Limb adaptive control system

Abstract

The invention relates to the technical field of intelligent control, in particular to a limb adaptive control system, which comprises a first-level control module, a second-level control module, a first-level feedback control module and a second-level feedback control module, the second-level control module is used for adjusting a single stimulation in a complete planned action based on an electric stimulation sequence; according to a comparison result of the actual muscle recruitment degree and a preset muscle recruitment degree base line in the current time window, dynamically adjusting the stimulation current of single stimulation in the next time window; and the third-level control module is used for adjusting a complete plan action based on the electrical stimulation sequence, updating the current strategy network model according to the effect of the current complete plan action based on the electrical stimulation sequence, and guiding the next complete plan action by the updated strategy network model. Passive stimulation is converted into active auxiliary collaboration, so that self-learning optimization of the system is realized.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to a limb adaptive control system. Background Technology

[0002] Limb motor dysfunction is a problem faced by many users. Traditional solutions, such as manual assisted stimulation, have some effect, but they suffer from problems such as scarcity of human resources, difficulty in ensuring the intensity and standardization of interactive stimulation, and low user participation, resulting in limited remodeling efficiency.

[0003] Despite advancements in individual technologies, existing technologies have consistently failed to provide a solution that can maintain high reliability in dynamic environments and achieve personalized, precise, and adaptive control due to inherent limitations in their system architecture. Summary of the Invention

[0004] The purpose of this application is to provide a limb adaptive control system, which innovatively constructs a multi-level closed-loop architecture that tightly couples the user's movement intention, neuromuscular response and actual action effect, realizing the transition from passive stimulation to active assistance and coordination, thereby achieving system self-learning optimization.

[0005] In some embodiments, this application provides a limb adaptive control system, the system comprising: a first-level control module, configured to initiate electrical stimulation via first-level feedforward control after detecting a user's valid movement intention; a second-level control module, configured to adjust individual stimuli in a complete planned action based on an electrical stimulation sequence after initiating electrical stimulation; during the interval between different individual stimuli, acquiring electromyographic signals within the current time window and evaluating the actual muscle recruitment degree, and dynamically adjusting the stimulation current of individual stimuli in the next time window based on a comparison between the actual muscle recruitment degree and a pre-set baseline of muscle recruitment degree within the current time window, wherein the magnitude of the stimulation current is related to the magnitude of muscle recruitment degree; a third-level control module, configured to adjust a complete planned action based on an electrical stimulation sequence; acquiring the user's current state parameters, making decisions based on the current policy network model to obtain the adjustment action and stimulation parameters corresponding to the maximum probability, performing adaptive stimulation based on the stimulation parameters, and updating the current policy network model based on the effect of the current complete planned action based on the electrical stimulation sequence, the updated policy network model guiding the next complete planned action.

[0006] In some embodiments, the first-level control module further includes a motion intention monitoring unit, a stimulation parameter initialization unit, and an execution unit: the motion intention monitoring unit is used to activate an electrical stimulation sequence based on the user's valid motion intention; the stimulation parameter initialization unit is used to obtain an initial electrical stimulation parameter sequence corresponding to the user from a pre-set mapping table; and the execution unit is used to perform electrical stimulation using the initial electrical stimulation parameter sequence, wherein the initial electrical stimulation parameter sequence includes at least one of personalized initial stimulation intensity, muscle recruitment baseline for the corresponding time window, and safety boundary information.

[0007] In some embodiments, the second-level control module further includes an evaluation unit and an adjustment unit: the evaluation unit is used to evaluate the actual muscle recruitment degree based on the electromyographic signals collected within the current time window during the stimulation interval; the adjustment unit is used to compare the actual muscle recruitment degree with a preset baseline for muscle recruitment degree within the current time window; if the actual muscle recruitment degree is less than the lower limit of the preset baseline, the stimulation current is increased according to a preset threshold to improve the actual muscle recruitment degree of a single stimulus in the next time window, and the adjusted stimulation current is constrained within a preset safety constraint range; if the actual muscle recruitment degree is greater than the upper limit of the preset baseline, the stimulation current is decreased according to a preset threshold to reduce the actual muscle recruitment degree of a single action in the next time window; if the actual muscle recruitment degree is within the range of the upper and lower limits of the preset baseline, the electrical stimulation of a single stimulus in the next time window continues according to the current stimulation current; wherein, the baseline for muscle recruitment degree and the preset threshold corresponding to a single stimulus in different time windows are different.

[0008] In some embodiments, the second-level control module further includes a muscle recruitment baseline setting unit, which is used to: acquire electromyographic signals of a user performing a standard movement within a complete planned movement cycle based on an electrical stimulation sequence before electrical stimulation; divide a complete planned movement cycle based on an electrical stimulation sequence into multiple time windows according to the physiological phase of the user's movement, and calculate the muscle recruitment baseline corresponding to the electromyographic signal of a single movement within each time window.

[0009] In some embodiments, the third-level control module further includes an update unit, which is configured to: obtain the concordance error and reward between a current complete planned action based on an electrical stimulation sequence and a standard stimulation action; update the policy network model and the value network model according to the concordance error and the reward; and use the updated policy network model and the value network model to guide the next complete planned action.

[0010] In some embodiments, the updating unit further includes a reward calculation subunit, which includes at least one of performance enhancement, safety margin, and stability penalty; performance enhancement is used to represent the degree of improvement in the consistency of the current action relative to the average of the previous few times, safety margin is used to represent operation within the safety boundary parameter space, and stability penalty is used to represent the penalty for large changes in the stimulation current.

[0011] In some embodiments, the system further includes a motion state determination module, which is used to: collect the user's electroencephalogram (EEG) signals and electromyogram (EMG) signals; determine the motion state of the EEG signals based on the EMG activity markers of the EMG signals; and when a valid motion intention is determined, adaptively control the limbs through a multi-level closed-loop adaptive control strategy, wherein the multi-level closed-loop adaptive control strategy includes at least one of a first-level feedforward control, a second-level compensation control, and a third-level iterative learning control.

[0012] In some embodiments, the motion state determination module further includes a joint determination unit, which is used to: calculate the phase lock value between whole-brain electrodes based on the multi-band data of the EEG signal, construct a functional connectivity matrix and extract network topology features to obtain an EEG feature vector; extract the root mean square value from the electromyographic envelope in the electromyographic signal as a muscle activation intensity feature, and determine the electromyographic activity marker of the EEG signal based on the muscle activation intensity feature; perform intention recognition on the EEG feature vector through a preset classifier to obtain the validity of the motion state; when the validity of the motion state corresponds to a valid motion intention and the electromyographic activity marker corresponds to significant activity, it is determined to be a valid motion intention; when the validity of the motion state corresponds to a valid motion intention and the electromyographic activity marker corresponds to no significant activity, it is determined to be an invalid motion intention.

[0013] In some embodiments, the system further includes a pre-configuration module, which is used to: determine the type of stimulation action; obtain the electrode application position corresponding to the type of stimulation action and a general initial stimulation current intensity reference range from a pre-set stimulation parameter template library; wherein the electrode application position is used to instruct the user to apply the electrode at the corresponding position for motion control, and the general initial stimulation current intensity reference range is used to instruct the user to dynamically adjust to obtain a personalized initial stimulation intensity corresponding to the user.

