Intelligent diopter training system and method based on multi-modal physiological feedback and reinforcement learning algorithm

Abstract

The invention discloses an intelligent diopter training system and method based on a multi-modal physiological feedback and reinforcement learning algorithm. The intelligent diopter training system comprises a multi-modal physiological data acquisition module, a central processing PC, a user terminal and a cloud data platform. Visual behavior data, ciliary muscle state data and autonomic nerve reaction data of a user are collected in real time through a multi-modal physiological data collection module, multi-modal physiological data fusion and feature extraction are carried out, and a refraction state evaluation model is used for calculating a comprehensive score of a refraction state grade, visual fatigue and adjustment sensitivity of the current user. A personalized training parameter set of the next training period is dynamically decided and generated based on a reinforcement learning algorithm, and the user terminal presents visual training stimulation according to the current personalized training parameter set. The cloud data platform stores the data and generates a visual training effect trend chart, a diopter change prediction report and personalized training suggestions to complete training scheme personalization, training means intellectualization and feedback evaluation comprehensiveness.

Description

Technical Field

[0001] This invention relates to the field of refractive error training technology, and in particular to an intelligent refractive error training system and method based on multimodal physiological feedback and reinforcement learning algorithms. Background Technology

[0002] With the deepening development of the digital age, visual health problems are becoming increasingly prominent worldwide, especially the prevalence of refractive errors (including myopia, hyperopia, and astigmatism), which continues to rise and shows a trend towards affecting younger people. Traditional refractive correction methods mainly rely on optical instruments, such as eyeglasses and contact lenses, as well as surgical procedures to change corneal curvature. However, these methods are all passive corrections or one-time interventions, failing to train and improve the eye's own accommodative function at its root. Among existing technologies, vision training, as a non-invasive auxiliary correction method, has been applied to a certain extent. Its basic principle is to stimulate eye muscles and enhance accommodative flexibility through a series of preset eye training movements. However, such training programs generally have significant limitations: 1. Rigid training programs lack personalization: Most training systems use a uniform training mode and intensity, which cannot be dynamically adjusted according to the user's real-time physiological feedback (such as accommodative lag, pupil dynamic changes, and binocular coordination); 2. Single feedback mechanism and incomplete evaluation: Traditional training effect evaluation relies heavily on periodic visual acuity chart checks, which cannot achieve continuous and accurate monitoring of key physiological indicators during the training process. The lack of multi-dimensional data support makes the optimization of training programs lack a scientific basis; 3. Low user dependence makes it difficult to persist in the long term: boring and repetitive training content can easily cause users, especially children and adolescents, to develop resistance, resulting in poor training compliance and difficulty in achieving the expected long-term corrective effect.

[0003] In recent years, the development of biosensing technology, artificial intelligence algorithms, and wearable devices has provided new possibilities for solving the aforementioned problems. For example, Chinese patent CN120549746A discloses a dynamic visual perception training device with a multi-focal adjustable lens, which adjusts the lens focus through a high-precision stepper motor to achieve visual perception training. Another example is Chinese patent CN120570774A, which discloses a visual training system and method. While this method dynamically adjusts the training scheme by acquiring basic data such as the user's intraocular pressure, refractive error, astigmatism axis, and pupillary distance, it still only achieves lens focusing, resulting in limited effectiveness. Therefore, how to effectively utilize artificial intelligence, comprehensively collect users' physiological training data using wearable sensors, and generate visualized training images using multimodal fusion and adaptive algorithms to achieve personalized training schemes for different users is an urgent technical problem that needs to be solved. Summary of the Invention

[0004] In view of this, the present invention proposes an intelligent refractive error training system and method based on multimodal physiological feedback and reinforcement learning algorithms to solve the problems of lack of personalization, single training methods and incomplete training feedback evaluation in existing training programs.

[0005] The technical solution of the present invention is implemented as follows: On the one hand, the present invention provides an intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithms, which includes a multimodal physiological data acquisition module, a central processing PC, a user terminal and a cloud data platform;

[0006] The multimodal physiological data acquisition module is used to collect users' visual behavior data, ciliary muscle state data, and autonomic nerve response data in real time and send them to the central processing PC.

