A rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning

Abstract

The present invention relates to the technical field of bridge earthquake damage performance assessment, and specifically discloses a method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning, comprising: real-time acquisition of multi-source data of high-speed railway bridge structures, historical earthquake databases, and BIM models of bridge structures through a heterogeneous data acquisition system; construction of a multimodal feature space comprising partial features including time-frequency domain features, spatial features, and probabilistic features; construction of a hierarchical machine learning architecture based on a dynamic Bayesian network damage propagation model, a hybrid model of a multi-scale convolutional neural network and a long short-term memory network, and an uncertainty quantification and dynamic update mechanism; and establishment of an earthquake damage performance assessment model based on the hierarchical machine learning architecture; use of the earthquake damage performance assessment model to perform earthquake damage performance assessment on the multimodal feature space and generate a damage probability distribution; and effective mining of potential earthquake damage-related information behind the data to more accurately simulate and predict earthquake damage to bridges.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge earthquake damage performance assessment, and in particular to a method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning. Background Art

[0002] High-speed railways, as crucial infrastructure in modern transportation systems, play a crucial role in ensuring mobility and economic development. However, the potential threat of earthquakes poses a significant challenge to the safety of high-speed railway bridges. Accurately and rapidly assessing bridge damage after an earthquake is crucial for implementing effective emergency response measures, ensuring the resumption of railway operations, and mitigating secondary disaster losses.

[0003] Traditional methods for assessing the seismic damage performance of high-speed railway bridges are primarily based on structural mechanics theory and numerical simulation. These methods establish complex mechanical models and calculate the bridge's response to earthquakes using theoretical formulas. While these methods can analyze bridge damage to a certain extent, the calculations are cumbersome, require extensive expertise and computing resources, and are time-consuming, making it difficult to quickly provide assessment results in emergency situations following an earthquake. With the rapid advancement of sensor technology, data acquisition and storage technologies, and machine learning algorithms, machine learning-based rapid assessment methods for the seismic damage performance of high-speed railway bridges have shown great potential. However, relying solely on structural mechanics models fails to fully integrate multi-source heterogeneous data, and the use of a single data type limits comprehensive and accurate assessments of bridge damage performance. Furthermore, earthquakes inherently have uncertainties, and the response of bridge structures to earthquakes also involves numerous uncertainties. Traditional assessment methods have limited means of addressing these uncertainties and cannot accurately quantify their impact on seismic damage performance assessments, resulting in reduced reliability.

[0004] Therefore, the present invention proposes a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning. Summary of the Invention

[0005] The present invention provides a machine learning-based rapid assessment method for earthquake damage performance of high-speed railway bridges. This method involves collecting multi-source data on bridge structures in real time through numerous distributed sensor arrays within a heterogeneous data acquisition system. This data is then combined with a historical earthquake database and a BIM model of the bridge structure to provide a rich and comprehensive data foundation for subsequent analysis. The multi-source data covers the bridge's real-time operational status, the historical earthquake database provides a reference for past earthquakes and bridge responses, and the BIM model presents bridge details from a three-dimensional spatial structural perspective. This combination provides a deeper and more comprehensive understanding of the bridge, laying a solid foundation for accurate earthquake damage performance assessment. A multimodal feature space is constructed, encompassing time-frequency, spatial, and probabilistic features, to extract and integrate features from bridge data across multiple dimensions. Time-frequency features reflect the temporal and frequency variations of the data, spatial features reflect the spatial characteristics of the bridge structure, and probabilistic features account for the uncertainty of event occurrence. The establishment of this multimodal feature space helps to more comprehensively and accurately characterize the characteristics of bridges under earthquakes and uncover potential earthquake damage-related information hidden within the data. A hierarchical machine learning architecture was constructed based on a dynamic Bayesian network damage propagation model, a multiscale CNN-LSTM (convolutional neural network and long short-term memory) hybrid model, and an uncertainty quantification and dynamic update mechanism. A seismic damage performance assessment model was developed. The dynamic Bayesian network effectively handles uncertainty and dynamic changes in information, analyzing the damage propagation process. The multiscale CNN-LSTM hybrid model excels at extracting spatial and temporal features at different scales. The uncertainty quantification and dynamic update mechanism ensure that the model adapts to dynamic changes and uncertainties in the data. This complex and sophisticated architecture empowers the assessment model with powerful learning and analysis capabilities, enabling more accurate simulation and prediction of bridge damage during earthquakes. The seismic damage performance assessment model evaluates damage performance in a multimodal feature space and generates a damage probability distribution, providing intuitive and quantitative assessment results for the seismic damage performance of high-speed railway bridges. The damage probability distribution clearly demonstrates the likelihood of different levels of damage, helping decision-makers quickly understand the risk of bridge damage after an earthquake. This allows them to formulate targeted maintenance, repair, or emergency response plans, effectively ensuring the safe operation of high-speed railway bridges and mitigating potential earthquake losses.

[0006] The present invention provides a method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning, comprising:

[0007] S1: Using numerous distributed sensor arrays deployed in a heterogeneous data acquisition system, multi-source data on high-speed railway bridge structures is collected in real time, along with historical earthquake databases and BIM models of bridge structures.

[0008] S2: Combine multi-source data with the historical earthquake database and the BIM model of the bridge structure to construct a multimodal feature space containing some features of time-frequency domain features, spatial features, and probabilistic features;

[0009] S3: A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model and a multi-scale CNN-LSTM hybrid model, along with uncertainty quantification and dynamic update mechanisms. A seismic damage performance assessment model is then established based on this hierarchical machine learning architecture.

[0010] S4: Use the earthquake damage performance assessment model to evaluate the earthquake damage performance of the multimodal feature space and generate the damage probability distribution of the high-speed railway bridge.

[0011] The preferred method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning is as follows: S1: using a large number of distributed sensor arrays deployed in a heterogeneous data acquisition system to collect multi-source data of high-speed railway bridge structures in real time, and obtaining a historical earthquake database and a BIM model of the bridge structure, including:

[0012] Through the numerous distributed sensor arrays deployed in the heterogeneous data acquisition system, the track plate deformation data, seismic wave response data, overall bridge displacement, and speckle images of each local structure under different intensities of earthquake action of the high-speed railway bridge structure are collected in real time as multi-source data of the high-speed railway bridge structure;

[0013] Obtain a historical earthquake database and a BIM model of a bridge structure, where the BIM model of the bridge structure includes bridge set parameters, material properties, and connection methods.

[0014] The preferred method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning is: S2: combining multi-source data with a historical earthquake database and a BIM model of the bridge structure to construct a multimodal feature space containing partial features of time-frequency domain features, spatial features, and probabilistic features, including:

[0015] Based on the spatiotemporal alignment algorithm, the multi-source data is spatiotemporally aligned with the BIM model of the bridge structure to obtain spatiotemporally aligned source data;

[0016] Extracting time-frequency domain features and spatial features from the spatiotemporal alignment source data;

[0017] Based on the historical earthquake database and the introduction of Monte Carlo simulation, parameter perturbation samples are generated, and then the sensor measurement error and model parameter uncertainty are fused based on evidence theory to obtain probabilistic characteristics;

[0018] A multimodal feature space is constructed based on partial features of time-frequency domain features, spatial features, and probabilistic features.

[0019] Preferably, the rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning extracts time-frequency domain features and spatial features from the spatiotemporal aligned source data, including:

[0020] The energy distribution of the preset frequency band is extracted from the multi-source data aligned in the spatiotemporal alignment source data by short-time Fourier transform, and then decomposed by Hilbert-Huang transform to obtain multiple intrinsic mode functions as time-frequency domain features;

[0021] Constructing the bridge structure stiffness matrix based on the BIM model of the aligned bridge structure in the spatiotemporal alignment source data;

[0022] Extracting damage-sensitive features from spatiotemporally aligned source data;

[0023] Among them, the spatial characteristics include the bridge structure stiffness matrix and damage sensitivity characteristics.

