A Multi-Scale Subway Disaster Risk Level Prediction Method Based on Multi-Source Data Fusion

Abstract

This invention presents a multi-scale subway disaster risk level prediction method based on multi-source data fusion. It obtains data collected during the construction, operation, and maintenance of subway projects, stores it in a database, and assigns it specific timestamps and spatial coordinates. After preprocessing, the data is divided into three categories: text, images, and sensor data, each containing multiple modalities. A multi-modal data analysis and fusion unit is constructed, including a multi-head cross-modal attention module and three parallel feature extraction branches, generating the final multi-modal fusion feature vector. Using the operating conditions as input and the physical variables as inputs, and the predicted physical quantities and risk level judgment results as outputs, a subway project risk level prediction model based on a physical information neural network is constructed. This method effectively integrates multi-source data related to subway projects, contributing to improved accuracy and practicality in disaster risk level prediction.

Description

Technical Field

[0001] This invention relates to the field of underground engineering technology, specifically a multi-scale subway disaster risk level prediction method based on multi-source data fusion. Background Technology

[0002] Subways are linearly distributed and have outstanding advantages such as large capacity, high efficiency, and environmental friendliness. Their stations have powerful integrated functions and can be combined with surrounding facilities to form multifunctional underground space complexes. They occupy an extremely important position in the urban transportation and underground space utilization system and play a supporting and leading role in urban development.

[0003] In addition, due to the large scale of subway projects, long construction periods, and complex construction environments, physical and mechanical factors such as soil and rock properties, building materials, and construction equipment will directly affect the safety of subway projects during construction, operation, and maintenance.

[0004] However, against the backdrop of extreme climate change, the normalization of influenza, and frequent population movement, the potential impacts of non-physical factors such as urban development, geological environment, and social operation on subway projects are becoming increasingly prominent. However, current research lacks necessary monitoring and evaluation methods for external environmental factors, including both natural and social factors. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a multi-scale subway disaster risk level prediction method based on multi-source data fusion. This prediction method uses artificial intelligence technology to deeply integrate dynamic multi-source data from macro and micro levels, establishes a universally applicable subway engineering risk level prediction model based on a physical information neural network, and assesses the long-term dynamic evolution of subway projects to comprehensively improve the safety of subway projects.

[0006] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows: In a first aspect, the present invention provides a multi-scale subway disaster risk level prediction method based on multi-source data fusion, the method comprising the following: Data collected during the construction, operation, and maintenance of the subway project is obtained, stored in a database, and assigned a specific timestamp. and spatial coordinates ; After preprocessing, the data is divided into three categories: text, image, and sensor data. Each category contains multiple modalities, totaling [amount missing]. One mode; Constructing a multimodal data analysis and fusion unit: including a multi-head cross-modal attention module and three parallel feature extraction branches: The first feature extraction branch is a knowledge processing branch, which uses a pre-trained language model based on the BERT architecture. It takes text data as input and outputs text data features of different modalities. The second feature extraction branch is a visual feature extraction branch, which consists of a modality discriminator and a parallel module composed of a convolutional neural network and a visual transformer. It takes image data as input and determines whether to use a convolutional neural network or a visual transformer for feature extraction based on the modality discriminator, outputting image data features of different modalities. The third feature extraction branch is a sequence data processing branch, which consists of principal component analysis, a multilayer perceptron, and a gating module. The gating module is embedded before the last layer output of the multilayer perceptron. It takes sensor data as input and uses principal component analysis to determine the number of principal components for sensor data of different modalities. The principal components are then concatenated and input into the multilayer perceptron. The multilayer perceptron selects the output features through the gating module to obtain sensor data features of different modalities. The text data features, image data features, and sensor data features output from each branch are each passed through a fully connected layer and mapped to a preset unified dimension to obtain feature vectors extracted from different modalities. ; The The feature vectors of each modality are input to the multi-head cross-modal attention module; the multi-head cross-modal attention module targets specific time points. Perform cross-modal attention computation on all modal features to obtain A modal feature vector rich in information Then take this Concatenate and concatenate the vectors to generate a representation of that time point. The final multimodal fusion feature vector ; The final multimodal fusion feature vector As input for operating conditions, and with a timestamp extracted from the database. and spatial coordinates Using physical variables as input and predicted physical quantities and risk level judgment results as output, an artificial neural network architecture is adopted and a physical information neural network is embedded to construct a subway engineering risk level prediction model based on the physical information neural network, which is used to assess the risk level of subway engineering.

[0007] Furthermore, the modality discriminator is implemented using a CNN.

[0008] Furthermore, a probability threshold is introduced as a benchmark for path selection: first, the modality discriminator calculates the attribute probability of the input image, and then compares the probability with the preset probability threshold; based on the comparison result, the system automatically selects and activates CNN or Vision Transformer for feature extraction, thereby outputting adapted image data features.

