Lightweight local ground motion parameter prediction model construction method and earthquake early warning system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INST OF ENG MECHANICS CHINA EARTHQUAKE ADMINISTRATION
- Filing Date
- 2025-01-06
- Publication Date
- 2026-07-03
Smart Images
Figure CN119716988B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of earthquake prediction technology, and more specifically, to a lightweight method for constructing on-site ground motion parameter prediction models and an earthquake early warning system. Background Technology
[0002] Predicting local ground motion parameters is an important task in earthquake early warning. Its purpose is to predict local ground motion parameters using P-wave signals recorded by stations after an earthquake occurs, so as to quickly issue warnings and effectively reduce casualties and economic losses caused by earthquake disasters.
[0003] Existing local ground motion parameter prediction models typically rely on multiple features manually extracted from P-waves for prediction, or utilize multi-layer neural network structures to extract feature data from the original seismic waveform and then predict local ground motion parameters based on this extracted data. Manually extracting features from P-waves is labor-intensive and time-consuming, impacting the speed of alert dissemination. Conversely, extracting feature data from the original seismic waveform using multi-layer neural networks requires significant memory, limiting the deployment environment for the corresponding prediction models. Summary of the Invention
[0004] In view of this, the purpose of this application is to provide a lightweight method for constructing a local ground motion parameter prediction model and an earthquake early warning system, so as to solve the technical problems of low parameter prediction efficiency or limited deployment scenarios of existing local ground motion parameter prediction models.
[0005] In a first aspect, embodiments of this application provide a method for constructing a lightweight on-site seismic motion parameter prediction model, the method comprising:
[0006] Based on the acceleration waveform data at the arrival of the earthquake P-wave, the initial prediction model is trained to obtain a pre-trained prediction model.
[0007] The pre-trained prediction model is used as a teacher model, and the knowledge of the teacher model is transferred to the student model through distillation;
[0008] A lightweight on-site ground motion parameter prediction model is constructed based on the student model.
[0009] In the above implementation process, this lightweight on-site ground motion parameter prediction model construction method trains an initial prediction model based on the acceleration waveform data at the arrival of the P-wave, obtaining a pre-trained prediction model. This pre-trained prediction model is then used as a teacher model, and its knowledge is transferred to a student model through distillation. A lightweight on-site ground motion parameter prediction model is then constructed based on this student model. This method reduces the complexity and computational load of the student model by transferring knowledge from the pre-trained model to the student model through distillation. Furthermore, constructing the lightweight on-site ground motion parameter prediction model based on the student model ensures efficient on-site ground motion parameter prediction while reducing the model's hardware resource requirements and expanding its deployment scenarios. This solves the technical problems of low parameter prediction efficiency or limited deployment scenarios in existing on-site ground motion parameter prediction models.
[0010] Furthermore, compared to existing local ground motion parameter prediction models based on multi-layer neural network structures, the lightweight local ground motion parameter prediction model constructed using the method provided in this application is simpler and can further improve the model's running speed, thereby improving the efficiency of local ground motion parameter prediction and achieving better earthquake early warning results.
[0011] Optionally, in this embodiment of the application, the step of training the initial prediction model based on the acceleration waveform data at the arrival of the earthquake P-wave to obtain a pre-trained prediction model includes: integrating the acceleration waveform data to obtain velocity waveform data; integrating the velocity waveform data to obtain displacement waveform data; calculating the local ground motion parameters at the arrival of the earthquake P-wave based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data; and training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model.
[0012] In the above implementation process, velocity waveform data can be obtained by integrating the acceleration waveform data; displacement waveform data can be obtained by integrating the velocity waveform data. Compared with existing local ground motion parameter prediction equations based on a single characteristic parameter, more characteristic parameters can be obtained based on acceleration waveform data, velocity waveform data, and displacement waveform data. This reduces the uncertainty of the local ground motion parameter prediction results obtained based on the constructed lightweight local ground motion parameter prediction model, thus improving the accuracy of the obtained local ground motion parameter prediction results. Therefore, the lightweight local ground motion parameter prediction model constructed based on the lightweight local ground motion parameter prediction model construction method provided in this application can improve the efficiency and accuracy of earthquake early warning.
[0013] Optionally, in this embodiment, the pre-trained prediction model includes a trained first feature extraction module and a first fully connected module; the step of using the pre-trained prediction model as a teacher model and transferring the knowledge of the teacher model to the student model through distillation includes: using the pre-trained prediction model as the teacher model and transferring the first feature extraction module to the student model through distillation; wherein, the student model includes the first feature extraction module and a second fully connected module;
[0014] The step of constructing a lightweight local ground motion parameter prediction model based on the student model includes: inputting the acceleration waveform data, the velocity waveform data, and the displacement waveform data into the student model to obtain the prediction parameters output by the student model; adjusting the parameters of the second fully connected module in the student model according to the local ground motion parameters and the prediction parameters until the similarity between the prediction parameters output by the adjusted student model and the local ground motion parameters reaches a preset similarity threshold; and constructing the lightweight local ground motion parameter prediction model based on the adjusted student model.
