Monitoring and early warning system for preventing juvenile myopia

Abstract

The invention relates to a monitoring and early warning system for preventing teenager myopia, which captures teenager eye using behavior characteristics in real time in a dynamic environment and accurately recognizes an abnormal state through cooperative work of multiple modules, and eliminates space-time dislocation interference among equipment by using three-dimensional topology modeling, so that the system can be used for preventing teenager myopia. The double-frequency attention mechanism effectively separates eye physiological signals and environmental noise, the electromagnetic constraint noise generation technology restores a real interference scene, the gradient coupling mechanism continuously optimizes the equipment relation network, finally high-precision eye use behavior recognition is achieved, early warning is conducted on myopia inducements such as short-distance eye use overtime and illumination discomfort, and the myopia identification accuracy is improved. Through a personalized intervention scheme, the false alarm rate is remarkably reduced, the user compliance is improved, the myopia development speed of teenagers is effectively delayed, and a'monitoring-early warning-intervention 'closed-loop prevention and control system is formed.

Description

Technical Field

[0001] This invention relates to the field of myopia prevention, and more specifically, to a monitoring and early warning system for preventing myopia in adolescents. Background Technology

[0002] With the development of smart hardware, smart desk lamps, smart bracelets, and other health monitoring devices are becoming increasingly popular among students, especially in learning and working environments. For students, the environment in which they use smart desk lamps is often dynamic and changing. For example, students may study on public transportation, where unstable factors such as vehicle movement or pedestrian activity can affect the lamp's stability. Similarly, in the home environment, students may study in open spaces like the living room, where surrounding activities can also cause environmental changes. These dynamic environments can lead to instability in the device's sensor data, thus affecting the accuracy of behavior recognition. Smart desk lamps, smart bracelets, and other devices often integrate multiple sensors, most commonly accelerometers and ambient light sensors, used to monitor changes in user posture and light levels. However, these sensors are susceptible to various interferences in complex usage environments. For instance, accelerometers are easily affected by external environmental factors such as vehicle vibrations when users actively adjust their posture, leading to inaccurate judgments. Ambient light sensors are easily affected by momentary occlusion (such as a user's arm blocking the light), causing fluctuations in light intensity and potentially misinterpreting it as poor eye use.

[0003] Currently, existing intelligent learning assistance systems often employ simple threshold segmentation or weighted averaging algorithms when processing multimodal sensor data. These methods frequently fail to consider the spatiotemporal correlations between different sensors. The collaboration between accelerometers and ambient light sensors is not effectively utilized, resulting in the system's inability to accurately identify students' actual learning postures and poor eye habits in dynamic environments. Accelerometers are prone to data interference because they cannot distinguish between students' active posture adjustments and passive swaying caused by external environmental factors such as vehicle movement. Ambient light sensors are susceptible to erroneous judgments when light levels change suddenly, such as when an arm blocks the light source of a desk lamp, leading to a rapid change in light intensity. This is due to poor eye-use behavior; such false alarms seriously affect the accuracy and reliability of the system; furthermore, existing technologies have failed to achieve dynamic sensor data fusion and interference suppression; due to the lack of in-depth analysis of spatiotemporal correlation in traditional algorithms, they cannot dynamically adjust weights based on the reliability of different sensors at different time points, thus failing to effectively filter instantaneous noise; frequent false alarms will lead to a decline in user trust, and may even cause alarm fatigue among users, ultimately leading to the abandonment of the system; therefore, how to design an intelligent, multimodal collaborative sensing architecture that can analyze and dynamically adjust sensor weights in real time and effectively eliminate environmental interference has become a key problem that urgently needs to be solved in current technology. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a monitoring and early warning system for preventing myopia in adolescents. Through the collaborative work of multiple modules, it captures the characteristics of adolescents' eye-use behavior in a dynamic environment in real time and accurately identifies abnormal states, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a topology construction module, an interference separation module, a noise generation module, and an optimization and coordination module, wherein... Topology construction module: Responding to real-time data streams collected from multiple sources, based on the three-dimensional spatial positions of the wristband, desk lamp, and vision device and the signal frequency domain basis, it creates a time-frequency-space three-dimensional dynamic coupled topology and generates a hypergraph adjacency matrix containing spatial attenuation factors and frequency domain coupling factors. Interference separation module: Based on the hypergraph adjacency matrix, it executes a dual-frequency branch attention mechanism, extracts eye micro-behavioral features through the physiological rhythm branch, and analyzes light intensity gradient changes through the noise branch, thus separating effective behavioral signals from interference noise; Noise generation module: It uses electromagnetic field dynamic equations to constrain the adversarial generative network and synthesizes frequency domain consistent noise based on wristband magnetometer data and device distance threshold; Optimize the collaboration module: update the hypergraph adjacency matrix weights through joint gradient backpropagation, and drive device topology reconstruction by decoupling the loss function and adversarial gradient coupling calculation.

