Makeup waterproof performance evaluation method and system based on deep learning

Abstract

The invention relates to the technical field of make-up waterproof performance evaluation, and discloses a make-up waterproof performance evaluation method and system based on deep learning, and the method comprises the steps: constructing a microstructure diagram; constructing and training a space-time diagram convolutional network; constructing a hierarchical reinforcement learning framework; constructing a structure sensitivity prediction module; implementing a test state self-adaptive conversion mechanism; the predictive learning ability is realized; the problems that a traditional static test method cannot reflect a real use scene, microscopic and macroscopic view angles are separated, the test efficiency is low, and the prediction capacity is insufficient are solved, and dynamic, efficient and accurate waterproof performance evaluation is achieved through the deep learning technology.

Description

Technical Field

[0001] This invention relates to the field of waterproof performance evaluation technology for color cosmetic products, and more specifically, to a method and system for evaluating the waterproof performance of color cosmetics based on deep learning. Background Technology

[0002] The waterproof performance of makeup products is a key indicator that consumers pay close attention to, directly affecting the user experience and market competitiveness. Currently, the industry mainly uses the following methods to evaluate the waterproof performance of makeup: First, static immersion testing, where makeup products are applied to a specific substrate and then immersed in water to observe color changes or peeling; second, simulated sweat testing, where makeup samples are exposed to artificial sweat to assess their stability; and third, mechanical friction combined with a humid environment test, simulating friction and water contact in daily use.

[0003] However, existing methods for evaluating the waterproof performance of makeup products have significant shortcomings:

[0004] 1. Limitations of Static Testing: Traditional methods for evaluating the waterproof performance of cosmetics typically employ static testing, assessing product performance only under fixed conditions. This fails to reflect the dynamic environmental changes and cumulative effects in real-world usage scenarios. For example, consumers may experience complex environments such as temperature variations, humidity fluctuations, and changes in motion during actual use, which static testing struggles to simulate.

[0005] 2. Disconnect between Microscopic and Macroscopic Effects: Existing evaluation methods either focus on macroscopic effects (such as overall waterproofing duration and color retention) or on microscopic structures (such as molecular arrangement stability and interfacial tension changes), lacking a method that organically combines the two. This disconnect prevents the establishment of a correlation mechanism between microscopic structural changes and macroscopic performance, limiting the understanding of the essence of waterproofing performance.

[0006] 3. Low testing efficiency: Traditional testing methods typically require numerous repetitive trials, are time-consuming and resource-intensive, and cannot optimize test paths for specific product characteristics. In an industry trend of increasingly shorter product development cycles, inefficient testing methods have become a bottleneck restricting the speed of innovation.

[0007] 4. Insufficient predictive capability: Existing methods struggle to predict a product's waterproof performance under long-term or extreme conditions based on short-term test results, potentially leading to unexpected performance issues after the product's market launch. This insufficient predictive capability increases product quality risks and raises subsequent improvement costs. Summary of the Invention

[0008] This invention provides a method and system for evaluating the waterproof performance of makeup based on deep learning, which solves the technical problems of traditional static testing methods in related technologies, such as inability to reflect real-world usage scenarios, separation of micro and macro perspectives, low testing efficiency, and insufficient predictive ability.

[0009] This invention provides a deep learning-based method for evaluating the waterproof performance of makeup, comprising:

[0010] High-resolution microscopic images of cosmetic products are acquired, the images are processed using an image segmentation algorithm to construct a microstructure graph, and attributes are assigned to the nodes and edges of the microstructure graph.

[0011] A spatiotemporal graph convolutional network model is constructed and trained to realize the modeling and analysis of the dynamic changes of the microstructure graph;

[0012] Construct a hierarchical reinforcement learning framework, including defining the state space, action space and reward function, building high-level and low-level policy networks, and generating intelligent testing policies;

[0013] A structure-sensitive prediction module was constructed to realize the mapping prediction from microstructural changes to macroscopic waterproof performance;

[0014] Implement an adaptive test state transition mechanism to dynamically adjust test conditions based on real-time feedback and simulate real-world usage scenarios;

[0015] It enables predictive learning capabilities, allowing the prediction of long-term waterproofing performance degradation through minute initial structural changes;

[0016] The hierarchical reinforcement learning framework optimizes the testing strategy by optimizing the objective function, which comprehensively considers the reward value obtained at each time step in the test trajectory and the accuracy constraint of microstructure prediction.

[0017] Furthermore, the step of processing the image using an image segmentation algorithm to construct a microstructure map includes:

[0018] The resolution of the image segmentation algorithm is set to 5-50 nanometers. The segmentation threshold is selected according to the type of makeup product. The segmentation threshold for mascara is 0.15-0.25, the segmentation threshold for foundation is 0.20-0.35, and the segmentation threshold for lipstick is 0.10-0.20.

[0019] The segmented image is converted into a graph data structure, where nodes represent molecular clusters or functional units in cosmetic products, and edges represent physical or chemical connections between nodes.

[0020] The node and edge attributes of the microstructure graph are preprocessed, including normalizing numerical features, encoding categorical features, smoothing and denoising time-series data, and imputing missing values.

[0021] Furthermore, the steps for constructing and training a spatiotemporal graph convolutional network model include:

[0022] Construct a multi-layer spatiotemporal graph convolutional network model, including spatial graph convolutional layers, temporal convolutional layers, and an attention mechanism;

[0023] Design node update functions, including aggregation functions and non-linear activation functions;

[0024] Set the key network parameters, including kernel size of 3-7, network depth of 4-8 layers, feature dimension of 64-256, learning rate of 0.0001-0.001, and batch size of 16-64.

[0025] The network was trained using historical test data that included standard environment datasets, extreme environment datasets, and mixed environment datasets.

[0026] Optimize network parameters, including structure prediction loss, performance prediction loss, and regularization term, by employing adaptive weight adjustment and gradient pruning techniques.

[0027] Furthermore, the steps to construct a hierarchical reinforcement learning framework include:

[0028] Define the state space, including the current state of the microstructure graph, historical test trajectories, and environmental parameters;

[0029] Define the action space, including test parameter adjustment actions and test termination judgment actions;

[0030] Define a reward function that comprehensively considers information gain, testing efficiency, and prediction accuracy. Calculate the overall reward value through a weighted combination method, where each weight coefficient is non-negative and sums to 1. The weights for information gain range from 0.3 to 0.5, testing efficiency from 0.2 to 0.4, and prediction accuracy from 0.2 to 0.4.

[0031] Build a high-level strategy network to determine the testing phase and overall strategy;

[0032] Construct a low-level policy network responsible for generating specific test parameters;

[0033] The model is trained using the policy gradient method to optimize the objective function, where time step t is defined as the discrete test time, the discount factor γ ranges from 0.9 to 0.99, and the test termination time step T is set to a maximum value according to different product types.

[0034] Furthermore, the steps for constructing the structure-sensitive prediction module include:

[0035] Define a key structural feature extraction function to extract features directly related to waterproof performance from the microstructure diagram, including molecular arrangement density, cross-linking strength, surface tension, degree centrality, clustering coefficient, bonding strength and intermolecular distance, and standardize all features to map them to the [-1,1] interval;

[0036] A structure-performance mapping model is constructed, including a shared feature extractor, a task-specific layer, a cross-task attention mechanism, and an output decoding and transformation module, which maps the extracted structural features to waterproof performance indicators, including waterproof durability, sweat resistance, and abrasion resistance.

[0037] The effectiveness of the structure-sensitive prediction module is verified by visualizing the correlation between the predicted values ​​and actual measured values ​​of the microstructure prediction model using scatter plots.

