Smokeless intelligent accurate control method for preventing and controlling juvenile myopia

Abstract

The invention provides a smokeless intelligent accurate control method for prevention and control of juvenile myopia. The smokeless intelligent accurate control method comprises the steps of eye movement and facial muscle multi-mode acquisition and noise filtering, dynamic graph convolutional network and spatial-temporal feature fusion analysis, blink and head movement prediction and intelligent moxibustion path adjustment. By constructing a resting state eye movement behavior mode library, coupling face and head dynamic data, utilizing multi-mode feature fusion and AI hierarchical control, accurately recognizing blinking and head disturbance, and generating an avoidance track and a dynamic moxibustion therapy power regulation strategy in advance, the method can effectively avoid periorbital sensitive areas, keep the temperature in the moxibustion therapy process constant, and improve the moxibustion therapy effect. The response speed and path stability are improved, and moxibustion safety and continuity are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of myopia prevention and control in adolescents and dynamic moxibustion trajectory control technology, and in particular to a smokeless intelligent precision control method for myopia prevention and control in adolescents. Background Technology

[0002] The intervention and prevention of myopia in adolescents has become an important global medical and public health issue. Traditional Chinese medicine moxibustion, due to its non-drug and side-effect-free characteristics, is gradually gaining attention and application in the treatment of myopia in adolescents. In recent years, with the development of smart hardware and visual perception technology, automated moxibustion-based eye and acupoint irradiation devices have seen extensive application research. These technologies typically integrate infrared / laser smokeless heat sources, using optical cameras for acupoint identification and automatic positioning to achieve targeted moxibustion at preset acupoints. Some advanced devices incorporate robotic gimbals, allowing the moxibustion head to adaptively move along a set path in three-dimensional space. They can also be combined with infrared temperature sensors or epidermal thermal imaging to achieve closed-loop temperature control of heat output.

[0003] However, existing smokeless intelligent moxibustion technologies generally rely on traditional PID feedback or delayed path fine-tuning algorithms. These algorithms can only passively replan the moxibustion path after detecting significant head displacement or cooperative instructions from the user, lacking immediate response to "non-cooperative" and "minor disturbance" scenarios. For example, when adolescent users are relaxing, distracted, or unconsciously swaying, they frequently exhibit involuntary, slight head movements, blinking, and other unconscious small actions. In such cases, the system's perception and path adjustment are delayed. Due to multi-layered delays in data acquisition, feature recognition, and motion control feedback, if the repositioning / obstacle avoidance trajectory of the moxibustion head to the visual target is not generated in time, moxibustion interruptions, acupoint deviations, or temperature fluctuations will occur, affecting continuity, comfort, and safety.

[0004] Furthermore, current head movement compensation strategies primarily employ static mapping or simple linear prediction, failing to deeply integrate with the dynamic eye-movement and facial behavior patterns unique to adolescents. This results in inaccurate path adjustment during minor head disturbances, and may even lead to over-recovery or false triggering. Simultaneously, solutions relying on external markers, head-mounted devices, or expensive motion capture systems suffer from poor comfort and are not suitable for widespread application to children and adolescents. Both approaches generally face technical bottlenecks in practical applications, such as poor user experience and limited control precision. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present invention provides a smoke-free intelligent and precise control method for myopia prevention and control in adolescents.

[0006] The technical solution of this invention is implemented as follows: A smoke-free, intelligent, and precise control method for myopia prevention and control in adolescents, comprising:

[0007] S1: Collect eye movement data of adolescents when they are looking at a fixed visual target in a relaxed state, including blink frequency, saccade amplitude, micro-flutter rhythm, and facial muscle activity signals, and construct a resting eye movement behavior pattern library;

[0008] S2: The eye-tracking behavior data is correlated with the facial key point displacement sequence to generate an eye-tracking-head motion coupling law database, which is used to predict the user's posture perturbation trend.

[0009] S3: The user's facial 3D model is acquired in real time through an optical positioning camera, the acupoints around the eyes and the whole body are identified, and the displacement sequence of key facial points is extracted as the node input of the dynamic anatomical topology map.

[0010] S4: Input the dynamic anatomical topology map into the lightweight graph neural network, and combine it with multimodal sensing data from an infrared temperature sensor and an optional macro thermal imager to perform spatiotemporal modeling analysis;

[0011] S5: Based on the eye movement prior knowledge base, determine whether the current eyelid closure speed exceeds the preset threshold, and detect whether the brow is raised to determine whether the user is about to blink actively;

[0012] S6: If it is determined that a blinking action is about to occur, further detect whether the head generates a continuous angular acceleration in a certain direction to activate the path offset compensation strategy in the corresponding direction.

[0013] S7: Based on the pre-stored eye-head coupling law database, the central controller generates the spatial avoidance trajectory of the moxibustion head 0.3~0.5 seconds in advance, so that the moxibustion spot withdraws from the sensitive area along the preset safety arc;

[0014] S8: After the blinking action is completed, the high-speed closed-loop temperature control and position servo are realized through FPGA to quickly return the moxibustion head to the original acupoint path, maintaining the continuity and stability of the moxibustion process.

[0015] S9: It adopts a hierarchical control architecture, which only triggers the upper-layer AI algorithm to perform full path replanning when significant attitude drift is detected, so as to reduce system resource consumption and improve response efficiency;

[0016] S10: During the path adjustment process, the skin surface temperature change is continuously monitored, and the output of the smokeless moxibustion energy source is adjusted through a closed-loop feedback mechanism to ensure that the moxibustion temperature is kept constant within the range of 42℃±2℃.

[0017] The present invention provides a smoke-free, intelligent, and precise control method for myopia prevention and control in adolescents, which has the following beneficial effects:

[0018] (1) Significantly improves the path response speed and stability of the dynamic suspension moxibustion system in the non-cooperative state of adolescents, effectively solves the problem of moxibustion interruption and positioning drift caused by uncontrollable factors such as head shaking and blinking, significantly improves the foresight and accuracy of the spatial trajectory planning of the suspension moxibustion head, and the moxibustion spot can reliably avoid the sensitive area of ​​the face, so as to realize the continuous and uninterrupted moxibustion process.

[0019] (2) The performance levels are clearly defined. The bottom-level servo and temperature control links are in a high-speed closed loop, while the upper-level multimodal AI algorithm is only triggered when there is a significant posture abnormality, which effectively reduces the consumption of system resources and greatly improves processing efficiency and dynamic tracking sensitivity. It eliminates the need for traditional head-mounted motion sensing devices or cumbersome external auxiliary devices. The operation process is non-invasive and highly comfortable, making it suitable for teenagers.

[0020] (3) Multimodal fusion perception improves the accuracy of posture prediction, triggers path adjustment in advance, and avoids blind spots and the risk of mis-moxibustion due to high temperature. Temperature closed-loop control ensures the safety of moxibustion, significantly reduces the hidden dangers of low temperature burns or insufficient moxibustion temperature, and improves the consistency of therapeutic effect;

[0021] (4) It has strong environmental adaptability and diverse moxibustion treatment scenarios, and can be widely used in clinical, home and rehabilitation fields, especially suitable for long-term management needs such as myopia prevention and vision maintenance in adolescents. The dynamic replanning mechanism greatly improves the user experience and reduces pain, moxibustion interruption or decreased comfort caused by movement interference. Attached Figure Description

[0022] Figure 1 This is a flowchart of a smoke-free intelligent precision control method for myopia prevention and control in adolescents according to the present invention;

[0023] Figure 2 This is a sub-flowchart of a smoke-free, intelligent, and precise control method for myopia prevention and control in adolescents according to the present invention.

[0024] Figure 3 This is another sub-flowchart of the present invention for a smoke-free intelligent precision control method for myopia prevention and control in adolescents. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0026] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0027] like Figure 1 As shown, this invention provides a smoke-free, intelligent, and precise control method for myopia prevention and control in adolescents, specifically including:

[0028] S1: Collect eye movement data of adolescents when they are looking at a fixed visual target in a relaxed state, including blink frequency, saccade amplitude, micro-flutter rhythm, and facial muscle activity signals, and construct a resting eye movement behavior pattern library;

[0029] S2: The eye-tracking behavior data is correlated with the facial key point displacement sequence to generate an eye-tracking-head motion coupling law database, which is used to predict the user's posture perturbation trend.

[0030] S3: The user's facial 3D model is acquired in real time through an optical positioning camera, the acupoints around the eyes and the whole body are identified, and the displacement sequence of key facial points is extracted as the node input of the dynamic anatomical topology map.

[0031] S4: Input the dynamic anatomical topology map into the lightweight graph neural network, and combine it with multimodal sensing data from an infrared temperature sensor and an optional macro thermal imager to perform spatiotemporal modeling analysis;

[0032] S5: Based on the eye movement prior knowledge base, determine whether the current eyelid closure speed exceeds the preset threshold, and detect whether the brow is raised to determine whether the user is about to blink actively;

[0033] S6: If it is determined that a blinking action is about to occur, further detect whether the head generates a continuous angular acceleration in a certain direction to activate the path offset compensation strategy in the corresponding direction.

[0034] S7: Based on the pre-stored eye-head coupling law database, the central controller generates the spatial avoidance trajectory of the moxibustion head 0.3~0.5 seconds in advance, so that the moxibustion spot withdraws from the sensitive area along the preset safety arc;

[0035] S8: After the blinking action is completed, the high-speed closed-loop temperature control and position servo are realized through FPGA to quickly return the moxibustion head to the original acupoint path, maintaining the continuity and stability of the moxibustion process.

[0036] S9: It adopts a hierarchical control architecture, which only triggers the upper-layer AI algorithm to perform full path replanning when significant attitude drift is detected, so as to reduce system resource consumption and improve response efficiency;

[0037] S10: During the path adjustment process, the skin surface temperature change is continuously monitored, and the output of the smokeless moxibustion energy source is adjusted through a closed-loop feedback mechanism to ensure that the moxibustion temperature is kept constant within the range of 42℃±2℃.

