A low-odor multilayer composite tape voc release prediction method and system

Abstract

The present application relates to a kind of low odor multilayer composite adhesive tape VOC release prediction method, comprising the following steps: S1: the original data of multilayer composite adhesive tape is obtained and preprocessed, and multimodal input data set is obtained;S2: atomic-bond diagram representation is constructed, and input graph neural network is carried out feature extraction, obtains molecular level characterization vector, molecular level feature is fused with material macroscopic attribute, and unified material characteristic representation vector X1 is obtained;S3: based on material characteristic representation vector and environmental parameter as input, establish the neural network model satisfying diffusion equation constraint, obtain VOC diffusion field data X2 for preliminary prediction;S4: the diffusion field data X2 obtained is converted into time series input, and bidirectional recurrent neural network model is constructed to carry out time series feature extraction, and the feature vector X3 containing time dependence is obtained;S5 based on integrated prediction model, output target prediction value.The present application realizes the high-precision, generalizable prediction of the VOC release behavior of multilayer composite adhesive tape.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction, and in particular to a method and system for predicting VOC emissions from low-odor multilayer composite tapes. Background Technology

[0002] With the widespread application of multilayer composite tapes in electronics, automotive interiors, and medical protective equipment, the volatile organic compounds (VOCs) released during product use are gradually becoming a significant factor affecting environmental quality and human health. Especially in enclosed or semi-enclosed spaces, the adhesive layer, substrate layer, and release layer of the tape may continuously release trace amounts of organic matter, leading to odor problems and even safety hazards. Therefore, developing low-odor, multilayer composite tapes and accurately predicting their VOC release behavior has become a key technological requirement in materials design and quality control.

[0003] In existing technologies, research on VOC release largely relies on experimental detection methods or simplified models based on Fick's diffusion law. While experimental methods offer high precision, they suffer from drawbacks such as long processing times, high costs, and difficulty in covering multiple operating conditions. Traditional mechanistic models typically assume homogeneous materials and ideal boundary conditions, making it difficult to accurately describe interfacial resistance, non-uniform diffusion, and multi-component coupling release behavior between different materials in multilayer composite structures. This results in significant discrepancies between predicted results and actual conditions.

[0004] In recent years, machine learning methods have been gradually applied to the field of materials property prediction. However, existing methods mostly focus on statistical fitting and lack effective constraints on physical mechanisms, which can easily lead to insufficient generalization ability under small sample or new material conditions. At the same time, for complex processes such as VOC release, which have significant time dependence and spatial distribution characteristics, traditional models struggle to simultaneously take into account the combined effects of molecular structure information, material properties, environmental factors, and dynamic evolution processes. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a method for predicting VOC release from low-odor multilayer composite tapes, achieving high-precision and generalizable prediction of VOC release behavior of multilayer composite tapes.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting VOC emissions from low-odor multilayer composite tapes includes the following steps: S1: Obtain the raw data of the multi-layer composite tape and preprocess it to obtain the multimodal input dataset; S2: The VOC components and key adhesive components in the multimodal input dataset are transformed into molecular graph structure data, an atomic-bond graph representation is constructed, and the graph neural network is input for feature extraction to obtain a molecular-level representation vector; at the same time, the macroscopic properties of the material are numerically encoded, and the molecular-level features are fused with the macroscopic properties of the material to obtain a unified material feature representation vector X1; S3: Based on the material feature representation vector and environmental parameters as input, a neural network model that satisfies the diffusion equation constraint is established to obtain VOC diffusion field data X2 for preliminary prediction; S4: Convert the obtained diffusion field data X2 into a time series input, construct a bidirectional recurrent neural network model to extract time series features, generate time-weighted feature representations, and obtain feature vectors X3 containing time dependencies; S5: The obtained multi-source features X1, X2, and X3 are fused to construct a unified input tensor, which is then input into the integrated prediction model. Through a fully connected network and a nonlinear mapping relationship, the target prediction value is output.

[0007] Furthermore, the raw data of the multilayer composite tape is obtained and preprocessed to obtain a multimodal input dataset, as follows: The raw data of the multilayer composite tape includes material layer structure information, initial VOC content data, environmental condition data, and time dimension data. The raw data is uniformly coded and standardized to form a structured input vector, in which material information is converted into category embedding, and continuous variables are normalized to obtain a multimodal input dataset in a unified format.