[0006] In a preferred embodiment, the specific operation of creating a time-frequency-space three-dimensional dynamically coupled topology in the topology construction module is as follows: A1. Construct a Euclidean distance matrix based on the real-time three-dimensional coordinate data reported by smart bracelets, smart desk lamps and smart vision testing devices, calculate the spatial straight-line distance between each pair of devices, and combine the distance values ​​between each pair of devices into a spatial position feature vector, while extracting the maximum distance value as a normalization benchmark. A2. Perform discrete cosine transform on the original sensor signal, retain the frequency domain coefficients of the preset number of main frequencies to form a frequency domain basis matrix; dynamically adjust the spatial scale coefficient of the Gaussian kernel function through environmental interference level parameters, and calculate the spatial attenuation factor between devices. A3. Accumulate the absolute value difference of the frequency domain basis matrix coefficients of each pair of devices, add the zero-prevention division constant, and take the reciprocal to generate the frequency domain coupling factor. A4. Multiply the spatial attenuation factor matrix and the frequency domain coupling factor matrix element by element to obtain the initial hypergraph adjacency matrix.

[0007] In a preferred embodiment, the update process of the dynamically coupled topology is as follows: The continuous time stream is segmented according to the time window sliding step size parameter set by the system, and the historical adjacency matrix state is weighted by the topological inertia coefficient. The new arrival signal data within the sliding window is recursively processed using the aforementioned distance matrix construction process and frequency domain basis matrix generation process to obtain the current adjacency matrix state. The historical weights and the current state are then combined using a weighted fusion formula to generate an updated hypergraph adjacency matrix.

[0008] In a preferred embodiment, the dynamic adjustment process of the spatial scale coefficient is as follows: The acceleration variance collected by the vibration sensor of the smart bracelet is used as the environmental interference level parameter; the preset static scale benchmark value is used as the base value, and the base value is multiplied by one plus the sum of 0.5 times the environmental interference level parameter to obtain the calculation result of the current spatial scale coefficient.

[0009] In a preferred embodiment, the specific process of executing the dual-frequency branch attention mechanism in the interference separation module is as follows: Based on the hypergraph adjacency matrix output by the topology construction module, the spatial relationships between device nodes are first spatially enhanced and reconstructed to generate spatial spectrum feature vectors. These spatial spectrum features are then combined with the original signals of eye microbehavior and light intensity gradient sequences. In the physiological rhythm branch, the principal component features of the physiological rhythm are extracted using a time-frequency transformation algorithm, and in the noise branch, the light intensity gradient change features are analyzed using a recursive filtering algorithm. Subsequently, dual-channel attention weight coefficients are generated, corresponding to the intensity distributions of the physiological rhythm branch and the noise branch, respectively. These dual-channel attention weight coefficients are used to weight the original signal, initially separating the effective behavioral signal components and the interference noise components.

[0010] In a preferred embodiment, the specific operation of separating the effective behavioral signal from the interference noise is as follows: Based on the aforementioned dual-channel attention weight coefficients and the synthetic adversarial noise provided by the noise generation module, a gradient-driven interference separation operation is performed. The interference noise component is suppressed by a dynamic noise attenuation operator, while the effective behavioral signal component is enhanced by a residual masking mechanism. This mechanism calculates the visual load influence factor between the original signal and the initially separated signal. This influence factor is quantized by the spatial spectral entropy function and adjusted by a temperature coefficient. Finally, the enhanced behavioral feature signal is output.

[0011] In a preferred embodiment, the specific operation of constraining the adversarial generation network using the electromagnetic field dynamic equation in the noise generation module is as follows: Based on the time-series data of raw magnetic induction intensity collected by the wristband magnetometer, the acceleration dimension of the rate of change of the magnetic field is calculated as a physical constraint benchmark parameter. Combined with the maximum Euclidean distance threshold between devices output by the topology construction module, a dynamic gradient upper bound constraint formula is constructed. This formula stipulates that the upper limit of the time-varying gradient change rate of the synthesized noise signal is proportional to the ratio of the magnetic field acceleration amplitude to the distance threshold. This constraint condition is used as the gradient regularization term of the generative adversarial network generator. At the same time, the aforementioned effective signal frequency domain basis is used as the condition vector input to guide the noise spectrum structure to be consistent with the real eye-use behavior environment. The magnetic field change acceleration is obtained by calculating the second derivative.

[0012] In a preferred embodiment, the synthesis operation of the frequency domain uniform noise is as follows: The noise spectrum is orthogonally expanded using Legendre orthogonal polynomial basis functions. Conditional vector concatenation of magnetic field constraint parameters and effective signal frequency domain basis are used as generator inputs. A preliminary noise time-domain sequence is generated by learning Legendre expansion coefficients through a feedforward network. An adversarial loss function based on Burg spectral entropy distance is constructed to measure the frequency domain similarity between synthetic noise and labeled real environmental interference noise. This loss function maximizes the spectral distribution similarity between synthetic noise and real noise while minimizing the difference in their autoregressive model residuals. Finally, the generator parameters are iteratively adjusted through an adversarial training optimizer to ensure that the Burg spectral entropy distance between the output noise and the real interference signal in the frequency domain is less than a set threshold. The Burg spectral entropy distance is defined as the weighted difference between the logarithmic values ​​of the synthetic noise spectral energy and the real noise spectral energy, minus the normal distribution entropy benchmark value.

