Whitening skin care AI recommendation method based on multi-source heterogeneous data fusion

Abstract

This invention relates to the field of big data analysis and discloses an AI recommendation method for skin whitening based on multi-source heterogeneous data fusion. The method includes: acquiring the user's intrinsic parameter sequence and the extrinsic parameter sequence of the geographic grid environment; acquiring the multidimensional spectral feature tensor of the target body part to construct a high-dimensional heterogeneous feature space; calculating individual parameter deviations by comparing historical group physiological data; performing integral calculations on high-frequency environmental parameters to generate an environmental stress cumulative integral; determining the active ingredient tolerance threshold based on the individual parameter deviation and the environmental stress cumulative integral; and then scheduling the target component distribution weight vector through a nonlinear penalty function. This invention solves the problem of mismatch between the recommendation scheme and the dynamic environmental carrying capacity by constructing a spatiotemporal environmental stress interference mechanism and an individual tolerance shift self-correction logic, thereby achieving dynamic constraints on the dose-effect boundary of active ingredients and avoiding the risk of tolerance collapse induced by drastic environmental changes.

Description

Technical Field

[0001] This invention relates to an AI recommendation method for skin whitening based on multi-source heterogeneous data fusion, belonging to the field of big data analysis technology. Background Technology

[0002] The metabolic mechanism and deposition patterns of melanin are currently the focus of research. Achieving whitening effects by intervening in tyrosinase activity and slowing down melanin transport is the main approach in the industry. With the evolution of information collection technology, using mobile terminals to collect users' environmental and physiological parameters, combined with skin image data obtained from three-spectrum skin detection devices, has become the core path for generating personalized skincare solutions. This approach typically uses data processing models to associate users' physiological indicators with image features, thereby identifying the degree of pigmentation, inflammation index, and skin barrier status. However, under dynamic environmental conditions and complex climatic conditions, the skin's tolerance to whitening active ingredients fluctuates significantly. Current solutions usually use static collection methods to associate skin characteristics with skincare components, that is, matching corresponding plant ingredients based on skin indicators collected at a single moment. This approach ignores an objective fact: human skin is not in an isolated static environment. Skin metabolic load and inflammatory pathways are affected by the cumulative integral of environmental parameters within the geographical grid. For example, when a specific area experiences continuous high-intensity ultraviolet radiation exposure or an increase in air pollution index, the overall defense baseline of the group's skin will shift, leading to a contraction of the individual's tolerance boundary to high concentrations of active ingredients.

[0003] To address the impact of environmental disturbances, the industry has attempted to aid decision-making by increasing data collection dimensions or adding discrete environmental variables. Analysis reveals that this linear superposition approach fails to resolve the conflict between individual phenotypic data and macroscopic spatiotemporal big data. Even with the introduction of environmental parameters, without calculations of historical stress data and comparisons with the dynamic baseline of the population, the system still cannot perceive the pre-stress state of the skin's underlying layers, easily identifying aggressive treatments with irritation risks. This technical logical flaw often leads to tolerance issues or post-inflammatory hyperpigmentation during the implementation of skincare solutions; for example, publication number CN121122... Chinese invention patent application 497A discloses a method and system for generating personalized skincare product formulas based on skin condition analysis. It simulates temperature, humidity, and ultraviolet radiation in an environmental chamber and uses feedback from skin parameters of skincare product samples before and after use to fit the environmental correlation. Analysis revealed that this approach is highly dependent on artificially preset simulated environments and feedback from physical samples. Not only is the process lengthy, but its core logic is based on an idealized environmental chamber model, failing to penetrate the deep correlation of multi-source heterogeneous data in the real world. It is also difficult to perceive the cumulative stress effect of environmental parameters (such as ultraviolet exposure and air pollution) dynamically evolving with geographic grids in natural environments.

[0004] Therefore, the technical problem to be solved by this invention is how to calculate the dynamic absorption tolerance of an individual by integrating multi-source spatiotemporal big data, and how to achieve adaptive adjustment of the compatibility weights of medicinal plant components at the data processing level, so as to ensure the safety of the recommended scheme under environmental interference. Summary of the Invention

[0005] To address the problems in the background technology, the technical solution of this invention is as follows: A whitening skincare AI recommendation method based on multi-source heterogeneous data fusion, comprising the following steps: Step S101: Obtain the sequence of internal parameters characterizing the user's physiological state and the sequence of external parameters characterizing the user's current geographic grid environment parameters; Step S102: Obtain the multidimensional spectral feature tensor of the user's target area, extract the phenotypic state features from it, and concatenate them with the internal factor parameter sequence and the external factor parameter sequence to construct a high-dimensional heterogeneous feature space. Step S103: In the high-dimensional heterogeneous feature space, call the historical population physiological data in the preset geographic grid, compare the individual transdermal water loss in the intrinsic parameter sequence with the baseline mean in the historical population physiological data, and calculate the individual parameter deviation. Step S104: Extract the time series of high-frequency environmental parameters within the geographic grid for a preset period, and generate an environmental stress cumulative integral characterizing the impact of environmental fluctuations through integral calculation; Step S105: Determine the active ingredient tolerance threshold of the target user under the current geographic grid based on the mapping relationship between individual parameter deviation and environmental stress cumulative integral. When the active ingredient tolerance threshold reaches the preset intervention threshold, the weight of the target component distribution weight vector is reduced by a nonlinear penalty function to reduce the weight of the main component that inhibits melanin and simultaneously compensate the weight of the repair auxiliary components. Step S106: Map the scheduled target component distribution weight vector to the library of authentic medicinal plant active ingredients to generate a personalized skin care solution with the ratio weight limited by the tolerance threshold of the active ingredients.