[0007] The central processing PC is connected to the multimodal physiological data acquisition module via USB port for receiving data and performing multimodal physiological data fusion and feature extraction. It uses a refractive state assessment model to calculate and output a comprehensive score of the current user's refractive state level, visual fatigue, and accommodation sensitivity. Based on a reinforcement learning algorithm, it dynamically makes decisions to generate a personalized training parameter set for the next training cycle and sends it to the user terminal.

[0008] The user terminal is connected to the USB port of the central processing PC to provide a human-computer interaction interface, receive the current personalized training parameter set, and present visual training stimuli.

[0009] The cloud data platform is connected to the central processing PC via the Internet and is used to store all users' historical training data, physiological feedback data and algorithm models, analyze and generate visualized training effect trend charts, refractive error change prediction reports and personalized training suggestions.

[0010] Preferably, the multimodal physiological data acquisition module includes: an eye tracker and a wearable physiological sensor;

[0011] The eye tracker is used to acquire visual behavior data, including pupil diameter, blink frequency, fixation stability index and eye movement trajectory, and to extract ciliary muscle state data and ciliary muscle tension.

[0012] The wearable physiological sensor includes an ECG & GSR sensor and an EEG electroencephalogram sensor, used to acquire autonomic nervous system response data;

[0013] The ECG & GSR sensors are used to simultaneously collect heart rate variability (HRV) and skin conductance response (GSR) data.

[0014] The EEG sensor is used to simultaneously acquire the alpha and theta wave energies of an EEG.

[0015] The central processing PC includes: a data fusion and feature extraction unit, a user status evaluation unit, and an adaptive training parameter decision unit;

[0016] The data fusion and feature extraction unit is used to perform time alignment and noise reduction on the received multimodal physiological data, and to extract features related to visual fatigue and regulatory ability based on time domain, frequency domain and nonlinear dynamics methods.

[0017] The user status assessment unit has a built-in refractive status assessment model trained on a large number of samples, which is used to output the user's current refractive status level, visual fatigue and accommodation sensitivity comprehensive score in real time based on the extracted features.

[0018] The adaptive training parameter decision unit is used to dynamically generate a personalized training parameter set for the next training cycle based on the comprehensive score output by the user status evaluation unit and the reinforcement learning algorithm.

[0019] The user terminal includes: an interaction logic control unit and a display unit;

[0020] The interactive logic control unit is used to control the display unit to operate according to the personalized training parameter set generated by the adaptive training parameter decision unit;

[0021] The display unit is used to present dynamically changing visual stimulus patterns, including but not limited to: moiré fringes, Gaussian sine gratings, vector graphics, three-dimensional stereograms, and random dot stereograms.

[0022] The personalized training parameter set includes at least: the spatial frequency of the visual stimulus pattern, the speed and magnitude of the contrast change of the pattern, the color combination and saturation of the pattern, the duration and interval of the training task, and the switching frequency and depth range of the three-dimensional depth of field.

[0023] On the other hand, the present invention provides an intelligent refractive power training method based on multimodal physiological feedback and reinforcement learning algorithms; comprising the following steps:

[0024] S1. The user starts training through the user terminal, the system collects the user's basic refractive data and loads the initial training parameters;

[0025] S2. The user terminal presents visual stimuli based on the current training parameter set, and the multimodal physiological data acquisition module begins to synchronously collect the user's visual behavior data, ciliary muscle state data, and autonomic nerve response data.

[0026] S3. The central processing PC performs multimodal physiological data fusion and feature extraction, and inputs it into the refractive state assessment model to calculate the user's current refractive state, visual fatigue and accommodation sensitivity in real time.

[0027] S4. The adaptive training parameter decision unit calculates and generates a personalized training parameter set for the next training cycle based on the comprehensive score using a reinforcement learning algorithm.

[0028] S5. The user terminal receives and applies the new personalized training parameter set, updates the displayed visual stimuli, and repeats steps S2 to S5 to form a closed-loop real-time adaptive training process.

[0029] S6. When the preset training time is reached or the system detects that the user's fatigue level exceeds the safety threshold, the training is terminated, and all data from this training is uploaded to the cloud data platform for storage and analysis.

[0030] The S7 cloud data platform uses long-term training data and the ARIMA model to generate visualized training effect trend charts, diopter change prediction reports, and personalized training suggestions for users to view.