[0013] In a preferred embodiment, the specific operation of updating the hypergraph adjacency matrix weights through joint gradient backpropagation in the optimization collaboration module is as follows: Based on the decoupling loss function output by the interference separation module and the adversarial gradient data provided by the noise generation module, a cross-mode gradient tensor is constructed. This tensor is generated by the outer product of the second-order Hessian matrix of the adversarial gradient with the partial derivative of the decoupling loss with respect to the topological adjacency matrix. The gradient tensor is normalized by Riemannian manifold constraints and an exponential decay term is superimposed. The regularized gradient tensor is mapped to the tangent space of the special orthogonal group Lie algebra, and the topological matrix weights are updated by spinor exponential mapping.

[0014] In a preferred embodiment, the decoupling loss function and adversarial gradient coupling calculation driving device topology reconstruction includes: calculating the Ricci curvature scalar of the topological adjacency matrix based on the spatial spectral feature vector; performing element-wise multiplication of the curvature scalar with the negative exponential term of the spatial feature gradient to generate a curvature gating mask matrix; performing a tensor shrinking operation on the updated topology matrix to synthesize the topology weights with the gating mask and the preset device basis vectors by dot product; and finally achieving parameter stability control of the topology matrix through spectral radius constraints and spatial feature infinite norm scaling to output the reconstructed device relationship topology.

[0015] The beneficial effects of this invention are as follows: This system, through the collaborative work of multiple modules, captures the characteristics of adolescents' eye-use behavior in real time in a dynamic environment and accurately identifies abnormal states. The system utilizes three-dimensional topology modeling to eliminate spatiotemporal misalignment interference between devices, a dual-frequency attention mechanism to effectively separate physiological signals of the eyes from environmental noise, electromagnetic constraint noise generation technology to restore real interference scenarios, and a gradient coupling mechanism to continuously optimize the device relationship network. Ultimately, it achieves high-precision eye-use behavior recognition, provides early warnings for myopia-inducing factors such as excessive close-range eye use and unsuitable lighting, significantly reduces the false alarm rate through personalized intervention programs, improves user compliance, effectively slows down the progression of myopia in adolescents, and forms a closed-loop prevention and control system of "monitoring-early warning-intervention". Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] This embodiment provides, for example Figure 1-2 The system shown is a monitoring and early warning system for preventing myopia in adolescents, specifically including: a topology construction module, an interference separation module, a noise generation module, and an optimization and coordination module. Topology Construction Module: Responding to real-time data streams collected from multiple sources, this module creates a time-frequency-space three-dimensional dynamic coupled topology based on the three-dimensional spatial positions of wristbands, desk lamps, and vision devices and the signal frequency domain basis. It generates a hypergraph adjacency matrix containing spatial attenuation factors and frequency domain coupling factors. To address the misjudgment of eye behavior caused by the spatiotemporal misalignment of data from multiple sources in dynamic environments, this module establishes a time-frequency-space three-dimensional dynamic coupled topology and achieves accurate modeling of device relationships through the synergy of physical constraints and signal transformation. Interference Separation Module: Based on the hypergraph adjacency matrix, a dual-frequency branch attention mechanism is implemented. The physiological rhythm branch extracts micro-behavioral features of the eyes, and the noise branch analyzes the changes in light intensity gradient, separating the effective behavioral signal from the interference noise. In order to overcome the misjudgment of eye behavior caused by sensor signal coupling in dynamic scenes, this module designs a dual-frequency branch attention mechanism to achieve high-precision behavior recognition through cross-domain decoupling of biological rhythm and physical noise. Noise generation module: This module uses electromagnetic field dynamic equations to constrain adversarial generative networks and synthesizes frequency-domain consistent noise based on wristband magnetometer data and device distance thresholds. To solve the frequency-domain coupling problem between electromagnetic interference and behavioral signals in dynamic environments, this module uses an adversarial generative architecture constrained by electromagnetic fields and guides the synthesis of frequency-domain consistent noise through physical equations, providing a noise basis that matches the real environment for interference separation. The optimization and collaboration module updates the hypergraph adjacency matrix weights through joint gradient backpropagation. By decoupling the loss function and adversarial gradient coupling calculation, it drives device topology reconstruction. To achieve closed-loop optimization of dynamic topology for multiple devices, this module utilizes a gradient coupling-tensor reparameterization mechanism. Through joint backpropagation, it feeds back the behavior recognition error to the device relationship topology, breaking through the iterative optimization bottleneck of traditional static graph models.

