Multi-source feature fusion-based intelligent screening method and system for waterproof roll material interlayer adhesive

Abstract

This invention relates to the field of building waterproofing materials technology, specifically to an intelligent screening method for interlayer adhesives in waterproof membranes based on multi-source feature fusion. The method includes: constructing a multi-source feature database comprising multi-dimensional feature data of a first waterproof membrane, a second waterproof membrane, and candidate adhesives; establishing a hybrid neural network model, pre-constructing a training dataset, and training the hybrid neural network model using the training dataset until the total loss function converges; inputting the multi-dimensional feature data of the first and second waterproof membranes into the trained hybrid neural network model to screen and obtain a superior adhesive formulation. This invention introduces Physical Information Neural Network (PINN) into the field of waterproof membrane bonding. Through an explicit physical parameter branching structure, it embeds interfacial chemistry and fracture mechanics mechanisms as constraints into the model, overcoming the shortcomings of purely data-driven models, such as lack of physical interpretability and poor extrapolation ability, thereby improving the accuracy and reliability of predictions.

Description

Technical Field

[0001] This invention relates to the field of building waterproofing materials technology, specifically to a method and system for intelligent screening of interlayer adhesives for waterproof membranes based on multi-source feature fusion. Background Technology

[0002] In modern building waterproofing projects, composite waterproofing methods are often used to improve the reliability and durability of the waterproofing layer. This involves laying two or more layers of waterproofing membranes made of different materials (such as a composite of bitumen-based membranes and polymer membranes). The reliability of the bond between the two layers of waterproofing membranes is crucial to the entire waterproofing system. Currently, adhesives are mainly used for interlayer bonding. However, due to the significant differences in the surface chemical properties, physical structure, and mechanical properties of the membrane materials (such as SBS modified bitumen, TPO, EPDM, PVC, etc.), the compatibility, bond strength, and long-term durability of the adhesive with the interface of the two membranes are difficult to guarantee.

[0003] Current technologies mainly rely on empirical methods or simple single-index screening (such as testing only peel strength). This approach has significant drawbacks: First, the adhesion failure mechanism is complex, involving multiple dimensions such as surface free energy, polarity matching, differences in thermal expansion coefficients, and chemical cross-linking reactions, which cannot be fully reflected by a single index; second, screening adhesives through a large number of trial-and-error experiments is costly and time-consuming, and it is difficult to cover different construction environments (temperature and humidity) and long-term aging conditions; finally, there is a lack of dynamic prediction capabilities for the long-term service performance of adhesives, which often leads to quality problems such as short-term adhesion that fails, but debonding and water seepage after experiencing damp heat or freeze-thaw cycles in engineering applications.

[0004] Therefore, there is an urgent need for a method that can comprehensively consider the properties of the roll material itself, the properties of the adhesive, and environmental and process factors, and efficiently and accurately screen and predict the performance of adhesives. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent screening of interlayer adhesives for waterproof membranes based on multi-source feature fusion. It introduces Physical Information Neural Network (PINN) into the field of waterproof membrane bonding, and embeds interfacial chemistry and fracture mechanics mechanisms as constraints into the model through an explicit physical parameter branch structure. This overcomes the shortcomings of pure data-driven models, such as lack of physical interpretability and poor extrapolation ability, and improves the accuracy and reliability of predictions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a smart screening method for interlayer adhesives in waterproof membranes based on multi-source feature fusion, comprising:

[0007] Construct a multi-source feature database that includes multi-dimensional feature data of the first waterproof membrane, the second waterproof membrane, and candidate adhesives;

[0008] A hybrid neural network model is established. A training dataset containing multi-source features and corresponding bonding performance labels is pre-constructed based on publicly available literature data. The hybrid neural network model is trained using the training dataset until the total loss function converges.

[0009] The multidimensional feature data of the first and second waterproof membranes are input into the trained hybrid neural network model to screen for the optimal adhesive formulation and output a recommended solution.

[0010] The actual performance data collected on-site is used to iterate and update the hybrid neural network model.

[0011] Furthermore, the hybrid neural network model is a multi-task learning model, which includes:

[0012] The input layer is used to receive multidimensional feature data of the first waterproof membrane, the second waterproof membrane, and candidate adhesives;

[0013] Feature embedding layer, used to vectorize and embed discrete features;

[0014] The physical information constraint layer employs an explicit physical parameter branching structure to embed the physical mechanism of interface bonding failure into the model training process.

[0015] A multi-task learning layer is used to predict short-term adhesion performance indicators and long-term durability levels.

[0016] The output layer is used to output predicted values ​​for short-term adhesion performance and long-term durability levels.