[0006] Preferably, step S104 is further refined into the following sub-steps: Step S1041, obtain the time series data of high-frequency environmental external factor parameters for the past 14 days within the geographic grid where the user is currently located, where the side length of the geographic grid is 5km; Step S1042, perform cumulative integration calculation on the PM2.5 index and ultraviolet intensity among the high-frequency environmental external factor parameters for the past 14 days to generate the cumulative environmental stress integral characterizing the environmental stress baseline shift within the geographic grid.

[0007] Preferably, step S105 involves weight reduction scheduling of the target component distribution weight vector using a nonlinear penalty function, which is further refined into the following sub-steps: Step S1051, dividing the elements in the target component distribution weight vector into aggressive feature operators and defensive feature operators; Step S1052, constructing a deep Q-network model that includes the current skin state feature vector, the component adjustment action space, and a reward function based on the objective indicator of improvement degree; Step S1053, using a reinforcement learning feedback mechanism, dynamically adjusting the distribution weights of the aggressive and defensive feature operators based on the determination result of the active ingredient tolerance threshold.

[0008] Preferably, the calculation of individual parameter deviation in step S103 includes the following sub-steps: step S1031, retrieving the baseline mean of transdermal water loss of the population matching the current user's age group; step S1032, when the measured value of individual transdermal water loss in the intrinsic parameter sequence is greater than the baseline mean, marking the individual parameter deviation as a positive shift.

[0009] Preferably, in step S102, obtaining the multidimensional spectral feature tensor of the user's target area includes: acquiring white light images, polarized light images, and ultraviolet light images of the facial skin; extracting texture features and analyzing pigment deposition depth on the white light images, polarized light images, and ultraviolet light images; and mapping the extracted depth features to the quantization dimension in the multidimensional spectral feature tensor.

[0010] Preferably, the intrinsic parameter sequence includes: age, sex, inflammatory status, genotype, lifestyle habits, individual transdermal water loss, and glycation level; the extrinsic parameter sequence includes: seasonal characteristics, geographic grid coordinates, ultraviolet exposure index, and air pollution index.

[0011] Preferably, in step S106, generating a personalized skincare plan includes: matching the active concentration ratio of Siegesbeckia orientalis, Dendrobium nobile, Curcuma longa, Eucommia ulmoides, Polygonatum sibiricum, and oleanolic acid according to the scheduled target component distribution weight vector; and generating a technical list characterizing the compatibility ratio of skincare components based on the active concentration ratio.

[0012] Preferably, under the logic loop of generating personalized skin care solutions, the method further includes step S107, which involves periodically collecting changes in the user's physiological data after applying the personalized skin care solution, and inputting this data as a feedback signal into the deep Q-network model to correct the preset response adjustment coefficient.

[0013] Preferably, in step S101, the intrinsic parameter sequence is obtained by collecting the user's structured physiological data through a mobile terminal; the extrinsic parameter sequence is obtained by calling the open data interfaces of the geographic information system and environmental monitoring stations.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In AI-powered recommendations for skin whitening and moisturizing, this invention addresses the technical problem of the disconnect between recommendation logic and dynamic environmental carrying capacity by constructing a spatiotemporal big data interference mechanism. This invention collects multi-source environmental parameters within the user's geographic grid, including air quality index and UV intensity time series data from the past 14 days, and calculates the environmental stress integral. Through deep fusion of this integral with individual multi-omics parameters, the system can identify the pre-excitation state of inflammatory pathways in the skin's underlying layers caused by environmental fluctuations. This overcomes the limitations of traditional solutions that rely solely on single static phenotypic data collection for product mapping. This dynamic scheduling mechanism based on big data trends ensures that the solution output can respond in real-time to drastic changes in the spatiotemporal environment, avoiding tolerance collapse caused by blindly distributing high concentrations of active ingredients during extreme climates or periods of surging environmental pollution.

[0015] 2. By utilizing population baseline deviation calculation, this invention achieves precise dimensionality reduction calculation of individual absorption tolerance. It calls the average skin physiological parameters of people of the same geographical region and age group in the cloud database and sets them as a dynamic baseline for comparison with the transdermal water loss measured by individuals. By calculating the deviation of individual parameters relative to the population baseline, the system can quantify the dynamic defense margin of the skin barrier. This deviation mapping logic from the population to the individual enables the recommendation system to move beyond the qualitative identification of isolated features and reveal potential barrier damage risks through data mining. This mechanism ensures that the subsequent component matching logic can impose boundary constraints based on the individual's actual metabolic load, eliminating the solution suitability deviation caused by individual differences.