[0038] Step S1: Collect eye movement data of adolescents in a relaxed state while fixating on a fixed visual target, including blink frequency, saccade amplitude, micro-flutter rhythm, and facial muscle activity signals, to construct a resting-state eye movement behavior pattern database. Specifically, this includes:

[0039] S1.1: Based on an eye-tracking system and electromyography signal acquisition equipment, the raw data of eye movements of adolescents when they fixate on a fixed visual target in a relaxed state are acquired simultaneously, including blink frequency, saccade amplitude and micro-tremor rhythm, in order to generate a raw eye movement behavior feature dataset.

[0040] Based on an eye-tracking system and electromyography (EMG) signal acquisition equipment, tests were conducted on adolescents under fixed visual target presentation conditions. Raw eye movement signals and synchronous EMG signals were collected over a continuous period of time to obtain raw data on eye movement characteristics in a resting state.

[0041] A high-resolution video eye-tracking system (sampling rate 250~500Hz, angular resolution ≤0.1°) is used to record the timing of blinking events, including the degree of eyelid opening and closing. The horizontal and vertical eye position coordinates of each sampling point are obtained through the pupil center localization algorithm and corneal reflection vector calculation within the system.

[0042] Furthermore, the displacement change rate of consecutive frames is calculated by the eye position coordinate difference algorithm (parameter: time window length 4ms) to generate a saccade amplitude sequence, so as to characterize the micro displacement range of the fixation point in the resting state.

[0043] Furthermore, a spectrum analysis algorithm (Fast Fourier Transform, FFT, resolution 0.5Hz) was used to extract the dominant frequency component of the micro-tremor rhythm, forming the micro-tremor feature spectrum of the eyeball in the resting state, which serves as an important input for subsequent pattern analysis;

[0044] During the synchronization of eye movement signal acquisition, the raw electromyographic signals of the frontalis muscle and orbicularis oculi muscle are acquired through a multi-channel surface electromyography acquisition device (sampling rate ≥1kHz, signal-to-noise ratio ≥60dB). The eye movement and electromyographic data are precisely synchronized using the system timestamp to achieve consistent acquisition of multimodal signals.

[0045] Through the above chain processing, a raw eye movement feature dataset containing blink frequency, saccade amplitude, micro-tremor rhythm and synchronous electromyographic signal is generated, which serves as the basic data preparation for the subsequent construction of the resting-state eye movement pattern library.

[0046] For example, in a resting-state test scenario, an infrared video eye-tracking system with a sampling rate of 500Hz was used to record the adolescent subjects' gaze at a fixed crosshair target at a distance of 40cm from the center of the screen for 2 minutes. The system output the horizontal coordinate (x) and vertical coordinate (y) of each sampling point. Based on the blink frequency characteristics, the instantaneous closing velocity of the eyelid height sequence was calculated using differential calculation. The total number of blink events within 2 minutes was 42, and the blink frequency was [missing information]. Hz. For the saccade amplitude characteristics, a 4ms time window was used to calculate the horizontal eye position change rate, yielding an average saccade amplitude of 0.3° and a maximum amplitude of 1.1°. For the micro-flicker rhythm, the eye position coordinate sequence was subjected to FFT transformation to extract the power spectrum within the 0–40Hz range, detecting a significant peak (power intensity 0.85 units) at the dominant frequency of 3.2Hz. Simultaneously, an electromyography system with a sampling rate of 1kHz was used to record the activity of the frontalis muscle and orbicularis oculi muscle. After synchronization processing, a brief contraction with a peak amplitude of 72μV was observed in the frontalis muscle signal before the blink event. The final output is a raw eye movement behavior feature dataset containing parameters such as blink frequency 0.35Hz, average saccade amplitude 0.3°, micro-flicker dominant frequency 3.2Hz, and corresponding electromyography peak values, providing complete input for subsequent wavelet filtering and pattern library construction.

[0047] S1.2: Perform time-frequency domain joint filtering on the original eye movement behavior feature dataset, and use wavelet transform algorithm to remove electrooculography interference and electromyography noise, so as to extract a clean eye movement event sequence as the eye movement feature parameter set;

[0048] S1.3: Based on a multi-channel surface electromyography signal acquisition device, electromyographic activity data of key facial muscle groups (such as frontalis muscle and orbicularis oculi muscle) are acquired, and an adaptive filtering algorithm is used to denoise and normalize the electromyographic signals in order to extract facial muscle activity feature parameters.

[0049] Based on a multi-channel surface electromyography (EMG) signal acquisition device, synchronous signal acquisition of key facial muscle groups is performed in a resting state. The input signals include the original EMG voltage time sequence of the frontalis muscle and orbicularis oculi muscle in a relaxed state of gazing at a fixed visual target.

[0050] A bandpass filtering algorithm (parameters: cutoff frequency range 20Hz~450Hz) was used to perform preliminary frequency domain denoising on the original electromyographic voltage time series, preserving the effective frequency band signals related to muscle contraction and suppressing low-frequency motion artifacts and high-frequency power frequency interference.

[0051] Furthermore, an adaptive filtering algorithm (parameters: reference input is the environmental interference signal, step size μ=0.01) is used to suppress background noise in the bandpass filtered signal, thereby achieving dynamic suppression of environmental electromagnetic interference and electrooculography cross noise.

[0052] Furthermore, a normalization method (parameter: each channel is independently standardized with zero mean and unit variance) is used to map the electromyographic signal amplitude of each channel after adaptive filtering to a uniform dimension interval, so as to eliminate the influence of differences in electromyographic amplitude among different individuals on feature extraction.

[0053] A time-domain statistical feature extraction algorithm (parameters: window length 250ms, sliding step size 50ms) was used to extract feature indicators reflecting the intensity and rhythm of facial micromuscular movements from normalized electromyography sequences, such as root mean square value, mean absolute value, waveform length, and zero crossover rate, and to generate facial muscle activity feature parameters.

[0054] The above algorithm transforms the results of the previous step into structured facial muscle activity feature data, achieving high signal-to-noise ratio quantization output of facial motion features, which can provide highly robust input parameters for subsequent eye-movement-facial muscle movement multimodal fusion modeling.

[0055] For example, during facial muscle activity acquisition, bipolar patch electrodes were used to collect data from the frontalis muscle and orbicularis oculi muscle, with a sampling rate set to 2000Hz. The raw electromyography (EMG) data were processed using a bandpass filter (20Hz~450Hz, fourth-order Butterworth) to remove unusable frequency bands, ensuring passband ripple was less than 0.5dB in the filter design. In the adaptive filtering stage, the least mean square (LMS) algorithm was used, with the reference signal taken from an environmental noise monitoring electrode 20cm from the facial muscle area. The step size parameter μ was set to 0.01, and the mean square error decreased significantly after iterative convergence. In the normalization stage, a zero-mean unit variance transformation was performed on each channel of data to ensure consistent amplitude ranges among different subjects. In the feature extraction stage, the root mean square value, mean absolute value, waveform length, and zero crossover rate were calculated within a 250ms time window to obtain a 4-dimensional feature vector for each window. These vectors were then concatenated to generate a complete EMG feature matrix. In testing and verification, the matrix exhibited high stability and low noise interference during subsequent DTW time-series alignment, which can significantly improve the classification accuracy of the resting-state eye-movement behavior pattern library.

[0056] S1.4: Perform multimodal fusion modeling on the eye movement feature parameter set and facial muscle activity feature parameters, and use dynamic time warping algorithm to temporally align the eye movement and facial muscle movement data to generate an eye movement-facial muscle movement coupling feature vector.

[0057] The eye movement feature parameter set and facial muscle activity feature parameters obtained by processing S1.2 and S1.3 are used to realize the representation mapping of cross-modal data in a unified feature space by adopting a multimodal feature fusion algorithm (parameters: feature dimension matching rules, normalization coefficients);

[0058] Furthermore, the Dynamic Time Warping (DTW) algorithm (parameters: Euclidean distance as the distance metric, Sakoe-Chiba bandwidth as the path constraint) is used. This allows for non-uniform length matching and alignment of eye movement event time series and facial muscle movement event time series, and yields the minimum cumulative distance path matrix.

[0059] Furthermore, based on the minimum cumulative distance path matrix, a temporal synchronization index vector is constructed, and the event frames of the two modalities are resampled according to the synchronization index to generate an eye movement and facial muscle movement event alignment sequence of equal length and temporal matching.

[0060] Furthermore, a feature concatenation and weighted fusion strategy is adopted (parameter: weight ratio). and (corresponding to eye-tracking and muscle-movement modalities respectively), achieving a linear combination of cross-modal features and forming a fusion dimension of... The feature matrix is ​​×d, where d is the single-modal feature dimension;

[0061] By normalizing the data (parameter: Z-score standardization), the numerical values ​​of each dimension of the feature matrix are converted into standardized data with zero mean and unit variance, ensuring the comparability of different modal features in terms of dimensions. Based on this, an eye-movement-facial muscle movement coupling feature vector is generated to achieve a high-precision description of the coordinated pattern of eye movement and facial muscle movement in the resting state.

[0062] For example, in one embodiment, the eye-tracking feature parameter set has a dimension of 1. The facial muscle activity feature parameters have the following dimensions: The minimum cumulative distance obtained after matching using the DTW algorithm is: The alignment sequence length is set to Frame. Feature fusion stage, according to weight ratio and Weighted splicing is performed to form a fusion dimension. The feature vectors are then used. After standardization, the coupling vectors are used for subsequent Gaussian mixture model clustering, with the model set to have 100 clusters. The discrimination rate is significantly improved, and it can stably identify four types of resting state sub-modes: flat blink, fast saccade, slow saccade, and micro-tremor, which verifies the effectiveness of the multimodal fusion and temporal alignment strategy.