[0009] Furthermore, the specific calculation process of the multi-head cross-modal attention module is as follows: S1: For Each of the modalities , the feature vector The corresponding modes are generated after linear transformation. In each head The query vector below Key vector Sum value vector There are a total of Size; Then, concatenate the key and value vectors of other modalities. For each modality... To obtain other modalities in splicing key vector under the head and concatenated value vectors ; S2: Calculate the query vector and the splicing key vector The inner product of the transpose, divided by the dimension of the key vector. Scaling the square root yields the attention score. ; S3: Set a learnable penalty factor The attention score Multiplied by the aforementioned penalty factor The attention weights are then obtained by normalizing the data using the softmax function, and then concatenated with the feature vector. Multiply, we get m The first modality h Cross-modal attention output of the head ; S4. Modal of All outputs of the head Perform splicing and then use a linear transformation matrix. Transform into modes m Information-rich modal feature vectors ; S5. Will The information-rich modal feature vector of each modality The final multimodal fusion feature vector is obtained by cascading and concatenating the features. .

[0010] Furthermore, the multi-task output layer of the subway engineering risk level prediction model based on physical information neural network is set with an independent risk assessment level node, and the displacement field, stress field and pore pressure are predicted simultaneously. The risk assessment level node classifies low, medium, and high risk levels based on the predicted physical quantity and a set of risk level boundary values. The model incorporates physical constraints, which, together with boundary condition constraints, guide the model's training process. The physical constraints include partial differential equations for soil-rock equilibrium, seepage field control, soil dynamic response, and structural dynamics. The boundary condition constraints include displacement and stress equilibrium equations for key action surfaces. The risk level boundary values ​​are defined as calibrable parameters and are set through an independent calibration stage.

[0011] Furthermore, the method also includes: The GAMI algorithm of XAI (Explainable Artificial Intelligence) is used to perform interpretability analysis on the multimodal data analysis and fusion unit and the subway engineering risk level prediction model based on physical information neural network. The GAMI algorithm quantifies the independent contribution and interaction effect of each input feature to the final risk level prediction through inverse feature deconstruction and interaction effect analysis. The results of the interpretability analysis are used to guide closed-loop optimization, including: pruning the three feature extraction branches in the multimodal data analysis fusion unit and reconstructing the internal network structure based on feature contribution; and verifying the effectiveness of the global parameter automatic optimization mechanism by visualizing the distribution and evolution of learnable parameters within the overall architecture of the multimodal data analysis fusion unit and the subway engineering risk level prediction model.

[0012] Secondly, the present invention also protects a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, enables the method described thereon.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The method of the present invention can effectively integrate multi-source data related to subway projects, make full use of the advantages of various types of data, and help improve the accuracy and practicality of disaster risk level prediction.

[0014] 2. This invention constructs a risk level prediction model for subway engineering based on a physical information neural network. This model integrates physical equations related to soil and rock masses into the neural network, enabling it to learn and predict based on physical laws. This effectively captures complex data dependencies across time and dimensions, providing a scientific basis for early warning and resource optimization, and reducing potential disaster impacts.

[0015] 3. This invention introduces an interpretable artificial intelligence (XAI) analysis method. This method can not only visualize the importance of each influencing factor to the prediction model results of subway engineering risk levels, enabling effective tracing of the causes of subway engineering disaster risks; it can also guide the pruning and structural reconstruction of the multimodal data fusion process, achieving closed-loop optimization of the model, thereby ensuring that the prediction results have high interpretability while further improving the model's operating efficiency and robustness. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the multimodal data analysis and fusion unit in this invention.

[0017] Figure 2 This is a schematic diagram of the overall structure of an embodiment of the subway engineering risk level prediction model based on physical information neural network in this invention. Detailed Implementation

[0018] Specific embodiments of the present invention are given below. These specific embodiments are only used to further illustrate the present invention in detail, and do not limit the scope of protection of this application.

[0019] This invention relates to a multi-scale subway disaster risk level prediction method based on multi-source data fusion, comprising the following steps: 1. Construct a multimodal data analysis and fusion unit to process and integrate high-dimensional, multi-scale physical property data and metaphysical data based on multimodal data analysis and fusion.

[0020] The physical and metaphysical data related to the construction, operation, and maintenance of subway projects are diverse, including but not limited to soil conditions, groundwater levels, train speeds, population flow, economic development, and extreme weather, and encompass multi-dimensional information formats such as text, images, and sensor data. Therefore, a multimodal data analysis and fusion unit is used for preprocessing, feature extraction, and fusion to select effective datasets, thereby providing data support for the subway project risk level prediction model.

