Neural feedback rehabilitation training method and system

Abstract

The invention relates to the technical field of neural rehabilitation training, and discloses a neural feedback rehabilitation training method and system. According to the method, electroencephalogram, near-infrared, myoelectricity and heart rate variability signals are acquired based on a multi-modal nerve-physiological signal acquisition device, and a nerve-physiological signal sequence is generated and input into a fusion processing system. And carrying out denoising, artifact removal and feature extraction on the data sequence to generate standardized feature data. And based on the standardized data, identifying the neural state of the user by utilizing a graph neural network, and generating personalized neural state trend data by adopting an attention mechanism in combination with historical data. Based on trend data and a training target, reinforcement learning is adopted to generate a personalized training strategy, training is implemented through an immersive interaction system, user behaviors and neural feedback are recorded, the training strategy is dynamically adjusted and uploaded to a cloud end, and a cloud-side collaborative optimization neural state model and algorithm are adopted, so that the intelligence and accuracy of rehabilitation training are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of neurorehabilitation, artificial intelligence and human-computer interaction, and in particular to a neurofeedback rehabilitation training method and system. Background Art

[0002] At present, neurofeedback rehabilitation training, as an important means of brain function rehabilitation, has been widely used in the intervention process of various neurological diseases such as cognitive impairment, motor dysfunction, attention deficit, etc. However, the existing neurofeedback training technology has the following prominent problems:

[0003] On the one hand, traditional rehabilitation training systems mostly rely on single modality signals (such as only based on EEG or heart rate signals), which cannot comprehensively reflect the user's complex neuro-physiological state during training, resulting in incomplete perception of the user's actual state and limited accuracy and personalization of training feedback.

[0004] On the other hand, the existing system lacks dynamic identification and real-time adjustment mechanisms, and is unable to dynamically adjust training tasks and feedback methods based on users' real-time behavioral responses and neural feedback, resulting in a lack of targetedness in the rehabilitation training process and difficulty in adapting to the user's rehabilitation status that changes over time, affecting rehabilitation effects and user compliance.

[0005] In addition, the lack of efficient intelligent algorithms, especially the lack of autonomous training strategy generation mechanism based on reinforcement learning, makes it difficult for existing rehabilitation training to automatically optimize personalized training plans based on user history and real-time data. At the same time, there is also a lack of immersive interactive methods that integrate multiple channels (visual, auditory, and tactile), which cannot fully mobilize user participation enthusiasm and the potential for reshaping neural functions.

[0006] Therefore, there is an urgent need for an intelligent rehabilitation training method and system that can dynamically identify the user's personalized neural state based on multimodal neural-physiological signals such as EEG, near-infrared, electromyography, and heart rate variability, combined with user behavioral responses, and generate rehabilitation training strategies that adapt to user state changes in real time through reinforcement learning algorithms, while achieving real-time feedback through immersive multi-channel interaction, so as to achieve a precise, dynamic, closed-loop rehabilitation training process, effectively improve the intelligence, adaptability, and personalization of neurofeedback training, and thus improve the rehabilitation training effect. Summary of the invention

[0007] The present invention provides a neurofeedback rehabilitation training method and system to solve the problem of how to dynamically identify the user's personalized neural state based on multimodal neural-physiological data such as EEG signals, near-infrared signals, electromyographic signals and heart rate variability, and combine real-time behavioral response to generate personalized rehabilitation training strategies that adapt to user state changes through reinforcement learning and immersive multi-channel feedback, thereby realizing an accurate, dynamic and closed-loop neurofeedback rehabilitation training process.

[0008] In order to solve the above technical problems, the present invention provides a neurofeedback rehabilitation training method, comprising:

[0009] Acquire the user's electroencephalogram (EEG) signals, near-infrared signals, myoelectric signals, and heart rate variability physiological indicators based on a multimodal neuro-physiological signal acquisition device, generate a neuro-physiological signal data sequence containing multi-dimensional time series data, and input the neuro-physiological signal data sequence into a fusion processing system in real time;

[0010] Performing joint preprocessing on the neuro-physiological signal data sequence, including noise removal, artifact removal, and abnormal band suppression based on a deep learning algorithm, extracting a multimodal joint feature vector including frequency domain features, time domain features, and coherence features, and generating a standardized multimodal feature data sequence;

[0011] Based on the standardized multimodal feature data sequence, a graph neural network modeling algorithm is used to identify the user's current cognitive state, motor state or emotional state, and a label sequence containing neural state labels is generated;

[0012] Based on the label sequence L and the user's historical training data, an attention mechanism is used to dynamically generate personalized neural state trend data;

[0013] Based on the personalized neural state trend data and training target parameters, a reinforcement learning algorithm is used to generate intelligent training strategy data including training task type, feedback method parameters and training difficulty level;

[0014] Inputting the intelligent training strategy data into an immersive interactive system to generate a virtual reality or augmented reality scene, implementing visual, auditory and tactile multi-channel interactive training tasks based on the training task type and feedback method parameters, and recording the user's behavioral responses and neural feedback signals during the training process in real time;

[0015] Based on the behavioral response and neural feedback signal, combined with personalized neural state data, dynamically evaluate the user's training effect, adjust the training difficulty level and feedback method parameters, and generate personalized training adjustment data;

[0016] The neural-physiological signal data sequence, standardized multimodal feature data sequence, personalized neural state data, intelligent training strategy data, user behavior response and training adjustment data are uploaded to the cloud, and the graph neural network and reinforcement learning algorithm are optimized based on the cloud-edge collaborative computing architecture, and updated to the local system for subsequent training.

[0017] Furthermore, it is characterized in that the step of jointly preprocessing the neural-physiological signal data sequence includes jointly denoising the time domain artifacts and frequency domain anomalies of the EEG signals, near-infrared signals, electromyographic signals, and heart rate variability signals based on a deep neural network, and outputting the denoised neural-physiological signal sequence.

[0018] Furthermore, it is characterized in that the step of generating the standardized multimodal feature data sequence includes extracting time-frequency features, analyzing the coherence between signals, and calculating the mutual information of the denoised neural-physiological signal sequence, and jointly forming a multimodal joint feature vector and then standardizing it to obtain a standardized multimodal feature data sequence.