[0063] S1.5: Based on the eye-movement-facial muscle movement coupling feature vector, a Gaussian mixture model is used to perform cluster analysis on the eye movement behavior patterns of adolescents in a relaxed state, so as to identify and classify the resting eye movement behavior pattern subclasses and construct a resting eye movement behavior pattern library.

[0064] Based on eye-movement-facial muscle motion coupling feature vectors, a Gaussian mixture model (parameters: number of components K, number of expectation-maximization iterations N, convergence threshold ε) is used to estimate the probability density and model the distribution of the feature vector set in the feature space.

[0065] Furthermore, through the Expectation-Maximization (EM) algorithm (parameter: initial mean) Covariance and mixed weights This enables iterative optimization of model parameters and yields the mixture probability vector for each component pair of samples.

[0066] Furthermore, each feature vector is assigned a pattern cluster label by the maximum a posteriori probability discrimination criterion, thereby realizing the pattern classification of the resting-state feature vector and generating a sample set within the pattern cluster;

[0067] Furthermore, by calculating the mean vector and covariance matrix within each pattern cluster, a statistical correlation model between intra-cluster eye movement behavior and facial muscle movement is achieved, thereby characterizing the physiological differences among different pattern clusters.

[0068] By constructing a resting-state eye movement behavior pattern library using the above clustering labels and intra-cluster statistical features, we can realize the subclassification and structured data storage of users' eye movement-facial muscle movement coupling behavior in a relaxed state.

[0069] For example, in the adolescent resting gaze experiment scenario, the input feature vector dimension is set to 20 dimensions, the number of components K is set to 4, the number of iterations N is set to 200, and the convergence threshold ε is set to... The initial mean vector was predicted using the K-means algorithm, the covariance matrix was set to diagonal matrix form, and the mixture weights were all set to 1. During the EM iteration, the E-step is performed first to calculate the membership probability of each sample i on each component k. Then, perform M steps to update the mean based on the member probabilities of all samples. Covariance and weight After the EM iteration converges, a pattern cluster label is assigned to each sample using the maximum a posteriori probability rule, and the mean eye closure velocity is calculated within each cluster. Mean of brow myocardial effect This process forms pattern cluster feature descriptions. The final generated resting-state pattern library contains four pattern clusters, each corresponding to different eye movement frequencies and muscle motion coupling features. These clusters can significantly improve prediction accuracy and response stability in non-cooperative states during subsequent pose prediction and path planning.

[0070] Step S2: The eye-tracking behavior data is correlated with facial key point displacement sequences to create an eye-tracking-head-movement coupling pattern database, used to predict user posture perturbation trends. Specifically, this includes:

[0071] S2.1: Perform time alignment processing on the eye movement data and facial key point displacement sequence collected in S1 to eliminate the temporal offset between multimodal signals;

[0072] S2.2: Based on the dynamic time warping algorithm, pattern matching is performed on eye movement behavior features and facial key point trajectories to extract the temporal coupling relationship features between blinking, saccades and head movements;

[0073] S2.3: The nonlinear dependency between eye movement parameters and facial displacement vector is modeled using mutual information analysis to generate the eye-head motion coupling strength matrix;

[0074] For the temporal coupling relationship between eye movement behavior features and facial key point trajectories after S2.2 pattern matching, the mutual information analysis method (parameters: feature sequence length N=256, time window width=50ms) is used to achieve quantitative modeling of the nonlinear dependence between the two types of signals.

[0075] Furthermore, by using the probability density estimation method (parameters: kernel density estimation type is Gaussian kernel, bandwidth value is automatically calculated according to Silverman's rule), the joint probability distribution P(X,Y) of the eye movement parameter sequence and the facial displacement vector sequence is estimated, and the datasets of marginal probability distributions P(X) and P(Y) are obtained.

[0076] Furthermore, the nonlinear dependence values ​​of the two types of signals are calculated using the mutual information formula, thereby obtaining the mutual information value between each set of eye-tracking parameters and the facial displacement vector. ;

[0077] Furthermore, for the full mutual information value matrix, normalization processing is adopted (parameter: the normalization method is Min-Max, and the value range is mapped to [0,1]) to generate the eye-movement-head motion coupling strength matrix, so as to realize a unified dimensional representation of the coupling strength between different parameter pairs;

[0078] Through mutual information analysis and normalization, the temporal coupling relationship features from the previous step are transformed into a quantified nonlinear dependency strength matrix, enabling an accurate characterization of the coupling pattern between eye movement and head movement, and providing highly consistent input features for subsequent dimensionality reduction and prediction model training.

[0079] For example, in the resting-state data acquisition of adolescent users, the selected eye movement parameters include blink frequency X1 (Hz), saccade amplitude X2 (°), and micro-tremor rhythm X3 (Hz), with corresponding facial displacement vectors including nasal tip displacement Y1 (mm), brow center displacement Y2 (mm), and outer canthus displacement Y3 (mm). A Gaussian kernel with bandwidth h=0.35 is used for joint probability density estimation to obtain the p(x,y) distribution in the sample space, and the mutual information value for each pair of parameters is calculated, such as I(X1;Y1)=0.42, I(X2;Y2)=0.68, and I(X3;Y3)=0.51. The obtained mutual information matrix is ​​then Min-Max normalized, resulting in a normalized coupling strength matrix with corresponding elements of 0.62, 1.00, and 0.75, representing the relative coupling strength between these parameter pairs. Model validation shows that when this matrix is ​​used as an input feature, the model accuracy for predicting head perturbation trends is significantly improved, and the eye-tracking-head-tracking coupling pattern has stable transferability among different users.

[0080] S2.4: Principal component analysis is used to reduce the dimensionality of the coupling strength matrix and extract the feature vectors of key coupling modes to construct a low-dimensional eye-movement-head-movement coupling feature space.

[0081] S2.5: Train a long short-term memory network model based on the coupled feature space, establish a predictive mapping relationship between eye movement sequence and head displacement trend, and generate an eye-guided head movement prediction model;

[0082] Based on the aforementioned coupled feature space, a Long Short-Term Memory (LSTM) network model (parameter settings: 128 hidden layer units, 20 time steps, tanh activation function) is used to establish a nonlinear predictive mapping relationship between eye movement behavior sequences and head displacement trends.

[0083] Furthermore, by serializing the low-dimensional coupled feature vectors in the training samples (method: arranged in the order of time windows to ensure the integrity of temporal dependencies between features), the model input vector is standardized to obtain the normalized temporal feature matrix;

[0084] Furthermore, using the mean squared error (MSE) as the loss function, the prediction error is calculated using the following formula:

[0085]

[0086] in, For predicted values, For the target value, The number of samples;

[0087] Furthermore, the weight matrix of the LSTM model is iteratively updated using the Adam optimization algorithm (parameters: learning rate 0.001, β1=0.9, β2=0.999) to achieve rapid convergence of errors and obtain stable prediction performance.

[0088] Furthermore, the impact of different hyperparameter combinations on prediction accuracy was evaluated using a cross-validation method (fold number: 5), and the parameter combination with the lowest overall prediction error on each fold validation set was selected as the final model configuration.

[0089] Through the above LSTM training and optimization process, the coupled feature space obtained in the previous step is transformed into an eye-tracking guided prediction model that can output head displacement trend prediction results in real time, thus realizing the ability to proactively identify user posture perturbations.

[0090] For example, for the constructed low-dimensional coupled feature space, eye-tracking and head-movement time-series data with 2000 samples are selected, with a time window length of 20 frames, and each frame contains 16-dimensional fused feature values. An LSTM network is used with 128 hidden layer units, linear activation in the output layer, a batch size of 64, and 50 training epochs. During training, the prediction error value for each sample is calculated using the MSE loss function. For example, for a certain batch of samples, the sum of squared errors between the actual value vector and the predicted value vector is calculated as follows:

[0091]

[0092] After iterative updates by the Adam optimizer, the prediction error tended to stabilize after 35 rounds of training. The trend prediction output in the test set was highly consistent with the actual head displacement trend, and it could achieve stable look-ahead prediction results in multiple sample scenarios, significantly improving the continuity and accuracy stability of the moxibustion path in adolescents with slight posture perturbations.

[0093] S2.6: Jointly train and optimize the prediction model with the three-dimensional facial key point coordinate sequence to form an eye-tracking prior knowledge base that can be used to predict user posture perturbation trends in non-cooperative states.

[0094] In the input conditions, the prediction model comes from the training results of the Long Short-Term Memory Network (LSTM) on the low-dimensional eye-movement-head-movement coupling feature space. The three-dimensional facial key point coordinate sequence is acquired in real time by an optical positioning camera and formed into continuous spatial trajectory data after time alignment and filtering.

[0095] A joint training method (parameters: the objective function is to minimize the prediction error of the posture perturbation trend, the learning rate is set to 0.001, and the batch size is 32) is adopted to achieve synchronous iterative updates of the prediction model parameters and the 3D facial key point coordinate sequence under the same optimization framework;

[0096] Furthermore, by using a multi-task learning algorithm (parameters: the main task is eye-tracked head movement trend prediction, and the auxiliary task is facial key point trajectory reconstruction), the robustness of the prediction model in the shared feature representation space is enhanced, and a feature weight matrix that is applicable to both trend prediction and trajectory estimation is obtained.

[0097] Furthermore, through optimization of the spatiotemporal attention mechanism (parameters: time window length 500ms, spatial adjacency matrix threshold 0.5), adaptive allocation of feature weights for highly relevant regions in the 3D keypoint sequence is achieved, and enhanced spatiotemporal context feature vectors are generated.

[0098] The gradient-weighted class activation mapping method is used to visualize and analyze the output of the optimized model, and the contribution distribution of each input feature in the perturbation trend prediction is calculated to verify the effectiveness of the model's feature selection.

[0099] By using parameter constraint regularization (parameter: L2 regularization coefficient 0.0005), the result of the previous step is transformed into an eye-tracking prior knowledge base that still maintains high prediction accuracy in non-cooperative states, thereby significantly improving the trend prediction ability for perturbation actions such as slight head shaking and rapid blinking.