[0021] 1.1 Establishing a database The database contains data collected during the construction, operation, and maintenance of the subway project. It consists of two parts: physical property data and super-physical property data. Physical property data includes soil conditions, structural integrity, ground settlement, construction temperature, train speed, and groundwater level. Soil conditions include soil type, density, water content, and permeability. Structural integrity data includes cracks, deformation, and corrosion. Super-physical property data includes population demographics, population movement, extreme weather, economic development, and urban planning. The database covers multi-dimensional information formats, including text, images, and sensor data.

[0022] The database has time-series storage layers and geographic-series storage layers to ensure that every piece of data entered (whether text, image, or sensor data) is assigned a clear timestamp. and spatial coordinates The spatiotemporal coordinates corresponding to the data are represented as follows: Furthermore, the database supports real-time updates, facilitating the storage of predicted risk level information. It also features visualization capabilities, providing an intuitive and dynamic display of time distribution, geographical distribution, and prediction accuracy.

[0023] 1.2 Data Preprocessing Data preprocessing is a crucial preliminary step in building a multimodal data analysis and fusion unit, aiming to improve the quality and usability of the raw data. The primary task of this process is data cleaning, which involves a unified review of data from all sources to remove irrelevant information, errors, and inconsistencies, such as anomalous labels, special characters, and extra spaces. These issues may stem from data entry errors or actual anomalies encountered during the data collection process. For text-based data (Modality A), this embodiment uses a BERT-based word segmenter to process it, converting it from unstructured natural language into a numerical sequence. The BERT-based word segmenter in this step is matched with the BERT-based pre-trained language model in the subsequent knowledge processing branch. Specifically, this word segmenter is used to standardize and index the original text, such as construction specifications and evaluation reports, mapping the text to numerical input IDs that the model can directly read, thus obtaining a standardized text ID sequence. This step only utilizes the BERT architecture's word list mapping function for digital preprocessing of the data and does not yet involve semantic feature extraction from deep neural networks.

[0024] For image-based data (Modal B), this embodiment performs a series of visual information standardization and enhancement operations. First, resizing is performed by rescaling and cropping all input images using image processing techniques to match their resolution to the requirements of the subsequent visual model input layer. Second, data augmentation is performed by creating variants of the original images through geometric transformations, color adjustments, and random erasure to enhance the model's generalization ability. Finally, filtering operations are performed to reduce image noise and enhance its key features.

[0025] For sensor-type data (Modal C), the core processing principle of this embodiment is to address the issues of missing data and scale inconsistencies. For missing values ​​in the monitoring data, this embodiment employs nearest-neighbor interpolation to impute the missing values, ensuring the integrity of the time series data. Immediately after imputation, Z-score standardization is performed on all numerical sensor data, adjusting the mean to 0 and the variance to 1. This step eliminates dimensional and numerical scale differences between different sensors (such as temperature and strain), ensuring the fairness of subsequent model analysis.

[0026] 1.3 Multimodal Data Analysis and Fusion In this embodiment, the multimodal data types are defined as follows: text data is A, image data is B, and sensor data is C. Text data includes construction specification text A1, evaluation report text A2, and maintenance record text A3; image data includes RGB image B1, grayscale image B2, and infrared thermal imaging image B3; sensor data includes earth pressure gauge data C1, hydrostatic level data C2, strain gauge data C3, total station data C4, and temperature and humidity sensor data C5, totaling 11 modes. In the model, the total number of modes is represented by... express( =11), the index of any modality is then used. express.

[0027] A multimodal data analysis and fusion unit is constructed, including a multi-head cross-modal attention module and three parallel branches: the first feature extraction branch is a knowledge processing branch; the second feature extraction branch is a visual feature extraction branch; and the third feature extraction branch is a sequence data processing branch.

[0028] The first feature extraction branch (knowledge processing branch) uses a pre-trained language model based on the BERT architecture as the feature extractor. This branch takes the generated standardized text ID sequence as input to the deep Transformer network of the pre-trained language model based on the BERT architecture, uses the self-attention mechanism to capture the contextual dependencies of the text, and finally extracts the hidden layer vectors that can represent the semantic information of the entire text as the output of this modality.

[0029] The second feature extraction branch (visual feature extraction branch): employs a modality discriminator and a parallel Vision Transformer and Convolutional Neural Network (CNN). This branch takes image data as input and introduces a probability threshold as a basis for path selection: first, the modality discriminator calculates the attribute probabilities of the input image, and then compares these probabilities with a preset probability threshold; based on the comparison result, the system automatically selects and activates either the CNN or the Vision Transformer for feature extraction, thereby outputting suitable image data features.

[0030] The third feature extraction branch (sequence data processing branch) consists of Principal Component Analysis (PCA), a Multilayer Perceptron (MLP), and a gating module. The gating module is embedded before the final output layer of the MLP. The input to this branch is sensor data. For sensor data of different modalities, PCA is first used for noise reduction and dimensionality reduction preprocessing. Then, the principal components are concatenated and input into the MLP. Before the final output layer, the MLP uses the gating module to select the output features, thus obtaining the sensor data features of different modalities.