[0024] The preferred method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning extracts damage-sensitive features from spatiotemporally aligned source data, including:

[0025] Based on the phase correlation method, the speckle images of the local structure in the aligned multi-source data in the spatiotemporal aligned source data under different intensity earthquakes are pre-registered at the pixel level, and the full-field displacement vector field is constructed based on all the speckle images after pixel level registration;

[0026] A two-dimensional Delaunay triangulated mesh is constructed based on the full-field displacement vector field, and the Green strain and tension are calculated by counting the displacement gradients of the mesh nodes to generate the strain cloud map of the corresponding level;

[0027] Calculate the warping deformation of the track slab, the damage to the track interlayer connectors, the damage to the supports, and the damage to the bridge piers based on the various strain components in the strain cloud diagram;

[0028] Modal analysis was performed on the time-space aligned source data to extract the first ten frequency change rates of the bridge piers;

[0029] The warping deformation of the track slab and the first ten-order frequency change rate of the bridge pier are regarded as damage-sensitive features.

[0030] Preferably, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning calculates the warping deformation of the track slab, the damage amount of the track interlayer connector, the damage amount of the support, and the damage amount of the bridge pier based on multiple strain components in the strain cloud map, including:

[0031] Establish a calculation model for warping deformation based on strain components and a calculation model for damage to each component;

[0032] De-noising and outlier processing are performed on the strain cloud map to obtain the pre-processed strain cloud map;

[0033] The preprocessed strain cloud map is input into the warping deformation calculation model to obtain the warping deformation of the track plate, and the preprocessed strain cloud map is input into the damage calculation model of each component to obtain the damage amount of the track interlayer connector, the support damage amount and the bridge pier damage amount.

[0034] Preferably, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning is used to perform modal analysis on spatiotemporally aligned source data to extract the first ten frequency change rates of bridge piers, including:

[0035] The bridge pier vibration response signal is extracted from the spatiotemporally aligned source data, and the partial bridge pier vibration response signal within the frequency band of interest is separated from the prying response signal by bandpass filtering.

[0036] The frequency domain decomposition method is used to perform spectrum analysis on the vibration response signals of some bridge piers within the frequency band of interest to identify the first ten natural frequencies;

[0037] Based on the first ten natural frequencies and the first ten natural frequencies of the pier in the reference state, the change rate of the first ten frequencies of the pier is calculated.

[0038] Preferably, the rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning constructs a multimodal feature space based on partial features of time-frequency domain features, spatial features, and probabilistic features, including:

[0039] Visualize the feature contribution status of time-frequency domain features, spatial features, and probability features based on the SHAP value analysis method;

[0040] Based on the recursive feature elimination method and the feature contribution status of time-frequency domain features, spatial features, and probability features, some features whose contribution is not less than the contribution threshold in the time-frequency domain features, spatial features, and probability features are retained and feature dimensionality reduction is performed to obtain a multimodal feature space.

[0041] The preferred method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning is S3: a hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model and a multi-scale CNN-LSTM hybrid model as well as an uncertainty quantification and dynamic update mechanism, and an earthquake damage performance assessment model is established based on the hierarchical machine learning architecture, including:

[0042] A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model, a multi-scale CNN-LSTM hybrid model, and an uncertainty quantification and dynamic update mechanism.

[0043] Construct a multimodal feature space of a large number of samples, and generate a system state vector for each sample based on the multimodal feature space of each sample;

[0044] A earthquake damage performance assessment model is established based on a hierarchical machine learning architecture and the system state vectors of all samples.

[0045] Preferably, the rapid earthquake damage performance assessment method for high-speed railway bridges based on machine learning establishes an earthquake damage performance assessment model based on a hierarchical machine learning architecture and the system state vectors of all samples, including:

[0046] Based on the dynamic Bayesian network damage propagation model in the hierarchical machine learning architecture, the system state vectors of all samples are decomposed into multiple damage levels, and the conditional damage probability matrix of all samples is generated based on the system state vectors of all samples and the corresponding damage levels;

[0047] Based on the conditional damage probability matrix of each sample, the transition probability of the local structural damage state of each bridge in all samples is defined;

[0048] The bridge damage propagation path is modeled based on a discrete-time Markov chain, and the joint probability of the bridge damage propagation path of each sample is calculated based on the transition probability of the local structural damage state of each bridge in each sample;

[0049] Based on the multi-scale CNN-LSTM hybrid model in the hierarchical machine learning architecture, the joint probability of the bridge damage propagation path of each sample is fused at multiple scales to obtain the damage probability distribution of each bridge sample.

[0050] At the same time, uncertainty quantification results are obtained based on the uncertainty quantification and dynamic update mechanism, and the model update frequency is driven based on the uncertainty quantification results until an earthquake damage performance evaluation model is established.

[0051] The beneficial effects of the present invention compared to the prior art are as follows: with the help of numerous distributed sensor arrays in a heterogeneous data acquisition system, multi-source data of the bridge structure can be obtained in real time, including vibration, strain, displacement and other information. At the same time, combined with historical earthquake data and the BIM model of the bridge structure, a rich multimodal feature space can be constructed. With its powerful data mining and pattern recognition capabilities, the machine learning algorithm can automatically learn the complex relationship between features and earthquake damage performance from massive data and establish an efficient and accurate earthquake damage performance assessment model. This method can not only quickly obtain evaluation results, but also improve the accuracy and reliability of the assessment. It is of great significance to improving the emergency response capabilities of high-speed railways in earthquake disasters, and is expected to be widely used in the field of railway engineering safety assurance in the future.

[0052] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the structures specifically pointed out in this application document.

[0053] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0055] Figure 1 This is an architecture diagram of a method for rapid earthquake damage performance assessment of high-speed railway bridges based on machine learning in an embodiment of the present invention;

[0056] Figure 2 1 is a diagram of a multimodal feature fusion architecture in an embodiment of the present invention;

[0057] Figure 3 This is a network architecture diagram of the multi-scale CNN-LSTM hybrid model in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0059] Example 1:

[0060] The present invention provides a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning. Figure 1 ,include:

[0061] S1: Using numerous distributed sensor arrays deployed in a heterogeneous data acquisition system, multi-source data on high-speed railway bridge structures is collected in real time, along with historical earthquake databases and BIM models of bridge structures.

[0062] This step aims to collect a comprehensive data foundation for the subsequent earthquake damage performance assessment. Through the numerous distributed sensor arrays in the heterogeneous data acquisition system, multi-source data of the high-speed railway bridge structure is collected in real time. These sensors may include accelerometers, strain gauges, displacement sensors, etc. The collected data, such as track plate deformation data, seismic wave response data, overall bridge displacement, pier material strain, etc., reflect the real-time status of the bridge during an earthquake from multiple dimensions. At the same time, a historical earthquake database is obtained, which contains relevant information on past earthquake events, such as earthquake intensity, duration, bridge response, etc., to provide a historical reference for analyzing the impact of earthquakes on bridges. In addition, a BIM (Building Information Model) of the bridge structure is obtained. This model covers detailed information such as the geometric parameters, material properties, and connection methods of the bridge, accurately presenting the bridge structure from a three-dimensional perspective, and assisting in understanding the physical characteristics of the bridge and potential earthquake damage associations.

[0063] S2: Combine multi-source data with the historical earthquake database and the BIM model of the bridge structure to construct a multimodal feature space containing some features of time-frequency domain features, spatial features, and probabilistic features;

[0064] This step first uses a spatiotemporal alignment algorithm to align multi-source data with the BIM model of the bridge structure, ensuring temporal and spatial consistency across data from different sources, thereby generating spatiotemporally aligned source data. For example, sensor data collected at different temporal and spatial locations is matched to corresponding locations and time points in the BIM model. Time-frequency and spatial features are then extracted from the spatiotemporally aligned source data. Time-frequency features reflect the temporal and frequency variations of the data. For example, short-time Fourier transforms are used to extract the energy distribution in a preset frequency band, followed by Hilbert-Huang transform decomposition to obtain multiple intrinsic mode functions as time-frequency features. Spatial features reflect the spatial characteristics of the bridge structure, including constructing a stiffness matrix based on the BIM model and extracting damage-sensitive features, such as track slab warping deformation and the frequency change rate of bridge piers through speckle image analysis. Based on a historical earthquake database, Monte Carlo simulation is used to generate parameter perturbation samples. Evidence theory is then used to fuse sensor measurement errors with model parameter uncertainties to generate probabilistic features that account for both earthquake and measurement uncertainties. Finally, a multimodal feature space is constructed based on the time-frequency domain features, spatial features, and probability features to comprehensively characterize the characteristics of the bridge under earthquake action and to explore potential information related to earthquake damage.

