Coating weather resistance prediction system based on multi-task supervised neural network

Abstract

The invention discloses a coating weather resistance prediction system based on a multi-task supervised neural network, and the system comprises a data obtaining module which is used for obtaining the mass ratio of a three-primary-color pigment of a target coating to an acrylic acid substrate and the coating thickness; the first neural network module is used for predicting the sunlight reflectivity of the target coating; the second neural network module is used for predicting the weather resistance of the coating; the network training module is used for training the first neural network module and the second neural network module; and the network training module is also used for completing initialization of a plurality of shared hidden layers and intermediate parameter output layers through a transfer learning mode, and completing training of the second neural network through a supervised learning mode. According to the paint weather resistance prediction system based on the multi-task supervised neural network, the limitations of single prediction target and simple model structure are solved, and an intelligent solution is provided for comprehensive evaluation and optimization of paint performance.

Description

Technical Field

[0001] This invention relates to the field of green building materials technology, and in particular to a coating weather resistance prediction system based on a multi-task supervised neural network. Background Technology

[0002] Architectural exterior wall coatings play a vital role in modern architecture. On one hand, they significantly enhance the aesthetics of buildings by offering a variety of colors, glosses, and textures. On the other hand, they effectively protect building exteriors from external factors such as ultraviolet radiation, rain, pollutants, and temperature and humidity fluctuations, making them an important component of green building. However, architectural exterior wall coatings are constantly exposed to complex environments, and their weather resistance is affected by a combination of factors, including ultraviolet radiation, temperature and humidity fluctuations, and air pollution. This directly determines the product's lifespan and performance.

[0003] With the rapid development of the coatings industry, coating weather resistance prediction systems have gradually become indispensable tools in the research and development process. Currently, most coating weather resistance prediction systems on the market are based on traditional empirical models or simple statistical methods. These systems often require a large amount of experimental data and have low prediction accuracy and efficiency. Especially when dealing with new coating formulations, the predictive ability of existing systems decreases significantly due to the lack of historical data support. Furthermore, existing prediction systems typically employ a single hardware architecture with limited computing power, making it difficult to support the real-time operation of complex deep learning models. This severely restricts the practical application effectiveness of these prediction systems.

[0004] The paper "Siru Q, Yuequan D, Xiaoxia L, et al. Prediction and influence of themesass proportion of trichromatic colourants and acrylic substrate on the optical and thermal performance of external wall coatings: An artificial neural network approach[J]. Solar Energy Materials and Solar Cells, 2022, 236" proposes a coating performance prediction method based on artificial neural networks (ANN). This study focuses on analyzing the influence of the proportions of trichromatic colourants and acrylic substrate on solar reflectivity, verifying the application potential of ANN in optical performance prediction. However, the system proposed in this paper still has the following technical shortcomings: First, the system structure is relatively simple, containing only a single neural network model, which cannot effectively handle multi-objective prediction tasks; second, the system lacks complete data processing and model training modules, making it difficult to guarantee the stability and reliability of the prediction results; finally, the system's research scope is limited to the single index of solar reflectivity, failing to predict the comprehensive index of coating weather resistance.

[0005] In practical applications, paint manufacturers and research institutions urgently need an integrated and intelligent paint weather resistance prediction system. Such a system must not only accurately predict various performance indicators of paints, but also possess comprehensive data management, model training, and result analysis capabilities. Simultaneously, the system should be adaptable to different hardware environments and have good scalability and maintainability. This requires us to fully consider the optimization of hardware and software architecture during system design, constructing a fully functional and reliable prediction platform. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing coating weather resistance prediction systems, such as their simple structure, limited functionality, and low prediction accuracy, by providing a coating weather resistance prediction system based on a multi-task supervised neural network. This system not only overcomes the limitations of existing research, such as single prediction targets and simple model structures, but also provides an intelligent solution for the comprehensive evaluation and optimization of coating performance, thereby effectively improving the accuracy and efficiency of coating weather resistance prediction.