[0031] The output of each branch is passed through a fully connected layer to map the features of each modality to a preset unified dimension. The feature vectors extracted from the data of different modalities are represented as follows: .

[0032] The multi-head cross-modal attention module is specifically implemented as described in S1-S5 below. This module receives... M A feature vector that has been mapped to a uniform dimension As input. Its core mechanism is: for M Each of the modalities m (As the query party), it computes cross-modal attention with all other modalities (as keys and values). Through this mechanism, each modality's All modalities are updated (i.e. enriched with information from all other modalities), resulting in information-enriched modal feature vectors. Finally, all of them M A modal feature vector rich in information Cascaded concatenation yields the final fused feature vector. .

[0033] Furthermore, the modality discriminator in the visual feature extraction branch is implemented using a CNN to determine whether the input image is more suitable for feature extraction using a CNN or a visual transformer. The modality discriminator outputs a probability value representing the likelihood that the image is suitable for feature extraction using a CNN. During the training phase, a cross-entropy loss function is used to measure the difference between the predicted probability and the true label. The true label is a binary classification: if the image has significant local features, it is labeled as suitable for a CNN; if the image has long-range dependencies and global features, it is labeled as suitable for a visual transformer. The modality discriminator parameters are iteratively updated using backpropagation and gradient descent to gradually converge the discriminator. The probability threshold is determined using a validation set; specifically, an ROC curve is plotted on the validation set, and the probability value corresponding to the optimal inflection point is selected as the final probability threshold to optimize system performance.

[0034] For different sensor data types, Principal Component Analysis (PCA) is used to perform PCA on each type of sensor data to extract its principal components. The covariance matrix of the data is calculated, and eigenvalues ​​and eigenvectors are solved. After determining the number of principal components based on the eigenvalues, the data is input into a Multilayer Perceptron (MLP) to extract one-dimensional vector features. A gating module is added between hidden layers to suppress feature activation in irrelevant regions, improving the MLP's focus on target features when processing various sensor data. Specifically: First, PCA is performed on each type of sensor data: its covariance matrix is ​​calculated, eigenvalue decomposition is performed, and the top principal components with a cumulative variance contribution rate of 85% are selected based on eigenvalue ranking. The original data is projected onto these principal components to obtain a dimensionality-reduced data representation. Then, the principal components extracted from each sensor are concatenated into a new feature matrix, which is input into the MLP for high-level feature extraction. The MLP consists of an input layer, hidden layers, and an output layer. The hidden layers use the ReLU activation function, and parameters are optimized through backpropagation and gradient descent to minimize the cross-entropy loss function. A gating module is embedded before the final output of the MLP. The gating module improves the feature weights through learnable parameters and the Sigmoid function, dynamically weights and filters the deep features extracted by the MLP, suppresses the activation of irrelevant features, enhances the model's attention to key information of the target, and finally outputs highly robust sensor data features.

[0035] The module then moves to the multi-head cross-modal attention module, which calculates cross-modal attention weights between different modalities, identifies and learns the correlations between modalities, and adaptively adjusts feature representations by mining deep feature relationships within and between modalities, thereby enhancing the robustness and generalization of features and improving the multi-modal feature fusion effect.

[0036] Specifically, the computation process of the multi-head cross-modal attention module is as follows: S1. Key Vector Generation: In each multi-head attention layer, for each modality... First, we need to generate each header. (Assuming there are a total of) Query vectors under (size) Key vector Sum value vector Specifically, it involves the feature vector After linear transformation, a multi-headed query vector, key vector, and value vector are generated, which can be expressed by the following formula:

[0037]

[0038]

[0039] in, It is the index of the head. This indicates a linear transformation operation. It is the first The mode corresponding to the size The learnable query weight matrix will transform the feature vectors Mapping to different query subspaces to obtain a suitable representation for the query, the system can extract the first value from multiple angles and dimensions using multiple different matrices. Key query information for modality. and They represent the first The mode corresponding to the size The learnable key weight matrix and value weight matrix.

[0040] Then, the key and value vectors of other modalities are concatenated to aggregate information from those modalities. For modalities... Calculation of other modes in splicing key vector under the head and concatenated value vectors ,for Head mode Query vector Prepare to interact with the key and value vectors of all other modalities.

[0041]

[0042]

[0043] in, It is the total number of modes, in this embodiment M The value is 11.

[0044] S2. Attention Score Calculation: For each head... Calculate the mode query vector and splicing key vector The inner product of the transposes yields a matrix representing the similarity between them. The inner product result is divided by... Scaling yields the attention score. ;in It is the first The scaling operation, which adjusts the dimension of the key vector, helps prevent the inner product from becoming too large or too small, ensuring computational stability and gradient update effectiveness, and improving training stability and convergence speed.

