Cloud-edge hybrid AI artificial spinal cord system and control method

Abstract

The invention discloses a clinical application-oriented cloud edge hybrid AI artificial spinal cord system and a control method, and relates to the technical field of neural engineering and brain-computer interfaces. The system comprises a neural signal acquisition module, a multi-mode perception fusion module, an edge real-time control unit, a cloud intelligent training unit, a nerve stimulation output module, a somatosensory feedback module, a clinical-level safety fusion module and a wireless communication module. According to the invention, a core architecture of low-delay real-time control at an edge end less than or equal to 15ms, cloud individualized AI model training, rehabilitation process dynamic adaptation and networking non-inductive upgrading is adopted, the motion intention decoding accuracy is improved through multi-modal sensing fusion, and the use security and data privacy are guaranteed through a clinical-level security compliance system; the motion function of the patient with the spinal cord transection injury can be reconstructed, the control strategy can be dynamically optimized according to the rehabilitation process of the patient, and the lifelong iteration upgrading of the system is realized. The problems that in the prior art, pure edge computing power is insufficient, pure cloud delay is too high, individual adaptation and clinical compliance core pain points are lacked are solved, the motion intention recognition accuracy is larger than or equal to 98%, clinical conversion can be directly achieved, the method is suitable for rehabilitation of diseases such as spinal cord injury, cerebral apoplexy and cerebral palsy and nerve function replacement, and the commercialized application prospect is wide.

Description

Technical Field

[0001] This invention relates to the fields of neural engineering, brain-computer interface, artificial intelligence and rehabilitation engineering, specifically to a clinically applicable, personalized, adaptive cloud-edge hybrid AI artificial spinal cord system and control method, which is particularly suitable for motor function reconstruction, rehabilitation process adaptation and long-term neuromodulation in patients with spinal cord transection injuries. Background Technology

[0002] Transverse spinal cord injuries permanently disrupt the signal transmission pathway between the brain and peripheral nerves, resulting in loss of motor, sensory, and autonomic functions below the level of injury. This poses a global challenge in clinical medicine. While advancements have been made in technologies such as neurostimulation, exoskeletons, and brain-computer interfaces, key limitations remain.

[0003] 1. Pure edge computing devices have limited computing power and cannot complete high-precision, individualized neural signal decoding. They also have low control precision and are difficult to adapt to the differences in neural characteristics of different patients.

[0004] 2. Pure cloud computing architecture is limited by network latency (usually >30ms), which cannot meet the real-time requirement of <15ms for natural human movement control, and is prone to safety risks such as imbalance and falls;

[0005] 3. It lacks network outage security redundancy and clinical-grade fault handling mechanisms, which does not meet the safety and compliance requirements of medical devices and makes it difficult to achieve clinical translation;

[0006] 4. The system model is a fixed version shipped from the factory and cannot be updated online for long-term iteration. Furthermore, the control strategy cannot be dynamically adjusted according to the patient's recovery process, and the control effect will gradually decrease as recovery progresses.

[0007] 5. Decoding relies solely on a single neural electrical signal, resulting in weak anti-interference capabilities, lack of integration of multimodal perception data, and insufficient control stability and naturalness;

[0008] 6. The lack of clinical-grade data anonymization and privacy protection mechanisms makes it impossible to meet the requirements of medical data security regulations.

[0009] Therefore, the industry urgently needs an artificial spinal cord replacement technology that combines real-time performance, intelligence, security, personalized adaptability, and clinical compliance, and supports long-term iterative upgrades, in order to achieve clinical translation and commercialization. Summary of the Invention

[0010] The technical problem to be solved by this invention overcomes all the shortcomings of the prior art, and provides a cloud-edge hybrid AI artificial spinal cord system and control method, which resolves the technical contradiction between real-time performance and intelligence, achieves personalized and precise control, and meets the safety and compliance requirements of medical devices.

[0011] Technical solution

[0012] A cloud-edge hybrid AI artificial spinal cord system includes: a neural signal acquisition module, a multimodal perception fusion module, an edge real-time control unit, a cloud-based intelligent training unit, a neural stimulation output module, a somatosensory feedback module, a clinical-grade safety and compliance module, and a wireless communication module.

[0013] 1. Nerve signal acquisition module: Using minimally invasive nerve sleeve electrodes or high-density epidermal electromyography electrodes, placed on the peripheral nerve upstream of the spinal cord injury plane, it is used to acquire nerve electrical signals of motor intention; the module includes low-noise amplification circuit, filtering circuit, analog-to-digital conversion unit, and supports 8 to 64 channels expandable configuration.

