An electro-oculogram signal acquisition and analysis system and method

Abstract

The application discloses an electrooculogram signal collection and analysis system and method, relates to the technical field of bioelectric signal collection, and comprises a data processing system, a CPU central processing circuit, a medium-frequency signal generation circuit, an output circuit, a signal collection system and an electrooculogram wave analysis system. The CPU central processing circuit is sequentially connected with the medium-frequency signal generation circuit and the output circuit. The output circuit is adapted to a target to be measured through the signal collection system. The signal collection system transmits collected signals to the electrooculogram wave analysis system, and feeds back the pretreated signals to the CPU central processing circuit. The CPU central processing circuit is in data intercommunication with the data processing system, and closed-loop feedback is performed through the data processing system. The application reduces the contact impedance of the flexible electrode, improves the motion artifact suppression capability, improves the prescription optimization efficiency, and improves the treatment efficiency through clinical verification.

Description

Technical Field

[0001] This invention relates to the field of bioelectric signal acquisition technology, and more specifically to an electrooculogram (EOG) signal acquisition and analysis system and method. Background Technology

[0002] Traditional electrooculography (EOG) acquisition systems employ a three-layer structure: 1) a silver / silver chloride metal electrode layer (0.3-0.5 mm thick); 2) a sodium chloride electrolyte wetting layer; and 3) a rubber retaining ring. The components are connected via physical snap-fit ​​connections, and the signal transmission path is: electrode → twisted-pair copper wire → preamplifier. The manufacturing process uses injection molding, which presents the following problems: the rigid electrode has a fixed radius of curvature (R=50mm), which does not match the average radius of curvature of the human skull (R=75mm), resulting in uneven contact pressure distribution (FEA simulation shows that the pressure peak in the postauricular region reaches 32kPa). The rigid electrode has poor compatibility with the skin contact surface, and long-term wear can easily cause pressure. Electrooculography (EOG) acquisition requires a separate silver chloride ring electrode (12mm in diameter), which is connected to the main control unit through a 3.5mm interface. Multiple devices working together generate a clock error of ±5ms. EOG signal acquisition requires additional dedicated electrodes, increasing the complexity of the equipment. Signal processing uses a Butterworth filter (3rd order, cutoff frequency 0.5-40Hz), which produces baseline drift (±15% amplitude distortion) under electromyographic interference (>100μV). Existing treatment systems use fixed parameter prescriptions and lack a dynamic optimization mechanism based on individual biofeedback. At the same time, multimodal signal fusion analysis relies on manual feature extraction, making it difficult to achieve high-dimensional data association modeling.

[0003] Therefore, how to propose an electrooculography signal acquisition and analysis system and method to solve the problem of biocompatibility of rigid electrodes and achieve synchronous high-precision acquisition of multimodal signals is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides an electrooculography (EOG) signal acquisition and analysis system and method to solve the problem of biocompatibility of rigid electrodes and achieve synchronous high-precision acquisition of multimodal signals. To achieve the above objectives, the present invention adopts the following technical solution:

[0005] An electrooculogram (EOG) signal acquisition and analysis system includes: a data processing system, a CPU central processing circuit, an intermediate frequency (IF) signal generation circuit, an output circuit, a signal acquisition system, and an EOG wave analysis system. The CPU central processing circuit is sequentially connected to the IF signal generation circuit and the output circuit. The output circuit is adapted to the target under test through the signal acquisition system. The signal acquisition system transmits the acquired signal to the EOG wave analysis system for preprocessing and then feeds it back to the CPU central processing circuit. The CPU central processing circuit communicates with the data processing system and performs closed-loop feedback through the data processing system.

[0006] Optionally, the signal acquisition system includes: a multilayer flexible circuit board and an adaptive contact module, wherein the multilayer flexible circuit board is adapted to the target under test through the adaptive contact module.

[0007] Optionally, the multilayer flexible circuit board includes: a base layer, a conductive layer, a shielding layer and an encapsulation layer connected in sequence, and a first electrooculogram (EEG) unit and a second electrooculogram (EOG) unit are embedded and integrated on the conductive layer.

[0008] Optionally, the adaptive contact module includes: fabricating a pyramid structure on PDMS material, optimizing the contact impedance within the strain range of the pyramid structure using finite element method, and adapting it to the target under test using a conductive gel.

