Earthquake safety risk prediction method based on data analysis

Abstract

This invention belongs to the field of neural network technology, specifically relating to a data analysis-based method for earthquake safety risk prediction. The method includes: acquiring multi-source monitoring data and building static data to construct a training sample set; replacing the fixed fundamental frequency with an adaptive fundamental frequency and constructing a preliminary physical constraint loss function based on the initial estimate of the fundamental frequency; introducing an asymmetric stiffness degradation mechanism to construct an optimized loss function for the anomalous positive change of the adaptive fundamental frequency; constructing a spatial graph structure based on geographic Euclidean distance and site condition parameters, and applying graph Laplace regularization constraints to obtain the final loss function; and inputting real-time ground motion sequences into the neural network model to output the results. This invention can overcome the physical distortion caused by the nonlinear evolution of buildings under strong earthquake scenarios, reduce the cumulative prediction error in the later stages of the sequence, eliminate the risk distribution jumps between adjacent homogeneous buildings, and output earthquake response prediction results for urban high-rise building clusters with mechanical self-consistency and spatial continuity.

Description

Technical Field

[0001] This invention relates to the field of neural network technology. More specifically, this invention relates to a data analysis-based method for predicting earthquake safety risks. Background Technology

[0002] In the field of seismic response prediction for urban high-rise building clusters, current mainstream methods are gradually shifting from purely data-driven neural network models to temporal prediction architectures that integrate physical information. Among these, lightweight models, such as Convolutional Long Time Expression Memory (ConvLEM), effectively balance prediction accuracy and computational efficiency thanks to their temporal memory mechanism, which combines spatial feature extraction capabilities with linear complexity. The core advantage of this approach lies in introducing structural dynamics equations as constraints. During model training, the predicted values ​​of displacement, velocity, and acceleration are coupled to a physically consistent solution space, significantly improving the physical plausibility and engineering interpretability of the prediction results compared to traditional black-box models. Simultaneously, the inherent low inference latency of ConvLEM makes it potential for real-time deployment at the urban fringe, providing technical support for second-level seismic response. This paradigm constitutes the current mainstream technical framework in this field, demonstrating significant progress in both physical information embedding and computational efficiency adaptation.

[0003] However, existing ConvLEM models based on physical information constraints still have certain limitations. While these models can achieve high-precision predictions for single buildings by introducing residual constraints in the equations of motion when training data is sufficient and the structure is within its elastic range, they still rely on fixed elastic parameter assumptions when facing strong earthquake scenarios involving high-rise building clusters. They cannot adaptively learn the nonlinear evolution of structural stiffness degradation during strong earthquakes, leading to a significant amplification of prediction errors in the later stages of the epicenter. Furthermore, existing methods typically perform isolated predictions for each building, ignoring the inherent spatial correlation of responses between adjacent buildings due to similar site conditions. This ultimately results in discontinuous and fluctuating risk distribution maps. Summary of the Invention

[0004] To address the technical problems of existing models relying on fixed parameters, which cannot adapt to the nonlinear degradation of structural stiffness under strong earthquakes, leading to amplified errors, and isolated predictions ignoring the spatial correlation between buildings, resulting in distorted risk distribution, this invention provides a data analysis-based earthquake safety risk prediction method, comprising: Acquire multi-source monitoring data and building static data, preprocess them, and construct a training sample set; The initial estimate of the fundamental frequency of the target high-rise building is obtained by using static building data. The fixed fundamental frequency in the motion equation of the single-degree-of-freedom system is replaced with an adaptive fundamental frequency. A preliminary physical constraint loss function is constructed based on the deviation between the adaptive fundamental frequency and the initial estimate of the fundamental frequency. The change in the adaptive fundamental frequency between adjacent time series is calculated, and an asymmetric stiffness degradation mechanism is introduced by penalizing abnormal positive changes, thereby constructing an optimized loss function; A spatial graph structure is constructed based on the geographical Euclidean distance between the target high-rise buildings and site condition parameters. The difference in the predicted response vectors between the target high-rise buildings is constrained by graph Laplace regularization using the spatial graph structure to obtain the final loss function. Real-time ground motion sequences are input into a neural network model optimized based on the final loss function for forward inference, and the results of seismic safety risk in the target area are output and visualized.

