LED personalized eye-protection illumination method based on AI learning

Abstract

The invention relates to the field of intelligent table lamps, and discloses an LED personalized eye protection lighting method based on AI learning, and the method comprises the steps: fusing the myopia degree manually input by a user, real-time ambient light intensity and desktop reflection characteristics, and generating an eye load adaptation matrix; performing cross-domain convolution recursion modeling on the eye load adaptation matrix, and constructing a user personalized visual tolerance tensor; based on the visual tolerance tensor, utilizing an adaptive threshold drift model to extract an illumination comfort trajectory, and identifying potential visual fatigue risk candidate points; mapping the potential asthenopia risk candidate points to an LED dimmable unit cluster, and performing iterative coupling optimization on brightness gradient, color temperature offset and an illumination incident angle to obtain light field adjustment solutions facing different myopia degrees; generating a non-uniform distributed control instruction among the brightness, the color temperature and the refresh frequency based on the light field adjustment solution; the method has the advantage of intelligently adjusting the light brightness and the color temperature by combining the myopia degree of the user and the ambient brightness.

Description

Technical Field

[0001] This invention relates to the field of smart desk lamps, specifically to a method for personalized eye-protection LED lighting based on AI learning. Background Technology

[0002] With the increasing prevalence of LED desk lamps in students' home learning environments, lighting products are gradually evolving from simple, fixed-brightness illumination towards intelligent and health-oriented designs. However, most existing smart desk lamps automatically adjust brightness based solely on ambient light intensity, neglecting individual differences among users, especially the visual sensitivity of nearsighted individuals. Users with different degrees of nearsightedness exhibit significant differences in their perception of light intensity and color temperature; excessive brightness or darkness can cause visual fatigue and even exacerbate eye strain. While existing technologies include solutions that use sensors to detect ambient light and adjust lamp brightness and color temperature, they lack individualized adjustment mechanisms for nearsighted individuals, failing to provide more precise lighting matching based on the user's actual nearsightedness degree. Especially among students, where the nearsightedness rate is high, prolonged use of lighting incompatible with their vision may negatively impact eye health. Therefore, designing an AI-based personalized eye-protection LED lighting method that intelligently adjusts light brightness and color temperature based on the user's nearsightedness degree and ambient brightness is essential. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an AI-based personalized LED eye-protection lighting method that intelligently adjusts light brightness and color temperature based on the user's myopia degree and ambient brightness, thus solving the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of intelligently adjusting light brightness and color temperature based on the user's myopia degree and ambient brightness, this invention provides the following technical solution: A personalized LED eye-protection lighting method based on AI learning, comprising the following steps: The system integrates the myopia degree manually input by the user, real-time ambient light intensity, and desktop reflection characteristics, and performs nonlinear perturbation coding on the multi-source signals based on the vision correction demand function to generate an eye load adaptation matrix. Cross-domain convolutional recursive modeling is performed on the eye load adaptation matrix, and a user-personalized visual tolerance tensor is constructed by combining historical eye use pattern trajectories, circadian rhythm shift factors and spectral imbalance index. Based on the visual tolerance tensor, an adaptive threshold drift model is used to extract the illumination comfort trajectory. By detecting local brightness oversaturation areas and color temperature change slopes, potential candidate points for visual fatigue risk are identified. The potential visual fatigue risk candidate points are mapped to LED tunable unit clusters, and degree-sensitive correction coefficients are embedded in the spatial light distribution network. Iterative coupling optimization is performed on the brightness gradient, color temperature shift and light incident angle to obtain light field adjustment solutions for different myopia degrees. Based on the light field modulation solution and combined with the dynamic evolution trend of the global light comfort index, the control strategy of the LED driver layer is reconstructed to generate non-uniform distributed control commands among brightness, color temperature and refresh frequency.

[0005] Preferably, the process of nonlinear perturbation coding of multi-source signals based on the vision correction demand function is as follows: The system collects the myopia degree input by the user and converts it into optical correction coefficients, while also connecting the optical intensity curves obtained from the ambient light intensity sensor and the desktop reflectance spectral unit. By using a multi-channel feature superposition method, the optical correction coefficient, optical intensity curve and reflectivity distribution of the writing surface are combined into a multi-source signal set; A nonlinear perturbation function, including pseudo-random phase shift, weighted normalization and polynomial convolution, is applied to a multi-source signal set to output an initial signal encoding result with correction weights and environmental adaptation features.

[0006] Preferably, the process of generating the eye load adaptation matrix is ​​as follows: The initial signal encoding results are mapped to a three-dimensional index space according to the user's myopia degree, ambient light intensity, and reflection characteristics, respectively. A multidimensional interpolation method was used to correct missing points, and robust regression was used to remove abnormal interference values ​​during the environmental data collection process. A dynamic weighting mechanism is introduced, assigning high weight to the myopia degree factor, medium weight to the instantaneous light intensity factor, and low weight to the reflectance spectrum factor. Multidimensional coupling relationships are obtained through tensor projection, and the hierarchical distribution of visual load intensity is marked in the matrix to form an eye load adaptation matrix.