[0017] Furthermore, the physical information constraint layer adopts an explicit physical parameter branching structure, including:

[0018] The surface feature encoding branch of the roll material is used to map the surface free energy and polar components of the roll material into the interface wetting latent variable h1;

[0019] The adhesive feature encoding branch is used to map the surface tension and glass transition temperature of the adhesive to the latent variable h2 of the adhesive.

[0020] The physical formula calculation unit, with built-in Young's equation and interfacial fracture energy formula, is used to calculate the theoretical wetting coefficient based on h1 and h2. Theoretical interface fracture energy ;

[0021] Constraint loss calculation unit, used to calculate model predictions. , Compared with theoretical value , The deviation between them is calculated and added as a regularization term to the total loss function.

[0022] Furthermore, the multi-task learning layer includes:

[0023] The first output task is a regression task used to predict short-term bonding performance indicators.

[0024] The second output task is a classification task used to predict the long-term durability level.

[0025] The short-term adhesion performance index includes at least one of peel strength, shear strength and initial tack after 24 hours; the long-term durability grade includes at least one of peel strength retention rate after heat aging, adhesion strength decay rate after water immersion, and interface failure mode after freeze-thaw cycles.

[0026] Furthermore, the multi-dimensional feature data of the first and second waterproof membranes are input into a trained hybrid neural network model to screen for optimal adhesive formulations and output recommended solutions, including:

[0027] The multidimensional feature data of the first and second waterproof membranes to be screened, as well as the preset construction environment conditions, are input into the trained hybrid neural network model.

[0028] Using a Bayesian optimization algorithm, a global search is performed in the continuous feature space of candidate adhesives to maximize the comprehensive score of bonding performance output by the hybrid neural network model, and to iteratively find a better combination of virtual formulation parameters for the adhesive.

[0029] The optimal virtual formula parameter combination obtained through optimization is compared with the feature similarity of the preset finished adhesive database. According to the Euclidean distance principle, the finished adhesive model with the closest formula characteristics that is already available on the market is matched. The Euclidean distance between the optimal virtual formula and the matched finished adhesive is calculated. If the Euclidean distance is greater than the preset deviation threshold, it is determined that the existing finished adhesive cannot meet the requirements, and a "custom development is recommended" prompt is output, and a recommended formula is output for the manufacturer's reference.

[0030] Output the matched finished adhesive model, supplier information, and the corresponding optimal combination of construction process parameters.

[0031] Furthermore, it also includes: using the SHAP method to calculate the contribution value of each input feature to the output prediction result, and outputting the ranking of key factors affecting the bonding reliability and the corresponding explanatory text.

[0032] Furthermore, actual performance data is collected from the field, and the hybrid neural network model is iteratively updated and feedback is performed, including:

[0033] Collect actual performance data after on-site bonding using the superior adhesive;

[0034] The actual performance data is compared with the model predictions to calculate the prediction deviation;

[0035] The prediction bias is used as a feedback signal to incrementally learn the hybrid neural network model and update the model parameters.

[0036] A multi-source feature fusion-based intelligent screening system for interlayer adhesives in waterproof membranes includes:

[0037] The feature extraction module is used to obtain multi-dimensional feature data of the first waterproof membrane, the second waterproof membrane, and the candidate adhesives;

[0038] A database module, connected to the feature extraction module, is used to store the multidimensional feature data;

[0039] The model training and inference module is equipped with a pre-trained hybrid neural network model.

[0040] The optimization recommendation module is used to execute optimization algorithms and output optimal adhesive and construction process parameters.

[0041] The feedback iteration module is used to collect actual performance data on-site and trigger incremental learning of the model.

[0042] Furthermore, the feature extraction module includes:

[0043] A contact angle measuring instrument is used to measure the surface free energy of waterproof membranes;

[0044] Fourier transform infrared spectroscopy is used to analyze the chemical composition of adhesives;

[0045] A rotational viscometer is used to measure the viscosity of adhesives.

[0046] Furthermore, it also includes a visualization module, which displays the optimal adhesive information, construction process parameters, and feature importance analysis charts generated using the SHAP method in a graphical interface.

[0047] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0048] 1. The intelligent screening method and system for interlayer adhesives of waterproof membranes based on multi-source feature fusion. This invention introduces physical information neural network (PINN) into the field of waterproof membrane bonding. Through explicit physical parameter branching structure, the interfacial chemistry and fracture mechanics mechanism are embedded as constraints into the model, which overcomes the shortcomings of pure data-driven models that lack physical interpretability and have poor extrapolation ability, and improves the accuracy and reliability of prediction.

[0049] 2. This intelligent screening method and system for interlayer adhesives of waterproof membranes based on multi-source feature fusion constructs a multi-source feature database covering the membrane body, adhesive, and environmental processes, and adopts a multi-task learning architecture to achieve simultaneous prediction of short-term bond strength and long-term durability, thus solving the pain point of "emphasizing short-term and neglecting long-term" in the existing technology.