[0019] Furthermore, it is characterized in that the step of identifying the user's current neural state based on the standardized multimodal feature data sequence includes adopting a graph neural network algorithm, taking the features of different physiological signals as the nodes of the graph, and the feature correlations as the edges of the graph, establishing a graph structured neural state graph, and outputting a neural state label sequence.

[0020] Furthermore, it is characterized in that the step of generating personalized neural state trend data based on the label sequence and historical training data includes assigning different weights to each label according to the historical training data based on the attention mechanism, and dynamically calculating the personalized neural state trend data.

[0021] Furthermore, it is characterized in that the step of generating intelligent training strategy data based on personalized neural state trend data and training goals includes:

[0022] Construct a training task decision function based on policy gradient to generate a candidate set of training tasks;

[0023] Based on the reinforcement learning algorithm, the optimal task is selected from the candidate set of training tasks, and intelligent training strategy data including training task type, feedback method parameters and training difficulty level are output.

[0024] Furthermore, it is characterized in that the step of inputting the intelligent training strategy data into the immersive interactive system includes generating a virtual reality or augmented reality scene according to the training task type, and outputting a multi-channel interactive training task based on feedback mode parameters, and feeding back training information to the user through vision, hearing and touch.

[0025] Furthermore, it is characterized in that the step of dynamically evaluating the user training effect comprises:

[0026] Based on behavioral response data and neural feedback signals, combined with personalized neural state data, the user's training effect is calculated to generate instant training effect data;

[0027] According to the instant training effect data, the training difficulty level and feedback method parameters are adjusted to generate personalized training adjustment data for subsequent training calls.

[0028] Furthermore, it is characterized in that the step of feeding back the training adjustment data to the reinforcement learning algorithm comprises:

[0029] The parameters of the training task decision function are dynamically modified according to the training adjustment data, and the adaptability and accuracy of the training task are optimized through continuous reinforcement learning.

[0030] Furthermore, it is characterized in that the data uploaded to the cloud includes a neuro-physiological signal data sequence, a standardized multimodal feature data sequence, personalized neural state data, intelligent training strategy data, user behavior response data and training adjustment data;

[0031] Based on the federated learning algorithm, the graph neural network model and the reinforcement learning strategy generation algorithm are optimized to generate optimized model parameters, which are distributed to the local system to update the training strategy and neural state recognition algorithm.

[0032] The following are its main beneficial effects:

[0033] (1) Realize accurate identification based on multimodal neuro-physiological signals and improve the comprehensiveness and accuracy of neuro-state monitoring. The present invention can comprehensively reflect the user's neural activity state during rehabilitation training by synchronously acquiring multiple neuro-physiological signals such as EEG, near infrared, electromyography, heart rate variability, etc., and after joint preprocessing and standardized fusion. Compared with the traditional single signal monitoring method, it effectively improves the accuracy and dynamic adaptability of neural state recognition, and provides a scientific basis for rehabilitation training.

[0034] (2) Dynamically generate personalized neural state trend data based on the attention mechanism and graph neural network to enhance the pertinence and adaptability of training strategies. This invention combines the attention mechanism with the graph neural network algorithm for the first time, dynamically generates personalized neural state trends based on the user's current neural state and historical training data, and uses them as key parameters for the generation of subsequent training strategies, so that training tasks can dynamically adapt to the user's current ability changes, avoiding the problem that fixed training plans cannot adapt to user changes, thereby effectively improving the personalization and scientific nature of rehabilitation training.

[0035] (3) Based on the personalized rehabilitation training strategy of reinforcement learning and the immersive multi-channel interactive system, a complete closed-loop training mechanism is constructed to improve the rehabilitation training effect. The present invention automatically generates training tasks that match the user's neural state through a reinforcement learning algorithm, combines virtual reality or augmented reality to generate immersive interactive training scenes, and adjusts the task difficulty and feedback method in real time according to training feedback to form a dynamic closed-loop training system. Compared with the traditional method of pre-setting fixed training tasks, it can dynamically adjust the training plan according to the user's real-time feedback, effectively improve the rehabilitation training effect and sustainability, and enhance the user's active participation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flowchart of a neurofeedback rehabilitation training method provided in an embodiment of the present application;

[0037] Figure 2 A structural block diagram of a neurofeedback rehabilitation training system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] Example 1: Reference Figure 1 , is a flowchart of a neurofeedback rehabilitation training method provided by an embodiment of the present invention, and the process may at least include steps S100-S700:

[0039] S100, acquiring the user's electroencephalogram (EEG) signals, near-infrared signals, myoelectric signals, and heart rate variability physiological indicators based on a multimodal neuro-physiological signal acquisition device, generating a neuro-physiological signal data sequence including multi-dimensional time series data, and inputting the neuro-physiological signal data sequence into a fusion processing system in real time;

[0040] S200, performing joint preprocessing on the neural-physiological signal data sequence, including noise removal, artifact removal, and abnormal band suppression based on a deep learning algorithm, and extracting a multimodal joint feature vector including frequency domain features, time domain features, and coherence features to generate a standardized multimodal feature data sequence;

[0041] S300, based on the standardized multimodal feature data sequence, using a personalized neural state modeling algorithm to perform real-time recognition of the user's current cognitive state, motor state or emotional state, and obtain personalized neural state data including neural state labels and dynamic change trends;

[0042] S400, based on the personalized neural state data and the training target parameters, a reinforcement learning algorithm is used to dynamically generate a personalized rehabilitation training strategy for the current neural state, and output intelligent training strategy data including training task type, feedback method parameters and training difficulty level;

[0043] S500, inputting the intelligent training strategy data into an immersive interactive system, generating a virtual reality or augmented reality scene, and implementing visual, auditory and tactile multi-channel interactive training tasks for the user based on the training task type and feedback method parameters, while recording the user's behavioral response and neural feedback signals during the training process in real time;

[0044] S600, based on the user behavior response and the neural feedback signal, combined with the personalized neural state data, dynamically evaluate the user's immediate training effect during the training process, and adjust the training difficulty level and feedback method parameters contained in the intelligent training strategy data according to the evaluation result to generate updated personalized training adjustment data;

[0045] S700. Upload the neural-physiological signal data sequence, standardized multimodal feature data sequence, personalized neural state data, intelligent training strategy data, user behavior response and training adjustment data to the cloud. Based on the cloud-edge collaborative computing architecture, analyze the user's long-term training trend and optimize the personalized neural state modeling algorithm and reinforcement learning training strategy generation algorithm, and return the optimized model parameters to the local system for subsequent iterative training.