[0065] S3: A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model and a multi-scale CNN-LSTM hybrid model, along with uncertainty quantification and dynamic update mechanisms. A seismic damage performance assessment model is then established based on this hierarchical machine learning architecture.

[0066] A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model, a multi-scale CNN-LSTM hybrid model, and an uncertainty quantification and dynamic update mechanism. The dynamic Bayesian network excels at processing uncertainty and dynamically changing information, enabling analysis of damage propagation within bridge structures. The multi-scale CNN-LSTM hybrid model extracts spatial and temporal features at different scales, fully exploring patterns in the data. The uncertainty quantification and dynamic update mechanism ensure that the model can adapt to dynamic data changes and uncertainties.

[0067] A multimodal feature space of a large number of samples is constructed, and a system state vector is generated based on the multimodal feature space of each sample, which comprehensively describes the characteristics of the sample.

[0068] A seismic damage performance assessment model is established using a hierarchical machine learning architecture and the system state vectors of all samples. For example, a dynamic Bayesian network damage propagation model is used to decompose the system state vector into multiple damage levels, generating a conditional damage probability matrix that defines the transition probabilities of local structural damage states at various locations on the bridge. A discrete-time Markov chain is used to model the bridge damage propagation path and calculate the joint probability. A multi-scale CNN-LSTM hybrid model then fuses the joint probabilities at multiple scales to obtain a damage probability distribution. Simultaneously, uncertainty quantification and a dynamic update mechanism drive model updates, ultimately leading to the establishment of a comprehensive seismic damage performance assessment model.

[0069] S4: Use the earthquake damage performance assessment model to evaluate the multimodal feature space and generate the damage probability distribution of high-speed railway bridges;

[0070] Using the established earthquake damage performance assessment model, the multimodal feature space is analyzed to evaluate the earthquake damage performance of high-speed railway bridges, ultimately generating a damage probability distribution. This damage probability distribution intuitively demonstrates the likelihood of different levels of damage occurring. For example, it can display the probability of bridges experiencing minor damage, moderate damage, severe damage, and complete damage. This provides decision makers with quantitative assessment results, allowing them to quickly understand the damage risk of bridges after an earthquake. This allows them to develop targeted maintenance, repair, or emergency response plans to ensure the safe operation of high-speed railway bridges and reduce losses caused by earthquakes.

[0071] Example 2:

[0072] Based on Example 1, a rapid earthquake damage performance assessment method for high-speed railway bridges based on machine learning is provided. S1: Real-time multi-source data of high-speed railway bridge structures is collected through numerous distributed sensor arrays deployed in a heterogeneous data acquisition system, and a historical earthquake database and a BIM model of the bridge structure are obtained, including:

[0073] Through the numerous distributed sensor arrays deployed in the heterogeneous data acquisition system, the track plate deformation data, seismic wave response data, overall bridge displacement, and speckle images of each local structure under different intensities of earthquake action of the high-speed railway bridge structure are collected in real time as multi-source data of the high-speed railway bridge structure;

[0074] Obtain a historical earthquake database and a BIM model of a bridge structure, where the BIM model of the bridge structure includes bridge set parameters, material properties, and connection methods.

[0075] In this example:

[0076] Multi-source data acquisition: Utilize numerous distributed sensor arrays in a heterogeneous data acquisition system to collect specific types of data on high-speed railway bridge structures in real time.

[0077] Track slab deformation data: As the part of a high-speed railway bridge that directly bears train loads, track slab deformation is crucial to the overall performance of the bridge and the safety of train operations. Real-time monitoring of track slab deformation data using sensors can promptly detect earthquake impacts on the track slab, such as warping and twisting, which can further affect the smoothness and safety of train travel.

[0078] Seismic wave response data: Seismic waves are the fluctuations that transmit energy during earthquakes. Bridges react in various ways to earthquakes. Collecting seismic wave response data can reveal a bridge's response to different earthquake wave characteristics, such as vibration amplitude and frequency. This information helps analyze the intensity and mode of earthquake impact on bridge structures and is a key component of evaluating bridge damage performance.

[0079] Bridge overall displacement: Monitoring bridge overall displacement data provides a direct indicator of the bridge's spatial position changes during an earthquake. Excessive overall displacement may indicate a threat to the stability of the bridge structure. By acquiring this data in real time, the overall condition of the bridge during an earthquake can be quickly determined, providing important evidence for subsequent assessments.

[0080] Bridge pier strain refers to the relative deformation of a pier when subjected to external forces (such as earthquakes and vehicle loads). In practice, when external forces act on a pier, such as the impact of seismic waves during an earthquake or the pressure exerted by a passing train, the pier material undergoes changes in shape or size, and the extent of this change is measured by strain. Strain, typically expressed as a dimensionless ratio, reflects the degree of elongation or contraction of the pier material in the direction of the force. Information about pier strain is crucial for assessing earthquake damage to bridges. By monitoring pier strain, we can understand the stress state of the pier and determine whether it exceeds the material's tolerance. This allows us to infer the presence and extent of damage, providing crucial information for timely maintenance or repair measures. For example, if strain in a specific area of ​​a pier suddenly increases beyond the normal range, it may indicate damage such as cracks, requiring further inspection and treatment.

[0081] Speckle images of each local structure under earthquakes of varying intensities: Speckle imaging technology is commonly used to measure minute surface deformations. Speckle images of each local structure of a bridge under earthquakes of varying intensities are collected and analyzed to reveal subtle changes in the local structure, such as strain and displacement. Seismic exposures of varying intensities simulate the various conditions a bridge might experience in actual operation. Combined with speckle image analysis, this provides a more comprehensive understanding of the damage to the local structure under the combined effects of earthquakes and varying loads.

[0082] Other data acquisition:

[0083] A historical earthquake database contains records of past earthquake events and corresponding bridge responses. This historical data serves as a reference for analyzing the similarities between the current earthquake and historical earthquakes, as well as the damage patterns and extent of bridges in similar earthquakes, providing empirical support for this earthquake damage performance assessment.

[0084] BIM model of the bridge structure: This model includes the bridge's geometric parameters (such as the length and height of the bridge and track structures), material properties (such as the mechanical properties of concrete and steel), and connection methods (such as the connection between the piers and the bridge deck, and the characteristics of the connection nodes between the various layers of the track structure). This detailed information accurately describes the bridge structure at a physical level and is crucial for understanding the mechanical response and potential damage mechanisms of the bridge under earthquakes. It also provides the foundational structural information for the subsequent construction of a multimodal feature space and the development of a seismic damage performance assessment model.

[0085] Example 3:

[0086] Based on Example 1, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning is proposed. S2: Multi-source data is combined with a historical earthquake database and a BIM model of the bridge structure to construct a multimodal feature space containing partial features of time-frequency domain features, spatial features, and probabilistic features, including:

[0087] Based on the spatiotemporal alignment algorithm, the multi-source data is spatiotemporally aligned with the BIM model of the bridge structure to obtain spatiotemporally aligned source data;

[0088] Extracting time-frequency domain features and spatial features from the spatiotemporal alignment source data;

[0089] Based on the historical earthquake database and the introduction of Monte Carlo simulation, parameter perturbation samples are generated, and then the sensor measurement error and model parameter uncertainty are fused based on evidence theory to obtain probabilistic characteristics;

[0090] A multimodal feature space is constructed based on partial features of time-frequency domain features, spatial features, and probabilistic features.

[0091] In this example:

[0092] Spatiotemporal alignment: Using a spatiotemporal alignment algorithm, the multi-source data collected by the distributed sensor array is spatiotemporally aligned with the BIM model of the bridge structure. Because the multi-source data comes from different types of sensors, the time and spatial dimensions of their collection may differ, while the BIM model is a digital representation of the bridge structure from a three-dimensional spatial and temporal perspective. Through spatiotemporal alignment, the multi-source data and the BIM model can correspond to each other in time and space, forming a unified spatiotemporally aligned source data. For example, the vibration data collected at a certain moment in time for a certain part of the bridge can be accurately matched to the corresponding spatial position and time point in the BIM model. This ensures the consistency and relevance of the subsequently extracted features, laying the foundation for accurate analysis of the bridge's status.

[0093] Extracting time-frequency domain and spatial features includes:

[0094] Time-frequency domain feature extraction: Time-frequency domain features reflect the changing characteristics of data along the time and frequency dimensions. Extracting time-frequency domain features from spatiotemporally aligned source data helps capture the dynamic changes in bridge structural response over time and frequency. For example, during an earthquake, the vibration response of a bridge exhibits specific patterns at different times and frequencies. By analyzing these patterns, the damage to the bridge structure can be inferred.