[0021] In this embodiment, the specific operation of creating a time-frequency-space three-dimensional dynamically coupled topology is as follows: A1. Real-time three-dimensional coordinate data reported by smart bracelets, smart desk lamps, and smart vision testing devices. ( Represents the three-dimensional coordinates of the smart bracelet (e.g., [0.2, 1.0, 0.5] in meters). Represents the three-dimensional coordinates of the smart desk lamp. Constructing a Euclidean distance matrix (representing the three-dimensional coordinates of the intelligent vision testing device) Calculate the straight-line distance between any two devices, and combine the distance values ​​between any two devices into a spatial position feature vector, while extracting the maximum distance value. ( The maximum distance between devices is used as the normalization benchmark. A2. Regarding the original sensor signal ( This indicates the real-time signal of the smart bracelet (data collected by the bracelet from a six-axis sensor (accelerometer + gyroscope) at time t). This represents the real-time signal of the smart desk lamp (ambient illuminance and spectral data of the desk lamp at time t). This indicates the real-time signal from the intelligent vision testing device (micro-behavioral features of the eye captured by the embedded camera, such as blinking frequency). Indicates the starting point of the time window (the starting timestamp of the current data processing cycle). The discrete cosine transform (DCT) is performed on the signal processing time window length (the duration of a single signal processing iteration (matching the device sampling rate)), which ranges from 5 to 10 seconds. The formula is as follows: ; in, Indicates that device i is in the first... Transformation coefficients of each frequency component Indicates the first in the window Number of sampling points (total number of points = ), Indicates that device i is in signal values ​​and {Smart bracelets, smart desk lamps, and smart vision testing devices} The frequency index (low frequency → high frequency) represents the discrete cosine transform and k represents the number of main frequencies to be retained and (Dimensionality reduction constraint) The frequency domain coefficients of the preset number of main frequencies are retained to form the frequency domain basis matrix. ( (Represents the frequency domain basis vector of the smart bracelet device); through environmental interference level parameters. (0 = no interference; 1 = severe interference) Dynamically adjust the spatial scaling coefficient of the Gaussian kernel function to calculate the spatial attenuation factor between devices. The calculation formula is as follows: ; Where D represents the Euclidean distance matrix between devices (a real symmetric matrix). The dynamic spatial scale coefficient is represented by the following calculation method: , This represents the static scale reference value (its value range is >0). Indicates the level of environmental interference (a real number in the interval [0,1]); A3. The absolute value differences of the frequency domain basis matrix coefficients of each pair of devices are accumulated, and the reciprocal of the sum of these differences, after superimposing the zero-prevention division constant, is used to generate the frequency domain coupling factor. The formula is as follows: ; in, Indicates that device i is in the first... The frequency band coefficient, k, represents the number of main frequencies retained (its value range is...). (signal length) express Norm (sum of absolute values) This represents the division-to-zero protection constant (its value is set to...). ); A4. The initial hypergraph adjacency matrix is ​​obtained by element-wise multiplying the spatial attenuation factor matrix and the frequency domain coupling factor matrix. Its expression is: ; in, It represents the Hadamah accumulation. This represents the hypergraph adjacency matrix; the elements of the Euclidean distance matrix are composed of the straight-line distances between devices, and the maximum distance value is used for distance normalization in subsequent calculations; the spatial location feature vector is composed of the distance between the wristband and the desk lamp. Wristband - Vision device distance Distance between desk lamp and vision-enhancing equipment Arranged in sequence ; The update process of the dynamically coupled topology is as follows: Based on the system-defined time window sliding step parameters Segmenting continuous time streams, using topological inertia coefficients to weight the historical adjacency matrix state. The newly arrived signal data within the sliding window is recursively processed using the aforementioned distance matrix construction process and frequency domain basis matrix generation process to obtain the current adjacency matrix state. The historical weights and the current state are then weighted and fused using a formula to generate an updated hypergraph adjacency matrix, the formula of which is: ; in, Represents the state of the historical adjacency matrix. This represents the updated hypergraph adjacency matrix. This represents the time window sliding step size parameter. This represents the topological inertia coefficient (historical state weight). This represents a recursive computation function. Indicates newly arrived signal data; The dynamic adjustment process of the spatial scale coefficient is as follows: The acceleration variance collected by the vibration sensor of the smart bracelet is read as the environmental interference level parameter; the preset static scale benchmark value is used as the base value, and the base value is multiplied by one and then added to the sum of 0.5 times the environmental interference level parameter as the calculation result of the current spatial scale coefficient. The environmental interference level parameter takes values ​​from zero to one and is obtained by linear mapping from the variance of the bracelet acceleration data.