[0008] Furthermore, the VOC components and key adhesive components in the multimodal input dataset are transformed into molecular graph structure data, an atom-bond graph representation is constructed, and then input into a graph neural network for feature extraction to obtain molecular-level representation vectors, as follows: First, molecular structure information of VOC components and key adhesive components is extracted from the multimodal input dataset and represented as a graph structure G=(V,E), where the set of nodes V represents atoms in the molecule and the set of edges E represents chemical bonds between atoms; for each atomic node v i Construct the initial feature vector h ∈V i (0), the features include atom type, electronegativity, atomic radius, hybrid orbital type, and partial charge; for each chemical bond e ij Construct edge feature vectors for E, including bond type, bond length, and bond energy; A graph neural network model is constructed to extract features from the molecular graph, and a graph attention network (GAT) is used for information propagation. The node update process is represented as follows: ; in, This represents the feature vector of node i in the l-th layer. Let W represent the set of neighboring nodes of node i. (l) Let be a trainable weight matrix, and σ(·) be a non-linear activation function; This represents the feature vector of node i in the (l+1)th layer; The attention coefficients in layer l; The attention coefficient Calculated as follows:

[0009] in, a For trainable attention weight vectors, Represents vector concatenation operation; LeakyReLU is the modified linear unit activation function with leakage; e ij Let be the feature vector of the edge between node i and node j; After completing the L-layer graph convolution, readout operations are performed on the features of all nodes to obtain the molecular-level representation vector h. G , represented as: ; Wherein, weight w i Obtained through self-attention mechanism: ; Where q is a trainable parameter vector; Let i be the feature representation of node i in the Lth layer; For multiple VOC components, a weighted polymerization method is used to obtain the overall molecular characteristic representation: ; in, β represents the graph embedding vector of the k-th VOC molecule. k This represents the mass fraction of the corresponding component; the final output is a molecular-level characterization vector. .

[0010] Furthermore, the macroscopic properties of the material are numerically encoded, and the molecular-level features are fused with the macroscopic properties to obtain a unified material feature representation vector, as follows: The macroscopic properties of the material include at least polarity parameters and glass transition temperature T. g The solubility parameter δ, density ρ, and free volume fraction were used to normalize the macroscopic properties of the continuous material. For discrete attributes, an embedding encoding method is used to map them into low-dimensional dense vectors. e c =Embedding(c); Where c represents the category label, e c For the corresponding embedding vector; e c and Concatenation forms the initial macroscopic feature vector ; ; Where W1 and W2 are weight matrices, and b1 and b2 are bias terms; f MLP For multilayer perceptron functions; Obtaining molecular-level characteristics With macroscopic features X macro Then, a gated fusion mechanism is used for fusion to obtain a unified material feature representation vector X1: ; Where g is the gate vector, and ⊙ represents element-wise multiplication.

[0011] Furthermore, based on the material feature representation vector and environmental parameters as input, a neural network model satisfying the diffusion equation constraints is established to obtain preliminary VOC diffusion field data for prediction, as follows: After obtaining the unified material feature representation vector X1 in step S2, the structural parameters of the multilayer composite tape are combined, including the number of layers N and the thickness of each layer. Based on sequence relation, temperature T, relative humidity RH, and airflow velocity v, a diffusion path model of VOCs in a multi-layered structure is constructed, defining spatial coordinates as x∈[0,L]. This indicates the overall thickness, with different layers corresponding to different spatial ranges. For the A diffusion governing equation is established for the layer material to describe the mass transfer process of VOCs in the layer:

[0012] in, Indicates the first VOC concentration distribution in the layer The effective diffusion coefficient of this layer is determined by both the material eigenvector and environmental parameters, and is specifically expressed as: D i′ =f θ (X 1,i′ ,T,RH); Where, f θ (·) represents the nonlinear mapping function learned by the neural network; X 1,i′ Indicates the first The unified material feature representation vector input in the layer; Furthermore, continuity constraints are introduced at the interlayer interfaces to describe the mass transfer coupling relationship between different materials. For the first... Layer and First Intersection The following conditions must be met: (1) Concentration continuity: ; in, The allocation coefficient; (2) Flux continuity: ; Introduce convective mass transfer boundary conditions at the outer surface boundary, at the release interface x=L:

[0013] Where h is the surface mass transfer coefficient, and C air VOC concentration in the air; Based on this, a physical information neural network PINN is constructed, using spatial coordinates x, time t, material feature vector X1, and environmental parameters (T, RH, v) as joint inputs: C θ (x,t)=NN θ (x,t,X1,T,RH,v); By using an automatic differential calculation network to output the time and space derivatives, the residuals of the diffusion equation are constructed: ; By training the neural network parameters θ by minimizing the loss function, a VOC concentration prediction model that satisfies physical constraints is obtained; after training, the model outputs a space-time continuous distribution function: C(x,t)=C θ (x,t); Finally, for discrete sampling points (x) k ,t k The values ​​are evaluated to construct the VOC diffusion field dataset: .