[0015] In the above implementation process, the first feature extraction module is transferred to the student model through distillation, and the parameters of the second fully connected module in the student model are adjusted using acceleration waveform data, velocity waveform data, displacement waveform data, and local ground motion parameters. This can reduce the complexity and computational load of the constructed lightweight local ground motion parameter prediction model, improve the efficiency of local ground motion parameter prediction, and further improve the accuracy of local ground motion parameter prediction.
[0016] Optionally, in this embodiment, the first fully connected module includes a linear activation function layer; the second fully connected module includes a nonlinear activation function layer and a linear activation function layer; the step of constructing the lightweight local ground motion parameter prediction model based on the adjusted student model includes: performing a pruning operation on the nonlinear activation function layer in the adjusted student model until the sparsity of the student model after the pruning operation reaches a preset sparsity threshold; and determining the student model after the pruning operation as the lightweight local ground motion parameter prediction model.
[0017] In the above implementation process, the second fully connected module includes a nonlinear activation function layer and a linear activation function layer. By pruning the nonlinear activation function layer in the adjusted student model, the complexity and computational load of the constructed lightweight ground motion parameter prediction model can be further reduced while ensuring the accuracy of the local ground motion parameter prediction of the constructed model, thereby improving the efficiency of local ground motion parameter prediction.
[0018] Optionally, in this embodiment, the first feature extraction module includes a convolutional layer and a pooling layer; training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model includes: performing feature extraction on the acceleration waveform data, the velocity waveform data, and the displacement waveform data based on the convolutional layer to obtain feature extraction data; performing pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data; outputting initial model prediction parameters based on the dimensionality-reduced feature data using the first fully connected module; and adjusting the internal parameters of the initial prediction model based on the initial model prediction parameters and the local ground motion parameters to obtain the pre-trained prediction model.
[0019] In the above implementation process, feature extraction data can be obtained from acceleration waveform data, velocity waveform data and displacement waveform data through convolutional layers; and pooling layers can select the maximum and minimum values in local regions to reduce the length of the input sequence, thereby extracting key feature data and improving the robustness of the pre-trained prediction model obtained through training.
[0020] Optionally, in this embodiment, the feature extraction data includes acceleration feature data, velocity feature data, and displacement feature data; the first feature extraction module further includes: a random deactivation layer; before performing pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data, the step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model further includes: randomly determining effective feature data in the feature extraction data based on the random deactivation layer; the effective feature data includes at least one of the acceleration feature data, velocity feature data, and displacement feature data;
[0021] The step of performing a pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data includes: performing a pooling operation on the effective feature data based on the pooling layer to obtain the dimensionality-reduced feature data.
[0022] In the above implementation process, the effective feature data in the feature extraction data can be randomly determined by the random deactivation layer. In other words, based on the random deactivation layer, some feature extraction data can be randomly discarded during the training process. This reduces the dependence of the pre-trained prediction model on specific feature data, reduces the possibility of overfitting of the pre-trained prediction model, and improves the generalization ability of the pre-trained prediction model obtained through training.
[0023] Optionally, in this embodiment, the first feature extraction module further includes: a bidirectional gating unit layer; before the first fully connected module outputs the initial model prediction parameters based on the dimensionality reduction feature data, the step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model further includes: obtaining feature representation data of the dimensionality reduction feature data based on the bidirectional gating unit layer;
[0024] The step of outputting initial model prediction parameters based on the dimensionality reduction feature data by the first fully connected module includes: outputting the initial model prediction parameters based on the feature representation data by the first fully connected module.
[0025] In the above implementation process, the bidirectional gating unit layer can help the pre-trained prediction model better capture long-term dependencies in the sequence, thereby improving the performance of the pre-trained prediction model obtained through training.
[0026] Optionally, in this embodiment, the first feature extraction module includes: a convolutional layer, a random deactivation layer, a pooling layer, and a bidirectional gated unit layer; the step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model includes: performing feature extraction on the acceleration waveform data, the velocity waveform data, and the displacement waveform data based on the convolutional layer to obtain feature extraction data; the feature extraction data includes: acceleration feature data, velocity feature data, and displacement feature data; based on the random deactivation layer... The system first determines effective feature data from the extracted feature data using a pooling layer. The effective feature data includes at least one of the acceleration feature data, velocity feature data, and displacement feature data. The effective feature data is pooled using the pooling layer to obtain dimensionality-reduced feature data. Feature representation data of the dimensionality-reduced feature data is obtained using the bidirectional gating unit layer. Initial model prediction parameters are output based on the feature representation data using the first fully connected module. The internal parameters of the initial prediction model are adjusted according to the initial model prediction parameters and the local ground motion parameters to obtain the pre-trained prediction model.