[0046] Step S100 at least includes steps S110-S130:

[0047] S110, acquiring the user's electroencephalogram signal, near infrared signal, electromyographic signal and heart rate variability signal, and performing synchronous timing calibration on the signals to obtain a raw data sequence of a neuro-physiological signal containing a timestamp label.

[0048] Specifically, based on the neuro-physiological signal acquisition subsystem in the neurofeedback rehabilitation training system, the user's electroencephalogram (EEG) signal, functional near-infrared spectroscopy (fNIRS) signal, electromyography (EMG) signal and heart rate variability (HRV) signal are collected at the same time, and the different types of signals are recorded as

[0049] Furthermore, in order to ensure the consistency of multimodal signals in the time dimension, a unified timestamp T is embedded in the signals. k Perform timing calibration, where T k Represents the time label corresponding to the kth time point, and obtains the original neural-physiological signal data sequence D containing the timestamp k , specifically expressed as:

[0050]

[0051] Where K is the total length of the sampling sequence, are the neuro-physiological signals at the kth time point respectively.

[0052] Furthermore, the data sequence D k As the input data for subsequent synchronous alignment and standardization processing, it ensures that all signal sources are processed collaboratively according to the unified timeline.

[0053] S120, performing inter-channel synchronous registration and amplitude normalization on the original data sequence of the neural-physiological signal to obtain a standardized neural-physiological signal data sequence, wherein the amplitude normalization is performed according to a formula.

[0054] Specifically, based on the original data sequence D containing the timestamp tag obtained in step S110 k , firstly, the timestamp T of the multimodal signal k Perform unified registration to ensure that the four signals of EEG, fNIRS, EMG, and HRV are synchronized at the same time scale and recorded as

[0055] After completing the synchronous registration, each type of signal is further normalized according to the following normalization formula:

[0056]

[0057] in, represents the normalized data of the c-th type signal at the k-th time point, represents the raw data of the c-th signal (where c∈{EEG,fNIRS,EMG,HRV}) after synchronous registration, μ c represents the mean value of the c-th type of signal, calculated at all K time points, σ c Represents the standard deviation of the c-th type signal, calculated based on all K time points.

[0058] Furthermore, the standardized processing of all types of signals is combined to obtain the final standardized neural-physiological signal data sequence

[0059]

[0060] in As the input data of step S130, it is used for subsequent multimodal fusion processing.

[0061] S130, inputting the standardized neural-physiological signal data sequence into a fusion processing system in real time to generate a multimodal neural-physiological signal data set for subsequent feature extraction.

[0062] Specifically, based on the standardized neuro-physiological signal data sequence obtained in step S120 The sequence is input into the multimodal fusion processing unit in the neurofeedback rehabilitation training system in real time, and a complete multimodal data set is constructed in the fusion processing unit.

[0063] The multimodal neural-physiological signal dataset Contains the following:

[0064]

[0065] in,

[0066]

[0067]

[0068] T={T 1 ,T 2 ,...,T K} is a timestamp sequence.

[0069] Furthermore, the multimodal dataset As the input data of step S210, it is used for subsequent noise removal, abnormal band suppression and joint feature extraction processing.

[0070] In addition, the standardized neuro-physiological signal data sequence With time label T k The dynamic correspondence will continue to serve as the time series basis in the subsequent joint feature extraction module to ensure that the multimodal features are processed in a time-series consistent manner.

[0071] Step S200 at least includes steps S210-S230:

[0072] S210, removing noise and suppressing abnormal bands from the multimodal neural-physiological signal dataset, based on a trained self-supervised deep denoising network Clean the signal to obtain the denoised neural-physiological signal sequence

[0073] Specifically, based on the standardized multimodal neural-physiological signal data set obtained in step S130 Through the trained self-supervised deep denoising network De-noise and suppress abnormal bands for each type of signal to obtain a denoised neural-physiological signal sequence

[0074] Where i represents the signal category number, i∈{1,2,3,4}, corresponding to EEG, fNIRS, EMG and HRV signals respectively; j represents the channel number of the signal, for multi-channel signals such as EEG and EMG, j∈{1,2,...,C i}, C i is the number of channels of the i-th signal.

[0075] The denoising process formula is specifically:

[0076]

[0077] Furthermore, the denoised signal It will be used as the input signal of the subsequent joint feature extraction step S220 to ensure the effectiveness and robustness of the extracted features.

[0078] S220, jointly extracting time-frequency features, coherence features and mutual information features from the denoised neural-physiological signal sequence to generate a multimodal joint feature vector F = {f EEG ,f fNIRS ,f EMG ,f HRV}.

[0079] Specifically, based on the denoised neural-physiological signal sequence obtained in step S210 First, the EEG and EMG signals are transformed into the time-frequency domain to extract the frequency domain energy features. Specifically, the short-time Fourier transform (STFT) algorithm is used. Calculate and obtain the frequency domain energy feature f of EEG EEG , and its expression formula is:

[0080]

[0081] Among them, K 1 is the time series length of the EEG signal, C 1 is the number of EEG signal channels.

[0082] Furthermore, for HRV signals According to the RR interval sequence n , extract the mean square error feature f of its changes HRV , the specific calculation formula is:

[0083]

[0084] Among them, N 4 is the number of RR intervals contained in the HRV signal.

[0085] Furthermore, the inter-channel coherence characteristics of EEG, EMG and fNIRS signals were analyzed. The temporal coherence of the cross-channel signal is calculated based on the Pearson correlation coefficient ρ, and the formula is as follows:

[0086]

[0087] Among them, j 1 ,j 2 For different channel numbers, are the means of the corresponding channels, is the standard deviation.

[0088] Finally, the above time-frequency features, coherence features and HRV mean square error features are combined to form a complete multimodal joint feature vector F = {f EEG ,f fNIRS ,f EMG ,f HRV}, used to describe the user's neural state characteristics in the current period for standardization processing in step S230.