[0007] Preferably, the process of performing cross-domain convolutional recursive modeling on the eye load adaptation matrix is ​​as follows: The eye load adaptation matrix is ​​divided into local sub-blocks, each sub-block corresponding to the range of myopia degree, the light intensity distribution interval and the reflectance characteristic bandwidth; Each sub-block is input to a convolutional recursive unit to extract the temporal pattern, and a cross-domain coupling operator is embedded in the convolutional kernel to synchronously process the coupling relationship between signals from different sources in the temporal and frequency domains. A memory gating mechanism is introduced in the recursive modeling process, which enables the historical eye use pattern trajectory generated in the previous stage to dynamically adjust the update rate of the current state. Furthermore, it is combined with a preset circadian rhythm offset factor to perform weighted correction on the differences in light sensitivity at different time periods. In the output layer of the convolution result, the residual correction network compensates for the spectral imbalance index caused by the mixing of ambient light and LED light source, and outputs the feature sequence after cross-domain modeling.

[0008] Preferably, the process of constructing the user-personalized visual tolerance tensor is as follows: The feature sequence after cross-domain modeling is embedded into the tensor draft in three dimensions according to the time axis, myopia factor and optical interference factor. Threshold normalization was applied to the initial tensor to ensure dimensional consistency across different dimensions. The main tolerance patterns are extracted using the principal component mapping method, and feature alignment is performed by combining the user's historical eye behavior. A visual tolerance tensor with individual differences is generated through multi-level tensor decomposition.

[0009] Preferably, the process of extracting the lighting comfort trajectory using an adaptive threshold drift model is as follows: The brightness and color temperature change curves of the visual tolerance tensor at different time periods are sampled frame by frame. An adaptive drift function is introduced to dynamically adjust the reference thresholds for brightness and color temperature to match the user's visual tolerance range; The sliding window analysis method is used to identify the comfort zone and the oversaturation zone, and the duration and intensity changes are calculated. The results are combined into a time series to form a light comfort trajectory.

[0010] Preferably, the process for identifying potential candidate points of visual fatigue risk is as follows: Perform a second-order differential analysis on the illumination comfort trajectory to extract the slope of color temperature abrupt change and brightness acceleration; Two judgment conditions are set: the first condition is that the rate of change of brightness exceeds a preset threshold, and the second condition is that the slope of color temperature change exceeds the fatigue sensitivity threshold. If both conditions are met, the corresponding time segment is marked as a potential risk candidate point.

[0011] Preferably, the process of mapping potential visual fatigue risk candidate points to LED tunable unit clusters and embedding degree-sensitive correction coefficients in the spatial light distribution network is as follows: The desktop and paper areas corresponding to potential risk candidate points are divided into several illuminated subdomains. Each subdomain is assigned a corresponding LED dimmable unit, and the light sensitivity factor is calculated based on the user's refractive state parameters, including myopia degree, astigmatism axis and pupil diameter. After normalizing the light sensitivity factor, an individualized degree sensitivity correction coefficient is generated; In the spatial light distribution network, brightness gradient, color temperature shift and light incidence angle and degree-sensitive correction coefficient are coupled through multidimensional edge weight functions; The edge weight coefficients are iteratively updated to obtain the light field distribution result after individualized optical correction.

[0012] Preferably, the process of obtaining the light field modulation solution for different degrees of myopia is as follows: Based on the spatial light distribution results, the illumination adaptability index under different myopia degrees is calculated to characterize the overall matching degree of light field distribution with user visual clarity and comfort. Using the light adaptability index as the optimization feedback quantity, and combining it with a multi-objective optimization function, joint optimization is carried out among brightness uniformity, color temperature comfort and glare suppression. In each optimization iteration, user vision correction factors are dynamically introduced to ultimately obtain a light field modulation solution that covers people with various degrees of myopia.

[0013] Preferably, the process of generating non-uniform distributed control commands among brightness, color temperature, and refresh rate is as follows: By combining the light field modulation solution with the dynamic evolution trend of the global light comfort index, multi-parameter joint constraints are generated. Distributed control decomposition is performed on the brightness channel, color temperature channel, and refresh channel of the LED driver layer; A non-uniform weighting algorithm is used to assign independent control step sizes to different channels to adapt to the distribution characteristics of risk candidate points. Integrate the control signals output from all channels to form a non-uniformly distributed control command.

[0014] Compared with existing technologies, this invention provides a personalized LED eye-protection lighting method based on AI learning, which has the following beneficial effects: This invention establishes an eye load adaptation matrix by integrating individual user refractive parameters such as myopia degree, astigmatism axis, and pupil diameter with real-time ambient light intensity and desktop reflection characteristics. This matrix scientifically quantifies the visual load level of each user under different lighting conditions, thereby achieving personalized lighting adjustment. Through a combination of cross-domain convolutional recursive modeling and historical eye usage pattern trajectories, the constructed visual tolerance tensor not only reflects the user's dynamic adaptability to brightness and color temperature but also takes into account the impact of circadian rhythms and spectral imbalances on visual comfort, improving the accuracy and continuity of lighting adjustment. By utilizing an adaptive threshold drift model to extract the lighting comfort trajectory and identifying risks in local oversaturated brightness areas and color temperature abrupt change slopes, potential eye fatigue risk points can be detected in a timely manner, enabling proactive protection against light environment issues. By mapping potential risk points to LED tunable unit clusters and embedding individualized illumination correction coefficients into the spatial light distribution network, the brightness gradient, color temperature shift, and illumination incident angle are precisely controlled through iterative coupling optimization. This results in a light field adjustment solution tailored to users with different degrees of myopia, achieving personalized, non-uniform, and dynamically responsive lighting effects. Distributed control commands are generated by combining the dynamic evolution trend of the global light comfort index, simultaneously adjusting LED brightness, color temperature, and refresh rate to achieve refined control of different illumination subdomains. This effectively reduces the risk of eye fatigue, improves visual clarity, and enhances visual comfort and eye protection during prolonged work, study, or reading. It also improves the energy efficiency and adaptability of LED lighting systems, demonstrating significant practical value and promotional potential. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a method for personalized eye-protection LED lighting based on AI learning includes the following steps: S1: Integrate the myopia degree manually input by the user, real-time ambient light intensity and desktop reflection characteristics, and perform nonlinear perturbation coding on the multi-source signals based on the vision correction demand function to generate an eye load adaptation matrix.