[0014] 2. Multimodal perception fusion module: integrates electromyography signal sensor, plantar pressure sensor, joint angle sensor, muscle tension sensor, tactile sensor array and posture gyroscope, used to synchronously collect nerve signals, limb movement status and environmental feedback data; and filters out interference noise from single signals through data fusion algorithm.

[0015] 3. Edge Real-Time Control Unit: Composed of an FPGA real-time controller compliant with medical device electromagnetic compatibility standards and a low-power edge AI chip, it is used for real-time decoding of neural signals and multimodal data, generation of individualized motion commands, and closed-loop control, with a control latency of ≤15ms. It supports independent operation without network access and can download and load the latest individualized AI model from the cloud when the patient is stationary, enabling seamless iterative upgrades.

[0016] 4. Cloud-based intelligent training unit: Composed of a cloud-based GPU server that meets medical data security standards, a personalized neural database, an AI training platform, and a model release and management system, it is used to receive anonymized personalized data uploaded from the edge, train and dynamically optimize personalized AI models and rehabilitation adaptation models, and release the latest version of the AI ​​model to support edge updates.

[0017] 5. Neurostimulation output module: Multi-channel, independently controllable clinical-grade neurostimulation cuff electrodes or functional neuromuscular stimulation units are placed on the nerves or target muscle groups downstream of the injury plane; supports individualized adjustment of stimulation intensity, frequency, and pulse width, and outputs precise electrical pulses after receiving instructions from the edge unit.

[0018] 6. Somatosensory feedback module: In conjunction with the multimodal perception fusion module, it collects limb movement status, muscle tension changes, and tactile feedback data in real time to form a closed-loop control of "movement command - limb response - feedback adjustment".

[0019] 7. Clinical-grade safety and compliance module: Integrates a safety fault-tolerance unit, a data anonymization unit, an access control unit, and a fault self-diagnosis unit. It monitors network latency, abnormal nerve signals, limb posture imbalance, and equipment failure in real time. When network latency exceeds 20ms, a dangerous posture occurs, or equipment failure occurs, it automatically triggers emergency braking, cuts off nerve stimulation output, and issues an audible and visual alarm. It performs real-time anonymization processing on data uploaded from the edge. It enables hierarchical access control for doctors, patients, and technicians. It performs real-time fault diagnosis on electrodes, chips, and communication modules.

[0020] 8. Wireless communication module: Adopts 5G / Wi-Fi 6 low-latency and high-security communication protocol, supports asynchronous data transmission between edge and cloud, personalized model download and iterative upgrade; integrates communication encryption module.

[0021] A cloud-edge hybrid AI artificial spinal cord control method includes the following steps:

[0022] (1) Multi-source data acquisition: Acquire motor intention neural electrical signals upstream of the spinal cord injury plane, and simultaneously acquire multimodal data such as electromyography, limb posture, plantar pressure, muscle tension, and tactile feedback; (2) Real-time preprocessing at the edge: Denoise, extract features, and fuse neural signals and multimodal data; (3) Personalized real-time inference: Load a lightweight AI model specific to the patient, complete real-time decoding of motor intention, generate personalized motor control commands, and control delay ≤15ms; (4) Precise output of neural stimulation: Output electrical pulses adapted to the patient's neural characteristics to activate downstream nerves and muscles and drive limb movement; (5) Dynamic adjustment of somatosensory closed loop: Collect motor response data in real time and dynamically adjust the somatosensory closed loop. (6) Asynchronous upload of desensitized data: After desensitizing the multi-source data, it is asynchronously uploaded to the cloud intelligent training unit; (7) Cloud-based individualized model training and rehabilitation adaptation optimization: Based on the desensitized data, an individualized AI model is trained, and the model parameters are dynamically optimized to generate a staged rehabilitation adaptation model according to the patient's rehabilitation process; (8) Cloud-based model release and edge-end seamless upgrade: The latest individualized AI model and rehabilitation adaptation model are released in the cloud, and the edge-end downloads and completes incremental updates when the patient is stationary; (9) Full-process clinical-grade safety monitoring: Real-time monitoring of device status, network status, patient posture and data security, and immediate triggering of emergency braking and alarm when abnormality occurs.