[0007] To achieve the above-mentioned objectives, the technical solution provided by this invention includes: A coating weather resistance prediction system based on a multi-task supervised neural network includes: The data acquisition module is used to obtain the mass ratio of the three primary color pigments to the acrylic substrate and the coating thickness of the target coating. The first neural network module is used to receive the quality ratio output by the data acquisition module and predict the solar reflectance of the target coating. The second neural network module is used to receive the solar reflectance output by the first neural network module and the coating thickness output by the data acquisition module, and predict the weather resistance of the coating. The output of the second neural network module includes the ultraviolet absorption and surface temperature change of the target coating as intermediate parameters, and the comprehensive weather resistance index as the weather resistance prediction result of the target coating. The second neural network module includes a shared hidden layer, an intermediate parameter output layer, an independent hidden layer, and a prediction output layer connected in sequence. The network training module is connected to the first neural network module and the second neural network module respectively, and is used to train the first neural network module and the second neural network module; the network training module is also used to initialize several shared hidden layers and intermediate parameter output layers through transfer learning, and to train the second neural network through supervised learning.

[0008] Preferably, the network training module includes: The parameter migration unit is used to migrate the network parameters of the first neural network module to the shared hidden layer of the second neural network module by freezing the weights. The weight initialization unit is used to perform adaptive initialization of the weights of the intermediate parameter output layer of the second neural network module.

[0009] Preferably, the network training module further includes: The parameter freezing unit is used to freeze the parameters of the shared hidden layer and intermediate parameter output layer of the second neural network module during training. A pre-training unit is used to pre-train the second neural network module using a publicly available material property dataset.

[0010] Preferably, the network training module further includes: The fine-tuning unit is used to acquire the quality ratio data and coating thickness data used to train the first neural network module as a first training set, and to fine-tune the second neural network module based on the first training set.

[0011] Preferably, the fine-tuning unit includes: The test data acquisition subunit is used to conduct standard weathering tests based on the training data of the first neural network module and obtain the corresponding weathering evaluation index. The training set construction subunit is used to combine the quality ratio data, coating thickness data and corresponding weather resistance evaluation indicators to construct the first training set. A supervised training subunit is used to perform supervised training on the second neural network module based on the first training set.

[0012] Preferably, the network training module further includes a data augmentation unit, used to augment the first training set by random perturbation.

[0013] Preferably, the data expansion unit includes: The perturbation generation subunit is used to randomly perturb the mass ratio data and coating thickness data within a preset range. The verification subunit is used to verify whether the disturbed data meets the preset physical and technological constraints.

[0014] The present invention also provides a coating weather resistance prediction device, comprising: a processor; a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, is used to load the coating weather resistance prediction system based on a multi-task supervised neural network as described in any one of claims 1 to 7. Beneficial effects

[0015] 1. A systematic intelligent prediction solution is provided: This invention achieves automated prediction of coating weather resistance by constructing a complete system including a data acquisition module, a neural network module, and a training module. The modular design of the system not only makes the prediction process more standardized and efficient, but also significantly reduces reliance on traditional laboratory testing, effectively reducing R&D cycle and cost investment.

[0016] 2. Multi-level performance evaluation is achieved: The system architecture design of this invention innovatively introduces a dual neural network module, which can not only predict the basic parameter of solar reflectivity, but also further output intermediate parameters such as ultraviolet absorption and surface temperature changes, ultimately providing a comprehensive weather resistance index. This multi-level evaluation mechanism makes coating performance prediction more comprehensive and reliable, overcoming the limitations of existing systems with a single objective.

[0017] 3. Improved processing capability for small sample data: This invention incorporates a dedicated network training module within the system, utilizing a transfer learning mechanism to transfer knowledge and experience from the first neural network module to the second neural network module. This design fully leverages existing data resources, significantly enhancing the system's predictive ability under small dataset conditions.