[0050] 3. The intelligent screening method and system for interlayer adhesives of waterproof membranes based on multi-source feature fusion utilizes Bayesian optimization algorithm for intelligent screening of adhesives and reverse recommendation of process parameters. It also innovatively adds a finished adhesive matching unit, which enables the recommendation results to be directly applied, significantly shortening the research and development and application cycle and reducing costs.

[0051] 4. The intelligent screening method and system for interlayer adhesives of waterproof membranes based on multi-source feature fusion introduces incremental learning and interpretability analysis mechanisms, enabling the system to continuously evolve and providing engineers with clear decision-making basis, thereby improving the system's practicality and intelligence level.

[0052] 5. This intelligent screening method and system for interlayer adhesives in waterproof membranes, which integrates multiple features, uses a hybrid architecture of physical constraints and data-driven approaches. The model can identify performance relationships under multivariate coupling conditions and output accurate critical condition predictions, providing quantifiable decision-making basis for adhesive screening under extreme working conditions. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the structure of the present invention;

[0054] Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1-2 The intelligent screening method for interlayer adhesives of waterproof membranes based on multi-source feature fusion in this embodiment includes the following steps:

[0057] S1. Construct a multi-source feature database that includes multi-dimensional feature data of the first waterproof membrane, the second waterproof membrane, and candidate adhesives;

[0058] S2. Establish a hybrid neural network model. Based on publicly available literature data, pre-construct a training dataset containing multi-source features and corresponding bonding performance labels. Train the hybrid neural network model using the training dataset until the total loss function converges.

[0059] S3. Input the multi-dimensional feature data of the first and second waterproof membranes into the trained hybrid neural network model, screen out the optimal adhesive formulation, and output the recommended solution.

[0060] S4. Collect actual performance data from the site and iterate and update the hybrid neural network model accordingly.

[0061] In step S1 of this embodiment, the multidimensional characteristic data of the waterproof membrane includes surface free energy and its polar component, material type, thickness, surface roughness, tensile strength, and elongation at break, while the multidimensional characteristic data of the adhesive includes matrix resin type, glass transition temperature, viscosity, type and ratio of curing agent, type and content of tackifying resin, open time, and curing time.

[0062] In addition, the multi-source feature database also includes construction and environmental feature data, specifically including construction environment temperature, construction environment relative humidity, adhesive coating thickness, curing time, and interface compaction pressure.

[0063] In step S2 of this embodiment, the hybrid neural network model is a multi-task learning model, which includes:

[0064] The input layer is used to receive multidimensional feature data of the first waterproof membrane, the second waterproof membrane, and candidate adhesives;

[0065] Feature embedding layer, used to vectorize and embed discrete features;

[0066] The physical information constraint layer employs an explicit physical parameter branching structure to embed the physical mechanism of interface bonding failure into the model training process.

[0067] A multi-task learning layer is used to predict short-term adhesion performance indicators and long-term durability levels.

[0068] The output layer is used to output predicted values ​​for short-term adhesion performance and long-term durability levels.

[0069] The physical information constraint layer adopts an explicit physical parameter branching structure, including:

[0070] 1. The surface feature encoding branch of the roll material is used to map the surface free energy and polar components of the roll material into the interface wetting latent variable h1;

[0071] The surface free energy of roll A and roll B ( (Unit: mN / m, Range: 20-50), Polar component ( Features such as surface roughness (Ra, unit mN / m, range 0-15) and surface roughness (Ra, unit μm, range 0-10) are first normalized in terms of dimensionlessness: each feature value is divided by its upper limit of physical range to obtain the dimensionless normalized value. , , The values ​​range from [0,1].

[0072] The normalized value is input into a fully connected network layer and mapped to the interface wetting latent variable h1. The dimension of h1 is 1, and its physical meaning is the normalized comprehensive index of the surface energy of the roll material.

[0073] 2. Adhesive feature encoding branch, used to map the surface tension and glass transition temperature of the adhesive to the latent variable h2 of the adhesive;

[0074] The surface tension of the adhesive ( ,unit The characteristics, such as glass transition temperature (Tg, in °C, range -60~50) and curing agent ratio (c, range 0-0.5), are also normalized in terms of dimensions. , , The values ​​range from [0,1].

[0075] The normalized values ​​are input into a fully connected network layer and mapped to the adhesive latent variable h2. The dimension of h2 is 1, and its physical meaning is the normalized comprehensive performance index of the adhesive.