[0018] The process of nonlinear perturbation coding of multi-source signals based on the vision correction demand function in S1 is as follows: The system collects the myopia degree input by the user and converts it into optical correction coefficients, while also connecting the optical intensity curves obtained from the ambient light intensity sensor and the desktop reflectance spectral unit. By using a multi-channel feature superposition method, the optical correction coefficient, optical intensity curve and reflectivity distribution of the writing surface are combined into a multi-source signal set; A nonlinear perturbation function, including pseudo-random phase shift, weighted normalization and polynomial convolution, is applied to a multi-source signal set to output an initial signal encoding result with correction weights and environmental adaptation features.

[0019] The process of generating the eye load adaptation matrix in S1 is as follows: The initial signal encoding results are mapped to a three-dimensional index space according to the user's myopia degree, ambient light intensity, and reflectance characteristics. The original optical signal from the smart glasses or eye-tracking acquisition device is acquired and converted into a standardized feature vector through a time-series coding algorithm. Based on the user's individual refractive power (myopia degree parameter, in diopters D), the ambient light intensity (in lux), and reflectance characteristics (represented by the spectral reflectance curve), the three types of parameters are used as coordinate axes in the three-dimensional index space to perform multi-dimensional mapping on the initial signal, ensuring that each point in the index space corresponds to a specific combination of the user's eye state and external environmental conditions. Multidimensional interpolation methods are used to correct missing points, and robust regression is used to remove abnormal interference values ​​during environmental acquisition. In the index space, some acquisition points may be missing due to sensor jitter, sampling delay, or environmental occlusion. Multidimensional interpolation methods (such as trilinear interpolation or radial basis function-based interpolation methods) are used to perform numerical compensation between known points to fill in the missing parameter values. Considering that ambient light acquisition is easily affected by reflected noise or instantaneous strong light, robust regression algorithms (such as Huber regression or RANSAC method) are further used to remove outliers in the acquisition sequence that deviate too much from the mean, so as to improve signal stability and effectiveness. A dynamic weighting mechanism is introduced, assigning high weight to the myopia degree factor, medium weight to the instantaneous light intensity factor, and low weight to the reflectance spectrum factor. Based on the differences in the sensitivity of visual load to different parameters, dynamic weighting factors are set. Myopia degree directly determines the ease of visual accommodation and focusing, therefore it is assigned the highest weight (e.g., 0.5-0.6). Instantaneous light intensity directly affects pupil constriction and visual acuity, but this can be partially alleviated by blinking or optical accommodation, therefore it is assigned a medium weight (e.g., 0.3-0.35). The reflectance spectrum factor mainly affects color shift and visual fatigue during prolonged viewing, and has a smaller impact on short-term visual load, therefore it is assigned a low weight (e.g., 0.1-0.15). The dynamic weighting mechanism is adjusted according to real-time conditions during operation to ensure that the matrix results more accurately describe visual load in accordance with physiological characteristics. Multidimensional coupling relationships are obtained through tensor projection, and the hierarchical distribution of visual load intensity is marked in the matrix to form an eye load adaptation matrix. The weighted and corrected three-dimensional index space parameters are transformed into high-dimensional tensor representations, and the coupling relationships between different parameters are extracted using tensor decomposition or principal component projection methods. For example, the nonlinear interaction between myopia degree and light intensity, and the superposition effect of spectral reflectance characteristics on long-term visual fatigue are extracted. These coupling relationships are mapped into the matrix structure and hierarchically marked according to visual load intensity (divided into three or more levels from low to high: mild, moderate, and severe). The final matrix not only reflects the independent effects of different factors, but also reveals the comprehensive visual load level under the interaction of multiple factors.

[0020] S2: Perform cross-domain convolutional recursive modeling on the eye load adaptation matrix, and construct a user-personalized visual tolerance tensor by combining historical eye use pattern trajectories, circadian rhythm shift factors and spectral imbalance index.