[0023] Beneficial effects

[0024] 1. Real-time performance and accuracy are both achieved: edge control latency is ≤15ms, and the motion intention recognition accuracy is ≥98% when combined with multimodal perception fusion and individualized AI model.

[0025] 2. Strong individualized adaptation capability: The cloud-based AI model is customized for each patient and dynamically optimized according to the rehabilitation process to adapt to the different neurological characteristics and rehabilitation stages of different patients.

[0026] 3. Lifetime upgradeable: Supports automatic download of the latest personalized AI models and algorithms via the internet, making the system more accurate the more it is used.

[0027] 4. Clinical-grade safety and compliance: It has functions such as independent operation without network access, emergency braking, fault self-diagnosis, data anonymization, and hierarchical access control, and complies with medical device safety standards and medical data security regulations.

[0028] 5. Complete pathway and strong scalability: It realizes a complete closed loop of neural function reconstruction and can be extended to the fields of stroke rehabilitation, cerebral palsy rehabilitation and other fields.

[0029] 6. Great commercialization potential: It can be directly licensed to medical device companies to achieve mass production. Attached Figure Description

[0030] Figure 1 Overall architecture diagram of the cloud-edge hybrid AI artificial spinal cord system (including multimodal fusion and compliance modules): arranged linearly from top to bottom, with the brain / afferent nerve icon on the left, and the neural signal acquisition module, multimodal perception fusion module, and edge real-time control unit on the right in sequence; the edge real-time control unit is connected to the cloud intelligent training unit via bidirectional arrows; the edge real-time control unit is connected downwards to the neural stimulation output module, and then points to the outgoing nerve / muscle / limb icon; the limb icon is connected to the somatosensory feedback module, and its output arrow returns to the multimodal perception fusion module; a dashed box covers the edge real-time control unit, cloud intelligent training unit, wireless communication module, and neural stimulation output module, with the box labeled "Clinical-grade safety and compliance module (safety fault tolerance, data anonymization, access control level, emergency braking)".

[0031] Figure 2 Neural signal and multimodal data flow and closed-loop control flowchart: The main closed-loop flowchart on the left, from top to bottom: Multi-source data acquisition → Real-time preprocessing → Multimodal fusion → Personalized AI decoding → Stimulus output → Limb movement → Somatosensory feedback → Return to multimodal fusion. Next to personalized AI decoding, it is marked "Control delay ≤ 15ms," and between stimulus output and limb movement, it is marked "Emergency braking trigger node." The asynchronous upgrade process on the right branches off from multi-source data acquisition with dashed arrows: Edge-end desensitized data → Cloud storage → Personalized model training → Rehabilitation adaptation optimization → Model release → Edge-end download → Seamless update → Pointing to the personalized AI decoding module.

[0032] Figure 3Hardware structure diagram of the edge real-time control unit (including fault self-diagnosis): From left to right: Electrode signal interface, multimodal data interface → Filtering and amplification circuit → FPGA controller → Edge AI chip (labeled "Individualized model storage") → Fault self-diagnosis unit → Storage unit → Stimulation-driven output (right output port) and wireless communication module (labeled "Encrypted transmission"). The fault self-diagnosis unit has an additional control line directly connected to the stimulation-driven output.

[0033] Figure 4 Schematic diagram of cloud-based personalized model training, rehabilitation adaptation optimization, and iterative updates: From left to right: Electrode signal interface, multimodal data interface → Filtering and amplification circuit → FPGA controller → Edge AI chip (labeled "Personalized Model Storage") → Fault self-diagnosis unit → Storage unit → Stimulation-driven output (right output port) and wireless communication module (labeled "Encrypted Transmission"). The fault self-diagnosis unit has an additional control line directly connected to the stimulation-driven output. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0035] Example 1: Specific Implementation of System Modules

[0036] 1. The nerve signal acquisition module uses a 32-channel minimally invasive nerve cuff electrode, placed above the lumbar plexus nerve above the patient's injury level. The electrode interface connects to a low-noise amplifier circuit (60dB gain, 5-500Hz bandwidth), which outputs a digital signal after 24-bit analog-to-digital conversion. The electrode configuration can be adjusted to 8-64 channels according to the patient's nerve distribution.

[0037] 2. The multimodal perception fusion module integrates a high-density electromyography sensor (sampling rate 2kHz), a plantar pressure sensor (100Hz), a knee / hip joint angle sensor (50Hz), a posture gyroscope (100Hz), and a tactile sensor array. All sensor data are aligned through a clock synchronization mechanism and spatiotemporal fusion is performed using Kalman filtering to eliminate noise and drift from individual signals.