[0095] Spatial feature extraction: Spatial features primarily reflect the spatial characteristics of bridge structures, including information such as the bridge's geometry, structural stiffness distribution, and local structural variations. These features are crucial for understanding the spatial mechanical response and damage propagation of bridges. For example, by constructing a bridge structural stiffness matrix, the ability of each bridge component to resist deformation can be quantified, allowing analysis of the stress conditions and potential damage areas at different locations during earthquakes.

[0096] Probabilistic feature extraction: Monte Carlo simulation to generate parameter perturbation samples: Based on a historical earthquake database, the Monte Carlo simulation method is introduced. Monte Carlo simulation is a technique that simulates uncertainty through random sampling. In this case, various parameters in historical earthquake data (such as earthquake intensity, duration, and spectral characteristics) are used to generate parameter perturbation samples through multiple random sampling to simulate the uncertainty in the earthquake process.

[0097] Uncertainty fusion based on evidence theory: Sensor measurement errors and model parameter uncertainties are combined using evidence theory to generate a probabilistic signature. Sensors inevitably have measurement errors when collecting data, and when building bridge models, model parameters also have a certain degree of uncertainty. Evidence theory can effectively process this uncertainty information, fusing it to produce a probabilistic signature that reflects the probability of bridge damage under earthquake conditions. This probabilistic signature quantifies the likelihood of different damage states occurring in a bridge under earthquake conditions, providing probabilistic information for earthquake damage performance assessment.

[0098] Constructing a multimodal feature space: The extracted time-frequency, spatial, and probabilistic features are integrated to construct a multimodal feature space. This multimodal feature space comprehensively describes the characteristics of bridges under earthquakes from multiple dimensions, fully exploring the potential damage-related information hidden in the data. This provides rich, comprehensive, and representative feature data for subsequent machine learning-based earthquake damage performance assessment models, helping to improve the accuracy and reliability of model assessments.

[0099] Example 4:

[0100] Based on Example 3, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning extracts time-frequency domain features and spatial features from the spatiotemporal aligned source data, including:

[0101] The energy distribution of the preset frequency band is extracted from the multi-source data aligned in the time-space aligned source data by short-time Fourier transform, and then decomposed by Hilbert-Huang transform to obtain multiple intrinsic mode functions as time-frequency domain features;

[0102] Constructing the bridge structure stiffness matrix based on the BIM model of the aligned bridge structure in the spatiotemporal alignment source data;

[0103] Extracting damage-sensitive features from spatiotemporally aligned source data;

[0104] Among them, the spatial characteristics include the bridge structure stiffness matrix and damage sensitivity characteristics.

[0105] In this example:

[0106] Short-time Fourier transform extracts the energy distribution of preset frequency bands: Short-time Fourier transform (STFT) can convert time domain signals into the time-frequency domain, thereby revealing the frequency composition of the signal at different time points. In this embodiment, STFT is applied to the aligned multi-source data in the spatiotemporal aligned source data, focusing on the energy distribution within the preset frequency band. Because the energy distribution in different frequency bands is often related to the specific response or damage state of the bridge structure. For example, changes in energy in certain frequency bands (such as 0-50 Hz) may reflect the intensification of vibration in a certain part of the bridge structure, which may be a signal of potential damage. By analyzing the energy distribution of these preset frequency bands, the characteristic information of the bridge structure response in the time and frequency dimensions can be preliminarily obtained.

[0107] Hilbert-Huang transform is used to obtain the intrinsic mode function: After obtaining the energy distribution of the preset frequency band, the Hilbert-Huang transform (HHT) is further used. HHT is a method suitable for nonlinear and non-stationary signal analysis. It can decompose complex signals into multiple intrinsic mode functions (IMFs). These IMFs represent oscillation modes at different time scales in the signal. Each IMF has a certain physical meaning and reflects the different characteristic components of the original signal. The multiple IMFs obtained by HHT decomposition as time-frequency domain features can more finely characterize the time-frequency characteristics of the bridge structure response, which helps to discover the complex vibration modes and potential damage characteristics of the bridge structure under earthquake action.

[0108] Constructing the bridge structure stiffness matrix based on the BIM model: Using the BIM model of the bridge structure aligned in the spatiotemporal alignment source data, the bridge structure stiffness matrix is ​​constructed. The bridge structure stiffness matrix is ​​an important parameter matrix that describes the bridge structure's ability to resist deformation. It reflects the relationship between the displacement of each node of the bridge and the load it bears. The bridge structure stiffness matrix can be accurately calculated through detailed information such as the bridge's geometric parameters, material properties, and connection methods provided by the BIM model. This matrix is ​​crucial for analyzing the mechanical response of a bridge under earthquakes. For example, the deformation of each part of the bridge under a given load (such as earthquake force) can be calculated through the stiffness matrix, thereby evaluating the stability of the bridge structure and potential damage areas. The formula is:

[0109]

[0110] Where K global is the bridge structure stiffness matrix, K j is the stiffness matrix of the j-th beam element, T j is the coordinate transformation matrix, is the transposed matrix of the coordinate transformation matrix, and m is the total number of beam elements in the bridge structure.

[0111] Extracting damage-sensitive features: Damage-sensitive features refer to characteristic parameters that are more sensitive to damage to the bridge structure. Changes in these features can effectively detect whether the bridge structure is damaged and the extent of the damage. Damage-sensitive features are extracted from spatiotemporally aligned source data, for example, by analyzing speckle images, vibration responses and other data. These damage-sensitive features together with the bridge structure stiffness matrix constitute spatial features, which can comprehensively reflect the state of the bridge structure from a spatial perspective and provide key information for the assessment of bridge seismic damage performance. For example, changes in certain damage-sensitive features may directly indicate damage to the local structure of the bridge. Combined with the analysis of the bridge structure stiffness matrix, the damage location can be more accurately located and the impact of the damage on the overall structure can be assessed.

[0112] Example 5:

[0113] Based on Example 4, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning extracts damage-sensitive features from spatiotemporally aligned source data, including:

[0114] Based on the phase correlation method, the speckle images of the local structure in the aligned multi-source data in the spatiotemporal aligned source data under different intensity earthquakes are pre-registered at the pixel level, and the full-field displacement vector field is constructed based on all the speckle images after pixel level registration;

[0115] A two-dimensional Delaunay triangulated mesh is constructed based on the full-field displacement vector field, and the Green strain and tension are calculated by counting the displacement gradients of the mesh nodes to generate the strain cloud map of the corresponding level;

[0116] Calculate the warping deformation of the track slab, the damage to the track interlayer connectors, the damage to the supports, and the damage to the bridge piers based on the various strain components in the strain cloud diagram;

[0117] Modal analysis was performed on the time-space aligned source data to extract the first ten frequency change rates of the bridge piers;

[0118] The warping deformation of the track slab and the first ten-order frequency change rate of the bridge pier are regarded as damage-sensitive features.

[0119] In this example:

[0120] Speckle image registration and full-field displacement vector field construction based on phase correlation method: First, the phase correlation method is used to perform preset pixel-level registration on the speckle images under different intensities of earthquakes in the spatiotemporal alignment source data. The speckle image records the random speckle distribution on the surface of the local structure of the bridge. Under different loads, these speckles will be displaced due to structural deformation. The phase correlation method can accurately calculate the relative displacement between adjacent speckle images by analyzing the phase information of the image, so as to accurately register them at the pixel level. For example, assuming that in the speckle image of a local structure, a certain speckle moves slightly before and after the load changes, the phase correlation method can accurately measure the distance and direction of this movement;

[0121] A full-field displacement vector field is constructed based on all speckle images after pixel-level registration. This field reflects the displacement of the entire local structural surface under earthquakes of varying intensities. Each point corresponds to a displacement vector containing both magnitude and direction. This full-field displacement vector field provides an intuitive understanding of the deformation trends of the local structure under load, providing foundational data for subsequent analysis.