[0021] The process of performing cross-domain convolutional recursive modeling on the eye load adaptation matrix in S2 is as follows: The eye load adaptation matrix is ​​divided into local sub-blocks, each corresponding to the myopia degree range, light intensity distribution interval, and reflectance feature bandwidth. The three dimensions of the eye load adaptation matrix are discretized: the myopia degree dimension is divided into multiple intervals with a step size of 0.25D or 0.5D to facilitate local modeling for users with different myopia levels; the ambient light intensity dimension is binned according to light intensity intervals (e.g., 0–50 lx, 50–200 lx, 200–500 lx) to ensure the model can capture differences in visual load under different brightness conditions; the reflectance feature dimension is divided according to the reflectivity or spectral bandwidth of the desktop or paper material (e.g., segmented in 10nm or 50nm increments) to fully reflect the material's influence on light distribution. The three-dimensional matrix is ​​combined in a grid pattern to generate several local sub-blocks. Each sub-block contains a sample set within the corresponding interval and retains time-series information for subsequent recursive modeling. Statistical information (mean, variance, sample size) is recorded for each sub-block for weighted calculation and outlier removal. Each sub-block is input to a recursive convolutional unit to extract temporal patterns. A cross-domain coupling operator is embedded in the convolutional kernel to synchronously process the coupling relationships between signals from different sources in the temporal and frequency domains. The multimodal signals of each sub-block are arranged in temporal order to form a tensor (dimension of time window length × spatial or feature dimension × number of modal channels), which serves as the input to the recursive convolutional unit. During the convolution operation, the convolutional kernel is modified by the cross-domain coupling operator. This operator generates a weight vector based on the short-term cross-correlation or mutual information between different modalities (myopia degree, light intensity, reflection features) within the current sub-block, and uses this weight to scale or adjust the convolutional kernel channel by channel, thereby synchronously extracting cross-modal coupling features during the convolution process. The convolution output is then input to the recursive unit, which captures dynamically changing patterns through time-step iteration, realizing continuous encoding of temporal information within the sub-block. A memory gating mechanism is introduced during the recursive modeling process, enabling the historical eye-use pattern trajectory generated in the previous stage to dynamically adjust the update rate of the current state. Furthermore, a pre-set circadian rhythm offset factor is used to weight and correct the differences in light sensitivity across different time periods. A memory gating mechanism is set within the recursive unit, accepting three inputs: the current sub-block features, the hidden state of the previous time step, and the historical eye-use pattern vector. This vector contains information such as the user's past fixation time, blink frequency, and usage duration statistics. The gating mechanism calculates update coefficients using the sigmoid function, dynamically controlling the degree to which the current state absorbs historical information. The circadian rhythm offset factor is added to the gating calculation as a periodic function to adjust the user's sensitivity to light at different time periods, allowing the model to reduce its response to short-term light intensity changes at night or during periods of fatigue, prioritizing cumulative visual load and long-term trends, and ensuring that the modeling results more closely match actual visual physiological states. In the output layer of the convolution result, the residual correction network compensates for the spectral imbalance index caused by the mixing of ambient light and LED light sources, and outputs the feature sequence after cross-domain modeling. After the convolution recursive unit outputs the original feature sequence, it obtains the current ambient light spectral distribution and the nominal spectrum of the LED light source from the sensor, and calculates the difference between the two to form a spectral imbalance vector. This vector is input into the residual correction network (composed of several fully connected layers or residual blocks) to generate a compensation term, and is added element-wise to the original feature sequence to obtain the spectrally compensated feature sequence. This sequence not only contains the temporal features of cross-modal coupling, but also undergoes spectral bias correction, which can truly reflect the changes in visual load under mixed light environment.

[0022] The process of constructing the user-personalized visual tolerance tensor in S2 is as follows: The feature sequences obtained from cross-domain modeling are embedded into a three-dimensional tensor draft according to the time axis, myopia factor, and optical interference factor. The feature sequences obtained through cross-domain convolutional recursive modeling are arranged according to time steps to form the time axis dimension. The user's myopia information corresponding to each time step is mapped to the myopia factor dimension. At the same time, signals affecting visual comfort, such as ambient light spectrum, desktop reflection, and LED adjustment parameters, are combined into the optical interference factor dimension. These three types of information are used to construct a three-dimensional tensor draft according to time step × myopia factor × optical interference factor, so that each tensor unit contains temporal information, individual visual acuity differences, and environmental optical interference features. Metadata, including sampling timestamp, light intensity unit, and reflectance calibration value, is also added to the tensor. Threshold normalization was applied to the initial tensor to ensure dimensional consistency across different dimensions. The minimum, maximum, and mean values ​​were calculated for each dimension of the tensor, and outliers were truncated based on preset thresholds to remove excessively large or small outliers, thus reducing the impact of environmental noise and measurement errors on the model. A normalization method was used to map the values ​​of the time series dimension, myopia factor, and optical interference factor to the same dimensional range (e.g., 0–1 or a mean of 0 and a variance of 1), ensuring that the weight ratios of each dimension would not deviate due to differences in numerical scale during subsequent tensor operations. During normalization, the time dimension was dynamically weighted based on the user's historical eye strain intensity to enhance the feature representation of periods of high visual load. This study extracts key tolerance patterns using principal component mapping (PCM) and aligns them with users' historical eye-use behavior. A multi-level tensor decomposition method generates a visual tolerance tensor with individual differences. The normalized tensor is flattened or expanded along a specified dimension to adapt to higher-order PCM algorithms. By calculating the covariance matrix and extracting the first few principal components, key tolerance patterns that explain most changes in visual load are obtained, while retaining the weight coefficients and contribution information of each principal component. These key patterns are then aligned with users' historical eye-use behavior data (such as daily fixation time distribution, blink frequency changes, and continuous eye-use duration). A similarity matching method is used to correct the temporal offset of the principal components, ensuring the tensor reflects the user's long-term visual tolerance trend. A multi-level tensor decomposition strategy maps the principal components back to the original three-dimensional tensor structure, generating a final visual tolerance tensor with individual differences. Each tensor unit retains not only temporal information, myopia differences, and optical interference, but also the user's personalized tolerance pattern.