[0038] Radar charts are used to display the performance of different products across multiple waterproofing dimensions, enabling multi-dimensional waterproofing performance evaluation.

[0039] Furthermore, the steps to implement the test state adaptive transition mechanism include:

[0040] Configure a status monitoring module to collect microscopic images and performance data of makeup samples in real time and update the microstructure diagram;

[0041] Construct a state evaluation function to evaluate the information value and potential weaknesses of samples under the current test state. This includes calculating the similarity between the current microstructure state and samples in the historical test database, analyzing the gap between the current state and the expected performance target, evaluating the stability of the current state, and calculating the expected gain of acquiring new information under the current test conditions.

[0042] Using a hierarchical reinforcement learning model, the system generates the next optimal test action based on the state evaluation results. The system automatically determines whether manual intervention is needed based on the confidence threshold. When the prediction confidence is below 0.75, the system automatically triggers an intervention request.

[0043] The test conditions are automatically adjusted by the actuator, including changing the temperature, humidity, water pressure or simulating motion.

[0044] Build a test trajectory recording system to record the complete test process and provide data support for subsequent analysis.

[0045] Furthermore, the steps to achieve predictive learning capabilities include:

[0046] Construct a multi-scale time decomposition network to separate short-term fluctuations from long-term trends;

[0047] Configure a sensitivity analysis module to identify microstructural change patterns that have the greatest impact on long-term performance;

[0048] Construct a long-term prediction model, extrapolating long-term performance based on short-term test data, using the formula:

[0049] S(t long )=S(t short )·f decay (t long -t short G short ,ΔG)

[0050] Calculate the long-term waterproof performance score, where S(t) long S(t) represents the long-term waterproof performance score. short ) represents the short-term waterproof performance score, t long t represents a long-term time interval. short Indicates a short-term test moment, G short This represents a microstructure diagram in a short-term test, where ΔG represents the rate of change of the microstructure, and f decay This is the performance degradation function;

[0051] Wherein, the performance degradation function f decay The value range is strictly constrained within the interval (0,1), and the value range of the time scale parameter α is [0.001,0.1], which is adjusted according to different types of makeup products.

[0052] Furthermore, the performance degradation function f decay The specific implementation is as follows:

[0053] f decay (t long -t short G short ,ΔG)=exp(-α·R eval (G short ,ΔG)·(t long -t short ))

[0054] Where α is the time scale parameter, exp represents the exponential function, and R0... eval This is a structural change evaluation function used to assess the impact of microstructural changes on performance;

[0055] Structural change evaluation function R eval The output value is strictly constrained within the range of [0.1, 10], where 0.1 indicates that the structural change has a very small impact on the performance, and 10 indicates that the structural change has a very large impact on the performance.

[0056] When the topological change rate of the microstructure exceeds 30%, the system automatically adjusts R. evalThe value is set to the maximum of 10;

[0057] When the rate of change of the critical node exceeds 15%, R eval The value is not lower than 5.0;

[0058] When the microstructure remains stable, R eval The value should not exceed 0.5.

[0059] Furthermore, the technical effects of the method include:

[0060] Testing efficiency is improved by using spatiotemporal graph convolutional network technology to analyze microstructural changes in real time, resulting in a significant reduction in the testing cycle.

[0061] The testing accuracy has been greatly improved by using microstructure analysis technology to guide the optimization of testing strategies.

[0062] Achieving dynamic scenario simulation, the test state adaptive transition mechanism successfully simulates complex dynamic changes in the real-world usage environment;

[0063] It has predictive capabilities; through predictive learning technology, the system can predict the waterproof performance of a product under long-term or extreme conditions based on short-term test data.

[0064] The intelligent testing path, through hierarchical reinforcement learning technology, enables the system to automatically explore and identify critical points and weaknesses in product performance;

[0065] Resource utilization is optimized by using microstructure diagram construction technology to guide test path selection and reduce the exploration of invalid test paths.

[0066] This invention provides a deep learning-based system for evaluating the waterproof performance of makeup, used to execute the aforementioned deep learning-based method for evaluating the waterproof performance of makeup, including:

[0067] The image acquisition module is used to acquire high-resolution microscopic images of makeup products;

[0068] The image processing module is used to process images using image segmentation algorithms and construct microstructure maps.

[0069] The attribute assignment module is used to assign attributes to nodes and edges of the microstructure graph;

[0070] The spatiotemporal graph convolutional network module is used to model and analyze the dynamic changes of microstructures.

[0071] A hierarchical reinforcement learning module is used to generate intelligent testing strategies;

[0072] The structure-sensitive prediction module is used to realize the mapping prediction from microstructural changes to macroscopic waterproof performance;

[0073] The test state adaptive switching module is used to dynamically adjust test conditions based on real-time feedback;

[0074] A predictive learning module is used to predict long-term degradation of waterproofing performance through minute initial structural changes;

[0075] Test execution device, used to automatically adjust test conditions;

[0076] The test trajectory recording system is used to record the entire test process and provide data support for subsequent analysis.

[0077] The beneficial effects of this invention are that it achieves dynamic, efficient, and accurate evaluation of waterproof performance through deep learning technology. Attached Figure Description

[0078] Figure 1 This is a flowchart of the deep learning-based method for evaluating the waterproof performance of makeup according to the present invention;

[0079] Figure 2 This is a line graph showing the degradation of the waterproof performance of different eyeliner gel products in a water immersion environment, illustrating the degradation of the waterproof performance of three different eyeliner gel products over time in a continuous water immersion environment.

[0080] Figure 3 This is a bar chart comparing the performance of the traditional testing method of this invention with the micro-guided self-evolving test network, showing the performance differences between the traditional testing method and the micro-guided self-evolving test network of this patent in multiple dimensions.

[0081] Figure 4 This is a scatter plot showing the correlation between the predicted values ​​and actual measured values ​​of the microstructure prediction model of the present invention, illustrating the correlation between the waterproof performance score predicted by the microstructure prediction model and the actual measured waterproof performance score.

[0082] Figure 5 This is a tree diagram illustrating the weight distribution of the hierarchical reinforcement learning reward function of the present invention, showing the weight distribution of the reward function in the hierarchical reinforcement learning framework. Detailed Implementation

[0083] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0084] Example 1

[0085] This implementation provides a supply chain procurement optimization method based on big data, such as... Figure 1 As shown, the method includes the following steps:

[0086] Step 1, Construction of microstructure diagram;

[0087] High-resolution image data of makeup samples were collected to construct a microstructure diagram model, which is used to represent the microstructure characteristics and interaction relationships of makeup products.

[0088] Step 1.1: Acquire microscopic images of makeup samples using high-resolution microscopic imaging equipment;

[0089] Acquire raw image data containing information such as particle distribution and molecular arrangement;

[0090] Optionally, the collection process can be carried out under different lighting conditions to obtain more comprehensive sample information.

[0091] Step 1.2: Apply an image segmentation algorithm to process the original image;

[0092] Identify and extract microscopic structural units such as particles and molecular clusters in makeup;

[0093] In some implementations, deep learning segmentation algorithms such as U-Net or Mask R-CNN can be used for more accurate structural unit identification.

[0094] Furthermore, the resolution of the image segmentation algorithm is set to 0.1-0.5μm to ensure the capture of the smallest particle structure;

[0095] The segmentation threshold ranges from [0.65, 0.85], where 0.65 represents a loose segmentation (prioritizing structural integrity) and 0.85 represents a strict segmentation (prioritizing the accuracy of structural boundaries).