[0076] 3. Physical formula calculation unit, with built-in Young's equation and interface fracture energy formula, used to calculate the theoretical wetting coefficient based on h1 and h2. Theoretical interface fracture energy ;

[0077] This unit incorporates Young's equations and an interface fracture energy formula based on fracture mechanics, but expresses them in normalized form:

[0078] Theoretical wetting coefficient: , with a value range of [-1, 1], characterizes the quality of wettability;

[0079] Theoretical interfacial fracture energy: , where k is an empirical constant (with a value of 1.0, and its dimensions are aligned with the predicted value), and its value range is [0,1].

[0080] The inputs h1 and h2 to the above formula are both normalized dimensionless values, and the output is... and It is also a dimensionless value, and is consistent with the predictions of the main branch of the model. , (Also normalized) They are consistent in dimensions.

[0081] 4. Constraint loss calculation unit, used to calculate model predictions. , Compared with theoretical value , The deviation between them is calculated and added as a regularization term to the total loss function.

[0082] The main branch of the model predicts the normalized wetting coefficient during training. and normalized fracture energy (The activation function is obtained through the output of an additional fully connected layer.) Ensure the output is in [-1, 1] or Ensure the output is in [0,1].

[0083] Will and , and The mean square error is calculated to obtain the physical constraint loss. .

[0084] It should be noted that although h1 and h2 are obtained by neural network mapping, their numerical range is constrained to [0,1], corresponding to the normalized representations of the comprehensive surface energy index of the roll material and the comprehensive performance index of the adhesive, respectively. Therefore, the simplified form of Young's equation and the interfacial fracture energy formula ( , It is reasonable to impose mechanistic constraints because this form preserves the monotonicity and interaction of the original physical formula in the normalized space. Those skilled in the art will understand that when h1 and h2 are monotonically positively correlated with the surface energy of the roll material and the surface tension of the adhesive, respectively, the above simplified formula can effectively reflect the changing trends of wettability and interfacial toughness.

[0085] Meanwhile, the multi-task learning layer includes:

[0086] The first output task is a regression task used to predict short-term bonding performance indicators.

[0087] The second output task is a classification task used to predict the long-term durability level.

[0088] The short-term adhesion performance index includes at least one of peel strength, shear strength and initial tack after 24 hours; the long-term durability grade includes at least one of peel strength retention rate after heat aging, adhesion strength decay rate after water immersion, and interface failure mode after freeze-thaw cycles.

[0089] In this embodiment, the multidimensional feature data of the first and second waterproof membranes are input into a trained hybrid neural network model to screen for a better adhesive formulation and output a recommended solution, including:

[0090] 1. Input the multidimensional feature data of the first and second waterproof membranes to be screened, as well as the preset construction environment conditions, into the trained hybrid neural network model;

[0091] 2. Using a Bayesian optimization algorithm, a global search is performed in the continuous feature space of candidate adhesives to maximize the comprehensive score of bonding performance output by the hybrid neural network model, and the optimal combination of virtual formulation parameters for the adhesive is iteratively sought.

[0092] Specifically, the Bayesian optimization algorithm includes the following steps:

[0093] 1) First, randomly select several sets of adhesive formulation parameters (such as curing agent ratio, tackifying resin content, etc.), input them into the trained hybrid neural network model, and obtain the corresponding comprehensive score of bonding performance;

[0094] 2) Based on these known recipes and scores, construct a "guessing model" (i.e., a Gaussian process surrogate model) to predict the possible score values ​​of other untried recipes and the reliability of the prediction;

[0095] 3) Using a “selection rule” (i.e., collection function), find the next set of recipe parameters that are “most likely to achieve a high score” in the candidate recipe space, and use it as the next candidate recipe to try.

[0096] 4) Input the newly selected recipe parameters into the hybrid neural network model to obtain the true predicted score, and add this new sample (recipe parameters + score) to the known dataset;

[0097] 5) Repeat steps 2) to 4) above to continuously update the "guessing model" and gradually approach the optimal solution of the model; when the number of iterations reaches the upper limit, or when the score improvement is lower than the preset threshold for several consecutive times, stop the search and output the optimal combination of formula parameters that has been found.

[0098] 3. The optimal virtual formula parameter combination obtained through optimization is compared with the feature similarity of the preset finished adhesive database. According to the Euclidean distance principle, the finished adhesive model with the closest formula characteristics that is already available on the market is matched. The Euclidean distance between the optimal virtual formula and the matched finished adhesive is calculated. If the Euclidean distance is greater than the preset deviation threshold, it is determined that the existing finished adhesive cannot meet the requirements, and a "custom development is recommended" prompt is output. A recommended formula is also output for the manufacturer's reference.

