Seismic oscillation time-history field generation method and device based on three-component broadband simulation data

Abstract

The invention discloses a seismic oscillation time history field generation method based on three-component broadband simulation data, and the method comprises the steps: collecting three-component seismic oscillation time history data recorded by a station, generating the three-component seismic oscillation time history data of a target field point through a simulation method, and constructing a three-component seismic oscillation time history data set through the combination of the recorded data and the simulation generated data; constructing a spatial-temporal feature extraction module based on a ConvLSTM3D neural network, wherein the spatial-temporal feature extraction module is used for extracting spatial-temporal features in the seismic oscillation time history data; constructing a seismic oscillation time history space-time prediction model based on a space-time feature extraction module and an LSTM neural network; training a seismic oscillation time-history space-time prediction model by adopting the three-component seismic oscillation time-history data set; and generating a seismic oscillation time-history characteristic time sequence of the sparse region target field point through the trained seismic oscillation time-history space-time prediction model, and further generating a sparse region seismic oscillation time-history field model. Limited stations are fully utilized to record data, and the three-component seismic oscillation time history of a large-scale field point is quickly generated.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data generation, and in particular to a method and device for generating a seismic motion time history field based on three-component broadband simulation data. Background Art

[0002] In recent decades, numerous catastrophic earthquakes worldwide have demonstrated the critical importance of seismic time-history data for the seismic design of engineering structures and disaster emergency response. However, due to the sparse distribution of seismic stations, areas with insufficient network density lack sufficient seismic motion data, making it difficult to respond effectively to earthquakes. Given that my country still has a number of areas with sparse seismic stations, research is underway to address the rapid generation of seismic time histories in these areas during earthquakes. Scientifically predicting seismic time-history data at unrecorded locations is crucial for improving the seismic safety of engineering structures and enhancing disaster response capabilities.

[0003] Traditionally, predicting seismic motion data at unrecorded sites relies primarily on numerical simulation methods. While these physics-based numerical simulation methods (such as the finite difference method and the finite element method) can accurately reproduce the propagation characteristics of seismic wave fields, they are computationally expensive, time-consuming, and involve complex parameter determination, making them difficult to rapidly apply across large regions. Therefore, there is an urgent need to explore more efficient and universally applicable alternative methods to meet the needs of rapid prediction at large sites.

[0004] In recent years, deep learning methods have shown great potential in the field of earthquake time history prediction. Deep learning methods are a data-driven modeling tool that can extract complex spatiotemporal features from a large amount of historical data and establish a nonlinear mapping relationship between input and output. Compared with traditional numerical simulation methods, deep learning methods have the advantage of rapid prediction and can generate earthquake time history data for large-scale sites in a short period of time, significantly improving prediction efficiency. Deep learning methods commonly used to process time history data include convolutional neural networks, recurrent neural networks, long short-term memory networks, convolutional long short-term neural networks, etc. Among them, convolutional long short-term neural networks combine the spatial feature extraction capabilities of convolutional neural networks and the time series modeling capabilities of long short-term memory networks. They can simultaneously capture the spatial correlation and temporal dynamic changes of acceleration time history data, and are particularly suitable for processing earthquake time history data with spatiotemporal characteristics.

[0005] However, due to the sparseness of stations, deep learning models often find it difficult to obtain sufficient training data, which limits their performance. Summary of the Invention

[0006] In order to solve the technical problems existing in the known technology, the present invention provides a method and device for generating a seismic time history field based on three-component broadband simulation data.

[0007] The technical solution adopted by the present invention to solve the technical problems existing in the known technology is:

[0008] A method for generating earthquake time history fields based on three-component broadband simulation data comprises the following steps: selecting an area with sparse earthquake stations and determining the spatial distribution positions of target points in the area without station records; collecting three-component earthquake time history data recorded by stations, and generating three-component earthquake time history data of target points using a simulation method, and constructing a three-component earthquake time history dataset by combining the recorded data and the simulated data; constructing a spatiotemporal feature extraction module based on a ConvLSTM3D neural network to extract spatiotemporal features reflecting spatial correlations between stations and components in earthquake time history data of multiple stations and multiple components; constructing a spatiotemporal earthquake time history prediction model based on the spatiotemporal feature extraction module and the LSTM neural network; training the spatiotemporal earthquake time history prediction model using the three-component earthquake time history dataset; generating a time series of earthquake time history features of target points in the sparse area without station records using the trained spatiotemporal earthquake time history prediction model, and further generating an earthquake time history field model for the sparse area.