[0018] 4. Optimized prediction performance and applicability: The system of this invention is based on a multi-task supervised neural network architecture, employing a combination of shared and independent hidden layers. This allows the system to handle multiple prediction tasks simultaneously, and the training set is expanded through a data augmentation unit. These designs not only improve the system's prediction accuracy but also enhance its adaptability to different coating formulations and environmental conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the coating weather resistance prediction system based on a multi-task supervised neural network provided in a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the second neural network module structure provided in another preferred embodiment of the present invention; Figure 3 This is a schematic diagram of the network training module structure and its relationship with the second neural network module provided in another preferred embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings. In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.

[0021] like Figure 1 As shown, this embodiment provides a coating weather resistance prediction system based on a multi-task supervised neural network, specifically including: The data acquisition module is used to obtain the mass ratio of the three primary color pigments to the acrylic substrate and the coating thickness of the target coating.

[0022] The data acquisition module is responsible for obtaining the mass ratio of the three primary color pigments to the acrylic substrate and the coating thickness data of the target coating. The three primary color pigments refer to the three basic pigments used to mix colors using the subtractive color mixing principle, specifically including cyan, magenta, and yellow. The acrylic substrate refers to the resin matrix formed by polymers of acrylic acid or its derivatives, serving as a carrier for pigments, additives, and solvents in the coating. These two key parameters constitute the core components of the coating formulation and play a decisive role in predicting the optical, thermal, and weather resistance properties of the coating.

[0023] In their earlier paper, "Siru Q, Yuequan D, Xiaoxia L, et al. Prediction and influence of the mass proportion of trichromatic colourants and acrylic substrate on the optical and thermal performance of external wall coatings: An artificial neural network approach [J]. Solar Energy Materials and Solar Cells, 2022, 236," the inventors conducted an in-depth study on the influence of the proportion of trichromatic colourants on the solar reflectivity of coatings, and performed detailed analysis through experiments and an artificial neural network (ANN) model. The results show that the proportion of trichromatic colourants is a key factor determining the solar reflectivity of coatings. By absorbing and reflecting light of different wavelengths, adjusting the proportion of trichromatic colourants significantly affects the optical properties of the coatings, thereby optimizing their solar reflectivity.

[0024] Furthermore, the acrylic base plays a crucial regulatory role in coating formulations. Its transparency, refractive index, and pigment dispersion not only enhance the optical performance of the three primary color pigments but also further improve the coating's ability to reflect sunlight. The synergistic effect of the acrylic base and the three primary color pigments has been proven to be significant in improving coating performance.

[0025] By optimizing the ratio of pigments to substrate, this study successfully explored an effective method to improve coating formulations, thereby significantly enhancing solar reflectivity. This research provides a solid theoretical basis for the optical performance optimization design of exterior wall coatings and lays a scientific foundation for further research on weather resistance prediction and coating performance improvement. The introduction of artificial neural networks in this study not only improved the efficiency of formulation optimization but also provided important insights for the development of intelligent coating performance prediction technologies.

[0026] Based on this, the present invention considers the close relationship between coating thickness and the optical properties, thermal properties, and weather resistance of the coating. Furthermore, it can form a complete coating formulation description with the ratio of the three primary color pigments and the acrylic substrate, providing multi-dimensional data support and possessing significant technical value and practical significance. The selection of coating thickness is explained in detail below from the three aspects directly affecting optical properties, thermal properties, and weather resistance: 1. The effect of coating thickness on optical performance: If the coating is too thin, light may pass through the coating without being effectively reflected, resulting in reduced reflectivity. If the coating is too thick, light scattering may be weakened, while increasing material costs. Obviously, reflectivity is directly related to the absorption of different wavelengths of light; higher reflectivity leads to lower absorption, and vice versa.

