Reasoning method of thermal response of tunnel fire based on MGraphFomer

Abstract

The present invention discloses a tunnel fire thermal response reasoning method based on MGraphFomer, including: constructing a tunnel structure fire thermal response database; according to the tunnel classification design and structural damage judgment principle, with the help of a branch convolutional neural network, adaptively constructing the structural thermal response level judgment standard of different types of tunnels, and according to the structural thermal response level judgment standard, combining the structural thermal response levels of different types of tunnels to obtain a coarse-grained judgment model of the mechanical response of the structural unit; based on the heterogeneous graph processing method of the dual encoder and the coarse-grained judgment model of the mechanical response of the structural unit, constructing a fine-grained model of the multi-scale response of the structural unit to obtain a digital prediction model of the thermal response of the tunnel fire structure; using the digital prediction model of the thermal response of the tunnel fire structure to predict the thermal response of the tunnel fire structure, and realizing the reasoning of the thermal response of the tunnel fire structure. The thermal response prediction method of the present invention has high accuracy and stronger generalization ability.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular to an MGraphFomer-based tunnel fire thermal response reasoning method. Background Art

[0002] The cascaded graph analysis neural network MGraphFomer is a graph neural network (GNN)-based architecture that can effectively process graph-structured data. In the context of tunnel fire thermal response reasoning, the tunnel system can be abstracted as a graph structure, where nodes can represent different locations in the tunnel and edges can represent the connection relationship between these locations.

[0003] A tunnel refers to a spatial building that is not completely closed, has one or more sides that are relatively open or semi-open, and has limited air exchange with the external environment.

[0004] Thermal response refers to the phenomenon that when a structure is subjected to high temperature loads, the thermal expansion or thermal contraction effect caused by temperature changes causes the structure to produce stress and deformation.

[0005] Predicting the stress and deformation of tunnel structures caused by high temperature loads under fire loads is the basis for evaluating the functional loss of tunnel structures and is of great significance.

[0006] At present, the prediction of the thermal response of tunnel structure under fire load is described in the paper "Simulation Analysis of the Forces on Lining Structure of Shield Tunnels and Fire Resistance Measures during Fire[J]. " (XIAO Mingqing, YANG Wenqian, FENG Kun, et al. Simulation Analysis of the Forces on Lining Structure of Shield Tunnels and Fire Resistance Measures during Fire[J]. Modern Tunnelling Technology, 2023, 60(3): 199-207.), which discloses a numerical simulation method for tunnels under high temperature conditions during fire, including: establishing a tunnel finite element model according to design parameters such as tunnel lining structure, calculated loads, and thermal parameters; using a numerical simulation method to perform high temperature simulation on the tunnel finite element model; and using the results of numerical simulation to further evaluate the stress, deformation, and degradation of the shield tunnel segments under fire.

[0007] However, the above methods rely on numerical simulation and require many assumptions. Their evaluation results are limited to the fire load response under specific working conditions and their generalization ability is insufficient.

[0008] At present, the prediction of the thermal response of three-dimensional structures under fire loads is described in the paper "Research on Thermal Response and Ultimate Strength of Ship Cabin Fire" (Hao Junkai, Xue Hongxiang, Huang Jie, Liang Le, Yang Shangsheng, Miao Yiran, Yu Qiao, Wang Yuyang. Applied Science and Technology, 2022, 49 (03): 1-10.), which discloses a method for evaluating the thermal response of the structure of a ship cabin deck under fire, including: obtaining the ambient temperature under cabin fire through fire dynamics simulation software; mapping the time-varying ambient temperature load to the structural finite element model through a self-compiled program; and then considering a series of factors such as the heat conduction mode of the structure and boundary conditions, the thermal response and ultimate strength of the structure are studied in the finite element software.

[0009] However, the above methods rely on simple self-written programs, and the accuracy of the evaluation will decrease when the multi-scale structural responses vary greatly.

[0010] In summary, the problems existing in the prior art are that the prediction generalization ability and accuracy of the thermal response of tunnel structures under fire loads are not strong. Summary of the invention

[0011] The purpose of the present invention is to provide a tunnel fire thermal response reasoning method based on MGraphFomer with high accuracy and stronger generalization ability.

[0012] The technical solution to achieve the purpose of the present invention is:

[0013] In a first aspect, a tunnel fire thermal response reasoning method based on MGraphFomer is provided, comprising the following steps:

[0014] Construct a tunnel structure fire thermal response database, which includes tunnel fire structure thermal response simulation data, test data and historical data;

[0015] Based on the tunnel structure fire thermal response database, the cascade graph analysis neural network MGraphFomer is trained and optimized to build a digital prediction model for the thermal response of tunnel fire structures.

[0016] The digital intelligent prediction model of thermal response of tunnel fire structure is used to predict the thermal response of tunnel fire structure and realize the reasoning of thermal response of tunnel fire structure.

[0017] Preferably, the step of constructing a tunnel fire structure thermal response database includes:

[0018] Through GIS information technology and the building digital twin method, the digital twin of the tunnel is obtained;

[0019] According to the constructed fire events, numerical simulation methods are used to obtain the simulation data of thermal response of fire structures. The simulation data of thermal response of fire structures, test data and historical data are integrated to obtain the thermal response database of tunnel fire structures.