[0009] Optionally, the data processing system includes: an individual optimization module, a group analysis module, and a dynamic prescription engine;

[0010] The individual optimization module is used to perform feature encoding on the real-time acquired EEG / EOG signals and output feature vectors;

[0011] The group analysis module is used to calculate the similarity matrix between the current user's features and the group feature library by comparing features;

[0012] The dynamic prescription engine receives individual feature vectors and similarity matrices, generates initial parameters through an attention weight allocation mechanism, and adjusts the parameters based on real-time biological feedback through reinforcement learning.

[0013] Optionally, the individual optimization module includes: collecting data and performing data preprocessing; constructing a Transformer feature extraction network; optimizing the parameters of the Transformer feature extraction network based on the preprocessed data to obtain an optimized feature extraction network.

[0014] Optionally, the group analysis module includes:

[0015] A contrastive learning framework with a dual-tower structure is constructed by using a feature encoding tower and a relation modeling tower. The feature encoding tower uses the ResNet-34 architecture to extract time-frequency map features of user EEG signals, while the relation modeling tower uses a graph attention network to build physiological feature associations between users. The contrastive learning framework with the dual-tower structure is used to calculate the similarity matrix between the current user features and the group feature library.

[0016] Optionally, the dynamic prescription engine includes: constructing a reinforcement learning framework based on proximal policy optimization; defining a state space, action space, and reward function; and training the reinforcement learning framework using a dual-delay deep deterministic policy gradient algorithm.

[0017] Optionally, a method for acquiring and analyzing electrooculogram (EOG) signals includes:

[0018] Apply a baseline stimulus and record the slope of the spectral change to establish an individual response model;

[0019] A temporal feature extraction network is constructed based on the individual response model and the Transformer architecture, and a parameter-feedback mapping relationship is established.

[0020] Cross-user feature transfer is performed using parameter-feedback mapping and contrastive learning to generate an optimized fundamental envelope;

[0021] Initial parameters are generated based on the optimized fundamental envelope, and a reinforcement learning framework is constructed to perform real-time closed-loop adjustment of the parameters.

[0022] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an electrooculography signal acquisition and analysis system and method, which has the following beneficial effects:

[0023] This invention proposes an electrooculography (EOG) signal acquisition and analysis system, comprising: a data processing system, a CPU central processing circuit, an intermediate frequency (IF) signal generation circuit, an output circuit, a signal acquisition system, and an EOG wave analysis system. The CPU central processing circuit is sequentially connected to the IF signal generation circuit and the output circuit. The output circuit is adapted to the target being measured through the signal acquisition system. The signal acquisition system transmits the acquired signal to the EOG wave analysis system for preprocessing before feeding it back to the CPU central processing circuit. The CPU central processing circuit and the data processing system communicate with each other, and closed-loop feedback is achieved through the data processing system. This invention reduces the contact impedance of flexible electrodes, improves the ability to suppress motion artifacts, and increases the efficiency of prescription optimization. Clinical validation has shown that it improves the treatment effectiveness rate. Attached Figure Description

[0024] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 The present invention provides a schematic diagram of the structure of an electrooculography signal acquisition and analysis system.

[0026] Figure 2 The schematic diagram of the traditional intermediate frequency therapy device provided by this invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention discloses an electrooculography (EOG) signal acquisition and analysis system, such as... Figure 1 As shown, it includes: a data processing system, a CPU central processing circuit, an intermediate frequency signal generation circuit, an output circuit, a signal acquisition system, and an electrooculogram (EOG) analysis system. The CPU central processing circuit is connected to the intermediate frequency signal generation circuit and the output circuit in sequence. The output circuit is adapted to the target under test through the signal acquisition system. The signal acquisition system transmits the acquired signal to the EOG analysis system for preprocessing and then feeds it back to the CPU central processing circuit. The CPU central processing circuit communicates with the data processing system and performs closed-loop feedback through the data processing system.

[0029] Furthermore, the signal acquisition system includes a multilayer flexible circuit board and an adaptive contact module, wherein the multilayer flexible circuit board is adapted to the target under test through the adaptive contact module.

[0030] Furthermore, the multilayer flexible circuit board includes a base layer, a conductive layer, a shielding layer and an encapsulation layer connected in sequence, and a first electrooculogram (EEG) unit and a second electrooculogram (EOG) unit are embedded and integrated on the conductive layer.