[0027] 2. The impact of coating thickness on thermal properties: Thick coatings lengthen the heat conduction path, slowing heat transfer and reducing the surface temperature of the exterior wall. Thin coatings allow heat to penetrate more easily, resulting in higher surface temperatures. Furthermore, thicker coatings can absorb more ultraviolet radiation, thus slowing down the aging of the substrate. However, excessively thick coatings may increase heat accumulation, which is detrimental to the paint's heat dissipation performance.

[0028] 3. The Impact of Coating Thickness on Weather Resistance: Besides the indirect impacts on weather resistance caused by the effects on optical and thermal properties mentioned above, coating thickness is a key factor determining the weather resistance of a coating. Thinner coatings may crack or wear under long-term exposure, reducing weather resistance. Thicker coatings offer better protection against ultraviolet radiation, humidity, and chemical corrosion. Furthermore, excessively thick coatings may cause cracking or peeling due to stress concentration, affecting the overall weather resistance index.

[0029] In summary, in addition to considering the influence of the proportions of the three primary color pigments and the acrylic substrate on the optical and thermal properties of the coating, this invention also considers the feature input of increasing the coating thickness to further supplement the spatial distribution characteristics of the coating in practical applications, which is a necessary parameter for predicting coating performance.

[0030] The first neural network module is used to receive the quality ratio output by the data acquisition module and predict the solar reflectance of the target coating. The first neural network module can directly adopt the model from the aforementioned paper; its construction and training methods will not be elaborated here.

[0031] The second neural network module is used to receive the solar reflectance output by the first neural network module and the coating thickness output by the data acquisition module, and predict the weather resistance of the coating. The output of the second neural network module includes the ultraviolet absorption and surface temperature change of the target coating as intermediate parameters, and the comprehensive weather resistance index as the prediction result of the weather resistance of the target coating.

[0032] It should be understood that the comprehensive weather resistance index is a quantitative indicator used to evaluate the overall performance stability and anti-aging ability of coatings under long-term environmental exposure. It comprehensively considers the optical, physical, and chemical properties of the coating, providing a unified evaluation standard through a weighted or unweighted combination of various key performance parameters. Those skilled in the art should know that the weather resistance data can be obtained by testing the target coating using weather resistance test methods specified in national standards (such as GB / T 14522) or industry standards. Such weather resistance data is authoritative and reliable. Since this invention uses a neural network for black-box prediction, it cannot provide conclusions as authoritative as national standard test results. Therefore, this invention considers embedding experimental data into the components and training method of the second neural network, using the weather resistance evaluation index obtained through standard weather resistance tests as the training set to train the second neural network, thereby enhancing its robustness within the standard range.

[0033] Specifically, such as Figure 2 As shown, the second neural network employs a multi-layered architecture, including a shared hidden layer, an intermediate parameter output layer, independent hidden layers, and a prediction output layer connected sequentially. This architecture design has the following advantages: The shared hidden layer, serving as the fundamental processing unit of the second neural network module, primarily performs preliminary feature extraction from the input solar reflectivity and coating thickness data. This layer shares a similar feature processing mechanism with the first neural network module, enabling it to fully understand and process data related to optical properties. Specifically, the shared hidden layer can identify basic patterns and regularities in the input data, laying the foundation for subsequent complex analyses.

[0034] The intermediate parameter output layer is responsible for predicting two key intermediate parameters: UV absorption and surface temperature change. This layer is unique in that it not only outputs the predicted results but also provides important reference data for subsequent weather resistance assessments. By analyzing these two intermediate parameters, we can better understand the coating's performance in real-world environments. For example, higher UV absorption usually indicates potentially accelerated material aging, while a larger surface temperature change may lead to thermal stress problems in the coating.

[0035] The independent hidden layer is a dedicated processing unit for deep feature analysis. It receives prediction results from the intermediate parameter output layer while maintaining a connection with the shared hidden layer. Through this fusion of multi-source information, a comprehensive analysis of coating performance is achieved. The importance of this layer lies in its ability to discover potential correlations between different features, analyzing the interaction between solar reflectivity, ultraviolet absorption, and temperature changes, thereby providing more comprehensive information support for the final weather resistance prediction.