[0038] 3. The edge real-time control unit hardware employs a Xilinx Zynq UltraScale+ FPGA (XCZU9EG) and a low-power AI chip (such as an NVIDIA Jetson TX2 NX). The FPGA handles real-time signal preprocessing (filtering, feature extraction), while the AI ​​chip runs a lightweight inference model. The system's measured inference latency is 11ms, meeting the ≤15ms requirement. The unit has a built-in fault self-diagnosis program that checks electrode impedance, chip temperature, and communication status every 10ms, triggering emergency braking in case of abnormalities.

[0039] 4. Cloud-based Intelligent Training Unit: Deployed in the cloud using a medical-grade GPU server (NVIDIA A100) and an encrypted database. Based on anonymized data uploaded by patients, a personalized model is trained using transfer learning. The model architecture is CNN-LSTM-Attention, with inputs consisting of multi-channel neural signals and multimodal features, and outputs including movement intention categories (standing, walking, turning, climbing stairs, etc.) and stimulus parameters (intensity, frequency, pulse width). Training uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64.

[0040] 5. Specific implementation of the multimodal fusion algorithm

[0041] Data preprocessing: Data from all sensors were resampled to 100Hz. Nerve signals were bandpass filtered to extract peak sequences. Electromyographic signals had their envelopes extracted. Pressure and angle signals were normalized. The sliding window length was 500ms, and the step size was 50ms.

[0042] Feature extraction: Extract time-domain features (root mean square, zero-crossing rate, waveform length), frequency-domain features (energy of power spectral density in the alpha, beta, and gamma bands), and sample entropy from neural signal windows; extract mean absolute value and wavelength from electromyographic envelopes; pressure and angle signals are directly used as time-series features.

[0043] Fusion Network: A multimodal fusion network based on an attention mechanism is employed. First, local features of each modality are extracted using a one-dimensional CNN (kernel size 3, stride 1, 64 channels), then input into a bidirectional LSTM (128-dimensional hidden layers) to capture temporal dependencies. The outputs of each modality's LSTM are then processed through a multi-head attention layer (8 heads) to calculate intermodal correlation weights, and weighted fusion is performed to obtain a joint representation. Finally, classification and regression results are output through a fully connected layer. During training, cross-entropy loss (classification) and mean squared error loss (regression) are jointly optimized, with a weight coefficient λ=0.5.

[0044] 6. Specific Implementation of Personalized AI Model Training and Rehabilitation Adaptation Optimization

[0045] Base model pre-training: A general decoding model is pre-trained using public datasets (such as NINAP), and robust neural features are extracted using self-supervised contrastive learning (SimCLR).

[0046] Personalized fine-tuning: For each patient, one hour of standard action data was collected, and the model was fine-tuned through transfer learning. The parameters of the underlying CNN were fixed, and the LSTM and fully connected layers were fine-tuned, with the learning rate reduced to 0.0001. At the same time, a reinforcement learning branch (PPO algorithm) was trained, using action completion, energy consumption, and safety as reward functions to optimize the stimulus parameter output strategy.

[0047] Dynamic optimization during the rehabilitation phase: Model retraining is triggered every two weeks based on rehabilitation physician assessments (such as ASIA scores). Incremental learning (elastic weight consolidation) is used to integrate new data and generate a phased model. The model is then deployed to the edge after A / B testing validation.

[0048] 7. The nerve stimulation output module uses 64-channel clinical-grade stimulation electrodes, placed on target points such as the sciatic nerve and common peroneal nerve in the lower limbs. The adjustable stimulation parameters are: intensity 0-10mA, frequency 20-100Hz, and pulse width 50-500μs. The module features current monitoring and overload protection.

[0049] 8. Clinical-grade safety and compliance module

[0050] Data anonymization: Differential privacy algorithm (ε=1.0) is used to remove identity information before data is uploaded.

[0051] Access control: Role-based access control (RBAC) allows doctors to adjust rehabilitation parameters, patients can only view data, and technicians are responsible for maintenance.

[0052] Emergency braking: When network latency >20ms, joint angle exceeds limits, or equipment malfunctions, immediately cut off stimulation output and trigger an audible and visual alarm. Fault logs are stored locally for later review.