[0020] Preferably, after obtaining the tunnel digital twin, the method further includes:

[0021] Through tunnel design specifications and combined with a three-dimensional convolutional neural network, a tunnel digital twin model with the ability to automatically generate internal structures is constructed to obtain a detailed digital twin of the tunnel;

[0022] The steps to obtain the detailed digital twin of the tunnel include:

[0023] Clean the data contained in the tunnel design specifications and the tunnel digital twin, integrate data from different sources and formats, perform standardization and normalization operations, and screen out feature subsets that meet the preset requirements; the feature subsets include at least the geometric features, material features, safety design parameters, and functional zoning requirements of the tunnel;

[0024] Design a 3D convolutional neural network, whose input layer converts the received data into tensor data, whose hidden layer is composed of multiple 3D convolutional layers, pooling layers and fully connected layers, and whose output layer outputs the prediction information of the internal structure of the tunnel; the prediction information at least includes the lining thickness and the distribution of the supporting structure in the unmodeled area of ​​the tunnel;

[0025] The designed three-dimensional convolutional neural network is trained, and the trained three-dimensional convolutional neural network is used to complete the prediction of the internal structure of the tunnel and generate the internal structure model of the tunnel;

[0026] The internal structure model of the tunnel is integrated with the digital twin of the tunnel to obtain a detailed digital twin of the tunnel.

[0027] Preferably, the step of acquiring simulation data of thermal response of fire structure comprises:

[0028] By designing tunnels by classification, we can obtain the three-dimensional data of tunnel structures, design the fire source location and temperature load, and construct fire events;

[0029] Numerical simulation is performed on fire events to obtain the performance degradation parameters of structural materials, structural deformation coefficient, and burst depth under fire loads, and obtain simulation data on the thermal response of fire structures.

[0030] Preferably, the steps of constructing a digital intelligent prediction model for thermal response of tunnel fire structures include:

[0031] According to the principles of tunnel classification design and structural damage judgment, with the help of branched convolutional neural network, the structural thermal response level judgment criteria of different types of tunnels are adaptively constructed. According to the structural thermal response level judgment criteria, the structural thermal response levels of different types of tunnels are combined to obtain the coarse-grained judgment model of the mechanical response of the structural unit.

[0032] Based on the heterogeneous graph processing method of dual encoders and the coarse-grained determination model of the mechanical response of structural units, a fine-grained model of multi-scale response of structural units is constructed;

[0033] The tunnel fire structure thermal response database is used to train and optimize the multi-scale response fine-grained model of the structural unit, and a digital intelligent prediction model for the thermal response of the tunnel fire structure is obtained.

[0034] Preferably, the step of constructing a coarse-grained determination model of the mechanical response of a structural unit includes:

[0035] According to the tunnel classification design, different types of tunnels are given specific numbers;

[0036] According to the empirical relationship of thermal response of tunnel fire, the preliminary structural thermal response is obtained;

[0037] According to the preliminary structural thermal response and the principle of structural damage judgment, the structural damage level is obtained and quantitatively described to construct a hierarchical label tree to provide a label basis for subsequent branch convolutional neural network training; among them, structural damage includes four levels: slight damage, moderate damage, severe damage, and complete damage;

[0038] Embed the hierarchy of categories in the convolutional model and build a branched convolutional neural network with internal output branches. The branched convolutional neural network uses convolutional components as building blocks and contains alternating convolutional and pooling layers, where each convolutional layer is followed by a batch normalization layer.

[0039] Small sample data are selected for different types of tunnels, and a branch training strategy is used to train structural damage. During the training process, different branches generate predictions at the corresponding levels of the label tree, and the structural damage level of a certain type of tunnel is predicted based on the distribution contribution of the loss weight. Through the construction and training of branch convolutional neural networks, the structural thermal response level judgment criteria of different types of tunnels are adaptively constructed.

[0040] According to the structural thermal response level evaluation criteria for different types of tunnels obtained through training, the structural thermal response levels corresponding to different types of tunnels are combined, and then integrated into a coarse-grained judgment model for the mechanical response of structural units according to pre-set combination rules and logical relationships.

[0041] Preferably, the step of constructing a multi-scale response fine-grained model of structural units includes:

[0042] Constructing a parallel dual-encoder graph analysis neural network, the parallel dual-encoder graph analysis neural network includes multiple dual encoders in parallel, a graph analysis neural network processing layer, and a graph analysis neural network decoding layer; wherein the dual encoder includes a geometric feature encoder and a deformation feature encoder with different inputs but consistent structures, and a feedforward neural network, the input of the geometric feature encoder is structural geometric feature data, the input of the deformation feature encoder is structural deformation features, and the outputs of the geometric feature encoder and the deformation feature encoder are weighted and combined after being given weights by the feedforward neural network;

[0043] The number of dual encoders is consistent with the structural thermal response level, and each structural thermal response level corresponds to a dual encoder.

[0044] Preferably, it also includes combining the large deformation unit mask mechanism and embedding the mask error term L in the loss function of the structural unit multi-scale response fine-grained model. mask :

[0045] ,

[0046] In the formula, the unit Damage weight , , , , Respectively represent units The weight coefficients under the four structural thermal response levels of slight damage, moderate damage, severe damage, and complete damage are: , n is the number of units, For unit The percentage of damage, , For unit is the temperature, t is the time, , , For unit exist ， , The coordinates of the direction, Representation unit The heat generated during time t.

[0047] Preferably, it also includes adding a physical information error term based on the loss function of the multi-scale response fine-grained model of the structural unit :

[0048] ,

[0049] ,

[0050]

[0051] in, is the heat conduction loss term, is the structural response loss term, ρ is the material density, C is the specific heat capacity, For unit is the temperature, t is the time, , , for ， ， The thermal conductivity of the material in the direction, Q is the intensity of the heat source, represents the normal direction of the unit face, Indicates the direction of stress and strain, For unit exist , Stress in the direction; E is the elastic modulus of the material; For unit exist , The strain in the direction; α is the thermal expansion coefficient; For unit according to , The result after the Kroneck function is used to determine the direction.

[0052] In a second aspect, a computer-readable storage medium storing one or more programs is also provided. The one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform the method as described above.

[0053] According to a third aspect, an electronic device is also provided, comprising one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method as described above.