[0027] Optionally, in this embodiment of the application, the step of calculating the local ground motion parameters at the arrival of the earthquake P-wave based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data includes: filtering the acceleration waveform data, the velocity waveform data, and the displacement waveform data respectively to obtain filtered acceleration waveform data, velocity waveform data, and displacement waveform data; and calculating the local ground motion parameters at the arrival of the earthquake P-wave based on the processed acceleration waveform data, velocity waveform data, and displacement waveform data.
[0028] The step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model includes: training the initial prediction model based on the acceleration waveform processing data, the velocity waveform processing data, the displacement waveform processing data, and the local ground motion parameters to obtain the pre-trained prediction model.
[0029] In the above implementation process, by filtering the acceleration waveform data, velocity waveform data, and displacement waveform data, noise data in the acceleration waveform data, velocity waveform data, and displacement waveform data can be removed, thereby improving the data quality of the obtained filtered acceleration waveform data, velocity waveform data, and displacement waveform data.
[0030] Optionally, in this embodiment of the application, before training the initial prediction model based on the acceleration waveform data at the arrival of the earthquake P-wave to obtain the pre-trained prediction model, the method further includes: automatically picking up the monitored vertical acceleration data at the arrival of the P-wave, and obtaining the acceleration waveform data at the arrival of the earthquake P-wave based on the monitored triaxial acceleration waveform data.
[0031] Secondly, embodiments of this application provide an earthquake early warning system, the system comprising: a server and an early warning unit;
[0032] The server is configured to predict local ground motion parameters based on a lightweight local ground motion parameter prediction model constructed according to any of the methods described in the first aspect above.
[0033] The early warning unit is configured to provide earthquake early warning based on the local ground motion parameters.
[0034] Thirdly, embodiments of this application also provide an electronic device; the electronic device includes:
[0035] Memory;
[0036] processor;
[0037] The memory stores a computer program executable by the processor. When the computer program is executed by the processor, it performs the lightweight on-site ground motion parameter prediction model construction method described in any of the first aspects.
[0038] Fourthly, embodiments of this application provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the lightweight on-site ground motion parameter prediction model construction method as described in any of the first aspects above.
[0039] Fifthly, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, perform the lightweight on-site ground motion parameter prediction model construction method as described in any of the first aspects.
[0040] Sixthly, embodiments of this application provide a lightweight on-site seismic motion parameter prediction model construction device, the device comprising:
[0041] The training module is used to train the initial prediction model based on the acceleration waveform data when the earthquake P-wave arrives, so as to obtain a pre-trained prediction model.
[0042] The distillation module is used to use the pre-trained prediction model as a teacher model and transfer the knowledge of the teacher model to the student model through distillation.
[0043] The model building module is used to build a lightweight on-site ground motion parameter prediction model based on the student model.
[0044] The beneficial effects of this application include at least the following: the lightweight on-site ground motion parameter prediction model construction method trains an initial prediction model based on the acceleration waveform data at the arrival of the P-wave, obtaining a pre-trained prediction model; the pre-trained prediction model is used as a teacher model, and the knowledge of the teacher model is transferred to the student model through distillation; a lightweight on-site ground motion parameter prediction model is constructed based on the student model. This lightweight on-site ground motion parameter prediction model construction method reduces the complexity and computational load of the student model by transferring knowledge from the pre-trained prediction model to the student model through distillation; and the lightweight on-site ground motion parameter prediction model is constructed based on the student model, which reduces the model's hardware resource requirements while ensuring the efficiency of on-site ground motion parameter prediction, thus expanding the model's deployable scenarios. This solves the technical problem of low parameter prediction efficiency or limited deployment scenarios in existing on-site ground motion parameter prediction models.
[0045] Furthermore, compared to existing local ground motion parameter prediction models based on multi-layer neural network structures, the lightweight local ground motion parameter prediction model constructed using the method provided in this application is simpler and can further improve the model's running speed, thereby improving the efficiency of local ground motion parameter prediction and achieving better earthquake early warning results. Attached Figure Description
[0046] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0047] Figure 1 A flowchart illustrating a lightweight on-site ground motion parameter prediction model construction method provided in this application embodiment;
[0048] Figure 2 A flowchart illustrating a method for obtaining a pre-trained prediction model provided in an embodiment of this application;
[0049] Figure 3 A flowchart illustrating a method for constructing a lightweight local ground motion parameter prediction model based on a student model, provided in an embodiment of this application;
[0050] Figure 4This is a schematic diagram of the structure of a pre-trained prediction model provided in an embodiment of this application;
[0051] Figure 5 A schematic diagram of a lightweight on-site ground motion parameter prediction model provided in an embodiment of this application;
[0052] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0053] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.