[0014] In some embodiments, the third-level control module further includes a state parameter setting unit and an action parameter setting unit: the state parameter setting unit is used to set the state parameters to include at least one of the following: consistency of historical motion actions, current intensity, state change, fatigue, exercise load, muscle excitability, and safety boundary compliance; the action parameter setting unit is used to set the action parameters to include at least one of the following: control direction and control intensity.

[0015] In the above embodiments, an innovative multi-level closed-loop architecture tightly couples the patient's motor intention, neuromuscular response, and actual movement effect, achieving a transition from passive stimulation to active assisted coordination, thereby enabling system self-learning optimization. It comprises three progressively layered control loops with different time scales. The first level is a feedforward control loop, initiating electrical stimulation from intention. When the decoder confirms a valid motor state, the system immediately retrieves the corresponding electrical stimulation parameter sequence from the personalized parameter mapping table, driving the electrical stimulation module to output, achieving a rapid response to the motor intention. The second level is a rapid feedback loop, a real-time electromyographic compensation. During the electrical stimulation interval, the system utilizes the electromyographic acquisition function of the electrical stimulation module to monitor the recruitment status of the target muscle in real time, such as the amplitude of the electromyographic signal. If the recruitment level is lower than the expected threshold, parameters such as current intensity are fine-tuned within the current stimulation cycle. This achieves micro-control over the execution of a single movement, ensuring that each neural command is effectively converted into muscle contraction, solving the problem of whether stimulation is effective. The third level is a lag optimization closed loop, based on visually guided iterative learning. After a complete action is performed, the system generates a skeletal animation of the user's execution in real time via video stream and performs quantitative matching with a standard model to calculate the action conformity. This metric is not used as the basis for adjusting the action but is instead input as a performance indicator into the iterative learning controller. The controller updates the electrical stimulation parameter sequence corresponding to the action in the personalized parameter mapping table based on the learning algorithm. By adjusting the initial stimulation current at different electrode positions, it optimizes the user's execution of the standard movement. This enables the system to have memory and learning capabilities, allowing it to optimize future strategies based on historical performance, driving the user's action quality to continuously approach the standard model, thus addressing the issue of whether the action is standard from a macroscopic perspective and achieving adaptive personalized motion control. Attached Figure Description

[0016] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein:

[0017] Figure 1 This is a block diagram of a limb adaptive control system provided in one embodiment of this application;

[0018] Figure 2This is a schematic diagram of a closed-loop real-time control system based on a first level and a second level, provided in one embodiment of this application;

[0019] Figure 3 This is a schematic diagram of the control system flow of the long-term optimization closed loop, i.e., the third-level closed loop, provided in one embodiment of this application;

[0020] Figure 4 This is a schematic diagram of the initialization and personalized modeling process provided in one embodiment of this application;

[0021] Figure 5 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] The technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0024] This solution does not aim to obtain disease diagnosis results or health status. It is a system that processes the user's electroencephalogram (EEG) and electromyogram (EMG) data to achieve adaptive limb control. All steps are performed by information processing methods implemented by computers and other devices.

[0025] It should be fully understood that the user information involved in this application (including but not limited to EEG signals and EMG signals) is information and data authorized by the user or fully authorized by all parties. The use of user information shall comply with the privacy policies and practices of the industry that are generally considered to meet or exceed the requirements for maintaining user privacy. The collection, use and processing of related data shall comply with relevant laws, regulations and standards, and provide corresponding operation access points for users to choose to authorize or refuse.

[0026] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0027] In some embodiments, see Figure 1 This application provides a limb adaptive control system 100, the system comprising:

[0028] The first-level control module 101 is used to initiate electrical stimulation through first-level feedforward control after detecting the user's valid movement intention.

[0029] The second-level control module 102 is used to adjust individual stimuli in a complete planned action based on an electrical stimulation sequence after the electrical stimulation is initiated; during the interval between different individual stimuli, it collects electromyographic signals within the current time window and evaluates the actual muscle recruitment degree; and dynamically adjusts the stimulation current of individual stimuli in the next time window based on the comparison result between the actual muscle recruitment degree and the preset baseline of muscle recruitment degree within the current time window, wherein the magnitude of the stimulation current is related to the magnitude of muscle recruitment degree.

[0030] The third-level control module 103 is used to adjust a complete planned action based on an electrical stimulation sequence; it obtains the user's current state parameters, makes a decision on the current state parameters according to the current policy network model to obtain the adjustment action and stimulation parameters corresponding to the maximum probability, performs adaptive stimulation according to the stimulation parameters, and updates the current policy network model according to the effect of the current complete planned action based on the electrical stimulation sequence. The updated policy network model guides the next complete planned action.

[0031] Specifically, this application provides a limb adaptive control system, which is a multi-level closed-loop adaptive control system. In some embodiments, it may include three progressively advancing control loops with different time scales. Details are as follows:

[0032] The first-level control module acts as a feedforward control closed loop, initiating electrical stimulation upon detecting the user's movement intention. Initiation includes turning on and executing electrical stimulation. Specifically, when the decoder determines that the user's current movement is a valid movement intention, it immediately sends a command to the intelligent control module. The control module, based on the current movement task, queries the corresponding initial electrical stimulation parameter sequence from the personalized parameter mapping table associated with the user and drives the electrical stimulator to output according to this initial electrical stimulation parameter sequence, achieving a rapid response to the movement intention. The electrical stimulation module provides an electrical stimulation sequence with a certain current intensity, with each stimulation interval being ΔT_stim.

[0033] The second-level control module acts as a rapid feedback loop, providing real-time compensation via electromyography (EMG). Specifically, after initiating electrical stimulation, this module adjusts individual stimuli within a complete planned action based on an electrical stimulation sequence. During the intervals between different individual stimuli, i.e., the stimulation interval ΔT_stim, the system utilizes the EMG acquisition function of the electrical stimulation module to monitor the muscle recruitment level of the target muscle in real time and assess the muscle recruitment status, such as the amplitude of the EMG signal. It then compares the currently acquired actual muscle recruitment level with a pre-set baseline for muscle recruitment within the current time window. If the current recruitment level is lower than the expected threshold, parameters such as the current intensity are fine-tuned within the stimulation cycle of the current time window. This allows for dynamic adjustment of the stimulation current for individual stimuli in the next time window, achieving micro-control of the execution of a single action and ensuring that each neural command is effectively converted into muscle contraction, thus solving the problem of whether the stimulation is effective.

[0034] The third-level control module acts as a lagging optimization loop, implemented through visually guided iterative learning. Specifically, this module adjusts a complete planned action based on an electrical stimulation sequence. After a complete planned action is completed, the system generates a skeletal animation of the user performing the entire planned action in real-time via video stream and performs quantitative matching with a standard model to calculate the degree of conformity between the current actual action and the standard action. However, this indicator is not used as the basis for adjusting the current complete planned action; instead, it is input as a performance indicator to the iterative learning controller. The controller updates the initial electrical stimulation parameter sequence corresponding to the action in the personalized parameter mapping table according to the learning algorithm. By adjusting the initial stimulation current at different electrode positions, it optimizes the user's completion rate of the standard stimulation action. This method enables the system to possess memory and learning capabilities, allowing it to optimize future strategies based on historical performance, driving the user's action quality to continuously approach the standard model, thus macroscopically addressing the issue of whether the action is standard and achieving adaptive personalized motion control.