[0089] The key innovations of the present invention include: based on the technical architecture of multimodal neural-physiological signal joint time series acquisition and standardized fusion processing, it realizes multi-dimensional, real-time and synchronous user neural state data accurate acquisition and preprocessing, avoiding the data one-sidedness and noise interference problems caused by traditional single signal acquisition.

[0090] S230, performing standardization processing based on the multimodal joint feature vector to generate a standardized multimodal feature data sequence for subsequent neural state recognition.

[0091] Specifically, based on the multimodal joint feature vector F calculated in step S220, each type of feature is normalized according to the normalization standard to eliminate the problem of inconsistent feature dimensions between different signal sources. The normalization process is performed according to the following formula:

[0092]

[0093] in,

[0094] is the standardized result of the i-th category feature,

[0095] f i is the joint eigenvalue of the i-th class,

[0096] are the mean and standard deviation of the i-th feature, estimated based on historical data.

[0097] Furthermore, all standardized features are combined to obtain a standardized multimodal feature data sequence

[0098] The standardized multimodal feature data sequence F norm As the input data of the subsequent step S300 (neural state modeling module), it ensures that the subsequent personalized neural state recognition algorithm based on graph neural network has a unified data foundation.

[0099] Step S300 at least includes steps S310-S330:

[0100] S310, based on the standardized multimodal feature data sequence, using a graph neural network (GNN) to model a neural state graph The nodes Represents the characteristics of each channel, edge Represents the correlation between features, and obtains the neural state label sequence L = {l 1 ,l 2 ,...,l n}.

[0101] Specifically, based on the standardized multimodal feature data sequence obtained in step S230 A graph neural network (GNN) is used to model the graph structure of the multimodal features and construct a neural state graph.

[0102] Among them, the node set represents the standardized characteristics of each signal channel of EEG, fNIRS, EMG, and HRV, that is,

[0103]

[0104] Edge Set Represents the feature correlation between any two nodes, which is defined based on the Pearson correlation coefficient ρ. The formula is:

[0105]

[0106] Among them, μ i , μ j The nodes v i 、v j The mean value, σ i , σ j is the standard deviation.

[0107] Furthermore, the neural network is used to Perform graph embedding learning to obtain the embedding vector h corresponding to each node i , and then aggregate all node embeddings to generate the corresponding neural state label sequence L = {l 1 ,l 2 ,...,l n}, where ln Represents the current neural state label of the user at the nth moment.

[0108] The neural state label sequence L is used as input data of step S320 to participate in the user dynamic neural trend analysis.

[0109] S320, based on the neural state label sequence, combined with the user's historical training data H = {h 1 ,h 2 ,...,h m}, using the attention mechanism to weight the dynamic change trend and generate personalized neural state trend data.

[0110] Specifically, based on the neural state label sequence L obtained in step S310, 1 ,l 2 ,...,l n}, combined with the historical training data sequence H = {h 1 ,h 2 ,...,h m}, through the attention mechanism, the labels at different times are weighted, the individual characteristics and change trends of users are extracted, and personalized neural state trend data T is generated.

[0111] Among them, the attention weight α t The calculation formula is:

[0112]

[0113] Among them, e t For historical data h t The relevant attention score, combined with the current label l t and historical labels, using the following attention function:

[0114] e t =Attn(l t ,h t )=tanh(W 1 l t +W 2 h t +b)

[0115] Among them, W 1 ,W 2 is the weight matrix and b is the bias term.

[0116] Finally, based on the weighted labels, personalized neural state trend data T is generated:

[0117]

[0118] Wherein T reflects the overall neural state change trend of the user in the recent period of time. Said T will be used as the input data of step S330 for personalized neural state fusion processing.

[0119] S330, fusing the personalized neural state trend data with the label sequence to generate personalized neural state data for subsequent training strategy generation module to call.

[0120] Specifically, based on the personalized neural state trend data T obtained in step S320 and the neural state label sequence L obtained in step S310, feature fusion processing is further performed to obtain the personalized neural state data S that is ultimately used to describe the user's current and historical comprehensive state. state .

[0121] The fusion process adopts the combination of linear weighting and nonlinear activation. The specific fusion formula is:

[0122]

[0123] in,

[0124] W T , W L is the linear fusion weight matrix,

[0125] is the historical label mean,

[0126] b s is the bias term,

[0127] σ(·) is a nonlinear activation function, preferably a Sigmoid or Tanh function.

[0128] The final generated personalized neural state data S state As input data for the subsequent step S400 (intelligent training strategy generation module), it participates in the formulation of targeted rehabilitation training tasks.

[0129] Step S400 at least includes steps S410-S430:

[0130] S410, based on personalized neural state data and training target parameters, construct a training task decision function π based on policy gradient θ (a|s), generate a set of candidate training tasks Where s is the current state and a is the action (training task).

[0131] Specifically, based on the personalized neural state data S obtained in step S330 state , combined with the preset training target parameter G = {g 1 ,g 2 ,...,gm}, using the policy gradient reinforcement learning model, establish the training task decision function π θ (a|s), where s = S state As the personalized neural state of the current user, a represents the optional rehabilitation training tasks.

[0132] The decision function π θ (a|s) represents the probability of taking training task a given the current personalized neural state s, specifically:

[0133]

[0134] Among them, f θ (a,s) is the function of the policy network to jointly score action a and state s, θ is the network parameter, It is a set of all candidate training tasks, including different rehabilitation training contents, training forms (such as EEG-based attention training, fNIRS-based cognitive load regulation, etc.), and training difficulty levels.

[0135] Furthermore, the decision function π θ (a|s) based on the current S state Perform multiple sampling with the training target G to dynamically generate a set of candidate training tasks Provide a data basis for subsequent screening and optimization.

[0136] S420, based on the reinforcement learning algorithm, the optimal action is screened for the training task set, and the following strategy optimization formula is adopted:

[0137]

[0138] Among them, R(a,s) is the reward function obtained based on state s after executing task a;

[0139] Specifically, based on the candidate training task set generated in step S410 And the decision function π θ (a|s), for each training task a, after being executed in the simulation environment or the user's historical training, the reward value R(a,s) corresponding to the task is obtained. The reward function R(a,s) is weighted and calculated based on multiple indicators such as the user's neural state improvement, training completion, and behavioral response accuracy in historical training, and is defined as:

[0140] R(a,s)=β 1 ΔS state +β 2 ·Acc(a)+β 3 ·Eff(a)

[0141] in,

[0142] ΔS state Indicates the change in the user's neural state before and after training, which comes from the dynamic update of personalized neural state data.