[0016] 3. By using adaptive compatibility weight calculation, a dose-effect safety protection mechanism based on logical closed loop is established. At the data processing level, this invention divides medicinal plant extracts into aggressive feature operators and defensive feature operators. Based on the interaction results of individual data and environmental integrals, the distribution weight of the operators is dynamically adjusted through a nonlinear penalty function. When the system determines that the user is in a high-sensitivity spatiotemporal interference window, it automatically reduces the weight of the main active component that inhibits melanin and proportionally increases the weight of auxiliary components that repair and anti-oxidation. This adaptive scheduling mechanism of weights makes the generation of personalized skin care solutions no longer a rigid stacking of rules, but a dynamic optimization process based on data risk constraints, achieving an automatic balance between effectiveness and safety in the recommendation results. Attached Figure Description

[0017] Figure 1 This is a flowchart of the AI ​​recommendation method for whitening skincare based on multi-source heterogeneous data fusion according to the present invention; Figure 2 This is a diagram illustrating the decision logic architecture of the deep Q-network and environmental stress integral of this invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A skin whitening and skincare AI recommendation method based on multi-source heterogeneous data fusion includes the following steps: Step S101: Obtain the sequence of internal parameters characterizing the user's physiological state and the sequence of external parameters characterizing the user's current geographic grid environment parameters; Step S102: Obtain the multidimensional spectral feature tensor of the user's target area, extract the phenotypic state features from it, and concatenate them with the internal factor parameter sequence and the external factor parameter sequence to construct a high-dimensional heterogeneous feature space. Step S103: In the high-dimensional heterogeneous feature space, call the historical population physiological data in the preset geographic grid, compare the individual transdermal water loss in the intrinsic parameter sequence with the baseline mean in the historical population physiological data, and calculate the individual parameter deviation. Step S104: Extract the time series of high-frequency environmental parameters within the geographic grid for a preset period, and generate an environmental stress cumulative integral characterizing the impact of environmental fluctuations through integral calculation; Step S105: Determine the active ingredient tolerance threshold of the target user under the current geographic grid based on the mapping relationship between individual parameter deviation and environmental stress cumulative integral. When the active ingredient tolerance threshold reaches the preset intervention threshold, the weight of the target component distribution weight vector is reduced by a nonlinear penalty function to reduce the weight of the main component that inhibits melanin and simultaneously compensate the weight of the repair auxiliary components. Step S106: Map the scheduled target component distribution weight vector to the library of authentic medicinal plant active ingredients to generate a personalized skin care solution with the ratio weight limited by the tolerance threshold of the active ingredients.

[0021] Preferably, step S104 is further refined into the following sub-steps: Step S1041, obtain the time series data of high-frequency environmental external factor parameters for the past 14 days within the geographic grid where the user is currently located, where the side length of the geographic grid is 5km; Step S1042, perform cumulative integration calculation on the PM2.5 index and ultraviolet intensity among the high-frequency environmental external factor parameters for the past 14 days to generate the cumulative environmental stress integral characterizing the environmental stress baseline shift within the geographic grid.

[0022] Preferably, step S105 involves weight reduction scheduling of the target component distribution weight vector using a nonlinear penalty function, which is further refined into the following sub-steps: Step S1051, dividing the elements in the target component distribution weight vector into aggressive feature operators and defensive feature operators; Step S1052, constructing a deep Q-network model that includes the current skin state feature vector, the component adjustment action space, and a reward function based on the objective indicator of improvement degree; Step S1053, using a reinforcement learning feedback mechanism, dynamically adjusting the distribution weights of the aggressive and defensive feature operators based on the determination result of the active ingredient tolerance threshold.

[0023] Preferably, the calculation of individual parameter deviation in step S103 includes the following sub-steps: step S1031, retrieving the baseline mean of transdermal water loss of the population matching the current user's age group; step S1032, when the measured value of individual transdermal water loss in the intrinsic parameter sequence is greater than the baseline mean, marking the individual parameter deviation as a positive shift.

[0024] Preferably, in step S102, obtaining the multidimensional spectral feature tensor of the user's target area includes: acquiring white light images, polarized light images, and ultraviolet light images of the facial skin; extracting texture features and analyzing pigment deposition depth on the white light images, polarized light images, and ultraviolet light images; and mapping the extracted depth features to the quantization dimension in the multidimensional spectral feature tensor.

[0025] Preferably, the intrinsic parameter sequence includes: age, sex, inflammatory status, genotype, lifestyle habits, individual transdermal water loss, and glycation level; the extrinsic parameter sequence includes: seasonal characteristics, geographic grid coordinates, ultraviolet exposure index, and air pollution index.

