A Method and System for Predicting Seismic Damage in Urban Agglomerations Based on Unmanned Aerial Vehicle Technology and Deep Learning Models

Abstract

This invention discloses a method and system for predicting seismic damage in urban agglomerations based on UAV technology and deep learning models. The method includes: constructing a database of multi-angle aerial photographs of typical buildings and a database of seismic damage to buildings under different load coupling effects; training an urban building agglomeration identification model based on the aerial photograph database to identify building type, geometry, and texture information; constructing and training horizontal and vertical seismic amplification factor prediction models to correct for seismic intensity under different terrain conditions; constructing and training a seismic damage prediction model for typical buildings, and predicting building damage levels based on the corrected seismic intensity; finally, achieving three-dimensional visualization of the damage status of urban building agglomerations through a damage level imaging module. This invention solves the problems of inaccurate urban building agglomeration modeling information and the large workload and low accuracy of seismic damage analysis in existing technologies, providing a scientific basis for the construction and assessment of resilient cities.

Description

Technical Field

[0001] This invention relates to the field of urban seismic analysis and prediction, and in particular to a method and system for predicting seismic damage to urban clusters based on UAV technology and deep learning models, which combines advanced remote sensing technology and artificial intelligence to rapidly model urban building clusters. Background Technology

[0002] With the accelerating pace of global urbanization, urban building complexes are expanding in scale and becoming increasingly complex. Against the backdrop of frequent earthquakes, the ability to quickly and accurately assess the potential damage to urban building complexes under seismic forces is crucial for developing disaster prevention and mitigation strategies, protecting people's lives and property, and building resilient cities. Traditional methods of obtaining urban building information primarily rely on manual on-site surveys and the compilation of drawings and documents. This approach is not only inefficient and costly, but also struggles to guarantee the real-time and comprehensiveness of the information, making it particularly challenging for mountainous cities with complex terrain.

[0003] Current technologies for analyzing building seismic response largely rely on limited numerical simulations or historical earthquake damage statistics. These methods often fail to adequately consider the diversity of building types, structural complexity, and heterogeneity of site conditions in actual urban environments. More importantly, traditional methods typically treat seismic motion as a uniform input, neglecting the significant amplification effect of complex terrain (such as ridges, valleys, and basins) on seismic motion (including both horizontal and vertical components), leading to substantial discrepancies between damage predictions and actual conditions. Furthermore, existing technologies only consider horizontal seismic motion, ignoring the impact of vertical seismic motion.

[0004] Furthermore, existing seismic analysis methods mostly focus on detailed simulations of individual important buildings or small local areas, lacking effective means for rapid and unified seismic assessment of entire building complexes at a macro-city scale. Directly applying traditional methods to city-level analysis faces significant challenges, including extremely high data acquisition costs, unbearable computational burdens, and difficulties in integrating and visualizing results. Therefore, there is an urgent need in this field for a highly efficient and reliable macro-level seismic damage analysis method capable of automation at the city cluster level to meet the strategic needs of urban disaster prevention planning, emergency deployment, and resilience assessment. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for urban cluster modeling and seismic damage prediction based on UAV technology and deep learning models. This method can achieve efficient and accurate modeling of the entire urban building cluster at a macro-urban scale. Furthermore, by comprehensively considering the coupling effect of vertical and horizontal ground motions and the amplification effect of complex terrain on both, it can achieve macro-level, accurate prediction and intuitive visualization of seismic damage to urban building clusters.

[0006] To solve the above-mentioned technical problems, the technical solution proposed in this application is as follows:

[0007] This invention provides a method for predicting seismic damage in urban agglomerations based on unmanned aerial vehicle (UAV) technology and deep learning models, comprising the following steps:

[0008] S1: Construct a database of multi-angle aerial photographs of typical buildings covering the target urban area;

[0009] S2: Construct a database of typical building seismic damage that simultaneously considers the coupling effects of vertical and horizontal loads;

[0010] S3: Based on the database constructed in step S1, train and optimize the urban building cluster recognition model;

[0011] S4: Based on the database constructed in step S2, train and optimize the earthquake damage prediction model for typical buildings;

[0012] S5: Construct a database of horizontal and vertical ground motion amplification factors, and train a prediction model for horizontal and vertical ground motion amplification factors based on this database;

[0013] S6: Use the amplification factor prediction model obtained in step S5 to correct the ground motion intensity, and input the corrected horizontal and vertical ground motion parameters into the damage prediction model trained in step S4 to obtain the damage level of each building.

[0014] S7: Input the damage level prediction results obtained in step S6 into the urban building complex damage level imaging module for three-dimensional visualization rendering.