[0031] In a specific embodiment, the multilayer flexible circuit board specifically includes:

[0032] 1. Structural composition:

[0033] Substrate: 25μm polyimide (PI) film (dielectric constant 3.5, loss factor 0.002);

[0034] Conductive layer: serpentine copper wiring (line width 80μm, spacing 120μm, elongation ≥30%).

[0035] Shielding layer: concentric aluminum foil layer (thickness 50nm, porosity 85%);

[0036] Encapsulation layer: Medical-grade silicone (Shore hardness 20A).

[0037] 2. Embedded Integration:

[0038] EEG unit: 6 triangular electrode arrays (3mm side length, 2.5mm spacing), connected to the analog-to-digital converter chip via vias;

[0039] EOG unit: 2 sets of differential electrode pairs (pitch 8mm±0.1mm), directly connected to the op-amp;

[0040] Common ground: Star topology, impedance <1Ω (@10kHz).

[0041] Furthermore, the adaptive contact module includes: fabricating a pyramid structure on PDMS material, optimizing the contact impedance within the strain range of the pyramid structure using finite element method, and adapting it to the target under test using conductive gel.

[0042] In a specific implementation, the adaptive contact module specifically includes:

[0043] Micropillar array: Pyramid structure made of PDMS material (base 50μm, height 80μm, spacing 100μm);

[0044] Conductive gel: PVA hydrogel containing 0.9% NaCl (conductivity 3.2 S / m, water content 75%).

[0045] Contact mechanics: Contact impedance stabilization (CV<8%) within the 0-15% strain range was achieved through finite element optimization.

[0046] Furthermore, the data processing system includes: an individual optimization module, a group analysis module, and a dynamic prescription engine;

[0047] The individual optimization module is used to perform feature encoding on the real-time acquired EEG / EOG signals and output feature vectors;

[0048] The group analysis module is used to calculate the similarity matrix between the current user's features and the group feature library by comparing features;

[0049] The dynamic prescription engine receives individual feature vectors and similarity matrices, generates initial parameters through an attention weight allocation mechanism, and adjusts the parameters based on real-time biological feedback through reinforcement learning.

[0050] Furthermore, the individual optimization module includes: collecting data and performing data preprocessing; constructing a Transformer feature extraction network; and optimizing the parameters of the Transformer feature extraction network based on the preprocessed data to obtain an optimized feature extraction network.

[0051] In a specific embodiment, the method for constructing the individual optimization module includes:

[0052] 1. Data preprocessing stage:

[0053] Input: Raw EEG signal (1000Hz sampling, 16-bit resolution);

[0054] Noise reduction: An improved CEEMDAN algorithm is used for noise reduction, with adaptive decomposition layer number (N=6-10 layers).

[0055] Standardization: z-score normalization is followed by Box-Cox transformation to eliminate skewed distribution.

[0056] 2. Transformer Feature Extraction Network:

[0057] Encoder structure: 6-layer multi-head attention mechanism (number of heads = 8, hidden layer dimension = 512);

[0058] Position coding: Learnable dynamic position coding (DPE) is used instead of traditional sinusoidal coding;

[0059] Loss function: L=α MSE+β KL-Divergence, (α=0.7,β=0.3). .

[0060] Specifically, KL-Divergence (Kullback-Leibler divergence) represents a measure of the dissimilarity between two probability distributions P and Q. For two discrete probability distributions P and Q, the KL divergence is defined as:

[0061] ;

[0062] Where P(x) is the probability of the true distribution (such as the label distribution in a classification task), and Q(x) is the probability of the model's prediction or approximate distribution (such as the probability distribution of the neural network output). The KL divergence is asymmetric (DKL(P|Q)=DKL(Q|P)), the summation range covers all possible events x, and it is non-negative, taking a value of 0 if and only if P=Q.

[0063] The KL divergence, which measures the mean of the squared differences between predicted and actual values, is expressed as an integral for a continuous probability distribution:

[0064] ;

[0065] in, It is the actual value. This is the predicted value, and n is the number of samples.