[0035] See details Figure 2 The diagram below illustrates a closed-loop real-time control system based on a first-level and a second-level system, representing one embodiment. During movement, the system simultaneously acquires the user's electroencephalogram (EEG) and electromyographic (EMG) signals. The system decodes the movement intention based on the acquired multimodal sensor signals. If the movement intention is deemed valid, the first-level closed-loop control is triggered and initiated. After electrical stimulation is triggered, the system continuously monitors the user's EMG signals within a single stimulation cycle (i.e., a single movement cycle) within a complete planned movement based on an electrical stimulation sequence. The system calculates muscle recruitment intensity based on the detected signals to determine the recruitment status. Stimulation parameters are then fine-tuned in real-time according to the recruitment status, ensuring that the adjusted parameters remain within safe boundaries, thus guaranteeing that each stimulation is effective and safe.

[0036] See Figure 3 This is a flowchart of a long-term optimization closed-loop, or third-level closed-loop, control system provided in one embodiment of this application. The diagram illustrates the long-term learning capability of the control system. After a complete planned action is completed, i.e., after each electrical stimulation, the system calculates a reward based on the user's performance quality, drives the reinforcement learning algorithm to update the strategy, and optimizes the personalized parameter mapping table for control of the subsequent complete planned action, achieving continuous improvement of the control strategy. Specifically, after a complete planned action is completed, video data and performance indicators of the complete action are collected, and real-time skeletal animation is generated. The user's skeletal animation is compared with a model of a standard action to calculate the degree of conformity between the complete action and the standard action. This is used to generate a reward signal, update the reinforcement learning agent accordingly, and update the personalized electrical stimulation parameter mapping table. Thus, in a complete planned action, the updated parameters can be used for motion control.

[0037] In the above embodiments, a dynamic limb control system based on electromyographic signals is constructed. This system tightly couples the user's movement intention, neuromuscular response and actual movement effect through an innovative multi-level closed-loop architecture, realizing the transition from passive electrical stimulation to active assisted coordination, thereby achieving system self-learning optimization.

[0038] In some embodiments, the system further includes a motion state determination module, which is used to collect the user's electroencephalogram (EEG) signals and electromyogram (EMG) signals; determine the motion state of the EEG signals based on the EMG activity markers of the EMG signals; and when a valid motion intention is determined, adaptive control of the limbs is performed through a multi-level closed-loop adaptive control strategy, wherein the multi-level closed-loop adaptive control strategy includes at least one of a first-level feedforward control, a second-level compensation control, and a third-level iterative learning control.

[0039] During adaptive motion control, the user watches a guided animation matched to the task and imagines movement. The system acquires the user's EEG signals during the intervals between electrical stimulation pulses, and simultaneously acquires surface electromyography (EMG) signals. Subsequently, a decoding strategy fusing feature and decision layers is employed. First, EEG functional network features and EMG features are fused for initial recognition. Then, EMG activity markers are introduced for secondary verification of the recognition results to determine whether the user has a genuine motor intention. Only when the user is determined to have a valid, genuine motor intention will the first level of feedforward control be triggered. This approach significantly improves the robustness and security of motion intention decoding in dynamic environments, effectively preventing false triggering caused by EEG artifacts.

[0040] In some embodiments, the motion state determination module further includes a joint determination unit, which is used to: calculate the phase lock value between whole-brain electrodes based on the multi-band data of the EEG signal, construct a functional connectivity matrix and extract network topology features to obtain an EEG feature vector; extract the root mean square value from the electromyographic envelope in the electromyographic signal as a muscle activation intensity feature, and determine the electromyographic activity marker of the EEG signal based on the muscle activation intensity feature; perform intention recognition on the EEG feature vector through a preset classifier to obtain the validity of the motion state; when the validity of the motion state corresponds to a valid motion intention and the electromyographic activity marker corresponds to significant activity, it is determined to be a valid motion intention; when the validity of the motion state corresponds to a valid motion intention and the electromyographic activity marker corresponds to no significant activity, it is determined to be an invalid motion intention.

[0041] The following embodiments form the basis for ensuring the reliability of the limb adaptive control system input. The specific implementation includes the following content. It should be noted that the parameters involved in the following content are only illustrative examples and do not constitute a limitation on this application. Different parameter values ​​can be set as needed in different scenarios.

[0042] In some specific embodiments, the limb adaptive control system includes a time-division multiplexing acquisition module, specifically using hardware synchronization signals to strictly control the data acquisition timing. The electrical stimulator operates in pulse mode, for example, setting each pulse cycle to last 20ms with a 30ms interval, and acquiring EEG signals during the 25ms interval after the pulse, thereby effectively avoiding stimulation artifacts. Simultaneously with EEG acquisition, the electrical stimulation electrodes switch to electromyography (EMG) acquisition mode to acquire surface EMG signals of the target muscle.

[0043] Signal preprocessing includes preprocessing of EEG and EMG signals. Specifically, the raw EEG signals are subjected to bandpass filtering (0.5-45Hz), notch filtering (50Hz), and independent component analysis (ICA) to remove physiological artifacts such as those from electrooculography (EOG) and electrocardiography (ECG). The acquired EMG signals are subjected to bandpass filtering (10-500Hz), full-wave rectification, and their linear envelope is calculated.

[0044] Then, feature extraction and fusion decoding are performed. For EEG feature extraction, the preprocessed multi-channel EEG signal is decomposed into θ (4-8Hz), α (8-13Hz), and β (13-30Hz) frequency bands. For each frequency band, the phase-locked value (PLV) between whole-brain electrodes is calculated, a functional connectivity matrix is ​​constructed, and network topology features (such as node degree centrality) are extracted to form the EEG feature vector F_EEG. For EMG feature extraction, the root mean square (RMS) value is extracted from the EMG envelope as a feature of muscle activation intensity. Simultaneously, an adaptive threshold, such as twice the baseline mean standard deviation, is set to binarize the EMG signal into an EMG activity flag Flag_EMG, where 1 indicates significant activity and 0 indicates no activity.

[0045] Next comes the decision-level fusion. F_EEG is input into a pre-trained Support Vector Machine (SVM) classifier to obtain a preliminary intent recognition result, such as whether the user is in a moving or resting state. This result is then logically compared with Flag_EMG. Only when the SVM output indicates a moving state and Flag_EMG is 1 is the intent deemed valid, triggering subsequent control, such as activating the first level of motion control. If the SVM output indicates a moving state but Flag_EMG is 0, the intent is deemed suspicious, and the system either does not trigger a stimulus or only provides a very low-intensity stimulus, logging the results for algorithm optimization.

[0046] In the above embodiments, combining both electroencephalogram (EEG) and electromyogram (EMG) signals to comprehensively assess whether a valid motor intention state is present improves the accuracy of intention recognition.

[0047] In some embodiments, the first-level control module further includes a motion intention monitoring unit, a stimulation parameter initialization unit, and an execution unit. The motion intention monitoring unit is used to initiate an electrical stimulation sequence based on the user's valid motion intention; the stimulation parameter initialization unit is used to obtain an initial electrical stimulation parameter sequence corresponding to the user from a pre-set mapping table; the execution unit is used to perform electrical stimulation using the initial electrical stimulation parameter sequence, wherein the initial electrical stimulation parameter sequence includes at least one of personalized initial stimulation intensity, a muscle recruitment baseline for a corresponding time window, and safety boundary information.

