Earthquake disaster scenario recognition method and system based on deep learning

Abstract

The application belongs to the technical field of earthquake disaster scenario identification, and discloses an earthquake disaster scenario identification method based on deep learning, and the specific steps are as follows: S1, data acquisition and preprocessing S1.1, multi-source heterogeneous data acquisition, a comprehensive database containing seismic waveform data, ground deformation data, building structure data, geographic information data and historical disaster record data is established. Through the fusion of 3D convolution network, graph attention mechanism, space-time LSTM and adaptive cross-modal attention fusion technology, the joint modeling of seismic waveform space-time evolution law, ground deformation spatial distribution characteristics, building group topology vulnerability and disaster chain time sequence correlation is realized, the representation ability of complex nonlinear disaster mode is effectively improved, the disaster assessment response time is shortened to sub-second level through mixed precision quantization and edge computing deployment, and high identification accuracy is still maintained in the scene of strong noise and data loss by combining a multi-task classifier and a physical constraint verification mechanism.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of earthquake disaster scenario identification, and particularly relates to an earthquake disaster scenario identification method and system based on deep learning. BACKGROUND

[0002] Earthquake disaster scenario identification refers to comprehensively and scientifically identifying and evaluating the potential scenario of an earthquake disaster through systematic analysis of the scenario, process and influence of the earthquake disaster, so as to provide a basis for formulating disaster prevention and mitigation strategies, emergency plans and resource allocation.

[0003] Traditional earthquake disaster evaluation methods mainly rely on physical model simulation and statistical experience analysis, and have significant technical bottlenecks. The method based on the seismic wave propagation equation and the structural mechanics model is limited by the accuracy of parameter assumptions, and is difficult to adapt to the nonlinear dynamics characteristics under the coupling action of complex geological environment and building groups. The statistical learning method based on threshold discrimination faces the curse of dimensionality when processing multi-source heterogeneous data, and cannot effectively fuse the correlation rules of waveform time sequence characteristics, spatial deformation field and building topology, resulting in rough disaster element identification granularity. At the same time, the existing data fusion technology mainly adopts simple weighted superposition or rule splicing, lacks cross-modal feature alignment and deep semantic association modeling, and the physical coupling mechanism between InSAR deformation data, BIM structure parameters and real-time waveform data is not well analyzed. The traditional evaluation system relies on offline numerical simulation and manual interpretation, and has low computational efficiency when processing high-resolution remote sensing images and massive sensor data. The response delay of more than a minute cannot meet the real-time decision-making needs of emergency command. In addition, the disaster chain prediction model based on fixed rules has weak generalization ability for complex scenarios such as abnormal seismic source mechanism and secondary disaster interaction, and is prone to misjudgment when data is missing or noisy. These limitations seriously restrict the accuracy and timeliness of earthquake disaster scenario identification, and therefore need to be improved. SUMMARY

[0004] The purpose of the present application is to provide an earthquake disaster scenario identification method and system based on deep learning to solve the problems raised in the background art.

[0005] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: an earthquake disaster scenario identification method based on deep learning, the specific steps are as follows:

[0006] S1, data acquisition and preprocessing

[0007] S1.1, multi-source heterogeneous data acquisition

[0008] An integrated database is established, which includes seismic waveform data, ground deformation data, building structure data, geographic information data, and historical disaster record data, specifically including: P-wave and S-wave original waveform data recorded by the seismic network, with a sampling frequency not less than 100 Hz; millimeter-level ground deformation data obtained by InSAR satellite remote sensing; high-resolution disaster site image data obtained by unmanned aerial vehicle aerial photography; structural parameter data of the building information model (BIM); historical seismic intensity distribution map and disaster damage assessment report;

[0009] S1.2, data standardization processing

[0010] In the data standardization processing stage, a unified data processing framework with spatiotemporal consistency needs to be constructed. First, all geographic spatial data is accurately registered using the WGS84 global geodetic system, and the geographic reference deviation caused by different sensors and observation platforms is eliminated through coordinate conversion algorithms to ensure accurate matching of seismic station coordinates, satellite remote sensing images, building model positions and other multi-source data in three-dimensional space. On this basis, a global time synchronization mechanism based on coordinated universal time (UTC) is established to implement millisecond-level time alignment of heterogeneous time series data such as seismic waveform recorders, satellite transit times, and unmanned aerial vehicle collection times, and to construct a data correlation system under a continuous spatiotemporal coordinate system. For high-resolution aerial images, disaster site text reports and other unstructured data, image vectorization algorithms are used to convert raster data into vector layers with clear topological relationships, and natural language processing techniques are used to extract key semantic features from the text and encode them as structured vectors. Finally, through multi-dimensional data normalization processing, the physical quantity dimension differences are eliminated, the seismic waveform amplitude, ground deformation value, and building structure parameters are respectively standardized using the Min-Max standardization and Z-score standardization methods for scale unification, and the classification variables are implemented with one-hot encoding and embedding representation, and finally a standardized data set with unified spatiotemporal reference, standardized data structure, and balanced feature distribution is formed, laying a high-quality data foundation for subsequent deep learning model construction;

[0011] S2, deep learning model construction

[0012] S2.1, mixed architecture neural network

[0013] In the design of the multi-modal feature extraction network in the deep learning model construction stage, a mixed neural network architecture that works collaboratively needs to be constructed to fully exploit the multi-source features of the earthquake disaster data.

[0014] For spatio-temporal continuous seismic waveform data, a three-dimensional convolutional neural network is used for processing, and a convolution kernel with adaptive receptive field adjustment function is designed, with a size of 5x5x3. The hierarchical spatio-temporal feature extraction module captures the key seismic phase features such as P-wave first arrival and S-wave amplitude variation, and integrates the hollow convolution with an expansion rate of 2 to expand the time context perception range of the waveform sequence;

[0015] For high-resolution remote sensing image data, an improved U-Net architecture is constructed. In the encoder part, ResNet-50 is used as the backbone network and the channel attention mechanism is introduced. In the decoder stage, multi-scale feature fusion is implemented. Through the skip connection, the low-level texture features and high-level semantic features are aggregated across layers. Especially for disaster markers such as building collapse and ground fissure, a directional feature enhancement module is designed.

[0016] When processing building group structure data, a graph convolution network is used to establish a topological relationship analysis model. A single building is abstracted as a graph node, and the node attributes include structure type, floor height, construction year, etc. The spatial adjacency relationship is constructed as edge connection basis through Voronoi diagram. The graph attention mechanism (Graph Attention Layer) is used to dynamically learn the mechanical interaction weight between building groups, and the graph pooling operation based on physical constraints is designed to retain the key structure vulnerability features.