[0054] 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 application belongs; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this application.
[0055] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.
[0056] Please see Figure 1 The illustrated diagram shows a flowchart of a lightweight on-site ground motion parameter prediction model construction method provided in an embodiment of this application. This lightweight on-site ground motion parameter prediction model construction method may include the following steps:
[0057] S101. Based on the acceleration waveform data when the earthquake P-wave arrives, train the initial prediction model to obtain a pre-trained prediction model.
[0058] S102. Using the pre-trained prediction model as a teacher model, the knowledge of the teacher model is transferred to the student model through distillation;
[0059] S103. Construct a lightweight on-site ground motion parameter prediction model based on the student model.
[0060] In step 101, the seismic P-wave, also known as the seismic longitudinal wave, propagates in the same direction as the vibration of the medium particles. The acceleration waveform data may include triaxial acceleration waveform data (including vertical acceleration a). ud East-west acceleration a ew and the north-south acceleration ans (Waveform data). The initial prediction model can be an untrained neural network model, such as a convolutional neural network model or a recurrent neural network model.
[0061] In step 102, knowledge from the teacher model can be transferred to the student model via online or offline distillation. For example, if the initial prediction model is implemented using an untrained convolutional neural network model, the pre-trained prediction model may include a trained feature extraction module and a fully connected module; the feature extraction module from the teacher model can be transferred as knowledge to the student model.
[0062] In step 103, the student model after knowledge transfer can be used as a lightweight local ground motion parameter prediction model; alternatively, based on acceleration waveform data and the corresponding local ground motion parameters, some parameters in the student model can be fine-tuned, and the fine-tuned model can be used as a lightweight local ground motion parameter prediction model. After transferring the knowledge from the pre-trained prediction model to the student model through distillation, since the student model possesses the knowledge from the pre-trained prediction model and has a simpler network structure, a lightweight local ground motion parameter prediction model that can achieve local ground motion parameter prediction with low complexity, low computational cost, and low hardware resource requirements can be constructed based on the student model.
[0063] Therefore, the lightweight local ground motion parameter prediction model construction method provided in this application reduces the complexity and computational load of the student model by distilling the knowledge from the pre-trained prediction model to the student model. Furthermore, constructing a lightweight local ground motion parameter prediction model based on the student model ensures high prediction efficiency while reducing hardware resource requirements and expanding the model's deployment scenarios. This solves the technical problem of low prediction efficiency or limited deployment scenarios in existing local ground motion parameter prediction models. Moreover, compared to existing local ground motion parameter prediction models based on multi-layer neural network structures, the lightweight local ground motion parameter prediction model constructed using the method provided in this application is simpler, further improving its running speed and thus increasing its prediction efficiency for local ground motion parameters, resulting in better earthquake early warning effects. Additionally, because the lightweight local ground motion parameter prediction model constructed using the method provided in this application is simpler, it can operate effectively under limited resource conditions. Therefore, the lightweight ground motion parameter prediction model constructed based on the lightweight ground motion parameter prediction model construction method provided in this application can be deployed in mobile applications, IoT devices, edge computing and other resource-constrained environments, balancing the performance and resource consumption of the prediction model.
[0064] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating a method for obtaining a pre-trained prediction model, as provided in an embodiment of this application.
[0065] In some optional embodiments, S101, training the initial prediction model based on the acceleration waveform data at the arrival of the earthquake P-wave to obtain a pre-trained prediction model includes: S1011, integrating the acceleration waveform data to obtain velocity waveform data; S1012, integrating the velocity waveform data to obtain displacement waveform data; S1013, calculating the local ground motion parameters at the arrival of the earthquake P-wave based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data; S1014, training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model.