[0026] Preferably, in step S106, generating a personalized skincare plan includes: matching the active concentration ratio of Siegesbeckia orientalis, Dendrobium nobile, Curcuma longa, Eucommia ulmoides, Polygonatum sibiricum, and oleanolic acid according to the scheduled target component distribution weight vector; and generating a technical list characterizing the compatibility ratio of skincare components based on the active concentration ratio.

[0027] Preferably, under the logic loop of generating personalized skin care solutions, the method further includes step S107, which involves periodically collecting changes in the user's physiological data after applying the personalized skin care solution, and inputting this data as a feedback signal into the deep Q-network model to correct the preset response adjustment coefficient.

[0028] Preferably, in step S101, the intrinsic parameter sequence is obtained by collecting the user's structured physiological data through a mobile terminal; the extrinsic parameter sequence is obtained by calling the open data interfaces of the geographic information system and environmental monitoring stations.

[0029] Example 1: In industrial port cities continuously affected by extreme weather fluctuations, the system faces the challenge of users being exposed to both high-intensity ultraviolet radiation and high concentrations of PM2.5 for extended periods. Because static acquisition of phenotypic data in this environment cannot promptly reflect changes in the dynamic stress load of the skin's underlying layers, when the monitoring system detects a non-linear surge in the ultraviolet exposure index and air pollution index within the target geographic grid over the past 14 days, the user's physiological parameter sequence shows that the individual's transepidermal water loss (TEWL) deviates from its historical stable range. If a high concentration of whitening active ingredients is then distributed using static logic, it carries the risk of inducing skin barrier tolerance collapse and post-inflammatory hyperpigmentation. To address this technical challenge triggered by drastic environmental changes, this invention utilizes spatiotemporal big data with a side length of 5km within a preset geographic grid, extracting data from the past 1... A 4-day time series of high-frequency environmental parameters was used to calculate the cumulative integral of environmental stress, characterizing the impact of environmental fluctuations. The specific transmission pathway of this macro-meteorological integral into the skin's micro-stress state lies in the fact that continuously accumulating high doses of UVB and ultrafine particulate matter penetrating the stratum corneum induce mitochondrial oxidative damage within the epidermis. Based on the linear stimulation of epidermal keratinocytes by the integral dose to release inflammatory mediators such as interleukin-1α and tumor necrosis factor-α, a transformation function targeting this cumulative biochemical reaction pathway was systematically established. The energy of meteorological fluctuations within a 5-kilometer grid was equivalently mapped to the expected increase in the concentration of pro-inflammatory factors in the local microcirculation of the living organism. A correlation mapping model between external meteorological parameters and the boundary regulation of cellular receptor tolerance was established. Within the constructed high-dimensional heterogeneous feature space, the system accurately represented the measured values ​​of individual transdermal water loss. Compared with the baseline mean of the same age group in the cloud database Real-time comparison and calculation of individual parameter deviations; although the final output of this step is a scalar operation result, before this, more than ten structured feature values, such as age, genotype, and sex, which are not related to water loss, contained in the high-dimensional heterogeneous feature space have been used as strong constraints to perform a multi-dimensional vector distance search in the cloud-based physiological parameter database; the system uses a polynomial kernel function to filter out homogeneous sample clusters that are highly consistent with the target individual in all heterogeneous dimensions, and uses the statistical center value of this filtered specific sample cluster as the final dynamic baseline mean. This allows the single-line comparison calculation of water loss to implicitly incorporate the classification and calibration results of multi-dimensional, multi-source data.

[0030] To achieve dynamic constraints on the dose-response boundary of active components, the system utilizes a nonlinear penalty function to perform weight drop scheduling on the target component distribution weight vector, and calculates the adjusted aggressive feature operator weights, the expression of which is as follows: ,in, The adjusted weights for the aggressive feature operators. These are the initial weight values ​​generated from the initial feature space. This represents the measured value of transdermal water loss for an individual. This represents the baseline mean of transdermal water loss in the historical population. The initial feature operator weights are determined based on the environmental degradation factor, which is based on the cumulative integral of environmental stress. These weights are applied before the aforementioned weighted pressure reduction is performed. The precise generation logic is as follows: The system takes various physiological parameters and phenotypic feature values ​​in the high-dimensional heterogeneous feature space as inputs and feeds them into a gradient boosting decision tree model pre-trained using an in vitro transdermal absorption experimental dataset. This model calculates the information gain ratio of each dimension's feature value to the inhibition rate of tyrosinase activity in the melanin synthesis pathway, and normalizes the corresponding output feature gain ratios to directly construct the initial target component distribution weight vector without environmental stress feedback intervention. During the weight vector scheduling process, the system divides the target components into aggressive feature operators and defensive feature operators, and utilizes a feature vector containing the current skin state, the component adjustment action space, and a reward based on the degree of improvement of objective indicators. The deep Q-network model of the function dynamically suppresses the proportion of melanin-inhibiting main active components in Siegesbeckia orientalis, Dendrobium nobile, Curcuma longa, Eucommia ulmoides, Polygonatum sibiricum, and oleanolic acid, while simultaneously compensating for the proportion weight of repair-type auxiliary components. Through a reinforcement learning feedback mechanism, the system transforms the scheduled target component distribution weight vector to a library of authentic medicinal plant active ingredients based on the determination of the active ingredient tolerance threshold, generating a personalized skincare solution with proportion weights limited by the active ingredient tolerance threshold. At the execution level, this solution automatically switches to a low-irritation and high-repair component combination strategy during periods of high environmental stress, maintaining a dynamic balance in the skin microcirculation state facing inflammation risk without interrupting whitening intervention.