[0054] Compared with the prior art, the present invention has the following significant advantages:

[0055] 1. Stronger generalization ability: The present invention establishes a fire high temperature structural response database with random regional characteristics, random tunnel characteristics, random fire source positions and random fire source intensities, thereby ensuring the diversity of the database and the robustness of the neural network, thereby realizing the generalized prediction of the regional characteristics, tunnel structure characteristics and structural mechanical response under high temperature conditions of fire;

[0056] 2. Higher accuracy: The present invention establishes a physical information embedding graph of the structural thermal response under fire load to analyze a multi-order series neural network, which can not only ensure the model's learning ability for the physical problems of the structural thermal response under fire, but also ensure the model's learning ability for three-dimensional unstructured data, thereby achieving high convergence of the three-dimensional model and accurate prediction of the tunnel structure response under fire.

[0057] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the main flow chart of the tunnel fire thermal response reasoning method based on MGraphFomer of the present invention.

[0059] Figure 2 It is a schematic diagram of data transmission for coarse-grained determination of mechanical response of structural nodes.

[0060] Figure 3 It is a structural diagram of a dual encoder.

[0061] Figure 4 It is a graph analysis neural network.

[0062] Figure 5 It is a schematic diagram of the architecture of the processing layer and the decoding layer in a parallel dual encoder graph analysis neural network. DETAILED DESCRIPTION

[0063] like Figure 1 As shown, the tunnel fire thermal response reasoning method based on MGraphFomer of the present invention includes the following steps:

[0064] S10. Tunnel digital twin: Through GIS information technology, with the help of the multi-scale high-definition building digital twin method (MHD-BDT), the tunnel digital twin is obtained. And through the tunnel design specifications, combined with the three-dimensional convolutional neural network, a tunnel digital twin model with the ability to automatically generate internal structures is constructed to construct a fine digital twin of the tunnel.

[0065] The tunnel digital twin steps include:

[0066] S11. Construction of fine digital twin of tunnel: Through GIS information technology, the three-dimensional geographic information of the tunnel is obtained, and the digital twin of the tunnel is obtained with the help of the multi-scale high-definition building digital twin method (MHD-BDT).

[0067] Multi-scale High-Definition Building Digital Twin Method (MHD-BDT): Construct a digital twin of the tunnel at different scales, from the macroscopic tunnel location to the microscopic internal details of the tunnel. Use high-precision data acquisition technologies such as 3D laser scanning and drone photography to create a detailed three-dimensional model for the tunnel. Use sensors (such as stress, temperature, and vibration sensors) to collect data inside the tunnel in real time and integrate it into the digital twin model to achieve dynamic monitoring of the tunnel.

[0068] S12. Construction of tunnel digital twin model: Based on the existing tunnel design specifications and the three-dimensional model of the tunnel with internal information, data processing is performed, and a three-dimensional convolutional neural network is used to construct a tunnel digital twin model with the ability to automatically generate internal structures.

[0069] S121. Data processing: Clean the data contained in the tunnel design specifications and the three-dimensional tunnel model with internal information, integrate data from different sources and formats, perform standardization and normalization operations, extract key features such as tunnel geometry, material properties, safety design parameters, etc. from the original data, and screen out the most valuable feature subsets.

[0070] S1211. Data cleaning and integration: Clean the data from the tunnel design specification documents and the tunnel 3D model to remove noise, errors or redundant information. At the same time, integrate data from different formats and sources so that various types of data can correspond to each other and form a complete data system.

[0071] S1212. Data standardization and normalization: Since the data may have different dimensions and value ranges, in order to facilitate subsequent neural network processing, the data needs to be standardized or normalized, and the physical quantity data collected by various sensors should be mapped to a specific numerical range to ensure that the data is analyzed and calculated at the same scale, thereby improving the efficiency and accuracy of model training.

[0072] S1213. Feature extraction and selection: Extract key features related to the construction of the tunnel digital twin model from the original data. For the three-dimensional model data of the tunnel, the geometric features and material features of the tunnel may be extracted as important features; and features such as safety design parameters and functional zoning requirements may be extracted from the design specification data. The feature selection algorithm is used to select the feature subset that has the greatest impact on the model construction, reduce the data dimension, avoid overfitting, and highlight the driving effect of key information on the model.

[0073] S122. Construction of three-dimensional convolutional neural network: Design the network architecture, including the input layer, hidden layer, and output layer, reasonably initialize the parameters such as weights and biases in the network, select a suitable optimization algorithm, update the parameters based on the gradient information to minimize the loss function, adjust the hyperparameters such as the convolution kernel size, step size, and number of hidden layers through cross-validation, and determine evaluation indicators such as the mean square error to obtain a network model with excellent performance.

[0074] S1221. Network architecture design: For the construction of the tunnel digital twin model, the input layer receives the three-dimensional data of the tunnel and converts it into tensor data, which serves as the basis for the three-dimensional convolution layer to extract local spatial features, the pooling layer to reduce the dimension and reduce the computational complexity, and the fully connected layer to integrate the features and map the output dimension. The hidden layer processes the data in sequence. Finally, the output layer receives the data output by the hidden layer and the fully connected layer, and outputs the prediction information of the internal structure of the tunnel.

[0075] 1. Input layer: It is the entrance for the three-dimensional convolutional neural network to receive data. In the context of building a tunnel digital twin model, it is designed to adapt to the processing of three-dimensional data. After processing, the three-dimensional data is formed into tensor data, which can effectively store and represent data in three-dimensional space.