[0015] Furthermore, the step S1 of constructing a database of multi-angle aerial photographs of typical buildings specifically includes:

[0016] S1.1: Acquire orthogonal, oblique, and circumferential aerial images of buildings using a drone equipped with a multi-lens camera, and simultaneously record camera parameters and geospatial reference data;

[0017] S1.2: Based on multi-view images, a 3D model of the building is generated using a motion reconstruction algorithm, and the boundary, height, and volume coordinate data of the building are extracted; a deep learning semantic segmentation model is used to generate a building outline mask and damage type is labeled.

[0018] S1.3: Associate the structural type, materials, service life, and historical damage records of the buildings to form a multimodal database.

[0019] Furthermore, the construction of a typical building earthquake damage database in step S2 specifically includes:

[0020] S2.1: Determine the ground motion parameters, including selecting and classifying ground motion intensity indices;

[0021] S2.2: Obtain damage data of buildings under different seismic intensities through at least one of the following methods: experimental data acquisition, measured data acquisition, and simulation data acquisition;

[0022] S2.3: Design the database structure and establish data tables containing basic building information, seismic parameters, test data, measured data and simulation data, as well as their relationships;

[0023] S2.4: Perform data integration and management, including data import, consistency checks, and regular backups and updates.

[0024] Furthermore, the training of the horizontal and vertical seismic amplification factor prediction model in step S5 specifically includes:

[0025] S5.1: Based on a database of horizontal and vertical seismic amplification factors;

[0026] S5.2: Design a neural network model based on a dedicated database. The neural network model has an input layer with 12 nodes, a shared feature layer using the Inception module, and a dual-branch output layer that outputs the horizontal magnification factor and the vertical magnification factor respectively.

[0027] S5.3: Employs a database-driven training strategy, including dynamic data sampling and progressive training from simple shapes to complex terrains;

[0028] S5.4: Perform model optimization and database collaboration, and eliminate features with insignificant impact through multiple rounds of training;

[0029] S5.5: Prediction and deployment are based on the database. During prediction, a weighted average is calculated by combining the model prediction results and the database retrieval values.

[0030] Furthermore, the training and optimization of the urban building cluster recognition model in step S3 specifically includes:

[0031] S3.1: Data loading and preprocessing: Dynamically load data from the typical building multi-angle aerial image database, and perform image standardization, geometric annotation conversion and texture feature standardization;

[0032] S3.2: Model architecture initialization, using EfficientNet-B2 as the backbone network and configuring multi-task branches;

[0033] S3.3: Optimize strategy configuration by using the AdamW optimizer and combining it with cosine annealing learning rate scheduling;

[0034] S3.4: Loss function combination, the total loss function is composed of the weighted sum of the classification task loss, regression task loss and texture task loss;

[0035] S3.5: Execute training loops and monitoring, and employ an early stopping mechanism;

[0036] S3.6: Implement online data augmentation and test-time enhancement;

[0037] S3.7: Perform post-training processing and deployment, including model quantization, performance evaluation, and continuous optimization.

[0038] On the other hand, this application also claims protection for a seismic damage prediction system for urban agglomerations based on unmanned aerial vehicle (UAV) technology and deep learning models, comprising:

[0039] The urban building cluster identification module is used to identify the type, geometric information, and texture features of buildings based on drone aerial imagery.

[0040] A typical building earthquake damage prediction module is used to predict the damage level of a building based on building structural information and ground motion parameters.

[0041] The horizontal and vertical ground motion amplification factor prediction module is used to predict the amplification effect of complex terrain on horizontal and vertical ground motions.

[0042] The urban building complex damage level imaging module is used to visualize the damage level prediction results output by the typical building earthquake damage prediction module in three dimensions.

[0043] The urban building cluster identification module, the typical building earthquake damage prediction module, and the urban building cluster damage level imaging module are connected in sequence, while the horizontal and vertical ground motion amplification coefficient prediction module and the urban building cluster damage level imaging module are connected independently.

[0044] Furthermore, the urban building cluster identification module includes:

[0045] The data loading and preprocessing unit is used to load and preprocess aerial image data from a database of multi-angle aerial photographs of typical buildings.

[0046] The model architecture unit adopts a backbone network based on EfficientNet-B2 and is configured with multi-task branches, which include a building type classification branch, a geometric information regression branch, and a texture feature encoding branch.

[0047] The model training and optimization unit is used to train, optimize, and deploy the model architecture unit.

[0048] Furthermore, the geometric information regression branch includes a 1×1 convolutional layer, a sigmoid activation function, and parallel output area and aspect ratio regression units; the texture feature encoding branch includes a 1×1 convolutional layer and an L2 normalization unit.

[0049] Furthermore, the typical building earthquake damage prediction module includes:

[0050] The data preprocessing and feature standardization unit is used to integrate building information and seismic motion parameters and to standardize them.