[0031] Example 2: In a machine learning verification environment deployed on a distributed high-performance heterogeneous computing platform, the system calls a database of 1200 skin physiological samples with multi-source heterogeneous attributes as the data source. This database covers industrial grid data with a geographical span of no less than 50km and a gradient distribution of environmental stress characteristics. The aim is to verify the accuracy of the environmental stress accumulation integral and individual parameter deviation in scheduling the distribution weight vector of active components under the drive of a nonlinear penalty function. The data acquisition cycle is set to 24h. The determination of this cycle is controlled by the technical trade-off between the timeliness of capturing physiological parameter fluctuations and the system data processing load. When the data acquisition frequency is lower than this threshold, the system cannot capture the transient stress peak caused by sudden environmental changes, while higher than this frequency leads to redundant addressing overhead in the high-dimensional heterogeneous feature space. The experimental signal source actively superimposes Gaussian white noise with a signal-to-noise ratio of 20dB to simulate the measurement random error of physical sensors in complex electromagnetic environments, and records the physical response of each sample group as the UV index increases from 3 to 11.

[0032] In the baseline setting phase, the control group used static mapping logic without environmental stress regulation, while the sample group of this invention was connected to a scheduling model that included cumulative environmental stress integration, with its initial individual transdermal water loss. The mean is Historical population physiological data baseline mean for Initial radical feature operator weights The calibration value was set at 0.85. As the external PM2.5 index increased from 50 to 250, the proportion of the main melanin-inhibiting component in the skincare regimen output by the control group remained between 15% and 18%. The skin erythema index of this group increased from 12 to 38, reaching a performance degradation inflection point on day 7. The skin barrier damage score rose from the initial 12 to 38, indicating that the static mapping logic could not suppress the negative feedback generated by environmental stress. The sample group of this invention initiated nonlinear penalty function scheduling, with an environmental decay factor... As the environmental stress intensity increases, it decreases from 100 to 25. Substituting these values ​​into the weighted pressure drop formula, intermediate state data is obtained.

[0033] When individual parameters deviate for and When the value is 25, the adjusted aggressive feature operator weights generated by the system The pressure dropped from 0.85 to 0.12, where, The adjusted weights for the aggressive feature operators. These are the initial weight values. This represents the measured value of transdermal water loss for an individual. This represents the baseline mean of historical population physiological data. The environmental degradation factor is determined based on the cumulative integral of environmental stress. The weighting of Siegesbeckia orientalis extract and Curcuma longa extract decreases synchronously in the output of the scheme, while the weighting of repair-related auxiliary components increases from 5% compensation to 22%. As the intensity of environmental stress increases sequentially, the output of the scheme of this invention... The results showed a downward trend with a monotonically positive correlation with the environmental attenuation factor. When the cumulative integral of environmental stress exceeded 1.5 times the preset intervention threshold, the weight entered the saturation inhibition zone, and the rate of change tended to level off, thus avoiding secondary intervention of the highly active components on the damaged barrier. The data showed that the erythema index of the audience's skin remained stable below 15 on the 14th day, confirming the effectiveness of the dynamic weight allocation algorithm driven by physical perception data in maintaining the homeostasis of the physiological barrier.

[0034] Example 3: In the application of border high-altitude monitoring stations deployed at altitudes exceeding 3000m and where ultraviolet radiation intensity is consistently above level 10, the system faces the challenge of continuous decline in skin barrier function due to extremely low air humidity and extremely high photochemical stress. Under this environment, the transdermal water loss... Transient fluctuations can easily mask long-term physiological deterioration trends. If the system fails to discretize and weight the environmental exposure over the past 14 days, it will lead to undersampling errors in the calculation of the cumulative integral of environmental stress. Consequently, the generated tolerance threshold for active ingredients cannot accurately map the user's true tolerance boundary, resulting in the risk of issuing highly active and aggressive components and inducing physical barrier damage. To address the technical challenges triggered by this specific environmental characteristic, this invention obtains meteorological observation sequences with a sampling frequency of 1 hour within a geographic grid, and continuously calculates the ultraviolet index. With fine particulate matter concentration The system maps the stress to discrete time-step operators and calculates the cumulative integral of environmental stress through summation. The system pre-defines the environmental degradation factor under high-altitude conditions using initial state definitions in the background. The baseline value is determined by measuring the deviation of atmospheric transparency from saturated water vapor pressure within a geographic grid, ensuring... The value of this environmental attenuation factor τ provides real-time feedback on the weight of skin tolerance loss in the current spatiotemporal dimension. In the actual system calibration, the boundary setting of this environmental attenuation factor τ is supported by high and low temperature alternating exposure experiments on in vitro skin models. The experiments show that when the applied external stress energy exceeds the limit and causes a thermodynamic phase transition in the lipid barrier, the skin's bearing capacity undergoes a nonlinear jump-like weakening. Based on the peak point of barrier damage endothermic point determined by this experiment, the system defines the ideal steady-state parameter without additional environmental load intervention as an upper limit of 100 at the physical algorithm level, and calibrates the lower limit of 25 when atmospheric stress increases drastically and causes a large amount of core structural protein degradation in a decompensated state. The slope of its change follows the attenuation rate equation of the stratum corneum impedance test, ensuring that the system has a real engineering judgment limit.

