Seismic motion parameter prediction method based on multi-modal deep learning and physical characteristic parameters

Abstract

The application relates to the technical field of earthquake parameter prediction, in particular to a seismic motion parameter prediction method based on multi-modal deep learning and physical characteristic parameters, which comprises the following steps: obtaining a multi-modal physical characteristic set of a target earthquake event; performing consistent constraint transformation on the multi-modal physical characteristic set based on a seismic motion physical law to obtain an embedding characteristic set; identifying a nonlinear coupling relationship between different physical characteristic subsets according to the embedding characteristic set to obtain a global seismic motion representation vector; inputting the global seismic motion representation vector into a multi-task prediction network to output an initial joint prediction result; performing multi-dimensional physical law conformity checking on the initial joint prediction result based on real-time scene attributes of the target earthquake event, correcting prediction components that do not conform to a dynamic constraint threshold, and outputting a final prediction result. Through the application, the seismic motion parameter prediction method can simultaneously have physical consistency and high-precision nonlinear fitting capability.

Description

Technical Field

[0001] This invention relates to the technical field of earthquake parameter prediction, and in particular to a method for predicting seismic motion parameters based on multimodal deep learning and physical characteristic parameters. Background Technology

[0002] Seismic ground motion parameter prediction is a crucial foundational task in earthquake engineering and seismic design, as its results directly impact the seismic fortification safety of engineering structures and the early warning effectiveness of earthquake early warning systems. Existing methods for predicting seismic ground motion parameters mainly fall into two categories: one is based on physical statistical models of seismic ground motion attenuation relationships. These models establish empirical relationships between parameters such as magnitude, distance, and site conditions and seismic ground motion intensity through regression analysis, possessing clear physical meaning and good extrapolation capabilities. The other category is based on data-driven models using deep learning. These models automatically learn the nonlinear mapping relationship between seismic waveform data and seismic ground motion parameters through neural networks, exhibiting powerful capabilities for fitting complex patterns.

[0003] However, physical statistics-based models, limited by their regression function forms, struggle to fully characterize the high-order nonlinear coupling relationships between physical mechanisms such as the source rupture process, wave propagation path complexity, and site nonlinear response. While data-driven deep learning models possess powerful fitting capabilities, their prediction process lacks effective constraints from physical laws, potentially leading to prediction results that violate fundamental seismic energy attenuation laws or the physical consistency requirements between parameters.

[0004] Therefore, there is an urgent need for a method for predicting seismic motion parameters that can simultaneously possess physical consistency and high-precision nonlinear fitting capabilities. Summary of the Invention

[0005] This invention provides a method for predicting seismic motion parameters based on multimodal deep learning and physical feature parameters, which can effectively solve the problems in the background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Seismic motion parameter prediction methods based on multimodal deep learning and physical feature parameters include: Obtain the multimodal physical feature set of the target seismic event, including the source physical feature subset, path physical feature subset, site physical feature subset, and time-series feature subset of the seismic waveform; Based on the physical laws of seismic motion, a consistency constraint transformation is performed on the multimodal physical feature set to obtain an embedded feature set; Based on the embedded feature set, the nonlinear coupling relationship between different physical feature subsets is identified to obtain the global seismic motion characterization vector; The global ground motion characterization vector is input into a multi-task prediction network, and the initial joint prediction result is output. Based on the real-time scene attributes of the target seismic event, the corresponding dynamic constraint threshold is matched from the preset physical consistency threshold library, and the initial joint prediction results are checked for compliance with physical laws in multiple dimensions. Prediction components that do not comply with the dynamic constraint threshold are corrected, and the final prediction results are output.

[0007] Furthermore, the subset of source physical characteristics includes at least one of magnitude, source depth, source mechanism solution, and stress drop; The path physical feature subset includes at least one of fault distance, average shear wave velocity along the propagation path, and crustal medium quality factor. The subset of site physical characteristics includes at least one of site equivalent shear wave velocity, site overburden thickness, and site category. The temporal feature subset of the earthquake waveform includes three-component acceleration time history data within a preset time window after the earthquake P-wave is triggered.

[0008] Furthermore, based on the physical laws of seismic motion, a consistency constraint transformation is performed on the multimodal physical feature set, including: Based on the preset ground motion attenuation model, the reference ground motion intensity value corresponding to the multimodal physical feature set is calculated; Using the reference ground motion intensity value as a physical constraint, a nonlinear transformation is performed on the multimodal physical feature set, so that the transformed embedded feature set carries the physical information of the reference ground motion intensity value.

[0009] Furthermore, the nonlinear coupling relationships between different subsets of physical features are identified, including: By using multiple parallel modal-specific feature extraction branches, feature extraction is performed on the source physical feature subset, the path physical feature subset, the site physical feature subset, and the temporal feature subset of the seismic waveform, respectively, to obtain high-dimensional features for each mode; By utilizing a cross-modal attention fusion mechanism, the cross-attention weights between the high-dimensional features of each modality are calculated, and the high-dimensional features of each modality are fused based on the cross-attention weights to generate the global seismic motion characterization vector.

[0010] Furthermore, a multilayer perceptron network is used to extract features from the subset of physical features of the earthquake source and / or the subset of physical features of the path. For the soil profile data in the aforementioned subset of site physical features, a graph neural network is used for feature extraction. For the temporal feature subset of the earthquake waveform, a recurrent neural network or a convolutional neural network is used for feature extraction.