[0066] KL-Divergence is used to constrain the similarity between the model's output distribution and the true distribution, and is often used in generative models (such as VAEs) or knowledge distillation tasks. MSE is suitable for regression tasks, directly optimizing the accuracy of numerical predictions. Combining loss functions (e.g., α=0.7, β=0.3) can balance the optimization objectives of distribution alignment and numerical prediction.

[0067] 3. Parameter optimization process:

[0068] Step 1: Initialize parameters θ = {frequency f, intensity I, duty cycle D} ∈ [0, 1]^ 3 ;

[0069] Step 2: Collect real-time biofeedback data Φ = {α wave power, blink frequency, skin impedance};

[0070] Step 3: Generate parameter adjustment action a_t through the policy network π(a|s);

[0071] Step 4: Calculate the reward function r_t = w1·ΔΦ + w2·user comfort score;

[0072] Step 5: Update the Q-value function: Q(s,a) ← Q(s,a) + η[r + γmaxa′Q(s′,a′)] Q(s,a)].

[0073] Furthermore, the group analysis module includes:

[0074] A contrastive learning framework with a dual-tower structure is constructed by using a feature encoding tower and a relation modeling tower. The feature encoding tower uses the ResNet-34 architecture to extract time-frequency map features of user EEG signals, while the relation modeling tower uses a graph attention network to build physiological feature associations between users. The contrastive learning framework with the dual-tower structure is used to calculate the similarity matrix between the current user features and the group feature library.

[0075] Furthermore, the dynamic prescription engine includes: constructing a reinforcement learning framework based on proximal policy optimization; defining the state space, action space, and reward function; and training the reinforcement learning framework using a dual-delay deep deterministic policy gradient algorithm.

[0076] In a specific embodiment, an electrooculogram (EOG) signal acquisition and analysis system includes the following structure:

[0077] 1. Signal acquisition system:

[0078] (1) Multilayer flexible circuit board (PI / PET material): includes embedded first electrooculogram (EEG) unit and second electrooculogram (EOG) unit;

[0079] (2) Adaptive contact module: Microstructured conductive gel is used to improve the stability of skin contact impedance;

[0080] (3) Intelligent switching circuit: supports EEG independent acquisition mode and EEG+EOG synchronous acquisition mode.

[0081] 2. Data Processing System:

[0082] (1) Individual optimization module: The Transformer architecture is used to construct a temporal feature extraction network and establish a parameter-biological feedback mapping relationship;

[0083] (2) Group Analysis Module: Cross-user feature transfer is achieved through contrastive learning to generate optimized fundamental envelope;

[0084] (3) Dynamic prescription engine: Real-time closed-loop adjustment of parameters based on reinforcement learning (RL) framework.

[0085] 3. Key technical parameters:

[0086] Sampling rate: EEG 1000Hz / EOG 500Hz;

[0087] Common-mode rejection ratio: >110dB;

[0088] Wireless transmission: Supports BLE5.0 / Wi-Fi6 dual-mode communication.

[0089] Compared to the structure of traditional intermediate frequency therapy devices, such as Figure 2 As shown, this embodiment uses a signal acquisition system and a data processing system in combination to achieve effective acquisition and analysis of electrooculography (EOG) signals.

[0090] In a specific embodiment, an electrooculogram (EOG) signal acquisition and analysis system includes the following implementation steps:

[0091] 1. Hardware configuration:

[0092] (1) Integrated bioelectricity acquisition module:

[0093] Forehead electrode assembly (Fp1 / Fp2): Flexible dry electrodes were used to acquire EEG data from the frontal lobe (δ / θ band).

[0094] Periorbital electrode array (EOG-L / EOG-R): symmetrically distributed at 5mm from the outer canthus of both eyes to acquire vertical / horizontal eye movement signals;

[0095] Reference electrode: placed behind the ear mastoid process (A1 / A2), using Ag / AgCl conductive gel.

[0096] (2) Stimulus output module:

[0097] Intermediate frequency carrier generator (4.5kHz, compliant with GB 9706.1-2020);

[0098] Dynamic modulator: Adjusts the waveform envelope in real time based on EEG-EOG characteristics.

[0099] Specifically, the electrode manufacturing process includes:

[0100] Photolithography process: SU-8 3050 photoresist was used, with an exposure dose of 350 mJ / cm²;

[0101] Electroplating parameters: copper electroplating current density 2A / dm², time 30min, thickness 15μm;

[0102] Annealing treatment: Anneal at 250℃ for 2 hours in a nitrogen atmosphere to eliminate internal stress.