[0009] Furthermore, the method comprises the following steps:

[0010] Step 1: Collect three-component seismic motion time history data recorded by stations, generate three-component seismic motion time history data of target sites through simulation using existing historical record data, and compile a three-component seismic motion time history sample data set by combining the existing record data and the simulated three-component seismic motion time history data;

[0011] Step 2: Construct a spatiotemporal prediction model for earthquake motion history, which includes a spatiotemporal feature extraction module and an LSTM neural network connected in sequence.

[0012] Step 3: Using the three-component earthquake motion time history sample data set to train the earthquake motion time history spatiotemporal prediction model;

[0013] Step 4: The trained spatiotemporal earthquake motion history prediction model is used to predict the high-order characteristic time series of the earthquake motion history, thereby generating complete time series data of the earthquake motion history at unrecorded field points; and a multi-component earthquake motion history field model is constructed based on the generated time series data.

[0014] Furthermore, in step one, the simulation method includes a physical simulation method and / or a computational simulation method.

[0015] Furthermore, the physical simulation method includes the following steps:

[0016] In the selected sparsely populated area of seismic stations, the source parameters are set, a physical model is constructed, and one or a combination of the following physical simulation methods is used: spectral element method, finite difference method, and random method; the three-component seismic motion time history data of the unrecorded points in the sparsely populated area of seismic stations are calculated.

[0017] Furthermore, the computational simulation method includes the following steps:

[0018] Set up a number of sampling points for measuring lithology and geological structure in the area with sparse seismic observation stations. The locations of the sampling points include the spatial locations of recorded sites and the spatial locations of target sites without station records.

[0019] The mapping relationship between the three-component seismic motion history and the lithology and geological structure parameters of the sampling point is fitted based on the lithology and geological structure parameters of the sampling point and the existing three-component seismic motion time history historical records; the three-component seismic motion time history data of the target site are generated based on the mapping relationship between the fitted three-component seismic motion history and the lithology and geological structure parameters of the sampling point.

[0020] Furthermore, in step 2, the spatiotemporal feature extraction module includes a ConvLSTM3D module, a Conv3D module, and a Flatten layer connected in sequence;

[0021] The ConvLSTM3D module is used to capture the spatial correlation between stations and components in multi-component seismic time history data. It inputs the three-component seismic time history data of each station and outputs a 3D feature map reflecting the correlation between stations and components.

[0022] The Conv3D module is used to convert 3D features into 2D features;

[0023] The Flatten layer is used to flatten the 2D features into a one-dimensional tensor;

[0024] The Conv3D module has a built-in 3D convolution kernel, which uses the 3D convolution kernel to perform a weighted sum operation on the spatial features of the 3D feature map to reduce the dimension to a 2D feature map; then, it is flattened into a one-dimensional vector in chronological order through the Flatten layer.

[0025] Furthermore, the LSTM neural network includes an LSTM layer and a fully connected layer connected in sequence. The LSTM layer includes multiple LSTM units, and each LSTM unit includes an input gate, a forget gate, and an output gate. The Flatten layer inputs the one-dimensional vector flattened in chronological order into the LSTM layer. Each LSTM unit receives the feature representation of the corresponding time step, and captures the short-term and long-term time series characteristics of the seismic motion time series, as well as the long-term and short-term dependencies through the forget gate, input gate, and output gate, and outputs a feature vector. The number of output nodes of the fully connected layer is set according to the number of target sites without station records and the component dimension. Each neuron in the fully connected layer receives the feature vectors of all LSTM units, and through the linear transformation of the weight matrix and the nonlinear mapping of the activation function, the predicted value of each time step is gradually generated in chronological order. The predicted value of each time step includes the target site number and the seismic motion time series data of each component corresponding to the target site number. These predicted values are spliced in sequence to form a complete target site seismic motion time series.

[0026] Furthermore, in step three, the method of using three-component earthquake motion time history sample data to train the earthquake motion time history spatiotemporal prediction model includes the following method steps:

[0027] First, the recorded earthquake time history data were compiled into a three-dimensional data matrix, in which the horizontal axis represents time, the vertical axis represents the station order, and the depth axis represents the three-component direction. This matrix serves as the input data for the spatiotemporal earthquake time history prediction model. The three-component earthquake time history data generated by the simulation method were stacked row by row in the order of the target sites into a two-dimensional matrix, which served as the output data of the spatiotemporal earthquake time history prediction model.

[0028] The key hyperparameters of the earthquake motion history spatiotemporal prediction model are set, including the learning rate, number of iterations, and batch size. The Adam optimizer is used to update the parameters of the earthquake motion history spatiotemporal prediction model. The predicted value is calculated through forward propagation. The loss function is calculated through backpropagation and the weights of the earthquake motion history spatiotemporal prediction model are updated. The iterations are repeated until the earthquake motion history spatiotemporal prediction model converges.