[0022] In this embodiment, it is specifically necessary to explain the interference separation module. The specific process of executing the dual-frequency branch attention mechanism is as follows: Hypergraph adjacency matrix based on the output of the topology construction module First, spatial enhancement and reconstruction of the relationships between device nodes are performed to generate spatial spectral feature vectors. ; in, The symbol represents the original multimodal signal, and * represents the graph convolution operator (spatial aggregation). This represents the SELU activation function. This represents the original multimodal signal set (including timing data from smart bracelets, smart desk lamps, and smart vision testing devices), i.e. , Represents the spatial spectral eigenvector, and (d represents the feature dimension); this spatial spectral feature Original signals of eye microbehavior (Data from intelligent vision testing devices, including changes in blink frequency) Dynamic fluctuation data of pupil diameter and light intensity gradient sequence In combination, the principal component features of the physiological rhythm are extracted from the physiological rhythm branch using a time-frequency transformation algorithm, and the formula is as follows: ; in, This represents the updated hypergraph adjacency matrix. The graph convolution can be trained using a parameter matrix. Represents the normalized exponential function, Represents the normalized graph convolution kernel; The formula for the branching of physiological rhythms is: ; in, This represents the attention weight of the physiological rhythm, with a value range of [value range missing]. (Weight allocation for three device nodes), U represents the physiological branch weight matrix. Represents the spatial spectral eigenvectors. This represents the vector concatenation operator. This represents discrete wavelet transform (extracting the 0.5-5Hz physiological frequency band). This represents the micro-behavioral signal of the eye, with a sampling rate ≥100Hz (blinking / pupil oscillation). tanh represents the hyperbolic tangent activation function (output normalized to (−1,1)). The light intensity gradient change characteristics are analyzed in the noise branch using a recursive filtering algorithm. The noise branch formula is: ; in, This represents the noise attention weight, with a value range of [value range missing]. V represents the noise branch weight matrix. This represents a residual recursive filter. The light intensity gradient sequence (rate of change of ambient illuminance) is calculated as follows: , Represents the element-multiplication operator. This represents the sigmoid activation function (output compressed to the [0,1] interval); thus generating the dual-channel attention weight coefficients. These correspond to the intensity distributions of the physiological rhythm branch and the noise branch, respectively. The dual-channel attention weight coefficients are used to weight the original signal, initially separating the effective behavioral signal component and the interference noise component, ensuring that the effective behavioral signal is enhanced and recognized in subsequent operations. The specific steps for separating effective behavioral signals from interference noise are as follows: Based on the aforementioned dual-channel attention weight coefficients and the synthetic adversarial noise provided by the noise generation module, a gradient-driven interference separation operation is performed. The interference noise component is suppressed by a dynamic noise attenuation operator, while the effective behavioral signal component is enhanced by a residual masking mechanism, the expression of which is: ; in, This represents the attention weight of the circadian rhythm branch, with a value range of [0,1]. This represents the attention weight for the noise branch, with values ​​ranging from [0,1]. This represents synthetic adversarial noise (multidimensional temporal noise signal). It represents the original signal set of multiple devices (three-dimensional time-series signal matrix). This indicates channel-weighted reinforcement (amplifying physiological behavioral signals according to weights), and its calculation method is as follows: , This represents element-level noise suppression (attenuating interference based on the rate of change of noise weights), and its calculation method is as follows: , The rate of change in attention (reflecting the dynamic intensity of sudden disturbances) is calculated as follows: , This represents the instantaneous rate of change of the noise attention weight (a larger value indicates a stronger sudden disturbance). This represents a gradient stability constant (to prevent system crashes caused by a denominator of zero), and its value is set to... , This indicates a pure behavioral signal (1. preserving visual behavior characteristics (such as changes in viewing distance and pupil fluctuations), 2. filtering out interference such as arm occlusion / sudden changes in light); this mechanism calculates the visual load influence factor between the original signal and the initially separated signal. The formula for calculating the visual load influence factor is: ; in, This represents the raw multimodal signal set (raw sensor data streams from smart bracelets, smart desk lamps, and smart vision testing devices). This represents the initially separated pure behavioral signal (the effective behavioral signal after separation by gradient-driven interference). express Norm operators are used to calculate the difference in magnitude between the original signal and the separated signal. This represents the spatial spectral entropy function, used to quantify device association features. Information complexity, This represents the temperature coefficient, used to adjust the residual strengthening intensity (the larger the value, the more abrupt changes are preserved). The visual load influence factor (used to control the weight of the residual signal, specifically: Λ≈1 (strong interference) → retain more original signal features, Λ≈0 (weak interference) → use a pure signal) is quantized by the spatial spectral entropy function and adjusted by the temperature coefficient; the final output is the enhanced behavioral feature signal, whose expression is: ; in, This represents the element-wise multiplication operator, used to achieve weighted fusion of residual signals. This indicates that the final enhanced behavioral characteristics are preserved; the signal retains the characteristics of behavioral mutations completely through residual weighted fusion, while thoroughly filtering out sudden interference components, thus achieving high-precision eye behavior recognition.