[0103] Specifically, key signal processing parameters include:

[0104] Common-mode rejection ratio: 112dB (@60Hz);

[0105] Input reference noise: 0.8μVpp (0.5-100Hz bandwidth);

[0106] Phase synchronization accuracy: PLL bandwidth 5Hz, capture range ±15%.

[0107] 2. Signal processing flow:

[0108] (1) EEG feature extraction:

[0109] Visual cortex activation calculation:

[0110] VCA=(β_band_power) / (α_band_power)(13-30Hz / 8-12Hz);

[0111] When VCA > 1.2, it is considered a state of ciliary muscle tension (increased stimulation is required).

[0112] (2) EOG feature extraction:

[0113] Adjust the sensitivity parameters:

[0114] Blink frequency (BF): <15 times / minute triggers a 10% increase in stimulus intensity;

[0115] Sagation Amplitude (SA): Automatically switches stimulation mode when horizontal saccades > 30°.

[0116] 3. Dynamic adjustment algorithm:

[0117] Fundamental envelope generation formula:

[0118] A(t) = K1·(1+e^(t / t)) （-VCA） ) + K2log(BF+1);

[0119] K1 / K2 are individual calibration coefficients, determined through an initial 10-minute biofeedback training.

[0120] Phase synchronization mechanism: The stimulation waveform is synchronized with the α band (8-12Hz) to achieve phase-locked loop synchronization, with a phase difference of <π / 6.

[0121] 4. Closed-loop control of treatment:

[0122] A baseline stimulus (2 mA, 50% modulation) was applied and the slope of the EEG spectrum change was recorded;

[0123] Establish an individual response model: Δα_power vs. stimulus intensity curve;

[0124] Individual optimization module: A temporal feature extraction network is constructed using the Transformer architecture to establish a parameter-biological feedback mapping relationship;

[0125] Population Analysis Module: Achieves cross-user feature transfer through contrastive learning to generate optimized fundamental envelope;

[0126] Dynamic prescription engine: Real-time closed-loop adjustment of parameters based on reinforcement learning (RL) framework.

[0127] Specifically, the system adopts a cascaded-feedback architecture. The workflow is as follows: The individual optimization module first performs feature encoding on the real-time acquired EEG / EOG signals, outputting a feature vector with a dimension of 256. The group analysis module calculates the similarity matrix S∈R^ between the current user's features and the group feature library through feature comparison. （N×K） Where N is the number of users and K=5 is the number of pattern categories. The dynamic prescription engine receives individual feature vectors and similarity matrices, and generates initial parameters through an attention weight allocation mechanism. Reinforcement learning adjusts the parameters based on real-time biological feedback, and simultaneously uploads the optimized parameter-effect pairs to the population feature database. Population feature clustering is updated every 24 hours, and data timeliness is maintained through a sliding window mechanism (T=72 hours).

[0128] Specifically, the method for constructing the individual optimization module includes:

[0129] 1. Data preprocessing stage:

[0130] Input: Raw EEG signal (1000Hz sampling, 16-bit resolution);

[0131] Noise reduction: An improved CEEMDAN algorithm is used for noise reduction, with adaptive decomposition layer number (N=6-10 layers).

[0132] Standardization: z-score normalization is followed by Box-Cox transformation to eliminate skewed distribution.

[0133] 2. Transformer Feature Extraction Network:

[0134] Encoder structure: 6-layer multi-head attention mechanism (number of heads = 8, hidden layer dimension = 512);

[0135] Position coding: Learnable dynamic position coding (DPE) is used instead of traditional sinusoidal coding;

[0136] Loss function: L=α MSE+β KL-Divergence (α=0.7, β=0.3).

[0137] 3. Parameter optimization process:

[0138] Step 1: Initialize parameters θ = {frequency f, intensity I, duty cycle D} ∈ [0, 1]^3;

[0139] Step 2: Collect real-time biofeedback data Φ = {α wave power, blink frequency, skin impedance};

[0140] Step 3: Generate parameter adjustment action a_t through the policy network π(a|s);

[0141] Step 4: Calculate the reward function r_t = w1·ΔΦ + w2·user comfort score;

[0142] Step 5: Update the Q-value function: Q(s,a) ← Q(s,a) + η[r + γmaxa′Q(s′,a′)] Q(s,a)].