[0045] S3. Adjust attention weights: Adjust the obtained attention scores... Multiply by the penalty factor The penalty factor here It is a related to the first The error information related to the modality is a learnable parameter that can be automatically optimized through backpropagation and the Adam algorithm during training. Specifically, if a perturbation of a modality under a certain head causes a large error in the final predicted risk level, it indicates that the modality has a critical influence in the corresponding query subspace. In this case, the corresponding penalty factor will be automatically increased. The larger the penalty factor, the higher the weight of that modality in the multi-head cross-modal attention calculation, thus playing a more important role in the final information fusion. Next, the similarity matrix, adjusted for scaling and penalty factors, is normalized to obtain the attention score using the softmax function, yielding the attention weights. These attention weights are then weighted against the value vectors of other modalities to obtain the output of this head, which is then concatenated with the value vector. Multiply to get the first... m The first modality Cross-modal attention output of the head The softmax function ensures that the weights within each head are normalized, so that the sum of the attention weights for different modalities in their respective query subspaces is 1, thus allocating attention resources reasonably.

[0046]

[0047] S4. Multi-head stitching and feature fusion: combining modalities The outputs of all the heads are concatenated and then transformed using a linear transformation matrix. Transform the splicing result into a modality fusion feature vector For each of the 11 modalities, the fused feature vector for each modality is calculated sequentially:

[0048] in, This indicates a splicing operation. It is modal The learnable linear transformation matrix.

[0049] Combining the previous steps, The overall expression is as follows:

[0050] S5. Full-modal feature concatenation: for all M For each modality, steps S1 to S4 are performed to obtain their respective fused feature vectors. Then, all of them M Fusion feature vectors of each modality By cascading and concatenating, the final multimodal fusion feature vector corresponding to specific time and space nodes is generated. This serves as the input of working conditions for the subsequent subway project risk level prediction model.

[0051]

[0052] in, This indicates a splicing operation. This indicates the result calculated in step S4. The fused feature vector of each modality.

[0053] 2. Construct a subway engineering risk level prediction model based on Physics-Informed Neural Network (PINN) to achieve dual-drive of physics and data.

[0054] The subway engineering risk level prediction model based on Physical Information Neural Network (PINN) adopts an artificial neural network architecture, incorporating the key ideas of PINN, and includes an input layer, a physical constraint embedding layer, and a multi-task output layer. The final multimodal fusion feature vector... As input for operating conditions, and simultaneously using the spatiotemporal coordinates corresponding to the physical equations extracted from the database in step 1.1. As physical independent variables, the physical partial differential equations describing the response of the soil and rock mass, along with the boundary conditions, are transformed into built-in constraints of the network. These constraints are imposed through a total loss function, which consists of a data fitting loss term. Physical constraint loss item and boundary constraint loss term The weighted summation is used to obtain the predicted risk value. During training, the total loss function is minimized to ensure that the final output risk prediction strictly follows the basic physical laws of geotechnical mechanics. The specific steps are as follows: S1 Input Layer: This layer receives the final multimodal fused feature vector. Time and space coordinates It contains information related to physical property data and metaphysical data.

[0055] S2 Physical Constraint Embedding Layer: This layer utilizes an automatic differentiation mechanism to solve for the partial derivatives of the input spatiotemporal coordinates with respect to physical quantities such as displacement and stress. These derivatives are then substituted into the physical partial differential equations and boundary conditions describing the static and dynamic responses of the soil and rock mass to calculate the deviation between the model's predicted values ​​and the physical conservation laws (i.e., the equation residuals). These residuals are instantiated as physical constraint loss terms. and boundary constraint loss term Both and the data fitting loss term They work together to form the total loss function, thereby guiding the learning of neural network parameters during the backpropagation optimization process.

[0056] Physical constraints are imposed in residual form through a set of partial differential equations describing the coupled behavior of soil and rock masses, and together with boundary condition constraints, guide the training process of the neural network. This invention categorizes these constraints into two main types: static response constraints and dynamic response constraints. Each type includes physical constraints (i.e., governing equations) and boundary condition constraints. Specifically: For static response theory, physical constraints include the equilibrium control equations of the soil and rock medium and the control equations for the seepage field.

[0057] 1. The governing equations for soil and rock equilibrium (Navier-Cauchy equations) are as follows:

[0058] in, and It is the Lamé constant, which can be obtained from the elastic modulus. Compared to Poisson Seek; and yes direction and Displacement components in direction; subscript , Represents spatial dimension index (values ​​range from 1, 2, to 3, corresponding to three orthogonal spatial directions respectively). , Represents spatial coordinate components, The first of the net volume forces per unit volume Each component.

[0059] 2. The governing equations for the seepage field control are as follows:

[0060] in, This is the water storage coefficient; It is the permeability coefficient; It is pore water pressure; For fluid viscosity; It is time; It represents the net outflow rate of fluid mass due to seepage per unit volume of rock and soil per unit time. For the Hamiltonian operator (Nabla operator), in the formula Gradient calculation refers to the direction and rate of change of pressure in space; It represents divergence calculation, referring to the net flux of a fluid per unit volume.