[0005] Preferably, the step of acquiring multi-source monitoring data and building static data, performing preprocessing, and constructing a training sample set includes: Acquire real-time ground motion monitoring data and extract the structural type and latitude and longitude coordinates of the target high-rise building; perform noise reduction processing on the real-time ground motion monitoring data, extract the effective waveform segments within a specific time period after the arrival of the first arrival P-wave, and unify the effective waveform segments with different sampling rates to the same time reference to obtain aligned waveform segments; Normalize the aligned waveform segments to obtain standardized ground motion segments, and obtain site condition parameters based on the average shear wave velocity at the location of the target high-rise building. A training sample set is constructed by combining and pairing standardized ground motion segments, the structural type of the target high-rise building, latitude and longitude coordinates, and site condition parameters.

[0006] Preferably, the preliminary physical constraint loss function satisfies the expression: ; In the formula, This represents the initial physical constraint loss function; This represents the mean squared error loss function; Indicates the total number of time-series sampling points; Indicates the time sequence number; Indicates the time sequence number as The predicted value of the acceleration response; Indicates the time sequence number as Predicted speed response value; Indicates the time sequence number as The predicted displacement response; Indicates the time sequence number as The input acceleration value of the seismic motion; Indicates the structural damping ratio; Indicates adaptive base frequency; This represents the initial estimated value of the fundamental frequency; This represents the hyperparameter of the first physical constraint strength; This represents the hyperparameter of the second physical constraint strength.

[0007] Preferably, the optimization loss function satisfies the expression: ; In the formula, This represents the optimization loss function; This represents the initial physical constraint loss function; Indicates the asymmetric penalty hyperparameter; Indicates the total number of time-series sampling points; Indicates the time sequence number; Indicates the time sequence number as Adaptive base frequency; Indicates the time sequence number as Adaptive base frequency; Indicates the time step; This represents the linear rectification function.

[0008] Preferably, the final loss function satisfies the expression: ; In the formula, Represents the final loss function; This represents the optimization loss function; Indicates the spatial cooperative hyperparameters; This represents the total number of edges in the spatial graph structure; Represents the set of edges in a spatial graph structure; Indicates the serial number of the first high-rise building; Indicates the serial number of the second high-rise building; This indicates the spatial coordination weight between the first and second tallest buildings. This represents the predicted response vector corresponding to the first high-rise building; This represents the predicted response vector corresponding to the second tallest building.

[0009] Preferably, the spatial collaborative weights satisfy the expression: ; In the formula, This indicates the spatial coordination weight between the first and second tallest buildings. This represents the geographical Euclidean distance between the first and second tallest buildings. Indicates the distance-bandwidth parameter; This represents the average shear wave velocity at 30 meters above the ground corresponding to the tallest building. This represents the average shear wave velocity at 30 meters above the ground corresponding to the second tallest building. Indicates the bandwidth parameter of the site conditions; This represents an exponential function with the natural constant as its base.

[0010] Preferably, the step of inputting the real-time seismic motion sequence into the neural network model optimized based on the final loss function for forward inference includes: The real-time three-component acceleration waveform is continuously acquired within a specific time after the arrival of the first arrival longitudinal wave, and the real-time three-component acceleration waveform is fused with the static configuration data of the target high-rise building and then input into the neural network model. By activating the forward propagation computation link of the neural network model, skipping all loss determination nodes, the response time history and probabilistic damage index of the target high-rise building are inferred and output. The response time history and probabilistic damage index were compared with the threshold of the benchmark test library of the seismic design code, and the safety level of the target high-rise building was determined by combining the confidence interval.