[0031] The intelligent refractive error training system and method based on multimodal physiological feedback and reinforcement learning algorithms of this invention have the following advantages over existing technologies:

[0032] (1) The intelligent refractive power training scheme was realized by using multimodal physiological data fusion and feature extraction technology, refractive state assessment model, and reinforcement learning algorithm dynamic decision-making technology;

[0033] (2) By using an eye tracker and wearable physiological sensors to acquire visual behavior data, ciliary muscle status data and autonomic nerve response data of users during refractive error training in real time, the real-time and comprehensive refractive error training feedback was achieved.

[0034] (3) Using cloud data platforms based on long-term training data, ARIMA models are used to generate visualized training effect trend charts, diopter change prediction reports, and personalized training suggestions. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a structural diagram of the intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithm of the present invention;

[0037] Figure 2 This is a flowchart illustrating the workflow of the intelligent refractive error training method based on multimodal physiological feedback and reinforcement learning algorithms of the present invention.

[0038] Figure 3 This is a flowchart illustrating the reinforcement learning algorithm in the intelligent refractive power training method based on multimodal physiological feedback and reinforcement learning algorithm of the present invention.

[0039] Figure 4 The image shows the application effect of the intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithm of the present invention. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] like Figure 1 As shown, the intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithm of the present invention includes a multimodal physiological data acquisition module, a central processing PC, a user terminal and a cloud data platform;

[0043] The multimodal physiological data acquisition module is used to collect users' visual behavior data, ciliary muscle state data, and autonomic nerve response data in real time and send them to the central processing PC.

[0044] The central processing PC is connected to the multimodal physiological data acquisition module via USB port for receiving data and performing multimodal physiological data fusion and feature extraction. It uses a refractive state assessment model to calculate and output a comprehensive score of the current user's refractive state level, visual fatigue, and accommodation sensitivity. Based on a reinforcement learning algorithm, it dynamically makes decisions to generate a personalized training parameter set for the next training cycle and sends it to the user terminal.

[0045] The user terminal is connected to the USB port of the central processing PC to provide a human-computer interaction interface, receive the current personalized training parameter set, and present visual training stimuli.

[0046] The cloud data platform is connected to the central processing PC via the Internet and is used to store all users' historical training data, physiological feedback data and algorithm models, analyze and generate visualized training effect trend charts, refractive error change prediction reports and personalized training suggestions.

[0047] This system employs a multimodal physiological data acquisition module, a central processing PC, user terminals, and a cloud data platform to implement an intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithms. Utilizing multimodal physiological data fusion and feature extraction technologies, a refractive state assessment model, and reinforcement learning algorithm dynamic decision-making technology, the system enables real-time acquisition of visual behavior, ciliary muscle state, and autonomic nervous system response data during user training. It provides intelligent training plan decision-making, visualized reports, and personalized suggestions. Through design, simulation, and verification, a modular product is formed, allowing for rapid portability across different platforms and accelerating the product development process.

[0048] The multimodal physiological data acquisition module includes an eye tracker and a wearable physiological sensor.

[0049] The eye tracker is used to acquire visual behavior data, including pupil diameter, blink frequency, fixation stability index and eye movement trajectory, and to extract ciliary muscle state data and ciliary muscle tension.

[0050] The wearable physiological sensor includes an ECG & GSR sensor and an EEG electroencephalogram sensor, used to acquire autonomic nervous system response data;

[0051] The ECG & GSR sensors are used to simultaneously collect heart rate variability (HRV) and skin conductance response (GSR) data.

[0052] The EEG sensor is used to simultaneously acquire the alpha and theta wave energies of an EEG.

[0053] The central processing PC includes: a data fusion and feature extraction unit, a user status evaluation unit, and an adaptive training parameter decision unit;

[0054] The data fusion and feature extraction unit is used to perform time alignment and noise reduction on the received multimodal physiological data, and to extract features related to visual fatigue and regulatory ability based on time domain, frequency domain and nonlinear dynamics methods.

[0055] The user status assessment unit has a built-in refractive status assessment model trained on a large number of samples, which is used to output the user's current refractive status level, visual fatigue and accommodation sensitivity comprehensive score in real time based on the extracted features.

[0056] The adaptive training parameter decision unit is used to dynamically generate a personalized training parameter set for the next training cycle based on the comprehensive score output by the user status evaluation unit and the reinforcement learning algorithm.

[0057] The user terminal includes: an interaction logic control unit and a display unit;

[0058] The interactive logic control unit is used to control the display unit to operate according to the personalized training parameter set generated by the adaptive training parameter decision unit;

[0059] The display unit is used to present dynamically changing visual stimulus patterns, including but not limited to: moiré fringes, Gaussian sine gratings, vector graphics, three-dimensional stereograms, and random dot stereograms.