[0061] For dynamic response theory, physical constraints include the governing equations for soil dynamic response and structural dynamics: 1. The governing equation for the dynamic response of the soil is:

[0062] in, It is the soil density; It is soil displacement; It is time; For the Hamiltonian operator (Nabla operator), in the formula This represents the tensor divergence operation, which calculates the sum of the partial derivatives of the stress tensor in spatial coordinates, representing the stress gradient force inside the soil. The total volumetric force acting on a unit volume of soil; where , It is the soil elastic tensor; Let the soil strain tensor be defined as follows: In the formula It is the gradient of the displacement vector. It is the transpose of the gradient of the displacement vector.

[0063] 2. The governing equations for structural dynamics are as follows:

[0064] in, It is a quality matrix. It is the damping matrix. It is the stiffness matrix. It is time. It is structural displacement. This refers to the equivalent external load acting on the entire subway tunnel structure.

[0065] Boundary condition constraints include the displacement and stress equilibrium equations for the key action surfaces, and are also divided into static and dynamic cases: Static response boundary conditions are divided into displacement boundary (Dirichlet boundary) and stress boundary (Neumann boundary).

[0066] 1. Under static response conditions, the displacement equilibrium equations are set at the contact surface between the tunnel lining and the surrounding rock, and at the interface between the ground and the tunnel, and the displacements satisfy the given displacement values. .in, Indicates the boundary nodes under static response. exist Displacement components in the direction; This represents the first boundary case under static response; Represents the first boundary node Each node is numbered; Indicates the direction of displacement; At the boundary node superior The displacement component with a predetermined direction.

[0067] 2. Under static response conditions, the stress equilibrium equations are set at the upper surface boundary of the computational domain and on an unconstrained free surface. At the upper surface boundary of the computational domain, the stresses satisfy... .in, Represents the static stress tensor at the nodes Location, along the outer normal ,exist Surface force components in the direction; This represents the second boundary case under static response; It acts on the boundary nodes. In The surface force component in the direction; while on an unconstrained free surface, zero surface force (i.e., = 0).

[0068] Under dynamic response boundary conditions, there are also two cases: displacement boundary conditions and stress boundary conditions. 1. Under dynamic response conditions, the key surfaces of the displacement boundary conditions are located at the contact surfaces between the structure and the soil, and at internal constraints within the structure. The displacement boundary conditions are set as a known periodic function. To simulate dynamic displacement constraints during motion, where, Represents boundary nodes under dynamic response. exist Direction over time Changing dynamic displacement; This represents the first boundary case under dynamic response; express Displacement amplitude in the direction; The load frequency; express Phase angle in the direction; Indicates time.

[0069] 2. Under dynamic response conditions, the key application surface of the stress boundary condition is located at positions such as the inner surface of the tunnel floor. Its physical essence is the application of a periodic dynamic surface force representing the train load on this boundary. The rigorous mathematical expression of this condition is the stress equilibrium equation: ,in, Represents the dynamic stress tensor at the nodes Location, along the outer normal ,exist Surface force components in the direction; It is applied to the node exist The periodic dynamic surface force vector in the direction can be expressed as: ,in, This represents the second boundary case under dynamic response; and These are parameters related to material constants and load characteristics; The load frequency is denoted as .

[0070] S3 Multi-Task Output Layer: The output of this layer includes predicted physical quantities and risk level judgment results. The predicted physical quantity is specifically the displacement field. Equivalent force field pore pressure The multi-task output layer includes an independent risk assessment level node. This node uses a risk level assessment architecture to determine the risk level of the subway project. The risk level is based on the maximum displacement. (Unit: mm) Ratio of maximum equivalent stress to material yield strength and maximum pore water pressure The classification is divided into three levels as follows: when and and At that time, it was considered low risk. (Level 1). Under this risk level, the subway engineering system is in a stable, low-risk state. Its dynamic evolution process is smooth and completely controllable. The system operation has no abnormal fluctuations, and there is no risk accumulation phenomenon.

[0071] when At that time, it was considered high risk. At this risk level, the risks in the dynamic evolution of the subway engineering system accumulate rapidly, and the stability is significantly reduced. Timely intervention measures must be taken to prevent further deterioration.

[0072] When the "low risk" condition is not met Level 1) and does not meet the criteria of "high risk". When the condition is met, it is considered medium risk. At this risk level, the dynamic evolution of the subway engineering system exhibits certain fluctuations, requiring close monitoring of risk trends. The risk level assessment framework is as follows:

[0073] in, The yield strength of the material; , , , , , It is the boundary threshold used to classify risk levels. , , The upper bound for determining low risk. , , This serves as a lower bound for determining high risk, and is usually... < , < , < ).