[0143] In a specific embodiment, the group analysis module specifically includes:

[0144] The contrastive learning framework employs a dual-tower structure, consisting of an Encoder Tower and a Relation Tower. The Encoder Tower uses the ResNet-34 architecture to extract time-frequency map features of user EEG signals (128×128 grayscale images after STFT transformation), while the Relation Tower constructs physiological feature associations between users through a Graph Attention Network (GAT).

[0145] Specifically, the data augmentation strategy is to apply Gaussian noise (σ=0.05), random frequency band masking (masking rate 15%), and timing misalignment (±200ms) to the original signal.

[0146] Contrast loss function: Improved NT-Xent loss is adopted, with temperature coefficient τ=0.1 and negative sample weight decay coefficient λ=0.01;

[0147] Feature alignment mechanism: Cross-domain feature matching is achieved through the Sinkhorn algorithm in optimal transport theory;

[0148] Fundamental envelope generation: The spectral clustering algorithm is used to classify the population characteristics into patterns and generate K=5 typical envelope templates. Dynamic time warping (DTW) is used to adapt to individual differences.

[0149] In a specific embodiment, the steps for implementing group feature transfer specifically include:

[0150] (1) Establish a distributed feature database:

[0151] Storage structure: Anonymized feature vectors are stored using a time series database (TSDB) with a time resolution of 10 seconds per sample;

[0152] Data encryption: Homomorphic encryption technology is used to encrypt and store sensitive biometric data.

[0153] (2) Cross-device collaborative learning:

[0154] Federated learning framework: Employs the FedAvg algorithm, performing parameter aggregation every 6 hours;

[0155] Differential privacy protection: Add Laplace noise (ε=0.5, δ=1e-5) to the gradient update.

[0156] In a specific embodiment, the dynamic prescription engine specifically includes:

[0157] Constructing a reinforcement learning framework based on proximal policy optimization (PPO):

[0158] (1) State space S: Defines a feature vector with dimension 12, including α / β wave power ratio, EOG saccade velocity, skin impedance change rate, etc.

[0159] (2) Action space A: Three-dimensional continuous action space, corresponding to stimulation frequency (0.5-10Hz), intensity (0.1-5mA), and phase delay (0-360°);

[0160] (3) Reward function R: R = ω1·ln(1+Δα wave power) + ω2·H(user comfort score) + ω3·DKL(Pcurrent||Pgroup);

[0161] Where ω1=0.6, ω2=0.3, ω3=0.1, H(·) is the scoring quantification function, and DKL is the KL divergence constraint term for the population distribution;

[0162] (4) Policy network: The dual-delay deep deterministic policy gradient (TD3) algorithm is adopted, which includes an Actor network (3 layers and 512 nodes) and a Critic network (2 independent Q networks).

[0163] (5) Training mechanism: Historical data is used for pre-training in the offline stage, and ε-greedy exploration strategy is adopted in the online stage (ε=0.2 decay coefficient γ=0.99).

[0164] In specific implementations, the policy network directly generates actions or action probability distributions through neural networks (such as Actor networks), forming the core architecture of the model. For example, the Actor network in TD3 outputs deterministic actions (continuous space, such as the TD3 Actor network) or action probabilities (discrete space, such as the policy gradient method). The ε-greedy exploration policy dynamically balances exploration (random trial) and utilization (selecting the known optimal action) during training. For example, it selects a random action with probability ε and selects the optimal action output by the policy network with probability 1-ε. Specifically, the output is: action selection probability (ε-probability random action, 1-ε probability optimal action).

[0165] Furthermore, the policy network aims to learn the optimal policy for the environment, maximizing cumulative rewards. In the continuous action space, a balance between randomness and determinism needs to be struck (e.g., TD3 uses a dual-criteria network to reduce overestimation). A deep network (e.g., a 3-layer, 512-node Actor) is used to fit the policy function. The output is the action or action distribution parameters (e.g., mean and standard deviation). The network weights (e.g., the 512-node parameters of the Actor) need to be optimized.

[0166] Specifically, the training phase includes:

[0167] Offline phase: Based on historical data pre-training, learn basic strategies.