[0014] Furthermore, the loss function includes the physical equation residual loss L. phys Boundary condition loss L bc Interface continuity loss L int and initial condition loss L init The details are as follows: ; ; ; ; ; Where λ1, λ2, λ3, λ4 are weighting coefficients; By minimizing the aforementioned loss function L, the neural network parameters θ are trained to obtain a VOC concentration prediction model that satisfies the constraints of physical laws.

[0015] Furthermore, the obtained diffusion field data X2 is converted into a time series input, and a bidirectional recurrent neural network model is constructed to extract temporal features, generate a time-weighted feature representation, and obtain a feature vector X3 containing time dependencies, as follows: The VOC diffusion field data X2 obtained in step S3 is reconstructed along the time dimension to form an input format suitable for sequence modeling. Under a fixed spatial sampling strategy, the concentration distribution at each spatial location is collected at each time t to construct the time step feature vector S. The time-step feature vector S is input into the bidirectional recurrent neural network model Bi-LSTM to extract time-dependent features. The forward network models the historical evolution of VOC diffusion in forward time order, while the backward network models global context information in reverse time order. The two are concatenated to obtain the bidirectional hidden state representation of each time step. After obtaining the bidirectional hidden state representation, an attention mechanism is introduced to adaptively model the importance of each time step, calculate the attention score for each time step, and obtain the weight coefficient by normalizing the attention score; then, the features of all time steps are weighted and summed to obtain the global time representation vector X3.

[0016] Furthermore, the obtained multi-source features X1, X2, and X3 are fused to construct a unified input tensor, which is then input into the ensemble prediction model. Through a fully connected network and a nonlinear mapping relationship, the target prediction value is output, as follows: After obtaining the multi-source features X1, X2, and X3, the three types of features are embedded and encoded, mapped to a unified latent space, and corresponding representation vectors are obtained. These vectors are then used as tokens for different modalities to construct a fusion sequence input model. Through a multi-head self-attention mechanism, the model learns the correlation and interaction relationships between different modalities, enabling it to automatically capture the coupling features between material properties, diffusion behavior, and temporal dynamics, as shown below: ; Where Q, K, and V are the query, key, and value matrices, respectively; d is the dimension of the key vector; T represents the transpose; and softmax(·) is the normalization function. By stacking multiple attention Transformer structures, a high-order cross-modal fusion representation is obtained: Z = Transformer(X1, X2, X3); Based on the fusion representation Z, a diffusion probability model is introduced to construct a prediction model. By progressively adding noise to the target variable and learning the reverse denoising process, the complex nonlinear distribution is characterized. The forward diffusion process is defined as progressively adding Gaussian noise: ; Where y0 is the true target variable, y t This represents the noise state at step t. It follows a Gaussian normal distribution; β t Let be the noise scheduling parameters for step t; I is the identity matrix; Given y t-1 Under the condition y t The probability distribution; The reverse process is modeled parametrically using a neural network: ; in, The probability distribution modeled by the neural network parameters θ; The model training objective is to minimize the noise prediction error. ; in, For time step t, initial data y0 and noise Expectations; It is an L2 norm; After training the diffusion model, a backsampling process is used to generate the final prediction result, achieving joint modeling of multiple target variables. The final prediction result is expressed as follows: ; Among them, Y VOC Y represents the predicted total VOC emissions. comp For the concentration prediction vectors of each key component; Y time Y represents the predicted peak release time. odor This is the predicted value for the odor intensity index.

[0017] A VOC emission prediction system for low-odor multilayer composite tape includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the VOC emission prediction method for low-odor multilayer composite tape described above.

[0018] A computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method steps described above.