[0029] Furthermore, in the selected area with sparse seismic stations, the area is divided using kilometer grids to determine the spatial distribution locations of target field points without station records in the area.

[0030] The present invention also includes a device for a method for generating a seismic motion time history field based on three-component broadband simulation data, comprising a memory and a processor, wherein the memory is used to store a computer program; the processor is used to execute the computer program and implement the steps of the method for generating a seismic motion time history field based on three-component broadband simulation data as described above when executing the computer program.

[0031] The advantages and positive effects of this invention include: Based on deep learning, it achieves time-history output for large-scale sites within a defined area. By combining the ConvLSTM3D network with the LSTM network, it overcomes the limitations of traditional models in effectively learning the spatiotemporal characteristics of multi-station, multi-component time-history data. By leveraging the learning capabilities of the deep learning model and fully utilizing limited station record data, it can rapidly generate three-component seismic motion time histories for large-scale sites, providing data support for earthquake risk assessment and disaster emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flowchart of the method for generating a multi-component seismic time history field model.

[0033] Figure 2 Schematic diagram of the spatial distribution of earthquake observation stations and unrecorded sites.

[0034] Figure 3 Schematic diagram of the process of compiling recorded earthquake motion time history data into a three-dimensional data matrix.

[0035] Figure 4 Schematic diagram of the spatiotemporal feature extraction structure of the ConvLSTM3D module.

[0036] Figure 5 Schematic diagram of the structure of the spatiotemporal prediction model for earthquake motion history.

[0037] Figure 6 Schematic diagram of the earthquake time history field model in a sparse area.

[0038] In the picture:

[0039] H t-1 : The hidden state of the previous time step;

[0040] C t-1 : cell state at the previous time step;

[0041] g t : Contents of candidate memory cells at the current time step;

[0042] X t : Input of the current time step;

[0043] f t : The output of the forget gate at the current time step;

[0044] i t : The output of the input gate at the current time step;

[0045] O t : The output of the output gate at the current time step;

[0046] C t : cell state at the current time step;

[0047] H t : hidden state at the current time step;

[0048] t0: the initial time point of the first sliding window;

[0049] t n : The initial time point of the n+1th sliding window;

[0050] σ: sigmoid activation function;

[0051] tanh: Hyperbolic tangent activation function. DETAILED DESCRIPTION

[0052] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0053] The Chinese annotations for the English words, phrases and abbreviations in this application are as follows:

[0054] LSTM is the abbreviation of Long Short-Term Memory, and its Chinese meaning is long short-term memory network.

[0055] ConvLSTM is the abbreviation of Convolutional Long Short-Term Memory, and its Chinese meaning is Convolutional Long Short-Term Memory Network.

[0056] The Chinese interpretation of Flatten layer is flatten layer.

[0057] The Chinese interpretation of the Conv3D module is three-dimensional convolution layer.

[0058] Adam is the abbreviation of Adaptive Moment Estimation, which means adaptive momentum estimation in Chinese.

[0059] NS is the abbreviation of North-South, which means north-south direction in Chinese.

[0060] EW is the abbreviation of East-West, which means east-west direction in Chinese.

[0061] UD is the abbreviation of Up-Down, which means vertical direction in Chinese.

[0062] See Figures 1 to 6, a method for generating earthquake time history field based on three-component broadband simulation data, which selects a sparse area of earthquake stations and determines the spatial distribution position of target field points without station records in the area; collects three-component earthquake time history data with station records, and uses a simulation method to generate three-component earthquake time history data of target field points, and constructs a three-component earthquake time history data set by combining the recorded data and the simulated data; constructs a spatiotemporal feature extraction module based on the ConvLSTM3D neural network to extract spatiotemporal features reflecting the spatial correlation between stations and components in the earthquake time history data of multiple stations and multiple components; constructs a spatiotemporal prediction model for earthquake time history based on the spatiotemporal feature extraction module and the LSTM neural network; uses the three-component earthquake time history data set to train the spatiotemporal prediction model for earthquake time history; generates a time series of earthquake time history features of target field points without station records in the sparse area through the trained spatiotemporal prediction model for earthquake time history, and further generates a earthquake time history field model for the sparse area.

[0063] Preferably, the method may include the following steps:

[0064] Step 1: Collect the three-component seismic motion time history data recorded by the stations, use the existing historical record data to generate the three-component seismic motion time history data of the target site through simulation methods, and compile the three-component seismic motion time history sample data set by combining the existing recorded data and the three-component seismic motion time history data generated by simulation.