[0011] Preferably, determining the safety level of the target high-rise building by combining the confidence interval further includes: The safety level of all target high-rise buildings is hierarchically nested with the original spatial latitude and longitude coordinates and imported into the underlying geographic information system to generate a spatially continuous and smooth risk heat map. By combining the emergency management logic protocol stack, the system simultaneously outputs a list of high-risk buildings, an electronic fence for evacuating dangerous areas, and a coordinate scheme for the pre-deployment of rescue resources.

[0012] The beneficial effects of this invention are as follows: By transforming the fixed fundamental frequency in the traditional dynamic equation into an adaptive fundamental frequency constrained by the initial estimate, and by imposing a penalty on its abnormal positive change to introduce an asymmetric stiffness degradation mechanism, the neural network can adaptively learn during training and strictly follow the nonlinear physical law of irreversible loss of structural stiffness under strong earthquakes. This effectively overcomes the physical distortion caused by fixed parameters in traditional models and significantly reduces the cumulative prediction error in the later stages of the epicenter. At the same time, this invention constructs a spatial graph structure based on geographical distance and site conditions and performs graph Laplace regularization constraints, breaking the limitations of existing isolated reasoning with single or multiple samples. By forcing adjacent building groups with similar geological properties to maintain response coordination and smoothness, it effectively eliminates the discrete jumps in risk distribution caused by local small noises. Finally, it can output earthquake safety risk prediction results for the target area with underlying physical self-consistency and macroscopic spatial continuity based on forward inference. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating the process of a data analysis-based earthquake safety risk prediction method according to the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] This invention discloses a data analysis-based method for predicting earthquake safety risks, referring to... Figure 1 This includes steps S1-S5: S1. Acquire multi-source monitoring data and building static data and preprocess them to construct a training sample set.

[0017] It should be noted that ground motion monitoring data and building digital archives have different data formats and sampling frequencies. Therefore, time synchronization, alignment, and standardization are essential to ensure the temporal consistency and dimensional uniformity of the input features to the neural network model, thereby eliminating the interference of magnitude differences on the stability of subsequent model calculations. Furthermore, traditional pure data-driven networks lack the ability to handle the complex spatial correlations and long-term nonlinear evolution characteristics of urban high-rise building clusters, resulting in severely insufficient accuracy in predicting responses after strong earthquakes. Therefore, this invention constructs a convolutional long-term representation memory neural network, utilizing its spatial feature extraction capabilities and temporal memory mechanism as the foundational framework for subsequent physical constraints and feature learning.

[0018] Specifically, real-time ground motion monitoring data is obtained from strong earthquake monitoring stations and microelectromechanical system accelerometer arrays deployed in urban areas, while the structural type and latitude and longitude coordinates of the target high-rise building are extracted from pre-established building digital archives and geological survey databases.

[0019] The three-component acceleration time history data acquired by strong-motion stations and microelectromechanical system (MEMS) accelerometers are denoised to remove environmental noise and anomalous spikes, and effective waveform segments within a specific time period after the arrival of the first-arrival P-wave are extracted. Effective waveform segments with different sampling rates are then interpolated to the same time base to obtain aligned waveform segments, ensuring strict alignment between the ground motion input sequence and the building response time sequence.

[0020] The aligned waveform segments are mapped to a standard interval of 0 to 1 using the maximum and minimum value normalization method to obtain standardized ground motion segments.

[0021] The average shear wave velocity at 30 meters above the ground at the location of the target high-rise building is extracted from the geological survey database and standardized as an input parameter for spatial constraints to obtain site condition parameters.

[0022] Standardized ground motion segments are combined and paired with the structural type, latitude and longitude coordinates, and site condition parameters of the target high-rise building to construct a training sample set for supervised learning. The training sample set is then divided into a training set, a validation set, and a test set according to a conventional preset ratio.

[0023] It should be noted that the specific duration refers to a time window of 1 to 3 seconds after the arrival of the initial P-wave. In other embodiments, implementers can dynamically adjust the selection range of the specific duration and the alignment benchmark of the data sampling rate according to the actual station deployment and network transmission characteristics. The conventional preset ratio is 8:1:1, that is, the training set accounts for 0.8, the validation set accounts for 0.1, and the test set accounts for 0.1. This ratio ensures that the neural network obtains sufficient samples to fit complex feature mappings, while reserving independent data for structure verification and generalization ability evaluation, effectively preventing overfitting. In other embodiments, implementers can adaptively adjust this ratio according to the total data volume of the training sample set.