[0051] A BP neural network architecture unit includes an input layer, two hidden layers, and an output layer, wherein the output layer is configured to output a probability distribution of five damage levels or a continuous damage index.

[0052] The training strategy optimization unit is used to configure the loss function, optimizer, and training strategy to optimize the BP neural network architecture unit.

[0053] Furthermore, the horizontal and vertical ground motion amplification factor prediction module includes:

[0054] The database unit stores topographic and geological data and corresponding horizontal and vertical seismic amplification factors;

[0055] The neural network model unit has an input layer that adapts to terrain feature input, a shared feature layer containing an Inception module, and a dual-branch output layer that outputs horizontal and vertical magnification factors respectively.

[0056] The prediction process optimization unit is used to combine the prediction results of the neural network model unit and the historical similar terrain data retrieved from the database unit to generate a weighted average final prediction result.

[0057] Furthermore, this application also claims protection for an electronic device comprising:

[0058] processor;

[0059] Memory, used to store computer programs;

[0060] The processor is configured to implement the method described in any of the above when executing the computer program.

[0061] Compared with the prior art, the present invention achieves the following beneficial technical effects:

[0062] This application achieves efficient and automated modeling and seismic analysis of urban building complexes by integrating UAV aerial photography and deep learning technologies. The method innovatively considers the coupling effect of vertical and horizontal ground motions, as well as the amplification effect of complex terrain, significantly improving the accuracy and scientific rigor of urban-scale damage prediction. Finally, it visually presents the macroscopic damage distribution through 3D visualization, providing reliable technical support for urban disaster prevention planning and emergency decision-making. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 A flowchart illustrating a method for predicting seismic damage in urban agglomerations based on unmanned aerial vehicle (UAV) technology and a deep learning model, provided as an embodiment of the present invention.

[0065] Figure 2 This is a flowchart for constructing a database of multi-angle aerial photographs of typical buildings, provided as an embodiment of the present invention.

[0066] Figure 3 The flowchart illustrates the construction of a typical building earthquake damage database that simultaneously considers the coupling effects of vertical and horizontal loads, as provided in an embodiment of the present invention.

[0067] Figure 4 The diagram below illustrates a structural block diagram of an urban agglomeration seismic damage prediction system based on unmanned aerial vehicle (UAV) technology and a deep learning model, as provided in an embodiment of the present invention.

[0068] Figure 5 The diagram shows the structural block diagram of the horizontal and vertical ground motion amplification coefficient prediction module of an urban agglomeration seismic damage prediction system based on UAV technology and deep learning model, which is provided in an embodiment of the present invention.

[0069] Figure 6 The diagram shows the structural block diagram of the urban building cluster identification module of an urban cluster seismic damage prediction system based on UAV technology and deep learning model, which is provided in an embodiment of the present invention.

[0070] Figure 7 This is a structural block diagram of a typical building earthquake damage prediction module in an urban agglomeration earthquake damage prediction system based on UAV technology and deep learning model, provided as an embodiment of the present invention.

[0071] Figure reference numerals: Urban building cluster identification module 100, data loading and preprocessing unit 110, model architecture unit 120, model training and optimization unit 130, typical building earthquake damage prediction module 200, data preprocessing and feature standardization unit 210, BP neural network architecture unit 220, training strategy optimization unit 230, horizontal and vertical ground motion amplification coefficient prediction module 300, database unit 310, neural network model unit 320, prediction process optimization unit 330, urban building cluster damage level imaging module 400. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] It should be noted that relational terms such as "first" and "second" used herein are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Please see Figure 1 A method for predicting seismic damage in urban agglomerations based on UAV technology and deep learning models includes the following steps S1 to S7.

[0075] Step S1: Construct a database of multi-angle aerial photographs of typical buildings. This step forms the data foundation for all subsequent analyses.

[0076] Step S2: Construct a database of typical building seismic damage under different vertical and horizontal load coupling effects. This database is used to train the damage prediction model.

[0077] Step S3: Based on the database constructed in Step S1, train and optimize the urban building cluster recognition model. This model is used to automatically identify building information from newly acquired aerial data.

[0078] Step S4: Based on the database constructed in Step S2, train and optimize the typical building earthquake damage prediction model. This model is the core of the damage prediction.

[0079] Step S5: Construct a database of horizontal and vertical ground motion amplification factors, and train a prediction model for the horizontal and vertical ground motion amplification factors based on this database. This model is used to correct the ground motion parameters input into the damage prediction model.

[0080] Step S6: Using the amplification factor prediction model obtained in Step S5, the seismic intensity is corrected according to the actual urban terrain conditions. Then, the corrected seismic parameters are input into the damage prediction model trained in Step S4 to obtain the damage level of each building.