[0019] The present invention has the following beneficial effects: 1. This invention combines molecular-level diagram structure representation with macroscopic material properties to achieve cross-scale feature learning from microscopic mechanisms to macroscopic phenomena. It can deeply explore the physicochemical essence of VOC release process. At the same time, it introduces the physical constraint neural network PINN to model the diffusion process, combining data-driven methods with physical laws to ensure that the prediction results conform to the basic laws of diffusion and mass transfer, which significantly improves the interpretability and physical consistency of the model. The combination of bidirectional recurrent neural network and attention mechanism further enhances the ability to capture time-series dependencies, enabling the model to accurately identify key time nodes and dynamic evolution laws in the VOC release process. 2. This invention, through the deep fusion of multi-source heterogeneous information, can make full use of multi-dimensional data such as molecular structure information, material macroscopic properties, physical diffusion fields and temporal dynamic features to form a richer and more complete feature representation. The end-to-end learning of molecular structure by the graph neural network avoids the subjectivity and incompleteness of traditional manual feature engineering and can automatically discover molecular structure patterns related to VOC release behavior. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a method for predicting VOC emissions from low-odor multilayer composite tape is provided, comprising the following steps: S1: Obtain the raw data of the multi-layer composite tape and preprocess it to obtain the multimodal input dataset; S2: The VOC components and key adhesive components in the multimodal input dataset are transformed into molecular graph structure data, an atomic-bond graph representation is constructed, and the graph neural network is input for feature extraction to obtain a molecular-level representation vector; at the same time, the macroscopic properties of the material (polarity, glass transition temperature, solubility parameters, etc.) are numerically encoded, and the molecular-level features are fused with the macroscopic properties of the material to obtain a unified material feature representation vector X1; S3: Based on the material feature representation vector and environmental parameters as input, a neural network model that satisfies the diffusion equation constraint is established to obtain VOC diffusion field data X2 for preliminary prediction; S4: Convert the obtained diffusion field data X2 into a time series input, construct a bidirectional recurrent neural network (Bi-LSTM) model for temporal feature extraction, and automatically identify key time stages in the release process (such as the initial rapid release stage and the later slow release stage) by introducing an attention mechanism, generate time-weighted feature representations, and obtain feature vectors X3 containing time dependencies; S5: The obtained multi-source features X1, X2, and X3 are fused to construct a unified input tensor, which is then input into the integrated prediction model. Through a fully connected network and a nonlinear mapping relationship, the target prediction value is output, including: total VOC release, concentration of each key component, peak release time, and odor intensity index.

[0022] In this embodiment, the raw data of the multilayer composite tape is acquired and preprocessed to obtain a multimodal input dataset, as follows: The raw data of the multilayer composite tape includes material layer structure information (substrate type, adhesive type, release layer type and thickness of each layer), initial VOC content data (obtained by GC-MS), environmental condition data (temperature, humidity, airflow velocity), and time dimension data (release time series). The raw data is uniformly encoded and standardized to form a structured input vector, in which material information is converted into category embedding, and continuous variables are normalized to obtain a multimodal input dataset in a unified format.

[0023] In this embodiment, the VOC components and key adhesive components in the multimodal input dataset are transformed into molecular graph structure data, an atom-bond graph representation is constructed, and the data is input into a graph neural network for feature extraction to obtain molecular-level representation vectors, as follows: First, molecular structure information of VOC components and key adhesive components is extracted from the multimodal input dataset and represented as a graph structure G=(V,E), where the set of nodes V represents atoms in the molecule and the set of edges E represents chemical bonds between atoms; for each atomic node v i Construct the initial feature vector h ∈V i (0), the features include atom type, electronegativity, atomic radius, hybrid orbital type, and partial charge; for each chemical bond e ij Construct edge feature vector e ∈E ij This includes bond type (single bond, double bond, etc.), bond length, and bond energy; A graph neural network model is constructed to extract features from the molecular graph, and a graph attention network (GAT) is used for information propagation. The node update process is represented as follows: ; in, This represents the feature vector of node i in the l-th layer. Let W represent the set of neighboring nodes of node i. (l) Let be a trainable weight matrix, and σ(·) be a non-linear activation function; This represents the feature vector of node i in the (l+1)th layer; The attention coefficients in layer l; The attention coefficient Calculated as follows:

[0024] in, a For trainable attention weight vectors, Represents vector concatenation operation; LeakyReLU is the modified linear unit activation function with leakage; e ij Let be the feature vector of the edge between node i and node j; After completing the L-layer graph convolution, a readout operation is performed on the features of all nodes to obtain the molecular-level representation vector h. G , represented as: ; Wherein, weight w i Obtained through self-attention mechanism: ; Where q is a trainable parameter vector; Let i be the feature representation of node i in the Lth layer; For multiple VOC components, a weighted polymerization method is used to obtain the overall molecular characteristic representation: ; in, β represents the graph embedding vector of the k-th VOC molecule. k This represents the mass fraction of the corresponding component; the final output is a molecular-level characterization vector. .