[0066] The acceleration waveform data can include: vertical acceleration data, east-west acceleration data, and north-south acceleration data; the velocity waveform data can include: vertical velocity data, east-west velocity data, and north-south velocity data; and the displacement waveform data can include: vertical displacement data, east-west displacement data, and north-south displacement data. Early warning can be issued by deploying stations in the target area and utilizing the arrival time difference between the first and subsequent destructive seismic waves. The lightweight local ground motion parameter prediction model constructed based on the method provided in this application can be used to predict local ground motion parameters. The local ground motion parameters can be adjusted according to the actual earthquake early warning scenario. For example, the local ground motion parameters can include at least one of the following: high-speed rail early warning parameters, gas early warning parameters, nuclear power early warning parameters, instrument intensity parameters, and aftershock rescue parameters. It can be based on... Calculation of local ground motion parameters PGA for high-speed rail earthquake early warning GT ;a ew,gt a represents the east-west acceleration after bandpass filtering in the 0.05-5Hz range. ns,gt This represents the north-south acceleration after a 0.05-5Hz bandpass filter. It can be based on... Calculate the local ground motion parameters SI and S for gas-fired earthquake early warning. v,ew S represents the velocity response spectrum in the east-west direction. v,ns The velocity response spectrum represents the north-south direction, and ξ represents the damping ratio. This can be based on... Calculation of local ground motion parameters CAV for nuclear power plant earthquake early warning;t max This indicates the duration of the earthquake motion time history. It can be based on... Calculation of local ground motion parameters PGA for Chinese instrumental seismic intensity I and PGV I ;a ud,I a ew,I and a ns,I These represent the vertical acceleration, east-west acceleration, and north-south acceleration after bandpass filtering in the 0.1-10Hz range, respectively. ud,I v ew,I and v ns,I These represent the vertical, east-west, and north-south velocities after bandpass filtering in the 0.1-10Hz range. This can be based on... The peak ground motion (Pi) parameter for aftershock relief is calculated. Velocity waveform data can be obtained by integrating acceleration waveform data; displacement waveform data can be obtained by integrating velocity waveform data. Compared to existing ground motion parameter prediction equations based on a single characteristic parameter, this method, using acceleration, velocity, and displacement waveform data, yields more characteristic parameters, reducing the uncertainty of ground motion parameter prediction results obtained from the constructed lightweight ground motion parameter prediction model. Therefore, the lightweight ground motion parameter prediction model constructed based on the method provided in this application can improve the efficiency and accuracy of earthquake early warning.
[0067] Please refer to Figure 3 , Figure 3 This is a flowchart illustrating a method for constructing a lightweight local ground motion parameter prediction model based on a student model, as provided in an embodiment of this application.
[0068] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of a pre-trained prediction model provided in an embodiment of this application. Figure 4The diagram shown is a structural schematic of the pre-trained prediction model, also known as the teacher model. In some optional embodiments, the pre-trained prediction model includes a trained first feature extraction module and a first fully connected module; S102, using the pre-trained prediction model as a teacher model, and transferring the knowledge of the teacher model to the student model through distillation, including: using the pre-trained prediction model as the teacher model, and transferring the first feature extraction module to the student model through distillation; wherein, the student model includes the first feature extraction module and a second fully connected module; S103, constructing a lightweight local ground motion parameter prediction model based on the student model, including: S1031, inputting the acceleration waveform data, the velocity waveform data, and the displacement waveform data into the student model, and obtaining the prediction parameters output by the student model; S1032, adjusting the parameters of the second fully connected module in the student model according to the local ground motion parameters and the prediction parameters, until the similarity between the prediction parameters output by the adjusted student model and the local ground motion parameters reaches a preset similarity threshold; S1033, constructing the lightweight local ground motion parameter prediction model based on the adjusted student model.
[0069] Among them, such as Figure 4 As shown, the first feature extraction module may include feature processing layers such as convolutional layers and pooling layers. In the process of constructing a lightweight on-site ground motion parameter prediction model based on the student model, the parameters of the first feature extraction module in the student model are no longer adjusted; that is, the first feature extraction module does not participate in the training process of the student model, but instead provides feature processing data for the student model. In this way, while ensuring the parameter prediction accuracy of the constructed lightweight on-site ground motion parameter prediction model, the training difficulty and model complexity are reduced. The first feature extraction module is transferred to the student model through distillation, and the parameters of the second fully connected module in the student model are adjusted using acceleration waveform data, velocity waveform data, displacement waveform data, and on-site ground motion parameters. This can further improve the accuracy of on-site ground motion parameter prediction while reducing the complexity and computational load of the constructed lightweight on-site ground motion parameter prediction model and improving the efficiency of on-site ground motion parameter prediction.
[0070] Please refer to Figure 5 , Figure 5 This is a schematic diagram of a lightweight on-site ground motion parameter prediction model provided in an embodiment of this application. In some optional embodiments, the first fully connected module includes a linear activation function layer; the second fully connected module includes a nonlinear activation function layer and a linear activation function layer (e.g., ...). Figure 5(as shown); S1033, Constructing the lightweight local ground motion parameter prediction model based on the adjusted student model, including: performing a pruning operation on the nonlinear activation function layer in the adjusted student model until the sparsity of the student model after the pruning operation reaches a preset sparsity threshold; determining the student model after the pruning operation as the lightweight local ground motion parameter prediction model.