[0011] Furthermore, the multi-task prediction network includes a shared feature layer and multiple parallel task-specific output layers; The shared feature layer is used to perform nonlinear mapping on the global seismic motion characterization vector and extract multi-task shared features; Each of the task-specific output layers corresponds to a seismic motion target parameter, which is used to output the corresponding predicted value based on the multi-task shared features. The target parameters for ground motion include at least two of the following: peak ground acceleration, peak velocity, and response spectrum.

[0012] Furthermore, the multi-task prediction network is trained using a composite loss function, which includes: The multi-task prediction error term measures the deviation between the predicted value and the true value of the dedicated output layer for each task. A physical regularization term is used to constrain the global ground motion characterization vector and the multi-task shared feature to conform to a preset physical law of ground motion attenuation. Structural risk term, used to control the complexity of the network model.

[0013] Furthermore, the construction process of the physical consistency threshold library includes: Historical earthquake events are categorized into scenarios based on magnitude range, distance range, and site type. For each scenario, based on the statistical distribution of historical strong earthquake observation data, the maximum physical reasonable threshold, minimum physical reasonable threshold, and correlation constraint threshold between parameters are determined for each region. For each scenario, a dynamic adaptation factor based on real-time source parameters and real-time path parameters is configured to generate the dynamic constraint threshold.

[0014] Furthermore, the multi-dimensional physical law conformity verification includes at least one of the following verification rules: Amplitude compliance verification is used to determine whether each predicted component is within the preset amplitude threshold range; The attenuation law consistency check is used to determine whether the attenuation trend of the predicted component conforms to the preset attenuation rate threshold. Energy conservation check is used to determine whether the energy relationship between each predicted component meets the preset energy deviation threshold. Parameter correlation verification is used to determine whether the correlation coefficients between each predicted component conform to the preset correlation range.

[0015] Further, based on a preset seismic attenuation model, the reference seismic intensity value corresponding to the multimodal physical feature set is calculated, including: The ground motion intensity attenuation model in the field of earthquake resistance engineering is selected as the basic model. Based on the magnitude, fault distance and site category of the multimodal physical feature set, the attenuation model correction parameters verified by historical strong earthquake observation data under the corresponding earthquake scenario are retrieved. The coefficients of the basic model are adapted to the scenario to obtain a special ground motion intensity attenuation model that matches the target earthquake event. The magnitude, fault distance, site type, and equivalent shear wave velocity of the multimodal physical feature set are input into the dedicated ground motion intensity attenuation model to calculate the reference peak ground acceleration and reference peak velocity corresponding to the target seismic event, which are used as the reference ground motion intensity value. The reference ground motion intensity value is normalized so that its numerical range matches the parameter numerical range of the multimodal physical feature set.

[0016] The technical solution of this invention can achieve the following technical effects: This invention acquires a full-chain multimodal physical feature set covering the source, path, site, and time-series waveforms, and performs a consistency constraint transformation on the multimodal physical feature set based on the physical laws of seismic motion. On the one hand, it fully covers the influencing factors of the entire physical process of seismic motion from source energy release and medium propagation to site response, overcoming the shortcomings of existing physical statistical models that are limited by regression function forms and cannot characterize the high-order nonlinear coupling relationship between multiple physical factors. On the other hand, through the pre-conformity constraint transformation, the feasible domain of the basic physical laws of seismic motion is anchored from the feature input stage, so that the model always performs feature learning and mapping within the framework of seismic motion physical mechanism. This avoids the problem that pure data-driven models are prone to learning false numerical correlations and the separation between feature physical meaning and fitting ability, and achieves a synergistic balance between feature physical meaning anchoring and nonlinear fitting ability.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0018] 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 only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] like Figure 1 As shown, the seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters of the present invention specifically includes the following steps: Step S1: Obtain the multimodal physical feature set of the target seismic event, including the source physical feature subset, path physical feature subset, site physical feature subset, and time-series feature subset of the seismic waveform; Step S2: Based on the physical laws of seismic motion, perform consistency constraint transformation on the multimodal physical feature set to obtain an embedded feature set with enhanced physical consistency; Step S3: Based on the embedded feature set, identify the nonlinear coupling relationship between different physical feature subsets to obtain a global seismic motion characterization vector that integrates the physical information of the entire chain; Step S4: Input the global ground motion characterization vector into the multi-task prediction network and output the initial joint prediction result; Step S5: Based on the real-time scene attributes of the target seismic event, match the corresponding dynamic constraint threshold from the preset physical consistency threshold library, perform multi-dimensional physical law conformity verification on the initial joint prediction results, correct the prediction components that do not conform to the dynamic constraint threshold, and output the final prediction results.