[0074] Traditionally, these boundary values ​​are determined based on engineering experience. However, due to differences in geological conditions, design standards, construction techniques, and operational requirements among various subway projects, their values ​​exhibit significant ambiguity and uncertainty. Therefore, this invention innovatively treats these boundary values ​​as calibrable parameters. These calibrable parameters are set through an independent calibration phase. Specifically, these parameters can be preset based on the experience of domain experts, or calibrated using systematic optimization methods such as grid search on validation datasets to determine the threshold combination that best reflects the actual engineering situation. This approach ensures that the PINN model focuses on accurate fitting of physical laws while also providing the risk level assessment framework with flexibility and interpretability for specific projects.

[0075] The total loss function of the subway engineering risk level prediction model based on physical information neural network during training is:

[0076] in, , and These are the weighting coefficients for the physical constraint term, data term, and boundary condition term. , and The loss terms are physical constraint loss, data fitting loss, and boundary constraint loss. These three losses are constructed by calculating the mean squared error (MSE) and are used to measure the residuals of the model predictions to the physical equations, the deviation from the actual data, and the degree of non-compliance on the boundaries, respectively.

[0077] Based on Explainable AI (XAI), this study analyzes the importance of input features to the prediction results of a subway engineering risk level prediction model based on a physical information neural network.

[0078] An interpretable generalized additive model and its interaction effect analysis algorithm (GAMI) are introduced. The model decision-making is made transparent through inverse feature deconstruction and interaction effect analysis. The internal decision-making process of the multimodal data analysis fusion unit and the subway engineering risk level prediction model is visualized and analyzed. The model is deeply analyzed from two dimensions: feature fusion and model decision-making. Based on the analysis results, the model is optimized in a closed loop. The process of the model evaluating the risk level according to the input features and physical constraints is more intuitively displayed, which enhances the trust and understanding of the model and meets the transparency and interpretability requirements of the multimodal feature fusion process.

[0079] This step of XAI feature analysis and model optimization is specifically carried out from two aspects: feature importance analysis and visualization of the multimodal fusion process. In feature analysis, the GAMI algorithm can accurately quantify each input feature (i.e., the final multimodal fusion feature vector) through inverse deconstruction. The independent contribution of each component in the multimodal fusion model to the final risk level prediction, as well as the interaction effects between different features, are analyzed. For visualization of the multimodal fusion process, a heatmap is constructed to display the dynamic correlation strength between modes by analyzing the attention weights of multiple heads across modalities, identifying the dominant mode, and analyzing the penalty factor. The distribution was used to verify the cross-modal feature iteration process.

[0080] The analysis results of this step are used to guide the closed-loop optimization of the multimodal data analysis fusion unit and the PINN prediction model. This optimization is performed at two levels: structural optimization and parameter verification.

[0081] In terms of structural optimization, based on the feature contribution values ​​analyzed after multiple training iterations using the GAMI algorithm, pruning and network structure reconstruction are performed on feature extraction branches corresponding to features (or modalities) with consistently negative or near-zero contribution values. This simplifies the internal network structure of the multimodal data analysis fusion unit, reducing overfitting risk and computational complexity. Regarding parameter validation, this verifies the interpretability of the automatic optimization mechanism of the multimodal data analysis fusion unit and the PINN prediction model. The defined total loss function is the sole basis for training the PINN prediction model, and its gradient backpropagation is used to automatically update all learnable parameters, including all learnable parameters within the multimodal data analysis fusion unit (e.g., penalty factors). and each linear transformation matrix , , The network parameters of the subway project risk level prediction model (etc.) and the model itself are considered. The XAI analysis in this step (especially "analyzing the distribution of penalty factors" and "constructing heatmaps") is used to verify the effectiveness of this automatic optimization mechanism. Taking the penalty factor as an example, the effectiveness is verified by visualizing the penalty factor. The changes in the value during training can intuitively demonstrate this automatic optimization mechanism: when a perturbation of a certain mode affects the total loss function... If a modal has a significant impact (i.e., the modal information is critical, and its error is highly correlated with the total error), gradient descent will automatically optimize its corresponding penalty factor. The feature importance of a feature is increased, thus amplifying its attention weight; conversely, if its contribution is low or negative, its weight is suppressed. The feature importance analysis of GAMI complements this visualization verification, jointly confirming the adaptive optimization capability of the multimodal data analysis fusion unit and the subway engineering risk level prediction model during the training process.