[0053] Example 2: Individualized functional reconstruction in patients with complete T10 spinal cord transection

[0054] The patient is a 35-year-old male with a complete T10 lesion and an ASIA grade A. The following steps were performed:

[0055] 1. Implant 32-channel nerve cuff electrodes into the lumbar plexus, and wear multimodal sensors (electromyography, plantar pressure, joint angle, posture gyroscope, tactile array).

[0056] 2. An initial general-purpose model is deployed on the edge unit (FPGA + Jetson TX2 NX) for real-time decoding and stimulus output. In the initial stage, standing and slow walking were achieved with a recognition accuracy of 85%.

[0057] 3. Data was collected continuously for one week, anonymized, and then uploaded to the cloud. An individualized Transformer+DRL model was trained in the cloud, improving accuracy to 98.5%.

[0058] 4. The model was optimized every two weeks based on rehabilitation progress. By week 4, the patient could walk and turn steadily; by week 8, the patient could climb 5cm steps, with a 45% improvement in movement fluency and a 30% improvement in rehabilitation efficiency.

[0059] 5. End-to-end security module monitoring: During network outage testing, the edge unit independently ran the basic control program, completing standing and slow walking maneuvers without any safety incidents. Data anonymization and access control passed medical compliance review.

[0060] Industrial applicability

[0061] This invention can be applied to fields such as spinal cord injury rehabilitation, post-stroke motor function reconstruction, and cerebral palsy rehabilitation, and has good prospects for clinical translation and commercial value.

Claims

1. A cloud-edge hybrid AI artificial spinal cord system, characterized in that, include: The system comprises a neural signal acquisition module, a multimodal perception fusion module, an edge real-time control unit, a cloud-based intelligent training unit, a neural stimulation output module, a somatosensory feedback module, a clinical-grade safety and compliance module, and a wireless communication module. The neural signal acquisition module acquires motor intention neural electrical signals upstream of the spinal cord injury plane and converts them into digital signals. The multimodal perception fusion module simultaneously acquires and fuses multimodal data such as electromyography, limb posture, plantar pressure, muscle tension, and tactile feedback. The edge real-time control unit consists of an FPGA real-time controller compliant with medical device electromagnetic compatibility standards and a low-power edge AI chip. It decodes neural signals and multimodal data in real time, generates individualized movement commands, and implements closed-loop control with a control latency ≤15ms. It supports independent operation of basic movement control programs even when offline and can download and load the latest individualized AI models published by the cloud-based intelligent training unit via the wireless communication module when the patient is stationary, enabling seamless iterative upgrades. The cloud-based intelligent training unit includes a medical-grade GPU server, an individualized... The system includes a neural database, AI training platform, and model release management system. These systems receive anonymized, individualized data uploaded by the edge real-time control unit, train and dynamically optimize individualized AI models and rehabilitation adaptation models, and release the latest version of the AI ​​model for edge-end updates. The neural stimulation output module is a multi-channel, independently controllable clinical-grade electrical stimulation unit that receives instructions from the edge real-time control unit and outputs individualized electrical pulses to activate nerves or muscles downstream of the injury plane. The somatosensory feedback module collects limb movement status, muscle tension, and tactile feedback data in real time and feeds the data back to the edge real-time control unit for closed-loop control. The clinical-grade safety and compliance module includes a safety fault-tolerance unit, a data anonymization unit, a permission management unit, and a fault self-diagnosis unit. It monitors device status, network latency, patient posture, and data security in real time, and automatically triggers emergency braking, cutting off neural stimulation output and issuing audible and visual alarms when network latency exceeds 20ms, dangerous postures occur, or device malfunctions. It also performs data anonymization and hierarchical permission management. The wireless communication module uses the 5G / Wi-Fi 6 protocol for low-latency, encrypted data transmission and model download iteration between the edge real-time control unit and the cloud-based intelligent training unit.

2. The system according to claim 1, characterized in that, The neural signal acquisition module uses 8-64 channel expandable minimally invasive neural sleeve electrodes or high-density epidermal electromyography electrodes, and has built-in low-noise amplification circuit, filtering circuit and analog-to-digital conversion unit.

3. The system according to claim 1, characterized in that, The multimodal perception fusion module integrates electromyography (EMG) sensors, plantar pressure sensors, joint angle sensors, muscle tension sensors, tactile sensor arrays, and attitude gyroscopes. It also performs spatiotemporal alignment and feature-level fusion of multi-source data through Kalman filtering or deep learning fusion networks to filter out interference noise from single signals.