[0023] S3: Based on the visual tolerance tensor, an adaptive threshold drift model is used to extract the illumination comfort trajectory. By detecting local oversaturated brightness areas and color temperature abrupt change slopes, potential candidate points for visual fatigue risk are identified.

[0024] The process of extracting the lighting comfort trajectory using the adaptive threshold drift model in S3 is as follows: The brightness and color temperature change curves of different time periods in the visual tolerance tensor are sampled frame by frame; the time axis data is extracted from the constructed user-personalized visual tolerance tensor to obtain the brightness and color temperature value sequence for each time step; each time period is sampled frame by frame at a sampling frequency of milliseconds or seconds to ensure that high-frequency illumination changes (such as LED PWM dimming or flicker fluctuations) are accurately captured, while preserving the low-frequency change trend; timestamps, light intensity units, and absolute color temperature values ​​can be recorded during the sampling process; An adaptive drift function is introduced to dynamically adjust the reference thresholds for brightness and color temperature to match the user's visual tolerance range. The reference threshold ranges for brightness and color temperature are initialized based on the user's myopia, astigmatism, and pupil diameter. The adaptive drift function is applied to the sampled brightness and color temperature sequences. This function automatically adjusts the upper and lower limits of the thresholds based on the average value and fluctuation amplitude of historical visual tolerance tensors, as well as short-term eye load intensity, to match the current user's visual comfort range in real time. The drift function also considers circadian rhythms and changes in ambient light intensity, allowing the reference thresholds to dynamically shift over time, thus more accurately reflecting the user's visual load tolerance at different times. The sliding window analysis method identifies comfortable and oversaturated zones, calculates their duration and intensity changes, and combines the results into a time series to form a lighting comfort trajectory. An appropriate sliding window length (e.g., 10 or 30 seconds) is set, and the brightness and color temperature curves are scanned frame by frame. Within each window, the average brightness, color temperature deviation, and instantaneous fluctuation amplitude are calculated. Based on a threshold adjusted by an adaptive drift function, the data within the window is marked as a comfortable or oversaturated zone, and the start and end times and the range of light intensity changes corresponding to that window are recorded. The entire time series is traversed, consecutive comfortable or oversaturated zones are merged, their duration and intensity changes are calculated, and all window results are combined in chronological order to generate a complete lighting comfort trajectory, reflecting the user's visual comfort and potential fatigue risk at different time periods.

[0025] The process of identifying potential candidate points for visual fatigue risk in S3 is as follows: A second-order differential analysis was performed on the illumination comfort trajectory to extract the color temperature abrupt change slope and luminance acceleration. Sequence data of luminance and color temperature changes over time were extracted from the generated illumination comfort trajectory. The first-order differential of the luminance sequence was calculated to obtain the instantaneous rate of change, and then the second-order differential was performed on the first-order result to obtain the luminance acceleration curve, reflecting the degree of abrupt change in illumination intensity. The second-order differential was also performed on the color temperature sequence to obtain the color temperature abrupt change slope curve, used to capture the drastic degree of light color change. During the calculation process, a smoothing filter was used to remove high-frequency noise to ensure the reliability of the luminance acceleration and color temperature abrupt change slope. A dual-judgment condition is set: the first condition is that the rate of change in brightness exceeds a preset threshold, and the second condition is that the slope of the color temperature change exceeds a fatigue sensitivity threshold. Based on the user's refractive parameters such as myopia degree, astigmatism axis, and pupil diameter, combined with ambient light intensity and historical visual tolerance data, the brightness change threshold and color temperature change threshold applicable to the user are calculated. The first condition, that the brightness acceleration exceeds the threshold set by the user individually, indicates that a rapid change in light intensity may cause visual discomfort; the second condition, that the slope of the color temperature change exceeds the fatigue sensitivity threshold, indicates that a rapid change in light color may lead to visual fatigue. The thresholds can be dynamically adjusted according to the user's eye use status, circadian rhythm, and historical comfort range to improve recognition accuracy. If both conditions are met simultaneously, the corresponding time segment is marked as a potential risk candidate point. Each sampling time point of the illumination comfort trajectory is traversed, and the luminance acceleration and color temperature abrupt change slope are compared with their respective thresholds. If the luminance acceleration is greater than both the luminance threshold and the color temperature abrupt change slope is greater than the color temperature threshold, then the time segment is considered to have a potential risk of visual fatigue. During the marking process, the start and end times of the time segment, the corresponding luminance and color temperature values, the slope and acceleration values, and the risk level are recorded. Time segments that continuously meet the conditions are merged to form a complete sequence of potential risk candidate points.

[0026] S4: Map potential visual fatigue risk candidate points to LED tunable unit clusters, embed degree-sensitive correction coefficients in the spatial light distribution network, and perform iterative coupling optimization on brightness gradient, color temperature shift and incident light angle to obtain light field modulation solutions for different myopia degrees.