[0076] 2. Hidden layer: It is composed of multiple three-dimensional convolutional layers, pooling layers and fully connected layers. Three-dimensional convolutional layer: Its main function is to extract local spatial features from the data. In tunnel data processing, the convolution kernel scans the input data by sliding in three-dimensional space. Pooling layer: Its main function is to reduce the data dimension and reduce the amount of calculation. It can make the data more compact without losing the main feature information when processing a large amount of tunnel data. Fully connected layer: Integrate the feature information extracted by the previous layers. After the convolution layer extracts local features and the pooling layer reduces the dimension, the fully connected layer comprehensively processes these scattered feature information, linearly combines all feature information according to the weights and biases pre-learned by the network, and maps it to the final required output dimension.

[0077] 3. Output layer: It is the last layer of the network, which outputs the prediction information of the internal structure of the tunnel, predicts the lining thickness of the unmodeled area of ​​the tunnel, the distribution of the supporting structure, etc.

[0078] S1222. Parameter initialization and optimization algorithm selection: Reasonably initialize the weights and biases in the 3D convolutional neural network. Select a suitable optimization algorithm to adjust the network parameters, and update the network parameters according to the gradient information of the loss function to minimize the loss function so that the network output gradually approaches the true value.

[0079] S1223. Hyperparameter adjustment and model evaluation index determination: Adjust the hyperparameters of the three-dimensional convolutional neural network, including the convolution kernel size, step size, padding method, number of hidden layers, number of neurons, etc. Use the cross-validation method to divide the data set into training set, validation set and test set, train the network on the training set, evaluate the network performance and adjust the hyperparameters on the validation set, and finally perform a performance test on the finalized network model on the test set. Determine the appropriate model evaluation index, use the mean square error (MSE) to measure the average error between the predicted value and the true value, and the accuracy rate is used to judge the correctness ratio of the model prediction results in the classification task.

[0080] S123. Construction of digital twin model of tunnel: Use the trained 3D convolutional neural network to automatically generate the internal structure of the tunnel. Input the processed data into the network. The network predicts the internal structure information based on the learned features and generates a high-detail model. The automatically generated internal structure model will be integrated with the existing 3D model of the tunnel to improve the model so that it can dynamically reflect the actual status of the tunnel.

[0081] S1231. Structure generation based on three-dimensional convolutional neural network: The internal structure of the tunnel is automatically generated using a trained three-dimensional convolutional neural network. The processed tunnel data is input into the network, and the network outputs prediction information of the internal structure of the tunnel based on the learned features and patterns.

[0082] S1232. Model fusion and improvement: The automatically generated tunnel internal structure model is integrated with the existing tunnel 3D model to form a complete digital twin model of the tunnel. During the fusion process, different parts of the model are spliced, aligned and integrated to ensure the coherence and consistency of the model.

[0083] S13. Construction of fine digital twins of tunnels: With the help of three-dimensional convolutional neural networks, three-dimensional point cloud prediction is performed on all tunnel digital twins to automatically obtain the internal structure information of the tunnel. Through point cloud reconstruction, the structural unit surface information is obtained and the information is combined to provide a basis for subsequent numerical simulation and prediction.

[0084] S20. Construction of tunnel fire structure thermal response database: Based on the constructed fire events, a numerical simulation method is used to obtain fire structure thermal response simulation data, and the simulation data is integrated with test and historical fire data to obtain a tunnel fire structure thermal response database.

[0085] The steps of constructing the tunnel fire structure thermal response database include:

[0086] S21. Fire event construction: By classifying and designing tunnels, obtaining structural three-dimensional data, designing fire source locations and temperature loads, and constructing fire events.

[0087] Classification design refers to classifying tunnels according to different tunnel types, such as mountain tunnels and open-cut and backfill tunnels.

[0088] S22. Numerical simulation of thermal response of fire structure: numerical simulation is performed on the fire event to obtain the performance degradation parameters of the structural material, the structural deformation coefficient, and the burst depth under the fire load, and obtain simulation data of thermal response of the fire structure.

[0089] S23. Construction of fire structure thermal response database: Integrate the simulation data and test and historical fire data, and perform data processing to obtain multiple groups of fire load data and structural thermal response data corresponding to each group of fire load data, combine the fire load data and structural thermal response data according to fire cases, and obtain a tunnel fire structure thermal response database.

[0090] The fire structure thermal response database construction step includes:

[0091] S231. Processing of test and historical fire data: Collect test data of thermal response of fire load structure and historical data of thermal response of fire load structure, screen and reconstruct the collected data, and combine the data to obtain test and historical fire data.

[0092] The test and historical fire data processing steps include:

[0093] S2311. Data collection: Collect test data on thermal response of fire load structures by designing tests and reviewing literature; collect historical data on thermal response of fire load structures by reviewing literature, databases, and accident investigation reports.

[0094] S2312. Data screening: Classify the test and historical fire data according to data characteristics, delete the erroneous, incomplete or inconsistent data in the above classified data, and extract the explosion data that meets the requirements.

[0095] S2313, data reconstruction: interpolate the data deleted in the data screening to obtain the correct data;

[0096] S2314. Data combination: Classify and combine data based on fire cases, regional characteristics, tunnel structure characteristics, and fire load characteristics.

[0097] S232. Construction of fire structure response database: Classify and clean the test and historical fire data and numerical simulation fire data, extract the fire data that meets the requirements, perform data conversion and data combination, obtain multiple sets of fire load input data and structural mechanics response data, and establish a fire load structure response database.

[0098] The fire structure response database construction step includes:

[0099] S2321. Data classification: classify the collected regional characteristics, tunnel structure characteristics and fire load characteristics according to data features.

[0100] S2322. Data cleaning: Delete the erroneous, incomplete or inconsistent data in the above classified data to achieve data cleaning.

[0101] S2323, fire data extraction: extract the fire data that meets the requirements through S2331, data classification and S2332, data cleaning.