[0024] A convolutional long-term representation memory neural network is constructed, and the constructed training sample set is used as the input basis for the convolutional long-term representation memory neural network to perform subsequent forward feature propagation and backward parameter optimization.

[0025] S2. Use static building data to obtain the initial estimate of the fundamental frequency of the target high-rise building, replace the fixed fundamental frequency in the motion equation of the single-degree-of-freedom system with the adaptive fundamental frequency, and construct a preliminary physical constraint loss function based on the deviation between the adaptive fundamental frequency and the initial estimate of the fundamental frequency.

[0026] It should be noted that existing physical information neural networks have a standard mean squared error loss. When structural dynamics equations are introduced as constraints, the fundamental frequency is forcibly set to a fixed value. This fixed assumption only holds true in the elastic physics stage and cannot reflect the actual evolution of stiffness degradation and damping dissipation that occurs after a building enters a nonlinear state under strong earthquakes. This leads to a significant accumulation of prediction errors as the earthquake intensity increases. To address this issue, this invention transforms the fundamental frequency in the equations of motion of a single-degree-of-freedom system from a fixed parameter into an adaptive learnable parameter that participates in backpropagation optimization along with the network weights. Simultaneously, a parameter deviation penalty is introduced, enabling the model to adaptively adjust the dynamic parameters during training to match the dynamic characteristics of different buildings. This also prevents the adaptive fundamental frequency from deviating excessively from the physical prior after iterative adjustment, thus avoiding a decrease in accuracy.

[0027] Specifically, the initial physical constraint loss function satisfies the expression: ; In the formula, This represents the initial physical constraint loss function; This represents the mean squared error loss function; Indicates the total number of time-series sampling points; Indicates the time sequence number; Indicates the time sequence number as The predicted value of the acceleration response; Indicates the time sequence number as Predicted speed response value; Indicates the time sequence number as The predicted displacement response; Indicates the time sequence number as The input acceleration value of the seismic motion; Indicates the structural damping ratio; Indicates adaptive base frequency; This represents the initial estimated value of the fundamental frequency; This represents the hyperparameter of the first physical constraint strength; This represents the hyperparameter of the second physical constraint strength; This indicates a penalty term for parameter deviation.

[0028] Wherein, mean square error loss function As a baseline for fitting loss to the underlying data; It is a commonly used physical constraint loss formula, which is itself the equation of motion for a single-degree-of-freedom system. It is derived by using the acceleration response predicted by the network. Predicted speed response value and displacement response prediction value Substituting the equations of motion for a single-degree-of-freedom system, the residuals resulting from non-compliance with these equations are calculated. If all three satisfy the equations of motion for a single-degree-of-freedom system, the summation of the residuals is zero. If the fundamental physical laws of structural dynamics are not satisfied, the summation of the residuals produces a positive loss, forcing the network to adaptively adjust node parameters to conform to the physical laws of structural dynamics. This invention uses the fixed original fundamental frequency of the physical constraint loss formula as the adaptive fundamental frequency. Parameter deviation constraints are applied to it, thereby introducing a parameter deviation penalty term. ,in, This is the initial estimate of the fundamental frequency, obtained directly from the building height using existing empirical formulas for natural vibration periods, with the addition of a parameter deviation penalty term. It can prevent adaptive base frequency Excessive deviation from physical priors leads to a decrease in accuracy.