[0035] In the process of generating decision logic in the deep Q-network model, the system implants a process judgment quantification procedure and defines a reward function based on the skin barrier repair rate. Its calculation logic relies on the descent gradient of the erythema index and Synchronous monitoring of regression curves; specifically, when the system identifies... Persistent deviation from the population baseline mean When the amplitude exceeds 1.5 times the standard deviation, the system triggers a pressure drop command within the action space, using a nonlinear penalty function to exponentially decay the weights of the aggressive feature operator; specifically, the tolerance threshold of the active ingredient is determined by calculating the ratio of individual parameter deviation to the environmental attenuation factor, and the adjusted weights of the aggressive feature operator are determined according to the following formula: ,in, The adjusted weights for the aggressive feature operators. These are the initial weight values ​​generated from the initial feature space. This represents the measured value of transdermal water loss for an individual. This represents the baseline mean of transdermal water loss in the historical population. It is a dimensionless environmental degradation factor calibrated based on the cumulative integral of environmental stress.

[0036] After completing the weighted scheduling, the system will The physicochemical characteristic matrix of the authentic medicinal plant active ingredient library is injected into the system. This matrix stores the effective mass percentage distribution of sieboldii from Siegesbeckia orientalis, curcumin from Curcuma longa, and oleanolic acid. By calculating the dot product of the weight vector and the component concentration vector, a personalized skincare solution with the ratio weight limited by the tolerance boundary is generated. On the physical execution side, this solution reduces the concentration of the main whitening component from 2.5% under normal conditions to 0.8%, and simultaneously introduces the mucopolysaccharide component from Dendrobium nobile extract to enhance the water retention capacity of the stratum corneum. Through the synergistic effect of this discretized integral and nonlinear penalty scheduling, the system reduces the transdermal water loss of individual users on the 21st day after the surge in ultraviolet stress at high altitudes. The physiological barrier homeostasis was protected by dynamically allocating algorithm weights driven by physical perception data, which stabilized and returned to a safe range within 1.1 times the baseline mean.

[0037] Example 4: When the system faces signal drift in the acquisition accuracy of physical sensors due to environmental electromagnetic noise, the system uses a calibration procedure with a sampling window length of 72 hours in an offline environment. It then uses an information entropy filtering algorithm to remove abnormal discrete points with a signal-to-noise ratio below 15dB, establishing a feature distribution characterizing the skin barrier features of different age groups. Specifically, the system acquires the measured transdermal water loss of 1000 physiological samples within the target geographic grid, and determines the baseline mean of historical group physiological data by calculating the second-order statistical moments of the sample sequences. Furthermore, a dynamic drift weight is established for the cumulative integral of environmental stress with respect to the baseline mean, expressed mathematically as follows: ,in, These are the corrected physiological baseline values. This represents the baseline mean of historical population physiological data. The sensitivity coefficient is calibrated based on the sensor's sampling accuracy. The cumulative integral of the current environmental stress is obtained by accumulating the environmental external factors within the discrete time step. This is a preset stress intervention threshold.

[0038] When the system encounters quantification biases in the component library processing mapping, it uses a parameter calibration process based on component concentration gradient verification to determine the conversion ratio between the target component distribution weight vector (composed of Siegesbeckia orientalis, Clematis chinensis, Dendrobium nobile, Curcuma longa, Eucommia ulmoides, Polygonatum sibiricum, and oleanolic acid) and the active monomer concentration. An incremental permeation experiment with a concentration gradient of 0.1% is then applied to an in vitro reconstructed skin model to measure the transdermal absorption rate and extracellular acidification rate of the active components under different weight distributions, establishing the adjusted aggressive feature operator weights. The system establishes a mapping relationship to mass concentration; it utilizes the reward function of a deep Q-network model to impose discretization constraints on the action space, locking the output concentration range of the main active ingredient inhibiting melanin between 0.5% and 3.5%. The environmental decay factor of the nonlinear penalty function is calibrated by monitoring physiological response mutation points triggered by the tolerance threshold of the active ingredient. This ensures that the component ratios of the generated personalized skincare solutions remain consistent through automatic calibration when raw materials are replaced in different batches. The algorithm logic driven by physical calibration data completes the positioning of the dose-effect boundary of active ingredients.