[0060] The personalized training parameter set includes at least: the spatial frequency of the visual stimulus pattern, the speed and magnitude of the contrast change of the pattern, the color combination and saturation of the pattern, the duration and interval of the training task, and the switching frequency and depth range of the three-dimensional depth of field.

[0061] The eye tracker is a Tobii Pro Spectrum type eye tracker, the ECG & GSR sensor is a Shimmer3 GSR+ECG sensor module, the EEG sensor is a BioSemi ActiveTwo sensor, the central processing PC is a Microsoft Surface Pro 11 computer, and the user terminal is a Lenovo 75-inch touch conference all-in-one machine.

[0062] The intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithms in this embodiment employs a multimodal physiological data acquisition module, a central processing PC, a user terminal, and a cloud data platform. This system utilizes multimodal physiological data fusion and feature extraction technology, a refractive state assessment model, and reinforcement learning algorithm dynamic decision-making technology to achieve real-time acquisition of visual behavior, ciliary muscle state, and autonomic nervous system response data during user training. It enables intelligent decision-making for training plans, visualized reports, and personalized suggestions. Through design, simulation, and verification, a modular product is formed, enabling rapid portability between different platforms and accelerating the product development process.

[0063] Example 2

[0064] A smart refractive error training method based on multimodal physiological feedback and reinforcement learning algorithms is provided. This method employs the smart refractive error training system based on multimodal physiological feedback and reinforcement learning algorithms as described in Example 1, and includes the following steps:

[0065] S1. The user starts training through the user terminal, the system collects the user's basic refractive data and loads the initial training parameters;

[0066] S2. The user terminal presents visual stimuli based on the current training parameter set, and the multimodal physiological data acquisition module begins to synchronously collect the user's visual behavior data, ciliary muscle state data, and autonomic nerve response data.

[0067] S3. The central processing PC performs multimodal physiological data fusion and feature extraction, and inputs it into the refractive state assessment model to calculate the user's current refractive state, visual fatigue and accommodation sensitivity in real time.

[0068] S4. The adaptive training parameter decision unit calculates and generates a personalized training parameter set for the next training cycle based on the comprehensive score using a reinforcement learning algorithm.

[0069] S5. The user terminal receives and applies the new personalized training parameter set, updates the displayed visual stimuli, and repeats steps S2 to S5 to form a closed-loop real-time adaptive training process.

[0070] S6. When the preset training time is reached or the system detects that the user's fatigue level exceeds the safety threshold, the training is terminated, and all data from this training is uploaded to the cloud data platform for storage and analysis.

[0071] The S7 cloud data platform uses long-term training data and the ARIMA model to generate visualized training effect trend charts, diopter change prediction reports, and personalized training suggestions for users to view.

[0072] The multimodal physiological data fusion and feature extraction includes: first, performing timestamp alignment, filtering and denoising, and outlier removal on the physiological data of each modality; second, extracting key features from the preprocessed data, including the fluctuation entropy of pupil diameter, the power spectral density of blink frequency, the fractal dimension of eye movement trajectory, the low-frequency / high-frequency power ratio of HRV, the rising slope of GSR, and the energy ratio of α wave to θ wave of EEG.

[0073] For example: denoising based on the Butterworth filter design uses the following transfer function:

[0074]

[0075] Among them, s c The cutoff frequency is n, and the filter order is n.

[0076] Outlier removal is based on an improved Z-score method, when data point x... i satisfy

[0077] When it is considered an outlier, is the median, and MAD is the absolute deviation from the median;

[0078] The fluctuation entropy of pupil diameter is based on the multi-scale entropy formula:

[0079]

[0080] in, The reconstructed sequence of the original signal, scale factor ;

[0081] Here, SampleEn is the sample entropy algorithm, specifically:

[0082]

[0083] Among them, A m (r) and B m (r) represents the number of template matches in the m+1 dimension and the m dimension, respectively;

[0084] The power spectral density of blink frequency is calculated based on the Welch method, specifically:

[0085]

[0086] in, It involves signal segmentation; FFT stands for Fast Fourier Transform, which extracts the main frequency. ;

[0087] The fractal dimension of the eye-tracking trajectory is calculated using the Higuchi method, specifically:

[0088]

[0089] Here, the fractal dimension D is right The slope of the linear regression;

[0090] The low-frequency / high-frequency power ratio of HRV is calculated by integrating the power of the low-frequency band (LF: 0.04-0.15Hz) and the high-frequency band (HF: 0.15-0.4Hz) from the HRV power spectrum.