[0048] Specifically, this application also includes constructing an initial personalized parameter mapping table. Specifically, the system creates a personalized electrical stimulation parameter mapping table for the user. The data structure of this mapping table uses actions as keys, and the corresponding values ​​include at least one of the following: the personalized initial stimulation intensity obtained after calibration, the muscle recruitment baseline for the corresponding time window, and safety boundary information.

[0049] Each electrode channel corresponds to individualized safety boundary information, such as sensing threshold, tolerance threshold, and safe operating limit. This safety information will serve as a hard constraint that must not be violated in all subsequent adaptive adjustment processes (including the second and third level closed loops), thus embedding safety at the system level.

[0050] The multi-level closed-loop adaptive control system provided in this application is a dynamic limb control system based on electromyography (EMG) and electroencephalography (EEG) signals. Its core lies in a multi-level closed-loop adaptive control process, including the following steps: First, the system performs initialization and personalized modeling. Initially, based on the user's movement stage (e.g., bedside, sitting, walking), the system guides the placement of electrical stimulation electrodes on the target muscle groups. Crucially, a video acquisition module records standard movements under human guidance and assistance, generating a personalized standard skeletal animation model for the user. Simultaneously, the system initializes a personalized electrical stimulation parameter mapping table, which presets initial electrical stimulation parameter sequences for different movement actions, such as current intensity and frequency curves over time. Then, when the user is determined to have a valid movement intention, the system can directly initiate and execute the first level of electrical stimulation control based on the pre-built initialized personalized parameter mapping table.

[0051] In some embodiments, the system further includes a pre-configuration module, which is used to determine the type of stimulation action; obtain the electrode application position corresponding to the type of stimulation action and a general initial stimulation current intensity reference range from a pre-set stimulation parameter template library; wherein, the electrode application position is used to instruct the user to apply the electrode at the corresponding position for motion control, and the general initial stimulation current intensity reference range is used to instruct the user to dynamically adjust to obtain a personalized initial stimulation intensity corresponding to the user.

[0052] See Figure 4 This diagram illustrates the initialization and personalized modeling process provided in one embodiment of this application, showcasing the key steps of system initialization. First, electrodes are configured and safety boundaries are determined according to the stage of movement to ensure safe motion control. Then, a standard motion model is recorded under calibrated stimulation parameters. Finally, a personalized parameter mapping table is generated, laying the foundation for adaptive motion. Specifically, this includes: first, system initialization is performed, and the stage of movement is selected. Based on the selected stage, instructions are given to guide the electrodes to be applied to the target muscle group, and the electrodes are applied according to the instructions. Next, the user's personalized safety boundaries are determined, including perception and tolerance thresholds, and the user's personalized stimulation parameters are calibrated within these boundaries. A standard motion corresponding to this stimulation is recorded, and a standard skeletal animation model is generated, along with the initialization of the personalized electrical stimulation parameter mapping table.

[0053] Specifically, the solution provided in this application also includes initial and personalized modeling. Before users begin using the system, personalized configuration is required, including the selection of the exercise stage. Professionals select the current exercise stage through the system's human-computer interaction interface, such as a bedside ankle pump exercise control, seated leg raise exercise control, or simulated walking exercise control. Then, electrode application guidance is performed. The system has a pre-built stimulus parameter template library based on historical information. This template library defines the target muscle group, suggested electrode application locations, and the corresponding initial stimulation current intensity reference range for each movement type. Specific embodiments are as follows. It should be noted that the parameter ranges provided in the following embodiments are only illustrative examples and can be flexibly set according to needs in different scenarios.

[0054] For ankle pump motor control, such as ankle flexion and extension, the system guides the application of electrode patches to the belly of the tibialis anterior muscle (dominant in ankle dorsiflexion) and the gastrocnemius muscle (dominant in ankle plantarflexion). The initial stimulation current intensity reference range given by the template is 15-25 mA for the tibialis anterior muscle and 18-30 mA for the gastrocnemius muscle, with the frequency typically set at 20-40 Hz and the pulse width at 200-400 μs.

[0055] For seated leg raise control, such as knee extension: the system guides the electrodes to be applied to the main muscle bellies of the quadriceps, such as the rectus femoris and vastus lateralis. The initial stimulation current intensity is typically within the range of 20-35 mA.

[0056] For simulated walking motion control, the system guides the placement of electrodes on multiple muscle groups in both lower limbs, such as the gluteus medius, quadriceps femoris, hamstrings, and tibialis anterior, and presets an alternating stimulation sequence to simulate stepping movements. The initial current intensity is set in stages within the range of 15-35 mA according to muscle function and the magnitude of the effect.

[0057] Furthermore, this includes establishing safety boundaries. Before beginning individualized parameter calibration, the system must set safe and comfortable current stimulation boundaries for each user. This process is automatically guided by the system, as shown below:

[0058] Global safety limit setting. The system first sets an absolute safety limit based on international standards and clinical consensus, such as 40mA, which must not be exceeded under any circumstances.

[0059] Individualized tolerance threshold determination involves a progressive sensory test initiated for each applied electrode patch. Starting with an extremely low current intensity, such as 1 mA, the stimulation intensity is gradually increased in small increments, such as 0.5 mA. Users provide feedback via handheld button or voice command when they first experience a noticeable tingling sensation (sensory threshold) and when they feel discomfort but can tolerate it (tolerance threshold). The system records the sensory threshold and tolerance threshold for each electrode location.

[0060] The system defines safe operating boundaries by setting the initial safe operating limit for each electrode plate to a value slightly lower than its tolerance threshold, for example, 90% of the tolerance threshold. Simultaneously, the system sets a dynamic safe upper limit, which can be cautiously and gradually increased by the system under the authorization of professionals, based on the user's adaptation during subsequent motion control.

[0061] Individualized parameters are determined. Based on the selected movement, electrodes are applied according to the template guidance. The system then initiates a parameter calibration process, which is strictly limited to the defined individualized safety boundaries. Starting from the lower limit suggested by the template, the stimulation current is gradually increased within the safety boundaries until a clearly visible contraction of the target muscle or a standard range of motion in the joint is observed, which can be determined with the assistance of a video module. At this point, the current intensity is determined as the user's personalized initial value for that movement and recorded in the personalized electrical stimulation parameter mapping table. The system ensures that this value is far from the safety upper limit, leaving sufficient buffer space.

[0062] The multi-level closed-loop adaptive control process provided in the above embodiments is illustrated in the following schematic diagram: Figure 2 and Figure 4 As shown, where Figure 4 This is a flowchart of system initialization and personalized modeling. Figure 2 This is a flowchart of the system's real-time control closed loop (first-level and second-level closed loops). Figure 3 This is a flowchart of the system's long-term optimization closed loop (third-level closed loop). These three diagrams together illustrate the complete workflow and control logic of the system described in this application. Figure 4 This laid the foundation for the system's operation. Figure 2 This demonstrates the system's real-time, precise control capabilities on a second or millisecond scale. Figure 3 This reveals the system's long-term self-learning and optimization capabilities at daily or weekly scales. The combination of these three aspects clearly demonstrates the system-level innovation of this application's multi-timescale collaborative optimization.