[0096] For different types of makeup products, the system will automatically select the optimal segmentation parameters: a threshold range of 0.75-0.85 is used for powder products, and a threshold range of 0.65-0.75 is used for liquid products.

[0097] Step 1.3, based on the extracted microstructural units;

[0098] Construct a graph data structure: G = (V, E);

[0099] Where G represents the graph data structure, V represents the set of nodes (corresponding to microstructural units), and E represents the set of edges (corresponding to physical or chemical interactions between units);

[0100] Step 1.4: Assign attributes to the nodes and edges of the graph structure;

[0101] Node attributes include features such as size, shape, and material, while edge attributes include information such as interaction type and strength.

[0102] It should be noted that the properties of the edges can be calculated through a physical model or inferred from the image data through a deep learning model.

[0103] Before assigning attributes, comprehensive data preprocessing of the original data is required, including:

[0104] Numerical feature preprocessing: Min-Max normalization is performed on node features of different dimensions (such as size in micrometers and material in density values) to map all values ​​to the [0,1] interval, ensuring that different features have the same weight in subsequent calculations; For features with outliers, such as particle size distribution, Robust normalization is used to reduce the impact of outliers on the model.

[0105] Categorical feature encoding: For categorical attributes such as interaction type and molecular configuration type, one-hot encoding is used to convert them into numerical representations for easier network processing; for categorical features with order relationships, such as structural stability level, label encoding is used to preserve the order relationship.

[0106] Time series data processing: For sequence data that characterizes the changes in microstructure over time, a sliding window technique is used for segmentation, and differential transformation is used to eliminate trends and highlight short-term change characteristics;

[0107] Missing value handling: For missing data that may occur during the collection process, the mean / median imputation (for continuous features) or mode imputation (for discrete features) methods are used to fill in the missing data according to the feature type to ensure data integrity.

[0108] Step 2, construction and training of the spatiotemporal graph convolutional network;

[0109] A spatiotemporal graph convolutional network is constructed to analyze the dynamic evolution of microstructure graphs over time and with changes in the environment.

[0110] Step 2.1: Construct a multi-layer spatiotemporal graph convolutional network model;

[0111] The multi-layer spatiotemporal graph convolutional network model includes spatial graph convolutional layers and temporal convolutional layers, which are used to capture the spatial features and temporal evolution features of the microstructure of makeup.

[0112] The specific implementation includes:

[0113] Spatial graph convolutional layer: Employs a message-passing algorithm to update the representation of the central node by aggregating neighbor node information;

[0114] Temporal convolutional layer: Employs one-dimensional convolution operations to handle the evolution of node features over time;

[0115] Attention mechanism: Introduce node-level attention and edge-level attention to highlight important structural changes;

[0116] It should be understood that the spatiotemporal graph convolutional network provided in this application is not limited to the above structure, and can be flexibly adjusted according to specific application scenarios, such as adding skip connections or residual connections to improve network performance.

[0117] Step 2.2, Design the network structure; the network structure includes node update functions:

[0118]

[0119] in, and Let N(v) represent the features of node v in layer l and layer l+1, respectively, and let W represent the set of neighbors of node v. (l) and B (l) Let these represent the weight matrix and bias matrix of the l-th layer, respectively. Let represent the features of the l-th layer neighbor node u, where u represents the index of the neighbor node, AGGREGATE represents the aggregation function used to summarize the feature information of the neighbor nodes, and σ is a non-linear activation function. The specific implementation of the aggregation function AGGREGATE is as follows:

[0120] First, the attention weights from each neighbor node u to the center node v are calculated by inputting the node pair features into a small multilayer perceptron and normalizing them using the softmax function.

[0121] Then, these weights are used to sum the features of neighboring nodes to obtain an aggregated representation of the neighboring information;

[0122] Finally, the aggregated neighbor information is combined with the central node's own information, and a new feature representation of the node is obtained through nonlinear transformation;

[0123] The specific implementation of the nonlinear activation function σ is the ReLU function, that is:

[0124] σ(x) = max(0,x)

[0125] This function can introduce nonlinear transformation capabilities, enhancing the network's ability to express complex features.

[0126] Furthermore, the key parameter constraints for the spatiotemporal graph convolutional network are as follows:

[0127] The kernel size of the graph convolutional layer ranges from [1,3], with a standard configuration of 2.

[0128] The kernel size of the temporal convolutional layer ranges from [3, 9], with a standard configuration of 5.

[0129] The network depth (number of layers) ranges from [4, 12] layers, with a standard configuration of 8 layers;

[0130] The feature dimension has a value range of [32, 256], with a standard configuration of 128.

[0131] The initial learning rate is set in the range of [1e-4, 1e-2], the standard value is 5e-3, and a cosine annealing strategy is used for dynamic adjustment. The minimum learning rate is no less than 1 / 10 of the initial value.

[0132] The batch size range is [8, 64], and the standard configuration is 32.

[0133] These parameter constraints ensure that the network has sufficient expressive power while avoiding overfitting and training instability.

[0134] Step 2.3: Train the network using historical test data;

[0135] Historical test data training network includes sequences of changes in the microstructure of cosmetics under different environmental conditions and corresponding data on changes in waterproof performance;

[0136] The training data includes:

[0137] Standard environmental dataset: structural evolution sequences under normal temperature and humidity conditions;

[0138] Extreme environment dataset: structural evolution sequences under extreme conditions such as high temperature, high humidity, and violent exercise;

[0139] Hybrid environment dataset: Data simulating environmental condition transitions during daily use;

[0140] Before training the network, environmental parameters (temperature, humidity, water pressure, etc.) are standardized and converted into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate the dimensional differences between different environmental parameters.

[0141] For waterproof performance indicators, a percentage format is used to uniformly represent them, and logical transformation processing is performed to improve the linearity of data distribution;

[0142] For the topological features of microstructure graphs, a spectral normalization method is used to ensure that graph structures of different sizes are comparable.

[0143] Furthermore, according to another embodiment of this application, a synthetic dataset can also be generated using data augmentation techniques to expand the diversity of training samples.

[0144] Step 2.4, optimize network parameters;

[0145] Optimize network parameters to enable it to accurately predict the evolution trend of the microstructure of cosmetics and the corresponding changes in waterproof performance under given environmental conditions;

[0146] The optimization objective function includes:

[0147] Structural prediction loss: the KL divergence between the predicted structure and the actual structure;

[0148] Performance prediction loss: the mean square error between predicted performance and actual performance;

[0149] Regularization term: L2 regularization to prevent overfitting;

[0150] During the optimization process, to ensure the consistency of dimensions and reasonable contribution of different loss terms, the following measures are taken:

[0151] Loss term normalization: Divide the structural prediction loss (KL divergence) and the performance prediction loss (mean squared error) by the average value of their initial training batches to keep the different loss terms consistent in magnitude;

[0152] Adaptive weight adjustment: The weights of different loss terms are dynamically adjusted based on the performance of the validation set to ensure that the model performs well in both structure prediction and performance prediction tasks.

[0153] Gradient clipping: Norm clipping is performed on the gradient to prevent gradient explosion caused by differences in units, thus ensuring the stability of the training process;

[0154] In the process of network parameter optimization, the difference between the current structure prediction and the actual structure is first quantified as the KL divergence value, which reflects the degree of deviation between the predicted distribution and the actual distribution.

[0155] Simultaneously, the mean square error between the predicted waterproof performance and the measured performance is calculated, which reflects the accuracy of the performance prediction; then, the two losses are weighted and summed, and an L2 regularization term is added to prevent overfitting. The magnitude of the regularization term is proportional to the sum of squares of the network parameters.