[0017] A bidirectional long short-term memory network is constructed to capture the temporal dynamics of disaster evolution. In the time dimension, 128 hidden units are set and a hierarchical propagation mechanism is used. The forward layer captures the propagation characteristics of the source rupture process, and the backward layer inversely analyzes the disaster chain reaction path. Through the time attention mechanism, the feature contribution degree of different time phases is weighted and fused. Finally, a multi-modal feature expression system is formed, which considers spatial topology, time evolution and physical mechanism. It provides a strong discriminative basic feature representation for subsequent cross-modal feature fusion.

[0018] S2.2, Feature Fusion

[0019] After the preliminary extraction of multi-modal features, there is a significant modal gap between the original heterogeneous features. To effectively integrate these cross-domain features, an adaptive attention mechanism needs to be designed to integrate the deep features of multi-source heterogeneous data.

[0020] Firstly, the heterogeneity of seismic waveform, remote sensing image, building structure and other data is eliminated by the cross-modal feature alignment layer. The dynamic time warping (DTW) algorithm is used to align the time series features with different sampling rates, and the feature projection matrix is constructed to realize the semantic matching in high-dimensional space. On this basis, the multi-head self-attention mechanism is introduced to establish the inter-modal correlation matrix, and the multi-dimensional dependency patterns such as the interaction between spatial deformation field and building group topology, the causal relationship between source rupture process and surface response, etc. are captured through parallel computing. Further, the differentiable feature selection gating unit is developed to dynamically evaluate the contribution of each modal feature based on the gated recurrent network (GRU), and the redundant features are soft shielded. The channel attention weighting strategy is used to focus on the key disaster identification features. Finally, the multi-scale features are mapped to the unified semantic space through the hierarchical feature aggregation module, and the feature dimension is compressed using the depth separable convolution. The 128-dimensional fusion feature vector is generated, which encodes the physical mechanism, spatial distribution and temporal evolution characteristics of the disaster, providing high discriminative input representation for subsequent classification tasks.

[0021] S2.3 Scenario recognition classifier

[0022] The 128-dimensional high-density feature vector generated through adaptive feature fusion not only preserves the physical characteristics of the original data, but also contains the potential laws of disaster evolution. To convert these abstract features into operable disaster scenario judgments, a hybrid classification decision system is constructed, and the joint reasoning of disaster elements is realized through a multi-task learning framework.

[0023] Firstly, the deep residual network (ResNet-50) is used as the backbone architecture to alleviate the gradient vanishing problem through jump connection. The global average pooling layer is used to map the fusion features to the disaster intensity level space, realizing the initial classification from Ⅵ degree to Ⅺ degree. For abnormal disaster patterns such as earthquake swarm and induced landslide, support vector data description (SVDD) is introduced to construct the hyper-spherical decision boundary, and the kernel function is used to detect dangerous scenarios deviating from the regular distribution. A conditional random field (CRF) post-processing module is designed to encode the spatial continuity constraints such as building damage degree and surface rupture zone direction into an energy function. The graph cut algorithm is used to optimize the regional consistency of the classification results. Finally, a multi-task learning framework is built to output the triple recognition results of disaster type, intensity level and impact range simultaneously through the collaborative design of shared bottom features and independent task heads. The uncertainty quantification module is used to evaluate the confidence of each prediction result, forming a decision output with physical interpretability and engineering practicality.

[0024] S3, Model training and optimization

[0025] S3.1, Training strategy design

[0026] In the model training stage, a progressive optimization strategy is adopted to improve learning efficiency and generalization ability: first, the network parameters are unfrozen in stages, and the feature extraction layer is trained preferentially to capture the basic pattern, and then the classifier parameters are optimized jointly; the curriculum learning mechanism is introduced, which gradually transitions from single source mechanism samples to multiple types of composite disaster scenes; synchronous implementation of adversarial training, through the generation of adversarial perturbation samples to enhance the robustness of the model to noise interference; for the long-tail distribution characteristics of earthquake disaster data, a dynamic weighted cross-entropy loss function is designed, and the weight coefficient is adaptively adjusted based on the class sample frequency to balance the class imbalance problem in the main and aftershock identification task;

[0027] S3.2, Hyperparameter Optimization

[0028] After the basic framework design of the training strategy is completed, the improvement of the model performance will depend on the precise exploration of the hyperparameter space, for which an intelligent parameter tuning system is constructed to achieve global optimal configuration of model parameters through multi-dimensional optimization strategies;

[0029] The Bayesian optimization algorithm is used to search for learning rate combinations (basic learning rate range 1e-5 to 1e-3), and the neural architecture search (NAS) is used to determine the Pareto optimal solution of network depth and convolution kernel size; genetic algorithm is used to optimize batch size (32-256) and L2 regularization coefficient (1e-4 to 1e-2), and multi-generation population evolution is used to screen anti-overfitting parameter groups; a hyperparameter response surface analysis model is constructed to quantify the influence of parameter interaction on recognition accuracy and inference speed, and finally the global optimal configuration that takes into account computational efficiency (FLOPs≤15G) and classification performance (F1-score≥0.92) is determined;

[0030] S4, Scenario Recognition and Decision Support Stage

[0031] S4.1, Real-time Inference Engine

[0032] To meet the real-time needs of earthquake emergency response, an inference system based on edge computing is designed: through mixed precision quantization technology, the floating point model is converted to 8-bit integer, realizing FPGA hardware acceleration deployment; a streaming data processing pipeline is constructed to support parallel access and online update of multi-source sensor data; a lightweight inference engine is integrated to ensure sub-second response delay (<500ms), and an uncertainty quantification module is developed simultaneously to output prediction confidence using the Monte Carlo Dropout method, providing reliability evaluation basis for decision-making;

[0033] S4.2, Visual Interaction Platform

[0034] A three-dimensional disaster situation awareness system is constructed, WebGL technology is used to realize real-time rendering on the browser side, dynamic superimposed display of the focal mechanism sphere, surface deformation field, and building damage heat map is supported, and the key evolution process from 10 seconds before the disaster to 30 minutes after the disaster can be traced back;

[0035] S4.3, auxiliary decision support

[0036] An intelligent auxiliary decision chain is established, an emergency response plan is automatically generated based on the scenario recognition result, the rescue path planning is optimized in combination with real-time traffic and aftershock probability prediction, the occurrence probability and spatio-temporal distribution of derivative risks such as landslides and barrier lakes are predicted through an LSTM network to warn of secondary disasters, a building reinforcement priority list is output, and the structure damage index, population density, and economic value are weighted and sorted to provide quantitative decision basis for post-disaster reconstruction.