[0102] S2324, data conversion and combination: format the above cleaned data according to the data type, standardize the formatted data, and combine the standardized data according to the regional characteristics of each fire case, the tunnel structure characteristics, the fire characteristics, and the structural mechanical response of each fire case to achieve data conversion and combination.

[0103] Data formatting: Convert the raw data in the digital-physics-driven intelligent model into a format that can be recognized by the neural network, and collect data features from discrete grid nodes.

[0104] For example,

[0105] 1. Reshape the original data into a continuous feature vector form, and expand or compress the data in a certain order into a storage form recognizable by the neural network through interpolation or sampling methods.

[0106] 2. Use graph-structured data to directly collect features from discrete grid nodes to avoid loss of accuracy during data reshaping, and inherit the spatial characteristics of these grids into the graph according to the graph data structure.

[0107] Data standardization: According to different data features of the formatted data, a data standardization method is used to map the numerical ranges of different data features to similar intervals.

[0108] S2325. Construct a fire structure response database: store the above converted and combined data as a tunnel fire structure force response database.

[0109] S30. Construction of a digital prediction model for structural thermal response: With the help of a branched convolutional neural network, a coarse-grained model of the mechanical response of the structural unit is constructed. A dual-encoder parallel graph analysis neural network model (Multi-GraphFomer) is used, combined with a large deformation unit mask mechanism, to construct a fine-grained judgment model for the multi-scale response of the structural unit. The multi-scale response fine-grained judgment model for the structural unit is trained and optimized to obtain a digital prediction model for the structural thermal response under fire loads.

[0110] The step of constructing a digital prediction model of structural thermal response in S30 includes:

[0111] S31. Construction of a coarse-grained model of the mechanical response of structural units: Based on the tunnel classification design and with the help of a branched convolutional neural network, a structural thermal response grade evaluation standard for different types of tunnels is adaptively constructed. According to the structural thermal response grade evaluation standard, the structural thermal response grades of tunnels of different types are combined to determine the damage grade of the structural unit nodes. The data are combined to obtain a coarse-grained model of the mechanical response of the structural units.

[0112] The step of constructing the coarse-grained determination model of the mechanical response of the structural unit comprises:

[0113] S311, branch convolutional neural network construction: embed the hierarchy of categories in the convolutional model to build a network with internal output branches. The network uses convolutional components as building blocks and contains a 5-layer structure. There are 3 convolutional layers, and each convolutional layer is followed by a batch normalization layer to accelerate convergence and stabilize the model; set 2 pooling layers, both using maximum pooling to reduce data dimensions and extract key features, and use the ReLU function as the activation function to enhance the nonlinear expression of the network.

[0114] S312, branch convolutional neural network training: According to the tunnel classification design, different types of tunnels are assigned specific labels, where mountain tunnels correspond to label 0 and open-cut backfill tunnels correspond to label 1, so as to distinguish different tunnel categories for subsequent targeted processing. Small sample data is selected from the database according to the tunnel type, for example, input data is selected for mountain tunnels with label 0 and open-cut backfill tunnels with label 1, for example: Where l represents the node number, represents the tunnel grid node position, ( ) is the fire source location, ten is the fire source intensity, T is the fire temperature, time is the fire time, is the displacement direction, is the displacement size, is the stress direction, is the stress magnitude, is the strain magnitude, The branch training strategy is used to train the damage level of tunnels under different tunnel types. During the training process, different branches of the branch convolutional neural network generate predictions at the corresponding levels of the label tree, and predict the damage level of the tunnel according to the distribution contribution of the loss weight, thereby realizing the coarse-grained judgment of the mechanical response of the structural node. The data transmission process is as follows: Figure 2 As shown in the figure. Through the construction and training of this branched convolutional neural network, the structural thermal response level evaluation criteria for different types of tunnels are adaptively constructed.

[0115] S313, constructing a coarse-grained model of mechanical response of structural units: according to the structural thermal response level evaluation criteria of tunnel structures under different types obtained by the above training, the structural thermal response levels corresponding to different types of tunnels are combined, and the damage levels of structural unit nodes under different conditions are determined according to the pre-set combination rules and logical relationships, and the obtained damage levels are combined with the input data to form , level is the damage level.

[0116] The damage level is divided into four levels: in case of slight damage, the deformation of the structural unit is very small, such as the deformation of the metal structure is at the millimeter level or even smaller, and the stress is in the lower proportion range of the allowable stress of the material; in case of moderate damage, the deformation of the metal structure is several millimeters to more than ten millimeters, and the concrete structure has cracks of 0.5 mm to 2 mm or even larger, and the stress is at 30% - 60% of the allowable stress of the material; in case of severe damage, the deformation of the metal structure can reach tens of millimeters or even larger, and the concrete structure has large-scale cracks and fragments, and the stress is close to or exceeds the allowable stress of the material. For high-strength materials, the stress can reach more than 80% of the yield strength or even completely yield; in case of complete damage, the structure completely loses its original shape, suffers severe distortion, fracture, and collapse, the deformation cannot be measured conventionally, the stress far exceeds the ultimate strength of the material, and the structure completely fails.

[0117] S32. Construction of a fine-grained model of multi-scale responses of structural units: Based on the geometric encoder and the deformation encoder, a parallel dual-encoder graph analysis neural network model is constructed. Combined with the large deformation unit mask mechanism and the coarse-grained model of the mechanical response of the structural unit, a fine-grained model of multi-scale responses of the structural unit is constructed.