[0122] 2D Delaunay triangular mesh construction and strain cloud map generation: 2D Delaunay triangular mesh is constructed based on the full-field displacement vector field. Delaunay triangular mesh is a method for triangulating discrete points on a plane. It can connect discrete points in the full-field displacement vector field into triangular meshes, and these triangles have the property of empty circumcircles, which can better reflect the spatial relationship between data points. For example, the key displacement points in the full-field displacement vector field are used as mesh nodes to construct the Delaunay triangular mesh;

[0123] The Green's strain is calculated by statistically analyzing the displacement gradients of the grid nodes, and then a strain cloud map of the corresponding level is generated. The displacement gradient reflects the rate of change of displacement in space. By calculating the displacement gradient of the grid nodes, the strain information within each triangular area can be obtained. Green's strain is a physical quantity that describes the deformation of a material. Calculating Green's strain based on the displacement gradient can more accurately reflect the strain state of the local structure. Based on these strain values, a strain cloud map is generated. The strain cloud map uses different colors to intuitively display the strain size of different areas of the local structure. Darker or lighter colors represent larger or smaller strains, which facilitates intuitive analysis of the strain distribution of the local structure.

[0124] Track slab warping calculation: Based on the various strain components in the strain cloud, the track slab warping, damage to track interlayer connectors, support damage, and pier damage are calculated. Track slabs are a crucial component of high-speed railway bridges, and their warping has a significant impact on train operation safety. First, a strain component-based warping calculation model is established. This model considers the combined effects of strain components in different directions in the strain cloud on track slab warping.

[0125] De-noising and outlier processing are performed on the strain contour to obtain a pre-processed strain contour. Since the actual collected data may be affected by noise or contain outliers, these noise and outliers may affect the accuracy of the track plate warping deformation calculation. De-noising and outlier processing can improve data quality and remove irrelevant or erroneous data points. For example, filtering algorithms can be used to remove high-frequency noise, and outliers can be removed by setting a reasonable data range.

[0126] The pre-processed strain nephogram is input into the warping deformation calculation model to obtain the track slab warping deformation. This pre-processed strain nephogram is also input into the component damage calculation model to determine the damage amounts of the track interlayer connectors, bearings, and piers. The processed strain nephogram data is more accurate and reliable. Inputting it into the calculation model yields more precise results for track slab warping deformation, damage amounts of track interlayer connectors, bearings, and piers. These results can serve as important indicators for assessing local bridge structural damage.

[0127] Extraction of the first ten frequency change rates of bridge piers: Extract the vibration response signals of bridge piers from the spatiotemporally aligned source data, and separate the vibration response signals of some bridge piers within the frequency band of interest from the vibration response signals through band-pass filtering. As the supporting structure of a bridge, the vibration characteristics of bridge piers are closely related to the overall stability of the bridge. Band-pass filtering can select signals in a specific frequency range, remove interference signals of other frequencies, and retain only signals within the frequency band of interest, because the signals in this frequency band may be related to the specific vibration mode or damage state of the bridge pier. For example, vibrations in certain frequency bands may be related to the first-order, second-order, or third-order vibration modes of the bridge pier;

[0128] Frequency domain decomposition was used to perform spectrum analysis on the vibration response signals of some bridge piers within the frequency band of interest, identifying the top ten natural frequencies. Frequency domain decomposition is a method that converts time domain signals into the frequency domain for analysis. This method can obtain the energy distribution of the signal at different frequencies, thereby identifying the top ten natural frequencies of the bridge piers. Natural frequencies are inherent characteristics of structures and tend to change when bridge piers are damaged.

[0129] The rate of change of the first ten natural frequencies of the pier is calculated based on the first ten natural frequencies and the first ten natural frequencies of the pier in the baseline state. The frequency rate of change is determined by comparing the first ten natural frequencies measured with the natural frequencies in the baseline state (e.g., when the bridge is unaffected by an earthquake or in normal operation). The frequency rate of change can intuitively reflect changes in the pier's structural state. A large rate of change may indicate damage to the pier structure, making it an important damage-sensitive characteristic.

[0130] Damage-sensitive features were identified: Track slab warping and the first ten-order frequency change rates of bridge piers were used as damage-sensitive features. These two features reflect the damage of the bridge structure from different perspectives: track slab warping reflects local damage to the bridge superstructure, while the frequency change rates of bridge piers reflect damage to the bridge support structure. Together, they provide key characteristic data for rapid, machine learning-based assessment of high-speed railway bridge seismic damage performance, helping to more accurately assess the extent and location of bridge damage during earthquakes.

[0131] Example 6:

[0132] Based on Example 5, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning calculates the warping deformation of the track slab, the damage amount of the track interlayer connector, the damage amount of the support, and the damage amount of the bridge pier based on multiple strain components in the strain cloud map, including:

[0133] Establish a calculation model for warping deformation based on strain components and a calculation model for damage to each component;

[0134] De-noising and outlier processing are performed on the strain cloud map to obtain the pre-processed strain cloud map;

[0135] The preprocessed strain cloud map is input into the warping deformation calculation model to obtain the warping deformation of the track plate, and the preprocessed strain cloud map is input into the damage calculation model of each component to obtain the damage amount of the track interlayer connector, the support damage amount and the bridge pier damage amount.

[0136] In this example:

[0137] Establish a calculation model for warping deformation based on strain components: Under the action of external forces such as earthquakes, the internal strain state of the track plate is complex and diverse, and strain components of different directions and types jointly affect the warping deformation of the track plate. Therefore, it is necessary to establish a special calculation model to comprehensively consider the relationship between these strain components and warping deformation. This model may be based on theories such as material mechanics and structural mechanics. Through mathematical formulas and algorithms, various strain components in the strain cloud map (such as normal strain, shear strain, etc.) are used as input parameters to calculate the warping deformation of the track plate. For example, according to the principles of elastic mechanics, combined with the geometric shape of the track plate, material properties and the interaction between strain components, a corresponding mathematical model is constructed to accurately describe the conversion process from strain to warping deformation.

[0138] Establishing models for calculating damage to various components: Similarly, based on the strain components in the strain nephogram, models are developed to calculate damage to track interlayer connectors, supports, and bridge piers. These models account for the damage mechanisms of each component under different strain states. For example, track interlayer connectors may experience damage such as loosening or fracture under specific strain combinations. The model calculates damage by analyzing strain information in the areas related to the connectors in the strain nephogram, combining the mechanical properties and structural characteristics of the connectors. For supports and bridge piers, similar principles are used to calculate damage using strain components based on their respective structural functions and stress characteristics.

[0139] Denoising and outlier processing of strain cloud images include:

[0140] Denoising: Various noises are inevitably introduced during the actual collection and generation of strain contour maps. These noises may originate from sensor measurement errors, environmental interference, or errors in image processing. Noise can interfere with the accurate acquisition of true strain information and affect the accuracy of track slab warping deformation calculations. Therefore, appropriate denoising methods, such as filtering techniques (e.g., Gaussian filtering and median filtering), are necessary to remove noise signals from the strain contour map, resulting in a smoother image and highlighting the true strain trend.

[0141] Outlier processing: In addition to noise, strain contours may also contain outliers, which may be caused by sensor failure, data transmission errors, or local sudden interference. Outliers can seriously affect data analysis results. If left unprocessed, they may cause significant deviations in the calculated track plate warping deformation. By setting a reasonable threshold range or using statistical methods (such as those based on mean and standard deviation), these outliers can be identified, corrected, or removed, resulting in a preprocessed strain contour, providing reliable data for subsequent accurate calculation of track plate warping deformation.

[0142] The preprocessed strain contour map is input into the warping deformation calculation model. After denoising and outlier processing, the strain contour map significantly improves data quality and more accurately reflects the actual strain state of the track slab. This strain contour map is then fed into the previously established warping deformation calculation model. The model analyzes and calculates each strain component based on a predefined algorithm and formula, ultimately outputting the track slab's warping deformation. This result accurately reflects the extent of the track slab's warping under its current strain state, providing an important basis for assessing damage to bridge track slabs under earthquakes. For example, if the calculated track slab warping deformation exceeds a certain safety threshold, it indicates that the track slab may have sustained damage that could affect safe train operation.

[0143] The preprocessed strain cloud maps are simultaneously fed into the damage calculation models for each component. These models, based on their respective algorithms, combine the strain information from the strain cloud maps with the structural and mechanical properties of each component to calculate the damage to the track interlayer connectors, bearings, and piers. This damage data helps assessors comprehensively understand the extent of damage to key bridge components, providing a quantitative basis for developing targeted maintenance and repair strategies. For example, by comparing pier damage calculated over different periods, it is possible to determine the development trend of pier damage and take timely measures to ensure the safe operation of the bridge.