[0037] Preferably, the sample imbalance problem existing in the earthquake disaster scenario recognition in S1.1 needs to construct a multi-level composite data augmentation system. First, based on the conditional generative adversarial network (CGAN) framework, by inputting the historical earthquake parameter condition constraint, combined with latent space feature decoupling technology, rare disaster scenario samples with physical rationality are synthesized, including low-occurrence high-intensity earthquakes, secondary disasters in special geological structure regions, and other rare cases.

[0038] Preferably, after the disaster scenario samples are generated, on this basis, spatial enhancement processing is performed on the multi-source observation data, random rotation, translation, affine transformation, and other geometric deformation operations are used, and Gaussian noise, random occlusion, color disturbance, and other texture disturbance strategies are introduced to enhance the robustness of the model to sensor errors and field environment interference.

[0039] Preferably, the angle range of random rotation is ±30°, the maximum offset of translation is 20%, the signal-to-noise ratio of Gaussian noise is controlled at 30-50dB, the occlusion ratio of random occlusion is 15%-40%, and the HSV space of color disturbance is ±10% adjustment.

[0040] Preferably, the sample imbalance problem commonly existing in the earthquake disaster scenario identification in S1.1 is further addressed by applying an improved Mixup algorithm to implement cross-class data fusion. By adjusting the linear interpolation coefficient λ (Beta distribution parameter α = 0.4), adjacent intensity level samples are smoothly transitioned in the feature space to construct transition-type training samples with continuous disaster intensity changes. Meanwhile, a three-dimensional spatio-temporal data cube is constructed, with continuous seismic phase waveform data stacked along the time dimension (time step 0.1 seconds) and kilometer-level gridded ground deformation fields integrated along the spatial dimension (resolution 100 m x 100 m). The spatio-temporal joint features are extracted by a three-dimensional convolution kernel (5 x 5 x 3 size), and the time series length is extended to 2-3 times the original data by data resampling technology. Finally, a comprehensive data augmentation scheme is formed, including geometric transformation, noise injection, feature mixing, and spatio-temporal expansion, which increases the training sample size by 3-5 times and significantly improves the model's representation ability for long-tail distribution data and generalization performance for complex disaster scenarios.

[0041] Preferably, after the system optimization in the hyperparameter space in S3.2, a multi-dimensional evaluation system including physical constraint verification is established to ensure the model's generalization ability in complex earthquake scenarios. This system not only verifies the model's theoretical performance but also focuses on its applicability in actual earthquake engineering. A multi-dimensional comprehensive evaluation system is constructed, and the data set is partitioned by epicenter location and occurrence time through spatio-temporal cross-validation strategy to avoid overfitting and test the generalization ability. A quantitative index system including disaster identification accuracy, false alarm rate, spatial positioning error, and response time delay is developed. Meanwhile, a physical constraint verification mechanism is introduced to reasonably check the prediction results using the seismic wave propagation equation and the law of conservation of energy. The robustness of the model under noise interference and data anomalies is tested by combining with the adversarial sample generation technology, and finally a reliability evaluation report is formed, which considers both the data-driven performance and the compliance with physical laws.

[0042] The earthquake disaster scenario identification system based on deep learning includes the following modules:

[0043] Data acquisition and preprocessing module: used for integrating seismic waveform, ground deformation, building structure, geographic information, and historical disaster records, and pre-processing the data;

[0044] Deep learning model construction module: used for designing a hybrid neural network architecture, integrating an adaptive attention mechanism to realize cross-modal feature fusion, and constructing a multi-task classifier to simultaneously output disaster type, intensity level, and impact range;

[0045] Model training module: used for designing a model training strategy and performing hyperparameter search using Bayesian optimization and genetic algorithm. Meanwhile, a physical constraint verification mechanism is introduced to establish a multi-dimensional evaluation system to ensure that the model's prediction results comply with the seismic propagation law and engineering practical requirements.

[0046] Visualization interaction module: for real-time superimposed display of focal mechanism sphere, surface deformation field and building damage thermal map, combined with time axis to trace key evolution process before and after disaster occurrence, to provide intuitive space-time analysis interface;

[0047] Auxiliary decision module: for generating emergency response plan.

[0048] Preferably, the model training module is based on a multi-modal heterogeneous data fusion framework, a multi-level feature extraction system is constructed by integrating 3D convolution network, U-Net, graph convolution network and bidirectional LSTM through a hybrid architecture neural network, cross-modal alignment and correlation matrix modeling of seismic waveform space-time features, surface deformation spatial features, building group topological features and disaster evolution time sequence features are realized by using adaptive attention fusion mechanism, a hybrid classification decision system is constructed by combining deep residual network, support vector data description and conditional random field, disaster type identification, intensity classification and influence range prediction tasks are simultaneously optimized by using multi-task learning framework, and 128-dimensional high discriminative feature vectors are dynamically generated by using differentiable feature selection gate unit, and finally an end-to-end earthquake disaster scenario recognition model with physical law embedding and space-time continuity constraint is formed.

[0049] Preferably, the auxiliary decision module can optimize rescue path planning and resource scheduling strategy, and output building reinforcement priority list by comprehensively considering structure damage index, population density and economic value, to provide quantitative decision support for post-disaster reconstruction.