[0156] Finally, the network parameters are iteratively updated using the gradient descent method, which gradually reduces the total loss function value, thereby improving the network's prediction accuracy.

[0157] In specific application scenarios, taking eyeliner gel products as an example, this spatiotemporal graph convolutional network can predict the decay curve of the product's waterproof performance under a continuous immersion state of 30 minutes by analyzing the microstructure change sequence of the product in a simulated swimming environment, as well as the microstructure change patterns that may lead to waterproof failure, such as the rupture of the emulsion layer or the loss of hydrophobic components.

[0158] Step 3: Constructing a hierarchical reinforcement learning framework;

[0159] A hierarchical reinforcement learning framework is constructed to adaptively generate the optimal testing strategy based on the microstructure analysis results.

[0160] Step 3.1, define the state space;

[0161] State space S RL It includes the current test environment parameters (temperature, humidity, water pressure, etc.) and the current microstructure state of the makeup sample;

[0162] When constructing the state space, continuous environmental parameters are standardized using Z-score to eliminate dimensional differences between different parameters.

[0163] Perform one-hot encoding conversion on discrete environmental parameters (such as test scenario type);

[0164] Principal component analysis is used to reduce the dimensionality of the microstructure state characteristics, retaining principal components that explain more than 95% of the variance, reducing the dimensionality of the state space, and improving learning efficiency.

[0165] Step 3.2, define the action space;

[0166] The motion space A includes adjustable test parameters (temperature changes, humidity changes, water pressure changes, motion simulation, etc.);

[0167] Furthermore, the high-level action space is a discrete space, containing 5 basic test scenario types (standard environment, high temperature environment, high humidity environment, sports environment, and composite environment). Each scenario type has 3 difficulty levels (basic, intermediate, and challenge), resulting in a total of 15 discrete action choices.

[0168] The lower-level action space is a continuous space, and the range of values ​​for each parameter is strictly constrained as follows:

[0169] The temperature parameter range is [15℃, 45℃], and the standard test temperature is 25℃;

[0170] The relative humidity range is [30%, 95%], and the standard test humidity is 60%.

[0171] Exercise intensity is expressed as the increase in heart rate per minute, ranging from [0, 60] beats / minute, with a standard test intensity of 20 beats / minute.

[0172] The water flow rate range is [0, 100] mL / min, and the standard test water flow rate is 30 mL / min.

[0173] All continuous parameters are normalized to the [-1,1] interval when the neural network outputs, and are converted into actual physical quantities through linear mapping.

[0174] Step 3.3, Design the reward function;

[0175] The reward function r takes into account the following factors:

[0176] Information gain: The amount of new information provided by new test conditions;

[0177] Test efficiency: The time and resources required to complete the test;

[0178] Prediction accuracy: the accuracy of predicting changes in microstructure;

[0179] When calculating the reward value, the three indicators with different dimensions, information gain (in bits), testing efficiency (in terms of time and resource consumption) and prediction accuracy (in terms of error rate), are processed by Min-Max normalization and uniformly mapped to the [0,1] interval. Then, the weighted sum is performed according to the optimal weight verified by experiments to ensure that the contribution of different factors in the reward calculation is reasonably balanced.

[0180] Furthermore, the specific formula for calculating the reward function r is as follows:

[0181] r = w1·I norm +w2·E norm +w3·P norm

[0182] Among them, I norm E represents the normalized information gain. norm For the normalized test efficiency, P norm To determine the normalized prediction accuracy, w1, w2, and w3 are the corresponding weighting coefficients. These weighting coefficients must satisfy the following constraints:

[0183] All weights are non-negative real numbers: w i ≥0, i∈{1,2,3};

[0184] The sum of the weights equals 1: w1 + w2 + w3 = 1;

[0185] The information gain weight w1 ranges from [0.35, 0.45], with a standard value of 0.4;

[0186] The test efficiency weight w2 ranges from [0.30, 0.40], with a standard value of 0.35;

[0187] The prediction accuracy weight w3 ranges from [0.20, 0.30], with a standard value of 0.25;

[0188] The system also features a dynamic weight adjustment mechanism that automatically adjusts weight allocation based on the testing phase.

[0189] In the early stages of testing (completion rate <30%), increase the information gain weight to the maximum of 0.45;

[0190] Use the standard weight configuration during the middle of the test (30%-70% completion);

[0191] In the later stages of testing (completion > 70%), the prediction accuracy weight is increased to the maximum of 0.30.

[0192] In addition, when the information gain of three consecutive tests is lower than the threshold of 0.1, the system will temporarily increase the test efficiency weight to 0.4 to avoid getting stuck in an inefficient test loop.

[0193] Step 3.4: Construct a high-level policy network;

[0194] The high-level policy network is responsible for selecting the test scenario type (such as high temperature environment, sports environment, rainy environment, etc.);

[0195] The high-level policy network adopts an Actor-Critic architecture, including:

[0196] State encoder: Encodes the current microstructure state and environmental parameters into a fixed-dimensional vector representation;

[0197] Actor Network: Generates the policy distribution for scene selection, implemented using a multilayer perceptron;

[0198] Critic network: Evaluates the value of the current state to assist in policy updates;

[0199] Alternatively, the high-level policy network can also adopt a memory-enhanced architecture, which stores historical decision-making experience through memory modules to improve the long-term effectiveness of scenario selection.

[0200] Step 3.5: Construct the low-level policy network;

[0201] The low-level policy network is responsible for optimizing specific test parameters within the selected scenario type;

[0202] The low-level policy network is specifically configured for each scenario type, including:

[0203] Parameter generation network: Generates continuous parameter values ​​based on the current state, such as temperature, humidity, water pressure, etc.

[0204] Action evaluation network: predicts information gain and test cost for different parameter combinations;

[0205] Action decoding module: decodes the continuous values ​​output by the network into specific executable test parameters, including mapping the normalized output values ​​back to the actual physical quantity range (e.g., mapping the temperature output value in the range of [0,1] back to the actual temperature range of [20℃,40℃]), and performing quantization processing according to the accuracy requirements of the test equipment;

[0206] Step 3.6: Train the hierarchical reinforcement learning model using the policy gradient method;

[0207] Optimize the objective function:

[0208]

[0209] Furthermore, the optimization of the time dimension in the objective function is specifically reflected in the following ways:

[0210] The time step t is explicitly defined as a discrete time point in the test process, and each time step corresponds to one test state transition and decision;

[0211] The discount factor γ ranges from [0.9, 0.99] and is used to balance short-term and long-term rewards. A larger γ value makes the model focus more on long-term performance.

[0212] Maximum time step T max Different values ​​are set for different product types: 120 time steps for foundation (corresponding to 10 hours of testing), 96 time steps for mascara (corresponding to 8 hours of testing), and 72 time steps for lipstick (corresponding to 6 hours of testing).

[0213] The time step is dynamically adjusted according to the testing phase. The initial exploration phase is 5 minutes per step, the critical change phase is shortened to 2 minutes per step, and the stabilization phase is extended to 10 minutes per step, ensuring that more detailed test data is obtained at key time points.

[0214] Where J(θ, φ) represents the objective function; θ represents the parameters of the microstructure prediction model; φ represents the parameters of the policy network; E represents the expectation function; τ represents the state-action trajectory; π φ T represents a policy function with parameter φ; max Represents the maximum time step; γ is the discount factor; r t G is the reward at time t; t A microstructure diagram representing time t; s t Indicates the state at time t; a t p(G|s) represents the action at time t; t ) represents a given state s t Distribution of the microstructure diagram below; p θ (G t+1 |G t a t ) and p true (G t+1 |G t ,a t ) represent the predicted and actual structural transition probabilities, respectively; D KLdenoted by KL divergence function, used to measure the difference between two probability distributions; λ is the equilibrium parameter; ~ denotes the "follows" sign, indicating that the random variable follows a certain probability distribution; ∑ denotes the summation sign.