[0071] The nonlinear activation function layer can include 16 neurons (or other reasonable values), and can be implemented using a ReLU activation function layer; the linear activation function layer can include 1 neuron. The degree of pruning can be controlled by setting the final sparsity. For example, with an initial sparsity of 0.5, the final sparsity can be set to 0.8 or other reasonable values. The time range of the pruning operation can also be controlled by setting the start step (which can be 0) and the end step (which can be the total number of steps in the training process). The second fully connected module includes a nonlinear activation function layer and a linear activation function layer; by pruning the nonlinear activation function layer in the adjusted student model, the complexity and computational cost of the constructed lightweight ground motion parameter prediction model can be further reduced while ensuring the accuracy of the local ground motion parameter prediction, thereby improving the efficiency of local ground motion parameter prediction.
[0072] In some optional embodiments, the first feature extraction module includes: a convolutional layer and a pooling layer; S1014, training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model includes: performing feature extraction on the acceleration waveform data, the velocity waveform data, and the displacement waveform data based on the convolutional layer to obtain feature extraction data; performing pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data; outputting initial model prediction parameters based on the dimensionality-reduced feature data using the first fully connected module; and adjusting the internal parameters of the initial prediction model based on the initial model prediction parameters and the local ground motion parameters to obtain the pre-trained prediction model.
[0073] The convolutional layer can be implemented using a one-dimensional convolutional layer, specifically based on... Implement the corresponding convolution operation; X i+j Let i represent the element in the input sequence (acceleration waveform data, velocity waveform data, and displacement waveform data), k represent the position in the output sequence, and K represent the size of the convolution kernel. j Y represents the weights of the convolution kernel. iThis represents the output of the convolutional layer. In a one-dimensional convolutional layer, the number of kernels can be 16, the kernel size can be 3, and the stride can be 1. The pooling layer can be implemented using a one-dimensional max pooling layer, and the pooling window size of the one-dimensional max pooling layer can be 4. A pre-training similarity threshold can be set; if the similarity between the predicted parameters output by the adjusted initial prediction model and the actual ground motion parameters reaches the preset similarity threshold, the adjusted initial prediction model is determined as the pre-trained prediction model. Convolutional layers can extract feature data from acceleration waveform data, velocity waveform data, and displacement waveform data; and pooling layers can select extreme values in local regions to reduce the length of the input sequence, thereby extracting key feature data and improving the robustness of the pre-trained prediction model.
[0074] In some alternative embodiments, such as Figure 4 As shown, the first feature extraction module may further include a random deactivation layer. The feature extraction data may include acceleration feature data, velocity feature data, and displacement feature data. Before performing pooling operations on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data, S1014, the initial prediction model is trained based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model. This further includes: randomly determining effective feature data from the feature extraction data based on the random deactivation layer; the effective feature data includes at least one of the acceleration feature data, velocity feature data, and displacement feature data.
[0075] The step of performing a pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data includes: performing a pooling operation on the effective feature data based on the pooling layer to obtain the dimensionality-reduced feature data.
[0076] The acceleration feature data includes features extracted from acceleration waveform data using convolutional layers; the velocity feature data includes features extracted from velocity waveform data using convolutional layers; and the displacement feature data includes features extracted from displacement waveform data using convolutional layers. The drop rate of the random deactivation layer (Dropout layer) can be 0.2 or other reasonable values. The random deactivation layer randomly determines the effective features from the extracted data, meaning that some extracted features can be randomly discarded during training. This reduces the dependence of the pre-trained prediction model on specific feature data, lowers the possibility of overfitting, and improves the generalization ability of the pre-trained prediction model.
[0077] In some alternative embodiments, such as Figure 4 As shown, the first feature extraction module may further include a bidirectional gating unit layer. Before the first fully connected module outputs the initial model prediction parameters based on the dimensionality reduction feature data, S1014, the initial prediction model is trained based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model, which further includes: obtaining feature representation data of the dimensionality reduction feature data based on the bidirectional gating unit layer;
[0078] The step of outputting initial model prediction parameters based on the dimensionality reduction feature data by the first fully connected module includes: outputting the initial model prediction parameters based on the feature representation data by the first fully connected module.
[0079] The output dimension of the bidirectional gated unit layer (BIGRU layer) can be 64 or other reasonable values. The bidirectional gated unit layer helps the pre-trained prediction model better capture long-term dependencies in sequences, thereby improving the performance of the pre-trained prediction model.
[0080] In some optional embodiments, S1013, calculating the local ground motion parameters at the arrival of the earthquake P-wave based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data, includes: filtering the acceleration waveform data, the velocity waveform data, and the displacement waveform data respectively to obtain filtered acceleration waveform data, velocity waveform data, and displacement waveform data; calculating the local ground motion parameters at the arrival of the earthquake P-wave based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data; wherein, S1014, training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model, includes: training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model.