[0143] Acc(a) represents the completion accuracy of training task a.

[0144] Eff(a) represents the comprehensive efficiency index of the training process.

[0145] β 1 ,β 2 ,β 3 is the preset weight coefficient, which is adjusted according to different rehabilitation goals. 1 It can be 0.25, β 2 It can be 0.55, β 3 It can be 0.2, which is obtained through historical calculation.

[0146] Furthermore, using the above reward function R(a,s), the strategy function π θ (a|s) performs gradient update, and the update formula is:

[0147]

[0148] in,

[0149] α is the learning rate,

[0150] represents the gradient of the policy function with respect to the parameter θ,

[0151] θ t+1 are the updated policy parameters.

[0152] Finally, according to the action probability output after the policy function is optimized, the optimal training task a in the current state is obtained. * , for subsequent output.

[0153] S430. Output intelligent training strategy data including training task type, feedback method parameters and training difficulty level according to the selected optimal training task, for subsequent interactive system call.

[0154] Specifically, based on the optimal training task a screened in step S420 * , which is parsed as the training task type T type , Feedback method parameter P feedback , Training difficulty level D level .

[0155] The training task type T typeIncluding but not limited to attention concentration training, cognitive flexibility training, emotion regulation training, motor function training, etc., the feedback method parameter P feedback Including audio-visual feedback parameters (such as real-time graphics, sound prompts), tactile feedback parameters (such as vibration prompts), etc. Training difficulty level D level Indicates the complexity of the task, which is dynamically adjusted according to the user's neural state fitness, and can take multiple levels such as simple, medium, and difficult.

[0156] Furthermore, the intelligent training strategy data S is comprehensively generated plan , specifically expressed as:

[0157] S plan ={T type ,P feedback ,D level}

[0158] The intelligent training strategy data S plan The final result is outputted for the immersive interaction system in step S500 to call and guide the user to enter a targeted training scenario.

[0159] Step S500 at least includes steps S510-S530:

[0160] S510: Inputting the intelligent training strategy data into a virtual reality or augmented reality generation module to generate an immersive interactive interface including training task scenarios and feedback methods.

[0161] Specifically, based on the intelligent training strategy data S obtained in step S430 plan ={T type ,P feedback ,D level}, where T type Indicates the training task type, P feedback Indicates the feedback mode parameter, D level Indicates the difficulty level of the training.

[0162] The intelligent training strategy data S plan Input to the immersive virtual reality (VR) or augmented reality (AR) generation module, through the training task type T type The driving system automatically loads the immersive training scene C corresponding to the task scene , combined with the feedback mode parameter P feedback Configure a multi-channel feedback output mode, which includes visual feedback, auditory feedback, and tactile feedback, specifically recorded as:

[0163] P multi = {P vision ,P audio ,P tactile}

[0164] Further, according to the training difficulty level D level Set the specific interaction complexity, number of targets, time limit and other parameters in the training scenario, and finally generate an immersive interaction interface C interactive , which includes training task scenarios and multi-channel feedback configurations.

[0165] Finally, the immersive interactive interface C interactive As the basic interface for user interaction in the subsequent step S520, the visualization and interactive presentation of the intelligent training strategy is realized.

[0166] S520: Based on the immersive interactive interface, output visual, auditory and tactile multi-channel feedback content to the user to drive the user to participate in the training task.

[0167] Specifically, based on the immersive interactive interface C generated in step S510 interactive , according to the feedback parameter P multi = {P vision ,P audio ,P tactile}, and output multi-channel interactive feedback content to the user, driving the user to follow the training task T type A response is required.

[0168] Among them, visual feedback P vision Including dynamic images, presentation of training targets in virtual environments; auditory feedback audio Including voice prompts, background sound effects, prompt sounds for mission success or failure; tactile feedback tactile This includes instant tactile stimulation performed through vibration devices or peripherals, forming a multimodal collaborative feedback experience.

[0169] At the same time, based on the training difficulty level D level , dynamically adjust the visual complexity of the output (such as the number of targets), the frequency of audio prompts, and the intensity of tactile prompts to ensure that the training task matches the user's current neural state.

[0170] Furthermore, the system follows the preset time window T window ={t 1 ,t 2 ,...,t k Output feedback in stages, control the task rhythm, enable users to complete training interactions according to a periodic rhythm, and provide synchronization markers for subsequent behavioral data recording.

[0171] S530, during the user training process, real-time recording of user behavior response data B = {b 1 ,b 2 ,...,b n} and the neural feedback signal X f , for subsequent dynamic evaluation module to call.

[0172] Specifically, based on the user's interactive behavior under the guidance of the immersive multi-channel feedback output in step S520, the system records the user's behavior response data B in real time. 1 ,b 2 ,...,b n}, where b i Represents the specific behavioral action data of the user in the i-th interaction, including reaction time, action amplitude, number of interaction objects, etc.

[0173] At the same time, the system uses the neuro-physiological signal acquisition device to synchronously record the neural feedback signal during the training process in real time. f , where X f ={X f,EEG ,X f,fNIRS ,X f,EMG ,X f,HRV}, which are real-time data sequences of user’s feedback signals such as EEG, fNIRS, EMG and HRV during the execution of training tasks.

[0174] Specifically, the neural feedback signal X f The acquisition is synchronized and calibrated according to the following time series:

[0175] X f,k ={X f,EEG,k ,X f,fNIRS,k ,X f,EMG,k ,X f,HRV,k},k=1,2,...,K f

[0176] Where K f represents the total sampling length during training, X f,i,k is the value of the i-th type feedback signal at the k-th time point.

[0177] Furthermore, to ensure that the user behavior response data B and the neural feedback signal X f The system uses a unified time label T k Perform real-time synchronization to form a joint data sequence:

[0178] Z k = {b k ,X f,k ,T k}

[0179] Finally, the joint data sequence Z kIt will serve as input data for the training effect dynamic evaluation module in the subsequent step S600, supporting a comprehensive evaluation mechanism based on behavioral response and neural feedback.