[0091] The rising slope of GSR is calculated as the linear regression slope of the GSR signal within 1-5 seconds after the stimulus begins.

[0092] The energy ratio of α-wave to θ-wave in EEG is calculated by extracting the energy of α-wave (8-13Hz) and θ-wave (4-8Hz) from the EEG power spectrum.

[0093] The refractive state assessment model is based on a multi-head Transformer neural network input model with an attention mechanism, a classification probability based on softmax, and a tension score output model based on sigmoid.

[0094] The comprehensive score of refractive state, visual fatigue, and accommodative sensitivity includes: the probability of refractive state, visual fatigue, and accommodative sensitivity, and a visual system tension value.

[0095] For example, the specific implementation method of the refractive state assessment model is as follows:

[0096] 1. Input layer: Receives a standardized sequence of multimodal physiological feature vectors. ,in, (d is the feature dimension, t is the time series length), standardization uses the Z-score method: specifically: , where μ and σ are the mean and variance of the feature, respectively;

[0097] 2. Position Encoding Layer: Sine and cosine functions are used to add positional information to the input sequence. The position encoding formula is as follows:

[0098] , ,

[0099] Where pos is the position index and i is the dimension index;

[0100] 3. Multi-head self-attention layer: Consists of H attention heads, each independently computing the query (Q), key (K), and value (V) matrix.

[0101] , ,

[0102]

[0103] in, It is a trainable weight matrix. It represents the dimension of each head, and the softmax function is applied along the row direction;

[0104] 4. Attention Output Fusion: The outputs of all heads are concatenated and subjected to a linear transformation.

[0105] ,

[0106] in, It is the output weight matrix;

[0107] 5. Feedforward Neural Network Layer: A two-layer fully connected network is applied at each location:

[0108] ,

[0109] in, , , b1 and b2 are bias terms, and the activation function is ReLU;

[0110] 6. Output Layer: The final output is then fed into the softmax layer and regression layer after global average pooling.

[0111] ,

[0112] in, It is the probability of refractive state classification (C is the number of categories, including myopia, hyperopia, astigmatism, etc.). σ is the visual system tension score, σ is the sigmoid function, and Wc, Wr, bc, and br are trainable parameters.

[0113] The reinforcement learning algorithm includes the following steps:

[0114] N1. Encode the comprehensive score and feature vector output by the user state evaluation unit into the state vector required by the reinforcement learning algorithm;

[0115] N2. Input the state vector into the policy network and value network of the PPO algorithm. The network outputs a probability distribution in the action space and the state value of the action.

[0116] N3. A specific action is obtained by sampling according to the probability distribution. This action is represented as an instruction to increase or decrease each parameter in the current training parameter set.

[0117] N4. In the next evaluation cycle, the immediate reward for this action is calculated based on the change in the user's status and the reward function.

[0118] N5. Using the collected state-action-reward sequences, periodically update the policy network and value network parameters of the PPO algorithm to optimize long-term cumulative rewards;

[0119] For example, the specific implementation of reinforcement learning algorithms is as follows:

[0120] 1. State Encoding: Encode the comprehensive score and feature vector output by the user state assessment unit into a state vector. This includes refractive state probability, visual fatigue score, and accommodative sensitivity score;

[0121] 2. Policy Network Inference: This involves considering the state vector... Inputs are fed into the PPO policy network and value network, and the output is action distribution parameters. , and action value Then sample and perform actions. , ;

[0122] The specific strategy network structure is as follows:

[0123] ,

[0124] ,

[0125] ,

[0126] in, These are the policy network parameters, ReLU is the activation function, and the softplus function ensures that the standard deviation is positive;

[0127] The specific value network structure is as follows:

[0128]

[0129] in, These are value network parameters;

[0130] 3. Action execution: This involves executing the action. Decode into training parameter adjustment instructions, for example:

[0131] ,

[0132] ,

[0133] in, and It is the preset step size;

[0134] 4. Reward Calculation: In the next evaluation period, rewards are calculated based on changes in user status. The reward function is as follows:

[0135]