[0063] In some embodiments, the second-level control module further includes an evaluation unit and an adjustment unit: the evaluation unit is used to evaluate the actual muscle recruitment degree based on the electromyographic signals collected within the current time window during the stimulation interval; the adjustment unit is used to compare the actual muscle recruitment degree with a preset baseline for muscle recruitment degree within the current time window; if the actual muscle recruitment degree is less than the lower limit of the preset baseline, the stimulation current is increased according to a preset threshold to improve the actual muscle recruitment degree of a single stimulus in the next time window, and the adjusted stimulation current is constrained within a preset safety constraint range; if the actual muscle recruitment degree is greater than the upper limit of the preset baseline, the stimulation current is decreased according to a preset threshold to reduce the actual muscle recruitment degree of a single action in the next time window; if the actual muscle recruitment degree is within the range of the upper and lower limits of the preset baseline, the electrical stimulation of a single stimulus in the next time window continues according to the current stimulation current; wherein, the baseline for muscle recruitment degree and the preset threshold corresponding to a single stimulus in different time windows are different.

[0064] Specifically, the second-level closed-loop control enables the adjustment of individual stimuli within a complete planned action based on an electrical stimulation sequence, achieving real-time compensation of electromyography (EMG). Specifically, in the second-level closed-loop control, the system acquires EMG signals during the intervals (ΔT_stim) between different individual stimuli within a single stimulus in a complete planned action based on an electrical stimulation sequence, and uses the acquired EMG signals to achieve real-time compensation and regulation of the current action execution process.

[0065] In some embodiments, the second-level control module further includes a muscle recruitment baseline setting unit, which is used to: acquire electromyographic signals of a user performing a standard movement within a complete planned movement cycle based on an electrical stimulation sequence before electrical stimulation; divide a complete planned movement cycle based on an electrical stimulation sequence into multiple time windows according to the physiological phase of the user's movement, and calculate the muscle recruitment baseline corresponding to the electromyographic signal of a single movement within each time window.

[0066] The core innovation of this method lies in establishing a segmented electromyographic prediction model based on action type, and adjusting dynamic parameters accordingly. The specific implementation steps are as follows:

[0067] A library of motion-related electromyography (EMG) prediction models was established. During the system initialization phase, not only were standard skeletal animation models generated, but a personalized EMG prediction model library was also constructed simultaneously. The model construction is as follows:

[0068] Specifically, when a user completes a standard stimulation movement with the assistance of a calibrated current, the system simultaneously records the surface electromyography (EMG) signals of each target muscle throughout the entire movement cycle. Then, segmentation processing is performed, including dividing the entire movement cycle (e.g., a leg raise, which includes raising, holding, and lowering) into several time windows, such as one window every 100ms, based on the physiological phase of the stimulation movement. Feature extraction and modeling are then performed, including calculating the root mean square (RMS) value of the EMG signal within each time window, which serves as an indicator of muscle activation intensity, i.e., muscle recruitment. The RMS values ​​of multiple standard movements are averaged to obtain the baseline muscle recruitment intensity (expected EMG intensity baseline EMG_expected(t)) and its acceptable fluctuation range [EMG_min(t), EMG_max(t)] for each time window. These baseline data, organized by movement, muscle, and time window, constitute the user's personalized EMG prediction model library.

[0069] Next, real-time electromyography (EMG) monitoring and compensation logic is implemented. During exercise, the system achieves real-time compensation under safety constraints according to the following process: Timing synchronization: The system's time control unit ensures strict alternation between EMG acquisition and stimulation delivery. EMG signals are rapidly acquired within the interval ΔT_stim (e.g., 30ms) after each stimulation pulse. Then, real-time feature calculation is performed, calculating the RMS value of the EMG signal within the current time window, denoted as EMG_actual(t). Next, difference assessment and decision-making under safety constraints are performed, comparing the actual muscle recruitment EMG_actual(t) with the expected baseline muscle recruitment EMG_expected(t) corresponding to the current movement, current muscle, and current time window. Before proposing any adjustments, a safety check is performed first, as shown below:

[0070] When the actual muscle recruitment is less than the lower limit of the pre-set muscle recruitment baseline, i.e., insufficient recruitment (Case A), this includes: if the actual muscle recruitment EMG_actual(t) < the minimum value of the muscle recruitment baseline EMG_min(t), it indicates insufficient muscle recruitment. The system first pre-calculates the adjusted current I_proposed = the current actual current I_current + ΔI. Then, a safety check is performed, i.e., the system immediately checks whether the adjusted current I_proposed is less than or equal to the individualized safe operating limit of the electrode channel. If the adjusted current I_proposed is within the safe range, the adjustment is adopted, and the latest current I_new = the adjusted current I_proposed. If the adjusted current I_proposed exceeds the safe limit, the system forcibly sets the current to the safe limit value, i.e., the latest current I_new = the safe limit current I_safe_max, and records a safety boundary violation event, indicating that the system may need to re-evaluate the user's muscle state or model parameters.

[0071] If the actual muscle recruitment degree is greater than the upper limit of the preset muscle recruitment degree baseline, i.e., case B, over-recruitment, this includes: if the actual muscle recruitment degree EMG_actual(t) > the maximum value of the muscle recruitment degree baseline EMG_max(t), it indicates over-recruitment. The system calculates the adjusted current I_proposed = current actual current I_current - ΔI. Since the adjustment is downward, a safety lower limit check is usually not required, as the lower limit is generally 0 or a sensing threshold. The system directly executes the latest current I_new = adjusted current I_proposed.

[0072] If the actual muscle recruitment degree is within the range of the upper and lower limits of the preset muscle recruitment degree baseline, i.e., in case C, recruitment is normal, specifically including: if the minimum value of the muscle recruitment degree baseline EMG_min(t) ≤ the actual muscle recruitment degree EMG_actual(t) ≤ the maximum value of the muscle recruitment degree baseline EMG_max(t), then keep the current stimulation parameters unchanged, and the latest current I_new = the current actual current I_current.

[0073] It should be noted that the parameters differ for different limbs. For example, when the limb corresponds to the lower limb, the adaptive control parameter adjustment strategy for the lower limb will have a different expected baseline setting. The shape and amplitude range of the expected muscle recruitment baseline value EMG_expected(t) curve are completely different for the tibialis anterior muscle in ankle pump movement and the quadriceps femoris muscle in leg raise movement. The system will call the corresponding model for comparison to ensure the accuracy of the evaluation.

[0074] Furthermore, the corresponding parameters differ at different stages of motor control. For example, during the initiation phase of a leg-raising movement, the expected electromyographic value is low, and compensation is sensitive; during the middle of the raising motion, the expected value reaches its peak, and the compensation step size ΔI can be appropriately increased. Throughout the entire process, the safety upper limit serves as an absolute boundary, ensuring that even in phases requiring rapid increases in stimulation, the modulation remains safe and controlled. This method of fine-tuning according to the movement phase while strictly adhering to safety boundaries provides professional assistance with varying levels of support while ensuring user safety.

[0075] In some embodiments, the third-level control module further includes an update unit, which is configured to: obtain the concordance error and reward between a current complete planned action based on an electrical stimulation sequence and a standard stimulation action; update the policy network model and the value network model according to the concordance error and the reward; and use the updated policy network model and the value network model to guide the next complete planned action.

[0076] In some embodiments, the updating unit further includes a reward calculation subunit, which includes at least one of performance enhancement, safety margin, and stability penalty; performance enhancement is used to represent the degree of improvement in the consistency of the current action relative to the average of the previous few times, safety margin is used to represent operation within the safety boundary parameter space, and stability penalty is used to represent the penalty for large changes in the stimulation current.