[0029] First physical constraint strength hyperparameter With the second physical constraint strength hyperparameter The first physical constraint strength hyperparameter is used to balance the tension between the main task of data regression and the constraints of the physical equations. When its value is large, the model tends to obey the physical equations more, and the parameter adjustment range is larger; when its value is small, the model focuses more on the actual sample changes, and the parameter adjustment range is smaller. With the second physical constraint strength hyperparameter The value of is typically determined by the implementers using a grid search method within a general range of 0.001 to 1000 to find the optimal equilibrium point. Structural damping ratio Typically, the value is preset to 0.05 based on the material properties of reinforced concrete. In other embodiments, implementers can make targeted adjustments based on the specific physical properties of building materials in the actual scenario.

[0030] S3. Calculate the change in the adaptive fundamental frequency between adjacent time series, and introduce an asymmetric stiffness degradation mechanism by penalizing abnormal positive changes, thereby constructing an optimized loss function.

[0031] It should be noted that although the model can adaptively adjust the fundamental frequency to match the dynamic characteristics of different buildings after being modified with learnable parameters, the adaptive fundamental frequency may evolve freely and unrestrainedly over time during the continuous back propagation of a long-term strong earthquake sequence, resulting in an increasing trend over time. According to the basic laws of energy dissipation in real structural dynamics, the stiffness of a building under the cumulative damage of a strong earthquake can only undergo irreversible degradation. That is, the adaptive fundamental frequency reflecting stiffness can only decrease or remain unchanged over time, and there is absolutely no possibility of spontaneous enhancement at the physical level. Therefore, this invention introduces an asymmetric regularization term into the loss function, imposing a severe penalty only on cases where the adaptive fundamental frequency exhibits an abnormal increase, while not imposing any constraint on cases where the adaptive fundamental frequency decreases normally or remains unchanged. This embeds a priori physical constraints on irreversible stiffness degradation, complementing the parameter deviation penalty term in step S2. The parameter deviation penalty prevents the adaptive fundamental frequency from deviating excessively from the initial estimate, while the asymmetric degradation constraint prevents the adaptive fundamental frequency from exhibiting a physically impossible increasing trend in the time series. Together, they ensure that the learnable parameters not only have adaptive capabilities but also strictly follow the actual mechanical behavior of the structure under strong earthquakes, providing a physically correct single-building response basis for subsequent spatial collaborative constraints.

[0032] Specifically, the optimized loss function satisfies the expression: ; In the formula, This represents the optimization loss function; This represents the initial physical constraint loss function; Indicates the asymmetric penalty hyperparameter; Indicates the total number of time-series sampling points; Indicates the time sequence number; Indicates the time sequence number as The adaptive base frequency, i.e., the adaptive base frequency at the next moment; Indicates the time sequence number as The adaptive base frequency, i.e., the adaptive base frequency at the current moment; Indicates the time step; This represents the linear rectification function.

[0033] Among them, the adaptive base frequency at the next moment is utilized Subtract the adaptive base frequency at the current moment Obtain the adaptive fundamental frequency change and divide it by the time step. As a normalization coefficient, it is converted into a rate of change to eliminate dimensional differences caused by devices with different sampling frequencies. Since building stiffness is proportional to the square of the adaptive fundamental frequency, when this rate of change is greater than 0, it means that the stiffness is increasing in a way that defies physical laws. In this case, a linear rectification function is used. The abnormal positive rate of change is retained and amplified; when the rate of change is less than 0, it means that normal building stiffness loss has degraded, and the linear rectification function... We force it to zero so that this physically consistent state does not interfere with or penalize the total loss function. We then square the extracted anomalous rate of change so that the more severe the spontaneous reinforcement fallacy, the more severe the correction penalty. The asymmetric penalty term is used, and the arithmetic mean of the penalty term over all time steps is taken. This forcibly decouples the optimization logic of this loss from the total duration of the seismic motion sequence. This effectively prevents long input samples from having excessively large accumulated penalty gradients during backpropagation due to too many steps, which could incorrectly dominate the overall training convergence direction, thus maintaining the general stability of hyperparameters across sample types. Asymmetric penalized hyperparameters. The optimal equilibrium point is determined by the implementers using a grid search method within a general range of 0.001 to 1000.

