among different subjects. During the fine-tuning phase, the main parameters of the pre-trained main network are frozen, and the parameters of the subject-specific adapter and classification layer are quickly fine-tuned using only a small amount of individual calibration data. An adaptive update strategy is configured in the subsequent rehabilitation process, and the model parameters are continuously iterated and optimized based on long-term data performance, ultimately outputting accurate gait intention recognition results.

[0044] Example 5: No training algorithm module The training-free algorithm module employs a dynamic ERD threshold adaptive optimization method based on dual time scales, specifically including: Short-term optimization: After each trial, the ERD threshold is fine-tuned based on task success or failure, recent accuracy trends, and historical adjustment directions to achieve rapid response to the user's real-time status. Long-term scale optimization: After each set number of valid trials, a periodic cosine annealing process is triggered, forcing the ERD threshold to be perturbed by a cosine curve within a preset reasonable range to actively explore potential better solutions, and then weighted and fused with the adaptive threshold.

[0045] The specific process includes: First, short-term optimization is triggered immediately after each single trial. The system comprehensively evaluates the success or failure of the current trial, the accuracy trend of recent consecutive trials, and the adjustment direction of historical thresholds. It then makes rapid fine-tuning of the ERD threshold in small steps to track and respond to the user's instantaneous EEG state fluctuations in real time. (1.1) in, The current ERD threshold, The ERD threshold is the one updated shortly. This indicates that the ERD determination result of the current trial is less than the ERD threshold. A value of 1; greater than the ERD threshold. The value is 0. This indicates the accuracy rate for consecutive recent trials that fall below the ERD threshold. This indicates the preset target accuracy rate. Adjust the step size based on the base. This is an adaptive adjustment coefficient. Through the above method, the threshold can be dynamically adjusted trial by trial based on the user's recent EEG response status.

[0046] Secondly, after each set number of valid experiments is completed, long-term scale optimization is triggered. The system starts a periodic cosine annealing process, which forcibly generates a perturbation threshold that fluctuates smoothly according to the cosine curve within a preset reasonable range. In this way, it actively jumps out of the local optimum solution trapped due to short-term fine-tuning and explores potential better threshold ranges. (1.2) in, Let be the perturbation amplitude, and k be a periodic variable for long-term optimization, so as to periodically explore other threshold regions near the current threshold and reduce the possibility that the threshold will fall into a local stable state for a long time.

[0047] Finally, the short-term adaptive threshold and the long-term cosine annealing threshold are weighted and fused to obtain the final dynamic ERD threshold, which is used to determine the motor intention of the EEG signal and thus obtain control commands. This method achieves optimized configuration of decoding algorithm parameters without increasing the user's training burden.

[0048] Example 6: Exoskeleton Control and Closed-Loop Feedback Based on the motion intent recognition results, exoskeleton control commands are sent to drive the exoskeleton assistive motion module to lead the user's limbs to perform continuous, smooth and physiologically consistent movements. The exoskeleton's motion status and sensory feedback information are transmitted to the user's brain in real time, forming a closed-loop neural regulation and promoting neural function reconstruction.

[0049] While the exoskeleton performs actions, the feedback module is an audiovisual feedback module, which is used to transmit the motion assessment information generated by the exoskeleton movement to the user in real time in the form of audiovisual feedback.

[0050] Example 7: System Workflow This embodiment illustrates the complete workflow of the system, including the following steps: S1: Guide users to perform a compound motor imagery paradigm, that is, while imagining motor movement of one lower limb, introduce the opposite upper limb to perform the task.

[0051] S2: The EEG signal acquisition module acquires the user's EEG signal during motor imagery and performs filtering and noise reduction preprocessing through the EEG processing module.

[0052] S3: Determine the user's rehabilitation stage: If in the early or middle stage of rehabilitation, call the trained algorithm module, use the time-frequency cross-view cross-attention fusion algorithm to extract time-frequency features and perform dynamic reweighting and individual adaptation, and output the motion intention recognition classification result; if in the middle or late stage of rehabilitation, call the untrained algorithm module, use the dynamic ERD threshold adaptive optimization method based on dual time scales to recognize motion intention, and output the recognition result.