[0023] In this embodiment, the noise generation module specifically describes the operation of the adversarial generation network constrained by the electromagnetic field dynamic equation as follows: Based on the raw magnetic flux density time series data collected by the wristband magnetometer The acceleration dimension of the rate of change of its magnetic field is calculated as the physical constraint reference parameter; combined with the maximum Euclidean distance threshold D between devices output by the topology construction module, a dynamic gradient upper bound constraint formula is constructed, which is: ; Where k represents the electromagnetic coupling coefficient (the value range is set to 0.01-0.1, used to adjust the upper bound of the noise gradient). The noise signal to be generated is represented by the formula, which specifies that the upper limit of the time-varying gradient rate of the synthesized noise signal is proportional to the ratio of the magnetic field acceleration amplitude to the distance threshold. This constraint is used as the gradient regularization term of the generative adversarial network generator. This ensures that the output noise signal satisfies the physical laws of the electromagnetic field dynamic equation; simultaneously, the aforementioned effective signal frequency domain basis is used as the condition vector input to guide the noise spectrum structure to remain consistent with the actual eye-use behavior environment. The acceleration due to magnetic field change is obtained through second derivative calculation, and its calculation formula is as follows: ; in, Indicates the acceleration due to changes in the magnetic field; The synthesis operation for frequency domain uniform noise is as follows: The noise spectrum is orthogonally expanded using Legendre orthogonal polynomial basis functions. The generator input is a condition vector concatenated with magnetic field constraint parameters and the effective signal frequency domain basis. Its expression is: ; in, This represents the effective signal frequency domain basis (the eye behavior feature signal input to the generator). This represents the upper bound constraint on the gradient rate of change (constraining the time-varying characteristics of noise in accordance with electromagnetic laws), where t represents the timestamp used to identify the current processing moment, and its value range is... k represents the frequency domain basis dimension ( (Minimum number of dominant frequencies); Generate an initial noisy time-domain sequence by learning Legendre expansion coefficients through a feedforward network. Its expression is: ; in, This represents a synthetic noise sequence used to simulate real electromagnetic interference. This represents the generator weighting coefficients, used to control the spectral energy distribution of Legendre. This represents an nth-order Legendre polynomial used to analyze the frequency domain structure of control noise, and its value range is... , This represents the frequency point number, used to identify the discrete frequency index, and its value range is... , This represents the maximum frequency point, used to limit the range of noise spectrum analysis. Its calculation method is as follows: N represents the Legendre polynomial order, used to determine the accuracy of the spectrum fitting, and its value ranges from 1 to 2. (Overfitting suppression) Construct an adversarial loss function based on Burg spectral entropy distance, the expression of which is: ; in, The DCT frequency domain coefficients representing the synthesized noise (i.e., the synthesized noise at a given frequency). The energy amplitude is used to quantify the energy distribution of the noise spectrum. The DCT frequency domain coefficients representing the real noise (i.e., the real interference in the calibration dataset at the specified frequency points). energy amplitude), This represents the spectral entropy balance factor, used to adjust the weighting of real / synthetic noise. ), Indicates the frequency point index, used for traversing the discrete spectrum ( ), The theoretical entropy value representing the normal distribution is used as a baseline correction term to make... The range of values ​​is normalized, and K represents the total number of frequency points, which is used to control the spectral resolution (usually half the number of signal sampling points). The Burg spectral entropy distance represents the difference in frequency domain distribution between synthetic and real noise (a smaller value indicates greater similarity in the frequency domain). It measures the frequency domain similarity between synthetic noise and labeled real environmental interference noise. This loss function maximizes the spectral distribution similarity between synthetic and real noise and minimizes the difference in their autoregressive model residuals. Finally, the generator parameters are iteratively adjusted through an adversarial training optimizer, and its expression is: ; in, This represents a discriminator network used to evaluate the frequency domain realism of noise. (representing network parameters) This represents a generator network used to synthesize physically constrained noise. (representing network parameters) Let z represent real interference noise data (from an electromagnetic interference calibration dataset), and z represent the conditional input vector. This causes output noise In terms of frequency domain characteristics, the Burg spectral entropy distance from the real interference signal is less than a set threshold (-0.3). The Burg spectral entropy distance is defined as the weighted difference between the logarithmic values ​​of the synthetic noise spectral energy and the real noise spectral energy, minus the normal distribution entropy benchmark value.

[0024] In this embodiment, it is specifically necessary to explain the optimization and coordination module. The specific operation of updating the hypergraph adjacency matrix weights through joint gradient backpropagation is as follows: Based on the decoupling loss function output by the interference separation module ( This represents the decoupling loss function. Represents the true label vector. (representing the predicted output vector) and the adversarial gradient data provided by the noise generation module. A cross-modal gradient tensor is constructed; this tensor is generated by the outer product of the partial derivative of the decoupling loss with respect to the topological adjacency matrix and the adversarial gradient with the second-order Hessian matrix. The formula for synthesizing higher-order tensors is as follows: ; Where T represents the cross-modal gradient tensor, used to fuse decoupled and adversarial gradient information. This represents the decoupling loss function. Represents the topological adjacency matrix (dynamic weights of device relationships). This represents the tensor outer product operator, used to construct higher-order gradient correlations. This represents the adversarial loss function, used to evaluate differences in the noise spectrum. This represents the attention weight for the noise branch, used to control the strength of noise suppression. This represents the generator network parameters; the gradient tensor is normalized using Riemannian manifold constraints, and an exponential decay term is added to suppress gradient magnitude fluctuations. The Riemannian constraint normalization formula is: ; in, This represents a regularized gradient tensor used to suppress gradient magnitude fluctuations. Denotes the Frobenius norm. This indicates the adversarial loss gradient (the direction of generator parameter optimization). Represents the gradient sensitivity coefficient (within a range of values). ), used to control the exponential decay intensity; regularize the gradient tensor Mapped to the tangent space of the special orthogonal group Lie algebra, its expression is: ; in, Represents the topological adjacency matrix (dynamic weights of device relationships). This represents the result of tensor matrix transformation. This represents the gradient increment coefficient (with a value range of [0.01, 0.1]). Representing the Lie algebraic tangent space coordinates, the topology matrix weights are updated through spinor exponential mapping: This process ensures that the device topology always satisfies orthogonal transformation constraints, completely avoiding the risk of gradient explosion or vanishing. Its expression is: ; in, This represents the oblique symmetry operator. This represents the topology update step size, with a value range of (0, 0.1]. Represents the matrix exponential mapping, Represents the updated topology matrix; Decoupling loss function and adversarial gradient coupling calculation drive device topology reconstruction, including: calculating the Ricci curvature scalar of the topological adjacency matrix based on spatial spectrum eigenvectors, the calculation formula is: ; in, This represents spatial spectrum characteristics (behavioral signal characteristics associated with device correlation enhancement). Represents the updated topology matrix. Let represent the tensor product (Kronecker product), used to expand the dimension of topological action; I represents the identity matrix, used to preserve the identity transformation of the topological structure. , Let represent the Hessian operator (rate of change of spatial feature curvature), and R represent the Ricci curvature scalar, used to quantify the degree of local curvature of the topological manifold. The curvature scalar is multiplied element-wise by the negative exponent of the spatial feature gradient to generate a curvature gating mask matrix, calculated as follows: ; in, This represents the Sigmoid function, used to map curvature to a gated interval of [0,1]. The curvature mask matrix (values ​​in the range [0,1]) is used to control the update magnitude of the topology parameters. A tensor shrinking operation is performed on the updated topology matrix, combining the topology weights with the gate mask and the preset device basis vectors as a dot product. Its expression is: ; in, This represents the updated topological adjacency matrix. Represents the curvature gate mask matrix. Represents the device basis vector. Corresponding to: , , , Represents the row and column indices of the topological matrix. (Three-device system); ultimately, the parameter stability control of the topological matrix is ​​achieved through spectral radius constraints and spatial feature infinite norm scaling, the expression of which is: ; in, This represents the condensed topology matrix, generated by topology-mask condensation. Indicates the spectral radius operator, The infinity norm represents the maximum absolute value of the vector elements. Describes the minimum value function, with a constraint scaling factor. Output the reconstructed device relationship topology .