[0025] In this embodiment, the macroscopic properties of the material are numerically encoded, and the molecular-level features are fused with the macroscopic properties of the material to obtain a unified material feature representation vector, as follows: The macroscopic properties of the material include at least polarity parameters and glass transition temperature T. g The solubility parameter δ, density ρ, and free volume fraction were used to normalize the macroscopic properties of the continuous material. For discrete attributes (such as material category), an embedding encoding method is used to map them into low-dimensional dense vectors. e c =Embedding(c); Where c represents the category label, e c For the corresponding embedding vector; e c and Concatenation forms the initial macroscopic feature vector ; ; Where W1 and W2 are weight matrices, and b1 and b2 are bias terms; f MLP For multilayer perceptron functions; Obtaining molecular-level characteristics With macroscopic features X macro Then, a gated fusion mechanism is used for fusion to obtain a unified material feature representation vector X1: ; Where g is the gate vector, and ⊙ represents element-wise multiplication.

[0026] In this embodiment, a neural network model satisfying the diffusion equation constraints is established based on the material feature representation vector and environmental parameters as input to obtain preliminary VOC diffusion field data, as detailed below: After obtaining the unified material feature representation vector X1 in step S2, the structural parameters of the multilayer composite tape are combined, including the number of layers N and the thickness of each layer. Based on sequence relation, temperature T, relative humidity RH, and airflow velocity v, a diffusion path model of VOCs in a multi-layered structure is constructed, defining spatial coordinates as x∈[0,L]. This indicates the overall thickness, with different layers corresponding to different spatial ranges. For the A diffusion governing equation is established for the layer material to describe the mass transfer process of VOCs in the layer:

[0027] in, Indicates the first VOC concentration distribution in the layer The effective diffusion coefficient of this layer is determined by both the material eigenvector and environmental parameters, and is specifically expressed as: D i′ =f θ (X 1,i′ ,T,RH); Where, f θ (·) represents the nonlinear mapping function learned by the neural network; X 1,i′ Indicates the first The unified material feature representation vector input in the layer; Furthermore, continuity constraints are introduced at the interlayer interfaces to describe the mass transfer coupling relationship between different materials. For the first... Layer and First Intersection The following conditions must be met: (1) Concentration continuity: ; in, The allocation coefficient; (2) Flux continuity: ; Introduce convective mass transfer boundary conditions at the outer surface boundary, at the release interface x=L:

[0028] Where h is the surface mass transfer coefficient, and C air VOC concentration in the air; Based on this, a physical information neural network PINN is constructed, using spatial coordinates x, time t, material feature vector X1, and environmental parameters (T, RH, v) as joint inputs: C θ (x,t)=NN θ (x,t,X1,T,RH,v); By using an automatic differential calculation network to output the time and space derivatives, the residuals of the diffusion equation are constructed: ; By training the neural network parameters θ by minimizing the loss function, a VOC concentration prediction model that satisfies physical constraints is obtained; after training, the model outputs a space-time continuous distribution function: C(x,t)=C θ (x,t); Finally, for discrete sampling points (x) k ,t k The values ​​are evaluated to construct the VOC diffusion field dataset: .

[0029] In this embodiment, the loss function includes the physical equation residual loss L. phys Boundary condition loss L bc Interface continuity loss L int and initial condition loss L init The details are as follows: ; ; ; ; ; Where λ1, λ2, λ3, λ4 are weighting coefficients; By minimizing the aforementioned loss function L, the neural network parameters θ are trained to obtain a VOC concentration prediction model that satisfies the constraints of physical laws.

[0030] In this embodiment, the obtained diffusion field data X2 is converted into a time series input, a bidirectional recurrent neural network model is constructed to extract temporal features, a time-weighted feature representation is generated, and a feature vector X3 containing time dependencies is obtained, as follows: The VOC diffusion field data X2 obtained in step S3 is reconstructed along the time dimension to form an input format suitable for sequence modeling. Under a fixed spatial sampling strategy, the concentration distribution at each spatial location is collected at each time t to construct the time step feature vector S. The time-step feature vector S is input into the bidirectional recurrent neural network model Bi-LSTM to extract time-dependent features. The forward network models the historical evolution of VOC diffusion in forward time order, while the backward network models global context information in reverse time order. The two are concatenated to obtain the bidirectional hidden state representation of each time step. After obtaining the bidirectional hidden state representation, an attention mechanism is introduced to adaptively model the importance of each time step. The attention score of each time step is calculated, and the weight coefficient is obtained by normalizing the attention score. Then, the features of all time steps are weighted and summed to obtain the global time representation vector X3, which integrates the information of key time nodes and can highlight the stage features that play a dominant role in VOC diffusion behavior, thus forming a feature representation X3 containing complete time dependencies, which can be used for subsequent prediction or multimodal fusion modeling.