[0053] S4: Generate exoskeleton control commands based on the motion intention recognition results, and drive the exoskeleton assistive motion module to enable the user's limbs to perform continuous, smooth and physiologically consistent movements.

[0054] S5: Transmits the exoskeleton's motion status and sensory feedback information to the user's brain in real time, forming a closed-loop neural regulation and promoting neural function reconstruction.

[0055] In summary, the scope of protection of this invention is not limited to the specific embodiments described above. Equivalent substitutions or modifications made by those skilled in the art based on the concept of this invention are all within the scope of protection of this invention.

Claims

1. A non-invasive closed-loop brain-computer interface system for accurate recognition of left and right lower limb movement intentions from the contralateral upper limb, characterized in that, This includes both hardware devices and signal processing. The hardware device includes an electroencephalogram (EEG) signal acquisition module and an exoskeleton-assisted movement module. The signal processing terminal includes an EEG processing module, a module with a trained algorithm, and a module without a trained algorithm. The EEG signal acquisition module is used to acquire EEG signals generated by the user based on a compound motor imagery paradigm. The EEG processing module is used to preprocess the acquired EEG signals; The trained algorithm module is used to identify motor intentions from EEG signals in the early and middle stages of rehabilitation. The training-free algorithm module is used to identify motor intentions from EEG signals during the mid- and late-stages of rehabilitation. The exoskeleton-assisted motion module is used to receive the identified motion intention results and execute the corresponding motion actions.

2. The non-invasive closed-loop brain-computer interface system according to claim 1, characterized in that, The composite motion imagination paradigm is as follows: While performing unilateral lower limb motor imagery, the contralateral upper limb is introduced to perform tasks in coordination, thereby activating multiple functional areas in the brain's motor cortex and enhancing the functional connections between different limb motor representations. When the system instructs the user to step with their left lower limb, the user's brain visualizes the movement of "stepping with the left leg," while the right upper limb performs the actual movement of "swinging the arm forward" or "clenching the fist." When the system instructs the user to step with their right lower limb, the user's brain visualizes the movement of "stepping with the right leg," while the left upper limb performs the actual movement of "swinging the arm forward" or "clenching the fist." Through the coordination of the actual physical swing of the arms and the mental stepping of the opposite leg, a cross-gait rhythm of "right arm and left leg" and "left arm and right leg" is formed.

3. The non-invasive closed-loop brain-computer interface system according to claim 1, characterized in that, The preprocessing method of the EEG processing module is as follows: Adaptive notch filtering and independent component analysis are used together to remove power frequency interference and electromyography artifacts. The noise is then processed by co-space mode and continuous wavelet transform respectively before being input into the trained algorithm module and the untrained algorithm module.

4. The non-invasive closed-loop brain-computer interface system according to claim 1, characterized in that, The trained algorithm module employs a trained recognition algorithm based on time-frequency cross-view cross-attention fusion, specifically including: A dual-branch network is constructed for feature extraction. The temporal branch uses multi-scale convolution to extract local temporal waveform features and captures the global long-distance dependencies of EEG sequences through a multi-head self-attention mechanism to obtain a temporal feature view. Frequency domain branching extracts the degree of decrease or increase of key frequency band power relative to the baseline state and the energy ratio of key frequency bands. Based on this, multi-scale convolution is used to further extract local spectral patterns at different frequency scales to form a frequency domain feature view. Based on the attention mechanism, the temporal feature view output by the temporal branch and the frequency feature view output by the frequency branch are bidirectionally complementary. In this process, a cross-view consistency loss function is introduced to force the alignment of the feature distribution of the two views in the high-level semantic space. A dynamic reweighted fusion module is introduced to calculate weight values ​​in real time based on the classification contribution of the two views and to dynamically reweight the features after interaction. The fused features are fed into a classifier with a subject-specific adapter to complete motion intent recognition.

5. The non-invasive closed-loop brain-computer interface system according to claim 4, characterized in that, The classifier trains the main network using training data from multiple subjects during the pre-training phase, freezes the main parameters of the pre-trained main network during the new user phase, fine-tunes the subject-specific adapter and classification layer parameters using only a small amount of individual calibration data, and continuously iterates and optimizes the model parameters using an adaptive update strategy in the subsequent rehabilitation process, outputting gait intention recognition results.