[0023] In this embodiment, the present invention acquires a full-chain multimodal physical feature set covering the source, path, site, and time-series waveforms, and performs a consistency constraint transformation on the multimodal physical feature set based on the physical laws of seismic motion. On the one hand, it fully covers the influencing factors of the entire physical process of seismic motion from source energy release and medium propagation to site response, overcoming the shortcomings of existing physical statistical models that are limited by regression function forms and cannot characterize the high-order nonlinear coupling relationship between multiple physical factors. On the other hand, through the pre-conformity constraint transformation, the feasible domain of the basic physical laws of seismic motion is anchored from the feature input stage, so that the model always performs feature learning and mapping within the framework of seismic motion physical mechanism. This avoids the problems of pure data-driven models easily learning false numerical correlations and the separation between feature physical meaning and fitting ability, and achieves a synergistic balance between feature physical meaning anchoring and nonlinear fitting ability. This invention deeply embeds the physical laws of seismic motion into the entire deep learning prediction process, from input layer physical constraint transformation and feature layer nonlinear coupling relationship identification to output layer dynamic physical verification and correction. Rather than simply superimposing physical constraints and model fitting, it overcomes the inherent contradiction in existing technologies where physical constraints and model fitting capabilities are mutually exclusive, and where increasing physical constraints easily leads to a loss of prediction accuracy while relaxing constraints easily violates physical laws. The physical constraints throughout the entire process define a physically reasonable boundary for the model's learning and inference, enabling the model to eliminate noise interference that violates physical mechanisms and learn the real nonlinear coupling relationships between multiple physical factors. While ensuring the physical consistency of the prediction results, it also improves the model's generalization ability in untrained scenarios, achieving a two-way synchronous improvement in prediction accuracy and physical consistency. By identifying the nonlinear coupling relationship of the feature set embedded with physical constraints, a global seismic motion characterization vector integrating the entire chain of physical information is generated. Based on this homogeneous characterization vector, a multi-task prediction network is used to jointly predict multi-dimensional seismic motion parameters, overcoming the shortcomings of existing technologies where single-parameter independent regression severs the inherent physical correlation of parameters and multi-output models cannot guarantee the physical consistency between parameters. By pre-setting a physical consistency threshold library for different scenarios, dynamic constraint thresholds are matched based on the real-time scenario attributes of the target seismic event. The initial prediction results are then checked for compliance with physical laws in multiple dimensions and targeted corrections are made. This overcomes the shortcomings of existing technologies that use fixed constraint thresholds, which cannot adapt to different seismic scenarios and are prone to errors such as overly broad constraints leading to prediction anomalies or overly strict constraints leading to accuracy loss. The dynamically adapted constraint thresholds can provide accurately adapted physical constraint boundaries for seismic scenarios with different magnitudes, distances, and site conditions. Targeted component corrections only adjust prediction components that do not conform to the constraints, fully preserving accurate prediction results that conform to physical laws, and achieving dual protection of physical compliance and prediction accuracy across all scenarios.

[0024] In a specific implementation, as one example, regarding the acquisition of the multimodal physical feature set of the target seismic event in step S1, the acquisition of the multimodal physical feature set of the target seismic event is achieved by establishing a unified spatiotemporal calibration benchmark, setting a dynamic acquisition window adapted to the seismic source, and constructing a feature parameter screening mechanism oriented towards prediction needs. This ensures that the parameters of each feature subset match the requirements for seismic motion parameter prediction. The specific implementation is as follows: Step S11: Establish a unified spatiotemporal calibration benchmark, configure a unique seismic event identifier for the target seismic event, and configure a unique station identifier for the observation station; using the first arrival time of the seismic P wave as the time benchmark and the latitude, longitude, and elevation of the observation station as the spatial benchmark, assign spatiotemporal labels to the collected data of all feature subsets. The assigned information includes the seismic event identifier, station identifier, first arrival time stamp of the P wave, and station spatial coordinates. The collected data of all feature subsets are associated based on this spatiotemporal label. Step S12: Perform directional acquisition and screening of the source physical feature subset. Based on the seismic network's real-time source inversion system, obtain the magnitude, focal depth, focal mechanism solution, and original stress drop parameters of the target earthquake event. Construct source feature screening rules, which retain parameters directly related to ground motion intensity and spectral characteristics to form a source physical feature subset. The parameters within the subset are at least one of magnitude, focal depth, focal mechanism solution, and stress drop. Step S13: Targeted acquisition and screening of physical feature subsets of the propagation path. Based on the geographic information system and seismic wave propagation model, calculate the fault distance from the epicenter of the target earthquake event to the observation station, extract the shear wave velocity profile of the strata along the propagation path and calculate the average shear wave velocity, retrieve the crustal medium quality factor of the corresponding propagation path through the crustal medium quality factor database, and construct path feature screening rules. The screening rules are to retain parameters directly related to seismic wave propagation attenuation to form a physical feature subset of the path. The parameters in the subset are at least one of fault distance, average shear wave velocity of the propagation path, and crustal medium quality factor. Step S14: Perform directional acquisition and screening of site physical feature subsets. Based on the site geological survey database, retrieve the site 30m equivalent shear wave velocity and site overburden thickness at the observation station. Determine the site category according to the seismic design code for buildings. Construct site feature screening rules. The screening rules are to retain parameters directly related to the site's seismic amplification effect to form a site physical feature subset. The parameters in the subset are at least one of the site equivalent shear wave velocity, site overburden thickness, and site category. Step S15: Set a dynamic adaptation acquisition window and execute the acquisition of the seismic waveform time series feature subset. Based on the preliminary magnitude results of the source physical feature subset, set a dynamic acquisition window after P-wave triggering. For example, the acquisition window is 2s when the magnitude is < 5.0, 3s when the magnitude is 5.0 ≤ magnitude < 6.0, and 4s when the magnitude is ≥ 6.0. The north-south, east-west, and vertical acceleration time history data within this dynamic window are acquired through the three-component acceleration sensors of the seismic station to form the seismic waveform time series feature subset. Step S16: Complete the integration of the multimodal physical feature set. The source physical feature subset, path physical feature subset, site physical feature subset, and temporal feature subset of the seismic waveform under the same seismic event identifier and station identifier are associated and integrated to form the multimodal physical feature set of the target seismic event. The integrated feature set carries a unified spatiotemporal label and the filtered feature parameters.