[0081] The filtering of acceleration, velocity, and displacement waveform data can be achieved using a Butterworth high-pass filter (specifically a fourth-order Butterworth high-pass filter) or a Chebyshev filter. For example, a Butterworth high-pass filter can be used to filter vertical, east-west, and north-south acceleration data to obtain processed vertical, east-west, and north-south acceleration data, respectively. By filtering the acceleration, velocity, and displacement waveform data, noise can be removed, improving the data quality of the resulting filtered acceleration, velocity, and displacement waveforms.
[0082] In some optional embodiments, before training the initial prediction model based on the acceleration waveform data at the arrival of the earthquake P-wave in S101 to obtain the pre-trained prediction model, the method further includes: automatically picking up the monitored vertical acceleration data at the arrival of the P-wave, and obtaining the acceleration waveform data at the arrival of the earthquake P-wave based on the monitored triaxial acceleration waveform data.
[0083] Specifically, the monitored vertical acceleration data can be automatically picked up at the arrival time of the P-wave based on algorithms such as Long Short-Time Average (STA / LTA) and Akaike Criterion (AIC) to obtain the acceleration waveform data at the arrival of the earthquake's P-wave. Before automatically picking up the P-wave arrival time of the monitored vertical acceleration data, baseline correction can be performed on the monitored triaxial acceleration waveform data to improve the accuracy of the automatic P-wave arrival time picking results, thereby improving the accuracy of the lightweight local ground motion parameter prediction of the constructed model. By automatically picking up the P-wave arrival time of the monitored vertical acceleration data to obtain the acceleration waveform data at the arrival time of the earthquake's P-wave, the consistency between the input data during model training and the input data during application can be improved, thus improving the accuracy of the lightweight local ground motion parameter prediction of the constructed model.
[0084] This application embodiment also provides an earthquake early warning system, the system including: a server and an early warning unit;
[0085] The server is configured to predict local ground motion parameters based on a lightweight local ground motion parameter prediction model constructed according to any of the methods described in the first aspect above.
[0086] The early warning unit is configured to provide earthquake early warning based on the local ground motion parameters.
[0087] The server can be implemented using a microcontroller, a single-chip microcomputer, or a programmable logic control unit. The specific implementation of this earthquake early warning system can be found in the descriptions above.
[0088] This application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the lightweight on-site ground motion parameter prediction model construction method as described in any of the first aspects above.
[0089] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the structure of an electronic device 200 provided in an embodiment of this application. The electronic device 200 includes: a memory 202 and a processor 201; the memory 202 stores a computer program executable by the processor 201, and when the computer program is executed by the processor 201, it executes the lightweight on-site ground motion parameter prediction model construction method described in any one of the first aspects.
[0090] The memory 202 and the processor 201 can be interconnected and communicate with each other via a communication bus 203 and / or other forms of connection mechanism (not shown). The memory 202 stores a computer program executable by the processor 201, which, when executed by the processor 201, performs the lightweight on-site ground motion parameter prediction model construction method described in the first aspect above.
[0091] This application also provides a computer-readable storage medium storing computer program instructions, which, when executed by processor 201, perform the lightweight on-site ground motion parameter prediction model construction method described in the first aspect above.
[0092] The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0093] It should be understood that the disclosed apparatus / systems and methods can also be implemented in other ways, as provided in the embodiments of this application. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0094] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0095] The above description is only an optional implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application.
Claims
1. A method for constructing a light-weighted local ground motion parameter prediction model, characterized in that, The method includes: Based on the acceleration waveform data at the arrival of the earthquake P-wave, the initial prediction model is trained to obtain a pre-trained prediction model. The pre-trained prediction model is used as the teacher model. The knowledge of the teacher model is transferred to the student model through distillation. The student model is then trained based on the acceleration waveform data. A lightweight on-site ground motion parameter prediction model is constructed based on the trained student model. The student model includes a first feature extraction module and a second fully connected module. The first feature extraction module obtains features from the teacher model through distillation. The second fully connected module includes a non-linear activation function layer and a linear activation function layer. The first feature extraction module includes a one-dimensional convolutional layer, a random deactivation layer, a one-dimensional max pooling layer, and a bidirectional gated unit layer connected in sequence. The one-dimensional convolutional layer has 16 kernels, a kernel size of 3, and a stride of 1. The dropout rate of the random deactivation layer is 0.