[0100] For example, in a real-world dynamic moxibustion application scenario for teenagers, an optical positioning camera with a sampling frequency of 60Hz is selected to acquire a 3D coordinate sequence of 17 facial key points, including the tip of the nose, inner canthus, and outer canthus, and this sequence is simultaneously input into an LSTM prediction model. The main task loss weight in the multi-task learning model is set to 0.8, and the auxiliary task loss weight is set to 0.2. The Adam optimizer converges to a prediction error of less than [a certain value] within 100 iterations. In the spatiotemporal attention module, the adjacency weight of key points around the eyes and the glabella area is increased to [value missing]. This significantly improves the predictive sensitivity of the synchronization trend between eyelid closure and head micro-movements. In the output phase, the eye-tracking prior knowledge base targets the period before blinking. The prediction accuracy remains at a high level, effectively driving the moxibustion trajectory to make advance avoidance actions, thereby ensuring the continuity and comfort of the moxibustion process.

[0101] like Figure 2 As shown, step S3 involves: acquiring a real-time 3D model of the user's face using an optical positioning camera, identifying the locations of acupoints around the eyes and throughout the body, and extracting the displacement sequence of key facial points as node input for a dynamic anatomical topology map. Specifically, this includes:

[0102] S3.1: Perform multi-scale Gaussian filtering preprocessing on the user's facial RGB-D image data acquired by the optical positioning camera to eliminate image noise interference and obtain a smoothed facial depth map;

[0103] S3.2: Based on the pre-trained 3D facial key point detection model, feature extraction is performed on the smoothed facial depth map to identify and locate 17 facial reference anatomical points, including the tip of the nose, inner canthus of the eye, outer canthus of the eye, center of the eyebrows, and center of the lips, to generate the initial three-dimensional geometric topology of the face.

[0104] For the smoothed facial depth map obtained by S3.1 multi-scale Gaussian filtering preprocessing, a 3D facial key point detection model based on pre-trained parameters (the model training data covers the facial feature distribution of adolescents of multiple ages) is used to generate the feature response map of the facial region when the preprocessed depth map data is input, so as to capture the depth gradient pattern related to the anatomical reference point position.

[0105] Furthermore, spatial features are extracted from the feature response map using a three-dimensional geometric convolution algorithm (parameters: kernel size 5×5×5, stride 2, activation function ReLU), and the facial surface curvature and normal vector distribution are calculated at each spatial scale to achieve preliminary localization and prediction of potential anatomical reference point regions and obtain a set of key candidate point coordinates.

[0106] Furthermore, an iterative nearest point (ICP) registration algorithm based on shape constraints (parameters: maximum number of iterations is 50, convergence threshold is 1e-6) is adopted to match the coordinate set of key candidate points with the three-dimensional face reference template. The coordinate positions are optimized based on facial anatomy calibration rules to achieve precise spatial positioning of 17 facial reference anatomical points, such as the tip of the nose, inner canthus of the eye, outer canthus of the eye, glabella, and lips.

[0107] Furthermore, for each located reference point, its neighborhood topological connectivity and local surface normal change rate in the facial 3D mesh model are calculated as node weights for subsequent topological structure construction, ensuring enhanced node stability in dynamic tracking scenarios.

[0108] Using a 3D point cloud connection algorithm, the 17 pre-located facial reference anatomical points are connected in an orderly manner according to anatomical proximity, generating an initial 3D geometric topology of the face containing node coordinates, edge connection relationships, and node weights, thereby realizing a structured representation of the facial spatial morphology.

[0109] For example, in the scenario of detecting 3D facial models of teenagers, for the input depth map after Gaussian filtering, the convolution kernel size of the 3D facial keypoint detection model is configured as 5×5×5, the stride is set to 2, and the batch normalization layer is used to improve the generalization ability of the training model. The feature response map output by the model is analyzed by spatial curvature to obtain 38 candidate keypoints. The ICP registration algorithm uses the 3D reference facial template as the matching object, and the iterative operation converges on the 35th time. The final output includes a set of 17 facial reference point coordinates, including the tip of the nose (coordinates: X=12.4mm, Y=3.2mm, Z=58.7mm), the inner canthus of the left eye (coordinates: X=-15.1mm, Y=7.5mm, Z=55.3mm), and the outer canthus of the right eye (coordinates: X=17.8mm, Y=8.0mm, Z=56.1mm). In the calculation of the local normal change rate, the neighborhood search radius was set to 4mm, and nodes with a normal change rate exceeding 0.85 were assigned a weight coefficient of 2 in the topology construction to enhance their stability in dynamic recognition. The final generated initial facial 3D geometric topology contains 17 nodes and 28 connecting edges, with an average node connectivity of 3.29, achieving a high-precision structured representation of the spatial morphology of adolescent faces.

[0110] S3.3: Input the initial three-dimensional geometric topology of the face into the acupoint localization reasoning engine built based on graph convolutional network (GCN), perform acupoint coordinate mapping based on the acupoint prior distribution knowledge graph, and identify the three-dimensional spatial coordinates of acupoints around the eyes such as Jingming, Zanzhu, Yuyao, Sizhukong, Tongziliao, Chengqi, and Sibai, as well as distal acupoints such as Baihui, Fengchi, Hegu, Guangming, Ganshu, Pishu, and Shenshu.

[0111] The initial facial 3D geometric topology structure output from step S3.2 is input into the acupoint localization inference engine. A graph convolutional network (GCN) based method (parameters: 3 network layers, 9 convolutional kernels, ReLU activation function, 64 node feature dimensions) is used to realize graph signal processing and spatial context encoding of the topology structure.

[0112] Furthermore, by matching the three-dimensional geometric features of each node with the probability density function of the corresponding node in the prior distribution knowledge graph of acupoints (parameters: Euclidean distance threshold 0.5mm, probability matching coefficient λ=0.7), the preliminary screening of potential acupoint candidate nodes is achieved, and a set of high-confidence candidate acupoint nodes is obtained.

[0113] Furthermore, a multi-scale feature fusion method (parameters: scale range [2mm, 5mm, 10mm], fusion strategy is weighted average) is used on the candidate acupoint node set to comprehensively encode the local geometric curvature features, skin texture gradient features and spatial relationships of adjacent nodes, and generate multimodal node feature vectors.

[0114] Furthermore, by using a similarity measurement algorithm (parameter: cosine similarity threshold 0.85), the feature vectors of multimodal nodes are matched with the feature vectors of standard acupoints in the prior distribution knowledge graph of acupoints, so as to achieve accurate identification of acupoints around the eyes such as Jingming, Zanzhu, Yuyao, Sizhukong, Tongziliao, Chengqi, and Sibai, as well as distal acupoints such as Baihui, Fengchi, Hegu, Guangming, Ganshu, Pishu, and Shenshu.

[0115] Furthermore, a three-dimensional coordinate mapping algorithm (parameters: the coordinate system transformation matrix is ​​optimized by the ICP algorithm, and the number of iterations is 50) is used to convert the three-dimensional geometric position of the identified acupoint nodes into standardized three-dimensional spatial coordinates in the facial three-dimensional model coordinate system, forming a set of acupoint coordinates that can be used for moxibustion path planning;

[0116] By using the above-mentioned graph convolutional network feature extraction, prior matching and coordinate mapping processing methods, the initial geometric topology data in step S3.2 is transformed into high-precision three-dimensional spatial coordinates of acupoints, thereby achieving the technical effect of moxibustion positioning.

[0117] For example, in an embodiment targeting adolescent users, facial RGB-D data collected by an optical positioning camera is processed by multi-scale Gaussian filtering and then input into a three-layer GCN structure. The GCN convolution kernel size is 9, and the initial feature dimension of the nodes is 64. The standard coordinates of the Jingming acupoint in the acupoint prior knowledge graph are (12.5mm, 35.2mm, 4.8mm). Candidate acupoint nodes are matched using a probability density function. Under the condition of λ=0.7 matching coefficient, the Euclidean distance of the neighboring nodes at the inner canthus of the eye is 0.42mm, which meets the threshold condition. Multi-scale feature fusion extracts the mean local curvature at three scales, which are 0.015, 0.019, and 0.022, respectively, and the texture gradient vector lengths are 1.8, 2.1, and 2.0, respectively. The weighted average yields a fused feature value curvature of 0.0181 and a texture gradient length of 1.97. The cosine similarity with the standard Jingming acupoint features is 0.89, which is higher than the set threshold of 0.85, confirming successful recognition. The coordinate system was mapped using an ICP-optimized coordinate transformation matrix with 50 iterations. The final output of the standardized three-dimensional coordinates of the Jingming acupoint in the facial model coordinate system was (12.51mm, 35.19mm, 4.81mm). When used for moxibustion path planning, the deviation of the moxibustion spot was controlled within 0.05mm, which significantly improved the positioning accuracy and path stability.

[0118] S3.4: Perform a local coordinate system transformation on the three-dimensional spatial coordinates of the acupoints to convert them into relative spatial pose parameters relative to the end effector of the moxibustion head, so as to form an initial reference coordinate system for moxibustion path planning;

[0119] S3.5: Based on the spatial displacement vectors of facial key points in a continuous frame image sequence, calculate the temporal series features of the facial key point displacement sequence, including displacement amplitude, velocity change rate and acceleration fluctuation, as the node input feature vectors of the dynamic anatomical topology map, for subsequent eye movement intention prediction model analysis in the graph neural network.

[0120] like Figure 3 As shown, step S4 involves inputting the dynamic anatomical topology map into a lightweight graph neural network, and combining it with multimodal sensing data from an infrared temperature sensor and an optional macro thermal imager to perform spatiotemporal modeling analysis. Specifically, this includes:

[0121] S4.1: Perform graph structure modeling on the displacement sequence of facial key points in the dynamic anatomical topology graph, and construct an adjacency matrix based on the spatial adjacency relationship between graph nodes to characterize the relative motion pattern of each facial region in three-dimensional space.