[0099] The specific steps are as follows:

[0100] 1) Represent the optimal combination of virtual recipe parameters as a list, for example:

[0101]

[0102] 2) Represent the corresponding parameters of each finished adhesive in the finished adhesive database as a similar list, such as V1, V2, V3...;

[0103] 3) Calculate the "similarity" between the virtual formula and each finished adhesive. The calculation method is as follows: subtract the values ​​at corresponding positions in the two lists, square the results, sum them, and then take the square root. The result is the "Euclidean distance". The smaller the distance, the closer the two are.

[0104] 4) From all the finished adhesives, select the one that is closest to the virtual formula as the matching result;

[0105] 5) If the nearest distance is greater than a pre-set "tolerance threshold", it means that the existing finished adhesive is far from the ideal formula and there is no perfectly matching product. At this time, the prompt "custom development is recommended" will be output, and a virtual formula will be output for the manufacturer's reference. If the nearest distance is less than or equal to the threshold, the matching finished adhesive model and supplier information will be output directly.

[0106] 4. Output the matched finished adhesive model, supplier information, and the corresponding optimal combination of construction process parameters.

[0107] In addition, in this embodiment, step S3 also includes an interpretability analysis step:

[0108] The SHAP method is used to calculate the contribution of each input feature to the output prediction result, and the key factors affecting the adhesion reliability are ranked and the corresponding explanatory text is output.

[0109] Finally, in this embodiment, actual performance data from the field is collected to iterate and update the hybrid neural network model, including:

[0110] 1. Collect actual performance data after on-site bonding using the superior adhesive;

[0111] 2. Compare the actual performance data with the model predictions and calculate the prediction deviation;

[0112] 3. Use the prediction deviation as a feedback signal to perform incremental learning on the hybrid neural network model and update the model parameters.

[0113] In summary, this method first constructs a multi-source feature database encompassing the membrane itself, adhesives, and environmental processes. Then, it establishes and trains a hybrid neural network model, which includes an explicit physical parameter branch structure and a multi-task learning layer. The physical branch embeds interfacial fracture mechanics and surface chemistry mechanisms into the model, while the multi-task learning layer simultaneously predicts short-term bond strength and long-term durability levels. Simultaneously, a Bayesian optimization algorithm is used to intelligently select the optimal virtual adhesive formulation based on the trained model, and a finished adhesive matching unit outputs directly purchasable finished product models and construction process parameters. Finally, the model undergoes incremental learning using field feedback data. This invention, through a hybrid mechanism- and data-driven approach, solves the problems of low efficiency and poor reliability in selecting interlayer adhesives for heterogeneous waterproof membranes, achieving accurate prediction and intelligent recommendation of bonding performance, and significantly improving the durability and intelligence level of the waterproofing system.

[0114] This invention proposes an intelligent screening system for interlayer adhesives in waterproof membranes based on multi-source feature fusion, comprising a feature extraction module 100 and a database module 200.

[0115] The feature extraction module 100 is used to acquire multi-dimensional feature data of the first waterproof membrane, the second waterproof membrane, and the candidate adhesive. In practical applications, this module can acquire feature data by interfacing with the data interfaces of various testing instruments or by manual input by operators. For example, for the membrane, its surface free energy data can be obtained through a contact angle meter; for the adhesive, its chemical composition can be analyzed by a Fourier transform infrared spectrometer, and its viscosity can be measured by a rotational viscometer.

[0116] Database module 200, connected to feature extraction module 100, is used to store the multidimensional feature data, specifically storing the aforementioned feature data and corresponding bonding performance verification data, forming a structured dataset. To construct the initial training set, this embodiment adopts the following method:

[0117] Data Sources: Relevant literature on "waterproof membrane bonding," "peel strength," and "durability" was retrieved from publicly available academic databases (such as Web of Science and CNKI), and publicly available experimental data was extracted. For example, initial training samples could be constructed based on publicly available research data on the peel strength of SBS modified bitumen membranes and TPO polymer membranes with different adhesive combinations, as well as reports on bonding performance under different temperature and humidity conditions.

[0118] Data completion: For features required by the feature system of this invention but not directly provided in the literature (such as the polar component of surface free energy), the standard values ​​in the "Polymer Surface and Interface Handbook" are supplemented according to the type of roll material.

[0119] Data labeling: The short-term peel strength values ​​reported in the literature were used as labels for the regression task; based on the performance degradation after aging described in the literature, long-term durability was manually labeled according to the A / B / C three-level standard.

[0120] Data volume: Through the above methods, more than 2,000 valid samples were collected and organized, covering the combination of mainstream roll materials such as SBS, APP, TPO, EPDM, and PVC with major adhesives such as polyurethane, acrylic, butyl rubber, and MS glue.