[0050] The beneficial effects of the present application are as follows:

[0051] By fusing 3D convolution network, graph attention mechanism, space-time LSTM and adaptive cross-modal attention fusion technology, joint modeling of seismic waveform space-time evolution law, surface deformation spatial distribution characteristics, building group topological vulnerability and disaster chain time sequence correlation is realized, the representation ability of complex nonlinear disaster model is effectively improved, and the disaster assessment response time is shortened to sub-second level through mixed precision quantization and edge computing deployment, combined with multi-task classifier and physical constraint verification mechanism, high recognition accuracy is still maintained in strong noise and data missing scene, relying on multi-source heterogeneous data standardization integration and visual decision chain, a full-process closed loop from millimeter-level deformation monitoring, minute-level disaster assessment to dynamic rescue scheduling is formed, and the analysis efficiency of high-dimensional, multi-scale disaster data and the scientificity of emergency decision are significantly enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The system block diagram of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0054] As Figure 1 shown, the embodiment of the present application provides a seismic disaster scenario recognition method based on deep learning, and the specific steps are as follows:

[0055] S1, data acquisition and preprocessing

[0056] S1.1, multi-source heterogeneous data acquisition

[0057] An integrated database containing seismic waveform data, ground deformation data, building structure data, geographic information data, and historical disaster record data is established, specifically including: P-wave and S-wave original waveform data recorded by a seismic network, with a sampling frequency not less than 100Hz; millimeter-level ground deformation data obtained by InSAR satellite remote sensing; high-resolution disaster scene image data obtained by unmanned aerial vehicle aerial photography; structure parameter data of a building information model (BIM); historical seismic intensity distribution map and disaster loss assessment report;

[0058] S1.2, data standardization processing

[0059] In the data standardization processing stage, a unified data processing framework with spatiotemporal consistency needs to be constructed. First, all geographic spatial data is accurately registered using the WGS84 global geodetic system, and the geographic reference deviation caused by different sensors and observation platforms is eliminated through coordinate conversion algorithms to ensure accurate matching of seismic station coordinates, satellite remote sensing images, building model positions, and other multi-source data in three-dimensional space. On this basis, a global time synchronization mechanism based on coordinated universal time (UTC) is established to implement millisecond-level time alignment of heterogeneous time series data such as seismic waveform recorders, satellite transit times, and unmanned aerial vehicle acquisition times, and to construct a data correlation system under a continuous spatiotemporal coordinate system. For high-resolution aerial images, disaster site text reports, and other unstructured data, image vectorization algorithms are used to convert raster data into vector layers with clear topological relationships, and natural language processing techniques are used to extract key semantic features from the text and encode them as structured vectors. Finally, multi-dimensional data normalization is used to eliminate physical dimension differences, and Min-Max standardization and Z-score standardization methods are used to scale numerical data such as seismic waveform amplitudes, ground deformation values, and building structure parameters, and one-hot encoding and embedding representation are used for categorical variables to ultimately form a standardized dataset with unified spatiotemporal reference, standardized data structure, and balanced feature distribution, laying a high-quality data foundation for subsequent deep learning model construction.

[0060] S2, deep learning model construction

[0061] S2.1, mixed architecture neural network

[0062] In the design of the multi-modal feature extraction network in the deep learning model construction stage, a mixed neural network architecture that works collaboratively needs to be constructed to fully exploit the multi-source features of earthquake disaster data.

[0063] For spatiotemporally continuous seismic waveform data, a three-dimensional convolutional neural network is used for processing, and a convolution kernel with adaptive receptive field adjustment function is designed with a size of 5x5x3. Through hierarchical spatiotemporal feature extraction modules, key seismic phase features such as P-wave first arrival and S-wave amplitude changes are captured, and a dilated convolution is integrated with a dilation rate of 2 to expand the time context perception range of the waveform sequence.

[0064] For high-resolution remote sensing image data, an improved U-Net architecture is constructed. In the encoder part, ResNet-50 is used as the backbone network and channel attention mechanism is introduced. In the decoder stage, multi-scale feature fusion is implemented, and low-level texture features and high-level semantic features are aggregated through skip connections across layers. Special directional feature enhancement modules are designed for disaster markers such as building collapse and ground fissures.

[0065] In the process of processing building structure data, a topological relationship analysis model is established by using a graph convolution network. A single building is abstracted as a graph node, and the node attributes include structure type, floor height, construction year, etc. The spatial adjacency relationship is constructed by Voronoi diagram as the edge connection basis. The graph attention mechanism (Graph Attention Layer) is used to dynamically learn the mechanical interaction weight between the building group. The graph pooling operation based on physical constraints is designed to retain the key structure vulnerability features.

[0066] A bidirectional long short-term memory network is constructed to capture the time sequence dynamics of disaster evolution. 128 hidden units are set in the time dimension and a hierarchical propagation mechanism is used. The forward layer captures the propagation characteristics of the source rupture process, and the backward layer inversely analyzes the disaster chain reaction path. The time attention mechanism is used to weight and integrate the feature contribution degree at different time phases. Finally, a multi-modal feature expression system is formed, which takes into account the spatial topology, time evolution and physical mechanism, providing a strong discriminative basic feature representation for subsequent cross-modal feature fusion.

[0067] S2.2, feature fusion

[0068] After the preliminary extraction of multi-modal features, there is a significant modal gap between the original heterogeneous features. In order to effectively integrate these cross-domain features, an adaptive attention mechanism is designed to integrate the deep features of multi-source heterogeneous data.

[0069] Firstly, the cross-modal feature alignment layer is used to eliminate the heterogeneity difference of seismic waveform, remote sensing image, building structure and other data. The dynamic time warping (DTW) algorithm is used to align the time sequence features with different sampling rates, and a feature projection matrix is constructed to realize semantic matching in high-dimensional space. On this basis, the multi-head self-attention mechanism is introduced to establish the inter-modal correlation matrix. Through parallel computing, the interaction between spatial deformation field and building group topology, the causal relationship between source rupture process and surface response, and other multi-dimensional dependence patterns are captured. Further, a differentiable feature selection gating unit is developed to dynamically evaluate the contribution of each modal feature based on the gated recurrent unit (GRU), and soft shielding is implemented for redundant features. Channel attention weighting strategy is used to focus on key disaster identification features. Finally, the hierarchical feature aggregation module is used to map multi-scale features to a unified semantic space. Deep separable convolution is used to compress the feature dimension, and residual connection is used to retain the original information, generating a 128-dimensional fusion feature vector. This vector encodes the physical mechanism, spatial distribution and time evolution characteristics of the disaster, providing a high discriminative input representation for subsequent classification tasks.

[0070] S2.3 scenario recognition classifier

[0071] The 128-dimensional high-density feature vector generated by adaptive feature fusion not only retains the physical characteristics of the original data but also contains the potential laws of disaster evolution. To convert these abstract features into operable disaster scenario judgments, a hybrid classification decision system is constructed, and joint reasoning of disaster elements is achieved through a multi-task learning framework.