[0122] For the displacement sequence of facial key points in the dynamic anatomical topology map, a graph modeling method based on spatial adjacency relationship (parameters: node coordinate accuracy 0.1mm, sampling period 33ms) is used to extract the spatial correlation relationship of local facial regions.

[0123] Furthermore, by using the Euclidean distance calculation method (parameter: node pairing threshold 15mm), the spatial distance matrix between each key point pair is obtained, and according to the preset adjacency relationship determination rules, node pairs with a distance less than the threshold are marked as directly adjacent, thereby realizing the construction of the adjacency data set;

[0124] Furthermore, a weighted adjacency matrix generation algorithm (parameters: weight function type is Gaussian radial basis function, σ parameter is 5mm) is used to assign weights to the nodes that have been determined to be adjacent, so as to reflect the motion coupling strength between different regions and obtain a weighted spatial adjacency matrix;

[0125] Furthermore, by using a normalization method (parameter: row normalization mode), the sum of weights in each row of the weighted adjacency matrix is ​​transformed into a standardized ratio to ensure the comparability of weights for different nodes in network analysis.

[0126] Through the above weighting and normalization processing, the relative motion patterns of each key point on the face in three-dimensional space are represented in the form of a structured adjacency matrix, realizing the graph-structured input of facial region relationships.

[0127] For example, in the facial key point sequence input for adolescent users, node coordinates are acquired by an optical positioning camera with an accuracy of 0.1 mm and a sampling period of 33 ms. The Euclidean distance between the tip of the nose and the inner canthus of the left eye is also considered. The calculation is performed using the following formula: ,in , , For the three-dimensional coordinate components of the nose tip, , , This represents the coordinate component of the left inner canthus. When the calculated result is 14.2mm, which is less than the adjacency threshold of 15mm, the two are considered directly adjacent. In weight allocation, the Gaussian radial basis function is used to calculate the weights. The weights are 0.95, and after row normalization, the corresponding element values ​​of the adjacency matrix are 0.12. The resulting adjacency matrix clearly represents the spatial relationships of the facial nodes. When input into the downstream ST-GCN network, it can stably extract facial motion patterns and significantly improve the accuracy of posture perturbation trend prediction.

[0128] S4.2: Combine the adjacency matrix with the time series features and input them into a lightweight spatiotemporal graph convolutional network (ST-GCN) to extract the motion trend features of facial key points within a continuous time window, so as to generate a dynamic facial pose evolution map.

[0129] The adjacency matrix generated by S4.1 and the time series features calculated by S3.5 are structurally fused. Matrix concatenation and normalization operations are used (parameters: adjacency matrix size n×n, time series feature window length T) to achieve a joint representation of spatial topology and temporal dynamics.

[0130] Furthermore, a lightweight spatiotemporal graph convolutional network (ST-GCN, 3 layers, spatiotemporal separation structure of convolutional kernels) is used to perform the first layer of convolution operation on the fused input. The convolutional kernels are weighted according to the adjacency matrix structure in the spatial dimension and slide convolution according to the time series features in the temporal dimension to realize the extraction of local motion patterns of facial key points between adjacent frames and obtain the first stage of spatial-temporal feature mapping tensor.

[0131] Furthermore, through the second layer convolution operation of ST-GCN (parameters: temporal convolution kernel length is 5, spatial convolution kernel is set to 3 according to the Chebyshev polynomial approximation order), while maintaining the original spatial adjacency weights, cross-frame key point connection weights are introduced to realize long-range dependency feature mining across time windows and generate intermediate feature tensors containing multi-scale motion information.

[0132] Furthermore, through the third convolution operation of ST-GCN (parameters: output channel number is C (not less than 64), activation function is ReLU), feature denoising and nonlinear mapping are performed on the intermediate feature tensor to extract the global motion trend features of facial key points within the entire continuous time window;

[0133] The global motion trend features mentioned above are transformed into a dynamic facial pose evolution map data structure through a fully connected mapping layer, thereby realizing a visual representation of the spatial trajectory and pose changes of facial key points over time.

[0134] For example, in a practical application, the input adjacency matrix size is 20×20, corresponding to 20 identified facial key points. The temporal series features use a window of 15 frames with a frame interval of 33ms. The adjacency matrix and temporal features are Z-score normalized separately before fusion. The first layer of ST-GCN introduces direct weights of the adjacency matrix into the spatial convolution kernel, and the temporal convolution kernel has a length of 3 to capture short-term motion patterns. The second layer sets the temporal convolution kernel length to 5, and the spatial convolution kernel calculates the spatial convolution coefficients through a third-order unfolding approximation using Chebyshev polynomials, where the coefficient calculation formula is: ,in The largest eigenvalue of the graph Laplacian matrix is ​​used; the number of output channels in the third layer is set to 64, and ReLU activation is applied. The final generated dynamic facial pose evolution map accurately reflects the key pose change trends such as brow lift, eyelid closure, and slight head tilt within a continuous time window, significantly improving the response accuracy and stability of the subsequent eye-tracking intention prediction model to pose perturbations in non-cooperative states;

[0135] S4.3: Perform sliding window time-series analysis on the skin surface temperature data collected by the infrared temperature sensor, and use wavelet transform algorithm to extract the frequency domain features of temperature changes in order to obtain the local thermal response dynamic feature sequence of the moxibustion area;

[0136] S4.4: If a micro thermal imager is enabled, grayscale normalization and heat flow vector field modeling are performed on the thermal images it acquires, and the rate of change of local blood flow is calculated to obtain the quantitative thermal effect characteristics of the skin microcirculation state before and after moxibustion.

[0137] S4.5: Based on the dynamic facial posture evolution map, local thermal response dynamic feature sequence and optional thermal effect features, a cross-modal attention mechanism is modeled through a multimodal feature fusion network to generate a multimodal perception vector that integrates spatiotemporal features, so as to represent the joint representation of the user's current posture perturbation trend and moxibustion response state.

[0138] For the dynamic facial pose evolution map, local thermal response dynamic feature sequence and optional thermal effect features, a multimodal feature fusion network (parameters: cross-modal attention mechanism = enabled, feature normalization method = Z-score, number of fusion layers = 3) is used to achieve unified representation and encoding of features of different modalities.

[0139] Furthermore, through a cross-modal attention mechanism (parameters: query vector = dynamic facial pose features, key vector = thermal response and thermal effect features, attention dimension = 128), the weight allocation between features is realized, and the relevance score matrix of each modality at the current time is obtained;

[0140] Furthermore, based on the aforementioned correlation scoring matrix, a weighted aggregation algorithm (parameters: aggregation strategy = Softmax normalized weighting, weight update frequency = 50ms) is used to achieve weighted fusion of features of each modality and generate a fusion vector representing the joint feature vector of the current attitude perturbation trend and thermal response state.

[0141] Furthermore, the joint feature vector is subjected to temporal convolution processing (parameters: kernel size = 3, stride = 1, number of convolution layers = 2) to achieve temporal smoothing and local trend enhancement of features, and to generate a multimodal perception vector that integrates spatiotemporal features.

[0142] The multimodal feature set from the previous step is transformed into a multimodal perception vector that integrates spatiotemporal features through the cross-modal attention mechanism modeling and weighted aggregation algorithm, thereby achieving a joint representation of the user's current posture perturbation trend and moxibustion response state.

[0143] For example, in the dynamic facial pose evolution map input, the three-dimensional coordinate changes of 17 facial key points are used as pose feature vectors, with normalized coordinate ranges of [-1, 1]. In the local thermal response dynamic feature sequence, the sampling frequency of the temperature sequence output by the infrared temperature sensor is set to 10 Hz, with a temperature change range of 38~44℃. In the thermal effect feature, the intensity range of the heat flow vector field is 0.05~0.15 W / cm². The multimodal feature fusion network first maps the above three types of features to a 128-dimensional embedding space to generate a pose embedding matrix. Thermal response embedding matrix Thermal effect embedding matrix Each matrix has rows corresponding to time slices and columns corresponding to embedding dimensions. Subsequently, in the cross-modal attention mechanism, [the following is used]... For query, Calculate the attention weight matrix for keys and values. and The weight update period is 50ms. The weight matrix is ​​normalized using Softmax to obtain the weighted fusion vector. = Finally, in terms of the time dimension... Perform a temporal convolution operation with a kernel size of 3 and a stride of 1, and output a multimodal sensing vector of consistent length that fuses spatiotemporal features. This vector demonstrates a significant improvement in the joint representation accuracy when the system determines the user's posture perturbation trend and the state of the moxibustion thermal response, ensuring that the system can stably and continuously execute the moxibustion path control in a non-cooperative state.

[0144] Step S5: Based on the eye-tracking prior knowledge base, determine whether the current eyelid closure speed exceeds a preset threshold, and detect whether the brow is raised to determine whether the user is about to blink actively. Specifically, this includes:

[0145] S5.1: The three-dimensional coordinate sequence of facial key points acquired by the optical positioning camera is filtered and denoised to eliminate the impact of ambient light interference and image acquisition jitter on the accuracy of eyelid contour recognition, and to obtain smoothed eyelid motion trajectory data.

[0146] S5.2: Calculate the rate of change of eyelid closure velocity based on eyelid motion trajectory data, and use a sliding time window mechanism to perform differential calculation on the eyelid displacement difference of consecutive frames to extract instantaneous closure velocity features and obtain dynamic closure velocity sequence;

[0147] Based on the smoothed eyelid motion trajectory data output by S5.1, the three-dimensional coordinate difference method (parameters: sampling frame rate, Gaussian weighting factor) is used to preliminarily calculate the eyelid displacement of adjacent time frames to reflect the spatial displacement change characteristics of the eyelid per unit time.

[0148] Furthermore, a sliding time window mechanism (parameters: window length 20ms, step size 10ms) is used to perform first-order differential operations on the displacement differences of consecutive frames, and the central difference method is used to improve the robustness of velocity estimation and obtain the instantaneous closed velocity sequence.