[0025] In this embodiment, a unified spatiotemporal calibration benchmark and dual-identifier association mechanism are used to ensure accurate spatiotemporal correspondence of physical feature subsets under the same seismic event and the same observation station, eliminating spatiotemporal deviations between different feature subsets. The dynamic acquisition window of waveform time series features is adjusted according to the source magnitude to adapt to the evolution law of ground motion waveforms corresponding to the source characteristics, avoiding the loss of effective features or the introduction of redundant data caused by fixed time windows. The parameter screening of each feature subset is carried out based on the core requirements of ground motion parameter prediction, eliminating redundant parameters that are not significantly related to ground motion intensity, propagation attenuation, and site amplification, reducing the computational load of subsequent feature transformation and nonlinear coupling relationship identification, and improving the efficiency of feature processing. The parameters of each feature subset after directional acquisition and screening are directly related to the elements of ground motion parameter prediction, making the multimodal physical feature set highly consistent with the requirements of ground motion parameter prediction.

[0026] In a specific implementation, as one example, in step S2, by pre-setting a ground motion attenuation model and completing scenario-based parameter adaptation, a benchmark ground motion intensity value matching the target seismic event is calculated. This benchmark value is then incorporated into the nonlinear transformation process as a physical constraint condition to perform a constraint transformation on the multimodal physical feature set. This ensures that the resulting embedded feature set retains both the nonlinear information of the original features and carries the physical information of ground motion attenuation, thereby enhancing the physical consistency of the feature set. Specifically, as follows: Step S21: Preset the ground motion intensity attenuation model and complete the scenario-based parameter adaptation. Select a ground motion intensity attenuation model suitable for the field of earthquake engineering as the basic model. The model input parameters include magnitude, fault distance, and site category. The output parameters are peak ground acceleration and peak ground velocity ground motion intensity values. Based on the unified spatiotemporal label carried by the multimodal physical feature set in step S1 and the magnitude of the source physical feature subset, the fault distance of the path physical feature subset, and the site category of the site physical feature subset, retrieve the attenuation model correction parameters verified by historical strong earthquake observation data under this type of earthquake scenario, and perform scenario-based adaptation on the coefficients of the basic model to obtain a specific ground motion intensity attenuation model that matches the target earthquake event. Step S22: Input feature parameters to calculate the benchmark ground motion intensity value. Input the magnitude of the source physical feature subset, the fault distance of the path physical feature subset, the site category and equivalent shear wave velocity of the site physical feature subset from step S1 into the adapted dedicated ground motion intensity attenuation model. Through model iterative calculation, obtain the benchmark peak ground acceleration and benchmark peak velocity corresponding to the target earthquake event, which are used as the benchmark ground motion intensity value for this physical consistency constraint transformation. Normalize the benchmark ground motion intensity value to match the numerical range of the parameters of the multimodal physical feature set. Step S23: Construct a nonlinear transformation function with physical constraints. Based on the radial basis function, construct a nonlinear transformation fundamental function. Incorporate the normalized reference ground motion intensity value as a physical constraint term into the fundamental function. The constraint term is coupled with the output value of the fundamental function in a weighted form. The weight coefficients are determined according to the correlation between each physical feature subset and the ground motion intensity. The coupling weight coefficients of the source and path physical feature subsets are higher than those of the site and waveform time series physical feature subsets. The constructed nonlinear transformation function satisfies the physical correlation constraint between the transformed feature value and the reference ground motion intensity value, and retains the original feature nonlinear information. Step S24: Perform physical consistency constraint transformation by mode. The multimodal physical feature set obtained in step S1 is split into subsets of source, path, site, and seismic waveform time series features. The parameters of each feature subset are input into the constructed nonlinear transformation function. The physical consistency constraint transformation of each feature subset is completed through function operation. During the transformation process, the parameters of each feature subset are coupled with the physical constraint term of the same reference ground motion intensity value, so that the features after each mode transformation carry unified ground motion attenuation physical information. Step S25: Integrate the transformed features and generate an embedded feature set. Integrate the source, path, site, and seismic waveform temporal feature subsets that have undergone constraint transformation. The integration process retains the unified spatiotemporal label assigned in step S1, forming an embedded feature set with enhanced physical consistency. Verify the validity of the parameters of the embedded feature set. The verification standard is that the physical correlation trend between the parameter values ​​and the benchmark ground motion intensity values ​​conforms to the ground motion attenuation law. After verification, it is determined as the final embedded feature set for subsequent identification of nonlinear coupling relationships.

[0027] In this embodiment, by adapting the parameters of the basic ground motion intensity attenuation model to specific scenarios, the calculated benchmark ground motion intensity value is highly matched with the source, path, and site characteristics of the target earthquake event, ensuring the scenario-specificity of the physical constraints. The benchmark ground motion intensity value is incorporated as a physical constraint term into the nonlinear transformation function, anchoring the transformation process of the multimodal physical feature set to the core physical law of ground motion attenuation. This ensures that the resulting embedded feature set carries explicit ground motion physical information, avoiding the loss of physical meaning caused by purely numerical transformations. Each feature subset completes the constraint transformation based on the same benchmark ground motion intensity value, strengthening... The physical correlation between the source, path, site, and waveform time series modal features was established. A nonlinear transformation function was constructed based on the radial basis function. While incorporating physical constraints, the original nonlinear information of the multimodal physical feature set was preserved, avoiding the loss of effective feature information during the transformation process. Physical correlation trend verification was performed on the embedded feature set to ensure the physical consistency of the embedded feature set. The transformation process was performed by splitting the feature subsets to adapt to the structural characteristics of the multimodal physical feature set, while retaining a unified spatiotemporal label. This formed a connection with the technical solution in step S1, eliminating the need for additional spatiotemporal registration of features and ensuring the overall consistency of the method.