[0121] It should be noted that the above dataset construction method is entirely based on publicly available literature data. Anyone skilled in the art can reproduce the dataset based on the data sources and methods provided by this invention, thereby reproducing the model training process.

[0122] The model training and inference module 300 is deployed with pre-trained hybrid neural network models, such as... Figure 2 As shown, the structure of the model is as follows:

[0123] Input layer: Receives normalized feature vectors , where n is the total number of features.

[0124] Embedded layer: For category features such as "roll material type" and "adhesive matrix resin type", convert them into dense vectors.

[0125] Fully connected layer: Several layers of fully connected neural networks are used to automatically extract high-level nonlinear features.

[0126] Physical information constraint layer: adopts an explicit physical parameter branch structure.

[0127] Surface feature coding branch of roll material: The surface free energy of roll material A and roll material B ( Unit: mN / m, range: 20-50; polar component ( Dimensional normalization is performed on features such as surface roughness (Ra, unit mN / m, range 0-15) and surface roughness (Ra, unit μm, range 0-10): Each feature value is divided by its upper limit of physical range to obtain a dimensionless normalized value. , , The values ​​range from [0,1]. These normalized values ​​are then input into a fully connected network layer and mapped to the interface wetting latent variable h1. The dimension of h1 is 1, and its physical meaning is the normalized comprehensive index of the surface energy of the roll material.

[0128] Adhesive feature coding branch: The surface tension of the adhesive ( The dimensions of the following features were normalized: glass transition temperature (Tg, °C, range -60 to 50), curing agent ratio (c, range 0 to 0.5). , , The values ​​are all in the range [0,1]. These normalized values ​​are then input into a fully connected network layer and mapped to the adhesive latent variable h2. The dimension of h2 is 1, and its physical meaning is the normalized comprehensive performance index of the adhesive.

[0129] Physical Formula Calculation Unit: This unit incorporates Young's equations and an interfacial fracture energy formula based on fracture mechanics. It calculates the theoretical wetting coefficient based on h1 and h2. (Simulated spreading coefficient), and theoretical interfacial fracture energy (Simplified form is) (where k is an empirical constant). This unit does not contain trainable parameters and only performs forward computation.

[0130] Constraint loss calculation: The main branch of the model predicts a wetting coefficient during training. and a fracture energy (Output via an additional fully connected layer). and , and The mean square error is calculated to obtain the physical constraint loss. , where λ1 and λ2 are weighting coefficients. It is added as a regularization term to the total loss function.

[0131] Multi-task learning layer: The model is divided into two branches here.

[0132] Branch 1 (Regression): Outputs short-term adhesion performance indicators, such as the predicted 180° peel strength value (N / mm) after 24 hours.

[0133] Branch 2 (Classification): Output long-term durability rating. In this embodiment, durability is defined as three levels: Grade A (Excellent, peel strength retention rate >80% after aging, cohesive failure); Grade B (Acceptable, retention rate 50%-80%, interface failure rate <30%); Grade C (Unacceptable, retention rate <50% or interface failure rate >50%).

[0134] Output layer: Integrates the outputs of the two branches.

[0135] The model was trained using sample data from database module 200, employing the Adam optimizer, with the total loss function being... Train until the loss function converges to obtain a usable prediction model.

[0136] This example demonstrates the process of intelligently screening adhesives using a trained model. Suppose a user needs to find the optimal adhesive for "SBS modified bitumen rolls" (roll A) and "TPO polymer rolls" (roll B).

[0137] First, feature data for the two types of roll materials were obtained through the feature extraction module: the surface free energy of SBS roll material was 32 mN / m, with a polar component of 5 mN / m; the surface free energy of TPO roll material was 28 mN / m, with a polar component of 3 mN / m. The user set the construction environment to "temperature 25℃, humidity 60%".

[0138] The optimization recommendation module 400 calls the trained model and uses a Bayesian optimization algorithm to search in the continuous feature space of the adhesive (such as the curing agent content of 0.1-0.5% and the tackifying resin content of 5%-20%). After 30 iterations of optimization, the optimal virtual formulation parameter combination recommended by the model is as follows: the matrix resin is silane-modified polyether (MS), the curing agent ratio is 0.3%, the tackifying resin content is 12%, the surface tension is 28 mN / m, and the glass transition temperature is -45℃.

[0139] Subsequently, the finished adhesive matching unit compared the virtual formula with a database of commercially available finished adhesives based on feature similarity. Following the Euclidean nearest neighbor principle, the closest matched finished adhesive model was identified as "MS-3200" (a certain brand of silane-modified polyether adhesive), with a similarity of 92%, exceeding the preset threshold of 85%. The system output the finished adhesive model and supplier information, and suggested an application thickness of 0.8 mm, an open time of 15 minutes, and an application temperature range of 15-30℃. The model predicted a short-term peel strength of 4.2 N / mm and a long-term durability rating of A.