[0081] Step S7: Input the damage level prediction results obtained in step S6 into the urban building complex damage level imaging module for three-dimensional visualization rendering to generate an intuitive damage distribution map.

[0082] Please see Figure 2 Step S1, constructing a database of multi-angle aerial photographs of typical buildings, specifically includes the following sub-steps:

[0083] S1.1: Acquire orthographic, oblique, and surround aerial images of the building using a drone equipped with a multi-lens camera, and simultaneously record camera parameters and geospatial reference data. Aerial photography operations using a drone equipped with a multi-lens camera should acquire orthographic, oblique, and surround aerial images of the building, ensuring coverage of the entire facade and local details. Simultaneously record camera parameters (EXIF information, shooting position, orientation, altitude) and geospatial reference data (GIS map, administrative boundary, elevation model). A reasonable aerial photography route and parameters must be developed.

[0084] S1.2: Based on multi-view imagery, a 3D building model is generated using the Structure for Motion Reconstruction (SfM) algorithm, and the building's boundary, height, and volume coordinate data are extracted. A deep learning semantic segmentation model is used to generate a building contour mask, and damage types are labeled. Based on the acquired multi-view imagery, a 3D building model is generated using the SfM algorithm. The building's boundary, height, and volume coordinate data are extracted from the 3D model. Simultaneously, deep learning semantic segmentation models such as U-Net are used to process the imagery, generating a building contour mask, and labeling damage types (such as cracks, collapse, and deformation) at the pixel, instance, and scene levels.

[0085] S1.3: Establish a multimodal database by associating the building's structural type, materials, service life, and historical damage records. Obtain the building's structural type, materials, service life, and historical damage records from design drawings, historical maintenance archives, and other sources. Link and integrate this attribute information with the aerial imagery, 3D model data, and annotation data obtained in S1.1 and S1.2 to form a unified multimodal database. Each building should have a unique identifier.

[0086] Please see Figure 3 Step S2 is a flowchart for constructing a typical building seismic damage database that simultaneously considers the coupling effects of vertical and horizontal loads. This step aims to establish a database of the mapping relationship between seismic input and building damage response. The specific implementation is as follows:

[0087] S2.1: Determine seismic motion parameters, including selecting and classifying seismic intensity indices. Peak ground acceleration, peak ground velocity, and acceleration response spectrum are selected as key seismic intensity indices. Based on seismic design codes or research requirements, seismic motion intensity is classified into multiple levels.

[0088] S2.2: Obtain damage data of buildings under different seismic intensity through at least one of the following methods: experimental data acquisition, measured data acquisition, and simulation data acquisition. The following three methods are preferred for collecting damage data of buildings under seismic loading:

[0089] Data acquisition: Design and execute shaking table tests or quasi-static tests, use sensors to collect the acceleration, displacement and strain response of the structure under simulated seismic loads, and record in detail the damage modes such as crack propagation and component failure.

[0090] Data Acquisition: After historical earthquake events, conduct on-site earthquake damage surveys, systematically record the damage status of various types of buildings, and collect corresponding actual ground motion records.

[0091] Simulation data acquisition: Establish refined finite element models representing different structural types, input seismic waves of different intensities, perform nonlinear dynamic time history analysis, and extract response parameters such as inter-story drift angle, component internal forces, and damage index.

[0092] S2.3: Design the database structure, establishing data tables containing basic building information, seismic motion parameters, experimental data, measured data, and simulation data, along with their relationships. Design relational data tables conforming to database normalization, including a building information table, a seismic motion record table, an experimental data table, a seismic damage survey table, and a numerical simulation results table, establishing relationships between the tables using unique identifiers.

[0093] S2.4: Perform data integration and management, including data import, consistency checks, and regular backups and updates. Import cleaned and standardized data into the database management system, implement data consistency checks, and establish a regular data backup, update, and version management mechanism.

[0094] Step S3, Urban Building Cluster Recognition Model Training and Optimization, in this application specifically includes the following sub-steps:

[0095] S3.1: Data Loading and Preprocessing. PyTorch DataLoader was used to dynamically load data from a database of multi-angle aerial photographs of typical buildings, with a batch size of 16, supporting multi-threaded parallel processing. Preprocessing included: normalizing pixel values ​​to the [0,1] range; performing histogram equalization to enhance contrast; normalizing the coordinates of the building's minimum bounding rectangle; and normalizing texture features (such as the roof material color histogram).