[0072] First, a deep residual network (ResNet-50) is used as the backbone architecture to alleviate the gradient vanishing problem through skip connections. The global average pooling layer is used to map the fused features to the disaster grade space, realizing the initial classification of intensity from VI to XI. For abnormal disaster patterns such as earthquake swarms and induced landslides, support vector data description (SVDD) is introduced to construct a hyperspherical decision boundary. Through kernel function mapping, dangerous scenarios deviating from the regular distribution are detected. A conditional random field (CRF) post-processing module is designed to encode the spatial continuity constraints such as building damage and surface rupture zone direction into an energy function. The graph cut algorithm is used to optimize the regional consistency of the classification results. Finally, a multi-task learning framework is built to output the three-tuple recognition results of disaster type, intensity level, and impact range simultaneously through the collaborative design of shared bottom features and independent task heads. The uncertainty quantification module is used to evaluate the confidence of each prediction result, forming a decision output with physical interpretability and engineering practicality.

[0073] S3、Model training and optimization

[0074] S3.1、Training strategy design

[0075] During model training, a progressive optimization strategy is adopted to improve learning efficiency and generalization ability. First, the network parameters are unfrozen in stages, and the feature extraction layer is trained preferentially to capture the basic patterns. Then, the classifier parameters are optimized jointly. The curriculum learning mechanism is introduced to gradually transition from single focal mechanism samples to multiple types of complex disaster scenarios. Simultaneous implementation of adversarial training enhances the robustness of the model to noise interference by generating adversarial perturbation samples. For the long-tail distribution characteristics of earthquake disaster data, a dynamic weighted cross-entropy loss function is designed to adaptively adjust the weight coefficients based on the frequency of class samples, balancing the class imbalance problem in main and aftershock identification tasks.

[0076] S3.2、Hyperparameter optimization

[0077] After completing the basic architecture design of the training strategy, the improvement of model performance will depend on the precise exploration of the hyperparameter space. Therefore, an intelligent parameter tuning system is constructed to achieve global optimal configuration of model parameters through multi-dimensional optimization strategies.

[0078] The Bayesian optimization algorithm is used to search for the learning rate combination (the basic learning rate range is 1e-5 to 1e-3), and the Pareto optimal solution of the network depth and convolution kernel size is determined by combining neural architecture search (NAS); the genetic algorithm is used to optimize the batch size (32-256) and L2 regularization coefficient (1e-4 to 1e-2), and the anti-overfitting parameter group is screened through multi-generation population evolution; the super parameter response surface analysis model is constructed, the influence of parameter interaction on identification accuracy and reasoning speed is quantified, and finally the global optimal configuration considering the calculation efficiency (FLOPs≤15G) and classification performance (F1-score≥0.92) is determined;

[0079] S4, scenario recognition and decision support stage

[0080] S4.1, real-time inference engine

[0081] In order to meet the real-time demand of earthquake emergency response, an inference system based on edge computing is designed: through the mixed precision quantization technology, the floating point model is converted to 8-bit integer, realizing the FPGA hardware acceleration deployment; a stream data processing pipeline is constructed to support parallel access and online update of multi-source sensor data; a lightweight inference engine is integrated to ensure sub-second response delay (<500ms), and an uncertainty quantization module is developed simultaneously to output the prediction confidence by using the Monte Carlo Dropout method, providing reliability evaluation basis for decision-making;

[0082] S4.2, visual interaction platform

[0083] A three-dimensional disaster situation awareness system is constructed, and the WebGL technology is used to realize real-time rendering on the browser side, supporting dynamic superposition display of focal mechanism sphere, surface deformation field and building damage heat map, and the key evolution process from 10 seconds before disaster to 30 minutes after disaster can be traced back;

[0084] S4.3, auxiliary decision support

[0085] An intelligent auxiliary decision chain is established, and an emergency response plan is automatically generated based on the scenario recognition result, and the rescue path planning is optimized by combining real-time traffic and aftershock probability prediction; at the same time, the occurrence probability and spatio-temporal distribution of derivative risks such as landslides and barrier lakes are predicted by LSTM network, and secondary disasters are warned; the building reinforcement priority list is output, and the structure damage index, population density and economic value are weighted and sorted to provide quantitative decision basis for post-disaster reconstruction.

[0086] By multi-source heterogeneous data standardization and integration, a hybrid neural network architecture is constructed to extract spatio-temporal features, an adaptive attention mechanism is adopted to fuse cross-modal features, a multi-task classifier is designed to simultaneously output disaster type, intensity and range, and a progressive training strategy and Bayesian hyperparameter optimization are combined to improve model performance. An edge computing engine is deployed to realize sub-second real-time inference, and a visualization platform and intelligent decision chain are established to realize the whole-process closed-loop management from disaster identification, dynamic evolution analysis to emergency resource scheduling, effectively improving the accuracy, response speed and decision-making scientificity of earthquake disaster assessment.

[0087] Among them, in S1.1, to address the sample imbalance problem commonly existing in earthquake disaster scenario identification, a multi-level composite data augmentation system needs to be constructed. First, based on the conditional generative adversarial network (CGAN) framework, by inputting the conditional constraint of historical earthquake parameters, combined with latent space feature decoupling technology, rare disaster scenario samples with physical rationality are synthesized, including low-probability high-intensity earthquakes, secondary disasters in special geological structure regions, and other scarce cases.

[0088] By conditional generative adversarial network, physically reasonable rare disaster scenario samples are generated, effectively alleviating the sample imbalance problem and improving the model's ability to identify low-probability high-intensity earthquakes and special geological secondary disasters.

[0089] Among them, after the generation of disaster scenario samples, spatial enhancement processing is performed on multi-source observation data. Geometric deformation operations such as random rotation, translation, and affine transformation are adopted, and texture disturbance strategies such as Gaussian noise, random occlusion, and color disturbance are introduced to enhance the model's robustness to sensor errors and on-site environmental interference.

[0090] Combined with the composite data augmentation strategy of geometric deformation and texture disturbance, the model's adaptability to sensor errors, environmental occlusion, and illumination changes is enhanced, and the generalization performance in complex on-site environments is improved.

[0091] Among them, the angle range of random rotation is ±30°, the maximum offset of translation is 20%, the signal-to-noise ratio of Gaussian noise is controlled at 30-50dB, the occlusion ratio of random occlusion is 15%-40%, and the HSV space color disturbance is adjusted by ±10%.

[0092] By precisely controlling the data augmentation parameter range, the enhancement effect is maximized while maintaining physical rationality, avoiding feature distortion caused by excessive deformation or noise interference.