[0168] Online phase: Continuously optimize policy parameters (such as the weights of the Actor network).

[0169] The described policy network updates its parameters based on historical data or online interactive data. It is suitable for continuous control tasks and scenarios requiring end-to-end policy learning.

[0170] Furthermore, the ε-greedy exploration strategy aims to avoid the strategy getting trapped in local optima and improve sample efficiency. The ε value needs to be dynamically adjusted to adapt to different training stages (e.g., high ε exploration in the early stages, low ε utilization in the later stages).

[0171] Specifically, the training phase includes: effective only in the online phase, dynamically adjusting the action selection strategy. Exploration is gradually reduced and exploitation enhanced by decaying ε (e.g., γ=0.99). The ε-greedy exploration strategy does not directly depend on data; it selects actions solely based on the current policy network output and the ε value, independent of the policy network, and implements the action selection logic through a random number generator. Only ε needs adjustment (e.g., initial ε=0.2, decay rate γ=0.99). It is suitable for discrete action spaces and tasks requiring a rapid balance between exploration and exploitation.

[0172] In a specific embodiment, the policy network is the "brain" of the model, responsible for generating actions; ε-greedy is the "regulator" of the training process, controlling the trade-off between exploration and exploitation. In TD3, the policy network (Actor) outputs deterministic actions, while ε-greedy further promotes exploration through random perturbations during the online phase. The two work together to improve the robustness of the algorithm.

[0173] In a specific embodiment, the dynamic prescription engine optimization process includes:

[0174] (1) Initialization phase:

[0175] Load pre-trained models: Individual Transformer model (parameter θ_ind), Group comparison model (parameter θ_group);

[0176] Establish secure communication: Connect to the cloud service platform via the TLS 1.3 protocol.

[0177] (2) Real-time control stage:

[0178] A policy evaluation is performed every 60ms: the Q-value function Q(s,a) = V(s) + A(s,a) is calculated.

[0179] Policy updates are performed every 5 minutes: Trust domain optimization based on PPO (β=0.01, KL threshold δ=0.03).

[0180] (3) Exception handling mechanism:

[0181] When an impedance change is detected (ΔZ>50%), it automatically switches to safe mode (intensity is reduced to 30%, frequency is fixed at 1Hz).

[0182] Establish a three-level early warning system (warning / degradation / shutdown) corresponding to different physiological parameter thresholds.

[0183] In a specific implementation, a method for acquiring and analyzing electrooculogram (EOG) signals includes:

[0184] Apply a baseline stimulus and record the slope of the spectral change to establish an individual response model;

[0185] A temporal feature extraction network is constructed based on the individual response model and the Transformer architecture, and a parameter-feedback mapping relationship is established.

[0186] Cross-user feature transfer is performed using parameter-feedback mapping and contrastive learning to generate an optimized fundamental envelope;

[0187] Initial parameters are generated based on the optimized fundamental envelope, and a reinforcement learning framework is constructed to perform real-time closed-loop adjustment of the parameters.

[0188] In a specific embodiment, a contact performance test (n=30) was conducted, and the results are as follows:

[0189] Dynamic impedance: 3.2±0.5kΩ (conventional electrode 5.5±1.8kΩ);

[0190] Improved signal-to-noise ratio: Resting-state EEG SNR=8.7dB (traditional 6.2dB), motion-state SNR=5.1dB (traditional 2.3dB).

[0191] Clinical validation (double-blind trial):

[0192] Treatment efficacy rate: 89.3% in the experimental group vs. 60.7% in the control group (p<0.01);

[0193] Optimization speed: The number of iterations required to reach the optimal parameter combination is 4.2 ± 1.3 (the traditional method requires 9.8 ± 2.7).

[0194] This invention achieves a 42% reduction in flexible electrode contact impedance (measured average 3.2kΩ vs. traditional 5.5kΩ); a 37% improvement in motion artifact suppression (assessed by the MSE algorithm); improved prescription optimization efficiency: achieving the optimal parameter combination within 5 treatment cycles for a single user; clinical validation shows a 28.6% increase in treatment effectiveness; the population feature database can be dynamically expanded to the level of 100,000 users; real-time decision latency <15ms (from signal acquisition to parameter output); incremental learning time for individual models <2ms / epoch; and enhanced privacy protection performance: meeting GDPR standards with feature traceability error >98%.