[0027] The process in S4 of mapping potential visual fatigue risk candidate points to LED tunable unit clusters and embedding degree-sensitive correction coefficients in the spatial light distribution network is as follows: The desktop and paper areas corresponding to potential risk candidate points are divided into several illumination sub-domains. Based on the potential visual fatigue risk candidate points identified in the previous stage, the two-dimensional coordinate range of the user's gaze area is determined, including the desktop, paper, or electronic device display area. The two-dimensional coordinate range is divided into several illumination sub-domains, and the size of each sub-domain is adjusted according to the LED layout density and the user's visual sensitivity to ensure uniform coverage and accurate reflection of local lighting needs. The geometric information of each sub-domain (such as center position, area, and azimuth angle) is recorded. Each subdomain is assigned a corresponding dimmable LED unit, and the light sensitivity factor is calculated based on the user's refractive state parameters, including myopia degree, astigmatism axis, and pupil diameter. For each illumination subdomain, the set of controllable LED units for each illumination subdomain is determined based on the spatial position and emission angle of the LED beads. A light sensitivity model is established using the user's refractive parameters (myopia degree, astigmatism axis, and pupil diameter) to calculate the sensitivity factor of the illumination subdomain to user visual fatigue under different light intensities and colors. The calculation of the sensitivity factor considers the impact of refractive error on visual focus and brightness adaptation ability, and also incorporates the user's historical visual tolerance data as a weighted correction. After normalizing the illumination sensitivity factor, individualized degree sensitivity correction coefficients are generated. The illumination sensitivity factor calculated for each subdomain is normalized so that the sensitivity coefficients of all subdomains fall within the [0,1] interval, which facilitates uniform processing in subsequent network calculations. The normalization method can be min-max normalization or Z-score standardization, maintaining the relative proportion of sensitivity as needed. The normalized coefficients are the individualized degree sensitivity correction coefficients, which are used to weight the adjustment weights of brightness and color temperature of each subdomain during the light field distribution optimization process. In the spatial light distribution network, brightness gradient, color temperature shift, and illumination incident angle and degree sensitivity correction coefficient are coupled through a multidimensional edge weight function. The spatial light distribution network is constructed, treating each tunable LED unit and its corresponding illumination subdomain as nodes, with edges connecting them to represent the influence relationship between the LED unit and the illumination subdomain. A multidimensional edge weight function is introduced into the network, using brightness gradient (local light intensity change), color temperature shift (light color change), illumination incident angle (angle between the light direction and the eye's visual axis), and the normalized degree sensitivity correction coefficient as coupling parameters to perform weighted calculations on each edge. The weighted calculation uses linear or nonlinear functions to synthesize the parameters and generate the influence weight of each edge, thereby quantifying the contribution of the LED light source to the visual comfort of the subdomain. The edge weight coefficients are iteratively updated to obtain the light field distribution result after individualized optical correction. The edge weight coefficients in the spatial light distribution network are gradually updated through iterative optimization methods (such as gradient descent, genetic algorithm or reinforcement learning strategy) so that the illuminance and color temperature of each subdomain match the individualized needs as much as possible. In each iteration, the direction and magnitude of the edge weight adjustment are corrected by combining the user's light comfort index and historical visual tolerance data. After several iterations, the final light field distribution result containing the brightness and color temperature of each LED unit is output, completing the individualized optical correction, ensuring the optimal lighting conditions in potential risk areas, and reducing the risk of user visual fatigue.

[0028] The process of obtaining the light field modulation solution for different myopia degrees in S4 is as follows: Based on the spatial light distribution results, an illumination adaptability index is calculated under different myopia degrees to characterize the overall matching degree of the light field distribution with the user's visual clarity and comfort. According to the individualized optically corrected light field distribution results generated in the previous stage, the brightness, color temperature, and incident angle information of each illumination subdomain are extracted, and these light field parameters are mapped into the user's retinal imaging model. For users with different myopia degrees, the focusing effect of light on the retina after passing through refractive errors, astigmatism axis, and pupillary changes is simulated to evaluate the impact of brightness uniformity, color temperature comfort, and glare intensity on visual clarity. Finally, the above multiple visual indicators are weighted and integrated to form an illumination adaptability index, which is used to quantify the matching degree of the light field distribution with the visual comfort of users with specific myopia degrees. Using the illumination adaptability index as the optimization feedback quantity, combined with a multi-objective optimization function, joint optimization is performed among brightness uniformity, color temperature comfort, and glare suppression. The calculated illumination adaptability index is used as the feedback quantity of the objective function and input into a multi-objective optimization algorithm, such as particle swarm optimization or genetic algorithm. During the optimization process, brightness uniformity objective (ensuring uniform brightness distribution in each subdomain), color temperature comfort objective (keeping color temperature changes within the user's tolerance range), and glare suppression objective (suppressing the stimulation of the eyes by local high brightness and direct light) are defined respectively. Through iterative calculation, the output brightness and color temperature of each LED unit are adjusted so that the comprehensive optimization index reaches the optimal balance among different subdomains. The changes in the illumination adaptability index are dynamically evaluated during the optimization process to ensure that the overall visual comfort is improved with each iteration. In each optimization iteration, user vision correction factors are dynamically introduced to ultimately obtain a light field modulation solution covering various myopia levels. During the optimization iteration process, for users with different myopia levels, their refractive power, astigmatism axis, and pupil diameter are used as correction factors input into the optimization algorithm to dynamically adjust the contribution weight of LED brightness and color temperature in each subdomain. By conducting parallel simulations on multiple groups of myopia users, the impact of light field distribution on clarity and comfort under different visual conditions is evaluated. After optimization, a comprehensive light field modulation solution is output, which maintains good brightness balance, color temperature comfort, and glare suppression effects under different myopia conditions, thereby realizing a personalized eye-protection lighting solution for various myopia levels.