[0180] Step S600 at least includes steps S610-S630:

[0181] S610, based on the user behavior response data and the neural feedback signal, combined with the personalized neural state data, dynamically evaluate the current training effect and generate instant training effect data E = {e 1 ,e 2 ,...,e n}.

[0182] Specifically, based on the user behavior response data B obtained in step S530, 1 ,b 2 ,...,b n} and the neural feedback signal X f ={X f,EEG ,X f,fNIRS ,X f,EMG ,X f,HRV}, combined with the personalized neural state data S generated in step S330 state , dynamically calculate the user's instant training effect during the training process.

[0183] The instant training effect data E={e 1 ,e 2 ,...,e n} Includes the effectiveness, accuracy, and neural response matching of the user's completed tasks in each training cycle, calculated according to the following formula:

[0184] e k =λ 1 ·Acc(b k )+λ 2 NeuResp(X f,k ,S state )+λ 3 ·Eff(b k )

[0185] Among them, e k represents the instantaneous training effect value of the kth training cycle, Acc(b k ) represents the user's behavior accuracy in the kth interaction, NeuResp(X f,k ,S state ) represents the user's neural feedback signal X in this cycle f,k With the target neural state S state The matching degree is measured by cosine similarity, and the specific formula is:

[0186]

[0187] Eff(b k ) represents the execution efficiency of user behavior (such as reaction time), λ 1 ,λ 2 ,λ 3 is the weight coefficient of multi-index fusion. For example, 1 can be 0.5, λ 2 can be 0.3, λ 2 It can be 0.2, which is obtained through historical calculation.

[0188] Furthermore, the instant training effect data E is used as a basis for adjusting the training strategy in the subsequent step S620.

[0189] S620: Analyze the instant training effect data, and adjust the training difficulty level d and feedback method parameter p in the intelligent training strategy data according to a preset adjustment standard to obtain updated training adjustment data.

[0190] Specifically, based on the instant training effect data E obtained in step S610, 1 ,e 2 ,...,e n}, the system performs an overall trend analysis on the effect values ​​of all training cycles and calculates the average training effect and standard deviation σ e , which is used to determine whether the current training task needs to be adjusted. The specific calculation is as follows:

[0191]

[0192] Furthermore, based on the average training effect With the preset effect threshold E target Compare and determine whether it is necessary to adjust the training difficulty level d and the feedback method parameter p. The specific judgment logic is as follows:

[0193] like Then increase the training difficulty level d to d+1, and keep the feedback method parameter p unchanged;

[0194] like Then reduce the training difficulty level d to d-1, and enhance the feedback method parameter p (such as increasing the voice prompt frequency and enhancing visual assistance);

[0195] like The original training difficulty level d and feedback method parameter p are maintained.

[0196] Among them, δ is the adjustment sensitivity threshold, which is determined according to the system preset standard.

[0197] Finally, the generated training difficulty level d after adjustment is new and feedback mode parameter p new The training adjustment data S adjust ={d new ,p new}, as the input of the subsequent S630.

[0198] S630: Feedback the updated training adjustment data to the reinforcement learning algorithm to adjust the training task decision function π θ The parameter θ of (a|s) is used for subsequent strategy optimization.

[0199] Specifically, based on the training adjustment data S obtained in step S620 adjust ={d new ,p new}, as the environmental feedback signal, is transmitted back to the policy gradient-based training task decision function π constructed in step S400 θ (a|s) dynamically updates the parameter θ.

[0200] The training adjustment data participates in the target update of the strategy function through the reinforcement learning feedback mechanism, and dynamically adjusts the strategy parameters according to the following formula:

[0201]

[0202] in,

[0203] θ t is the strategy parameter of the current cycle,

[0204] α is the learning rate for strategy update,

[0205] R new (a,s,S adjust ) is the reward function redefined in combination with the training adjustment data, reflecting the effectiveness and adaptability of the adjusted training task.

[0206] The reward function R new (a,s,S adjust ) is calculated as follows:

[0207] R new (a,s,S adjust )=R(a,s)+γ·[η 1 ·d new +η 2 ·p new ]

[0208] in,

[0209] R(a, s) is the base reward after performing action a, which comes from step S420.

[0210] γ is the adjustment weight factor.

[0211] η 1 , η 2 are the weights for difficulty adjustment and feedback method adjustment respectively. η 1 can be 0.35, η 2 can be 0.65, obtained through historical determination.

[0212] Finally, the system optimizes the training task generation logic based on the updated policy parameter θ t+1 to provide dynamic and personalized adjustment capabilities for subsequent training task recommendations.

[0213] Step S700 includes at least steps S710 - S730:

[0214] S710. Upload the neuro - physiological signal data sequence, standardized multi - modal feature data sequence, personalized neural state data, intelligent training strategy data, user behavior response, and training adjustment data to the cloud computing platform to form a multi - modal training dataset.

[0215] Specifically, based on the user behavior response data B = {b 1 , b 2 ,..., b n} obtained in step S530 and the neuro - feedback signal X f = {X f,EEG , X f,fNIRS , X f,EMG , X f,HRV}, as well as the standardized multi - modal feature data sequence obtained in step S230 Combined with the personalized neural state data S formed in step S330 state , and the intelligent training strategy data S generated in step S430 plan = {T type , P feedback , D level}, and the training adjustment data S output in step S620 adjust = {d new , p new}, are synchronously uploaded to the cloud computing platform to construct a training dataset D that includes cross - stage and cross - modal cloud , specifically expressed as:

[0216] D cloud = {X f , F norm , S state , S plan , B, Sadjust}

[0217] Furthermore, the multimodal training dataset D cloud Contains multi-dimensional dynamic labels as the basic data source for subsequent federated learning optimization models.

[0218] S720: Based on the multimodal training data set, a federated learning algorithm is used to perform global model parameter optimization on the personalized neural state modeling algorithm and the reinforcement learning training strategy generation algorithm to obtain an optimized model parameter θ * ,φ * .

[0219] Specifically, based on the multimodal training data set D uploaded in step S710 cloud , the federated learning algorithm is used to perform distributed multi-center training on the graph neural network (GNN) parameters φ and the reinforcement learning strategy network parameters θ to obtain the global optimal model parameters φ * ,θ * .