[0077] In some embodiments, the third-level control module further includes a state parameter setting unit and an action parameter setting unit: the state parameter setting unit is used to set the state parameters to include at least one of the following: consistency of historical motion actions, current intensity, state change, fatigue, exercise load, muscle excitability, and safety boundary compliance; the action parameter setting unit is used to set the action parameters to include at least one of the following: control direction and control intensity.

[0078] Specifically, this includes: recording standard movements. With the assistance of appropriate electrical stimulation, the user completes several repetitions of standard movements. The system records video using a high-definition camera (video capture module) and extracts the two-dimensional or three-dimensional coordinates of key joints, such as the hip, knee, and ankle, using a deep learning-based human posture estimation algorithm to generate a personalized standard skeletal animation model for the user. This model includes spatiotemporal information such as joint angles and movement trajectories. In the third-level closed-loop motion control process, this method treats the execution of each movement as a sequential decision-making process. The system, like an intelligent agent, explores different stimulation parameter adjustment strategies and learns the optimal strategy based on the induced movement quality rewards. The ultimate goal is to achieve dynamic, refined, and personalized lower limb adaptive control based on electromyographic signals, while ensuring safety and robustness.

[0079] The basic elements of reinforcement learning are defined as follows: Regarding the state (s_k), at the beginning of each iteration (the k-th attempt), the system's state s_k is a vector containing: historical performance, i.e., the action consistency index of the past few movements, such as the 3rd to 5th movements; current parameters, the basic stimulation current intensity of each electrode in each phase of the current movement, from the current personalized parameter mapping table; and user context, enabling the agent to perceive real-time changes in the user's state and reflecting the system's adaptive care for the user. Specifically, it consists of: a real-time fatigue index, which is the median frequency of the electromyographic signals on the surface of the target muscle group continuously monitored by the system. It is well known that when muscles are fatigued, the electromyographic signal spectrum shifts to lower frequencies, and the median frequency decreases. The system calculates the percentage decrease in the current median frequency relative to the start of the session, as a quantified fatigue index. When this index exceeds a certain threshold, the system tends to adopt a more conservative strategy, such as reducing the target stimulation intensity or suggesting a shorter exercise duration to avoid excessive fatigue. Training Load Accumulation records the total number of exercise trials for the day, as well as the total amount of exercise in recent days (e.g., the past few days). This reflects the user's cumulative neuromuscular load. On high-load days, the system may be more inclined to choose consolidation rather than challenging exercise strategies to promote supercompensation and avoid overexertion. Neuromuscular Excitability Level is assessed based on the background electromyography (EMG) signal level at rest (without stimulation). Higher background EMG may indicate muscle tension or a tendency to spasticity. When abnormally high excitability is detected, the system prioritizes stability, avoids applying strong stimuli that may induce spasticity, and may trigger an alarm. Safety Boundary Compliance records the number and frequency of events that violate the safety boundary during the current exercise session. Frequent violations of the safety boundary indicate that the current stimulation parameter settings may be close to the user's limits or that the user's state is unstable. The system will receive a strong signal that a more cautious exploration strategy is needed.

[0080] Define the basic elements of reinforcement learning. Regarding actions (a_k), the action a_k taken by the agent in state s_k is defined as a joint adjustment strategy for the stimulation parameters of each electrode. For the m-th electrode, its adjustment scheme is a tuple (D_m, Δ_m). This includes the direction of adjustment (D_m), where D_m ∈ {increase, maintain, decrease}, and the magnitude of adjustment (Δ_m), where Δ_m ∈ {increase, maintain, decrease}. Here, the magnitude is a multiplier relative to a base step size (Δ_base). For example, if Δ_m is "increase", the actual adjustment amount is 1.5 * Δ_base; if it is "maintain", it is 1.0 * Δ_base; and if it is "decrease", it is 0.5 * Δ_base. Therefore, the action a_k of the entire system is a set of adjustment instructions for all M electrodes {(D_1, Δ_1), (D_2, Δ_2), ..., (D_M, Δ_M)}.

[0081] Define the basic elements of reinforcement learning. Regarding reward (r_k), after the k-th iteration, the system receives a reward signal r_k to evaluate the quality of action a_k. The reward function is designed as follows: r_k = w_1 * Improvement + w_2 * SafetyMargin - w_3 * Instability, where Improvement is the improvement in the fit of the current action relative to the average of the previous iterations. This is the main reward. SafetyMargin is used to encourage the system to operate far from the safety boundary in the parameter space to maintain high safety redundancy. It is calculated as follows: SafetyMargin = average((I_safe_max_m - I_k_m) / I_safe_max_m) for all muscles m, where I_safe_max_m is the individualized safe working limit for the m-th muscle channel, and I_k_m is the average stimulation current intensity used by that channel in this iteration. The larger this value, the farther the system's operating point is from the safety limit, the higher the safety redundancy, and the higher the reward. Instability penalty: If the adjustment amplitude of the stimulus current is too large or oscillates frequently, a small negative reward is given to encourage a smooth and stable strategy. w_1, w_2, w_3 are weighting coefficients used to balance different objectives.

[0082] In some specific embodiments, the implementation of reinforcement learning algorithms includes: since the action space is discrete, with 3 directions × 3 amplitudes, each electrode has 9 action choices, but the combined space is large, making it suitable to use Actor-Critic type policy gradient methods, such as Proximal Policy Optimization (PPO). These methods can effectively handle high-dimensional discrete action spaces, and the policy updates are more stable.

[0083] The network structures are: An Actor network (policy network π), which takes the state s_k as input and outputs a probability distribution representing the probability of taking each action across all possible joint actions a_k. The agent samples the actions to be taken based on this probability distribution. A Critic network (value network V), which takes the state s_k as input and evaluates the long-term expected cumulative reward for that state.

[0084] The online learning process randomly initializes the policy network π and value network V. A new network is created for each patient, or a network fine-tuned from a pre-trained model. For each motion iteration k, the system obtains the current state s_k based on the observed state. The policy network π calculates the action probability based on s_k, and the system samples the jointly adjusted action a_k. The system updates the stimulus parameters in the personalized parameter mapping table based on a_k. The user completes a motion action under these parameters. The system calculates the action fit error e_k and the reward r_k. The transformed tuple (s_k, a_k, r_k, s_{k+1}) is stored in the experience replay buffer. A batch of data is periodically sampled from the experience replay buffer to update the Actor and Critic networks, aiming to maximize the long-term cumulative reward.

[0085] In the above embodiments, the agent actively explores optimal strategies. Unlike passive response errors, the agent can proactively try combined strategies such as increasing stimulation of one muscle while decreasing stimulation of another, potentially discovering more effective stimulation patterns than human experts have experienced. Multi-objective optimization, through carefully designed reward functions, allows the system to simultaneously optimize multiple objectives such as movement quality, safety, and comfort, achieving more human-like motion control. Addressing non-stationary environments, the user's neural function dynamically changes during motion control. The agent can continuously track these changes and adjust its strategies, possessing true long-term adaptive capabilities. Achieving super-personalization, the ultimately learned policy network π is a highly personalized stimulation strategy for a specific patient, allowing the brain to deeply understand the patient's unique response to specific stimulation patterns. The third-level closed loop is responsible for long-term, strategic parameter template optimization, with a decision cycle of one complete movement. The second-level closed loop is responsible for short-term, tactical real-time compensation, with a decision cycle of millisecond-level stimulation intervals. The combination of the two constitutes an extremely sophisticated intelligent control system in both time and decision-making levels.