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A monitoring and early warning system for preventing myopia in adolescents, characterized in that, Specifically, it includes: The module includes a topology construction module, an interference separation module, a noise generation module, and an optimization and coordination module. Topology construction module: Responding to real-time data streams collected from multiple sources, based on the three-dimensional spatial positions of the wristband, desk lamp, and vision device and the signal frequency domain basis, it creates a time-frequency-space three-dimensional dynamic coupled topology and generates a hypergraph adjacency matrix containing spatial attenuation factors and frequency domain coupling factors. Interference separation module: Based on the hypergraph adjacency matrix, it executes a dual-frequency branch attention mechanism, extracts eye micro-behavioral features through the physiological rhythm branch, and analyzes light intensity gradient changes through the noise branch, thus separating effective behavioral signals from interference noise; Noise generation module: It uses electromagnetic field dynamic equations to constrain the adversarial generative network and synthesizes frequency domain consistent noise based on wristband magnetometer data and device distance threshold; Optimize the collaboration module: update the hypergraph adjacency matrix weights through joint gradient backpropagation, and drive device topology reconstruction by decoupling the loss function and adversarial gradient coupling calculation.

2. The monitoring and early warning system for preventing myopia in adolescents according to claim 1, characterized in that: In the topology construction module, the specific operation for creating a time-frequency-space three-dimensional dynamically coupled topology is as follows: A1. Construct a Euclidean distance matrix based on the real-time three-dimensional coordinate data reported by smart bracelets, smart desk lamps and smart vision testing devices, calculate the spatial straight-line distance between each pair of devices, and combine the distance values ​​between each pair of devices into a spatial position feature vector, while extracting the maximum distance value as a normalization benchmark. A2. Perform discrete cosine transform on the original sensor signal, retain the frequency domain coefficients of the preset number of main frequencies to form a frequency domain basis matrix; dynamically adjust the spatial scale coefficient of the Gaussian kernel function through environmental interference level parameters, and calculate the spatial attenuation factor between devices. A3. Accumulate the absolute value difference of the frequency domain basis matrix coefficients of each pair of devices, add the zero-prevention division constant, and take the reciprocal to generate the frequency domain coupling factor. A4. Multiply the spatial attenuation factor matrix and the frequency domain coupling factor matrix element by element to obtain the initial hypergraph adjacency matrix.

3. The monitoring and early warning system for preventing myopia in adolescents according to claim 2, characterized in that: The update process of the dynamically coupled topology is as follows: The continuous time stream is segmented according to the time window sliding step size parameter set by the system, and the historical adjacency matrix state is weighted by the topological inertia coefficient. The new arrival signal data within the sliding window is recursively processed using the aforementioned distance matrix construction process and frequency domain basis matrix generation process to obtain the current adjacency matrix state. The historical weights and the current state are then combined using a weighted fusion formula to generate an updated hypergraph adjacency matrix.

4. The monitoring and early warning system for preventing myopia in adolescents according to claim 3, characterized in that: The dynamic adjustment process of the spatial scale coefficient is as follows: The acceleration variance collected by the vibration sensor of the smart bracelet is used as the environmental interference level parameter; the preset static scale benchmark value is used as the base value, and the base value is multiplied by one plus the sum of 0.5 times the environmental interference level parameter to obtain the calculation result of the current spatial scale coefficient.