[0039] Example 5: In the system initialization calibration application deployed on multi-channel synchronous acquisition nodes, the system faces the condition of heterogeneous data dimension mismatch. A deep Q-network model containing three fully connected layers is used to construct the multidimensional spectral feature tensor of the target area. Converted to a 128-dimensional feature vector and compared with an 8-dimensional intrinsic parameter sequence and 6-dimensional external factor parameter sequence The features are concatenated to form a 142-dimensional fused feature vector, which serves as the input state for the deep Q-network model. The number of neurons in the three fully connected layers is set to 256, 128, and 64 respectively from the input side to the output side. This multidimensional spectral feature tensor is then generated. In the processing steps, the system acquires the facial image matrix of the audience under white light, polarized light, and ultraviolet light collected by the hardware photoelectric sensor array. It then uses the gray-level co-occurrence matrix algorithm to extract the surface texture energy and contrast values ​​of the stratum corneum from the white light and polarized light matrices. Simultaneously, for the specific wavelength absorption spectrum of the ultraviolet light image, the system calculates the inverted melanin concentration distribution using the Beer-Lambert law, transforming it into quantized depth coordinates characterizing the depth of melanin in the subcutaneous physical tomography. The extracted texture energy, contrast values, and depth coordinates are then concatenated along the channel dimension to generate a quantized spectral feature tensor for subsequent calculations. The system then uses a reward function based on the steady-state regression rate of the skin barrier. The mathematical expression for the reward function of the calibrated offline model is as follows: ,in, As a reward value, and These are the dimensionless weighting coefficients determined based on sensitivity analysis of the offline validation set. This represents the measured value of transdermal water loss for an individual. This represents the baseline mean of historical population physiological data. The rate of change of erythema index at the recipient site.

[0040] When the system encounters batch-to-batch variations in the physicochemical properties of the target components during the active component distribution generation process, the system uses an adaptive parameter matrix generation procedure to convert the components in the scheduled target component distribution weight vector into physical mass concentrations. The mapping relationship is as follows: ,in, For the first The percentage by mass of each active ingredient in the skincare regimen. For the adjusted number Weights of feature operators, The correction factor is determined by the purity of effective monomers in the library of active ingredients from authentic medicinal plants. For the first The system uses a hash mapping algorithm to de-identify user physiological characteristics and, when the weight of the main active ingredient inhibiting melanin in *Siegesbeckia orientalis*, *Dendrobium nobile*, *Curcuma longa*, *Eucommia ulmoides*, *Polygonatum sibiricum*, and oleanolic acid is reduced to below 0.15, it triggers the ratio compensation logic of *Dendrobium nobile* extract and *Eucommia ulmoides* extract, so that the transepidermal water loss rate of the target area is maintained at a certain level after the intervention of the peak environmental stress. Furthermore, with a variance of less than 0.5, the skincare solution generation logic and the skin's physical tolerance boundary are in a steady-state closed loop. In this execution phase, the hash mapping not only encrypts the data, but the system further inputs the generated fixed-length hash ciphertext into the access-controlled secure ratio controller, which serves as the unique index key for retrieving the offline pharmacological attribute comparison dictionary, matching the lower limit of the hard correction threshold representing the vulnerability of the underlying stratum corneum of this latent physiological group. When the system detects that the allocation weight of the melanin inhibition operator is close to the correction threshold obtained above and is reduced to below 0.15, the central processing unit directly converts this state into a digital control pulse and sends it to the mixing terminal actuator, thereby forcibly opening the servo intervention valve for the ratio of repair extracts at the physical component output end.

[0041] In the standard procedure involving the characterization of active ingredients in a database of authentic medicinal plants, the system traces the extraction process of effective monomers in Siegesbeckia orientalis, Dendrobium nobile, Curcuma longa, Eucommia ulmoides, Polygonatum sibiricum, and oleanolic acid. Ultrasonic-assisted extraction is used at a frequency of 40kHz and a solid-liquid ratio of 1:15 for 45 minutes. The effective mass percentage distribution of each component is obtained as the input benchmark for the physicochemical characteristic matrix. The system determines the purity of Siegesbeckia orientalis and curcumin using high-performance liquid chromatography (HPLC), compares the measured concentrations with the standard fingerprint spectrum in the cloud database, and calculates correction factors reflecting the differences in activity between batches of raw materials. This is used to compensate for quality concentration fluctuations caused by the place of origin, ensuring that the generated formulation weights objectively map the melanin inhibition effect at the physical level. After the system completes the mapping of the target component distribution weights to the active ingredient library, the system generates the final bill of materials based on the component ratios of the personalized skincare plan. This includes selecting Dendrobium nobile extract with epidermal repair function as a carrier for defensive feature operators, and forming a microemulsion system with a particle size distribution average of 200nm by adjusting the emulsification pressure within the range of 30MPa to 45MPa, so that the transdermal penetration rate of the active ingredients is maintained at a certain level. Moreover, the fluctuation range is less than 10%, and the risk of skin tolerance breakdown is avoided through the synergy of physical process parameters and algorithm weights.