[0195] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0196] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A system for acquiring and analyzing electrooculogram (EOG) signals, characterized in that, include: The system comprises a data processing system, a CPU central processing circuit, an intermediate frequency signal generation circuit, an output circuit, a signal acquisition system, and an electrooculogram (EOG) analysis system. The CPU central processing circuit is sequentially connected to the intermediate frequency signal generation circuit and the output circuit. The output circuit is adapted to the target under test through the signal acquisition system. The signal acquisition system transmits the acquired signal to the EOG analysis system for preprocessing and then feeds it back to the CPU central processing circuit. The CPU central processing circuit communicates with the data processing system and performs closed-loop feedback through the data processing system. The data processing system includes: an individual optimization module, a population analysis module, and a dynamic prescription engine; The individual optimization module is used to perform feature encoding on the real-time acquired EEG / EOG signals and output feature vectors; The group analysis module is used to calculate the similarity matrix between the current user's features and the group feature library by comparing features; The dynamic prescription engine is used to receive individual feature vectors and similarity matrices, generate initial parameters through an attention weight allocation mechanism, and adjust the parameters based on real-time biological feedback through reinforcement learning. The individual optimization module includes: collecting data and performing data preprocessing; constructing a Transformer feature extraction network; optimizing the parameters of the Transformer feature extraction network based on the preprocessed data to obtain an optimized feature extraction network; The group analysis module includes: A contrastive learning framework with a dual-tower structure is constructed by using a feature encoding tower and a relation modeling tower. The feature encoding tower uses the ResNet-34 architecture to extract time-frequency map features of user EEG signals, while the relation modeling tower uses a graph attention network to build physiological feature associations between users. The contrastive learning framework with the dual-tower structure is used to calculate the similarity matrix between the current user features and the group feature library. The loss function for the Transformer feature extraction network is: L=a MSE+β KL-Divergence, α=0.7, β=0.3; KL-Divergence is a measure of the difference between two probability distributions P and Q. For two discrete probability distributions P and Q, the KL divergence is defined as: ； Where P(x) is the probability of the true distribution, Q(x) is the probability of the model prediction or approximate distribution, the KL divergence is asymmetric, the summation range covers all events x, and it is non-negative, and takes a value of 0 if and only if P=Q. The KL divergence, which measures the mean of the squared differences between predicted and actual values, is expressed as an integral for a continuous probability distribution: ； in, It is the actual value. This is the predicted value, and n is the number of samples.

2. The electrooculogram (EOG) signal acquisition and analysis system according to claim 1, characterized in that, The signal acquisition system includes a multilayer flexible circuit board and an adaptive contact module, wherein the multilayer flexible circuit board is adapted to the target under test through the adaptive contact module.

3. The electrooculography signal acquisition and analysis system according to claim 2, characterized in that, The multilayer flexible circuit board includes a base layer, a conductive layer, a shielding layer and an encapsulation layer connected in sequence, and a first electrooculogram (EEG) unit and a second electrooculogram (EOG) unit are embedded and integrated on the conductive layer.

4. The electrooculogram (EOG) signal acquisition and analysis system according to claim 2, characterized in that, The adaptive contact module includes: fabricating a pyramid structure on PDMS material, optimizing the contact impedance within the strain range of the pyramid structure using finite element method, and adapting it to the target under test using conductive gel.

5. The electrooculogram (EOG) signal acquisition and analysis system according to claim 1, characterized in that, The dynamic prescription engine includes: constructing a reinforcement learning framework based on proximal policy optimization; defining the state space, action space, and reward function; and training the reinforcement learning framework using a dual-delay deep deterministic policy gradient algorithm.

6. A method for acquiring and analyzing electrooculogram (EOG) signals, applicable to the EOG signal acquisition and analysis system described in any one of claims 1-5, characterized in that, include: Apply a baseline stimulus and record the slope of the spectral change to establish an individual response model; A temporal feature extraction network is constructed based on the individual response model and the Transformer architecture, and a parameter-feedback mapping relationship is established. Cross-user feature transfer is performed using parameter-feedback mapping and contrastive learning to generate an optimized fundamental envelope; Initial parameters are generated based on the optimized fundamental envelope, and a reinforcement learning framework is constructed to perform real-time closed-loop adjustment of the parameters.

Images