[0036] The prediction output layer is the final decision-making unit of the second neural network module, and its core function is to generate a comprehensive weather resistance index for the coating. This layer integrates the analysis results of the preceding layers to output an evaluation index that can comprehensively reflect the weather resistance performance of the coating.

[0037] The network training module is connected to the first neural network module and the second neural network module respectively, and is used to train the first neural network module and the second neural network module; the network training module is also used to initialize several shared hidden layers and intermediate parameter output layers through transfer learning, and to train the second neural network through supervised learning.

[0038] Transfer learning is a method that improves the training efficiency and performance of a target task by transferring model parameters, structure, or knowledge from a source task to a target task. In this invention, considering the high cost and long time required to obtain complete experimental data, which may result in insufficient training data, transfer learning is used to leverage existing data and model experience, reducing the target task's dependence on large-scale training data. Therefore, considering the generalizability between the features learned in the first neural network and the intermediate layers of the multi-task learning model, the low-level feature extraction layer in the first neural network has successfully captured the relationship between input features and solar reflectivity. These features are also valuable for predicting ultraviolet absorption and surface temperature changes. Furthermore, the output (solar reflectivity) of the first neural network has a close physical relationship with the intermediate parameter predictions; for example, high solar reflectivity usually corresponds to lower ultraviolet absorption and smaller surface temperature changes. Therefore, initializing the shared hidden layer and intermediate parameter output layer through transfer learning can significantly reduce the training time required for random parameter initialization, accelerate convergence, and improve prediction performance.

[0039] In some preferred embodiments, a method for initializing several shared hidden layers and intermediate parameter output layers through transfer learning is provided, specifically including: like Figure 3 As shown, the network training module includes: The parameter transfer unit is used to transfer the network parameters of the first neural network module to the shared hidden layer of the second neural network module by freezing the weights. The parameter transfer unit is the core functional unit responsible for knowledge reuse in the network training module. Its main task is to transfer the trained network parameters from the first neural network module to the shared hidden layer of the second neural network module. This process can be compared to an experienced engineer transferring their expertise in optical performance prediction to a new system. Specifically, the parameter transfer unit achieves knowledge preservation and transfer through "freezing weights." Here, "freezing weights" can be understood as fixing the parameter values ​​in the first neural network module that have been fully trained and can accurately identify the optical properties of the coating, transforming mature experience into standard operating procedures.

[0040] The parameter transfer unit's operation can be divided into three stages: First, it comprehensively evaluates the parameters of the first neural network module and identifies those parameters that are most important for predicting solar reflectance; second, it establishes a parameter mapping relationship to ensure that these key parameters can find their corresponding positions in the shared hidden layer of the second neural network module; finally, it performs the actual parameter copying operation and activates the parameter locking mechanism to ensure that this experience will not be changed or lost during subsequent training.

[0041] The weight initialization unit is used to adaptively initialize the weights of the intermediate parameter output layer of the second neural network module. This unit handles network parameters that need to learn and adapt to new environments, particularly the weights of the intermediate parameter output layer of the second neural network module. The unit employs an "adaptive initialization" strategy, meaning it adjusts the initial values ​​of the parameters based on the new targets to be predicted (UV absorption and surface temperature changes).

[0042] The working mechanism of the weight initialization unit can be divided into the following steps: First, it analyzes the parameter distribution characteristics in the first neural network module to understand the structure of existing knowledge; second, it designs an appropriate weight initialization range based on the target that the intermediate parameter output layer needs to predict; finally, it performs the initialization operation and provides a good starting point for the subsequent training process. This adaptive initialization method maintains continuity with the original knowledge while reserving sufficient flexibility for learning new knowledge.

[0043] The collaborative work of these two units is crucial to the overall performance of the system. The parameter transfer unit ensures the effective use of existing experience, avoiding the waste of resources in repeated learning; while the weight initialization unit creates favorable conditions for acquiring new knowledge. This allows the system to both inherit existing successful experiences and adapt to the requirements of new prediction tasks. This design significantly improves the system's training efficiency and prediction accuracy, especially when training data is relatively limited, where this experience-based learning approach is particularly important.