[0065] Step 2: Construct a spatiotemporal prediction model for earthquake motion history. The spatiotemporal prediction model for earthquake motion history includes a spatiotemporal feature extraction module and an LSTM neural network connected in sequence.

[0066] Step three: Use the three-component seismic motion time history sample data set to train the seismic motion time history spatiotemporal prediction model.

[0067] Step 4: The trained spatiotemporal earthquake motion history prediction model is used to predict the high-order characteristic time series of the earthquake motion history, thereby generating complete time series data of the earthquake motion history at unrecorded field points; and a multi-component earthquake motion history field model is constructed based on the generated time series data.

[0068] Preferably, in step 1, the simulation method may include a physical simulation method and / or a computational simulation method.

[0069] Preferably, the physical simulation method may include the following method steps:

[0070] In the selected sparsely populated area of seismic stations, the source parameters are set, a physical model is constructed, and one or a combination of the following physical simulation methods is used: spectral element method, finite difference method, and random method; the three-component seismic motion time history data of the unrecorded points in the sparsely populated area of seismic stations are calculated.

[0071] Preferably, the computational simulation method may include the following method steps:

[0072] A number of sampling points for determining lithology and geological structure can be set up in areas with sparse seismic observation stations. The locations of the sampling points may include the spatial locations of recorded sites and the spatial locations of target sites without station records.

[0073] The mapping relationship between the three-component seismic motion history and the lithology and geological structure parameters of the sampling point is fitted based on the lithology and geological structure parameters of the sampling point and the existing three-component seismic motion time history historical records; the three-component seismic motion time history data of the target site are generated based on the mapping relationship between the fitted three-component seismic motion history and the lithology and geological structure parameters of the sampling point.

[0074] Preferably, in step 2, the spatiotemporal feature extraction module may include a ConvLSTM3D module, a Conv3D module and a Flatten layer connected in sequence.

[0075] The ConvLSTM3D module can be used to capture the spatial correlation between stations and components in multi-component seismic time history data. It inputs the three-component seismic time history data of each station and outputs a 3D feature map reflecting the correlation between stations and components.

[0076] The Conv3D module can be used to convert 3D features into 2D features.

[0077] The Flatten layer can be used to flatten 2D features into a one-dimensional tensor.

[0078] The Conv3D module can have a built-in 3D convolution kernel, which uses the 3D convolution kernel to perform a weighted sum operation on the spatial features of the 3D feature map to reduce the dimension to a 2D feature map; then, it is flattened into a one-dimensional vector in chronological order through the Flatten layer.

[0079] The above-mentioned 1D, 2D, and 3D refer to 1 dimension, 2 dimensions, and 3 dimensions respectively.

[0080] Preferably, the LSTM neural network may include an LSTM layer and a fully connected layer connected in sequence. The LSTM layer may include multiple LSTM units, and each LSTM unit includes an input gate, a forget gate, and an output gate. The Flatten layer inputs the one-dimensional vector flattened in chronological order into the LSTM layer. Each LSTM unit receives the feature representation of the corresponding time step, and captures the short-term and long-term time series characteristics of the seismic motion time series, as well as the long-term and short-term dependencies through the forget gate, input gate, and output gate, and outputs a feature vector. The number of output nodes of the fully connected layer can be set according to the number of target sites without station records and the component dimension. Each neuron in the fully connected layer receives the feature vectors of all LSTM units, and through the linear transformation of the weight matrix and the nonlinear mapping of the activation function, the predicted value of each time step is gradually generated in chronological order. The predicted value of each time step may include the target site number and the seismic motion time series data of each component corresponding to the target site number. These predicted values are spliced in sequence to form a complete target site seismic motion time series.

[0081] Preferably, in step 3, the method of using three-component seismic motion time history sample data to train the seismic motion time history spatiotemporal prediction model may include the following method steps:

[0082] First, the recorded seismic motion time history data can be compiled into a three-dimensional data matrix, in which the horizontal axis can represent time, the vertical axis can represent the station order, and the depth axis can represent the three-component direction; if the three-dimensional data matrix is represented by the XYZ three-dimensional coordinate system, the X-axis represents time, the Z-axis represents the station order, and the Y-axis represents the three-component direction, which serves as the input data of the seismic motion time history spatiotemporal prediction model; the three-component seismic motion time history data generated by the simulation method are stacked row by row into a two-dimensional matrix according to the target site order, which serves as the output data of the seismic motion time history spatiotemporal prediction model.

[0083] Set the key hyperparameters of the seismic motion history spatiotemporal prediction model, which may include learning rate, number of iterations, and batch size. Use the Adam optimizer to update the parameters of the seismic motion history spatiotemporal prediction model, calculate the predicted value through forward propagation, calculate the loss function through backpropagation and update the weights of the seismic motion history spatiotemporal prediction model, and repeat the iteration until the seismic motion history spatiotemporal prediction model converges.