[0034] S4. Construct a spatial graph structure based on the geographical Euclidean distance and site condition parameters between the target high-rise buildings. Use the spatial graph structure to apply graph Laplacian regularization constraints to the differences in the predicted response vectors between the target high-rise buildings to obtain the final loss function.

[0035] It should be noted that the independent single-building model prediction mechanism does not consider the spatial correlation and transmission effect that inevitably exists between adjacent high-rise buildings in urban building clusters due to the similarity of their foundation and geological site conditions. This may lead to the final disaster distribution not matching the spatial continuity characteristics of real disasters. Two adjacent high-rise buildings with the same height and structural type should bear almost the same seismic input and produce similar damage levels. However, if the model predicts independently, it may give extreme differences in results, with one risk index of 0.2 and the other of 0.9, due to random bias or small noise in the training data. To solve this problem, this invention introduces a graph Laplace regularized spatial constraint architecture, transforming urban building clusters into a spatial graph structure with geographical distance and site conditions as edge weights. By adjusting the original loss function through two-factor spatial collaborative constraints, it can ensure that building clusters that are not only geographically close but also have homogeneous geological foundations maintain smooth and coordinated responses. At the same time, it automatically blocks forced correlations between buildings that are geographically close but have completely different foundation soil layers, thereby eliminating discontinuous jumps in risk heat maps caused by small independent noise.

[0036] Specifically, a spatial graph structure encompassing the relationships of the target high-rise building complex is constructed based on the K-nearest neighbor rule to determine the set of edges. and their total quantity.

[0037] The final loss function satisfies the expression: ; ; In the formula, Represents the final loss function; This represents the optimization loss function; Indicates the spatial cooperative hyperparameters; This represents the total number of edges in the spatial graph structure; Represents the set of edges in a spatial graph structure; Indicates the serial number of the first high-rise building; Indicates the serial number of the second high-rise building; This indicates the spatial coordination weight between the first and second tallest buildings. This represents the predicted response vector corresponding to the first high-rise building; This represents the predicted response vector corresponding to the second tallest building. This represents the geographical Euclidean distance between the first and second tallest buildings. Indicates the distance-bandwidth parameter; This represents the average shear wave velocity at 30 meters above the ground corresponding to the tallest building. This represents the average shear wave velocity at 30 meters above the ground corresponding to the second tallest building. Indicates the bandwidth parameter of the site conditions; This represents an exponential function with the natural constant as its base.

[0038] Among them, the predicted response vector and predicted response vector It integrates the displacement time history, velocity time history, acceleration time history, and internal damage index output by the neural network; and the distance bandwidth parameter. The spatial span scale is directly determined by those skilled in the art through calculating the median of the geographic Euclidean distances between all building points in the sample set; site condition bandwidth parameters. The distribution variance is directly obtained by the implementers through the calculation of the variation function characteristics of the site survey data of the input area, and is used to standardize the geological tolerance scale.

[0039] Index Term This invention quantifies physical proximity; buildings that are closer together should have similar responses. It uses a Gaussian kernel function to map geographic Euclidean distance to weights between 0 and 1. The closer the value is to 0, the closer the weight is to 1, when the geographical Euclidean distance... When the distance bandwidth parameter is greater than the value of the exponential term, the weight approaches 0. The consistency of geological properties was quantified; when the site conditions of two buildings are similar, Approaching 1, when the site differences are large. Approaching 0. Spatial synergy weight only occurs when two high-rise buildings are not only geographically close but also exhibit highly consistent geological properties such as deep soil layers or bedrock. Only if the value is not zero can the difference between the predicted response vectors of the two be reduced by using the graph Laplacian regularization term; if either condition does not satisfy the similarity constraint, spatial co-weighting is used. The forced smoothing constraint automatically fails when the temperature rapidly decays to 0. The core logic of the graph Laplacian regularization term is to drive adjacent high-rise building nodes in the spatial association graph structure to output similar predicted response results. By averaging the accumulated results using the total number of edges $E$ in the spatial graph, the linear expansion of the total loss value caused by the surge in regional building size and edge count can be prevented, ensuring the stability of spatial coordination hyperparameters. It maintains stable versatility across prediction scenarios with varying building densities and topological scales.