[0140] This embodiment adds interpretability analysis functionality and demonstrates the invention's ability to identify multivariate coupling relationships.

[0141] After optimizing the recommendation module 400's output results, the system automatically calls the SHAP analysis tool to calculate the contribution of each input feature to the output results. The analysis results show that "the polar component of the surface free energy of membrane A," "the glass transition temperature (Tg) of the adhesive," and "the ambient temperature during construction" are the top three key factors affecting bond reliability. Based on this, the system generates an explanatory text: "The current SBS membrane has a low polar component of surface free energy (5 mN / m), requiring the selection of an adhesive with low surface tension and a Tg below -40℃. Construction is recommended at an ambient temperature of 15-30℃; otherwise, the risk of debonding will increase."

[0142] The model of this invention, through a hybrid architecture of physical constraints and data-driven approaches, can identify the performance relationships of different adhesive systems under multivariate coupling conditions. In the analysis of the bonding scenario between SBS and TPO roofing membranes, the model output results show:

[0143] When the ambient temperature is above 10℃, the predicted bond strength of high polarity polyurethane adhesive is higher than that of butyl rubber.

[0144] When the ambient temperature is below 10℃, the two performance curves intersect.

[0145] The specific critical conditions output by the model are: when the temperature is below 8.5℃ and the relative humidity is above 75%, the predicted bond strength of butyl rubber surpasses that of polyurethane.

[0146] When the temperature is below 5°C and the humidity is above 82%, the predicted durability grade of polyurethane is C (unacceptable), while the predicted durability grade of butyl rubber is A (excellent).

[0147] The above values ​​are examples of model outputs for this input scenario. Those skilled in the art can obtain corresponding prediction results using the method of this invention based on different input parameters (such as roll material type, adhesive properties, environmental conditions, etc.).

[0148] Traditional experimental methods require simultaneous control of multiple variables such as temperature, humidity, adhesive type, and coating thickness. The number of experimental combinations increases exponentially, resulting in extremely high costs and making it difficult to systematically cover the entire parameter space. This invention, through a physically constrained hybrid neural network model, successfully extrapolates and quantifies this multivariate coupling relationship based on limited training data and by utilizing embedded interfacial chemical mechanisms and fracture mechanics constraints. This provides a quantifiable decision-making basis for adhesive screening under extreme conditions.

[0149] This embodiment demonstrates the feedback iteration mechanism of the system. In a real-world engineering project, the optimal adhesive solution recommended by this invention was used for construction. After construction was completed, on-site personnel collected the actual peel strength value (e.g., a measured value of 4.0 N / mm) using a portable adhesive strength tester. This data was then uploaded to the feedback iteration module 500 via a mobile terminal.

[0150] The feedback iteration module 500 compares the measured value (4.0 N / mm) with the value predicted by the model at the time of recommendation (4.2 N / mm), and calculates the prediction deviation as -0.2 N / mm (approximately -4.8%). This deviation data, along with the actual construction environment temperature and humidity, adhesive thickness, and other process parameters recorded at the time, is stored as a new sample point in the database module 200.

[0151] When the accumulated feedback data reaches a preset threshold (e.g., 100 sets of new data) or the prediction deviation continues to exceed a preset error range (e.g., ±5%), the system automatically triggers an incremental learning process. The model training and inference module 300 uses the updated training set, including this new data, to fine-tune the original model and update the model parameters. The updated model will make more accurate predictions for subsequent new projects, achieving the system's self-evolution.

[0152] In summary, the method and system provided by this invention, through multi-source feature fusion, a hybrid neural network model with explicit physical constraints, a finished adhesive matching mechanism, and incremental learning, achieve efficient and accurate screening and performance prediction of interlayer adhesives for heterogeneous waterproof membranes, significantly improving the reliability and intelligence level of waterproof systems, and possessing extremely high industrial application value.

[0153] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0154] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent screening of waterproofing membrane interlayer adhesive based on multi-source feature fusion, characterized in that, include: Construct a multi-source feature database that includes multi-dimensional feature data of the first waterproof membrane, the second waterproof membrane, and candidate adhesives; A hybrid neural network model is established. A training dataset containing multi-source features and corresponding bonding performance labels is pre-constructed based on publicly available literature data. The hybrid neural network model is trained using the training dataset until the total loss function converges. The multidimensional feature data of the first and second waterproof membranes are input into the trained hybrid neural network model to screen for the optimal adhesive formulation and output a recommended solution. The actual performance data collected on-site is used to iterate and update the hybrid neural network model.