[0149] Furthermore, based on the instantaneous closed velocity sequence, a low-pass filtering algorithm (parameter: cutoff frequency 15Hz) is used to suppress high-frequency jitter and random noise from visual sensor acquisition, so as to ensure the smoothness and accuracy of velocity change rate calculation.

[0150] Furthermore, when calculating the rate of change of speed, normalization is used (parameter: the maximum speed calibration value is taken from the 95th percentile of the resting blink speed distribution) to map the speed values ​​obtained in each time window to the [0,1] interval, so as to achieve comparability of speed characteristics among different users;

[0151] Furthermore, using the rate of change calculation formula, the average velocity difference between adjacent time windows is divided by the time window length to obtain the dynamic closing velocity change rate, where the formula is as follows:

[0152]

[0153] in, Let be the normalized average closing velocity for the i-th time window. The normalized average closing velocity of the previous time window. The length of the time window;

[0154] By sorting and storing the rate of change matrix in time series, the results are transformed into a sequence of dynamic eyelid closure speeds, achieving the expected technical effect of capturing eyelid closure trends in real time.

[0155] For example, under the output condition of an optical positioning camera with a sampling frame rate of 120fps, the three-dimensional trajectory coordinate data of the eyelids after third-order Savitzky-Golay filtering and noise reduction were acquired. The window length was set to 24 frames (approximately 200ms), and the step size was set to 12 frames (approximately 100ms). The instantaneous closing velocity was calculated using the central difference within each sliding window, and the maximum normalized velocity was found to be 0.82. The velocity difference between adjacent windows was divided by the 0.1-second window length, and the rate of change was derived from the above formula as −1.4, indicating a rapidly decreasing trend in the closing velocity. When the rate of change for three consecutive windows was less than −1.0, the dynamic closing velocity sequence successfully predicted the signal that the blink was about to end. Under the trigger of this signal, the system combined with the eye movement prior judgment module to generate a blink event prediction output, realizing dynamic path control for the moxibustion head to avoid obstacles in advance. Clinical simulation tests verified that the stability of the moxibustion spot was significantly improved under the following conditions in adolescents.

[0156] S5.3: Compare the dynamic closing speed sequence with the preset blink trigger threshold. If the closing speed exceeds the threshold in multiple consecutive time windows, it is determined that an active blinking action is about to occur, and a blink event trigger signal is generated.

[0157] S5.4: Extract features from the facial muscle displacement vectors in the glabella region, calculate the glabella elevation amplitude and angular velocity based on facial key point displacement data, in order to identify the trend of facial micro-expression changes accompanying blinking and obtain the glabella elevation state identifier.

[0158] S5.5: Perform a logical AND operation between the blink event trigger signal and the eyebrow lift status flag. If both meet the preset conditions, output the comprehensive judgment result of the blink action as the activation input basis for the path offset compensation strategy.

[0159] Using the blinking event trigger signal and the eyebrow lifting status indicator as input conditions, the logic judgment module in the central processing unit is called to perform a comprehensive blinking action judgment calculation.

[0160] The system employs digital logic and arithmetic (parameters: input A is the blink event trigger signal, and input B is the brow lift status indicator) to achieve the function of synchronously detecting the occurrence of two physiological events.

[0161] Furthermore, by modeling using Boolean algebra expressions, inputs A and B are mapped to binary variables respectively. and ,implement Calculate and generate a comprehensive judgment output. ;

[0162] Furthermore, the consistency of the logical AND operation results is verified by time window (parameters: time window length is 200ms, sampling frequency is 100Hz). A sliding window is used to count the proportion of the common high level of the two input signals within the window to ensure the stability and validity of the judgment and to obtain the comprehensive judgment status signal of blinking action.

[0163] Furthermore, the blinking action comprehensive judgment state signal is input into the state machine control module and mapped to the activation trigger bit of the path offset compensation strategy. The next control strategy call sequence is determined through the state transition table.

[0164] By using the above Boolean logic judgment and time window verification processing method, the blinking event trigger signal and the eyebrow lifting state indicator are integrated into a unique comprehensive judgment result, so as to achieve accurate identification before active blinking occurs and provide a stable trigger basis for advance planning of the safe avoidance path of the moxibustion head.

[0165] For example, in a moxibustion system for myopia prevention in teenagers, the blink event trigger threshold is set to an eyelid closure speed exceeding 35 mm / s for three consecutive frames, and the brow elevation threshold is 0.5 mm with an angular velocity exceeding 8° / s. Facial key point data from an optical positioning camera is collected and filtered. At a certain moment, both the blink event trigger signal and the brow elevation status indicator are at a high level. The logic determination module will... =1, =1 Substitute ,get =1. The time window consistency verification detected that the common high level of the two signals accounted for more than 95% within a 200ms window, outputting a stable comprehensive judgment state signal. Upon receiving this signal, the state machine control module immediately triggers the central controller to invoke the corresponding path offset compensation strategy, causing the moxibustion head to avoid the periorbital area along a safe arc. Verification results show that the avoidance process is smooth and the return to position is accurate, significantly improving the dynamic moxibustion path stability of the system in non-cooperative states.

[0166] Step S6: If it is determined that a blinking action is about to occur, further detection is performed to determine whether the head generates a continuous angular acceleration in a certain direction, so as to activate the path offset compensation strategy in the corresponding direction. Specifically, this includes:

[0167] S6.1: Based on the blinking prediction signal output by the eye-tracking prior knowledge base, start the head posture dynamic monitoring module, and perform Kalman filtering on the three-dimensional coordinate sequence of facial key points output by the optical positioning camera to eliminate noise interference in the image acquisition process and obtain smooth facial motion trajectory data.

[0168] S6.2: Perform Euler angle transformation on the smooth facial motion trajectory data to calculate the change in the current user's head posture angle in three-dimensional space, so as to obtain the real-time angular displacement sequence of the head in the three degrees of freedom of pitch, yaw and roll.

[0169] S6.3: Input the head angular displacement sequence into a sliding window integrator, and calculate the rate of change of head angular acceleration based on a sliding window mechanism with a time window length of 100ms, in order to identify whether there is a continuous head movement trend;

[0170] S6.4: Perform a threshold comparison operation on the angular acceleration change rate. If the angular acceleration of the head in a certain direction exceeds the preset safety threshold and the duration exceeds 50ms, it is determined to be a valid head disturbance signal and the path offset compensation mechanism is triggered.

[0171] S6.5: Based on the directional information of the effective head disturbance signal, retrieve the path offset compensation strategy template of the corresponding direction from the eye-movement-head motion coupling law database, and output the path adjustment command to the central controller to drive the moxibustion module to perform dynamic trajectory replanning;

[0172] Based on the direction vector of the effective head perturbation signal, a pattern retrieval algorithm (parameters: eye-movement-head motion coupling pattern database, similarity threshold 0.85, direction weight coefficient 0.75) is used to retrieve and match the pre-stored path offset compensation strategy templates in the database.

[0173] Furthermore, by using a multi-dimensional feature similarity calculation method (parameters: coupled feature space coordinate set, perturbation direction three-degree-of-freedom components), the similarity score between the current perturbation direction and the policy template direction features is realized, and the optimal matching template index result is obtained;

[0174] Furthermore, a strategy parameter parsing algorithm (parameters: template ID, pose correction coefficient set) is used to extract the spatial trajectory adjustment parameters of the moxibustion head in the matching template and generate a trajectory control parameter vector;

[0175] Furthermore, by using an instruction encoding conversion method (parameters: position adjustment instruction set, power adjustment instruction set), the mapping of the moxibustion path adjustment data format to the central controller's recognition format is realized, and an executable path adjustment control instruction sequence is obtained;

[0176] By controlling the bus data transmission protocol, the generated path adjustment control command sequence is output to the central controller to realize the dynamic trajectory replanning triggering and synchronous execution of the moxibustion module;

[0177] For example, in a dynamic moxibustion operation performed on a teenage user, the direction of the effective head disturbance signal was determined to be leftward yaw, with a peak angular acceleration of [value missing]. rad / s² and duration At milliseconds, the pattern retrieval algorithm searches the coupling pattern database with a similarity threshold of 0.85 and matches the left-biased path offset compensation template numbered T023. The directional component of this template in the coupling feature space is pitch. rad, yaw rad, roll The rad (radius) and directional feature similarity score reached 0.91, meeting the optimal matching criteria. The strategy parameter parsing algorithm extracted the trajectory adjustment parameters of the moxibustion head from template T023: spatial pose correction coefficient. Power correction factor Both are encoded into 16-bit binary instructions using a control data structure to generate a trajectory control parameter vector [1.12, 0.95]. The instruction encoding conversion method formats this parameter vector into a path adjustment instruction sequence according to the central controller's CAN bus protocol and transmits it to the central controller via the control bus. In actual execution, the central controller drives the acupuncture head to adjust in the yaw left direction according to this instruction. mm, while reducing the power of the smokeless moxibustion energy source to the rated value. This ensures that the light spot avoidance trajectory is smooth and the moxibustion temperature remains stable within the range of 42℃±2℃ during head disturbances, significantly improving the continuity and dynamic stability of the moxibustion path in non-cooperative scenarios.

[0178] Step S7: Based on the pre-stored eye-head coupling law database, the central controller generates a spatial avoidance trajectory for the moxibustion head 0.3-0.5 seconds in advance, causing the moxibustion spot to withdraw from the sensitive area along a preset safety arc. Specifically, this includes:

[0179] S7.1: Based on the historical posture perturbation patterns in the eye-tracking-head movement coupling database, extract the perturbation sequence with the highest similarity to the current user's eye-tracking behavior to identify potential head movement trends;

[0180] S7.2: Perform dynamic trajectory prediction modeling on the extracted perturbation sequence, and use the Kalman filter algorithm to estimate the pose change of the user's head in the next 0.3~0.5 seconds to generate a predicted spatial coordinate sequence;

[0181] S7.3: Based on the predicted spatial coordinate sequence and combined with the current spatial position of the moxibustion head, the avoidance path of the moxibustion head in three-dimensional space is calculated by the inverse kinematics algorithm, forming a safe avoidance trajectory with minimal disturbance.