[0220] During federated learning, all local fusion processing systems synchronously perform local model training and upload local updates Δθ k ,Δφ k Go to the cloud aggregation center and perform global parameter aggregation according to the following formula:

[0221]

[0222] in:

[0223] K is the number of all edge devices,

[0224] n k is the number of training samples for the kth device,

[0225] is the total number of samples,

[0226] θ k ,φ k are local model parameters,

[0227] Δθ k ,Δφ k is the gradient parameter for local update.

[0228] Furthermore, the final global optimal model parameter θ is obtained through multiple rounds of iterations. * ,φ * , corresponding to:

[0229] θ * : Optimized reinforcement learning training strategy network parameters are used to generate more accurate and personalized training tasks;

[0230] φ * : Optimized graph neural network personalized neural state recognition model parameters for subsequent high-precision state detection.

[0231] The θ * ,φ * As the distribution object of the subsequent step S730.

[0232] The key innovations of the present invention include: innovative use of a graph neural network combined with an attention mechanism to perform a dynamic trend modeling algorithm on the user's neural state, by analyzing the correlation between neural-physiological characteristics and the dynamic changes of historical data, and outputting personalized data reflecting the user's state trend in real time, thereby solving the problem that existing rehabilitation systems cannot dynamically reflect the user's neural state.

[0233] S730, the optimized model parameter θ * ,φ * Distribute to the local fusion processing system to update the graph neural network model and reinforcement learning algorithm parameters to improve the adaptability and accuracy of subsequent training processes.

[0234] Specifically, based on the global optimal model parameter θ obtained by training in step S720 * ,φ * , through the cloud-edge synchronization mechanism, the parameters are distributed to all registered local fusion processing systems E k , so that each user-end device can obtain the latest optimization model synchronously.

[0235] After the parameters are distributed, the local system automatically replaces the original graph neural network model The parameter φ, and the reinforcement learning policy function π θ The parameter θ of (a|s) is updated as:

[0236] φ=φ * ,θ=θ *

[0237] Furthermore, the updated graph neural network model and policy network It will serve as the basic model for subsequent steps S310 and S410, and will continue to be used for new neural state recognition and personalized training task generation, forming a dynamic closed-loop update mechanism to ensure that the training tasks are highly matched and adapted to the user's current state in real time.

[0238] The key innovations of the present invention include: automatically generating personalized rehabilitation training strategies based on reinforcement learning algorithms, and a closed-loop training program linked with a virtual reality / augmented reality immersive system, which can dynamically adjust the difficulty of training tasks and feedback methods according to the user's real-time neural feedback and behavioral responses, breaking through the lack of dynamic adjustment capabilities and lack of precise matching in traditional rehabilitation training programs, and significantly improving the personalization and intelligence level of rehabilitation training.

[0239] The following are its main beneficial effects:

[0240] (1) Realize accurate identification based on multimodal neuro-physiological signals and improve the comprehensiveness and accuracy of neuro-state monitoring. The present invention can comprehensively reflect the user's neural activity state during rehabilitation training by synchronously acquiring multiple neuro-physiological signals such as EEG, near infrared, electromyography, heart rate variability, etc., and after joint preprocessing and standardized fusion. Compared with the traditional single signal monitoring method, it effectively improves the accuracy and dynamic adaptability of neural state recognition, and provides a scientific basis for rehabilitation training.

[0241] (2) Dynamically generate personalized neural state trend data based on the attention mechanism and graph neural network to enhance the pertinence and adaptability of training strategies. This invention combines the attention mechanism with the graph neural network algorithm for the first time, dynamically generates personalized neural state trends based on the user's current neural state and historical training data, and uses them as key parameters for the generation of subsequent training strategies, so that training tasks can dynamically adapt to the user's current ability changes, avoiding the problem that fixed training plans cannot adapt to user changes, thereby effectively improving the personalization and scientific nature of rehabilitation training.

[0242] (3) Based on the personalized rehabilitation training strategy of reinforcement learning and the immersive multi-channel interactive system, a complete closed-loop training mechanism is constructed to improve the rehabilitation training effect. The present invention automatically generates training tasks that match the user's neural state through a reinforcement learning algorithm, combines virtual reality or augmented reality to generate immersive interactive training scenes, and adjusts the task difficulty and feedback method in real time according to training feedback to form a dynamic closed-loop training system. Compared with the traditional method of pre-setting fixed training tasks, it can dynamically adjust the training plan according to the user's real-time feedback, effectively improve the rehabilitation training effect and sustainability, and enhance the user's active participation.

[0243] Embodiment 2: Figure 2 FIG. 2 shows a structural block diagram of a neurofeedback rehabilitation training system according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0244] The multimodal neuro-physiological signal acquisition module 10 is used to synchronously acquire physiological indicators such as brain electrical activity, blood oxygen concentration, myoelectric response, and heart rate fluctuation generated in real time during the user's neurofeedback rehabilitation training through an electroencephalogram (EEG) acquisition device, a near-infrared spectroscopy (fNIRS) detector, an electromyography (EMG) sensor, and a heart rate variability (HRV) acquisition device. The module uses a unified timestamp to perform time series calibration of multi-channel data to form an original neuro-physiological signal data sequence with a time tag, ensuring the consistency and accuracy of multi-source signals in the time dimension, and inputs them into the subsequent processing system in real time.

[0245] The multimodal data fusion and preprocessing module 20 is used to perform synchronous registration, amplitude normalization and noise removal on the neuro-physiological signal data sequence transmitted by the acquisition module 10. Based on the deep learning self-supervised denoising network, this module cleans the artifacts, background noise, motion interference, etc. in multi-source signals such as EEG, near infrared, electromyography and heart rate, retains highly correlated effective information, and extracts standardized multimodal joint feature data sequences. This module also fuses the extracted time-frequency features, coherence features, and mutual information features to generate a multimodal joint feature vector with comprehensive expression capabilities, providing a basis for subsequent personalized neural state recognition.