[0215] The first objective is the standard reinforcement learning objective, which optimizes the testing strategy by maximizing cumulative rewards:

[0216]

[0217] Where t is the lower bound index of the summation, starting from 0, T max Maximum time step;

[0218] The second term is the microstructure prediction constraint, which ensures the accuracy of microstructure prediction by reducing the KL divergence between the predicted structure transition probability and the actual transition probability.

[0219]

[0220] The specific implementation of the expectation function E is as follows:

[0221] In this system, the expectation function is approximated using the Monte Carlo sampling method. First, multiple trajectories are sampled from the policy distribution, and each trajectory contains a series of state-action-reward sequences.

[0222] Then calculate the cumulative reward or KL divergence value for each trajectory;

[0223] Next, the calculation results of all sampled trajectories are averaged to obtain an estimate of the expected value;

[0224] To reduce variance, the system also employs an importance sampling technique, which improves estimation accuracy by adjusting sampling weights.

[0225] In addition, to handle continuous state spaces, the system uses a density estimation method based on kernel functions, which can accurately calculate the probability distribution in high-dimensional spaces;

[0226] This function dynamically adjusts the sampling strategy during training, gradually increasing sampling efficiency as training progresses, thus ensuring a balance between the accuracy of the desired calculation and computational efficiency.

[0227] like Figure 5 As shown, the weight allocation of the reward function in the hierarchical reinforcement learning framework adopts a tree structure design, which ensures that the testing strategy achieves the best balance between information acquisition and resource utilization.

[0228] The reward function mainly considers three factors: information gain (40%), test efficiency (35%), and prediction accuracy (25%).

[0229] Information gain is further subdivided into microstructure changes (25%) and water resistance fluctuations (15%).

[0230] Test efficiency is broken down into time consumption (20%) and resource usage (15%).

[0231] Forecast accuracy is broken down into short-term forecasts (10%) and long-term forecasts (15%).

[0232] This scientific weight allocation ensures that the reinforcement learning model can prioritize exploring test paths with high information value while taking into account resource utilization efficiency, thereby achieving the automatic generation of the optimal test strategy.

[0233] KL divergence function D KL The implementation of (P||Q) is as follows:

[0234] When processing the probability distribution of a microstructure graph, this function first converts the graph structure into a probability distribution in the eigenvector space.

[0235] Then, a graph isomorphism detection algorithm is used to identify structural equivalence classes, ensuring that similar structures are classified into similar probability distributions;

[0236] Next, for discrete features, the logarithmic ratio of the probability mass function is directly calculated and summed;

[0237] For continuous features, the probability density function is approximated by kernel density estimation, and then the KL divergence is calculated by numerical integration.

[0238] To address the sparsity problem of high-dimensional data, the system introduces a regularization term to prevent numerical instability when the probability approaches zero.

[0239] In addition, the function implements an adaptive partitioning strategy, which dynamically adjusts the calculation precision according to the data distribution characteristics, thereby improving calculation efficiency while ensuring accuracy.

[0240] This implementation method can effectively measure the performance of microstructure prediction models and guide model optimization in the direction of improving prediction accuracy.

[0241] When optimizing the objective function, the reward value obtained at each time point in the test trajectory is first calculated, and the future reward is decayed according to the discount factor. The smaller the discount factor, the more attention is paid to the recent reward.

[0242] Then, the decaying rewards at each time point are summed to obtain the total reward value of the trajectory;

[0243] Next, a weighted average of all possible trajectories is calculated based on their probability of occurrence to obtain the expected return value;

[0244] Simultaneously, the KL divergence constraint term for microstructure prediction is calculated, which ensures the accuracy of structure prediction;

[0245] Finally, the two objectives are weighted and combined according to the balancing parameters to form the final optimization objective function. The policy network parameters are then updated using the policy gradient method, which gradually increases the objective function value.

[0246] In specific application scenarios, taking lipstick products as an example, the layered reinforcement learning framework can first identify that the product's waterproof performance is most vulnerable in high-temperature environments. Then, it automatically configures a series of refined test parameters that gradually increase from low temperature to high temperature, finds the critical temperature point of the product's waterproof performance within the smallest test steps, and records the microstructural changes near the critical point, providing precise guidance for product formula optimization.

[0247] Step 4: Construction of the structurally sensitive prediction module;

[0248] A structure-sensitive prediction module was constructed to realize the mapping prediction from microstructural changes to macroscopic waterproof performance.

[0249] Step 4.1, define the key structural feature extraction function;

[0250] Extract features directly related to waterproof performance from the microstructure diagram, such as molecular arrangement density, cross-linking strength, and surface tension;

[0251] The specific implementation of the key structural feature extraction function is as follows:

[0252] First, a multi-scale analysis of the microstructure diagram is performed to calculate the topological features and physical properties of the nodes at different scales.

[0253] Then, a feature selection algorithm is used to filter out a subset of features that are highly correlated with waterproof performance;

[0254] Next, these features are normalized and dimensionality reduced to decrease redundant information;

[0255] Finally, the processed features are organized into structured vectors, which serve as input for subsequent performance prediction.

[0256] This function pays special attention to key indicators such as the regularity of molecular arrangement, the stability of interface structure, and the integrity of cross-linked networks. These indicators are extracted from the microstructure diagram through graph theory algorithms and physical models, which can effectively characterize the microstructure characteristics of color cosmetic products.

[0257] Different preprocessing methods are used for different types of features:

[0258] Topological features (such as degree centrality, clustering coefficient, etc.) are logarithmically transformed and then standardized to handle their long-tail distribution characteristics;

[0259] Physical properties (such as bond strength and intermolecular distance) are Robust normalized to reduce the impact of outliers;

[0260] Classification features (such as molecular configuration type) are converted into numerical values ​​using tag encoding.

[0261] Furthermore, the value range constraints for key structural features are as follows:

[0262] Molecular density is defined as the number of molecules per unit volume, ranging from [1e3, 1e6] molecules per μm. 3 ;

[0263] Crosslinking strength is expressed as crosslinking point density, with a value ranging from [0.1, 10] points / nm. 2 ;

[0264] The surface tension ranges from [20, 72] mN / m;

[0265] The degree centrality value ranges from [0, N-1], where N is the total number of network nodes;

[0266] The clustering coefficient ranges from [0,1].

[0267] The bonding strength ranges from [1,500] kJ / mol;

[0268] The intermolecular distance ranges from [0.1, 10] nm.

[0269] All features are standardized before being input into the prediction model and mapped to the [-1,1] interval to ensure that features of different dimensions have the same weight in the model.

[0270] Step 4.2, construct the structure-performance mapping model;

[0271] The extracted structural features are mapped to waterproof performance indicators (such as waterproof durability, sweat resistance, and abrasion resistance).

[0272] This model employs a multi-task learning framework, including:

[0273] Shared feature extractor: Extracts general feature representations from microstructure maps;

[0274] Task-specific layer: A dedicated predictive layer for different waterproofing performance indicators;

[0275] Cross-task attention mechanism: capturing the interrelationships between different waterproof performance indicators;

[0276] Output Decoding and Transformation Module: Decodes the model's raw output (usually normalized numerical values) and transforms it into concrete and understandable waterproof performance indicators, including:

[0277] Classification output decoding: The probability distribution output of the classification task (such as waterproof rating prediction) is converted into specific rating labels through the argmax operation;

[0278] Regression output transformation: The continuous value output of the regression task (such as waterproof durability prediction) is mapped back to the original units (such as hours and minutes) through denormalization and quantified according to industry standards (such as rounding to the nearest 5-minute interval).