[0182] Based on the predicted spatial coordinate sequence and the current spatial position data of the moxibustion head, an inverse kinematics algorithm (parameters: joint degrees of freedom of the moxibustion manipulator, end effector constraints, spatial pose accuracy threshold) is used to realize the joint angle calculation function between the end pose of the moxibustion head and the target avoidance position.

[0183] Furthermore, by using the joint angle iterative optimization method (parameters: Jacobian matrix solution step size 0.01, damping factor 0.001), the position error and attitude error of the end effector are minimized simultaneously, and a joint angle vector that meets the preset safety distance is obtained.

[0184] Furthermore, a path smoothing interpolation algorithm (parameters: cubic spline interpolation, trajectory sampling resolution 1ms) is used to perform time-series processing on the joint angle vector, so as to realize the smooth transition of position and posture changes of the moxibustion head during continuous movement, and generate a set of trajectory points that do not produce abrupt acceleration peaks.

[0185] Furthermore, the spatial safety of the trajectory point set is verified by an obstacle avoidance trajectory optimization algorithm (parameters: spatial constraints are obstacle avoidance radius ≥ 15mm and speed limit 120mm / s), and a safe obstacle avoidance trajectory with minimal disturbance is generated.

[0186] By using the above-mentioned inverse kinematics and trajectory optimization processing method, the predicted spatial coordinate sequence of the previous step is transformed into three-dimensional motion trajectory parameters that meet the dynamic avoidance requirements, thereby achieving safe withdrawal of the sensitive area around the eyes during the moxibustion process.

[0187] For example, when performing inverse kinematics calculations on the predicted spatial coordinate sequence, the robotic arm is set to 6 degrees of freedom, and the end effector position accuracy threshold is set to... mm, attitude accuracy threshold set to °. Damped least squares method is applied in the solution process, and the Jacobian matrix update step size is set to . The damping factor is set to To avoid the impact of singular configurations on solution stability, cubic spline interpolation is used in the trajectory generation stage to interpolate the sequence of joint angle changes. Smoothing is performed at a millisecond time resolution to ensure that the rate of change of joint acceleration does not exceed [the specified value]. mm / s². During the trajectory optimization phase, the avoidance path is checked for safe distance and speed constraints, and the avoidance radius is set to be no less than [missing value]. mm, the speed at the tip of the moxibustion head does not exceed The speed is measured in mm / s to ensure no skin contact or discomfort during the exercise. Testing showed that this trajectory planning significantly improved the smoothness and comfort of avoidance responses to blinking and slight head movements in teenagers, and the temperature fluctuation of the moxibustion spot during withdrawal from sensitive areas was controlled within [specific parameters]. Within a temperature range, the stability of the biological effects of moxibustion was ensured;

[0188] S7.4: Convert the safety avoidance trajectory into a motor drive command sequence, and use FPGA to control the servo motor group to perform spatial pose adjustment, so that the moxibustion spot can withdraw from the sensitive area around the eye along a preset arc.

[0189] S7.5: During the execution of the avoidance trajectory, the feedback data from the infrared temperature sensor is integrated in real time, and the output power of the moxibustion head is dynamically corrected using a PID adjustment mechanism to maintain the temperature of the moxibustion area within the range of 42℃±2℃, ensuring the stability of the biological effect of moxibustion.

[0190] Step S8: After the blinking action is completed, high-speed closed-loop temperature control and position servoing are implemented through FPGA to quickly return the moxibustion head to the original acupoint path, maintaining the continuity and stability of the moxibustion process. Specifically, this includes:

[0191] S8.1: Based on the judgment signal of the completion of the blinking action, the FPGA controller is triggered to enter the moxibustion head return control mode to start the closed-loop servo control process;

[0192] S8.2: Update the facial 3D model data output by the optical positioning camera in real time, extract the current facial key point coordinates, and obtain the latest anatomical topology map of the user's facial spatial pose.

[0193] S8.3: Calculate the spatial deviation vector between the current position of the moxibustion head and the original acupoint path based on the anatomical topology map, so as to generate a target pose correction command;

[0194] S8.4: Input the pose correction command into the PID control module in the FPGA to execute high-precision motor drive control, so as to realize the spatial trajectory replanning and high dynamic response displacement of the moxibustion head;

[0195] Based on the input conditions of the target pose correction command, the PID control module inside the FPGA is called (parameters: Kp is the position proportional gain, Ki is the position integral gain, and Kd is the position derivative gain) to realize closed-loop position control of the moxibustion head drive motor.

[0196] Furthermore, a position error calculation method is adopted, which calculates the difference between the target pose vector and the current position vector of the moxibustion head to generate a position error vector, and uses each component as the input signal of the PID controller to drive the three-axis servo motor to achieve real-time adjustment of the spatial trajectory.

[0197] Furthermore, the integral and differential terms of the error are calculated using discrete-time integration and differentiation algorithms, and then weighted and synthesized with the proportional term to form the control quantity for motor drive. The formula for calculating the control quantity Q is as follows:

[0198]

[0199] in, This is the current position error. The sampling interval time. This is the time-cumulative sum of the position error. This represents the change in error between adjacent sampling periods;

[0200] Furthermore, inverse kinematics calculations are performed to map the control quantity Q into torque or speed commands for each drive joint, and output to the drive motor via the FPGA's PWM signal generation module to realize multi-degree-of-freedom spatial movement of the moxibustion head.

[0201] Furthermore, during trajectory execution, the high-speed position sampling module is invoked to collect the new position vector of the moxibustion head in real time, compare it with the target pose, form new error feedback, update the input of the PID controller, and realize closed-loop iterative control of high dynamic response displacement.

[0202] By using a high-precision drive control algorithm based on PID, the pose correction command generated in the previous step is converted into a precise control signal for the drive motor, thereby realizing the dynamic replanning and rapid smooth return of the moxibustion head trajectory, ensuring that there is no significant spatial jitter in the recovery process of the moxibustion path after the posture disturbance.

[0203] For example, in an embodiment of a moxibustion device for myopia prevention and control in adolescents, the PID parameters are configured as Kp=1.2, Ki=0.05, Kd=0.01, the sampling period Δt is 5ms, and the maximum response speed of the three-axis servo motor is 300mm per second. During the return process after one blink, the errors between the target point and the current position on the X, Y, and Z axes are 2.5mm, −3.0mm, and 1.0mm, respectively. The output results of the control quantity calculated by the PID on each axis are 3.0N·m, −2.8N·m, and 1.2N·m, respectively. After inverse kinematic mapping, the above control quantity is converted into the drive pulse width signal of the three-axis motor, with duty cycles of 62%, 58%, and 45%, respectively. The moxibustion head smoothly returns to the original acupoint trajectory within 180ms, and the position deviation converges to within 0.2mm. During the process, no perceptible temperature fluctuation occurs in the moxibustion spot, the trajectory recovery is smooth, and the thermal effect is stable, which greatly improves the return accuracy and comfort in non-cooperative states.

[0204] S8.5: Combining the skin surface temperature data fed back by the infrared temperature sensor, the output power of the smokeless moxibustion energy source is adjusted in a closed loop to maintain the moxibustion temperature at 42℃±2℃ during the return of the moxibustion head.

[0205] S8.6: Through the high-speed sampling and pulse width modulation (PWM) control mechanism inside the FPGA, synchronous and coordinated control of the moxibustion head drive motor and heat source output is achieved to ensure temperature and spatial consistency during the recovery process of the moxibustion path.

[0206] Step S9: Employing a hierarchical control architecture, the upper-layer AI algorithm is triggered to perform full path replanning only when significant attitude drift is detected, thereby reducing system resource consumption and improving response efficiency. Specifically, this includes:

[0207] S9.1: Based on the prediction results of user posture perturbation trends output by multimodal sensing data and eye-tracking prior knowledge base, the current facial key point displacement sequence is dynamically analyzed to identify whether there is significant posture drift, so as to determine whether it is necessary to trigger the upper-level AI algorithm to perform full path replanning.

[0208] S9.2: Classify the identified significant attitude drifts, perform drift type matching based on the preset drift type database, extract the corresponding path replanning strategy template, and form a path adjustment plan that matches the current disturbance type.

[0209] S9.3: Input the matched path adjustment plan into the lightweight graph neural network, combine the spatial relationship between the current spatial pose of the moxibustion head and the target acupoint path, and perform path replanning calculation to generate new motion trajectory parameters of the moxibustion head;

[0210] S9.4: Based on the generated motion trajectory parameters of the moxibustion head, high-speed closed-loop position servo control is executed through the FPGA underlying controller to adjust the spatial pose of the moxibustion head in real time, so as to achieve high-precision tracking and stable transition during the path adjustment process.

[0211] S9.5: During the path adjustment process, the skin surface temperature data fed back by the infrared temperature sensor is continuously collected, and the output power of the smokeless moxibustion energy source is dynamically adjusted based on the PID adjustment algorithm to maintain the moxibustion temperature within the range of 42℃±2℃, ensuring the continuity and safety of the moxibustion effect during the path adjustment period.

[0212] Step S10: During the path adjustment process, the skin surface temperature change is continuously monitored, and the output of the smokeless moxibustion energy source is adjusted through a closed-loop feedback mechanism to ensure that the moxibustion temperature remains constant within the range of 42℃±2℃. Specifically, this includes:

[0213] S10.1: Low-pass filtering is performed on the real-time skin surface temperature data collected by the infrared temperature sensor to eliminate environmental thermal radiation interference and transient noise, and obtain a smooth and representative local skin temperature sequence.