[0086] In a specific embodiment, the modules of the system provided in this application are not simply stacked, but are deeply integrated around the core objective of multi-level closed-loop adaptive control. Its structure is described below:

[0087] The multimodal signal acquisition module includes an EEG cap, electrical stimulation electrodes that can be switched to EMG acquisition mode, an IMU unit, and a video acquisition module. Its key construction lies in ensuring precise alignment of the timestamps of EEG, EMG, and video data through hardware synchronization design, providing a foundation for subsequent data fusion and closed-loop control.

[0088] Intelligent Information Processing and Decision-Making Center. Includes a processor and memory, storing computer programs. The innovative structure of this module lies in its software algorithm, specifically configured to execute the following units: Robust Decoding Unit: Implements the aforementioned multimodal fusion decoding algorithm. Iterative Learning Control Unit: The core of this unit carries the aforementioned third-level closed-loop algorithm, responsible for maintaining and updating the personalized parameter mapping table.

[0089] Adjustable electrical stimulation execution module. Includes a high-precision electrical stimulator. Its key feature is its ability to receive not only trigger commands from the first-level closed loop, but also real-time fine-tuning parameters from the second-level closed loop and optimized new parameter sequences from the third-level closed loop. Its advantage lies in its ability to act as an intelligent execution terminal, precisely responding to the scheduling of control strategies at different levels.

[0090] The human-computer interaction and guidance module includes a display unit for presenting guided animations based on a complex motion visualization paradigm. Its synergy with the system lies in the high degree of matching between the presented animation content and the decoding paradigm set by the information processing module, and the muscle group combinations invoked by the control module, ensuring consistency in the system's guidance, perception, decision-making, and execution.

[0091] The data management and evaluation module stores all data and personalized parameter mapping tables, and generates evaluation reports. Its role is not only as a recorder, but also as a data warehouse for the system's long-term learning and an output terminal for verifying effectiveness.

[0092] In a specific application scenario, lower limb motor dysfunction is a major sequela for patients with neurological diseases such as spinal cord injury and stroke, and their later motor processes are highly dependent on neural plasticity. Traditional motor control methods (such as manual assisted movement) have some effect, but they suffer from problems such as scarce human resources, difficulty in ensuring the intensity and standard of movement, and low patient participation, resulting in limited efficiency in neural remodeling. In recent years, technology-assisted motor control solutions have become a research hotspot, mainly going through the following stages:

[0093] Functional electrical stimulation (FESPS) devices. These devices can stimulate muscles through preset programs to assist patients in completing movements. However, their operating mode is mostly open-loop control, meaning that the application of stimulation is unrelated to the patient's actual movement intention and the performance effect. This results in a lack of active participation in movement and an inability to adjust in real time according to the quality of movement, leading to a lower ceiling for therapeutic efficacy.

[0094] The initial integration of non-invasive brain-computer interfaces (BCIs) with functional electrical stimulation (fEP) represents a preliminary step. These systems attempt to construct a basic "intention-stimulus" closed loop, triggering electrical stimulation by decoding the patient's motor intentions (such as motor imagery), thus increasing patient engagement. However, significant systemic bottlenecks remain: insufficient decoding reliability, as EEG signals are highly susceptible to artifacts from body movement, muscle contraction, and the stimulation itself during dynamic motion. Existing systems often employ a single EEG decoding algorithm, resulting in high false alarm and false negative rates without verification from other modalities. This leads to unreliable system control and may even trigger unexpected actions, posing safety risks. The control strategy is singular and rigid, with the system employing only "trigger-based" control—issuing stimuli with fixed patterns or parameters after detecting an intention. It lacks real-time monitoring and evaluation of command execution effectiveness. Key information such as the patient's muscle recruitment status and the standardity of joint movement trajectories are not incorporated into the control loop. Therefore, the system cannot provide the precise and adaptive assistance that manual intervention can offer, such as a "push" when the patient's effort is insufficient or a "correction" when movements are distorted. The lack of personalization and adaptability is a significant problem. Different patients, and even the same patient at different stages of motor development, exhibit vast differences in neurological function and motor ability. Existing systems struggle to quickly adapt to these changes, requiring numerous manual parameter settings and adjustments by professionals—a cumbersome and experience-dependent process. More importantly, these systems lack the ability to learn and optimize from long-term motor data, failing to grow alongside the patient.

[0095] Preliminary attempts at multimodal feedback have seen some cutting-edge research introducing electromyography (EMG) or kinematic sensors as feedback mechanisms. However, these attempts often remain at the level of simple information aggregation or parallel processing, failing to form a hierarchical and temporally sequential collaborative control logic. For example, EMG signals may only be used to trigger posterior analysis, without being integrated with EEG decoding to improve reliability at the decision-making level; visual feedback may only be used for display, without forming a closed-loop optimization algorithm with stimulus parameters. The fundamental problem lies in the fact that existing system architectures lack a "brain" capable of integrating multimodal information and making intelligent decisions.

[0096] In summary, while existing technologies have made progress in individual technical aspects, their inherent limitations in system architecture have prevented them from providing a solution that can maintain high reliability and achieve personalized, precise adaptive control in dynamic environments. Therefore, there is an urgent need in the field for an innovative system-level solution to address the complex and intertwined technical challenges mentioned above.

[0097] Compared with traditional technologies, this application brings groundbreaking progress through its unique system-level architecture and multi-level closed-loop adaptive control strategy, and has at least the following beneficial effects:

[0098] This application enhances system reliability and security by employing a decision-level fusion mechanism that combines initial EEG decoding with secondary EMG signal analysis, thus constructing a reliable security firewall. It effectively identifies and filters EEG misinterpretations caused by electrical stimulation artifacts or body movement, fundamentally preventing the triggering of unintended electrical stimulation and significantly improving patient safety in dynamic motion environments—a feat unmatched by systems solely reliant on EEG decoding.

[0099] This represents a paradigm shift from trigger-based control to precise adaptive control. The core contribution of this application lies in establishing a three-tiered closed-loop collaborative control system with rapid feedforward feedback and lag optimization. The first-tier closed loop addresses whether the user's brain is involved. The second-tier closed loop (real-time electromyography compensation) addresses the effectiveness of the stimulus, ensuring that every neural command is translated into effective muscle contraction, achieving precise microscopic control during a single movement. The third-tier closed loop (visually guided iterative learning) addresses the standardization of the movement. By continuously optimizing a personalized electrical stimulation parameter mapping table, it drives the user's movement quality to continuously approach the standard, achieving macroscopic adaptive optimization across movement cycles. This control paradigm overcomes the bottleneck of existing systems' rigid control strategies and inability to adjust in real-time based on execution results, significantly improving the effectiveness of single movements.

[0100] This application endows the system with the intelligence of long-term learning and personalized adaptation. By introducing iterative learning algorithms and personalized parameter mapping tables, the system is no longer a static tool, but one that can learn from historical data and evolve along with the patient's movement progress. It can automatically adapt to the dynamic changes in the patient's neurological function, continuously optimize stimulation strategies, reduce reliance on manual adjustments, and truly realize the personalization and automation of a dynamic lower limb control system based on electromyographic signals, making long-term home exercise possible.