5. A monitoring and early warning system for preventing myopia in adolescents according to claim 4, characterized in that: The specific process of executing the dual-frequency branch attention mechanism in the interference separation module is as follows: Based on the hypergraph adjacency matrix output by the topology construction module, the spatial relationships between device nodes are first spatially enhanced and reconstructed to generate spatial spectrum feature vectors. These spatial spectrum features are then combined with the original signals of eye microbehavior and light intensity gradient sequences. In the physiological rhythm branch, the principal component features of the physiological rhythm are extracted using a time-frequency transformation algorithm, and in the noise branch, the light intensity gradient change features are analyzed using a recursive filtering algorithm. Subsequently, dual-channel attention weight coefficients are generated, corresponding to the intensity distributions of the physiological rhythm branch and the noise branch, respectively. These dual-channel attention weight coefficients are used to weight the original signal, initially separating the effective behavioral signal components and the interference noise components.

6. A monitoring and early warning system for preventing myopia in adolescents according to claim 5, characterized in that: The specific operation for separating the effective behavioral signal from the interference noise is as follows: Based on the aforementioned dual-channel attention weight coefficients and the synthetic adversarial noise provided by the noise generation module, a gradient-driven interference separation operation is performed, wherein the interference noise component is suppressed by a dynamic noise attenuation operator, while the effective behavioral signal component is enhanced by a residual masking mechanism. The mechanism calculates the visual load influence factor between the original signal and the initially separated signal. This influence factor is quantified by the spatial spectral entropy function and adjusted by the temperature coefficient. Finally, it outputs the enhanced behavioral feature signal.

7. A monitoring and early warning system for preventing myopia in adolescents according to claim 6, characterized in that: In the noise generation module, the specific operation of constraining the adversarial generation network using the electromagnetic field dynamic equation is as follows: Based on the time-series data of the raw magnetic induction intensity collected by the wristband magnetometer, the acceleration dimension of the rate of change of the magnetic field is calculated as a physical constraint benchmark parameter; combined with the maximum Euclidean distance threshold between devices output by the topology construction module, a dynamic gradient upper bound constraint formula is constructed; this formula stipulates that the upper limit of the time-varying gradient change rate of the synthetic noise signal is proportional to the ratio of the magnetic field acceleration amplitude to the distance threshold. The constraint condition is used as the gradient regularization term of the adversarial generative network generator; at the same time, the aforementioned effective signal frequency domain basis is used as the input of the condition vector to guide the noise spectrum structure to be consistent with the real eye behavior environment, and the magnetic field change acceleration is obtained by calculating the second derivative.

8. A monitoring and early warning system for preventing myopia in adolescents according to claim 7, characterized in that: The synthesis operation of the frequency domain uniform noise is as follows: The noise spectrum is orthogonally expanded using Legendre orthogonal polynomial basis functions. Conditional vector concatenation of magnetic field constraint parameters and effective signal frequency domain basis are used as generator inputs. A preliminary noise time-domain sequence is generated by learning Legendre expansion coefficients through a feedforward network. An adversarial loss function based on Burg spectral entropy distance is constructed to measure the frequency domain similarity between synthetic noise and labeled real environmental interference noise. This loss function maximizes the spectral distribution similarity between synthetic noise and real noise while minimizing the difference in their autoregressive model residuals. Finally, the generator parameters are iteratively adjusted through an adversarial training optimizer to ensure that the Burg spectral entropy distance between the output noise and the real interference signal in the frequency domain is less than a set threshold. The Burg spectral entropy distance is defined as the weighted difference between the logarithmic values ​​of the synthetic noise spectral energy and the real noise spectral energy, minus the normal distribution entropy benchmark value.

9. A monitoring and early warning system for preventing myopia in adolescents according to claim 8, characterized in that: In the optimization and coordination module, the specific operation of updating the hypergraph adjacency matrix weights through joint gradient backpropagation is as follows: Based on the decoupling loss function output by the interference separation module and the adversarial gradient data provided by the noise generation module, a cross-mode gradient tensor is constructed. This tensor is generated by the outer product of the second-order Hessian matrix of the adversarial gradient with the partial derivative of the decoupling loss with respect to the topological adjacency matrix. The gradient tensor is normalized by Riemannian manifold constraints and an exponential decay term is superimposed. The regularized gradient tensor is mapped to the tangent space of the special orthogonal group Lie algebra, and the topological matrix weights are updated by spinor exponential mapping.

10. A monitoring and early warning system for preventing myopia in adolescents according to claim 9, characterized in that: The decoupling loss function and adversarial gradient coupling calculation drive device topology reconstruction includes: calculating the Ricci curvature scalar of the topological adjacency matrix based on the spatial spectral feature vector; performing element-wise multiplication of the curvature scalar with the negative exponential term of the spatial feature gradient to generate a curvature gating mask matrix; performing tensor shrinking operation on the updated topology matrix to synthesize the topology weights with the gating mask and the preset device basis vectors by dot product; and finally achieving parameter stability control of the topology matrix through spectral radius constraints and spatial feature infinite norm scaling, outputting the reconstructed device relationship topology.

Images