[0118] S321, such as Figure 3 The dual encoder construction shown is as follows: the original data is first divided into geometric feature data (derived from unit node grid information) and deformation feature data (derived from damage level) according to the source. For example: Among them, (x, y, z) is the geometric feature data, ( ) is the deformation feature data. These two types of data flow into the geometric feature encoder and the deformation feature encoder respectively. The two encoders have the same structure, but each focuses on processing different types of data. The geometric feature data enters the encoder branch that specializes in processing geometric information, and the deformation feature data enters the corresponding branch. Inside each encoder, the data undergoes a series of encoding operations to preliminarily extract feature information. With the help of a feedforward neural network, dynamic weights are assigned to the outputs of the two encoders according to pre-set rules, and they are combined to form an encoded data output that integrates geometric and deformation feature information. It serves as the final input of the graph analysis neural network encoding layer and is embedded in the original graph analysis neural network model. According to the different damage level information obtained in real time, the way and parameter configuration of the deformation feature data flowing into the deformation feature encoder are dynamically adjusted to achieve classification processing and preliminary integration of data at the encoder entrance.

[0119] S322, dual encoders in parallel: With the help of the unit response mode coarse-grained model, based on the structural unit information under different coarse-grained classifications, each damage level corresponds to a dual encoder, and multiple dual encoders are connected in parallel. The network architecture is as follows Figure 5 After the data is dynamically weighted by the feedforward neural network, it is processed according to the data processing flow of the parallel dual encoder graph analysis neural network. In this process, the dual encoder processes the data according to the damage level under the current coarse-grained classification, and dynamically adjusts the weight distribution of each data in the corresponding encoder. Specifically, for the input data , when the coarse-grained judgment result is mild damage (level is mild damage), the data will be classified as the corresponding mild damage and flow into the mild damage dual encoder. (x, y, z) is the geometric feature data, and its weight in the geometric feature encoder is relatively large. ( ) is the deformation feature data, which has a relatively large weight in the deformation feature encoder. The processed output will be used as the input of the graph analysis neural network to provide a basis for subsequent data analysis.

[0120] S323, data processing iteration of the graph analysis neural network processing layer: The data output by the parallel dual encoder flows into the graph analysis neural network processing layer. In this layer, the graph attention network and the transformer neural network are used to process information alternately. The graph attention network dynamically adjusts the attention weights between nodes according to the connection relationship between nodes and the data characteristics, focusing on the transmission and processing of key information. Subsequently, the transformer neural network further mines the potential associations and feature information in the data based on the results processed by the graph attention network. Through multiple such alternating iterative processes, the data is continuously updated and optimized, and the last layer of graph data is gradually obtained. In this process, the data circulates in the processing layer, and each iteration improves the quality and representativeness of the data until the iteration termination condition set by the processing layer is reached, and the processed graph data is output to the decoding layer.

[0121] S324, data restoration output of the graph analysis neural network decoding layer: The graph data output by the processing layer flows into the graph analysis neural network decoding layer. The decoding layer restores the graph data by combining the transformer neural network and the feedforward neural network. The transformer neural network gradually restores the original semantic information of the data based on the feature relationship learned in the previous encoding and processing process. The feedforward neural network adjusts and optimizes the restoration process, and finally converts the data into a prediction result that can be directly interpreted, completing the data processing flow of the entire graph analysis neural network and outputting the multi-scale response prediction information of the structural unit. Its network architecture and data flow path are shown as follows: Figure 4 shown.

[0122] S325. Construct a large deformation unit mask mechanism: Generate a mask matrix with the damage level obtained by the coarse-grained model of the unit response pattern to construct a large deformation unit mask mechanism.

[0123] Mask matrix: For example, suppose we have n cells and define A mask matrix M, whose unit positions are in rows 1-n and columns 1-4 of the matrix, respectively, represents slight damage, moderate damage, severe damage, and complete damage from left to right. , , , Respectively represent the unit The weight coefficients under slight damage, moderate damage, severe damage, and complete damage are: . Each element , , , Can be 0 or a positive real number, and when not 0, it satisfies , is the damage percentage of n units, which means 25% for slight damage, 50% for moderate damage, 75% for severe damage, and 100% for complete damage. It is a proportional value used to measure the degree of damage caused by fire to the tunnel, presented in percentage form. Yes Unit The damage weight calculated by the mask matrix is ​​calculated by Can amplify the effects of high levels of damage.

[0124] ,

[0125] ,

[0126] S326. Construct a fine-grained judgment model for the mechanical response of structural units: Combining the collaborative operation and data processing results of the above-mentioned parallel dual encoder graph analysis neural network, feedforward neural network and large deformation unit mask mechanism, we finally successfully construct a fine-grained model of the multi-scale response of structural units.

[0127] Physical information embedding: The loss term of the neural network consists of data loss term, physical loss term and mask error term.

[0128] ,

[0129] L PDE is the physical information error term, L data is the data error term, that is, the error between the model prediction value and the actual observed data, obtained using the MSE method, L mask is the mask error term, which weights the physical information error term by the mask matrix to reduce the influence of highly deformed units. PDE ,λ data ,λ mask are the weights of the three errors, and the sum of the three weights is 1.

[0130] ,

[0131] The physical loss term is composed of the heat conduction loss term Loss T and structural response loss term Loss u The heat conduction loss term comes from the heat conduction equation (Fourier's law), and the structural response loss term comes from the structural response equation.

[0132] ,

[0133] ,

[0134] in, is the heat conduction loss term, is the structural response loss term, ρ is the material density, and C is the specific heat capacity is the temperature predicted by the model, t is the time, , , for , , The thermal conductivity of the material in the direction, Q is the intensity of the heat source, represents the normal direction of the unit face, Indicates the direction of stress and strain, For unit exist , Stress in the direction; E is the elastic modulus of the material; For unit exist , The strain in the direction; α is the thermal expansion coefficient; For unit according to , The Kroneck function is used to determine the direction.

[0135] Mask error term L mask It can be further defined as:

[0136] ,

[0137] Where n is the total number of unit nodes, , , is the direction of fixed temperature transfer in three-dimensional space, is the node number, is the thermal diffusivity, Representation unit The heat generated during time t.