[0279] Multi-dimensional index integration: Multiple individual indicators (such as sweat resistance, rain resistance, etc.) are converted into a comprehensive waterproof performance score through weighted averaging or more complex aggregation functions, which facilitates horizontal comparison between products;

[0280] like Figure 4 As shown, the correlation between the predicted values ​​and actual measured values ​​of the microstructure prediction model is visualized using a scatter plot, verifying the effectiveness of the structure-sensitive prediction module of this patent. The figure shows a high correlation between the predicted and actual values, with the point set primarily distributed near the diagonal, indicating high prediction accuracy. This high correlation demonstrates that this method can accurately establish the mapping relationship between microstructural features and macroscopic waterproof performance, providing a reliable basis for the scientific evaluation of product waterproof performance.

[0281] Meanwhile, the structure-sensitive prediction module can perform multi-dimensional waterproof performance evaluation of foundation products. Through a structure-performance mapping model, the system simultaneously predicts the waterproof performance of two foundation products in five dimensions: sweat resistance, rain resistance, swimming resistance, friction resistance, and durability. As shown in the figure, foundation A performs better in sweat resistance and durability, while foundation B has advantages in rain resistance and swimming resistance. This multi-dimensional performance evaluation provides consumers with scientific guidance for choosing the right product for different usage scenarios.

[0282] In specific application scenarios, taking foundation products as an example, the structure-sensitive prediction module can simultaneously predict the waterproof durability of the product under sweat, rain and swimming scenarios, as well as the makeup removal mode after waterproof failure (such as patchy peeling or overall fading), to label the product with accurate waterproof performance labels and guide consumers on the best usage methods in different usage scenarios.

[0283] Step 5: Test the implementation of the adaptive state transition mechanism;

[0284] Implement an adaptive test state transition mechanism to dynamically adjust test conditions based on real-time feedback and simulate real-world usage scenarios.

[0285] Step 5.1, Configure the status monitoring module;

[0286] Real-time acquisition of microscopic images and performance data of makeup samples, and updating of microstructure diagrams;

[0287] The status monitoring module uses high-precision optical sensors and miniature environmental monitors to continuously monitor the surface morphology, color changes, and environmental parameters of the sample.

[0288] Step 5.2, construct the state evaluation function;

[0289] The state evaluation function assesses the information value and potential weaknesses of a sample under the current test state;

[0290] The specific implementation of the state evaluation function is as follows:

[0291] First, calculate the similarity between the current microstructure state and samples in the historical test database, and identify the closest historical case.

[0292] Then analyze the gap between the current state and the expected performance target, and quantify it as a performance gap indicator;

[0293] Next, the stability of the current state is assessed by calculating the degree of fluctuation of key microstructural parameters.

[0294] Simultaneously, calculate the expected gain of acquiring new information under the current test conditions, and estimate the value of the new test point based on the principle of information entropy;

[0295] Finally, by combining the above indicators, a multi-dimensional scoring vector is generated, reflecting the overall value and potential problems of the current state. This function can dynamically identify key nodes and weak points in the testing process, providing a basis for decision-making in adjusting subsequent testing strategies. When calculating the multi-dimensional scoring vector, the scoring indicators of different dimensions (such as similarity, performance gap, stability indicators, and information gain) are standardized to ensure that each indicator has a reasonable weight distribution in the comprehensive score.

[0296] Furthermore, the calculation of each index in the state evaluation function adopts the following constraints:

[0297] The similarity index ranges from [0, 1], where 0 represents complete dissimilarity and 1 represents complete similarity. When the similarity exceeds 0.85, the system will prioritize referencing the test paths of historical cases.

[0298] The performance gap index ranges from [0, ∞), and after standardization, it is mapped to the interval [0, 1]. When the standardized performance gap index exceeds 0.7, the system will mark the state as "high-risk state".

[0299] The stability index ranges from [0, 1], where 0 indicates extreme instability and 1 indicates complete stability. When the stability is below 0.4, the system will automatically increase the sampling frequency.

[0300] The information gain index ranges from [0, ∞), in bits, and is normalized to the interval [0, 1]. When the normalized information gain is less than 0.2, the system will consider switching the test scenario.

[0301] The comprehensive calculation of the multidimensional scoring vector adopts the weighted geometric mean method to ensure that the extremely low value of a single indicator can significantly affect the final score.

[0302] Step 5.3: Utilize a hierarchical reinforcement learning model;

[0303] Generate the optimal test action for the next step based on the status assessment results;

[0304] Optionally, in emergency situations, a manual intervention mechanism can be introduced, allowing testers to adjust test parameters based on experience;

[0305] The system automatically determines whether manual intervention is needed based on the confidence threshold. When the prediction confidence is lower than the preset threshold, an intervention request is automatically triggered.

[0306] Furthermore, the confidence threshold in the system is explicitly defined as 0.75 (75%), meaning that when the model's prediction confidence for the next optimal test action is lower than 0.75, the system will automatically trigger a manual intervention request.

[0307] Confidence is calculated based on the probability distribution entropy value output by the policy network. The higher the entropy value, the greater the uncertainty and the lower the confidence.

[0308] The confidence threshold can be adjusted appropriately for different types of test scenarios:

[0309] For standard testing scenarios, the confidence threshold is maintained at 0.75;

[0310] For high-risk testing scenarios (such as extreme temperature and humidity conditions), the confidence threshold is increased to 0.85;

[0311] For exploratory testing scenarios, the confidence threshold can be reduced to 0.65 to encourage the system to explore more autonomously.

[0312] In addition, the system has set up a confidence monitoring mechanism for continuous decision-making. When the average confidence of three consecutive decisions is lower than 0.7, even if the confidence of a single decision is higher than the threshold, a manual intervention request will be triggered to prevent the system from getting stuck in a local suboptimal decision-making loop.

[0313] Step 5.4: Automatic adjustment of test conditions is achieved through the execution device;

[0314] Automatically adjusts to changes such as temperature, humidity, water pressure, or simulates motion.

[0315] The actuators include a precision temperature control system (accuracy ±0.1℃), a humidity regulator (accuracy ±1%RH), a programmable water pressure simulator, and a multi-degree-of-freedom motion simulation platform, enabling precise control of the test environment;

[0316] Step 5.5: Construct a test trajectory recording system;

[0317] Record the complete testing process to provide data support for subsequent analysis;

[0318] The recording system uses a time-series database to store all test parameters and corresponding sample states, supporting efficient retrieval and visual analysis, and facilitating the tracing and reproduction of test results.

[0319] Step 6: Achieving predictive learning ability;

[0320] To enable the system to learn proactively and predict long-term degradation of waterproofing performance through minute initial structural changes.

[0321] Step 6.1: Construct a multi-scale time decomposition network;

[0322] Separate short-term fluctuations from long-term trends;

[0323] This network employs an architecture that combines wavelet transform and recurrent neural networks, enabling it to decompose signals at different time scales and identify temporary fluctuations and persistent change patterns.

[0324] Step 6.2, configure the sensitivity analysis module;

[0325] Identify the microstructural change patterns that have the greatest impact on long-term performance;

[0326] This module quantifies the contribution of different microstructural features to waterproof performance through gradient calculation and feature importance ranking, and generates a sensitivity heatmap to intuitively display key influencing factors.