[0144] Example 7:

[0145] Based on Example 5, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning is developed. Modal analysis is performed on the spatiotemporally aligned source data to extract the first ten frequency change rates of the bridge piers, including:

[0146] The bridge pier vibration response signal is extracted from the spatiotemporally aligned source data, and the partial bridge pier vibration response signal within the frequency band of interest is separated from the prying response signal by bandpass filtering.

[0147] The frequency domain decomposition method is used to perform spectrum analysis on the vibration response signals of some bridge piers within the frequency band of interest to identify the first ten natural frequencies;

[0148] Based on the first ten natural frequencies and the first ten natural frequencies of the pier in the reference state, the change rate of the first ten frequencies of the pier is calculated.

[0149] In this embodiment:

[0150] Extracting and filtering bridge pier vibration response signals: First, bridge pier vibration response signals are extracted from the spatiotemporally aligned source data. These signals contain information about the vibrations generated by the piers under various external forces (such as earthquakes and train movements). However, the original vibration response signals often contain components from multiple frequency bands, some of which may be irrelevant to the key vibration characteristics or damage conditions of the piers and may even interfere with the extraction of valid information. Therefore, bandpass filtering is used to isolate the portion of the pier vibration response signal within the frequency band of interest from the original pier vibration response signals. A bandpass filter can be set to a specific frequency range, allowing only signals within that range to pass while blocking signals at other frequencies. For example, if certain natural vibration frequencies of a bridge pier are known to be concentrated in a specific frequency band, a bandpass filter can be set to retain only the vibration response signals within this frequency band, thereby highlighting information related to the key vibration characteristics of the piers and laying the foundation for subsequent accurate analysis of the pier condition.

[0151] Spectral analysis identifies the top ten natural frequencies: Frequency domain decomposition is used to perform spectral analysis on the vibration response signals of a portion of the bridge pier within the frequency band of interest after bandpass filtering. Frequency domain decomposition is an effective method for converting time-domain signals to the frequency domain for analysis, revealing the signal's energy distribution at different frequencies. This analysis method can identify the top ten natural frequencies of the bridge pier from the complex vibration response signal. Natural frequency is an inherent property of the bridge pier structure, dependent on factors such as its geometry, material properties, and boundary conditions. Under normal circumstances, a bridge pier has a specific natural frequency value. However, when a bridge pier structure is damaged, its natural frequency tends to change, making accurate identification of the natural frequency crucial for assessing the damage condition of the pier.

[0152] Calculate the first ten frequency change rates: After identifying the first ten natural frequencies of the bridge pier in the current state (i.e., f 当前 ) and then compare it with the first ten natural frequencies of the pier in the reference state (i.e., f 基准 ) for comparison:

[0153]

[0154] The reference state usually refers to the situation when the bridge pier is undamaged or in normal working condition, and the natural frequency at this time serves as a reference standard. By calculating the difference between the current first ten natural frequencies and the corresponding frequencies in the reference state, and normalizing the difference with the reference frequency, the first ten frequency change rates of the bridge pier are obtained. This frequency change rate can intuitively reflect the degree of change in the structural state of the bridge pier. For example, if the change rate of a certain frequency is large, it may mean that the structural part of the bridge pier related to the vibration mode of this order has been damaged or changed, providing a quantitative key indicator for evaluating the earthquake damage of the bridge pier, helping to timely discover potential structural problems and providing an important basis for the safety assessment of high-speed railway bridges.

[0155] Example 8:

[0156] On the basis of Example 3, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning is proposed. A multimodal feature space is constructed based on some features of time-frequency domain features, spatial features, and probability features. Figure 2 ,include:

[0157] Visualize the feature contribution status of time-frequency domain features, spatial features, and probability features based on the SHAP value analysis method;

[0158] Based on the recursive feature elimination method and the feature contribution status of time-frequency domain features, spatial features, and probability features, some features whose contribution is not less than the contribution threshold in the time-frequency domain features, spatial features, and probability features are retained and feature dimensionality reduction is performed to obtain a multimodal feature space.

[0159] In this example:

[0160] Visualizing feature contributions (e.g., the influence coefficient of support stiffness on damage probability) using the SHAP value analysis method: SHAP (SHapley Additive ex Planations) value analysis is a technique used to interpret machine learning model predictions. It assesses the contribution of each feature to the model output. In constructing a multimodal feature space, this method analyzes time-frequency, spatial, and probabilistic features, providing a clear understanding of the importance of each feature in describing bridge seismic damage performance. For example, SHAP value analysis may reveal that certain time-frequency features (e.g., energy distribution in a specific frequency band) contribute significantly to predicting the probability of a particular bridge damage type, while certain spatial features (e.g., stiffness variations at specific locations) play a significant role in reflecting the overall structural stability of the bridge. By visualizing these feature contributions, presenting the magnitude and sign of each feature's SHAP value in a chart or graph, the relative importance of each feature in the model is intuitively demonstrated, providing a basis for subsequent feature selection. This helps understand the impact of different feature types on bridge seismic damage performance assessment, identifying key features and those with less significant impact.

[0161] Feature screening and dimensionality reduction based on recursive feature elimination method include:

[0162] Recursive feature elimination: This method gradually eliminates features that contribute little to model performance. Based on the feature contribution status obtained through SHAP value analysis, a contribution threshold (30 percent) is set. For time-frequency, spatial, and probabilistic features, only those features whose contribution is at least this threshold are retained. For example, if a time-frequency feature has a low SHAP value, indicating that it contributes little to the bridge damage performance assessment model, it may be eliminated at this step.

[0163] Feature dimensionality reduction: While retaining important features, this screening process actually also achieves feature dimensionality reduction. Feature dimensionality reduction is of great significance. On the one hand, too many redundant or unimportant features may increase the computational complexity and complexity of model training, and even lead to overfitting, affecting the model's generalization ability. On the other hand, retaining key features can better highlight factors that have a practical impact on bridge seismic damage performance, making the model more concise and efficient. In this way, a multimodal feature space is ultimately obtained. This space not only contains the most important feature information for bridge seismic damage performance assessment, but also avoids the interference of too many irrelevant or redundant features. This helps to improve the accuracy and efficiency of the seismic damage performance assessment model established based on this multimodal feature space.

[0164] The above methods can improve the inter-class distance of features, the cross-dataset test accuracy, and reduce the width of the 95% confidence interval.

[0165] Example 9:

[0166] Based on Example 1, a rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning is proposed. S3: A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model and a multi-scale CNN-LSTM hybrid model as well as an uncertainty quantification and dynamic update mechanism, and an earthquake damage performance assessment model is established based on the hierarchical machine learning architecture. Figure 3 ,include:

[0167] A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model, a multi-scale CNN-LSTM hybrid model, and an uncertainty quantification and dynamic update mechanism.

[0168] Construct a multimodal feature space of a large number of samples (e.g., 100,000+ samples) and generate the system state vector of each sample based on the multimodal feature space of each sample;

[0169] A earthquake damage performance assessment model is established based on a hierarchical machine learning architecture and the system state vectors of all samples.

[0170] In this example:

[0171] Building a hierarchical machine learning architecture involves:

[0172] Integrate multiple models and mechanisms: This step aims to build a comprehensive, multi-layered machine learning architecture that integrates a dynamic Bayesian network damage propagation model, a multi-scale CNN-LSTM hybrid model, and uncertainty quantification and dynamic update mechanisms.

[0173] Dynamic Bayesian Network Damage Propagation Model: Dynamic Bayesian Networks excel at processing information with time series characteristics and uncertainty, effectively analyzing the propagation of damage within a bridge structure. Based on changes in the bridge's structural state at different points in time, they can infer how damage propagates from one location to another, as well as the transition probabilities between different damage states. For example, during an earthquake, a dynamic Bayesian network can simulate the gradual propagation of damage from a pier to the bridge deck, as well as the probability of damage state changes at each stage.

[0174] Multi-scale CNN-LSTM hybrid model: Convolutional neural networks (CNNs) excel at capturing spatial features in data, while long-short-term memory (LSTM) networks excel at processing long-term dependencies in time series data. Combining the two to form a multi-scale CNN-LSTM hybrid model can extract spatiotemporal features of bridge-related data at different scales. For example, CNNs can extract local spatial features from image data of bridge structures (such as speckle images), while LSTMs can process time series data such as bridge vibration responses and learn their long-term trends. This combination of the two helps fully uncover complex patterns in bridge data.