[0044] Supervised learning is a machine learning method that trains a model to accurately predict target values ​​on unknown data by learning the mapping relationship between input features and target outputs. Its core idea is to provide a clearly labeled training dataset to guide the model in progressively optimizing its parameters, ultimately achieving effective modeling of complex tasks. Training data typically consists of pairs of input features X (such as experimental parameters or recipe proportions) and corresponding target outputs Y (such as experimental results or performance metrics). The model's goal is to minimize the error between the predicted and actual values ​​(such as mean squared error or cross-entropy), thereby obtaining a predictive model with good generalization ability.

[0045] In this invention, a supervised learning method is used to train a second neural network. Its core objective is to construct an intelligent predictive model for the weather resistance of coatings based on weather resistance evaluation indicators obtained from standard weather resistance tests. Specifically, through standard weather resistance tests, the input features of the coating (proportion of three primary color pigments, proportion of acrylic substrate, and coating thickness) and the corresponding weather resistance evaluation results are obtained (the specific evaluation indicators depend on the standards used and are not further limited here). These standard test data not only possess high scientific validity and authority but also provide clear supervisory signals for the model, guiding the adjustment of model parameters.

[0046] To fully utilize this high-quality experimental data, this invention uses data from standard experimental procedures as input to the model and the final weather resistance assessment results obtained from the experiment as the target output. By defining a reasonable loss function (such as the error of the comprehensive weather resistance index), the difference between the model's predicted results and the experimentally labeled values ​​is calculated, and the model's weights and bias parameters are gradually optimized. This approach ensures the scientific validity and reliability of the model's predictions while significantly improving its robustness within the standard experimental range.

[0047] Furthermore, by utilizing supervised learning methods, the model can not only learn the direct relationship between input features and target output, but also capture the inherent patterns and latent dynamics of the data through the training process. By learning the correlation between intermediate parameters such as pigment ratio, coating thickness, UV absorption, and surface temperature changes, the predictive ability of the comprehensive weather resistance index is further improved. Ultimately, the trained second neural network can accurately and efficiently predict the weather resistance of coatings with unknown formulations.

[0048] In some preferred embodiments, a specific structure is provided for a network training module to train a second neural network through supervised learning, including: The parameter freezing unit is used to freeze the parameters of the shared hidden layers and intermediate parameter output layers of the second neural network module during training. Since the shared hidden layers and intermediate parameter output layers have already acquired weights from the first neural network during the transfer learning phase, freezing these layers prevents their parameters from being updated during pre-training, thus reducing the number of parameters that need optimization and lowering training complexity. This allows computational resources to be concentrated on the newly introduced independent hidden layers and prediction output layers in the second neural network, enabling more efficient model training.

[0049] A pre-training unit is used to pre-train the second neural network module using a publicly available material property dataset.

[0050] Because obtaining weather resistance test data relies on standard experiments, requiring complex experimental conditions and long experimental cycles, the amount of experimental data is limited, making it difficult to meet the needs of deep learning models. Furthermore, small sample data may lead to ineffective model training, resulting in overfitting or inaccurate predictions. Therefore, this invention considers pre-training the model on publicly available large-scale material property datasets. This allows the model to learn the optical, thermal, and other properties of the materials in advance, compensating for the limitations of insufficient experimental data. After completing initial feature learning on the publicly available datasets, the model only requires a small amount of high-quality experimental data for fine-tuning to adapt to the target task.