[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A skin whitening and skincare AI recommendation method based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: Step S101: Obtain the sequence of internal parameters characterizing the user's physiological state and the sequence of external parameters characterizing the user's current geographic grid environment parameters; Step S102: Obtain the multidimensional spectral feature tensor of the user's target area, extract the phenotypic state features from it, and concatenate them with the internal factor parameter sequence and the external factor parameter sequence to construct a high-dimensional heterogeneous feature space. Step S103: In the high-dimensional heterogeneous feature space, call the historical population physiological data in the preset geographic grid, compare the individual transdermal water loss in the intrinsic parameter sequence with the baseline mean in the historical population physiological data, and calculate the individual parameter deviation. Step S104: Extract the time series of high-frequency environmental parameters within the geographic grid for a preset period, and generate an environmental stress cumulative integral characterizing the impact of environmental fluctuations through integral calculation; Step S105: Determine the active ingredient tolerance threshold of the target user under the current geographic grid based on the mapping relationship between individual parameter deviation and environmental stress cumulative integral. When the active ingredient tolerance threshold reaches the preset intervention threshold, the weight of the target component distribution weight vector is reduced by a nonlinear penalty function to reduce the weight of the main component that inhibits melanin and simultaneously compensate the weight of the repair auxiliary components. Step S106: Map the scheduled target component distribution weight vector to the library of authentic medicinal plant active ingredients to generate a personalized skin care solution with the ratio weight limited by the tolerance threshold of the active ingredients.

2. The skin whitening AI recommendation method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, Step S104 is further refined into the following sub-steps: Step S1041, obtain the time series data of high-frequency environmental external factors parameters in the geographic grid where the user is currently located for the past 14 days. The side length of the geographic grid is 5km. Step S1042, perform cumulative integration calculation on the PM2.5 index and ultraviolet intensity in the high-frequency environmental external factors parameters in the past 14 days to generate the cumulative environmental stress integral characterizing the environmental stress baseline shift in the geographic grid.

3. The AI ​​recommendation method for skin whitening based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In step S105, the weights of the target component distribution weight vector are scheduled by weight reduction through a nonlinear penalty function, which is further refined into the following sub-steps: Step S1051, the elements in the target component distribution weight vector are divided into aggressive feature operators and defensive feature operators; Step S1052, a deep Q-network model is constructed that includes the current skin state feature vector, the component adjustment action space, and a reward function based on the objective indicator of improvement degree; Step S1053, the distribution weights of the aggressive feature operators and the defensive feature operators are dynamically adjusted according to the determination result of the active ingredient tolerance threshold using a reinforcement learning feedback mechanism.

4. The AI ​​recommendation method for skin whitening based on multi-source heterogeneous data fusion according to claim 1, characterized in that, Step S103 calculates the individual parameter deviation, including the following sub-steps: Step S1031, retrieve the baseline mean of transdermal water loss of the population matching the current user's age group; Step S1032, when the measured value of individual transdermal water loss in the intrinsic parameter sequence is greater than the baseline mean, mark the individual parameter deviation as a positive shift.

5. The AI ​​recommendation method for skin whitening based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In step S102, the multidimensional spectral feature tensor of the user's target area is obtained, including: acquiring white light images, polarized light images, and ultraviolet light images of the facial skin; extracting texture features and analyzing pigment deposition depth on the white light images, polarized light images, and ultraviolet light images; and mapping the extracted depth features to the quantization dimension in the multidimensional spectral feature tensor.

6. The whitening skincare AI recommendation method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The intrinsic parameter sequence includes: age, sex, inflammatory status, genotype, lifestyle habits, individual transdermal water loss, and glycation level; the extrinsic parameter sequence includes: seasonal characteristics, geographic grid coordinates, ultraviolet exposure index, and air pollution index.

7. The AI ​​recommendation method for skin whitening based on multi-source heterogeneous data fusion according to claim 1, characterized in that, Step S106 generates a personalized skincare plan, including: matching the active concentration ratio of Siegesbeckia orientalis, Dendrobium nobile, Curcuma longa, Eucommia ulmoides, Polygonatum sibiricum and oleanolic acid according to the scheduled target component distribution weight vector; and generating a list of technologies characterizing the compatibility ratio of skincare components based on the active concentration ratio.

8. The AI ​​recommendation method for skin whitening based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The logic loop for generating personalized skincare solutions also includes step S107, which involves periodically collecting physiological data changes of users after applying personalized skincare solutions, inputting these data as feedback signals into a deep Q-network model, and correcting the preset response adjustment coefficients.

9. The AI ​​recommendation method for skin whitening based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In step S101, the intrinsic parameter sequence is obtained by collecting the user's structured physiological data through a mobile terminal; the extrinsic parameter sequence is obtained by calling the open data interfaces of the geographic information system and environmental monitoring stations.

Images