6. The non-invasive closed-loop brain-computer interface system according to claim 1, characterized in that, The training-free algorithm module employs a dynamic event correlation desynchronization threshold adaptive optimization method based on dual time scales, specifically including: Short-term optimization: After each trial, the ERD threshold is fine-tuned based on task success or failure, recent success rate trends, and historical adjustment directions to achieve a response to the user's real-time status. Long-term scale optimization: After each set number of valid trials, a periodic cosine annealing process is triggered, forcing the ERD threshold to be perturbed by a cosine curve within a preset reasonable range to actively explore potential better solutions, and then weighted and fused with the adaptive threshold.

7. The non-invasive closed-loop brain-computer interface system according to claim 6, characterized in that, The training-free algorithm module employs a dynamic event correlation desynchronization threshold adaptive optimization method based on dual time scales, specifically executed according to the following steps: First, short-term optimization is triggered immediately after each single trial. The system comprehensively evaluates the success or failure of the current trial, the accuracy trend of recent consecutive trials, and the adjustment direction of historical thresholds. It then makes rapid fine-tuning of the ERD threshold in small steps to track and respond to the user's instantaneous EEG state fluctuations in real time. in, The current ERD threshold, The ERD threshold is the one updated shortly. This indicates that the ERD determination result of the current trial is less than the ERD threshold. A value of 1; greater than the ERD threshold. The value is 0; This indicates the accuracy rate for consecutive recent trials that fall below the ERD threshold. This indicates the preset target accuracy rate. Adjust the step size based on the base. This is an adaptive adjustment coefficient; through the above method, the threshold can be dynamically adjusted trial by trial based on the user's recent EEG response status. Secondly, after each set number of valid experiments is completed, long-term scale optimization is triggered. The system starts a periodic cosine annealing process, which forcibly generates a perturbation threshold that fluctuates smoothly according to the cosine curve within a preset reasonable range. In this way, it actively jumps out of the local optimum solution trapped due to short-term fine-tuning and explores potential better threshold ranges. in, The value is the perturbation amplitude, and k is a periodic variable for long-term optimization, which periodically explores other threshold regions near the current threshold to reduce the possibility of the threshold falling into a local stable state for a long time. Finally, the short-term adaptive threshold and the long-term cosine annealing threshold are weighted and fused to obtain the final dynamic ERD threshold, which is used to determine the motor intention of the EEG signal.

8. The non-invasive closed-loop brain-computer interface system according to claim 1, characterized in that, The hardware device also includes an audiovisual feedback module, which is used to transmit motion assessment information generated by the exoskeleton movement to the user in real time in the form of audiovisual feedback, forming a closed-loop neural control.

9. A method for accurately recognizing and controlling the movement intentions of the left and right lower limbs based on the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: S1: Guide users to perform a compound motor imagery paradigm, that is, while imagining motor movements of one lower limb, introduce the coordinated movement of the contralateral upper limb to perform the task. S2: The EEG signal acquisition module acquires the user's EEG signal during motor imagery, and the EEG processing module performs filtering and noise reduction preprocessing. S3: Determine the user's rehabilitation stage: If the user is in the early or middle stage of rehabilitation, the trained algorithm module is called, and the algorithm based on time-frequency cross-view cross-attention fusion is used to extract time-frequency features and perform dynamic reweighting and individual adaptation, and output the classification result. If the patient is in the middle or late stage of rehabilitation, the training-free algorithm module is invoked, and a dynamic event correlation desynchronization threshold adaptive optimization method based on dual time scales is used to identify the motion intention and output the classification result. S4: Generate exoskeleton control commands based on the motion intention recognition results, and drive the exoskeleton assistive motion module to lead the user's limbs to perform continuous, smooth and physiologically consistent movements; S5: Transmits the exoskeleton's motion status and sensory feedback information to the user's brain in real time, forming a closed-loop neural regulation and promoting neural function reconstruction.