4. The system according to claim 1, characterized in that, The individualized AI model mounted in the edge real-time control unit is a lightweight CNN-LSTM and deep reinforcement learning inference model. This model is compressed through model pruning and quantization techniques and deployed on the edge AI chip, and supports incremental updates.

5. The system according to claim 1, characterized in that, The cloud-based intelligent training unit retrains or fine-tunes the parameters of the individualized AI model every 1 to 4 weeks based on the clinical assessment data collected periodically during the patient's rehabilitation process, generates a phased rehabilitation adaptation model, and pushes the updated model to the edge through the model release management system.

6. The system according to claim 1, characterized in that, The neural stimulation output module supports individualized adjustment of stimulation intensity, frequency, and pulse width, and has current monitoring and overload protection functions to avoid muscle damage caused by overstimulation.

7. The system according to claim 1, characterized in that, The data anonymization unit of the clinical-grade security and compliance module uses differential privacy or k-anonymity algorithms to remove patient identity information. The access control unit implements hierarchical access management for doctors, patients, and technicians based on a role-based access control model. The fault self-diagnosis unit monitors electrodes, chips, and communication modules in real time and records fault logs.

8. A cloud-edge hybrid AI artificial spinal cord control method, characterized in that, Includes the following steps: (1) Multi-source data acquisition: The neural signal acquisition module acquires the motor intention neural electrical signals upstream of the spinal cord injury plane, and the multimodal perception fusion module simultaneously acquires multimodal data such as electromyography, limb posture, plantar pressure, muscle tension, and tactile feedback; (2) Edge real-time preprocessing: The edge real-time control unit performs denoising, feature extraction and fusion processing on the neural signals and multimodal data to filter out interference noise; (3) Personalized real-time reasoning: The edge real-time control unit loads a lightweight AI model specific to the patient, completes real-time decoding of motor intention, and generates personalized motor control instructions with a total control delay ≤15ms; (4) Precise output of neural stimulation: The neural stimulation output module receives control instructions and outputs electrical pulses adapted to the patient's neural characteristics to activate the nerves and muscles downstream of the injury plane and drive limb movement; (5) Dynamic adjustment of somatosensory closed loop: The somatosensory feedback module acquires motor response data in real time, and the edge real-time control unit... (6) Asynchronous uploading of desensitized data: After the edge real-time control unit desensitizes the multi-source data, it asynchronously uploads it to the cloud intelligent training unit through the wireless communication module during the control gap; (7) Cloud-based individualized model training and rehabilitation adaptation optimization: The cloud intelligent training unit trains individualized AI models based on desensitized data, and dynamically optimizes model parameters to generate staged rehabilitation adaptation models based on clinical assessment data in the patient's rehabilitation process; (8) Cloud-based model release and edge-based seamless upgrade: The cloud intelligent training unit releases the latest individualized AI model and rehabilitation adaptation model, and the edge real-time control unit downloads and completes incremental updates when the patient is stationary or in a non-moving state; (9) Full-process clinical-grade safety monitoring: The clinical-grade safety compliance module monitors the device status, network status, patient posture and data security in real time, and immediately triggers emergency braking and alarm when abnormalities occur.

9. The method according to claim 8, characterized in that, The feature extraction in step (2) includes time-domain features, frequency-domain features and nonlinear features. The fusion processing adopts a multimodal fusion network based on the attention mechanism, which weights and fuses the features of different modalities and then inputs them into the decoding model.

10. The method according to claim 8, characterized in that, The individualized AI model training described in step (7) adopts a transfer learning strategy, starting with a basic model pre-trained on a large-scale public neural dataset, and fine-tuning it using the patient's own data. The training objective is to minimize the joint loss function of the motor intention classification error and the stimulus parameter prediction error. The rehabilitation adaptation optimization is based on a reinforcement learning framework, using the patient's motor fluency, energy consumption and safety as reward functions, and dynamically adjusting the model output strategy.

11. The method according to claim 8, characterized in that, In step (2), the basic motion control program that runs independently on the edge device when the network is disconnected is a preset general control mode, which can realize the safety control of standing and slow walking, and store local data to be retransmitted after the network is restored.

12. The method according to claim 8, characterized in that, The method can be extended to the reconstruction of neurological function in stroke rehabilitation and cerebral palsy rehabilitation, and can be achieved by adjusting the signal acquisition location and stimulation target point.

Images