[0084] Preferably, in an area with sparse seismic stations, the area can be divided into kilometer grids to determine the spatial distribution positions of target field points without station records in the area.

[0085] The present invention also includes a device for a method for generating a seismic motion time history field based on three-component broadband simulation data, comprising a memory and a processor, wherein the memory is used to store a computer program; the processor is used to execute the computer program and implement the steps of the method for generating a seismic motion time history field based on three-component broadband simulation data as described above when executing the computer program.

[0086] The following is a preferred embodiment of the present invention to further illustrate the workflow and working principle of the present invention:

[0087] A method for generating earthquake time history fields based on three-component broadband simulation data comprises the following steps: selecting an area with sparse earthquake stations and determining the spatial distribution positions of target points in the area without station records; collecting three-component earthquake time history data recorded by stations, and generating three-component earthquake time history data of target points using a simulation method, and constructing a three-component earthquake time history dataset by combining the recorded data and the simulated data; constructing a spatiotemporal feature extraction module based on a ConvLSTM3D neural network to extract spatiotemporal features reflecting spatial correlations between stations and components in earthquake time history data of multiple stations and multiple components; constructing a spatiotemporal earthquake time history prediction model based on the spatiotemporal feature extraction module and the LSTM neural network; training the spatiotemporal earthquake time history prediction model using the three-component earthquake time history dataset; generating a time series of earthquake time history features of target points in the sparse area without station records using the trained spatiotemporal earthquake time history prediction model, and further generating an earthquake time history field model for the sparse area.

[0088] This paper proposes a spatiotemporal prediction model for earthquake motion history that uses three-component time history data (NS, EW, UD) from regional seismic observation stations as input and outputs a large amount of unrecorded ground motion time history data. This model innovatively combines the technical advantages of convolutional long short-term memory networks and long short-term memory networks, and by combining physical simulation data with deep learning methods, it enables rapid generation of ground motion time histories in areas with sparsely populated stations. The method includes the following steps:

[0089] Step 1: Select an area with sparse seismic stations and determine the spatial distribution of target field points with no records in the area. Step 1 includes the following sub-steps:

[0090] A1. Select an area with sparse seismic stations and determine the scope of the area.

[0091] A2. Divide the area into kilometer grids.

[0092] A3. Determine the locations of unrecorded points where earthquake motion history prediction is required.

[0093] Step 2: Collect three-component seismic motion time history data recorded by a limited number of stations, and use physical simulation methods to generate three-component seismic motion time history data without recorded points in a sparse station area to construct a three-component seismic motion time history dataset. Step 2 includes the following sub-steps:

[0094] B1. Collect the earthquake motion time history records of observation stations in the region as input data for the earthquake motion time history spatiotemporal prediction model. Figure 3As shown in Figure 1, the three-component seismic motion time histories (NS, EW, UD) of multiple groups of observation stations are stacked to obtain a three-dimensional matrix, where the horizontal axis represents time, the vertical axis represents the station order, and the depth axis represents the direction of the three components. That is, the three-dimensional matrix composed of the three-component seismic motion time histories of multiple groups of observation stations is used to construct the model input data.

[0095] B2. Use physical simulation methods to simulate and obtain the time history data of unrecorded points as the output data of the spatiotemporal prediction model of earthquake motion time history. Select the source parameters of the set earthquake in the sparse seismic station area, build a physical model, and use physical simulation methods (such as deterministic spectral element method and finite difference method, random method and hybrid method, etc.) to calculate the three-component earthquake motion time history data of the unrecorded points in the sparse station area. Use multiple groups of unrecorded points as target points to construct multiple groups of three-component time history data in the order of target points, and arrange the three-component time history data in rows and stack them into a two-dimensional matrix, that is, a two-dimensional matrix composed of multiple groups of three-component earthquake motion time history data of unrecorded points to construct the model output data;

[0096] B3. Use the input data constructed in step B1 and the output data generated in step B2 to construct a three-component earthquake time history sample dataset for training the spatiotemporal earthquake time history prediction model. This dataset is used to train the spatiotemporal earthquake time history prediction model, learn the mapping relationship between earthquake time history data recorded from observation stations and unrecorded sites, and achieve prediction of regional earthquake time history fields.

[0097] Step 3: The spatiotemporal feature extraction module performs spatiotemporal feature extraction. Step 3 includes the following sub-steps:

[0098] C1. First, the input data in the training set, which includes the three-dimensional matrix of the three-component seismic time histories of multiple groups of observation stations, is input into the ConvLSTM3D module, and the size of the three-dimensional convolution kernel is set so that it can cover all station time histories and all components.