[0136] in, These are weighting coefficients. It's about adjusting the sensitivity. It is visual fatigue, I 完成 Indicates the completion rate of the training task. Indicates the amount of change between adjacent periods;

[0137] 5. Policy Update: The policy network and value network are updated using the PPO algorithm, specifically as follows:

[0138] Define the PPO objective function: A pruning objective function is used to ensure stable policy updates. The objective function is as follows:

[0139]

[0140] in, It's a probability ratio. It is a policy network. It is the estimation of the advantage function. These are the clipping parameters;

[0141] Dominance estimation: Calculated using generalized dominance estimation (GAE) :

[0142]

[0143] in, , It is a discount factor. These are GAE parameters;

[0144] Training process: After collecting samples at every K=2048 time steps, the Adam optimizer is used to minimize the total loss function:

[0145]

[0146] in, It is a value loss function. It is the policy entropy regularization term, and c1 and c2 are coefficients.

[0147] like Figures 2 to 4 As shown, the user terminal loads initial training parameters and presents visual stimuli. The multimodal physiological data acquisition module begins to synchronously collect data on the user's visual behavior, ciliary muscle state, and autonomic nervous system response. The central processing PC performs multimodal physiological data fusion and feature extraction, refractive state assessment, and reinforcement learning to generate a personalized training parameter set. This forms a closed-loop real-time adaptive training process to complete the training. The cloud data platform uses the ARIMA model to generate a visualized training effect trend chart, a refractive error change prediction report, and personalized training suggestions for the user to review.

[0148] The intelligent refractive error training system and method based on multimodal physiological feedback and reinforcement learning algorithms in this embodiment adopts a modular design approach. It uses a multimodal physiological data acquisition module, a central processing PC, a user terminal, and a cloud data platform to realize an intelligent refractive error training system based on multimodal physiological feedback and reinforcement learning algorithms. By utilizing multimodal physiological data fusion and feature extraction technology, refractive state assessment model, and reinforcement learning algorithm dynamic decision-making technology, it realizes real-time acquisition of visual behavior, ciliary muscle state, and autonomic nerve response data during user training, intelligent decision-making of training plans, visualization of reports, and personalized suggestions.

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent diopter training system based on multi-modal physiological feedback and reinforcement learning algorithm, mainly comprising: A multi-modal physiological data acquisition module, a central processing PC, a user terminal and a cloud data platform; The multi-modal physiological data acquisition module is used for real-time acquisition of visual behavior data, ciliary muscle state data and autonomic nervous response data of a user and sending to the central processing PC; The central processing PC is in communication connection with the multi-modal physiological data acquisition module USB port, used for receiving data and carrying out multi-modal physiological data fusion and feature extraction, using a refractive state evaluation model to calculate output of a current user refractive state grade, visual fatigue degree and accommodation sensitivity comprehensive score, and based on a reinforcement learning algorithm to dynamically decide generation of a personalized training parameter set of a next training period and sending to the user terminal; The user terminal is in communication connection with the central processing PC USB port, used for providing a man-machine interaction interface, receiving a current personalized training parameter set and presenting visual training stimulation; The cloud data platform is in internet communication connection with the central processing PC, used for storing all user historical training data, physiological feedback data and algorithm models, analyzing and generating visual training effect trend graphs, refractive power change prediction reports and personalized training suggestions.

2. The intelligent refractive power training system based on multi-modal physiological feedback and a reinforcement learning algorithm according to claim 1, characterized in that: The multi-modal physiological data acquisition module comprises an eye tracker and a wearable physiological sensor; The eye tracker is used for obtaining visual behavior data, including pupil diameter, blink frequency, fixation stability indicators and eye movement trajectories, and extracting ciliary muscle state data, ciliary muscle tension; The wearable physiological sensor comprises an ECG&GSR sensor and an EEG electroencephalogram sensor, used for obtaining autonomic nervous response data; The ECG&GSR sensor is used for synchronous acquisition of heart rate variability HRV and galvanic skin response GSR data; The EEG electroencephalogram sensor is used for synchronous acquisition of α wave and θ wave energy of electroencephalogram EEG.