[0214] S10.2: Based on the smooth skin temperature sequence, compare it with the preset ideal moxibustion temperature threshold of 42℃±2℃, calculate the current temperature deviation ΔT, and use it as the input error signal for closed-loop feedback control;

[0215] S10.3: The temperature deviation ΔT is processed by a proportional-integral-derivative (PID) control algorithm to generate a power adjustment command for the moxibustion energy source, thereby realizing dynamic control of the heat output of the smokeless moxibustion module.

[0216] S10.4: Based on the power adjustment command, the duty cycle of the PWM drive signal of the smokeless moxibustion energy source is controlled by the FPGA to precisely adjust the infrared radiation intensity and obtain a stable thermal output response;

[0217] S10.5: Perform secondary temperature verification on the adjusted infrared radiation thermal field, and continuously update the PID control parameters through a closed-loop feedback mechanism to maintain the temperature of the moxibustion area within a safe and effective range of 42℃±2℃, thus forming a thermal control closed loop.

[0218] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0219] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smokeless intelligent precise control method for preventing and controlling myopia in adolescents, characterized in that, The method comprises the following steps: S1: collecting eye movement data of teenagers when gazing at a fixed visual target in a relaxed state, and constructing a resting state eye movement behavior pattern library; S2: associating and modeling the eye movement behavior data with the facial key point displacement sequence to generate an eye movement-head movement coupling rule database; S3: acquiring a user's facial three-dimensional model in real time through an optical positioning camera, identifying eye and whole-body acupoint matching positions, and extracting facial key point displacement sequences as node inputs of a dynamic dissection topology graph; S4: inputting the dynamic dissection topology graph into a lightweight graph neural network, combining multi-modal sensing data of an infrared temperature sensor and an optional macro thermal imager, and performing time-space modeling analysis; S5: judging whether the current eyelid closing speed exceeds a preset threshold based on an eye movement prior knowledge base, and detecting whether the eyebrow center is raised to determine whether the user is about to perform an active blinking action; S6: if it is determined that the blinking action is about to occur, detecting whether the head produces a sustained angular acceleration in a certain direction, and activating a path offset compensation strategy in the corresponding direction; S7: according to the pre-stored eye movement-head movement coupling rule database, a central controller generates a spatial avoidance trajectory of a suspended moxibustion head in advance, so that the moxibustion light spot withdraws from the sensitive area along a preset safe arc line; S8: after the blinking action is completed, the suspended moxibustion head is quickly returned to the original acupoint path.

2. The smokeless intelligent precise control method for myopia prevention and control of adolescents according to claim 1, characterized in that, The step S8 further comprises: S9: using a hierarchical control architecture, when a significant posture drift is detected, triggering an upper-layer AI algorithm to perform full-path re-planning; S10: during the path adjustment process, continuously monitoring the skin surface temperature change, and adjusting the output of the smokeless moxibustion energy source through a closed-loop feedback mechanism.

3. The smokeless intelligent precise control method for myopia prevention and control of adolescents according to claim 1, characterized in that, The step S1 specifically comprises: Based on an eye movement tracking system and an electromyographic signal acquisition device, raw eye movement data of teenagers when gazing at a fixed visual target in a relaxed state is synchronously acquired, and an original eye movement behavior feature data set is generated; The original eye movement behavior feature data set is subjected to time-frequency domain joint filtering processing, a wavelet transform algorithm is used to remove electrooculogram interference and electromyographic noise, and a clean eye movement event sequence is extracted as an eye movement feature parameter set; A multi-channel surface electromyographic signal acquisition device is used to acquire electromyographic activity data of facial key muscle groups, an adaptive filtering algorithm is used to denoise and normalize the electromyographic activity data, and facial muscle activity feature parameters are extracted; The eye movement feature parameter set and the facial muscle activity feature parameters are subjected to multi-modal fusion modeling, a dynamic time warping algorithm is used to time-align the eye movement and facial muscle movement data, and an eye movement-facial muscle movement coupling feature vector is generated; Based on the eye movement-facial muscle movement coupling feature vector, a Gaussian mixture model is used to perform cluster analysis on the eye movement behavior patterns of teenagers in a relaxed state, identify and divide resting state eye movement behavior pattern sub-classes, and construct a resting state eye movement behavior pattern library.

4. The smokeless intelligent precise control method for myopia prevention and control of adolescents according to claim 3, characterized in that, The raw eye movement data comprises a blinking frequency, a saccade amplitude, and a micro-tremor rhythm.

5. The smokeless intelligent precision control method for myopia prevention and control of adolescents according to claim 1, characterized in that, The step S2 specifically comprises: The collected eye movement data and facial key point displacement sequences are subjected to time alignment processing; Based on the dynamic time warping algorithm, the eye movement behavior characteristics and the face key point trajectory are matched, and the time sequence coupling relationship characteristics between blinking, saccade and head movement are extracted; The nonlinear dependence relationship between eye movement parameters and facial displacement vectors is modeled by using mutual information analysis method, and an eye movement-head movement coupling strength matrix is generated; The eye movement-head movement coupling strength matrix is processed by dimension reduction using principal component analysis, the key coupling mode feature vector is extracted, and a low-dimensional eye movement-head movement coupling feature space is constructed; Based on the low-dimensional eye movement-head movement coupling feature space, a long short-term memory network model is trained to establish a prediction mapping relationship of eye movement behavior sequence to head displacement trend, and an eye movement guided head movement prediction model is generated; The head movement prediction model is jointly trained and optimized with the three-dimensional face key point coordinate sequence to form an eye movement prior knowledge base that can be used for predicting user posture disturbance trend in a non-cooperative state.

6. The smokeless intelligent precision control method for myopia prevention and control of adolescents according to claim 1, characterized in that, The step S3 specifically comprises: Performing multi-scale Gaussian filtering preprocessing on the user face RGB-D image data collected by the optical positioning camera to obtain a smoothed face depth map; Based on the pre-trained 3D face key point detection model, feature extraction is performed on the smoothed face depth map to identify and locate the facial reference anatomical points and generate an initial three-dimensional geometric topology structure of the face; The initial three-dimensional geometric topology structure of the face is input into an acupoint positioning reasoning engine based on a graph convolution network, acupoint coordinate mapping is performed based on an acupoint prior distribution knowledge graph, and the three-dimensional spatial coordinates of the periorbital acupoints and the distal matching acupoints are identified; The three-dimensional spatial coordinates of the acupoints are converted into relative spatial pose parameters relative to the end effector of the hanging moxibustion head through local coordinate system transformation, and an initial reference coordinate system for moxibustion therapy path planning is formed; Based on the spatial displacement vectors of the face key points in the continuous frame image sequence, the time sequence characteristics of the face key point displacement sequence are calculated as the node input feature vector of the dynamic anatomical topology graph.

7. The smokeless intelligent precision control method for myopia prevention and control of adolescents according to claim 6, characterized in that, The acupoint positioning reasoning engine adopts a three-layer graph convolution network structure with a convolution kernel size of 9 and a node feature dimension of 64, and candidate nodes are screened in combination with the acupoint prior distribution knowledge graph, with a similarity cosine threshold of 0.

85. The standard acupoint three-dimensional coordinates are optimized and generated by ICP coordinate mapping algorithm for moxibustion therapy path planning.

8. The smokeless intelligent precision control method for myopia prevention and control of adolescents according to claim 1, characterized in that, The step S4 specifically comprises: Modeling the face key point displacement sequence in the dynamic anatomical topology graph based on the spatial adjacency relationship between the graph nodes to construct an adjacency matrix; The adjacency matrix and the time sequence characteristics are combined and input into a lightweight spatio-temporal graph convolution network to extract the motion trend characteristics of the face key points within a continuous time window and generate a dynamic face posture evolution graph; Performing sliding window time series analysis on the skin surface temperature data collected by the infrared temperature sensor, and extracting the frequency domain characteristics of temperature change by using wavelet transform algorithm to obtain the local thermal response dynamic characteristic sequence of the moxibustion area; If a macro close-up thermal imager is enabled, the thermal imaging image collected by the macro close-up thermal imager is subjected to grayscale normalization and thermal flow vector field modeling to calculate the local blood flow rate of change and obtain the quantitative thermal effect characteristics of the skin microcirculation state before and after moxibustion therapy. Based on the dynamic facial posture evolution map, the local thermal response dynamic feature sequence and the optional quantified thermal effect feature, a multi-modal feature fusion network is used to model the cross-modal attention mechanism to generate a multi-modal perception vector with fused space-time features.

9. The smokeless intelligent precision control method for myopia prevention and control of adolescents according to claim 8, characterized in that, The space-time graph convolution network parameters include a spatial convolution kernel with an approximation order of 3 according to Chebyshev polynomial, a time domain convolution kernel length of 5, and an output channel of no less than 64.

10. The smokeless intelligent precision control method for myopia prevention and control of adolescents according to claim 1, characterized in that, The step S5 specifically includes: Filtering and denoising the facial key point three-dimensional coordinate sequence obtained by the optical positioning camera to obtain smoothed eyelid motion trajectory data; Calculating the eyelid closure speed change rate based on the smoothed eyelid motion trajectory data, performing differential operation on the eyelid displacement difference of consecutive frames using a sliding time window mechanism, extracting instantaneous closure speed features, and obtaining a dynamic closure speed sequence; Comparing the dynamic closure speed sequence with a preset blink trigger threshold, if the closure speed exceeds the threshold in consecutive time windows, it is determined that an active blink action is about to occur, and a blink event trigger signal is generated; Extracting features of the facial muscle displacement vector in the eyebrow center region, calculating the eyebrow center lifting amplitude and angular velocity based on the facial key point displacement data, identifying the facial micro-expression change trend accompanied by blinking, and obtaining an eyebrow center lifting state identifier; Performing logical AND operation on the blink event trigger signal and the eyebrow center lifting state identifier, if both satisfy the preset condition, outputting a blink action comprehensive determination result as an activation input for the path deviation compensation strategy.

Images