[0082] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A multi-scale subway disaster risk level prediction method based on multi-source data fusion, characterized in that, The method includes the following: Data collected during the construction, operation, and maintenance of the subway project is obtained, stored in a database, and assigned a specific timestamp. and spatial coordinates ; After preprocessing, the data is divided into three categories: text, image, and sensor data. Each category contains multiple modalities, totaling [amount missing]. One mode; Constructing a multimodal data analysis and fusion unit: including a multi-head cross-modal attention module and three parallel feature extraction branches: The first feature extraction branch is a knowledge processing branch, which uses a pre-trained language model based on the BERT architecture. It takes text data as input and outputs text data features of different modalities. The second feature extraction branch is a visual feature extraction branch, which consists of a modality discriminator and a parallel module composed of a convolutional neural network and a visual transformer. It takes image data as input and determines whether to use a convolutional neural network or a visual transformer for feature extraction based on the modality discriminator, outputting image data features of different modalities. The third feature extraction branch is a sequence data processing branch, which consists of principal component analysis, a multilayer perceptron, and a gating module. The gating module is embedded before the last layer output of the multilayer perceptron. It takes sensor data as input and uses principal component analysis to determine the number of principal components for sensor data of different modalities. The principal components are then concatenated and input into the multilayer perceptron. The multilayer perceptron selects the output features through the gating module to obtain sensor data features of different modalities. The text data features, image data features, and sensor data features output from each branch are each passed through a fully connected layer and mapped to a preset unified dimension to obtain feature vectors extracted from different modalities. ; The The feature vectors of each modality are input to the multi-head cross-modal attention module; the multi-head cross-modal attention module targets specific time points. Perform cross-modal attention computation on all modal features to obtain A modal feature vector rich in information Then take this Concatenate and concatenate the vectors to generate a representation of that time point. The final multimodal fusion feature vector ; The final multimodal fusion feature vector As input for operating conditions, and with a timestamp extracted from the database. and spatial coordinates Using physical variables as input and predicted physical quantities and risk level judgment results as output, an artificial neural network architecture is adopted and a physical information neural network is embedded to construct a subway engineering risk level prediction model based on the physical information neural network, which is used to assess the risk level of subway engineering.

2. The method according to claim 1, characterized in that, The modality discriminator is implemented using a CNN.

3. The method according to claim 2, characterized in that, A probability threshold is introduced as a benchmark for path selection: First, the modality discriminator calculates the attribute probability of the input image, and then compares the probability with the preset probability threshold. Based on the comparison result, the system automatically selects and activates CNN or Vision Transformer for feature extraction, thereby outputting the appropriate image data features.

4. The method according to claim 1, characterized in that, The specific calculation process of the multi-head cross-modal attention module is as follows: S1: For Each of the modalities , the feature vector The corresponding modes are generated after linear transformation. In each head The query vector below Key vector Sum value vector There are a total of Size; Then, concatenate the key and value vectors of other modalities. For each modality... To obtain other modalities in splicing key vector under the head and concatenated value vectors ; S2: Calculate the query vector and the splicing key vector The inner product of the transpose, divided by the dimension of the key vector. Scaling the square root yields the attention score. ; S3: Set a learnable penalty factor The attention score Multiplied by the aforementioned penalty factor The attention weights are then obtained by normalizing the data using the softmax function, and then concatenated with the feature vector. Multiply, we get m The first modality h Cross-modal attention output of the head ; S4. Modal of All outputs of the head Perform splicing and then use a linear transformation matrix. Transform into modes m Information-rich modal feature vectors ; S5. Will The information-rich modal feature vector of each modality The final multimodal fusion feature vector is obtained by cascading and concatenating the features. .

5. The method according to claim 1, characterized in that, The multi-task output layer of the subway engineering risk level prediction model based on physical information neural network sets an independent risk assessment level node and simultaneously predicts displacement field, stress field and pore pressure. The risk assessment level node classifies low, medium, and high risk levels based on the predicted physical quantity and a set of risk level boundary values. The model incorporates physical constraints, which, together with boundary condition constraints, guide the model's training process. The physical constraints include partial differential equations for soil-rock equilibrium, seepage field control, soil dynamic response, and structural dynamics. The boundary condition constraints include displacement and stress equilibrium equations for key action surfaces. The risk level boundary values ​​are defined as calibrable parameters and are set through an independent calibration stage.

6. The method according to claim 1, characterized in that, The method further includes: The GAMI algorithm of XAI (Explainable Artificial Intelligence) is used to perform interpretability analysis on the multimodal data analysis and fusion unit and the subway engineering risk level prediction model based on physical information neural network. The GAMI algorithm quantifies the independent contribution and interaction effect of each input feature to the final risk level prediction through inverse feature deconstruction and interaction effect analysis. The results of the interpretability analysis are used to guide closed-loop optimization, including: pruning the three feature extraction branches in the multimodal data analysis fusion unit and reconstructing the internal network structure based on feature contribution; and verifying the effectiveness of the global parameter automatic optimization mechanism by visualizing the distribution and evolution of learnable parameters within the overall architecture of the multimodal data analysis fusion unit and the subway engineering risk level prediction model.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can implement the method described in any one of claims 1-6.

Images