[0029] S5: Based on the light field modulation solution and combined with the dynamic evolution trend of the global light comfort index, the control strategy of the LED driver layer is reconstructed to generate non-uniform distributed control instructions among brightness, color temperature and refresh frequency.

[0030] The process of generating non-uniform distributed control commands among brightness, color temperature, and refresh rate in S5 is as follows: By combining the dynamic evolution trend of the light field adjustment solution with the global light comfort index, multi-parameter joint constraints are generated. The brightness, color temperature, and illumination distribution results of each LED unit in the obtained light field adjustment solution are mapped and compared with the global light comfort index to analyze the real-time changing trends of illumination uniformity, color temperature comfort, and glare suppression indices. A constraint matrix is ​​established for the illumination parameters of each subdomain in the light field, incorporating the upper and lower limits of brightness, color temperature, and refresh rate, the maximum rate of change, and the illumination sensitivity of each subdomain into the constraints. These constraints are integrated into multi-parameter joint constraints to provide an operable mathematical basis for subsequent distributed control, ensuring that visual comfort is satisfied while adhering to hardware driving capability limitations during the adjustment process. The brightness, color temperature, and refresh channels of the LED driver layer are decomposed into distributed control. The overall LED driver control task is divided into three independent channels: the brightness channel is responsible for adjusting the luminous intensity of each LED unit, the color temperature channel is responsible for adjusting the spectral distribution of each LED unit, and the refresh channel is responsible for adjusting the frequency and PWM duty cycle of the light source output. Each channel generates a sub-control task list based on the light field adjustment solution and multi-parameter constraints, including target value, adjustable range, current state, and priority information. This decomposition method enables different channels to independently calculate control strategies while ensuring their consistency in the final light field output. A non-uniform weighting algorithm is used to assign independent control step sizes to different channels to adapt to the distribution characteristics of risk candidate points. Based on the spatial distribution of potential visual fatigue risk candidate points on the desktop or paper surface, the adjustment step sizes of each LED unit in terms of brightness, color temperature, and refresh rate are dynamically weighted. Subdomains with high risk are assigned smaller step sizes to achieve fine control, while subdomains with low risk can be assigned larger step sizes to improve response efficiency. In the weighting calculation, the user's visual tolerance tensor, the illumination comfort trajectory, and the light field modulation solution are comprehensively considered to generate an independent control step size vector for each LED unit. Through the non-uniform weighting strategy, the control commands can take into account both the overall light field uniformity and the local individualized eye protection needs. The control signals output from all channels are integrated to form non-uniform distributed control commands. The sub-control signals of the brightness channel, color temperature channel, and refresh channel are weighted by step size and aggregated to the LED driver interface. The final control command set is generated according to the address and illumination subdomain order of each LED unit. During the aggregation process, the control signals of different channels are synchronized in time and conflict detected to ensure that the multi-channel commands sent at the same time do not cause overshoot or jitter. The final output control commands can be directly executed at the hardware level, realizing personalized non-uniform illumination adjustment for users with different subdomains and different myopia degrees, while continuously optimizing light comfort and eye protection effects.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for personalized eye-protection LED lighting based on AI learning, characterized in that, Includes the following steps: The system integrates the myopia degree manually input by the user, real-time ambient light intensity, and desktop reflection characteristics, and performs nonlinear perturbation coding on the multi-source signals based on the vision correction demand function to generate an eye load adaptation matrix. Cross-domain convolutional recursive modeling is performed on the eye load adaptation matrix, and a user-personalized visual tolerance tensor is constructed by combining historical eye use pattern trajectories, circadian rhythm shift factors and spectral imbalance index. Based on the visual tolerance tensor, an adaptive threshold drift model is used to extract the illumination comfort trajectory. By detecting local brightness oversaturation areas and color temperature change slopes, potential candidate points for visual fatigue risk are identified. The potential visual fatigue risk candidate points are mapped to LED tunable unit clusters, and degree-sensitive correction coefficients are embedded in the spatial light distribution network. Iterative coupling optimization is performed on the brightness gradient, color temperature shift and light incident angle to obtain light field adjustment solutions for different myopia degrees. Based on the light field modulation solution and combined with the dynamic evolution trend of the global light comfort index, the control strategy of the LED driver layer is reconstructed to generate non-uniform distributed control commands among brightness, color temperature and refresh frequency.

2. The method for personalized LED eye-protection lighting based on AI learning according to claim 1, characterized in that, The process of nonlinear perturbation coding of multi-source signals based on the vision correction demand function is as follows: The system collects the myopia degree input by the user and converts it into optical correction coefficients, while also connecting the optical intensity curves obtained from the ambient light intensity sensor and the desktop reflectance spectral unit. By using a multi-channel feature superposition method, the optical correction coefficient, optical intensity curve and reflectivity distribution of the writing surface are combined into a multi-source signal set; A nonlinear perturbation function, including pseudo-random phase shift, weighted normalization and polynomial convolution, is applied to a multi-source signal set to output an initial signal encoding result with correction weights and environmental adaptation features.