[0093] Among them, in S1.1, the sample imbalance problem existing in the earthquake disaster scenario identification is further solved by applying the improved Mixup algorithm to implement cross-class data fusion. Through the linear interpolation coefficient λ (Beta distribution parameter α=0.4) regulation, the adjacent intensity level samples are smoothly transitioned in the feature space, and the transition training samples with continuous disaster intensity change are constructed. At the same time, a three-dimensional spatio-temporal data cube is constructed, the continuous seismic phase waveform data (time step 0.1 seconds) is stacked along the time dimension, and the kilometer-level gridded ground deformation field (resolution 100m×100m) is integrated along the spatial dimension. Through the three-dimensional convolution kernel (5×5×3 size), the spatio-temporal joint features are extracted, and the data resampling technology is used to expand the time series length to 2-3 times of the original data. Finally, a comprehensive data augmentation scheme covering geometric transformation, noise injection, feature mixing, and spatio-temporal expansion is formed, which makes the training sample size increase by 3-5 times, and significantly improves the model's representation ability for long-tail distribution data and generalization performance for complex disaster scenarios.

[0094] By using the improved Mixup algorithm and spatio-temporal data resampling technology, continuous disaster intensity transition samples are constructed and time series length is expanded, which strengthens the model's representation ability for continuous disaster evolution process and weak features.

[0095] Among them, in S3.2, after the system optimization in the hyperparameter space, to ensure the model's generalization ability in complex earthquake scenarios, a multi-dimensional evaluation system including physical constraint verification is established. This system not only verifies the model's theoretical performance, but also focuses on its applicability in actual earthquake engineering. A multi-dimensional comprehensive evaluation system is constructed, which partitions the data set by epicenter location and occurrence time through spatio-temporal cross-validation strategy to avoid overfitting and test the generalization ability. A quantitative index system including disaster identification accuracy, false alarm rate, spatial positioning error, and response time delay is developed. At the same time, a physical constraint verification mechanism is introduced to reasonably check the prediction results using the seismic wave propagation equation and the law of conservation of energy. In combination with the adversarial sample generation technology, the model's robustness in noisy interference and data anomaly conditions is tested. Finally, a reliability evaluation report is formed, which takes into account both data-driven performance and compliance with physical laws.

[0096] Through physical constraint verification and multi-dimensional evaluation system, it is ensured that the model's prediction results comply with the law of seismic propagation, and the reliability and robustness in engineering application are improved through spatio-temporal cross-validation and adversarial testing.

[0097] The earthquake disaster scenario identification system based on deep learning includes the following modules:

[0098] Data acquisition and preprocessing module: used for integrating seismic waveform, ground deformation, building structure, geographic information, and historical disaster records, etc. multi-source heterogeneous data, and preprocessing the data;

[0099] The deep learning model construction module is used for designing a hybrid neural network architecture, integrating an adaptive attention mechanism to realize cross-modal feature fusion, and constructing a multi-task classifier to synchronously output disaster types, intensity levels, and influence ranges.

[0100] The model training module is used for designing a model training strategy, and performing hyperparameter search by using Bayesian optimization and a genetic algorithm, while introducing a physical constraint verification mechanism to establish a multi-dimensional evaluation system, so as to ensure that the model prediction results meet the earthquake propagation law and engineering actual demand.

[0101] The visual interactive module is used for real-time superimposed display of a focal mechanism sphere, a ground surface deformation field, and a building damage heat map, and provides an intuitive spatio-temporal analysis interface by combining a time axis to trace back key evolution processes before and after the disaster.

[0102] The auxiliary decision module is used for generating an emergency response plan.

[0103] Through the data acquisition and preprocessing module, the deep learning model construction module, the model training module, the visual interactive module, and the auxiliary decision module, full-process automatic management of earthquake disaster data acquisition, feature analysis, model training, and emergency response is realized, the model engineering practicability is ensured by the physical constraint verification and spatio-temporal cross-validation of the model training module, and the disaster situation awareness accuracy and emergency response efficiency are improved by designing the visual interactive module and the auxiliary decision module.

[0104] The model training module is based on a multi-modal heterogeneous data fusion framework, and a multi-level feature extraction system is constructed by integrating a 3D convolutional network, a U-Net, a graph convolutional network, and a bidirectional LSTM through a hybrid architecture neural network. An adaptive attention fusion mechanism is adopted to realize cross-modal alignment and correlation matrix modeling of seismic waveform spatio-temporal features, ground surface deformation spatial features, building group topological features, and disaster evolution time sequence features. A hybrid classification decision system is constructed by combining a deep residual network, a support vector data description, and a conditional random field. The multi-task learning framework is used to synchronously optimize disaster type identification, intensity grading, and influence range prediction tasks. A differentiable feature selection gating unit is used to realize dynamic generation of a 128-dimensional high-discriminative feature vector, and finally an end-to-end earthquake disaster scenario recognition model is formed, which has physical law embedding and spatio-temporal continuity constraints.

[0105] Through the multi-modal heterogeneous data fusion framework and the hybrid neural network architecture, deep feature interactive modeling of seismic waveforms, ground surface deformations, building topologies, and time sequence evolution is realized. High-discriminative features are dynamically generated by combining physical constraints and gating units, and an end-to-end disaster recognition system is formed, which has spatio-temporal continuity and mechanism explainability.

[0106] The auxiliary decision module can optimize the rescue path planning and resource scheduling strategy, output a building reinforcement priority list by comprehensively considering the structure damage index, population density and economic value, and provide quantitative decision support for post-disaster reconstruction.

[0107] By dynamically optimizing the rescue path and quantitatively evaluating the building reinforcement priority, the multi-dimensional indexes of structure damage, population density and economic value are fused to provide scientific decision basis for post-disaster emergency response and reconstruction planning, and the resource scheduling efficiency and the life and property safety guarantee capability are significantly improved.