[0138] In the formula, is the heat source term, which represents the amount of heat generated per unit volume per unit time.

[0139] S33. Construction of a digital prediction model for thermal response: Using the data in the tunnel fire structure response database, combined with the Adam algorithm and a supervised method, the multi-scale response fine-grained model of the structural unit is trained, and the multi-scale response fine-grained model of the structural unit is optimized through hyperparameter tuning to obtain a digital prediction model for the thermal response of the structure under fire load.

[0140] S331, model training: using the data in the tunnel fire structure response database to input into the structural unit multi-scale response fine-grained model, using the Adam algorithm to update the model parameters, and using a supervised approach to train the model.

[0141] The model parameters are randomly generated at the beginning of training. The back propagation algorithm is needed to backpropagate the value of the loss function and calculate the gradient of the model parameters relative to the loss value to update the model parameters. The selected back propagation algorithm is the Adam algorithm, which combines the characteristics of gradient descent and momentum optimization and has good performance and convergence speed.

[0142] S332. Model construction: By tuning the hyperparameters of the fine-grained model of multi-scale response of structural units and optimizing the model, a digital prediction model of the thermal response of the structure under fire loads is obtained.

[0143] When performing hyperparameter tuning, you need to pay attention to a series of key hyperparameters, including the number of iterations, batch size, learning rate, and adjustment of the model structure. The key points of adjustment are as follows:

[0144] 1. The number of iterations is between 50 and 300 epochs;

[0145] 2. The batch size can be adjusted between 2 and 128;

[0146] 3. The learning rate is usually adjusted between 0.0001 and 0.01;

[0147] 4. Adjustment of the model structure is also a key step in optimization. According to the gradient calculation formula, the design of the structure affects the gradient update direction. Therefore, the model structure needs to be modified according to the model prediction results, including adjusting the number of layers in the model, the number of units in each layer, and selecting the appropriate graph network layer configuration.

[0148] S40. Prediction of thermal response of tunnel fire structure: A fire test set is constructed through the digital twin of the real area tunnel and environment, and the fire test set data is used as the input of the digital intelligent prediction model of the thermal response of the tunnel structure under the fire load. The output is the thermal response of the tunnel structure under the required fire load.

[0149] The tunnel fire structure thermal response prediction step comprises:

[0150] S41. Construction of fire test set: Digital twin of real fire scene tunnel and environment, obtain fire data, and through data processing, combine to obtain real fire events and construct a fire test set.

[0151] The step of constructing a test set in step S41 includes:

[0152] S411. Acquisition of tunnel digital twin: With the help of geographic information data sources and the GIS digital twin technology platform, a three-dimensional digital twin of the on-site tunnel is constructed.

[0153] S412. Fire event combination: Collect environmental characteristics, tunnel structure characteristics and fire characteristics through the GIS information system to obtain fire data, classify, clean, convert and combine the fire data according to data characteristics, and combine the data with the three-dimensional digital twin of the on-site tunnel to obtain the real fire event.

[0154] The environmental characteristics include: area size, tunnel type, and tunnel distribution; the tunnel structure characteristics include: the geometry of each tunnel and the material type of each tunnel; the fire characteristics include: the time of fire ignition, duration, and fire center.

[0155] The fire event combination step comprises:

[0156] S4121. Data classification: classify the collected environmental characteristics, tunnel structure characteristics and fire characteristics according to data features.

[0157] S4122. Data cleaning: Delete the erroneous, incomplete or inconsistent data in the above classified data to achieve data cleaning.

[0158] S4123. Acquisition of preliminary fire structural thermal response: Based on the existing empirical relationship for thermal response to tunnel fires, obtain preliminary judgment of the structural thermal response as input to the coarse-grained judgment model of the mechanical response of the structural unit.

[0159] S4124. Data conversion and combination: Format the above cleaned data according to the data type, standardize the formatted data, and combine the standardized data according to the environmental characteristics of each fire case, the tunnel structure characteristics, the fire characteristics, and the structural mechanical response of each fire case, and combine with the three-dimensional digital twin of the on-site tunnel to obtain the real fire event.

[0160] S413, constructing a fire test set: combining multiple real fire events to obtain a fire test set.

[0161] S42. Structural thermal response prediction: The fire test set data is used as the input of the numerical prediction model of the structural thermal response under fire load to obtain the precise thermal response of the tunnel structure under fire load.

[0162] Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes the above-mentioned tunnel fire thermal response reasoning method.

[0163] Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-mentioned tunnel fire thermal response reasoning method.

[0164] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented 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.

[0165] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0166] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0167] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0168] The above embodiments are only used to illustrate the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Any modification made on the basis of the technical solution in accordance with the technical idea proposed by the present invention falls within the protection scope of the present invention.