[0327] Step 6.3: Construct a long-term prediction model;

[0328] Extrapolating long-term performance based on short-term test data:

[0329] S(t long )=S(t short )·f decay (t long -t short G short ,ΔG)

[0330] Where S(t) long ) and S(t short ) represent the long-term and short-term waterproof performance scores, respectively; t long t represents a long-term time interval. shortIndicates a short-term test moment; G short This represents a microstructure diagram in a short-term test; ΔG represents the rate of change of the microstructure, reflecting the degree of change of the microstructure over time; f decay This is the performance degradation function;

[0331] When calculating long-term forecasting models, the time difference (t) long -t short Perform logarithmic transformation and normalization to better capture nonlinear effects on the time scale;

[0332] The different components (such as topological change rate and physical property change rate) in the microstructure change rate ΔG are weighted and standardized, and weights are assigned according to their importance to waterproof performance.

[0333] The waterproof performance score S(t) is subjected to a logit transformation to improve its numerical stability in the extreme value region close to 0 or 1.

[0334] Performance degradation function f decay The specific implementation is as follows:

[0335] First, a recurrent neural network is used to model the short-term changing trends of the microstructure;

[0336] Then, an acceleration factor is calculated based on the trend of change, which reflects the rate of structural change.

[0337] Next, a nonlinear decay model is constructed based on the acceleration factor and time difference;

[0338] Finally, the decay model is applied to the short-term performance score to predict long-term performance.

[0339] Specifically, f decay The function can be represented as:

[0340] f decay (t long -t short G short ,ΔG)=exp(-α·R eval (G short ,ΔG)·(t long -t short ))

[0341] Where t long t represents a long-term time interval. short Indicates a short-term test moment; G short This represents a microstructure diagram in a short-term test; ΔG represents the rate of change of the microstructure, indicating the degree of change of the microstructure over time; f decayα is the performance degradation function; α is the time scale parameter; exp represents the exponential function, which is an exponential operation with the natural constant e as the base, i.e., exp(x) = e x ;R eval This is a structural change evaluation function used to assess the impact of microstructural changes on performance;

[0342] Furthermore, the performance degradation function f decay The value of is strictly constrained within the interval (0,1], where 1 represents no performance degradation and close to 0 represents almost complete performance degradation. The time scale parameter α ranges from [0.001,0.1], adjusted according to different types of makeup products: 0.005±0.001 for mascara, 0.008±0.002 for foundation, and 0.003±0.001 for lipstick. The time difference (t) long -t short The timeframe is in hours and must be positive. When the prediction time exceeds twice the product's design lifespan, the reliability of the prediction results will decrease significantly. In this case, the system will automatically mark the prediction result as a "reference value" instead of a "certain value".

[0343] Structural change evaluation function R eval A more detailed implementation is as follows:

[0344] This function employs an architecture combining graph neural networks and temporal analysis. First, it processes the microstructure graph G through a graph convolutional network. short Extract spatial features;

[0345] Simultaneously, the rate of change ΔG is processed through a gated loop unit to capture time dynamic characteristics;

[0346] Then the two features are fused, and key structural changes are highlighted through an attention mechanism;

[0347] Next, a multilayer perceptron is used to perform a nonlinear transformation on the fused features to learn the complex relationship between structural changes and performance degradation;

[0348] Finally, residual connections and layer normalization techniques are used to enhance the model's stability and expressive power, outputting a scalar value as the decay rate. This function pays special attention to vulnerable points and change-sensitive regions in the microstructure, and can identify key modes that can lead to long-term performance degradation even with minor structural changes, providing a reliable basis for long-term prediction of waterproof performance.

[0349] Furthermore, the structural change evaluation function R evalThe output value is strictly constrained within the range of [0.1, 10], where 0.1 indicates that the structural change has a negligible impact on performance, and 10 indicates that the structural change has a significant impact on performance. To ensure the stability of the evaluation results, a threshold constraint mechanism is introduced:

[0350] When the topological change rate of the microstructure exceeds 30%, the system automatically adjusts R. eval The value is set to the maximum of 10;

[0351] When the rate of change of key nodes (such as cross-linking points and interface structures) exceeds 15%, R eval The value is not lower than 5.0;

[0352] When the microstructure remains basically stable (the rate of change is less than 3%), R eval The value should not exceed 0.5.

[0353] In addition, the system also sets time window constraints for calculating the rate of change, with a standard window of 30 minutes, a minimum of 10 minutes, and a maximum of 120 minutes, to ensure that meaningful structural change trends are captured.

[0354] In the long-term prediction calculation process, the waterproof performance score and corresponding microstructure information from the short-term test are first obtained;

[0355] Then the rate of change of the microstructure is calculated, which reflects how fast the structure changes over time;

[0356] Next, the current microstructure state and rate of change are input into the structural change evaluation function to obtain a decay rate value;

[0357] The decay rate is then multiplied by the time difference and the time scale parameter, and a decay coefficient is calculated using an exponential function.

[0358] Finally, the short-term waterproof performance score is multiplied by the attenuation coefficient to obtain the predicted long-term waterproof performance score.

[0359] In this way, the system can reasonably predict the waterproof performance of a product in long-term use based on short-term test data.

[0360] In specific application scenarios, such as Figure 2 As shown, taking waterproof mascara as an example, the predictive learning module can accurately predict the change curve of the waterproof performance of a product during 8 hours of continuous wear using 2 hours of accelerated test data. This helps manufacturers to accurately label the waterproof duration on product packaging and avoid consumer complaints caused by exaggerated advertising.

[0361] like Figure 3As shown, the microscopic guided self-evolving test network of this patent significantly outperforms traditional testing methods in multiple dimensions. The bar chart comparison illustrates the performance differences between the two methods in four key indicators: test cycle, test accuracy, test coverage, and resource utilization. Using this patented method, the test cycle is shortened by 90%, test accuracy is improved by 35%, test coverage reaches 95%, and resource consumption is only 33% of that of traditional methods. These quantitative data intuitively demonstrate the comprehensive advantages of this patented method compared to traditional static testing methods, not only significantly improving test efficiency but also achieving more accurate waterproof performance evaluation.

[0362] It is understood that data preprocessing methods known to those skilled in the art include data cleaning, data transformation, and data reduction. Data transformation includes type conversion and normalization and standardization. Although the dimensions and types of data were omitted in the description of the preceding embodiments, data preprocessing is a technical knowledge known to those skilled in the art and a prerequisite step in data processing. Therefore, the previously described well-known data preprocessing steps were not described independently.

[0363] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for evaluating the waterproof performance of makeup based on deep learning, characterized in that, Includes the following steps: High-resolution microscopic images of cosmetic products are acquired, the images are processed using an image segmentation algorithm to construct a microstructure graph, and attributes are assigned to the nodes and edges of the microstructure graph. A spatiotemporal graph convolutional network model is constructed and trained to realize the modeling and analysis of the dynamic changes of the microstructure graph; Construct a hierarchical reinforcement learning framework, including defining the state space, action space and reward function, building high-level and low-level policy networks, and generating intelligent testing policies; A structure-sensitive prediction module was constructed to realize the mapping prediction from microstructural changes to macroscopic waterproof performance; Implement an adaptive test state transition mechanism to dynamically adjust test conditions based on real-time feedback and simulate real-world usage scenarios; It enables predictive learning capabilities, allowing the prediction of long-term waterproofing performance degradation through minute initial structural changes; The hierarchical reinforcement learning framework optimizes the testing strategy by optimizing the objective function, which comprehensively considers the reward value obtained at each time step in the test trajectory and the accuracy constraint of microstructure prediction.