[0028] Furthermore, existing methods for extracting and fusing multimodal ground motion features do not design dedicated feature extraction architectures for the physical differences of the four types of features: source, path, site, and waveform time series. General network structures cannot accurately capture the core physical features within each mode. The cross-modal fusion process does not incorporate the physical prior rules of the entire ground motion chain, and the calculation of cross-attention weights has no physical constraints. It cannot accurately identify the nonlinear physical coupling relationship between source rupture, wave propagation attenuation, site amplification, and waveform evolution. The fused representation vector has problems such as weak physical correlation and incomplete coverage of physical information across the entire chain. To address the aforementioned issues, and considering that the embedded feature set output in step S2 already carries unified physical information about ground motion attenuation and retains the original nonlinear information of each mode, modal-specific feature extraction branches are designed for the physical characteristics of the four modes. High-dimensional physical features within each mode are extracted, and a cross-modal attention fusion mechanism incorporating prior knowledge of ground motion physics is constructed. The cross-attention weights between the high-dimensional features of each mode are calculated to accurately identify the nonlinear coupling relationships between different feature subsets, ultimately generating a global ground motion representation vector that fuses the physical information of the entire ground motion chain. The specific implementation is as follows: Step S31: The architecture configuration of the modal-specific feature extraction branch is as follows: The embedded feature set output in Step S2 is split into four physical feature subsets: source, path, site, and seismic waveform time series. A dedicated feature extraction branch matching the physical characteristics is configured for each subset. For the source and path physical feature subsets, a multilayer perceptron network is configured with three fully connected layers, followed by a batch normalization layer and a ReLU activation function. A channel attention module is connected at the network endpoint. The weight allocation of the channel attention module is adapted to the correlation of the reference ground motion intensity value in Step S2, used to extract high-dimensional features of source rupture characteristics and wave propagation attenuation characteristics. For the soil profile data of the site physical feature subset, a graph neural network is configured to model the soil profile as a topological graph structure, with single soil layers as nodes, wave impedance differences between soil layers as edge weights, and soil shear wave velocity, thickness, and density as node features. High-dimensional features of the nonlinear response of the site soil layers are extracted through two graph convolutional layers. For the seismic waveform time series feature subset, a CNN-BiGRU is configured. The hybrid network has three one-dimensional convolutional layers at the front end to extract local time-domain and frequency-domain features of the waveform, and two bidirectional gated recurrent units at the back end to capture long-range temporal dependence features of the waveform and extract high-dimensional temporal features of waveform evolution. Step S32: Standardization and alignment of high-dimensional features of each modality. The high-dimensional features output from the four dedicated feature extraction branches are subjected to layer normalization to eliminate the differences in numerical distribution of the outputs from different branches. The high-dimensional features of each modality are mapped to the feature space of the same dimension through a linear mapping layer to generate the query vector, key vector and value vector corresponding to each modality. The mapping process retains the unified spatiotemporal label given in step S1 to ensure the spatiotemporal consistency of the features of each modality. Step S33: Construct a cross-modal attention fusion mechanism with physical prior constraints. Based on the chain-like physical evolution law of seismic motion source, path, site, and waveform, set prior weight constraints for cross-modal attention. The basic weight of cross-attention between the source and path modes is higher than that of other mode combinations. Input the query vector, key vector, and value vector of each mode into the multi-head cross-modal attention module to calculate the cross-attention weights between different modal features. Incorporate prior weight constraints during the calculation process to obtain an attention weight matrix that conforms to the physical laws of seismic motion. Based on the attention weight matrix, perform weighted fusion of high-dimensional features of each mode to generate a preliminary fused feature vector. Step S34: Residual optimization and global representation generation of fused features. The preliminary fused feature vector is connected to the residual connection module. The input of the residual connection is the normalized reference ground motion intensity value in step S2, which strengthens the physical consistency of the fused features. The fused features after residual optimization are nonlinearly mapped through a fully connected layer to generate a global ground motion representation vector that integrates the physical information of the source, path, site, and waveform. The physical consistency of the global ground motion representation vector is checked. The check standard is that the Pearson correlation coefficient between the representation vector and the reference ground motion intensity value meets the preset threshold. After passing the check, it is used as the input of the subsequent multi-task prediction network.

[0029] In this embodiment, dedicated feature extraction branches are configured for the physical characteristics of the four types of physical feature subsets to adapt to the structural characteristics and physical connotations of different modal data. A multilayer perceptron network is adapted to extract the intrinsic correlation of discrete physical parameters of the seismic source and path. A graph neural network is adapted to the topological structure and nonlinear response characteristics of the site soil profile. A CNN-BiGRU hybrid network is adapted to extract spatiotemporal features from waveform time-series data, avoiding the loss of effective features caused by a general network structure. The weight allocation of the channel attention module is adapted to the correlation with the benchmark ground motion intensity value, and the prior weight constraints of the chain-like physical evolution law of ground motion are incorporated into the cross-modal attention fusion mechanism. This ensures that the feature extraction and fusion process is always anchored to the physical evolution mechanism of ground motion. The calculation of cross-attention weights is no longer a pure numerical operation, accurately identifying the nonlinear physical coupling relationship between different feature subsets, rather than a false numerical correlation. The standardization and linear mapping process of each modality's high-dimensional features retains a unified spatiotemporal label, consistent with steps S1 and S2. The technical solutions form a complete connection, avoiding fusion deviations caused by spatiotemporal misalignment of different modal characteristics; the residual connection module incorporates the reference ground motion intensity value, further enhancing the physical consistency of the fused global representation vector, so that the global representation vector not only contains the nonlinear characteristic information of each mode, but also fully anchors the core physical law of ground motion attenuation; a physical consistency verification step is set for the global ground motion representation vector to ensure that the output representation vector conforms to the basic physical law of ground motion.