[0051] Publicly available material performance datasets refer to structured collections of data containing material performance parameters released by research institutions, standards organizations, industry associations, or other public sources. These datasets are typically used to study the physical, chemical, and thermal properties of materials and are widely applied in materials design, performance optimization, and the development of predictive models. Examples include the Open Quantum Materials Database (OQMD), the Matbench Dataset, and the MaterialsProject database. This invention considers utilizing authoritative material databases containing optical, thermal, and environmental performance data to provide a scientific basis for the pre-training of a second neural network, helping the model achieve higher initial performance and generalization ability in coating performance prediction, while reducing experimental costs and data requirements.

[0052] Fine-tuning refers to the process of retraining an existing pre-trained model using a small amount of specific data related to the target task. Its goal is to adjust some or all of the model's parameters to improve its accuracy and adaptability on the target task. In this invention, fine-tuning is performed using the mass ratio and coating thickness data from the first neural network training. This data is highly reliable and contains clear weather resistance evaluation results, enabling accurate prediction of the overall weather resistance index based on the coating's mass ratio and coating thickness. The implementation is as follows: The network training module further includes a fine-tuning unit, used to acquire quality ratio data and coating thickness data for training the first neural network module as a first training set, and to fine-tune the second neural network module based on the first training set. Data from the standard experimental process is used as the model input, and the final weather resistance evaluation result obtained from the experiment is used as the target output. By defining a reasonable loss function (such as the error of the comprehensive weather resistance index), the difference between the model's predicted results and the experimentally labeled values ​​is calculated, and the model's weights and bias parameters are gradually optimized. This approach ensures the scientific validity and credibility of the model's prediction results while significantly improving its robustness within the standard experimental range.

[0053] Specifically, the fine-tuning unit includes: The test data acquisition subunit is used to conduct standard weathering tests based on the training data of the first neural network module and obtain the corresponding weathering evaluation index. The training set construction subunit is used to combine the quality ratio data, coating thickness data and corresponding weather resistance evaluation indicators to construct the first training set. A supervised training subunit is used to perform supervised training on the second neural network module based on the first training set.

[0054] Standard weather resistance tests were conducted on coating samples with different ratios of three primary color pigments to acrylic substrates and varying coating thicknesses to obtain key performance data related to weather resistance. The mass ratio data, coating thickness data, and corresponding weather resistance evaluation indicators obtained from these experiments were compiled to construct a first training set. This set not only reflects the true physical and chemical relationship between coating formulation parameters and weather resistance indicators but also provides accurate and high-quality supervision signals for the training of a second neural network. Based on this, supervised learning methods were used to train the second neural network, enabling it to effectively learn the mapping relationship between input features (mass ratio, coating thickness) and output results (weather resistance evaluation indicators). Through this training process, the second neural network can gradually optimize its internal parameters and accurately predict the weather resistance performance of new coating formulations under specific experimental conditions. This method not only fully utilizes the scientific rigor and authority of the experimental data but also ensures the reliability and practicality of the model's prediction results, providing solid data and technical support for intelligent coating performance evaluation.

[0055] It should be understood that the first training set data mainly comes from standard weathering tests, which are costly and time-consuming, and the large-scale data obtained is limited, making it difficult to meet the training needs of deep learning models. Furthermore, insufficient data may lead to incomplete coverage of input features by the model, resulting in overfitting or inaccurate predictions. Therefore, in some preferred embodiments, to improve the system's predictive ability, the network training module is also equipped with a data augmentation unit, which expands the training set by randomly perturbing the existing data.

[0056] The data augmentation unit is a specialized functional unit within the network training module. Its core task is to generate richer training samples from limited raw data through intelligent data processing. In some preferred embodiments, it specifically includes two sub-units: The perturbation generation subunit is used to randomly perturb the mass ratio data and coating thickness data within a preset range. This subunit is responsible for performing regular variation processing on the original data. Specifically, it applies carefully designed random perturbations to the mass ratio data and coating thickness data. This process is akin to fine-tuning a formula in a laboratory: for the mass ratio of the three primary color pigments, the perturbation generation subunit applies a random variation of approximately ±5% to the original ratio. This range is carefully weighed—too small a variation may not provide effective data enhancement, while too large a variation may lead to an unrealistic formula. Similarly, for the coating thickness, the perturbation generation subunit applies a random fluctuation of approximately ±10%, a range that takes into account thickness errors that may occur during actual coating processes.