[0040] Spatial Cooperative Hyperparameters The optimal equilibrium point is determined by the implementers using a grid search method within a general range of 0.001 to 1000.

[0041] S5. Input the real-time ground motion sequence into the neural network model optimized based on the final loss function for forward inference, and output and visualize the seismic safety risk results of the target area.

[0042] It should be noted that the network model, optimized under the dual constraints of reshaping the dynamic degradation law of individual units and the spatial graph collaborative mechanism, not only deeply matches the seismic damage evolution theory in its internal nonlinear feature mapping path, but also possesses the macroscopic collaborative perception capability of the dynamic characteristics of cross-regional building clusters. During the critical window of disaster occurrence, the system no longer needs to perform any reverse gradient updates; it only needs to rely on a single, extremely low-latency forward feature propagation to instantly and in parallel output high-fidelity engineering indicators and damage probabilities for each node.

[0043] Specifically, at the moment when the strong seismic wave swarm contacts the pre-set urban defense line, real-time three-component acceleration waveforms are continuously acquired within a specific time period after the arrival of the first P-wave. The building height, structural system type, fundamental frequency estimate, and structural damping ratio of each target high-rise building are matched from the associated database. The real-time three-component acceleration waveforms are fused with the corresponding static configuration data and fed into a convolutional long-time representation memory neural network trained with the final loss function.

[0044] The system activates only the forward propagation calculation link, skips all loss determination nodes, and quickly infers and outputs the displacement response time history, velocity response time history, acceleration response time history, and probabilistic damage index of the target high-rise building.

[0045] The calculated displacement response time history, velocity response time history, acceleration response time history, and probabilistic damage index are introduced into the benchmark verification library in the current national technical specifications for seismic design for threshold comparison verification. Combined with the confidence interval of the probabilistic output, the safety level of each building is determined. The safety level includes safe, restricted use, and unsafe.

[0046] Finally, the safety status levels of all buildings are hierarchically nested with their original spatial latitude and longitude coordinates, imported into the underlying geographic information system, and a risk heat map with no jumps and spatial continuity is generated. Combined with the emergency management logic protocol stack, a list of high-risk buildings, electronic fences for evacuation of dangerous areas, and coordinate schemes for the pre-deployment of rescue resources are output simultaneously.

[0047] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

[0048] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A data analysis-based method for predicting earthquake safety risks, characterized in that, include: Acquire multi-source monitoring data and building static data, preprocess them, and construct a training sample set; The initial estimate of the fundamental frequency of the target high-rise building is obtained by using static building data. The fixed fundamental frequency in the motion equation of the single-degree-of-freedom system is replaced with an adaptive fundamental frequency. A preliminary physical constraint loss function is constructed based on the deviation between the adaptive fundamental frequency and the initial estimate of the fundamental frequency. The change in the adaptive fundamental frequency between adjacent time series is calculated, and an asymmetric stiffness degradation mechanism is introduced by penalizing abnormal positive changes, thereby constructing an optimized loss function; A spatial graph structure is constructed based on the geographical Euclidean distance between the target high-rise buildings and site condition parameters. The difference in the predicted response vectors between the target high-rise buildings is constrained by graph Laplace regularization using the spatial graph structure to obtain the final loss function. Real-time ground motion sequences are input into a neural network model optimized based on the final loss function for forward inference, and the results of seismic safety risk in the target area are output and visualized.

2. The earthquake safety risk prediction method based on data analysis according to claim 1, characterized in that, The process of acquiring multi-source monitoring data and building static data, preprocessing them, and constructing a training sample set includes: Acquire real-time ground motion monitoring data and extract the structural type and latitude and longitude coordinates of the target high-rise building; perform noise reduction processing on the real-time ground motion monitoring data, extract the effective waveform segments within a specific time period after the arrival of the first arrival P-wave, and unify the effective waveform segments with different sampling rates to the same time reference to obtain aligned waveform segments; Normalize the aligned waveform segments to obtain standardized ground motion segments, and obtain site condition parameters based on the average shear wave velocity at the location of the target high-rise building. A training sample set is constructed by combining and pairing standardized ground motion segments, the structural type of the target high-rise building, latitude and longitude coordinates, and site condition parameters.