2. The pooling window size of the one-dimensional max pooling layer is 4. The output dimension of the bidirectional gated unit layer is 64. The lightweight local ground motion parameter prediction model is obtained by pruning the nonlinear activation function layer in the student model; wherein, the sparsity of the student model in the lightweight local ground motion parameter prediction model reaches a preset sparsity threshold. The lightweight in-situ ground motion parameter prediction model is used for deployment in resource-constrained equipment. The process of training the initial prediction model based on the acceleration waveform data at the arrival of the earthquake P-wave to obtain a pre-trained prediction model includes: The acceleration waveform data is integrated to obtain velocity waveform data; The velocity waveform data is integrated to obtain displacement waveform data; Based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data, the local ground motion parameters at the arrival of the earthquake P-wave are calculated; wherein, the local ground motion parameters include at least one of the following: high-speed rail early warning parameters, gas early warning parameters, nuclear power early warning parameters, instrument intensity parameters, and aftershock rescue parameters; The initial prediction model is trained based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model.
2. The method according to claim 1, characterized in that, in, The pre-trained prediction model includes a trained first feature extraction module and a first fully connected module; the step of using the pre-trained prediction model as a teacher model and transferring the knowledge of the teacher model to the student model through distillation includes: The pre-trained prediction model is used as the teacher model, and the first feature extraction module is transferred to the student model through distillation; The construction of a lightweight on-site ground motion parameter prediction model based on the student model includes: The acceleration waveform data, the velocity waveform data, and the displacement waveform data are input into the student model to obtain the prediction parameters output by the student model. Based on the local ground motion parameters and the predicted parameters, the parameters of the second fully connected module in the student model are adjusted until the similarity between the predicted parameters output by the adjusted student model and the local ground motion parameters reaches a preset similarity threshold. The lightweight on-site ground motion parameter prediction model is constructed based on the adjusted student model.
3. The method of claim 2, wherein, in, The first fully connected module includes a linear activation function layer; the construction of the lightweight local ground motion parameter prediction model based on the adjusted student model includes: Pruning is performed on the nonlinear activation function layer in the adjusted student model until the sparsity of the student model after pruning reaches a preset sparsity threshold. The student model after pruning is determined as the lightweight on-site ground motion parameter prediction model.
4. The method of claim 2, wherein, The step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model includes: Based on the convolutional layer, feature extraction is performed on the acceleration waveform data, the velocity waveform data, and the displacement waveform data to obtain feature extraction data; The feature extraction data is pooled based on the pooling layer to obtain dimensionality-reduced feature data. Based on the first fully connected module, the initial model prediction parameters are output according to the dimensionality reduction feature data; Based on the initial model prediction parameters and the local ground motion parameters, the internal parameters of the initial prediction model are adjusted to obtain the pre-trained prediction model.
5. The method of claim 4, wherein, in, The feature extraction data includes acceleration feature data, velocity feature data, and displacement feature data; before performing pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data, the step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model further includes: Based on the random deactivation layer, valid feature data in the feature extraction data are randomly determined; the valid feature data includes at least one of the acceleration feature data, the velocity feature data, and the displacement feature data. The step of performing a pooling operation on the feature extraction data based on the pooling layer to obtain dimensionality-reduced feature data includes: The effective feature data is pooled based on the pooling layer to obtain the dimensionality-reduced feature data.
6. The method of claim 4, wherein, Before the first fully connected module outputs the initial model prediction parameters based on the dimensionality reduction feature data, the step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model further includes: The feature representation data of the dimensionality-reduced feature data is obtained based on the bidirectional gating unit layer; The step of outputting initial model prediction parameters based on the dimensionality reduction feature data by the first fully connected module includes: Based on the feature representation data, the first fully connected module outputs the initial model prediction parameters.
7. The method of claim 1, wherein, The step of calculating the ground motion parameters at the arrival of the earthquake P-wave based on the acceleration waveform data, the velocity waveform data, and the displacement waveform data includes: The acceleration waveform data, the velocity waveform data, and the displacement waveform data are filtered respectively to obtain filtered acceleration waveform data, velocity waveform data, and displacement waveform data. Based on the acceleration waveform processing data, the velocity waveform processing data, and the displacement waveform processing data, calculate the local ground motion parameters at the arrival of the earthquake P-wave. The step of training the initial prediction model based on the acceleration waveform data, the velocity waveform data, the displacement waveform data, and the local ground motion parameters to obtain the pre-trained prediction model includes: The initial prediction model is trained based on the acceleration waveform processing data, the velocity waveform processing data, the displacement waveform processing data, and the local ground motion parameters to obtain the pre-trained prediction model.
8. A seismic early warning system characterized by, The system includes: a server and an early warning unit; The server is configured to predict local ground motion parameters based on a lightweight local ground motion parameter prediction model constructed according to any one of claims 1-7. The early warning unit is configured to provide earthquake early warning based on the local ground motion parameters.
9. An electronic device, comprising: The electronic device includes: Memory; processor; The memory stores a computer program executable by the processor, which, when executed by the processor, performs the method described in any one of claims 1-7.