[0175] Uncertainty Quantification and Dynamic Update Mechanism: Taking into account the uncertainty of earthquakes and the variability of bridge structural responses, an uncertainty quantification and dynamic update mechanism is introduced. This mechanism quantifies various uncertainties in the model, such as uncertainty in earthquake parameters and sensor measurement errors. Simultaneously, the model is dynamically updated based on new data and actual conditions, enabling it to better adapt to changing conditions and maintain assessment accuracy.

[0176] Generating a sample system state vector includes:

[0177] Constructing multimodal feature space samples: First, construct a multimodal feature space of a large number of samples. These samples cover the various states of bridges under different earthquake scenarios. The multimodal feature space contains time-frequency domain features, spatial features, and probabilistic features, etc., which comprehensively describe the characteristics of bridges under earthquakes from multiple dimensions. For example, different samples may correspond to the response of bridges under earthquakes of different intensities. The multimodal feature space of each sample contains corresponding time-frequency domain features (such as energy distribution in different frequency bands), spatial features (such as changes in bridge structural stiffness and damage sensitivity characteristics), and probabilistic features (probability distribution after considering earthquake uncertainty and measurement errors).

[0178] Generate a system state vector: This vector is generated based on the multimodal feature space of each sample. The system state vector integrates and encodes various features in the multimodal feature space to form a vector that represents the state of the sample bridge. This vector contains key information about the bridge under a specific earthquake scenario and provides a unified data format for subsequent model learning and evaluation. For example, time-frequency domain features, spatial features, and probabilistic features are arranged in a specific order or fused into a single vector using an algorithm to serve as the system state vector for the sample bridge.

[0179] Establishing a seismic damage performance assessment model includes:

[0180] Model training foundation: We leverage a pre-built hierarchical machine learning architecture and the system state vectors of all samples to build a damage performance assessment model. The hierarchical machine learning architecture provides a powerful learning framework, while the large number of sample system state vectors provides rich data for model training.

[0181] Learning and Modeling: The model learns from these sample system state vectors, discovers the patterns and rules contained therein, and establishes a mapping relationship between input features (system state vectors) and output results (bridge earthquake damage performance). For example, by continuously adjusting the model parameters, the model can accurately predict the damage probability distribution, damage level, and other earthquake damage performance indicators of bridges under corresponding earthquake scenarios based on the input system state vectors, thereby effectively evaluating the earthquake damage performance of high-speed railway bridges.

[0182] Example 10:

[0183] Based on Example 9, a rapid earthquake damage performance assessment method for high-speed railway bridges based on machine learning establishes an earthquake damage performance assessment model based on a hierarchical machine learning architecture and the system state vectors of all samples, including:

[0184] Based on the dynamic Bayesian network damage propagation model in the hierarchical machine learning architecture, the system state vectors of all samples are decomposed into multiple damage levels, and the conditional damage probability matrix of all samples is generated based on the system state vectors of all samples and the corresponding damage levels;

[0185] Based on the conditional damage probability matrix of each sample, the transition probability of the local structural damage state of each bridge in all samples is defined;

[0186] The bridge damage propagation path is modeled based on a discrete-time Markov chain, and the joint probability of the bridge damage propagation path for each sample is calculated based on the transition probability of the local structural damage state of each bridge in each sample;

[0187] Based on the multi-scale CNN-LSTM hybrid model in the hierarchical machine learning architecture, the joint probability of the bridge damage propagation path of each sample is fused at multiple scales to obtain the damage probability distribution of each bridge sample.

[0188] At the same time, uncertainty quantification results are obtained based on the uncertainty quantification and dynamic update mechanism, and the model update frequency is driven based on the uncertainty quantification results until an earthquake damage performance evaluation model is established.

[0189] In this example:

[0190] Damage level decomposition: Utilizing the dynamic Bayesian network damage propagation model within a hierarchical machine learning architecture, the system state vectors of all samples are analyzed and decomposed into multiple damage levels. The system state vector contains comprehensive information extracted from the multimodal feature space of the bridge under a specific earthquake scenario. Based on this information, the dynamic Bayesian network determines the extent of damage to the bridge and categorizes it into different levels, such as minor damage, moderate damage, severe damage, and complete damage. This step lays the foundation for the subsequent quantification of damage probability. For example, by analyzing the strain, displacement, and frequency changes of the bridge structure in the system state vector, the damage level of the bridge can be determined.

[0191] Conditional damage probability matrix generation: Based on the system state vectors and corresponding damage levels of all samples, a conditional damage probability matrix is ​​generated. This matrix reflects the probability of the bridge being at different damage levels under the given system state vector. For example, for a specific system state vector, the conditional damage probability matrix will give the probability of the bridge being at a slight damage level, the probability of being at a moderate damage level, and so on. This matrix quantifies the probabilistic relationship between the system state and the damage level, which helps to further analyze the possibility of bridge damage. The formula is:

[0192]

[0193] Where, P(S t+1 |S t ,D t ) is the conditional damage probability matrix, S t is the system state vector at time t (including the damage states of n components), S t+1 is the system state vector at time t+1 (including the damage states of n components), D t is the set of parent node damage states at time t, D i,t+1 is the damage level of component i at time t, P(D i,t+1 |D pa(i),t+1 ) is the conditional probability table based on the damage of the parent node;

[0194] Define the transition probability of local structural damage states: Based on the conditional damage probability matrix for each sample, define the transition probability of the local structural damage state of each bridge in all samples. Bridges are composed of multiple local structures, and the damage states of different local structures will change under the action of an earthquake and affect each other. The conditional damage probability matrix can be used to determine the probability of a local structure transitioning from one damage state to another. For example, the probability of transitioning from an undamaged state to a slightly damaged state on a bridge pier, or the probability of transitioning from slightly damaged state to moderate damage on a bridge deck, etc. These transition probabilities describe the possibility of damage propagation in the local structure of the bridge and provide key parameters for subsequent modeling of damage propagation paths.

[0195] Discrete-time Markov chain modeling: A discrete-time Markov chain is used to model the propagation path of bridge damage. A Markov chain is a stochastic process that assumes that the damage state of a bridge at a given moment depends solely on the damage state at the previous moment, and is independent of earlier states. This assumption, combined with the previously defined local structural damage state transition probabilities, allows for the construction of a model that models the propagation path of bridge damage between different local structures and at different points in time. For example, this can simulate how damage propagates from one part of a bridge to another over the duration of an earthquake.

[0196] Calculating joint probabilities: Based on the transition probabilities of the local structural damage states at each bridge location within each sample, the joint probability of the bridge damage propagation paths for each sample is calculated. The joint probability represents the probability that each of the local structural damage states will occur simultaneously under a specific damage propagation path. By calculating the joint probability, the likelihood of different damage propagation paths can be assessed, providing a more comprehensive understanding of the evolution of bridge damage during earthquakes. For example, the joint probability of a propagation path starting with minor damage to a bridge pier and subsequently propagating to different damage states at different locations on the bridge deck can be calculated to determine the likelihood of this path occurring in an actual earthquake.

[0197] Obtaining Damage Probability Distributions Using Multi-Scale Fusion: A multi-scale CNN-LSTM hybrid model within a hierarchical machine learning architecture is used to perform multi-scale fusion on the joint probabilities of damage propagation paths for each bridge sample. This multi-scale CNN-LSTM hybrid model enables feature extraction and analysis of data at different scales. In this step, the joint probabilities are processed at different scales, for example, from the macroscopic scale of overall bridge damage propagation to the microscopic scale of local structural damage details, fully exploiting the inherent information. Through multi-scale fusion, a damage probability distribution for each bridge sample is obtained. This distribution demonstrates the probability of a bridge experiencing different levels of damage, providing intuitive and quantitative results for assessing bridge seismic damage performance. For example, the probability of a bridge experiencing minor damage is X%, the probability of moderate damage is Y%, and the probability of severe damage is Z%.

[0198] Multi-scale fusion principle:

[0199] F=α·F low +β·F high

[0200] In the formula, F is the feature obtained after multi-scale fusion, F low is a low-resolution feature (large receptive field), F high is a high-resolution feature (rich in details), and α and β are adaptive fusion weights.