[0096] S3.2: Model Architecture Initialization. The backbone network uses EfficientNet-B2, with its weights initialized using a Kaiming uniform distribution. The multi-task branch configuration is as follows: The building type classification branch consists of global average pooling, a 256-dimensional fully connected layer (ReLU activation), a Dropout layer (rate=0.5), and a Softmax output layer (number of categories N). The geometric information regression branch consists of a 1×1 convolutional layer (output channels=4), a Sigmoid activation function, and parallel output area and aspect ratio regression units (linear activation). The texture feature encoding branch consists of a 1×1 convolutional layer (output channels=64) and an L2 normalization layer, outputting a 64-dimensional texture feature vector.

[0097] S3.3: Optimization Strategy Configuration. The AdamW optimizer is used, with a learning rate of 3e-4, weight decay of 1e-4, β1=0.9, and β2=0.999. Cosine annealing is used for learning rate scheduling, with an initial learning rate of 3e-4, a minimum learning rate of 3e-6, a cycle of 50 rounds, and 5 warm-start rounds. Global norm gradient clipping is implemented with a threshold of 1.0. Classification label smoothing (coefficient = 0.1) and geometric parameter Huber Loss are used as regularization measures.

[0098] S3.4: Combination of Loss Functions. The total loss function L_total consists of three weighted components: cross-entropy loss L_cls for classification tasks (weight 0.6), loss L_reg for regression tasks (including Smooth L1 Loss for coordinate regression and Huber Loss for geometric parameters, weight 0.3), and cosine similarity loss L_texture for texture tasks (weight 0.1).

[0099] S3.5: Training Loop and Monitoring. The total number of training rounds is set to 100, and an early stopping mechanism is adopted (the training terminates if the validation set loss does not decrease for 10 consecutive rounds). After each round of training, the loss on the validation set, building type accuracy, geometric localization mIoU, and texture similarity are calculated. The model with the minimum validation set loss and the model from the last round are saved.

[0100] S3.6: Data Augmentation and Regularization. Online data augmentation is dynamically applied during training, including random rotation (±15°), horizontal / vertical flipping, color dithering (brightness / contrast ±20%), and Gaussian noise (σ=0.02). Test-Time Augmentation (TTA) is used during inference, fusing the prediction results of the five transformations. Bagging can also be used to integrate three independently trained models.

[0101] S3.7: Post-training processing and deployment. Use TensorRT for INT8 quantization to compress the model size and improve inference speed. Evaluate model performance (accuracy, recall, F1 score, etc.) on independent test sets. Export the model in ONNX format to adapt it for edge devices. Build an online learning module to regularly integrate new data and fine-tune the model.

[0102] Step S5 of this application, which trains the prediction model for horizontal and vertical ground motion amplification factors, specifically includes the following sub-steps:

[0103] S5.1: Based on a database of horizontal and vertical seismic amplification factors.

[0104] The database of horizontal and vertical seismic amplification factors includes:

[0105] S5.01: Data Acquisition and Preprocessing: Establish a three-dimensional near-surface topographic and geological model, collect strong earthquake observation records of typical areas, and perform filtering and noise reduction processing on the observation data;

[0106] Specifically, topographic and geological modeling includes: using image recognition technology to extract the contours of complex terrains such as ridges, valleys, and basins, and combining ground-penetrating radar and micro-logging data to establish a three-dimensional near-surface model.

[0107] The integration of strong earthquake observation data includes: collecting strong earthquake records from typical areas such as the Beichuan Valley and the Wenchuan Basin, and organizing peak horizontal and vertical acceleration (PGA) and peak horizontal and vertical displacement (PGD) data according to topographic classification (ridge, valley, basin). The observation data is then filtered and denoised to remove outliers, ensuring a data integrity rate of ≥95%.

[0108] S5.02: Calculation and verification of magnification factor: Calculate the horizontal and vertical magnification factors through numerical simulation analysis, and cross-compare the simulation results with the measured values ​​for verification;

[0109] Specifically, the numerical simulation analysis includes: using the finite difference method to simulate the propagation of ground motion in different terrains, and calculating the horizontal and vertical amplification factors.

[0110] Cross-validation includes comparing simulation results with measured values, and conducting extensive parameter analysis to collect horizontal and vertical seismic amplification factors.

[0111] S5.03: Database Structure Design and Construction: Design the database table structure, including fields for terrain type, elevation, seismic wave frequency, horizontal magnification factor, and vertical magnification factor, and store the verified data in the database;

[0112] Specifically, the table structure definition includes: designing database tables according to the third normal form, containing fields such as: terrain type (ridge / valley / basin), elevation, seismic wave frequency, horizontal magnification factor, vertical magnification factor, data source (observation / simulation), and verification status.

[0113] Further develop the data dictionary, clarifying the units and value ranges of each field. Simultaneously generate metadata files to record the data collection time, processing methods, and quality ratings.