[0030] As one embodiment, in step S4, a multi-task prediction network that fits the physical correlation of ground motion parameters is constructed. A physical anchoring mechanism is embedded in the shared feature layer, and a task-specific output layer is configured for different ground motion target parameters. At the same time, a composite loss function linked to the prior physical benchmark is constructed, so that while fitting the nonlinear mapping relationship, the network always anchors to the basic physical laws of ground motion, and outputs an initial joint prediction result that has both physical consistency and numerical accuracy. The specific method includes: Step S41: Construct the architecture of a multi-task prediction network. The network consists of a physical anchoring shared feature layer with sequentially connected signals and a dedicated output layer for multiple tasks. The physical anchoring shared feature layer has two fully connected layers, each followed by a batch normalization layer and a LeakyReLU activation function. The input to the fully connected layer is the global ground motion representation vector output in step S3, and the hidden layer dimension of the fully connected layer matches the dimension of the global ground motion representation vector. The residual branch of the physical anchoring shared feature layer is connected to the normalized reference ground motion intensity value from step S2. During the feature mapping process, the physical reference ground motion attenuation is anchored, and the multi-task output is generated. Shared features; the multi-branch task-specific output layer sets up parallel output branches that correspond one-to-one with the seismic ground motion target parameters. The seismic ground motion target parameters include at least two of the following: peak ground acceleration (PGA), peak ground velocity (PFR), and response spectrum values. For scalar seismic ground motion target parameters such as PGA and PFR, the output branch sets up two fully connected layers with an output dimension of 1. For sequential seismic ground motion target parameters such as response spectrum values, the output branch sets up one one-dimensional convolutional layer plus two fully connected layers, with the output dimension matching the preset number of response spectrum period points. Each parallel output branch independently completes the prediction of its corresponding seismic ground motion target parameter and outputs the initial joint prediction results. Step S42: Construction of the composite loss function. The composite loss function is composed of a weighted combination of the multi-task prediction error term, the physical regularization term, and the structural risk term, and the formula is L. total =L mse +λ1L phy +λ2L reg L mse For the multi-task prediction error term, L phy L is the physical regularization term. regThe structural risk term uses λ1 and λ2 as preset weight balancing coefficients. The multi-task prediction error term is obtained by weighted summation of the mean square error loss of each task's dedicated output branch. The weight coefficients of each branch are set according to the engineering importance of the corresponding ground motion target parameter. The physical regularization term includes two constraints: the first is a shared feature physical consistency constraint, which constrains the Pearson correlation coefficient between the shared features of multiple tasks and the benchmark ground motion intensity value in step S2 to meet a preset threshold; the second is a multi-parameter physical correlation constraint, which constrains the amplitude ratio and attenuation trend between each output parameter to conform to the empirical physical laws of ground motion. The structural risk term uses an L2 regularization term to control the complexity of the network weights and suppress model overfitting. Step S43: Training and convergence verification of the multi-task prediction network. The global ground motion representation vector corresponding to historical earthquake events is used as the training input, and the measured ground motion target parameters of the corresponding earthquake events are used as the training labels. The constructed composite loss function is used as the global optimization objective, and the AdamW optimizer is employed to complete the iterative training of the network. During training, validation set data is input into the network in preset batches to verify the prediction accuracy and physical consistency of the network. When the validation set loss does not decrease for a preset number of consecutive rounds, an early stopping mechanism is triggered to terminate the training. The convergence of the trained network is verified. The verification criteria are that the prediction error of each task-specific output branch meets the preset accuracy threshold, and the physical correlation between the output parameters conforms to the basic laws of ground motion. After passing the verification, it is determined as the final multi-task prediction network. Step S44: Generation of initial joint prediction results. The global ground motion representation vector output in step S3, which has passed the physical consistency verification, is input into the trained multi-task prediction network. The multi-task shared features are extracted through the physical anchoring shared feature layer, and then each parallel task-specific output branch completes independent prediction. The output contains the initial joint prediction results containing all target ground motion parameters. The initial joint prediction results retain the unified spatiotemporal label given in step S1 and form a unique correspondence with the target earthquake event and observation station.

[0031] This embodiment sets up a residual branch that accesses the benchmark ground motion intensity value in the shared feature layer, ensuring that the nonlinear mapping process of the shared features is always anchored to the physical benchmark of ground motion attenuation, thus avoiding the loss of the entire chain of physical information carried by the global ground motion representation vector during feature mapping. For scalar and sequential ground motion target parameters, suitable task-specific output branches are configured to match the data types and physical characteristics of different parameters, extracting the specific features of the corresponding parameters and avoiding feature extraction bias caused by homogeneous network structures. The physical regularization term of the composite loss function is linked with the benchmark ground motion intensity value and the physical laws of ground motion in the preceding steps, rather than being a generalized numerical constraint. This constrains both the physical consistency of the shared features and the intrinsic physical correlation between multiple output parameters, avoiding physical contradictions between multi-task output parameters at the loss function level. The setting of the structural risk term controls the complexity of network weights, suppresses model overfitting, and improves the network's generalization ability in untrained earthquake scenarios. The multi-task joint prediction architecture allows the network to learn the intrinsic physical correlation between different ground motion parameters while learning the mapping relationship of single parameters, further strengthening the mutual adaptability of the prediction results of each parameter.