[0099] C2. Through three-dimensional convolution operations, the spatial features of the three-component seismic time history data of the stations are extracted, the spatial correlation between stations and components is captured, and the gating mechanism of LSTM is used to capture the temporal change pattern to obtain the spatiotemporal characteristics at the current time.

[0100] like Figure 4 As shown in Figure 1, first, the processed 3D data matrix is input into the ConvLSTM3D module, and a 3D feature map is obtained through a 3D convolution operation and LSTM gating mechanism. At each time step, the ConvLSTM3D unit receives the cell state, hidden state, and data of the current time step as input, and regulates the information flow through the input gate, forget gate, and output gate, thereby achieving efficient modeling of spatiotemporal features. Specifically, the forget gate determines the memory cell C of the previous time step. t-1Which information needs to be forgotten to filter historical information? The specific method is to use the three-dimensional convolution operation to respectively perform the hidden state H of the previous time step. t-1 and the current input X t Perform feature extraction and perform element-by-element multiplication on the cell state of the previous time step, while adding bias b f , and then use the sigmoid function to map the calculation result to the (0,1) interval to obtain the output f of the forget gate t , and then with C t-1 Element-by-element multiplication realizes forgetting; the input gate is responsible for integrating the new information of the current time step into the memory cell. Its calculation is divided into two steps: first, a three-dimensional convolution operation is used to perform the hidden state H of the previous time step. t-1 and the current input X t Perform feature extraction, perform element-by-element multiplication on the cell state at the previous time step and add bias b i , the calculation result is mapped to the (0,1) interval through the sigmoid function to obtain the output i of the input gate t , then the hidden state H of the previous time step t-1 and the current input X t Perform convolution operation and add bias b c , and then mapped to the (-1,1) interval through the tanh function to generate the candidate memory cell content g t , the output of the input gate is multiplied by the content of the candidate memory cell element by element to obtain new information; the output gate controls how much information in the memory cell of the current time step will be passed to the hidden state. The specific method is to first use the three-dimensional convolution operation to respectively perform the hidden state H of the previous time step. t-1 and the current input X t Perform feature extraction, perform element-by-element multiplication on the cell state at the previous time step and add bias b o , mapped to the (0,1) interval through the sigmoid function, and the output O of the output gate is obtained t , apply the tanh function to the cell state and map it to (-1,1), and then multiply it element-by-element with the output of the output gate to obtain the hidden state of the current time step; finally, after all time steps are processed, the feature information of all time steps is extracted to form a 3D feature map as the output of the ConvLSTM3D module.

[0101] C3. Repeat step C2 to obtain the 3D spatiotemporal feature map at all times.

[0102] C4. Use the Conv3D module to synthesize and compress the 3D spatiotemporal feature map to generate a 2D feature map.

[0103] C5. Convert the 2D features into one-dimensional vectors through the Flatten layer to adapt to the input format of subsequent time series modeling.

[0104] Step 4, modeling the high-order characteristic time series of earthquake motion time history and generating multi-component earthquake motion time history based on LSTM neural network, step 4 includes the following sub-steps;

[0105] D1. Use an LSTM neural network to further model the spatiotemporal features, leveraging its gating mechanism to capture the short-term and long-term sequence features and long-term and short-term dependencies of the earthquake motion history. The LSTM neural network consists of sequentially connected LSTM layers and fully connected layers. The LSTM layer includes multiple LSTM units, each of which includes an input gate, a forget gate, and an output gate. The Flatten layer inputs a one-dimensional vector flattened in chronological order into the LSTM layer. Each LSTM unit receives the feature representation of the corresponding time step and captures the short-term and long-term sequence features and long-term and short-term dependencies of the earthquake motion history through the forget gate, input gate, and output gate, outputting a feature vector.

[0106] D2. Input the feature vector generated by the LSTM layer at each time step into the fully connected layer, generate the predicted value through linear transformation and activation function, and splice it in chronological order to form a complete earthquake motion sequence for the unrecorded site.

[0107] Step 5, training of the spatiotemporal prediction model for earthquake motion history and prediction of time history at unrecorded sites, includes the following sub-steps:

[0108] E1. Information is forward propagated. The three-component seismic time history data from multiple stations in the training set are input into the ConvLSTM3D module to extract spatiotemporal features. The data is then processed through the Conv3D module, Flatten layer, LSTM layer, and fully connected layer. After the network is forward propagated to the output layer, the prediction error is calculated using the loss function. The error then propagates backward from the output layer, passing through the fully connected layer, LSTM layer, Flatten layer, Conv3D module, ConvLSTM3D module, and other layers in sequence. The gradients of the parameters of each layer are calculated, and the parameters are updated using the Adam optimizer.