[0201] Updating the model based on uncertainty quantification includes:

[0202] Uncertainty Quantification: Utilizing uncertainty quantification and a dynamic update mechanism, we obtain uncertainty quantification results. There are many uncertainties in earthquakes and bridge structural responses, such as the randomness of the earthquake itself, sensor measurement errors, and model parameter uncertainties. The uncertainty quantification mechanism analyzes and quantifies these factors, producing results that reflect the degree of model uncertainty. For example, uncertainty analysis of earthquake parameters can provide an error range for model predictions.

[0203] Drive model updates: The frequency of model updates is driven by uncertainty quantification results. If the uncertainty quantification results indicate a high degree of model uncertainty, such as a large error range in the prediction results, the model may need to be updated more frequently to adapt to changes in the actual situation. By continuously adjusting model parameters and optimizing the model structure based on new data and uncertainty quantification results, a stable and accurate earthquake damage performance assessment model is established. This process ensures that the model can always maintain a high level of assessment accuracy in the face of complex and changing actual conditions, providing a reliable tool for earthquake damage performance assessment of high-speed railway bridges. The formula is:

[0204] τ=γ·max(U1,U2,U3)

[0205] Where τ is the model update frequency, γ is the threshold coefficient, and U1, U2, and U3 are the uncertainties of each module respectively.

[0206] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is intended to include these modifications and variations.

Claims

1. A rapid assessment method for earthquake damage performance of high-speed railway bridges based on machine learning, characterized by: include: S1: Using numerous distributed sensor arrays deployed in a heterogeneous data acquisition system, multi-source data on high-speed railway bridge structures is collected in real time, along with historical earthquake databases and BIM models of bridge structures. S2: Combine multi-source data with the historical earthquake database and the BIM model of the bridge structure to construct a multimodal feature space containing some features of time-frequency domain features, spatial features, and probabilistic features, including: Based on the spatiotemporal alignment algorithm, the multi-source data is spatiotemporally aligned with the BIM model of the bridge structure to obtain spatiotemporally aligned source data; The energy distribution of the preset frequency band is extracted from the multi-source data aligned in the time-space aligned source data by short-time Fourier transform, and then decomposed by Hilbert-Huang transform to obtain multiple intrinsic mode functions as time-frequency domain features; Constructing the bridge structure stiffness matrix based on the BIM model of the aligned bridge structure in the spatiotemporal alignment source data; Based on the phase correlation method, the speckle images of the local structure in the aligned multi-source data in the spatiotemporal aligned source data under different intensity earthquakes are pre-registered at the pixel level, and the full-field displacement vector field is constructed based on all the speckle images after pixel level registration; A two-dimensional Delaunay triangulated mesh is constructed based on the full-field displacement vector field, and the Green strain and tension are calculated by counting the displacement gradients of the mesh nodes to generate the corresponding level of strain cloud map; Calculate the warping deformation of the track slab, the damage to the track interlayer connectors, the damage to the supports, and the damage to the bridge piers based on the various strain components in the strain cloud diagram; Modal analysis was performed on the time-space aligned source data to extract the first ten frequency change rates of the bridge piers; The warping deformation of the track slab and the first ten-order frequency change rate of the bridge pier are regarded as damage-sensitive features; Among them, the spatial characteristics include the bridge structure stiffness matrix and damage sensitivity characteristics; Based on the historical earthquake database and the introduction of Monte Carlo simulation, parameter perturbation samples are generated, and then the sensor measurement error and model parameter uncertainty are fused based on evidence theory to obtain probabilistic characteristics; Construct a multimodal feature space based on partial features of time-frequency domain features, spatial features, and probabilistic features; S3: A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model and a multi-scale CNN-LSTM hybrid model, along with uncertainty quantification and dynamic update mechanisms. A seismic damage performance assessment model is then established based on this hierarchical machine learning architecture. S4: Use the earthquake damage performance assessment model to evaluate the earthquake damage performance of the multimodal feature space and generate the damage probability distribution of the high-speed railway bridge.

2. The method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning according to claim 1 is characterized in that: S1: Using numerous distributed sensor arrays deployed in a heterogeneous data acquisition system, we collect multi-source data on high-speed railway bridge structures in real time, and acquire a historical earthquake database and a BIM model of the bridge structure, including: Through the numerous distributed sensor arrays deployed in the heterogeneous data acquisition system, the track plate deformation data, seismic wave response data, overall bridge displacement, and speckle images of each local structure under different intensities of earthquake action of the high-speed railway bridge structure are collected in real time as multi-source data of the high-speed railway bridge structure; Obtain a historical earthquake database and a BIM model of a bridge structure, where the BIM model of the bridge structure includes bridge set parameters, material properties, and connection methods.

3. The method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning according to claim 1 is characterized in that: Based on the various strain components in the strain cloud diagram, the warping deformation of the track slab, the damage to the track interlayer connectors, the damage to the supports, and the damage to the bridge piers are calculated, including: Establish a calculation model for warping deformation based on strain components and a calculation model for damage to each component; De-noising and outlier processing are performed on the strain cloud map to obtain the pre-processed strain cloud map; The preprocessed strain cloud map is input into the warping deformation calculation model to obtain the warping deformation of the track plate, and the preprocessed strain cloud map is input into the damage calculation model of each component to obtain the damage amount of the track interlayer connector, the support damage amount and the bridge pier damage amount.

4. The method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning according to claim 1 is characterized in that: Modal analysis is performed on the time-space aligned source data to extract the first ten frequency change rates of the bridge pier, including: Extracting the bridge pier vibration response signal from the spatiotemporally aligned source data, and separating the partial bridge pier vibration response signal within the frequency band of interest from the bridge pier vibration response signal through band-pass filtering; The frequency domain decomposition method is used to perform spectrum analysis on the vibration response signals of some bridge piers within the frequency band of interest to identify the first ten natural frequencies; Based on the first ten natural frequencies and the first ten natural frequencies of the pier in the reference state, the change rate of the first ten frequencies of the pier is calculated.

5. The method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning according to claim 1 is characterized in that: A multimodal feature space is constructed based on some features of time-frequency domain features, spatial features, and probabilistic features, including: Visualize the feature contribution status of time-frequency domain features, spatial features, and probability features based on the SHAP value analysis method; Based on the recursive feature elimination method and the feature contribution status of time-frequency domain features, spatial features, and probability features, some features whose contribution is not less than the contribution threshold in the time-frequency domain features, spatial features, and probability features are retained and feature dimensionality reduction is performed to obtain a multimodal feature space.

6. The method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning according to claim 1 is characterized in that: S3: A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model, a multi-scale CNN-LSTM hybrid model, and an uncertainty quantification and dynamic update mechanism. A seismic damage performance assessment model is then established based on this hierarchical machine learning architecture, including: A hierarchical machine learning architecture is constructed based on a dynamic Bayesian network damage propagation model, a multi-scale CNN-LSTM hybrid model, and an uncertainty quantification and dynamic update mechanism. Construct a multimodal feature space of a large number of samples, and generate a system state vector for each sample based on the multimodal feature space of each sample; A earthquake damage performance assessment model is established based on a hierarchical machine learning architecture and the system state vectors of all samples.

7. The method for rapid assessment of earthquake damage performance of high-speed railway bridges based on machine learning according to claim 6 is characterized in that: A damage performance assessment model is established based on a hierarchical machine learning architecture and the system state vectors of all samples, including: Based on the dynamic Bayesian network damage propagation model in the hierarchical machine learning architecture, the system state vectors of all samples are decomposed into multiple damage levels, and the conditional damage probability matrix of all samples is generated based on the system state vectors of all samples and the corresponding damage levels; Based on the conditional damage probability matrix of each sample, the transition probability of the local structural damage state of each bridge in all samples is defined; The bridge damage propagation path is modeled based on a discrete-time Markov chain, and the joint probability of the bridge damage propagation path of each sample is calculated based on the transition probability of the local structural damage state of each bridge in each sample; Based on the multi-scale CNN-LSTM hybrid model in the hierarchical machine learning architecture, the joint probability of the bridge damage propagation path of each sample is fused at multiple scales to obtain the damage probability distribution of each bridge sample. At the same time, uncertainty quantification results are obtained based on the uncertainty quantification and dynamic update mechanism, and the model update frequency is driven based on the uncertainty quantification results until an earthquake damage performance evaluation model is established.

Images