2. The method according to claim 1, characterized in that, The steps of processing the image using an image segmentation algorithm to construct a microstructure map include: The resolution of the image segmentation algorithm is set to 5-50 nanometers. The segmentation threshold is selected according to the type of makeup product. The segmentation threshold for mascara is 0.15-0.25, the segmentation threshold for foundation is 0.20-0.35, and the segmentation threshold for lipstick is 0.10-0.

20. The segmented image is converted into a graph data structure, where nodes represent molecular clusters or functional units in cosmetic products, and edges represent physical or chemical connections between nodes. The node and edge attributes of the microstructure graph are preprocessed, including normalizing numerical features, encoding categorical features, smoothing and denoising time-series data, and imputing missing values.

3. The method according to claim 1, characterized in that, The steps for building and training a spatiotemporal graph convolutional network model include: Construct a multi-layer spatiotemporal graph convolutional network model, including spatial graph convolutional layers, temporal convolutional layers, and an attention mechanism; Design node update functions, including aggregation functions and non-linear activation functions; Set key network parameters, including kernel size, network depth, feature dimension, learning rate, and batch size; The network was trained using historical test data that included standard environment datasets, extreme environment datasets, and mixed environment datasets. Optimize network parameters, including structure prediction loss, performance prediction loss, and regularization term, by employing adaptive weight adjustment and gradient pruning techniques.

4. The method according to claim 1, characterized in that, The steps to build a hierarchical reinforcement learning framework include: Define the state space, including the current state of the microstructure graph, historical test trajectories, and environmental parameters; Define the action space, including test parameter adjustment actions and test termination judgment actions; Define a reward function that comprehensively considers information gain, testing efficiency, and prediction accuracy. Calculate the overall reward value through a weighted combination method, where each weight coefficient is non-negative and sums to 1. The weights for information gain range from 0.3 to 0.5, testing efficiency from 0.2 to 0.4, and prediction accuracy from 0.2 to 0.

4. Build a high-level strategy network to determine the testing phase and overall strategy; Construct a low-level policy network responsible for generating specific test parameters; The model is trained using the policy gradient method to optimize the objective function, where time step t is defined as the discrete test time, the discount factor γ ranges from 0.9 to 0.99, and the test termination time step T is set to a maximum value according to different product types.

5. The method according to claim 1, characterized in that, The steps to construct a structure-sensitive prediction module include: Define a key structural feature extraction function to extract features directly related to waterproof performance from the microstructure diagram, including molecular arrangement density, cross-linking strength, surface tension, degree centrality, clustering coefficient, bonding strength and intermolecular distance, and standardize all features to map them to the [-1,1] interval; A structure-performance mapping model is constructed, including a shared feature extractor, a task-specific layer, a cross-task attention mechanism, and an output decoding and transformation module, which maps the extracted structural features to waterproof performance indicators, including waterproof durability, sweat resistance, and abrasion resistance. The effectiveness of the structure-sensitive prediction module is verified by visualizing the correlation between the predicted values ​​and actual measured values ​​of the microstructure prediction model using scatter plots. Radar charts are used to display the performance of different products across multiple waterproofing dimensions, enabling multi-dimensional waterproofing performance evaluation.

6. The method according to claim 1, characterized in that, The steps to implement the test state adaptive transition mechanism include: Configure a status monitoring module to collect microscopic images and performance data of makeup samples in real time and update the microstructure diagram; Construct a state evaluation function to evaluate the information value and potential weaknesses of samples under the current test state. This includes calculating the similarity between the current microstructure state and samples in the historical test database, analyzing the gap between the current state and the expected performance target, evaluating the stability of the current state, and calculating the expected gain of acquiring new information under the current test conditions. Using a hierarchical reinforcement learning model, the system generates the next optimal test action based on the state evaluation results. The system automatically determines whether manual intervention is needed based on the confidence threshold. When the prediction confidence is below 0.75, the system automatically triggers an intervention request. The test conditions are automatically adjusted by the actuator, including changing the temperature, humidity, water pressure or simulating motion. Build a test trajectory recording system to record the complete test process and provide data support for subsequent analysis.

7. The method according to claim 1, characterized in that, The steps to achieve predictive learning capabilities include: Construct a multi-scale time decomposition network to separate short-term fluctuations from long-term trends; Configure a sensitivity analysis module to identify microstructural change patterns that have the greatest impact on long-term performance; Construct a long-term prediction model, extrapolating long-term performance based on short-term test data, using the formula: S(t long )=S(t short )·f decay (t long -t short ,G short ,△G) Calculate the long-term waterproof performance score, where S(t) long S(t) represents the long-term waterproof performance score. short ) represents the short-term waterproof performance score, t long t represents a long-term time interval. short Indicates a short-term test moment, G short This represents a microstructure diagram in a short-term test, where ΔG represents the rate of change of the microstructure, and f decay This is the performance degradation function; Wherein, the performance degradation function f decay The value range is strictly constrained within the interval (0,1), and the value range of the time scale parameter α is [0.001,0.1], which is adjusted according to different types of makeup products.

8. The method according to claim 7, characterized in that, Performance degradation function f decay The specific implementation is as follows: f decay (t long -t short ,G short ,△G) =exp(-α·R eval (G short ,△G)·(t long -t short )) Where α is the time scale parameter, exp represents the exponential function, and R0 is the time scale parameter. eval This is a structural change evaluation function used to assess the impact of microstructural changes on performance. Structural change evaluation function R eval The output value is strictly constrained within the range of [0.1, 10], where 0.1 indicates that the structural change has a very small impact on the performance, and 10 indicates that the structural change has a very large impact on the performance. When the topological change rate of the microstructure exceeds 30%, the system automatically adjusts R. eval The value is set to the maximum of 10; When the rate of change of the critical node exceeds 15%, R eval The value is not lower than 5.0; When the microstructure remains stable, R eval The value should not exceed 0.

5.

9. The method according to claim 1, characterized in that, The technical effects of the method include: Testing efficiency is improved by using spatiotemporal graph convolutional network technology to analyze microstructural changes in real time, resulting in a significant reduction in the testing cycle. The testing accuracy has been greatly improved by using microstructure analysis technology to guide the optimization of testing strategies. Achieving dynamic scenario simulation, the test state adaptive transition mechanism successfully simulates complex dynamic changes in the real-world usage environment; It has predictive capabilities; through predictive learning technology, the system can predict the waterproof performance of a product under long-term or extreme conditions based on short-term test data. The intelligent testing path, through hierarchical reinforcement learning technology, enables the system to automatically explore and identify critical points and weaknesses in product performance; Resource utilization is optimized by using microstructure diagram construction technology to guide test path selection and reduce the exploration of invalid test paths.

10. A deep learning-based system for evaluating the waterproof performance of makeup, used to perform the method described in any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire high-resolution microscopic images of makeup products; The image processing module is used to process images using image segmentation algorithms and construct microstructure maps. The attribute assignment module is used to assign attributes to nodes and edges of the microstructure graph; The spatiotemporal graph convolutional network module is used to model and analyze the dynamic changes of microstructures. A hierarchical reinforcement learning module is used to generate intelligent testing strategies; The structure-sensitive prediction module is used to realize the mapping prediction from microstructural changes to macroscopic waterproof performance; The test state adaptive switching module is used to dynamically adjust test conditions based on real-time feedback; A predictive learning module is used to predict long-term degradation of waterproofing performance through minute initial structural changes; Test execution device, used to automatically adjust test conditions; The test trajectory recording system is used to record the entire test process and provide data support for subsequent analysis.

Images