[0031] In this embodiment, the obtained multi-source features X1, X2, and X3 are fused to construct a unified input tensor, which is then input into the ensemble prediction model. Through a fully connected network and a nonlinear mapping relationship, the target prediction value is output, as detailed below: After obtaining the multi-source features X1, X2, and X3, the three types of features are embedded and encoded, mapped to a unified latent space, and corresponding representation vectors are obtained. These vectors are then used as tokens for different modalities to construct a fusion sequence input model. Through a multi-head self-attention mechanism, the model learns the correlation and interaction relationships between different modalities, enabling it to automatically capture the coupling features between material properties, diffusion behavior, and temporal dynamics, as shown below: ; Where Q, K, and V are the query, key, and value matrices, respectively; d is the dimension of the key vector; T represents the transpose; and softmax(·) is the normalization function. By stacking multiple attention Transformer structures, a high-order cross-modal fusion representation is obtained: Z = Transformer(X1, X2, X3); Based on the fusion representation Z, a diffusion probability model is introduced to construct a prediction model. By progressively adding noise to the target variable and learning the reverse denoising process, the complex nonlinear distribution is characterized. The forward diffusion process is defined as progressively adding Gaussian noise: ; Where y0 is the true target variable (such as VOC emission), y t This represents the noise state at step t. It follows a Gaussian normal distribution; β t Let be the noise scheduling parameters for step t; I is the identity matrix; Given y t-1 Under the condition y t The probability distribution; The reverse process is modeled parametrically using a neural network: ; in, The probability distribution modeled by the neural network parameters θ; The model training objective is to minimize the noise prediction error. ; in, For time step t, initial data y0 and noise Expectations; It is an L2 norm; After training the diffusion model, a backsampling process is used to generate the final prediction result, achieving joint modeling of multiple target variables. The final prediction result is expressed as follows: ; Among them, Y VOC Y represents the predicted total VOC emissions. comp For the concentration prediction vectors of each key component; Y time Y represents the predicted peak release time. odor This is the predicted value for the odor intensity index.

[0032] Furthermore, the predicted mean and variance are calculated: These are used to provide point estimation results and uncertainty assessments, respectively.

[0033] A VOC emission prediction system for low-odor multilayer composite tape includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the VOC emission prediction method for low-odor multilayer composite tape described above.

[0034] A computer storage medium storing a plurality of instructions adapted for loading and execution of the above method steps by a processor.

[0035] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0036] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0037] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0038] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting VOC emissions from low-odor multilayer composite tapes, characterized in that, Includes the following steps: S1: Obtain the raw data of the multi-layer composite tape and preprocess it to obtain the multimodal input dataset; S2: The VOC components and key adhesive components in the multimodal input dataset are transformed into molecular graph structure data, an atomic-bond graph representation is constructed, and the graph neural network is input for feature extraction to obtain a molecular-level representation vector; at the same time, the macroscopic properties of the material are numerically encoded, and the molecular-level features are fused with the macroscopic properties of the material to obtain a unified material feature representation vector X1; S3: Based on the material feature representation vector and environmental parameters as input, a neural network model that satisfies the diffusion equation constraint is established to obtain VOC diffusion field data X2 for preliminary prediction; S4: Convert the obtained diffusion field data X2 into a time series input, construct a bidirectional recurrent neural network model to extract time series features, generate time-weighted feature representations, and obtain feature vectors X3 containing time dependencies; S5: The obtained multi-source features X1, X2, and X3 are fused to construct a unified input tensor, which is then input into the integrated prediction model. Through a fully connected network and a nonlinear mapping relationship, the target prediction value is output.

2. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 1, characterized in that, The process of acquiring and preprocessing the raw data of the multilayer composite tape to obtain a multimodal input dataset is as follows: The raw data of the multilayer composite tape includes material layer structure information, initial VOC content data, environmental condition data, and time dimension data. The raw data is then uniformly encoded and standardized to form a structured input vector. Material information is converted into category embeddings, and continuous variables are normalized to obtain a multimodal input dataset in a unified format.

3. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 1, characterized in that, The process involves converting VOC components and key adhesive components from the multimodal input dataset into molecular graph structure data, constructing an atom-bond graph representation, and inputting it into a graph neural network for feature extraction to obtain molecular-level representation vectors, as detailed below: First, molecular structure information of VOC components and key adhesive components is extracted from the multimodal input dataset and represented as a graph structure G=(V,E), where the set of nodes V represents atoms in the molecule and the set of edges E represents chemical bonds between atoms; for each atomic node v i Construct the initial feature vector h ∈V i (0), the features include atom type, electronegativity, atomic radius, hybrid orbital type, and partial charge; for each chemical bond e ij Construct edge feature vectors for E, including bond type, bond length, and bond energy; A graph neural network model is constructed to extract features from the molecular graph, and a graph attention network (GAT) is used for information propagation. The node update process is represented as follows: ； in, This represents the feature vector of node i in the l-th layer. Let W represent the set of neighboring nodes of node i. (l) Let be a trainable weight matrix, and σ(·) be a non-linear activation function; This represents the feature vector of node i in the (l+1)th layer; The attention coefficients in layer l; The attention coefficient Calculated as follows: ； in, a For trainable attention weight vectors, Represents vector concatenation operation; LeakyReLU is the modified linear unit activation function with leakage; e ij Let be the feature vector of the edge between node i and node j; After completing the L-layer graph convolution, readout operations are performed on the features of all nodes to obtain the molecular-level representation vector h. G , is represented as: ； Wherein, weight w i Obtained through self-attention mechanism: ； Where q is a trainable parameter vector; Let i be the feature representation of node i in the Lth layer; For multiple VOC components, a weighted polymerization method is used to obtain the overall molecular characteristic representation: ； in, β represents the graph embedding vector of the k-th VOC molecule. k This represents the mass fraction of the corresponding component; the final output is a molecular-level characterization vector. .

4. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 3, characterized in that, The process involves numerically encoding the macroscopic properties of materials, fusing molecular-level features with macroscopic properties to obtain a unified material feature representation vector, as detailed below: The macroscopic properties of the material include at least polarity parameters and glass transition temperature T. g The solubility parameter δ, density ρ, and free volume fraction were used to normalize the macroscopic properties of the continuous material. For discrete attributes, an embedding encoding method is used to map them into low-dimensional dense vectors. e c =Embedding(c); Where c represents the category label, e c For the corresponding embedding vector; e c and Concatenation forms the initial macroscopic feature vector ; ； Where W1 and W2 are weight matrices, and b1 and b2 are bias terms; f MLP For multilayer perceptron functions; Obtaining molecular-level characteristics With macroscopic features X macro Then, a gated fusion mechanism is used for fusion to obtain a unified material feature representation vector X1: ； Where g is the gate vector, and ⊙ represents element-wise multiplication.

5. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 1, characterized in that, The neural network model, which satisfies the diffusion equation constraints, is established based on material feature representation vectors and environmental parameters as inputs to obtain preliminary VOC diffusion field data for prediction, as detailed below: After obtaining the unified material feature representation vector X1 in step S2, the structural parameters of the multilayer composite tape are combined, including the number of layers N and the thickness of each layer. Based on sequence relation, temperature T, relative humidity RH, and airflow velocity v, a diffusion path model of VOCs in a multi-layered structure is constructed, defining spatial coordinates as x∈[0,L]. This indicates the overall thickness, with different layers corresponding to different spatial ranges. For the A diffusion governing equation is established for the layer material to describe the mass transfer process of VOCs in the layer: ； in, Indicates the first VOC concentration distribution in the layer The effective diffusion coefficient of this layer is determined by both the material eigenvector and environmental parameters, and is specifically expressed as: D i′ =f θ (X 1,i′ ,T,RH); Where, f θ (·) represents the nonlinear mapping function learned by the neural network; X 1,i′ Indicates the first The unified material feature representation vector input in the layer; Furthermore, continuity constraints are introduced at the interlayer interfaces to describe the mass transfer coupling relationship between different materials. For the first... Layer and First Intersection The following conditions must be met: (1) Concentration continuity: ； in, For allocation coefficients; (2) Flux continuity: ； Introduce convective mass transfer boundary conditions at the outer surface boundary, at the release interface x=L: ； Where h is the surface mass transfer coefficient, and C air VOC concentration in the air; Based on this, a physical information neural network PINN is constructed, using spatial coordinates x, time t, material feature vector X1, and environmental parameters (T, RH, v) as joint inputs: C θ (x,t)=NN θ (x,t,X1,T,RH,v)? By using an automatic differential calculation network to output the time and space derivatives, the residuals of the diffusion equation are constructed: ； By training the neural network parameters θ by minimizing the loss function, a VOC concentration prediction model that satisfies physical constraints is obtained; after training, the model outputs a space-time continuous distribution function: C(x,t)=C θ (x,t); Finally, for discrete sampling points (x) k ,t k The values ​​are evaluated to construct the VOC diffusion field dataset: 。 6. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 5, characterized in that, The loss function includes the physical equation residual loss L. phys Boundary condition loss L bc Interface continuity loss L int and initial condition loss L init The details are as follows: ； ； ； ； ； Where λ1, λ2, λ3, λ4 are weighting coefficients; By minimizing the aforementioned loss function L, the neural network parameters θ are trained to obtain a VOC concentration prediction model that satisfies the constraints of physical laws.

7. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 1, characterized in that, The obtained diffusion field data X2 is converted into a time series input, a bidirectional recurrent neural network model is constructed to extract time-series features, a time-weighted feature representation is generated, and a feature vector X3 containing time dependencies is obtained, as detailed below: The VOC diffusion field data X2 obtained in step S3 is reconstructed along the time dimension to form an input format suitable for sequence modeling. Under a fixed spatial sampling strategy, the concentration distribution at each spatial location is collected at each time t to construct the time step feature vector S. The time-step feature vector S is input into the bidirectional recurrent neural network model Bi-LSTM to extract time-dependent features. The forward network models the historical evolution of VOC diffusion in forward time order, while the backward network models global context information in reverse time order. The two are concatenated to obtain the bidirectional hidden state representation of each time step. After obtaining the bidirectional hidden state representation, an attention mechanism is introduced to adaptively model the importance of each time step, calculate the attention score for each time step, and obtain the weight coefficient by normalizing the attention score; then, the features of all time steps are weighted and summed to obtain the global time representation vector X3.

8. The method for predicting VOC emissions from low-odor multilayer composite tape according to claim 1, characterized in that, The obtained multi-source features X1, X2, and X3 are fused to construct a unified input tensor, which is then input into the ensemble prediction model. Through a fully connected network and a nonlinear mapping relationship, the target prediction value is output, as detailed below: After obtaining the multi-source features X1, X2, and X3, the three types of features are embedded and encoded, mapped to a unified latent space, and corresponding representation vectors are obtained. These vectors are then used as tokens for different modalities to construct a fusion sequence input model. Through a multi-head self-attention mechanism, the model learns the correlation and interaction relationships between different modalities, enabling it to automatically capture the coupling features between material properties, diffusion behavior, and temporal dynamics, as shown below: ； Where Q, K, and V are the query, key, and value matrices, respectively; d is the dimension of the key vector; T represents the transpose; and softmax(·) is the normalization function. By stacking multiple attention Transformer structures, a high-order cross-modal fusion representation is obtained: Z = Transformer(X1, X2, X3); Based on the fusion representation Z, a diffusion probability model is introduced to construct a prediction model. By progressively adding noise to the target variable and learning the reverse denoising process, the complex nonlinear distribution is characterized. The forward diffusion process is defined as progressively adding Gaussian noise: ； Where y0 is the true target variable, y t This represents the noise state at step t. It follows a Gaussian normal distribution; β t Let be the noise scheduling parameters for step t; I is the identity matrix; Given y t-1 Under the condition y t The probability distribution; The reverse process is modeled parametrically using a neural network: ； in, The probability distribution modeled by the neural network parameters θ; The model training objective is to minimize the noise prediction error. ； in, For time step t, initial data y0 and noise Expectations; It is an L2 norm; After training the diffusion model, a backsampling process is used to generate the final prediction result, achieving joint modeling of multiple target variables. The final prediction result is expressed as follows: ； Among them, Y VOC Y represents the predicted total VOC emissions. comp For the concentration prediction vectors of each key component; Y time Y represents the predicted peak release time. odor This is the predicted value for the odor intensity index.

9. A VOC emission prediction system for low-odor multilayer composite tape, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the VOC release prediction method for low-odor multilayer composite tape as claimed in any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions adapted for loading by a processor and executing the method steps as claimed in any one of claims 1-8.

Images