2. The intelligent screening method of the multi-source characteristic fusion waterproof membrane interlayer adhesive according to claim 1, characterized in that: The hybrid neural network model is a multi-task learning model, which includes: The input layer is used to receive multidimensional feature data of the first waterproof membrane, the second waterproof membrane, and candidate adhesives; Feature embedding layer, used to vectorize and embed discrete features; The physical information constraint layer adopts an explicit physical parameter branching structure to embed the physical mechanism of interface adhesion failure into the model training process. A multi-task learning layer is used to predict short-term adhesion performance indicators and long-term durability levels. The output layer is used to output predicted values ​​for short-term adhesion performance and long-term durability levels.

3. The intelligent screening method for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 2, characterized in that: The physical information constraint layer adopts an explicit physical parameter branching structure, including: The surface feature encoding branch of the roll material is used to map the surface free energy and polar components of the roll material into the interface wetting latent variable h1; The adhesive feature encoding branch is used to map the surface tension and glass transition temperature of the adhesive to the latent variable h2 of the adhesive. The physical formula calculation unit, with built-in Young's equation and interfacial fracture energy formula, is used to calculate the theoretical wetting coefficient based on h1 and h2. Theoretical interface fracture energy ; Constraint loss calculation unit, used to calculate model predictions. , Compared with theoretical value , The deviation between them is calculated and added as a regularization term to the total loss function.

4. The intelligent screening method for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 3, characterized in that: The multi-task learning layer includes: The first output task is a regression task used to predict short-term bonding performance indicators. The second output task is a classification task used to predict the long-term durability level. The short-term adhesion performance index includes at least one of peel strength, shear strength and initial tack after 24 hours; the long-term durability grade includes at least one of peel strength retention rate after heat aging, adhesion strength decay rate after water immersion, and interface failure mode after freeze-thaw cycles.

5. The intelligent screening method for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 1, characterized in that: The multidimensional feature data of the first and second waterproof membranes are input into a trained hybrid neural network model to screen for optimal adhesive formulations and output recommended solutions, including: The multidimensional feature data of the first and second waterproof membranes to be screened, as well as the preset construction environment conditions, are input into the trained hybrid neural network model. Using a Bayesian optimization algorithm, a global search is performed in the continuous feature space of candidate adhesives to maximize the comprehensive score of bonding performance output by the hybrid neural network model, and to iteratively find a better combination of virtual formulation parameters for the adhesive. The optimal virtual formula parameter combination obtained through optimization is compared with the feature similarity of the preset finished adhesive database. According to the Euclidean distance principle, the finished adhesive model with the closest formula characteristics that is already available on the market is matched. The Euclidean distance between the optimal virtual formula and the matched finished adhesive is calculated. If the Euclidean distance is greater than the preset deviation threshold, it is determined that the existing finished adhesive cannot meet the requirements, and a "custom development is recommended" prompt is output. A recommended formula is also output for the manufacturer's reference. Output the matched finished adhesive model, supplier information, and the corresponding optimal combination of construction process parameters.

6. The intelligent screening method for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 1, characterized in that: Also includes: The SHAP method is used to calculate the contribution of each input feature to the output prediction result, and the key factors affecting the adhesion reliability are ranked and the corresponding explanatory text is output.

7. The intelligent screening method for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 1, characterized in that: Collect actual performance data from the field and iterate and update the hybrid neural network model accordingly, including: Collect actual performance data after on-site bonding using the superior adhesive; The actual performance data is compared with the model predictions to calculate the prediction deviation; The prediction bias is used as a feedback signal to incrementally learn the hybrid neural network model and update the model parameters.

8. A multi-source feature fusion-based intelligent screening system for interlayer adhesives in waterproof membranes, characterized in that, include: The feature extraction module is used to obtain multi-dimensional feature data of the first waterproof membrane, the second waterproof membrane, and the candidate adhesives; A database module, connected to the feature extraction module, is used to store the multidimensional feature data; The model training and inference module is equipped with a pre-trained hybrid neural network model. The optimization recommendation module is used to execute optimization algorithms and output optimal adhesive and construction process parameters. The feedback iteration module is used to collect actual performance data on-site and trigger incremental learning of the model.

9. The intelligent screening system for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 8, characterized in that: The feature extraction module includes: A contact angle measuring instrument is used to measure the surface free energy of waterproof membranes; Fourier transform infrared spectroscopy is used to analyze the chemical composition of adhesives; A rotational viscometer is used to measure the viscosity of adhesives.

10. The intelligent screening system for interlayer adhesives of waterproof membranes based on multi-source feature fusion according to claim 8, characterized in that: It also includes a visualization module, which displays the optimal adhesive information, construction process parameters, and feature importance analysis charts generated using the SHAP method in a graphical interface.

Images