[0032] In step S5, firstly, a physical consistency threshold library with scene-specific and real-time parameter linkage is constructed. Then, dynamic constraint thresholds are matched based on the real-time scene attributes of the target seismic event, and a progressive multi-dimensional physical law compliance check is performed. For prediction components that do not meet the constraints, targeted corrections are made. After correction, a second closed-loop check is completed, and finally, the final prediction result of seismic motion parameters that conform to the full-dimensional physical laws is output. The specific implementation is as follows: Step S51: Construct a physical consistency threshold library. First, divide the earthquake events corresponding to historical strong earthquake observation data into scenarios according to magnitude intervals, fault distance intervals, and site categories. Each scenario corresponds to a unique scenario code. For each segmented scenario, based on the statistical distribution of historical strong earthquake observation data within that scenario, determine the maximum and minimum physical reasonable thresholds for ground motion target parameters in that scenario. Simultaneously, based on the measured correlation patterns of ground motion parameters within that scenario, determine the correlation constraint thresholds between different ground motion parameters, the attenuation rate threshold for ground motion intensity varying with distance, and the energy relationship deviation thresholds between ground motion parameters. Configure dynamic adaptation factors based on real-time source parameters and real-time path parameters for the basic thresholds of each scenario. The values ​​of the dynamic adaptation factors match the changing trends of real-time magnitude and real-time fault distance. When the magnitude increases, the upper limit of the amplitude threshold is adjusted synchronously, and when the fault distance increases, the range of the attenuation rate threshold is adjusted synchronously. Finally, a physical consistency threshold library with dynamic adaptation factors is formed, and each scenario code in the threshold library forms a unique mapping relationship with the corresponding dynamic constraint threshold.

[0033] Step S52: Matching real-time scene attributes of the target seismic event and retrieving dynamic constraint thresholds. Based on the magnitude, fault distance, and site category of the target seismic event obtained in step S1, determine the scene code corresponding to the target seismic event, and retrieve the basic threshold corresponding to the scene code from the pre-built physical consistency threshold library. Based on the real-time source parameters and real-time path parameters of the target seismic event, calculate the corresponding dynamic adaptation factor, dynamically adjust the retrieved basic thresholds, and generate dynamic constraint thresholds that perfectly match the real-time attributes of the target seismic event. The dynamic constraint thresholds include four types of constraint rules: amplitude threshold range, attenuation rate threshold range, energy deviation threshold, and parameter correlation range.

[0034] Step S53: Multi-dimensional progressive physical law compliance verification of the initial joint prediction results. The initial joint prediction results output in step S4 are split into independent prediction components according to the seismic motion target parameters, and four types of verification rules are executed sequentially. First, amplitude compliance verification is performed, determining whether each prediction component is within the amplitude threshold range corresponding to the dynamic constraint threshold, and marking prediction components that exceed the range. Second, attenuation law consistency verification is performed, based on the fault distance and magnitude of the target earthquake event, determining whether the attenuation trend of each prediction component conforms to the attenuation rate threshold range corresponding to the dynamic constraint threshold, and marking prediction components that do not conform to the attenuation law. Third, energy conservation verification is performed, based on the inherent energy correlation of peak ground acceleration, Arias intensity, and effective duration, determining whether the energy relationship between each prediction component conforms to the energy deviation threshold corresponding to the dynamic constraint threshold, and marking prediction components that do not conform to the energy constraint. Finally, parameter correlation verification is performed, calculating the correlation coefficient between each prediction component, determining whether the correlation coefficient is within the parameter correlation range corresponding to the dynamic constraint threshold, and marking prediction components that do not conform to the correlation constraint.

[0035] Step S54: Oriented correction and closed-loop verification of non-constrained predicted components. For the predicted components marked during the verification process, the reference ground motion intensity value calculated in step S2 is used as the correction benchmark. Combined with the interval boundary of the dynamic constraint threshold, the marked predicted components are oriented numerically adjusted. During the adjustment process, the original values ​​of unmarked predicted components are retained. For all predicted components after directional correction, a complete multi-dimensional physical law compliance verification is performed again. If all predicted components comply with all rules of the dynamic constraint threshold, the verification is considered passed. If there are still non-constrained predicted components, the directional correction and secondary verification process is repeated until all predicted components pass the full-dimensional verification. The final prediction result output integrates all ground motion parameter predicted components that have passed the full-dimensional closed-loop verification. During the integration process, the unified spatiotemporal label given in step S1 is retained, forming a unique correspondence with the target earthquake event and observation station, generating the final prediction result of ground motion parameters.

[0036] In this embodiment, the constructed physical consistency threshold library is used to classify scenes based on magnitude, fault distance, and site category. The basic threshold is determined by combining the statistical distribution of historical strong earthquake observation data. At the same time, a dynamic adaptation factor linked with real-time source and path parameters is configured to ensure that the final generated dynamic constraint threshold fully matches the real-time scene attributes of the target earthquake event. A progressive multi-dimensional physical law compliance verification is adopted, performing verification in sequence from amplitude, attenuation law, energy conservation, and parameter correlation, covering the core physical laws of ground motion parameters and avoiding the omission of physical constraints caused by single-dimensional verification. For the marked non-constrained prediction components, directional correction is performed, adjusting only the non-constrained values ​​and fully preserving the original prediction results that conform to physical laws, avoiding the loss of effective prediction information caused by full adjustment. At the same time, the benchmark ground motion intensity value in step S2 is used as the correction benchmark, so that the correction process is linked with the physical constraints in the previous step.