3. The method for personalized LED eye-protection lighting based on AI learning according to claim 2, characterized in that, The process of generating the eye load adaptation matrix is ​​as follows: The initial signal encoding results are mapped to a three-dimensional index space according to the user's myopia degree, ambient light intensity, and reflection characteristics, respectively. A multidimensional interpolation method was used to correct missing points, and robust regression was used to remove abnormal interference values ​​during the environmental data collection process. A dynamic weighting mechanism is introduced, assigning high weight to the myopia degree factor, medium weight to the instantaneous light intensity factor, and low weight to the reflectance spectrum factor. Multidimensional coupling relationships are obtained through tensor projection, and the hierarchical distribution of visual load intensity is marked in the matrix to form an eye load adaptation matrix.

4. The method for personalized LED eye-protection lighting based on AI learning according to claim 3, characterized in that, The process of performing cross-domain convolutional recursive modeling on the eye load adaptation matrix is ​​as follows: The eye load adaptation matrix is ​​divided into local sub-blocks, each sub-block corresponding to the range of myopia degree, the light intensity distribution interval and the reflectance characteristic bandwidth; Each sub-block is input to a convolutional recursive unit to extract the temporal pattern, and a cross-domain coupling operator is embedded in the convolutional kernel to synchronously process the coupling relationship between signals from different sources in the temporal and frequency domains. A memory gating mechanism is introduced in the recursive modeling process, which enables the historical eye use pattern trajectory generated in the previous stage to dynamically adjust the update rate of the current state. Furthermore, it is combined with a preset circadian rhythm offset factor to perform weighted correction on the differences in light sensitivity at different time periods. In the output layer of the convolution result, the residual correction network compensates for the spectral imbalance index caused by the mixing of ambient light and LED light source, and outputs the feature sequence after cross-domain modeling.

5. The method for personalized LED eye-protection lighting based on AI learning according to claim 4, characterized in that, The process of constructing a user-personalized visual tolerance tensor is as follows: The feature sequence after cross-domain modeling is embedded into the tensor draft in three dimensions according to the time axis, myopia factor and optical interference factor. Threshold normalization was applied to the initial tensor to ensure dimensional consistency across different dimensions. The main tolerance patterns are extracted using the principal component mapping method, and feature alignment is performed by combining the user's historical eye behavior. A visual tolerance tensor with individual differences is generated through multi-level tensor decomposition.

6. The method for personalized LED eye-protection lighting based on AI learning according to claim 5, characterized in that, The process of extracting the lighting comfort trajectory using an adaptive threshold drift model is as follows: The brightness and color temperature change curves of the visual tolerance tensor at different time periods are sampled frame by frame. An adaptive drift function is introduced to dynamically adjust the reference thresholds for brightness and color temperature to match the user's visual tolerance range; The sliding window analysis method is used to identify the comfort zone and the oversaturation zone, and the duration and intensity changes are calculated. The results are combined into a time series to form a light comfort trajectory.

7. The method for personalized LED eye-protection lighting based on AI learning according to claim 6, characterized in that, The process of identifying potential candidate points for visual fatigue risk is as follows: Perform a second-order differential analysis on the illumination comfort trajectory to extract the slope of color temperature abrupt change and brightness acceleration; Two judgment conditions are set: the first condition is that the rate of change of brightness exceeds a preset threshold, and the second condition is that the slope of color temperature change exceeds the fatigue sensitivity threshold. If both conditions are met, the corresponding time segment is marked as a potential risk candidate point.

8. The method for personalized LED eye-protection lighting based on AI learning according to claim 7, characterized in that, The process of mapping potential visual fatigue risk candidate points to LED tunable unit clusters and embedding degree-sensitive correction coefficients into the spatial light distribution network is as follows: The desktop and paper areas corresponding to potential risk candidate points are divided into several illuminated subdomains. Each subdomain is assigned a corresponding LED dimmable unit, and the light sensitivity factor is calculated based on the user's refractive state parameters, including myopia degree, astigmatism axis and pupil diameter. After normalizing the light sensitivity factor, an individualized degree sensitivity correction coefficient is generated; In the spatial light distribution network, brightness gradient, color temperature shift and light incidence angle and degree-sensitive correction coefficient are coupled through multidimensional edge weight functions; The edge weight coefficients are iteratively updated to obtain the light field distribution result after individualized optical correction.

9. A method for personalized LED eye-protection lighting based on AI learning according to claim 8, characterized in that, The process of obtaining the optical field modulation solution for different myopia degrees is as follows: Based on the spatial light distribution results, the illumination adaptability index under different myopia degrees is calculated to characterize the overall matching degree of light field distribution with user visual clarity and comfort. Using the light adaptability index as the optimization feedback quantity, and combining it with a multi-objective optimization function, joint optimization is carried out among brightness uniformity, color temperature comfort and glare suppression. In each optimization iteration, user vision correction factors are dynamically introduced to ultimately obtain a light field modulation solution that covers people with various degrees of myopia.

10. A method for personalized LED eye-protection lighting based on AI learning according to claim 9, characterized in that, The process of generating non-uniform distributed control commands among brightness, color temperature, and refresh rate is as follows: By combining the light field modulation solution with the dynamic evolution trend of the global light comfort index, multi-parameter joint constraints are generated. Distributed control decomposition is performed on the brightness channel, color temperature channel, and refresh channel of the LED driver layer; A non-uniform weighting algorithm is used to assign independent control step sizes to different channels to adapt to the distribution characteristics of risk candidate points. Integrate the control signals output from all channels to form a non-uniformly distributed control command.

Images