Claims

1. A tunnel fire thermal response reasoning method based on MGraphFomer, characterized in that: The steps include: Construct a tunnel structure fire thermal response database, which includes tunnel structure fire thermal response simulation data, test data and historical data; Based on the tunnel structure fire thermal response database, the cascade graph analysis neural network MGraphFomer is trained and optimized to build a digital prediction model for the thermal response of tunnel fire structures. The tunnel fire structure thermal response digital prediction model is used to predict the tunnel fire structure thermal response and realize the reasoning of the tunnel fire structure thermal response; The steps for constructing a digital prediction model for thermal response of tunnel fire structures include: According to the principles of tunnel classification design and structural damage judgment, with the help of branched convolutional neural network, the structural thermal response level judgment criteria of different types of tunnels are adaptively constructed. According to the structural thermal response level judgment criteria, the structural thermal response levels of different types of tunnels are combined to obtain the coarse-grained judgment model of the mechanical response of the structural unit. Based on the heterogeneous graph processing method of dual encoders and the coarse-grained determination model of the mechanical response of structural units, a fine-grained model of multi-scale response of structural units is constructed; The tunnel fire structure thermal response database is used to train and optimize the multi-scale response fine-grained model of the structural unit, and a digital intelligent prediction model of the tunnel fire structure thermal response is obtained; The steps for constructing the coarse-grained determination model of the mechanical response of structural units include: According to the tunnel classification design, different types of tunnels are given specific numbers; According to the empirical relationship of thermal response of tunnel fire, the preliminary structural thermal response is obtained; According to the preliminary structural thermal response and the principle of structural damage judgment, the structural damage level is obtained and quantitatively described to construct a hierarchical label tree to provide a label basis for subsequent branch convolutional neural network training; among them, structural damage includes four levels: slight damage, moderate damage, severe damage, and complete damage; Embed the hierarchy of categories in the convolutional model and build a branched convolutional neural network with internal output branches. The branched convolutional neural network uses convolutional components as building blocks and contains alternating convolutional and pooling layers, where each convolutional layer is followed by a batch normalization layer. Small sample data are selected for different types of tunnels, and a branch training strategy is used to train structural damage. During the training process, different branches generate predictions at the corresponding levels of the label tree, and the structural damage level of a certain type of tunnel is predicted based on the distribution contribution of the loss weight. Through the construction and training of branch convolutional neural networks, the structural thermal response level judgment criteria of different types of tunnels are adaptively constructed. According to the structural thermal response level evaluation criteria of different types of tunnels obtained through training, the structural thermal response levels corresponding to different types of tunnels are combined, and a coarse-grained judgment model of the mechanical response of the structural unit is formed according to the pre-set combination rules and logical relationships; The steps for constructing a fine-grained model of multi-scale response of structural units include: Constructing a parallel dual-encoder graph analysis neural network, the parallel dual-encoder graph analysis neural network includes multiple dual encoders in parallel, a graph analysis neural network processing layer, and a graph analysis neural network decoding layer; wherein the dual encoder includes a geometric feature encoder and a deformation feature encoder with different inputs but consistent structures, and a feedforward neural network, the input of the geometric feature encoder is structural geometric feature data, the input of the deformation feature encoder is structural deformation features, and the outputs of the geometric feature encoder and the deformation feature encoder are weighted and combined after being given weights by the feedforward neural network; The number of dual encoders is consistent with the structural thermal response level, and each structural thermal response level corresponds to a dual encoder.

2. The method according to claim 1, characterized in that: The steps for constructing the tunnel fire structure thermal response database include: Through GIS information technology and the building digital twin method, the digital twin of the tunnel is obtained; According to the constructed fire events, numerical simulation methods are used to obtain the simulation data of thermal response of fire structures. The simulation data of thermal response of fire structures, test data and historical data are integrated to obtain the thermal response database of tunnel fire structures.

3. The method according to claim 2, characterized in that Obtaining the tunnel digital twin also includes: Through tunnel design specifications and combined with a three-dimensional convolutional neural network, a tunnel digital twin model with the ability to automatically generate internal structures is constructed to obtain a detailed digital twin of the tunnel; The steps to obtain the detailed digital twin of the tunnel include: Clean the data contained in the tunnel design specifications and the tunnel digital twin, integrate data from different sources and formats, perform standardization and normalization operations, and screen out feature subsets that meet the preset requirements; the feature subsets include at least the geometric features, material features, safety design parameters, and functional zoning requirements of the tunnel; Design a 3D convolutional neural network, whose input layer converts the received data into tensor data, whose hidden layer is composed of multiple 3D convolutional layers, pooling layers and fully connected layers, and whose output layer outputs the prediction information of the internal structure of the tunnel; the prediction information at least includes the lining thickness and the distribution of the supporting structure in the unmodeled area of ​​the tunnel; The designed three-dimensional convolutional neural network is trained, and the trained three-dimensional convolutional neural network is used to complete the prediction of the internal structure of the tunnel and generate the internal structure model of the tunnel; The internal structure model of the tunnel is integrated with the digital twin of the tunnel to obtain a detailed digital twin of the tunnel.

4. The method according to claim 2, characterized in that The steps for obtaining simulation data of thermal response of fire structure include: By designing tunnels by classification, we can obtain the three-dimensional data of tunnel structures, design the fire source location and temperature load, and construct fire events; Numerical simulation is performed on fire events to obtain the performance degradation parameters of structural materials, structural deformation coefficient, and burst depth under fire loads, and obtain simulation data on the thermal response of fire structures.

5. The method according to claim 1, characterized in that: It also includes combining the large deformation unit mask mechanism and embedding the mask error term L in the loss function of the structural unit multi-scale response fine-grained model. mask : ， In the formula, the unit Damage weight , , , , Respectively represent units The weight coefficients under the four structural thermal response levels of slight damage, moderate damage, severe damage, and complete damage are: , n is the number of units, For unit The percentage of damage, , For unit is the temperature, t is the time, , , For unit exist , , The coordinates of the direction, Representation unit The heat generated during time t.

6. The method according to claim 1, characterized in that It also includes adding physical information error terms to the loss function of the fine-grained model of the multi-scale response of the structural unit : ， ， ， in, is the heat conduction loss term, is the structural response loss term, ρ is the material density, C is the specific heat capacity, For unit is the temperature, t is the time, , , for , , The thermal conductivity of the material in the direction, Q is the intensity of the heat source, represents the normal direction of the unit face, Indicates the direction of stress and strain, For unit exist , Stress in the direction; E is the elastic modulus of the material; For unit exist , The strain in the direction; α is the thermal expansion coefficient; For unit according to , The result after the Kroneck function is used to determine the direction.

7. A computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, characterized in that: When the instructions are executed by a computing device, the computing device is caused to perform the method according to any one of claims 1 to 6.

Images