[0114] S5.2: Design a neural network model based on a dedicated database. The model's input layer has 12 nodes, corresponding to topographic and geological parameters such as elevation, slope, aspect, soil type encoding, shear wave velocity, magnitude, and epicentral distance. The shared feature layer uses the Inception module, employing 1×1, 3×3, and 5×5 convolutional kernels in parallel to extract multi-scale features, and incorporates a BatchNormalization layer. The dual-branch output layer outputs the horizontal magnification factor Hamp (linear activation) and the vertical magnification factor Vamp (ReLU6 activation), respectively. A task weight balancing coefficient β (default 0.6) is introduced to adjust the contribution of the dual-branch loss.

[0115] S5.3: Employ a database-driven training strategy. Implement dynamic data sampling, prioritizing samples from the database that show a large current prediction error (e.g., Hamp prediction bias > 20%). Use progressive training, initially using only simple terrain (e.g., isolated terrain) data, gradually introducing complex terrain (e.g., regional ridge terrain) data.

[0116] S5.4: Model optimization and database collaboration. Through multiple rounds of training, the importance of features is analyzed, and features that do not have a significant impact on the prediction results are eliminated, thereby simplifying the model and improving its efficiency.

[0117] S5.5: Database-based prediction and deployment. The prediction process is as follows: After the user inputs terrain parameters, the system retrieves historical data of similar terrain (e.g., elevation error <10%, slope error <5%) from the database. A weighted average is then calculated by combining the retrieved data and the model prediction results: H_final = γ•H_model + (1-γ)•H_db, where γ is the model confidence level (default 0.8), and H_db is the mean of the retrieved values ​​from the database. When outputting the prediction results, the measured value range of similar terrain in the database is displayed simultaneously. The database needs to be maintained and updated regularly, including adding measured data, upgrading the simulation model, and deleting low-quality data.

[0118] Please see Figure 4 Based on the above method, the present invention provides an urban agglomeration seismic damage prediction system based on UAV technology and deep learning model. The system is used to execute the aforementioned urban agglomeration seismic damage prediction method based on UAV technology and deep learning model, and includes four modules: urban building agglomeration identification module 100, typical building seismic damage prediction module 200, horizontal and vertical ground motion amplification coefficient prediction module 300, and urban building agglomeration damage level imaging module 400.

[0119] The urban building cluster identification module 100 is responsible for automatically identifying and extracting various features of buildings from the raw aerial images acquired by the UAV. The typical building earthquake damage prediction module 200 uses the building information provided by the urban building cluster identification module 100, combined with ground motion parameters, to predict the damage state of buildings under specific earthquake actions. The horizontal and vertical ground motion amplification factor prediction module 300 is specifically used to analyze the impact of complex terrain conditions on earthquake ground motion, providing more accurate ground motion input for the typical building earthquake damage prediction module 200. Finally, the urban building cluster damage level imaging module 400 visualizes the prediction results of the typical building earthquake damage prediction module 200.

[0120] Specifically, see Figure 6 The urban building cluster identification module 100 can be further divided into a data loading and preprocessing unit 110, a model architecture unit 120, and a model training and optimization unit 130. The data loading and preprocessing unit 110 is responsible for reading data from a pre-built database of multi-angle aerial photographs of typical buildings and performing preprocessing operations such as standardization. The model architecture unit 120 is the core, and its backbone network adopts EfficientNet-B2, which achieves a good balance between accuracy and efficiency. Based on this, the model architecture unit 120 sets up three parallel branch tasks: a building type classification branch (e.g., outputting frame structure, brick-concrete structure, etc.), a geometric information regression branch (used to regress the bounding box coordinates, area, aspect ratio, etc. of buildings), and a texture feature encoding branch (used to extract texture features such as roof material and output feature vectors). The model training and optimization unit 130 is responsible for configuring training parameters (e.g., optimizer, learning rate strategy, loss function), executing the training process, implementing data augmentation strategies, and finally completing the model evaluation, quantization, and deployment.

[0121] See Figure 7The typical building earthquake damage prediction module 200 includes a data preprocessing and feature standardization unit 210, a BP neural network architecture unit 220, and a training strategy optimization unit 230. The data preprocessing and feature standardization unit 210 fuses building features (such as structural type, geometric dimensions, and material properties) from the urban building cluster identification module 100 with seismic motion parameters (such as PGA, PGV, and response spectrum features) and standardizes them to form a unified feature vector. The BP neural network architecture unit 220 adopts a double-hidden-layer BP neural network structure. The first hidden layer contains 64 nodes and uses the ReLU activation function, and the second hidden layer contains 32 nodes and uses the Sigmoid activation function. The output layer uses the appropriate activation function depending on whether the task is classification (outputting a level 5 damage probability) or regression (outputting a damage index of 0-1). The training strategy optimization unit 230 is responsible for defining the loss function (such as cross-entropy loss or mean squared error), selecting the optimizer (such as Adam), setting the training batch and early stopping strategy, and can introduce hyperparameter tuning, data augmentation, and adversarial training to improve model performance.