[0109] E2. Iteratively optimize the model parameters of the earthquake motion time-space prediction model. In each iteration, the system repeatedly executes the forward propagation and backpropagation processes, and continuously updates the network weights and bias parameters through the Adam optimizer until the loss function value is less than the preset threshold or the maximum number of iterations is reached. The training is terminated to obtain the final model parameters, complete the model training, and obtain the trained earthquake motion time-space prediction model.

[0110] E3. By completing the trained spatiotemporal prediction model of seismic motion history, the three-component seismic motion history of unrecorded points in the region is synchronously and rapidly generated to construct a sparse regional seismic motion history field model.

[0111] like Figure 5As shown in the figure, this paper designs a high-order feature time series modeling module based on LSTM for generating multi-station, multi-component earthquake time histories. To meet the input requirements of the LSTM layer while avoiding dimensionality explosion and reducing feature information loss, a stepwise dimensionality reduction approach is used to convert 3D feature maps into 1D feature vectors. First, a weighted summation operation is performed on the 3D feature map using a 3D convolution kernel to reduce its dimensionality to a 2D feature map. Then, the map is flattened into a one-dimensional vector in chronological order through a flattening layer and input into the LSTM layer. Each unit of the LSTM network receives the feature representation of the corresponding time step, and captures the short-term and long-term temporal characteristics of the earthquake motion time series, as well as the long-term and short-term dependencies, through the forget gate, input gate, and output gate, and outputs a feature vector. The number of output nodes of the fully connected layer is set according to the number of regional target points and the component dimensions (NS, EW, UD). Each neuron in the fully connected layer receives the feature vectors of all LSTM units, and generates a predicted value for each time step through linear transformation of the weight matrix and nonlinear mapping of the activation function. The predicted value of each time step is generated step by step in chronological order, and these predicted values are spliced in sequence to form a complete earthquake motion time series without recorded points.

[0112] like Figure 6 As shown in the figure, the model is used to synchronously generate three-component seismic motion time history data of large-scale unrecorded points in the region, realizing a method for constructing earthquake time history fields in areas with sparse stations, effectively filling the gap in seismic motion time history data in areas with sparse stations.

[0113] Functional modules and algorithms such as the spatiotemporal feature extraction module, LSTM neural network, ConvLSTM3D module, Conv3D module, Flatten layer, fully connected layer, Adam optimizer, spectral element method, finite difference method, and random method can all adopt applicable functional modules and algorithms in the existing technology, or adopt applicable functional modules and algorithms in the existing technology and construct them using conventional technical means.

[0114] The embodiments described above are only used to illustrate the technical ideas and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of the patent of the present invention cannot be limited by these embodiments alone. That is, any equivalent changes or modifications made to the spirit disclosed by the present invention still fall within the scope of the patent of the present invention.

Claims

1. A method for generating earthquake time history fields based on three-component broadband simulation data, characterized in that: A sparsely populated area of earthquake stations was selected, and the spatial distribution of target sites with no station records in the area was determined; three-component earthquake motion history data recorded by stations were collected, and three-component earthquake motion history data of target sites were generated using simulation methods, and a three-component earthquake motion history dataset was constructed by combining the recorded data and the simulated data; a spatiotemporal feature extraction module was constructed based on the ConvLSTM3D neural network to extract spatiotemporal features reflecting the spatial correlation between stations and components in the earthquake motion history data of multiple stations and multiple components; a spatiotemporal earthquake motion history prediction model was constructed based on the spatiotemporal feature extraction module and the LSTM neural network; the three-component earthquake motion history dataset was used to train the spatiotemporal earthquake motion history prediction model; the earthquake motion history feature time series of the target sites with no station records in the sparse area was generated by the trained spatiotemporal earthquake motion history prediction model, and a sparse area earthquake motion history field model was further generated.

2. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 1, characterized in that: The method comprises the following steps: Step 1: Collect three-component seismic motion time history data recorded by stations, generate three-component seismic motion time history data of target sites through simulation using existing historical record data, and compile a three-component seismic motion time history sample data set by combining the existing record data and the simulated three-component seismic motion time history data; Step 2: Construct a spatiotemporal prediction model for earthquake motion history, which includes a spatiotemporal feature extraction module and an LSTM neural network connected in sequence. Step 3: Using the three-component earthquake motion time history sample data set to train the earthquake motion time history spatiotemporal prediction model; Step 4: The trained spatiotemporal earthquake motion history prediction model is used to predict the high-order characteristic time series of the earthquake motion history, thereby generating complete time series data of the earthquake motion history at unrecorded field points; and a multi-component earthquake motion history field model is constructed based on the generated time series data.

3. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 2, characterized in that: In step 1, the simulation method includes a physical simulation method and / or a computational simulation method.

4. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 3, characterized in that: The physical simulation method includes the following steps: In the selected sparsely populated area of seismic stations, the source parameters are set, a physical model is constructed, and one or a combination of the following physical simulation methods is used: spectral element method, finite difference method, and random method; the three-component seismic motion time history data of the unrecorded points in the sparsely populated area of seismic stations are calculated.

5. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 3, characterized in that: The computational simulation method includes the following steps: Set up a number of sampling points for measuring lithology and geological structure in the area with sparse seismic observation stations. The locations of the sampling points include the spatial locations of recorded sites and the spatial locations of target sites without station records. Based on the lithology and geological structure parameters of the sampling points and the existing historical records of three-component earthquake motion time history, the mapping relationship between the three-component earthquake motion time history and the lithology and geological structure parameters of the sampling points is fitted; The three-component seismic motion time history data of the target site are generated based on the mapping relationship between the fitted three-component seismic motion time history and the lithology and geological structure parameters of the sampling point.

6. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 2, characterized in that: In step 2, the spatiotemporal feature extraction module includes a ConvLSTM3D module, a Conv3D module, and a Flatten layer connected in sequence; The ConvLSTM3D module is used to capture the spatial correlation between stations and components in multi-component ground motion time history data; It inputs the three-component seismic time history data of each station and outputs a 3D characteristic map reflecting the correlation between stations and components; The Conv3D module is used to convert 3D features into 2D features; The Flatten layer is used to flatten the 2D features into a one-dimensional tensor; The Conv3D module has a built-in 3D convolution kernel, which uses the 3D convolution kernel to perform a weighted sum operation on the spatial features of the 3D feature map to reduce the dimension to a 2D feature map; then, it is flattened into a one-dimensional vector in chronological order through the Flatten layer.

7. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 6, characterized in that: The LSTM neural network includes a LSTM layer and a fully connected layer connected in sequence. The LSTM layer includes multiple LSTM units, and each LSTM unit includes an input gate, a forget gate, and an output gate. The Flatten layer inputs the one-dimensional vector flattened in chronological order into the LSTM layer. Each LSTM unit receives the feature representation of the corresponding time step, and captures the short- and long-term time series characteristics of the seismic motion time series, as well as the long- and short-term dependencies, through the forget gate, input gate, and output gate, and outputs a feature vector. The number of output nodes of the fully connected layer is set according to the number of target sites without station records and the component dimension. Each neuron in the fully connected layer receives the feature vectors of all LSTM units, and through the linear transformation of the weight matrix and the nonlinear mapping of the activation function, it gradually generates the predicted value of each time step in chronological order. The predicted value of each time step includes the target site number and the seismic motion time series data of each component corresponding to the target site number. These predicted values are spliced in sequence to form a complete target site seismic motion time series.

8. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 2, characterized in that: In step 3, the method of using three-component earthquake motion time history sample data to train the earthquake motion time history spatiotemporal prediction model includes the following method steps: First, the recorded earthquake time history data were compiled into a three-dimensional data matrix, in which the horizontal axis represents time, the vertical axis represents the station order, and the depth axis represents the three-component direction. This matrix serves as the input data for the spatiotemporal earthquake time history prediction model. The three-component earthquake time history data generated by the simulation method were stacked row by row in the order of the target sites into a two-dimensional matrix, which served as the output data of the spatiotemporal earthquake time history prediction model. The key hyperparameters of the earthquake motion history spatiotemporal prediction model are set, including the learning rate, number of iterations, and batch size. The Adam optimizer is used to update the parameters of the earthquake motion history spatiotemporal prediction model. The predicted value is calculated through forward propagation. The loss function is calculated through backpropagation and the weights of the earthquake motion history spatiotemporal prediction model are updated. The iterations are repeated until the earthquake motion history spatiotemporal prediction model converges.

9. The method for generating earthquake time history field based on three-component broadband simulation data according to claim 1, characterized in that: In the selected area with sparse seismic stations, the area is divided into kilometer grids to determine the spatial distribution locations of target field points without station records in the area.

10. A device for generating a seismic time history field based on three-component broadband simulation data, comprising a memory and a processor, characterized in that: The memory is used to store computer programs; the processor is used to execute the computer program and implement the steps of the method for generating a seismic time history field based on three-component broadband simulation data as described in any one of claims 1 to 9 when executing the computer program.

Images