[0037] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for predicting seismic motion parameters based on multimodal deep learning and physical feature parameters, characterized in that, include: Obtain the multimodal physical feature set of the target seismic event, including the source physical feature subset, path physical feature subset, site physical feature subset, and time-series feature subset of the seismic waveform; Based on the physical laws of seismic motion, a consistency constraint transformation is performed on the multimodal physical feature set to obtain an embedded feature set; Based on the embedded feature set, the nonlinear coupling relationship between different physical feature subsets is identified to obtain the global seismic motion characterization vector; The global ground motion characterization vector is input into a multi-task prediction network, and the initial joint prediction result is output. Based on the real-time scene attributes of the target seismic event, the corresponding dynamic constraint threshold is matched from the preset physical consistency threshold library, and the initial joint prediction results are checked for compliance with physical laws in multiple dimensions. Prediction components that do not comply with the dynamic constraint threshold are corrected, and the final prediction results are output.

2. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 1, characterized in that, The subset of source physical characteristics includes at least one of magnitude, source depth, source mechanism solution, and stress drop; The path physical feature subset includes at least one of fault distance, average shear wave velocity along the propagation path, and crustal medium quality factor. The subset of site physical characteristics includes at least one of site equivalent shear wave velocity, site overburden thickness, and site category. The temporal feature subset of the earthquake waveform includes three-component acceleration time history data within a preset time window after the earthquake P-wave is triggered.

3. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 2, characterized in that, Based on the physical laws of seismic motion, a consistency constraint transformation is performed on the multimodal physical feature set, including: Based on the preset ground motion attenuation model, the reference ground motion intensity value corresponding to the multimodal physical feature set is calculated; Using the reference ground motion intensity value as a physical constraint, a nonlinear transformation is performed on the multimodal physical feature set, so that the transformed embedded feature set carries the physical information of the reference ground motion intensity value.

4. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 3, characterized in that, Identifying nonlinear coupling relationships between different subsets of physical features, including: By using multiple parallel modal-specific feature extraction branches, feature extraction is performed on the source physical feature subset, the path physical feature subset, the site physical feature subset, and the temporal feature subset of the seismic waveform, respectively, to obtain high-dimensional features for each mode; By utilizing a cross-modal attention fusion mechanism, the cross-attention weights between the high-dimensional features of each modality are calculated, and the high-dimensional features of each modality are fused based on the cross-attention weights to generate the global seismic motion characterization vector.

5. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 4, characterized in that, For the subset of physical features of the earthquake source and / or the subset of physical features of the path, a multilayer perceptron network is used for feature extraction; For the soil profile data in the aforementioned subset of site physical features, a graph neural network is used for feature extraction. For the temporal feature subset of the earthquake waveform, a recurrent neural network or a convolutional neural network is used for feature extraction.

6. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 4, characterized in that, The multi-task prediction network includes a shared feature layer and multiple parallel task-specific output layers; The shared feature layer is used to perform nonlinear mapping on the global seismic motion characterization vector and extract multi-task shared features; Each of the task-specific output layers corresponds to a seismic motion target parameter, which is used to output the corresponding predicted value based on the multi-task shared features. The target parameters for ground motion include at least two of the following: peak ground acceleration, peak velocity, and response spectrum.

7. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 6, characterized in that, The multi-task prediction network is trained using a composite loss function, which includes: The multi-task prediction error term measures the deviation between the predicted value and the true value of the dedicated output layer for each task. A physical regularization term is used to constrain the global ground motion characterization vector and the multi-task shared feature to conform to a preset physical law of ground motion attenuation. Structural risk term, used to control the complexity of the network model.

8. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 7, characterized in that, The construction process of the physical consistency threshold library includes: Historical earthquake events are categorized into scenarios based on magnitude range, distance range, and site type. For each scenario, based on the statistical distribution of historical strong earthquake observation data, the maximum physical reasonable threshold, minimum physical reasonable threshold, and correlation constraint threshold between parameters are determined for each region. For each scenario, a dynamic adaptation factor based on real-time source parameters and real-time path parameters is configured to generate the dynamic constraint threshold.

9. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 8, characterized in that, The multi-dimensional physical law conformity verification includes at least one of the following verification rules: Amplitude compliance verification is used to determine whether each predicted component is within the preset amplitude threshold range; The attenuation law consistency check is used to determine whether the attenuation trend of the predicted component conforms to the preset attenuation rate threshold. Energy conservation check is used to determine whether the energy relationship between each predicted component meets the preset energy deviation threshold. Parameter correlation verification is used to determine whether the correlation coefficients between each predicted component conform to the preset correlation range.

10. The seismic motion parameter prediction method based on multimodal deep learning and physical feature parameters according to claim 3, characterized in that, Based on a preset seismic attenuation model, the reference seismic intensity value corresponding to the multimodal physical feature set is calculated, including: The ground motion intensity attenuation model in the field of earthquake resistance engineering is selected as the basic model. Based on the magnitude, fault distance and site category of the multimodal physical feature set, the attenuation model correction parameters verified by historical strong earthquake observation data under the corresponding earthquake scenario are retrieved. The coefficients of the basic model are adapted to the scenario to obtain a special ground motion intensity attenuation model that matches the target earthquake event. The magnitude, fault distance, site type, and equivalent shear wave velocity of the multimodal physical feature set are input into the dedicated ground motion intensity attenuation model to calculate the reference peak ground acceleration and reference peak velocity corresponding to the target seismic event, which are used as the reference ground motion intensity value. The reference ground motion intensity value is normalized so that its numerical range matches the parameter numerical range of the multimodal physical feature set.

Images