[0057] The verification subunit is used to verify whether the perturbed data meets the preset physical and technological constraints. The verification subunit is responsible for ensuring that the perturbation-generated data meets the actual physical and technological constraints, avoiding the generation of unreasonable samples.

[0058] In some preferred embodiments, the present invention also provides a coating weather resistance prediction device, comprising: a processor; a memory storing a computer program, wherein the computer program, when executed by the processor, is used to load the above-mentioned coating weather resistance prediction system based on a multi-task supervised neural network.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A coating weather resistance prediction system based on a multi-task supervised neural network, characterized in that, include: The data acquisition module is used to obtain the mass ratio of the three primary color pigments to the acrylic substrate and the coating thickness of the target coating. The first neural network module is used to receive the quality ratio output by the data acquisition module and predict the solar reflectance of the target coating. The second neural network module is used to receive the solar reflectance output by the first neural network module and the coating thickness output by the data acquisition module, and predict the weather resistance of the coating. The output of the second neural network module includes the ultraviolet absorption and surface temperature change of the target coating as intermediate parameters, and the comprehensive weather resistance index as the weather resistance prediction result of the target coating. The second neural network module includes a shared hidden layer, an intermediate parameter output layer, an independent hidden layer, and a prediction output layer connected in sequence. The network training module is connected to the first neural network module and the second neural network module respectively, and is used to train the first neural network module and the second neural network module; the network training module is also used to initialize several shared hidden layers and intermediate parameter output layers through transfer learning, and to train the second neural network through supervised learning.

2. The coating weather resistance prediction system based on a multi-task supervised neural network as described in claim 1, characterized in that, The network training module includes: The parameter migration unit is used to migrate the network parameters of the first neural network module to the shared hidden layer of the second neural network module by freezing the weights. The weight initialization unit is used to perform adaptive initialization of the weights of the intermediate parameter output layer of the second neural network module.

3. The coating weather resistance prediction system based on a multi-task supervised neural network as described in claim 1 or 2, characterized in that, The network training module also includes: The parameter freezing unit is used to freeze the parameters of the shared hidden layer and intermediate parameter output layer of the second neural network module during training. A pre-training unit is used to pre-train the second neural network module using a publicly available material property dataset.

4. The coating weather resistance prediction system based on a multi-task supervised neural network as described in any one of claims 1 to 3, characterized in that, The network training module also includes: The fine-tuning unit is used to acquire the quality ratio data and coating thickness data used to train the first neural network module as a first training set, and to fine-tune the second neural network module based on the first training set.

5. The coating weather resistance prediction system based on a multi-task supervised neural network as described in claim 4, characterized in that, The fine-tuning unit includes: The test data acquisition subunit is used to conduct standard weathering tests based on the training data of the first neural network module and obtain the corresponding weathering evaluation index. The training set construction subunit is used to combine the quality ratio data, coating thickness data and corresponding weather resistance evaluation indicators to construct the first training set. The supervised training subunit is used to perform supervised training on the second neural network module based on the first training set.

6. The coating weather resistance prediction system based on a multi-task supervised neural network as described in claim 4 or 5, characterized in that, The network training module further includes a data augmentation unit, used to augment the first training set through random perturbation.

7. The coating weather resistance prediction system based on a multi-task supervised neural network as described in claim 6, characterized in that, The data expansion unit includes: The perturbation generation subunit is used to randomly perturb the mass ratio data and coating thickness data within a preset range. The verification subunit is used to verify whether the disturbed data meets the preset physical and technological constraints.

8. A device for predicting the weather resistance of coatings, characterized in that, include: processor; A memory storing a computer program, which, when executed by the processor, is used to load the coating weather resistance prediction system based on a multi-task supervised neural network as described in any one of claims 1 to 7.

Images