3. The intelligent refractive power training system based on multi-modal physiological feedback and a reinforcement learning algorithm according to claim 1, characterized in that: The central processing PC comprises a data fusion and feature extraction unit, a user state evaluation unit and an adaptive training parameter decision unit; The data fusion and feature extraction unit is used for time alignment, noise reduction processing of the received multi-modal physiological data, and extraction of features related to visual fatigue and accommodation ability based on time domain, frequency domain and nonlinear dynamics methods; The user state evaluation unit has a refractive state evaluation model trained by a large number of samples, used for real-time output of a current refractive state grade, visual fatigue degree and accommodation sensitivity comprehensive score of a user according to the extracted features; The adaptive training parameter decision unit is used for dynamic decision generation of a personalized training parameter set of a next training period based on a reinforcement learning algorithm according to the comprehensive score output by the user state evaluation unit.

4. The intelligent refractive power training system based on multi-modal physiological feedback and a reinforcement learning algorithm according to claim 1, characterized in that: The user terminal comprises an interaction logic control unit and a display unit; The interaction logic control unit is used to control the display unit to run according to the personalized training parameter set generated by the adaptive training parameter decision unit; The display unit is used to present a dynamically changing visual stimulation pattern, including but not limited to: Moire fringes, Gaussian sinusoidal grating, vector graphics, three-dimensional graphics and random point stereograms.

5. The intelligent diopter training system based on multi-modal physiological feedback and reinforcement learning algorithm according to claim 4, wherein: The personalized training parameter set at least includes: spatial frequency of visual stimulation pattern, contrast change speed and amplitude of pattern, color combination and saturation of pattern, duration and intermittent period of training task, and switching frequency and depth range of three-dimensional depth of field.

6. The method of claim 1-5, wherein the method is based on multi-modal physiological feedback and reinforcement learning algorithm. Comprising the following steps: S1. The user starts training through the user terminal, the system collects the user's basic refraction data, and loads the initial training parameters; S2. The user terminal presents visual stimulation according to the current training parameter set, and the multi-modal physiological data acquisition module starts to synchronously collect the user's visual behavior data, ciliary muscle state data and autonomic nervous response data; S3. The central processing PC performs multi-modal physiological data fusion and feature extraction, and inputs into the refraction state evaluation model to calculate the user's current refraction state, visual fatigue degree and accommodation sensitivity in real time; S4. The adaptive training parameter decision unit calculates and generates a personalized training parameter set for the next training period according to the comprehensive score using the reinforcement learning algorithm; S5. The user terminal receives and applies the new personalized training parameter set, updates the displayed visual stimulation, and repeats steps S2 to S5 to form a closed-loop real-time adaptive training process; S6. When the preset training time is reached or the system detects that the user's fatigue degree exceeds the safety threshold, the training is terminated, and all data of this training are uploaded to the cloud data platform for storage and analysis to form a training report; S7. The cloud data platform generates visual training effect trend chart, refraction degree change prediction report and personalized training suggestions based on long-term training data using ARIMA model for users to consult.

7. The method of claim 6, wherein the method further comprises: The step S3 specifically comprises: The multi-modal physiological data fusion and feature extraction includes: first, time stamp alignment, filtering and denoising and outlier rejection of physiological data of each modality; second, extracting key features from preprocessed data, including fluctuation entropy of pupil diameter, power spectral density of blink frequency, fractal dimension of eye movement trajectory, low / high frequency power ratio of HRV, rising slope of GSR and energy ratio of alpha and theta waves of EEG; The refraction state evaluation model is a multi-head Transformer neural network input model based on attention mechanism, a classification probability output model based on softmax and a tension score output model based on sigmoid; The comprehensive score of refraction state, visual fatigue degree and accommodation sensitivity includes: probability of refraction state, visual fatigue degree and accommodation sensitivity and a visual system tension value. 8.The smart diopter training method based on multi-modal physiological feedback and reinforcement learning algorithm of claim 6, wherein, The step S4 specifically comprises: The reinforcement learning algorithm comprises the following steps: N1. encode the comprehensive score and feature vector output by the user state evaluation unit into a state vector required by the reinforcement learning algorithm; N2. input the state vector into the policy network and value network of the PPO algorithm, and the network outputs a probability distribution on the action space and the state value of the action; N3. sample a specific action according to the probability distribution, which is expressed as an increase or decrease instruction for each parameter in the current training parameter set; N4. in the next evaluation period, according to the change of the user state, calculate the immediate reward obtained by the action according to the reward function; N5. periodically update the parameters of the policy network and value network of the PPO algorithm using the collected state-action-reward sequence to optimize the long-term cumulative reward.

Images