3. The earthquake safety risk prediction method based on data analysis according to claim 1, characterized in that, The preliminary physical constraint loss function satisfies the expression: ； In the formula, This represents the initial physical constraint loss function; This represents the mean squared error loss function; Indicates the total number of time-series sampling points; Indicates the time sequence number; Indicates the time sequence number as The predicted value of the acceleration response; Indicates the time sequence number as Predicted speed response value; Indicates the time sequence number as The predicted displacement response; Indicates the time sequence number as The input acceleration value of the seismic motion; Indicates the structural damping ratio; Indicates adaptive base frequency; This represents the initial estimated value of the fundamental frequency; This represents the hyperparameter of the first physical constraint strength; This represents the hyperparameter of the second physical constraint strength.

4. The earthquake safety risk prediction method based on data analysis according to claim 1, characterized in that, The optimization loss function satisfies the expression: ； In the formula, This represents the optimization loss function; This represents the initial physical constraint loss function; Indicates the asymmetric penalty hyperparameter; Indicates the total number of time-series sampling points; Indicates the time sequence number; Indicates the time sequence number as Adaptive base frequency; Indicates the time sequence number as Adaptive base frequency; Indicates the time step; This represents the linear rectification function.

5. The earthquake safety risk prediction method based on data analysis according to claim 1, characterized in that, The final loss function satisfies the expression: ； In the formula, Represents the final loss function; This represents the optimization loss function; Indicates the spatial cooperative hyperparameters; This represents the total number of edges in the spatial graph structure; Represents the set of edges in a spatial graph structure; Indicates the serial number of the first high-rise building; Indicates the serial number of the second high-rise building; This indicates the spatial coordination weight between the first and second tallest buildings. This represents the predicted response vector corresponding to the first high-rise building; This represents the predicted response vector corresponding to the second tallest building.

6. The earthquake safety risk prediction method based on data analysis according to claim 5, characterized in that, The spatial collaborative weights satisfy the expression: ； In the formula, This indicates the spatial coordination weight between the first and second tallest buildings. This represents the geographical Euclidean distance between the first and second tallest buildings. Indicates the distance-bandwidth parameter; This represents the average shear wave velocity at 30 meters above the ground corresponding to the tallest building. This represents the average shear wave velocity at 30 meters above the ground corresponding to the second tallest building. Indicates the bandwidth parameter of the site conditions; This represents an exponential function with the natural constant as its base.

7. The earthquake safety risk prediction method based on data analysis according to claim 1, characterized in that, The step of inputting the real-time ground motion sequence into the neural network model optimized based on the final loss function for forward inference includes: The real-time three-component acceleration waveform is continuously acquired within a specific time after the arrival of the first arrival longitudinal wave, and the real-time three-component acceleration waveform is fused with the static configuration data of the target high-rise building and then input into the neural network model. By activating the forward propagation computation link of the neural network model, skipping all loss determination nodes, the response time history and probabilistic damage index of the target high-rise building are inferred and output. The response time history and probabilistic damage index were compared with the threshold of the benchmark test library of the seismic design code, and the safety level of the target high-rise building was determined by combining the confidence interval.

8. The earthquake safety risk prediction method based on data analysis according to claim 7, characterized in that, Determining the safety level of a target high-rise building based on confidence intervals also includes: The safety level of all target high-rise buildings is hierarchically nested with the original spatial latitude and longitude coordinates and imported into the underlying geographic information system to generate a spatially continuous and smooth risk heat map. By combining the emergency management logic protocol stack, the system simultaneously outputs a list of high-risk buildings, an electronic fence for evacuating dangerous areas, and a coordinate scheme for the pre-deployment of rescue resources.

Images