[0108] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A deep learning-based earthquake disaster scenario recognition method, characterized in that, The specific steps are as follows: S1. Data acquisition and preprocessing; Establish a comprehensive database that includes seismic waveform data, surface deformation data, building structure data, geographic information data, and historical disaster records; S2, Deep learning model construction; S2.1, Hybrid Neural Network Architecture; In the design of multimodal feature extraction networks during the deep learning model building phase, it is necessary to construct a collaborative hybrid neural network architecture to fully exploit the multi-source features of earthquake disaster data; For spatiotemporally continuous seismic waveform data, a three-dimensional convolutional neural network is used for processing. A convolutional kernel with adaptive receptive field adjustment function is designed with a size of 5×5×3. The hierarchical spatiotemporal feature extraction module captures key phase features of P-wave arrival and S-wave amplitude changes. At the same time, a hole convolution is integrated with an expansion rate of 2 to expand the temporal context perception range of the waveform sequence. For high-resolution remote sensing image data, an improved U-Net architecture is constructed. In the encoder part, ResNet-50 is used as the backbone network and a channel attention mechanism is introduced. In the decoder stage, multi-scale feature fusion is implemented. Low-level texture features and high-level semantic features are aggregated across layers through skip connections. A directional feature enhancement module is designed for disaster markers such as building collapse and surface cracks. When processing building complex structural data, a graph convolutional network is used to establish a topological relationship analysis model. Individual buildings are abstracted as graph nodes, and node attributes include structural type, floor height, and construction year parameters. Spatial adjacency relationships are constructed through Voronoi graph as the basis for edge connection. A graph attention mechanism is used to dynamically learn the mechanical interaction weights between building complexes, and a graph pooling operation based on physical constraints is designed to retain key structural vulnerability features. A bidirectional long short-term memory network was constructed to capture the temporal dynamics of disaster evolution. 128 hidden units were set in the time dimension and a hierarchical propagation mechanism was adopted. The forward layer captured the propagation characteristics of the earthquake source rupture process, and the backward layer reversed the disaster chain reaction path. The feature contribution of different time phases was weighted and fused through a time attention mechanism. S2.2 Feature fusion; S2.3, Scene Recognition Classifier; A 128-dimensional high-density feature vector generated by adaptive feature fusion is used to construct a hybrid classification decision system, which achieves joint reasoning of disaster elements through a multi-task learning framework. First, a deep residual network is adopted as the backbone architecture. Skip connections are used to alleviate the gradient vanishing problem. Global average pooling layers are used to map the fused features to the disaster level space to achieve initial classification of intensity from VI to XI. For abnormal disaster patterns, such as earthquake swarms and induced landslides, support vector data description is introduced to construct a hyperspherical decision boundary. Kernel function mapping is used to detect dangerous scenarios that deviate from the normal distribution. A conditional random field post-processing module is designed to encode the spatial continuity constraints of building damage degree and surface rupture zone orientation into energy functions. Graph cut algorithm is used to optimize the regional consistency of classification results. Finally, a multi-task learning framework is built. Through the collaborative design of shared low-level features and independent task heads, the triplet identification results of disaster type, intensity level, and impact range are output synchronously. The confidence of each prediction result is evaluated by an uncertainty quantification module, forming a decision output that combines physical interpretability and engineering practicality. S3, Model Training and Optimization; S4, Context Identification and Decision Support Stage.

2. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: A multi-layered, composite data augmentation system needs to be constructed. First, based on the conditional generative adversarial network framework, by inputting historical earthquake parameters as constraints and combining potential spatial feature decoupling technology, rare disaster scenario samples with physical rationality are synthesized, including high-intensity earthquakes with low probability of occurrence and scarce cases of secondary disasters in special geological structures.

3. The earthquake disaster scenario recognition method based on deep learning according to claim 2, characterized in that: After the disaster scenario sample is generated, spatial augmentation processing is performed on the multi-source observation data. Geometric deformation operations such as random rotation, translation, and affine transformation are adopted, and Gaussian noise, random occlusion, color perturbation, and texture perturbation strategies are introduced to enhance the robustness of the model to sensor errors and on-site environmental interference.

4. The earthquake disaster scene recognition method based on deep learning according to claim 3, characterized in that: The random rotation angle range is ±30°, the maximum translation offset is 20%, the signal-to-noise ratio of Gaussian noise is controlled at 30-50dB, the occlusion ratio of random occlusion is 15%-40%, and the HSV space of color perturbation is adjusted by ±10%.

5. The earthquake disaster scene recognition method based on deep learning according to claim 1, characterized in that: Further, an improved Mixup algorithm was applied to implement cross-category data fusion. By adjusting the linear interpolation coefficient λ, a smooth transition processing was performed on samples of adjacent intensity levels in the feature space to construct transitional training samples with continuous disaster intensity changes. At the same time, a three-dimensional spatiotemporal data cube was constructed, stacking continuous seismic phase waveform data along the time dimension and integrating kilometer-level gridded surface deformation fields in the spatial dimension. Spatiotemporal joint features were extracted through three-dimensional convolution kernels, and data resampling technology was used to extend the time series length to 2-3 times that of the original data. Finally, a comprehensive data augmentation scheme covering geometric transformation, noise injection, feature mixing, and spatiotemporal extension was formed.

6. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: By establishing a multi-dimensional evaluation system that includes physical constraint verification, a comprehensive multi-dimensional evaluation system is constructed. The dataset is partitioned according to the epicenter location and occurrence time through a spatiotemporal cross-validation strategy. A quantitative indicator system including disaster identification accuracy, false alarm rate, spatial positioning error, and response delay is formulated. At the same time, a physical constraint verification mechanism is introduced, and the rationality of the prediction results is checked by using the seismic wave propagation equation and the law of energy conservation. Furthermore, the robustness of the model under noise interference and data anomalies is tested by combining adversarial example generation technology. Finally, a credibility evaluation report that takes into account both data-driven performance and conformity to physical laws is formed.

7. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: The data acquisition and preprocessing process specifically includes raw P-wave and S-wave waveform data recorded by seismic networks, with a sampling frequency of not less than 100Hz; millimeter-level surface deformation data acquired by InSAR satellite remote sensing; and high-resolution disaster site image data taken by UAVs. The data acquisition and preprocessing process includes structural parameter data from the urban building information model; historical earthquake intensity distribution maps and disaster damage assessment reports. Preprocessing, particularly data standardization, requires the construction of a unified data processing framework with spatiotemporal consistency. First, the WGS84 global geodetic system is used to accurately register all geospatial data. Coordinate transformation algorithms are used to eliminate geographic benchmark biases caused by different sensors and observation platforms. Based on this, a global time synchronization mechanism based on Coordinated Universal Time (UTC) is established. Millisecond-level time alignment is implemented for heterogeneous time-series data from seismic waveform recorders, satellite transit times, and UAV acquisition times, constructing a data association system under a continuous spatiotemporal coordinate system. For unstructured data such as high-resolution aerial images and disaster site text reports, image vectorization algorithms are used to convert raster data into vector layers with clear topological relationships. Simultaneously, natural language processing techniques are used to extract key semantic features from the text and encode them into structured vectors. Finally, multi-dimensional data normalization eliminates differences in physical dimensions. Min-Max standardization and Z-score standardization methods are used to unify the scale of seismic waveform amplitude, surface deformation values, and numerical data of building structural parameters, respectively. One-hot encoding and embedding representation are implemented for categorical variables.

8. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: After the initial extraction of multimodal features, the S2.2 feature fusion process requires an adaptive attention mechanism to integrate deep features from multi-source heterogeneous data. First, a cross-modal feature alignment layer eliminates the heterogeneity differences between seismic waveforms, remote sensing images, and building structure data. A dynamic time warping algorithm aligns temporal features at different sampling rates, and a feature projection matrix is ​​constructed to achieve semantic matching in high-dimensional space. Based on this, a multi-head self-attention mechanism is introduced to establish an intermodal correlation matrix. Parallel computation captures the interaction between spatial deformation fields and building topology, and the causal relationship between the seismic source rupture process and surface response—multi-dimensional dependency patterns. Further, a differentiable feature selection gating unit is developed. Based on a gated recurrent network, the contribution of each modal feature is dynamically evaluated, redundant features are softly masked, and a channel attention weighting strategy is used to focus on key disaster identification features. Finally, a hierarchical feature aggregation module maps multi-scale features to a unified semantic space. Depthwise separable convolution is used to compress feature dimensions, and residual connections are combined to retain original information, generating a 128-dimensional fused feature vector.

9. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: The S3 model training and optimization includes training strategy design and hyperparameter optimization. The training strategy design involves: first, in the model training phase, unfreezing network parameters in stages, prioritizing the training of the feature extraction layer to capture basic patterns, and then jointly optimizing the classifier parameters; introducing a course learning mechanism to gradually transition from single-source earthquake mechanism samples to multi-type composite disaster scenarios; simultaneously implementing adversarial training to enhance the model's robustness to noise interference by generating adversarial perturbation samples; and designing a dynamically weighted cross-entropy loss function to address the long-tail distribution characteristics of earthquake disaster data, adaptively adjusting weight coefficients based on class sample frequency to balance the class imbalance problem in the mainshock and aftershock identification task. After completing the basic architecture design of the training strategy, the improvement of model performance will depend on the accurate exploration of the hyperparameter space. To this end, an intelligent hyperparameter tuning system is constructed to achieve the globally optimal configuration of model parameters through multi-dimensional optimization strategies. A Bayesian optimization algorithm is used to search for learning rate combinations, and a neural architecture search is combined to determine the Pareto optimal solution for network depth and convolutional kernel size. A genetic algorithm is used to optimize the batch size and L2 regularization coefficient, and a parameter set resistant to overfitting is selected through multi-generation population evolution. A hyperparameter response surface analysis model is constructed to quantify the impact of parameter interactions on recognition accuracy and inference speed, and finally determine the globally optimal configuration that balances computational efficiency and classification performance.

10. The earthquake disaster scene recognition method based on deep learning according to claim 1, characterized in that: The S4 scenario recognition and decision support stage includes a real-time inference engine, a visualization interaction platform, and auxiliary decision support. The real-time inference engine is designed based on an edge computing inference system: it converts floating-point models to 8-bit integers using mixed-precision quantization technology, enabling FPGA hardware acceleration deployment; it constructs a streaming data processing pipeline to support parallel access and online updates of multi-source sensor data; it integrates a lightweight inference engine to ensure sub-second response latency, and simultaneously develops an uncertainty quantization module, using the Monte Carlo Dropout method to output prediction confidence. The visualization interaction platform constructs a three-dimensional disaster situation awareness system, employing WebGL technology for real-time rendering on the browser side, supporting dynamic overlay display of the seismic source mechanism sphere, surface deformation field, and building damage heatmap, and can trace back the critical evolution process from 10 seconds before to 30 minutes after the disaster. The auxiliary decision support is used to establish an intelligent auxiliary decision chain, automatically generating emergency response plans based on scenario recognition results, optimizing rescue route planning by combining real-time road conditions and aftershock probability prediction; and simultaneously using an LSTM network to predict the probability and spatiotemporal distribution of derivative risks of landslides and barrier lakes, providing early warning of secondary disasters. Output a priority list for building reinforcement, and rank them by weighting the comprehensive structural damage index, population density and economic value to provide a quantitative basis for post-disaster reconstruction decisions.

11. A deep learning-based earthquake disaster scene recognition system, the system being used to implement the deep learning-based earthquake disaster scene recognition method according to any one of claims 1 to 10, characterized in that, Includes the following modules: Data acquisition and preprocessing module: used to integrate multi-source heterogeneous data such as seismic waveforms, surface deformation, building structures, geographic information and historical disaster records, and to preprocess the data; Deep learning model building module: used to design hybrid neural network architecture, integrate adaptive attention mechanism to achieve cross-modal feature fusion, and build multi-task classifier to synchronously output disaster type, intensity level and impact range; Model training module: Used to design model training strategies and use Bayesian optimization and genetic algorithms to search for hyperparameters. At the same time, a physical constraint verification mechanism is introduced to establish a multi-dimensional evaluation system to ensure that the model prediction results conform to the earthquake propagation law and the actual engineering needs. Visualization and Interactive Module: Used for real-time overlay display of the focal mechanism sphere, surface deformation field and building damage heat map, combined with the timeline to trace the key evolution process before and after the disaster, providing an intuitive spatiotemporal analysis interface; Decision Support Module: Used to generate emergency response plans.

12. The earthquake disaster scene recognition system based on deep learning according to claim 11, characterized in that: The model training module is based on a multimodal heterogeneous data fusion framework. It integrates 3D convolutional networks, U-Net, graph convolutional networks, and bidirectional LSTM to construct a multi-level feature extraction system through a hybrid neural network architecture. It adopts an adaptive attention fusion mechanism to achieve cross-modal alignment and correlation matrix modeling of seismic waveform spatiotemporal features, surface deformation spatial features, building cluster topological features, and disaster evolution temporal features. It combines deep residual networks, support vector data description, and conditional random fields to construct a hybrid classification decision system. It uses a multi-task learning framework to simultaneously optimize disaster type identification, intensity classification, and impact range prediction tasks. It achieves dynamic generation of 128-dimensional high-discriminative feature vectors through differentiable feature selection gating units. Finally, it forms an end-to-end earthquake disaster scenario recognition model with physical law embedding and spatiotemporal continuity constraints.

13. The earthquake disaster scene recognition system based on deep learning according to claim 11, characterized in that: The auxiliary decision-making module can optimize rescue route planning and resource allocation strategies, and output a priority list of building reinforcement based on the comprehensive structural damage index, population density and economic value, providing quantitative decision support for post-disaster reconstruction.

Images