[0246] The personalized neural state recognition and modeling module 30 is used to establish a neural state graph model with different physiological signal features as nodes and correlation between features as edges based on the multimodal joint feature vector output by the preprocessing module 20, using a graph neural network algorithm, to dynamically recognize the user's cognitive state, attention level, emotional response, etc. during rehabilitation training. This module further combines the user's historical training data and uses the attention mechanism to perform dynamic trend analysis on the recognition results, generating personalized neural state data reflecting the user's continuous neural state changes for subsequent personalized training program generation.

[0247] The intelligent training strategy generation module 40 is used to dynamically formulate personalized training tasks and feedback plans based on the reinforcement learning strategy network according to the personalized neural state data, training target parameters, and rehabilitation training standards. This module uses the policy gradient algorithm to comprehensively consider parameters such as the complexity of the training task, the user's state fitness, and the feedback method to generate intelligent training strategy data including the training task type, feedback method parameters, and the training difficulty level, to ensure that the training task is in line with the user's current neural state and rehabilitation goals.

[0248] The immersive multi-channel interaction module 50 is used to generate a virtual reality (VR) or augmented reality (AR) training scene according to the training strategy provided by the intelligent training strategy generation module 40, and interact with the user in real time through multi-modal channels such as vision, hearing, and touch. This module outputs dynamic interactive elements, voice prompts, tactile feedback, etc. according to the training task requirements to guide the user to complete the task, and at the same time collects the user's behavioral response data and new neural feedback signals in real time during the interaction process to provide basic data support for the dynamic evaluation module.

[0249] The dynamic evaluation and strategy adjustment module 60 is used to analyze the user behavior response data and neural feedback signals obtained by the immersive multi-channel interaction module 50, and dynamically evaluate the current training effect in combination with the personalized neural state data to generate real-time training effect data. Based on the evaluation results, this module dynamically adjusts the difficulty level of the training task and the feedback method parameters in the intelligent training strategy, outputs the updated training adjustment data, and feeds it back to the intelligent training strategy generation module 40 for continuous strategy optimization, forming a dynamic closed loop of training-evaluation-adjustment.

[0250] The cloud-edge collaborative analysis and model optimization module 70 is used to upload the multimodal data collected and generated in the system, including neuro-physiological signal data, standardized features, neural state data, training strategies, behavioral feedback and adjustment data, to the cloud to build a multimodal rehabilitation training data set. Based on the federated learning algorithm, this module globally optimizes the graph neural network model and the reinforcement learning strategy model to form optimized model parameters, and synchronously distributes them to the local fusion processing system to continuously update the model, improve the intelligence and accuracy of the training process, and ensure the long-term and effective operation of the system.

[0251] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A neurofeedback rehabilitation training method, characterized in that: The method comprises: Acquire the user's physiological indicators based on a multimodal neuro-physiological signal acquisition device and generate a neuro-physiological signal data sequence; Performing joint preprocessing on the neural-physiological signal data sequence to generate a standardized multimodal feature data sequence; Based on personalized neural state trend data and training target parameters, a reinforcement learning algorithm is used to generate intelligent training strategy data; Inputting the intelligent training strategy data into an immersive interactive system to generate a virtual reality and / or augmented reality scene, and recording the user's behavioral response and neural feedback signal during the training process in real time; Based on the behavioral response and neural feedback signal, combined with personalized neural state data, dynamically evaluate the user's training effect, adjust the training difficulty level and feedback method parameters, and generate personalized training adjustment data; The neural-physiological signal data sequence, standardized multimodal feature data sequence, personalized neural state data, intelligent training strategy data, user behavior response and training adjustment data are uploaded to the cloud, and the graph neural network and reinforcement learning algorithm are optimized based on the cloud-edge collaborative computing architecture, and updated to the local system for subsequent training.

2. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of jointly preprocessing the neural-physiological signal data sequence includes jointly denoising the time domain artifacts and frequency domain anomalies of the EEG signals, near-infrared signals, electromyographic signals, and heart rate variability signals based on a deep neural network, and outputting the denoised neural-physiological signal sequence.

3. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of generating the standardized multimodal feature data sequence includes extracting time-frequency features, analyzing the coherence between signals, and calculating mutual information of the denoised neural-physiological signal sequence, and jointly forming a multimodal joint feature vector and then standardizing it to obtain a standardized multimodal feature data sequence.

4. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of identifying the user's current neural state based on a standardized multimodal feature data sequence includes adopting a graph neural network algorithm, taking the features of different physiological signals as nodes of the graph, and feature correlations as edges of the graph, establishing a graph-structured neural state graph, and outputting a neural state label sequence.

5. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of generating personalized neural state trend data based on the label sequence and historical training data includes assigning different weights to each label according to the historical training data based on the attention mechanism, and dynamically calculating the personalized neural state trend data.

6. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of generating intelligent training strategy data based on personalized neural state trend data and training goals includes: Construct a training task decision function based on policy gradient to generate a candidate set of training tasks; Based on the reinforcement learning algorithm, the optimal task is selected from the candidate set of training tasks, and intelligent training strategy data including training task type, feedback method parameters and training difficulty level are output.

7. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of inputting the intelligent training strategy data into the immersive interactive system includes generating a virtual reality or augmented reality scene according to the training task type, and outputting a multi-channel interactive training task based on feedback mode parameters, and feeding back training information to the user through vision, hearing, and touch.

8. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of dynamically evaluating the user training effect includes: Based on behavioral response data and neural feedback signals, combined with personalized neural state data, the user's training effect is calculated to generate instant training effect data; According to the instant training effect data, the training difficulty level and feedback method parameters are adjusted to generate personalized training adjustment data for subsequent training calls.

9. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The step of feeding back the training adjustment data to the reinforcement learning algorithm includes: The parameters of the training task decision function are dynamically modified according to the training adjustment data, and the adaptability and accuracy of the training task are optimized through continuous reinforcement learning.

10. The neurofeedback rehabilitation training method according to claim 1, characterized in that: The data uploaded to the cloud includes neuro-physiological signal data sequences, standardized multimodal feature data sequences, personalized neural state data, intelligent training strategy data, user behavior response data and training adjustment data; Based on the federated learning algorithm, the graph neural network model and the reinforcement learning strategy generation algorithm are optimized to generate optimized model parameters, which are distributed to the local system to update the training strategy and neural state recognition algorithm.

Images