[0101] A new standard for objective and quantitative motion assessment. This application uses the motion matching accuracy based on quantified skeletal animation as the core feedback indicator and assessment benchmark. This provides a quantifiable and objective measure of motion control effectiveness that is independent of subjective experience, enabling precise and scientific horizontal and vertical comparisons of motion effects between different users and at different times for the same user, greatly improving the scientific rigor and credibility of motion assessment.

[0102] In addition, it should be noted that this application can be used not only for users with limb and lower limb motor dysfunction, but also for assistive motor control of other limbs.

[0103] It is understood that the computer equipment in the limb adaptive control system solution provided in this application can be a server, and its internal structure diagram can be as follows: Figure 5As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The database stores relevant data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the method provided in this application.

[0104] Those skilled in the art will understand that Figure 5 The structures shown are merely block diagrams of a portion of the structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. The computer device may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program may include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0105] It should be understood that the processor mentioned in the embodiments of this application can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A limb adaptive control system, characterized in that, The system includes: The first-level control module is used to initiate electrical stimulation through first-level feedforward control after detecting the user's valid movement intention. The second-level control module is used to adjust individual stimuli in a complete planned action based on an electrical stimulation sequence after the electrical stimulation is initiated; during the interval between different individual stimuli, it collects electromyographic signals within the current time window and evaluates the actual muscle recruitment degree; based on the comparison result of the actual muscle recruitment degree with the preset baseline of muscle recruitment degree within the current time window, it dynamically adjusts the stimulation current of individual stimuli in the next time window, wherein the magnitude of the stimulation current is related to the magnitude of muscle recruitment degree; The third-level control module is used to adjust a complete planned action based on an electrical stimulation sequence. It obtains the user's current state parameters, makes a decision based on the current policy network model to obtain the adjustment action and stimulation parameters corresponding to the maximum probability, performs adaptive stimulation based on the stimulation parameters, and updates the current policy network model based on the effect of the current complete planned action based on the electrical stimulation sequence. The updated policy network model guides the next complete planned action.

2. The system according to claim 1, characterized in that, The first-level control module also includes a motion intention monitoring unit, a stimulus parameter initialization unit, and an execution unit: The motion intention monitoring unit is used to activate the electrical stimulation sequence based on the user's valid motion intention; The stimulation parameter initialization unit is used to obtain the initial electrical stimulation parameter sequence corresponding to the user from a pre-set mapping table; An execution unit is configured to perform electrical stimulation using the initial electrical stimulation parameter sequence, wherein the initial electrical stimulation parameter sequence includes at least one of personalized initial stimulation intensity, muscle recruitment baseline for a corresponding time window, and safety boundary information.

3. The system according to claim 1, characterized in that, The second-level control module also includes an evaluation unit and an adjustment unit: The evaluation unit is used to evaluate the actual muscle recruitment degree based on the electromyographic signals collected within the current time window during the stimulation interval. The adjustment unit is used to compare the actual muscle recruitment degree with a preset muscle recruitment degree baseline within the current time window. If the actual muscle recruitment degree is less than the lower limit of the preset muscle recruitment degree baseline, the stimulation current is increased according to a preset threshold to improve the actual muscle recruitment degree of a single stimulus in the next time window, and the adjusted stimulation current is constrained within a preset safety constraint range. If the actual muscle recruitment degree is greater than the upper limit of the preset muscle recruitment degree baseline, the stimulation current is decreased according to a preset threshold to reduce the actual muscle recruitment degree of a single action in the next time window. If the actual muscle recruitment degree is within the range of the upper and lower limits of the preset muscle recruitment degree baseline, the electrical stimulation of a single stimulus in the next time window continues to be performed according to the current stimulation current. The muscle recruitment degree baseline and preset threshold corresponding to a single stimulus in different time windows are different.

4. The system according to claim 3, characterized in that, The second-level control module also includes a muscle recruitment baseline setting unit, which is used for: Prior to electrical stimulation, electromyographic signals were acquired from the user to complete a standard movement within a complete planned movement cycle based on an electrical stimulation sequence. Based on the physiological phase of the user's movements, a complete planned movement cycle based on an electrical stimulation sequence is divided into multiple time windows, and the baseline of muscle recruitment corresponding to the electromyographic signal of a single movement within each time window is calculated.

5. The system according to claim 1, characterized in that, The third-level control module further includes an update unit, which is used for: Obtain the concordance error and reward between the current planned action and the standard stimulus action based on an electrical stimulation sequence; The policy network model and value network model are updated based on the matching error and reward, and the updated policy network model and value network model are used to guide the next complete plan action.

6. The system according to claim 5, characterized in that, The update unit further includes a reward calculation subunit, which includes at least one of performance improvement, safety margin, and stability penalty; Performance improvement is used to indicate the degree of improvement in the consistency of the current action relative to the average of the previous few times; safety margin is used to indicate operation within the safety boundary parameter space; stability penalty is used to indicate the penalty for large changes in the stimulation current.

7. The system according to claim 1, characterized in that, The system further includes a motion state determination module, which is used for: Collect the user's electroencephalogram (EEG) and electromyogram (EMG) signals; The motion state of the electroencephalogram (EEG) signal is determined based on the electromyographic activity markers of the electromyographic signal. When a valid movement intention is determined, the limb is adaptively controlled through a multi-level closed-loop adaptive control strategy. The multi-level closed-loop adaptive control strategy includes at least one of the following: a first-level feedforward control, a second-level compensation control, and a third-level iterative learning control.

8. The system according to claim 7, characterized in that, The motion state determination module further includes a joint determination unit, which is used for: Based on the multi-band data of the EEG signal, the phase lock value between whole brain electrodes is calculated, a functional connectivity matrix is ​​constructed, and network topology features are extracted to obtain the EEG feature vector. The root mean square value is extracted from the electromyographic envelope of the electromyographic signal as a muscle activation intensity feature, and the electromyographic activity marker bit of the electromyographic signal is determined based on the muscle activation intensity feature. The validity of the motion state is obtained by performing intention recognition on the EEG feature vectors using a preset classifier; When the validity of the movement state corresponds to a valid movement intention and the electromyographic activity marker corresponds to significant activity, it is determined to be a valid movement intention; When the validity of the movement state corresponds to a valid movement intention and the electromyographic activity marker corresponds to no significant activity, it is determined to be an invalid movement intention.

9. The system according to claim 1, characterized in that, The system further includes a pre-configuration module, which is used for: Determine the type of stimulus action; Obtain the electrode application position corresponding to the stimulation action type and the general reference range of initial stimulation current intensity from the pre-set stimulation parameter template library; The electrode placement position is used to instruct the user to place the electrode at the corresponding position for motion control, and the general initial stimulation current intensity reference range is used to indicate the personalized initial stimulation intensity corresponding to the user through dynamic adjustment.

10. The system according to claim 1, characterized in that, The third-level control module also includes a status parameter setting unit and an action parameter setting unit: The state parameter setting unit is used to set the state parameters including at least one of the following: consistency of historical motion actions, current intensity, state change, fatigue, exercise load, muscle excitability, and safety boundary compliance. The motion parameter setting unit is used to set the motion parameters, including at least one of control direction and control intensity.

Images