[0122] See Figure 5 The horizontal and vertical ground motion amplification factor prediction module 300 includes a database unit 310, a neural network model unit 320, and a prediction process optimization unit 330. The database unit 310 stores detailed topographic and geological data (such as elevation, slope, soil type, shear wave velocity, etc.) and their corresponding horizontal and vertical ground motion amplification factors, derived from strong earthquake observations and numerical simulations. The neural network model unit 320 is a specially designed network with 12 nodes in its input layer, corresponding to 12 types of topographic and geological feature parameters. The shared feature layer uses the Inception module, extracting multi-scale topographic features with parallel convolutional kernels (1×1, 3×3, 5×5). The network ultimately branches into two output branches: one outputting the horizontal amplification factor Hamp (linear activation), and the other outputting the vertical amplification factor Vamp (ReLU6 activation, limiting the output range). During model prediction, the prediction process optimization unit 330 retrieves historical data similar to the input terrain parameters from the database unit 310. Then, it performs a weighted average of the model prediction value H_model and the mean H_db of the retrieved database values ​​to obtain the final prediction result H_final, where the weight γ represents the model confidence level. This approach combines the advantages of both data-driven and model-driven methods.

[0123] The urban building complex damage level imaging module 400 receives the damage level prediction results of each building from the typical building earthquake damage prediction module 200. It typically uses a 3D engine such as Unity to map these data onto a 3D model of the city, generating a dynamic damage level heat map, which allows users to interactively view the distribution of earthquake damage in the entire urban area from multiple angles.

[0124] After all the above steps, the urban building complex damage level imaging module 400 can optionally generate a damage level heat map on the urban 3D model. For example, different colors can be used to represent different damage levels (e.g., green represents basically intact, yellow represents minor damage, orange represents moderate damage, red represents severe damage, and black represents collapse). Users can clearly see the spatial distribution of damage to the entire urban building complex under earthquake action, providing intuitive and scientific decision support for pre-disaster prevention, emergency response, and post-disaster assessment.

[0125] This application also provides an electronic device for implementing the above method, comprising:

[0126] At least one processor; and

[0127] A memory that is communicatively connected to the at least one processor;

[0128] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method.

[0129] The electronic device may include one or more processors, memory, and transmission devices, and the terminal may also include input / output devices.

[0130] Those skilled in the art will understand that the terminal can be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile internet device (MID), a PAD, and other terminal devices.

[0131] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting earthquake damage of urban agglomeration based on unmanned aerial vehicle technology and deep learning model, characterized in that, The method comprises the following steps: S1: constructing a typical building multi-angle aerial photograph database covering a target urban area; S2: constructing a typical building earthquake damage database considering the coupling of vertical and horizontal loads; S3: training and optimizing a city building group recognition model based on the database constructed in step S1; S4: training and optimizing a typical building earthquake damage prediction model based on the database constructed in step S2; S5: constructing a horizontal and vertical ground motion amplification coefficient database, and training a horizontal and vertical ground motion amplification coefficient prediction model based on the database; the training of the horizontal and vertical ground motion amplification coefficient prediction model in step S5 specifically comprises: S5.1: taking the horizontal and vertical ground motion amplification coefficient database as the basis; S5.2: designing a neural network model based on a special database, the neural network model having an input layer of 12 nodes, a shared feature layer adopting an Inception module, and a double-branch output layer respectively outputting horizontal amplification coefficients and vertical amplification coefficients; S5.3: adopting a database-driven training strategy, including dynamic data sampling and progressive training from simple terrain to complex terrain; S5.4: performing model optimization and database cooperation, and eliminating non-significant features through multiple rounds of training; S5.5: performing prediction and deployment based on the database, and performing weighted averaging of the model prediction results and the database retrieval values during prediction; S6: using the amplification coefficient prediction model obtained in step S5 to modify the ground motion intensity, and combining the typical building seismic damage prediction model S7: inputting the damage grade prediction results obtained in step S6 into a city building group damage grade imaging module to perform three-dimensional visual rendering. The building information identified by the urban building group identification model is input into the damage prediction model trained in step S4 together with the modified horizontal and vertical ground motion parameters and the building information to obtain the damage grade of each building. The step S1 of constructing the typical building multi-angle aerial photograph database specifically comprises:

2. The method of claim 1, wherein, S1.1: acquiring orthographic, oblique and surround aerial image of the building by a multi-lens camera carried by a drone, and synchronously recording camera parameters and geographic space reference data; S1.2: generating a three-dimensional model of the building based on the multi-view image through a motion recovery structure algorithm, and extracting boundary, height and volume coordinate data of the building; generating a building contour mask by a deep learning semantic segmentation model, and performing damage type labeling; S1.3: associating the structure type, material, service life and historical damage record of the building to form a multi-modal database. The step S2 of constructing the typical building earthquake damage database specifically comprises:

3. The method of claim 1, wherein, S2.1: determining ground motion parameters, including selecting ground motion intensity indicators and grading; S2.2: acquiring damage data of the building under different ground motion intensities through at least one of experimental data acquisition, measured data acquisition and simulation data acquisition; S2.3: designing a database structure, establishing a data table containing building basic information, ground motion parameters, experimental data, measured data and simulation data, and the association relationship thereof; S2.4: performing data integration and management, including data import, consistency check, regular backup and update. The step S3 of training and optimizing the city building group recognition model specifically comprises:

4. The method of claim 1, wherein, S3.1: data loading and preprocessing, dynamically loading data from the typical building multi-angle aerial photograph database, and performing image standardization, geometric annotation conversion and texture feature standardization; ​ S3.2: Model architecture initialization, EfficientNet-B2 is adopted as the backbone network, and multi-task branches are configured; S3.3: Optimization strategy configuration, AdamW optimizer is adopted combined with cosine annealing learning rate scheduling; S3.4: Loss function combination, the total loss function is composed of classification task loss, regression task loss and texture task loss; S3.5: Execute training cycle and monitoring, and adopt early stopping mechanism; S3.6: Implement online data augmentation and test time augmentation; S3.7: Perform post-processing and deployment after training, including model quantization, performance evaluation and continuous optimization.

5. A system for implementing the method for predicting earthquake damage in urban agglomerations based on unmanned aerial vehicle technology and deep learning models according to claim 1, characterized in that, It comprises: a city building group identification module (100) for identifying the type, geometric information and texture features of buildings based on unmanned aerial vehicle aerial images; a typical building seismic damage prediction module (200) for predicting the damage level of buildings based on building structure information and ground motion parameters; a horizontal and vertical ground motion amplification coefficient prediction module (300) for predicting the amplification effect of complex terrain on horizontal and vertical ground motion; a city building group damage level imaging module (400) for three-dimensional visualization of the damage level prediction results output by the typical building seismic damage prediction module (200); wherein the city building group identification module (100), the typical building seismic damage prediction module (200) and the city building group damage level imaging module (400) are connected in sequence, and the horizontal and vertical ground motion amplification coefficient prediction module (300) and the city building group damage level imaging module (400) are independently connected.

6. The system of claim 5, wherein, The city building group identification module (100) comprises: a data loading and preprocessing unit (110) for loading and preprocessing aerial image data from a typical building multi-angle aerial image database; a model architecture unit (120) adopting an EfficientNet-B2-based backbone network and configuring multi-task branches, the multi-task branches including building type classification branches, geometric information regression branches and texture feature encoding branches; a model training and optimization unit (130) for training, optimizing and deploying the model architecture unit (120).

7. The system of claim 6, wherein, The geometric information regression branch includes a 1x1 convolutional layer, a Sigmoid activation function and an area and height-width ratio regression unit output in parallel; the texture feature encoding branch includes a 1x1 convolutional layer and an L2 normalization unit.

8. The system of claim 5, wherein, The typical building seismic damage prediction module (200) comprises: a data preprocessing and feature standardization unit (210) for integrating building information and ground motion parameters and standardizing them; a BP neural network architecture unit (220) including an input layer, double hidden layers and an output layer, the output layer being configured to output a probability distribution of five-level damage levels or a continuous damage index; a training strategy optimization unit (230) for configuring loss functions, optimizers and training strategies to optimize the BP neural network architecture unit (220).

9. The system of claim 5, wherein, The horizontal and vertical ground motion amplification coefficient prediction module (300) comprises: a database unit (310) storing topographic and geologic data and corresponding horizontal and vertical ground motion amplification coefficients; a neural network model unit (320) having an input layer adapted to input topographic features, a shared feature layer including an Inception module, and a double-branch output layer outputting horizontal and vertical amplification coefficients respectively; a prediction process optimization unit (330) configured to combine a prediction result of the neural network model unit (320) and similar topographic historical data retrieved from the database unit (310) to generate a weighted average final prediction result.

10. An electronic device, comprising: comprising: a processor; a memory for storing a computer program; wherein the processor is configured to implement the method of any one of claims 1 to 4 when executing the computer program.

Images