Bridge health monitoring and intelligent control method and system based on machine learning

Abstract

The invention provides a bridge health monitoring and intelligent control method and system based on machine learning, and the method comprises the steps: carrying out the distributed sensing data collection of a bridge structure, generating a structure monitoring data set, carrying out the spatial-temporal feature extraction, and obtaining a vibration response feature sequence and a traffic load feature sequence of the bridge structure; performing correlation analysis on the two sequences, generating a structural damage sensitive feature matrix, performing damage evolution trend prediction based on the structural damage sensitive feature matrix, generating damage development probability distribution of the bridge key component, and generating a dynamic traffic control instruction according to the damage development probability distribution. The dynamic traffic control instruction comprises a vehicle passing parameter dynamically adjusted based on a damage degree quantitative index, and sending the dynamic traffic control instruction to a traffic management system. According to the invention, the accuracy of bridge health monitoring and the dynamic adaptability of control measures can be improved.

Description

Technical Field

[0001] This invention relates to the field of bridge monitoring, specifically to a method and system for bridge health monitoring and intelligent control based on machine learning. Background Technology

[0002] As a crucial component of transportation infrastructure, the structural health of bridges directly impacts traffic safety and regional economic operation. Currently, bridge health monitoring and control involves deploying sensors at key bridge locations to collect structural response data. The structural condition is then assessed based on preset thresholds or empirical formulas. When monitoring data exceeds the threshold, an early warning is issued, and fixed traffic control measures are implemented. However, this method, based on fixed thresholds and empirical judgments, struggles to fully capture the dynamic evolution of bridge structures under the combined effects of complex traffic loads and environmental factors. Its ability to identify structural damage early and predict its development trends is limited, resulting in a lack of refined scientific basis for traffic control measures and hindering the achievement of a dynamic balance between bridge structural safety and traffic efficiency. Therefore, improving the accuracy of bridge health monitoring and the dynamic adaptability of control measures has become an urgent technical challenge. Summary of the Invention

[0003] This invention provides a method and system for bridge health monitoring and intelligent control based on machine learning.

[0004] According to one aspect of the present invention, a bridge health monitoring and intelligent control method based on machine learning is provided. The method includes: collecting distributed sensor data of the bridge structure to generate a structural monitoring data set containing multi-channel time series; extracting spatiotemporal features from the structural monitoring data set to obtain a vibration response feature sequence and a traffic load feature sequence of the bridge structure; performing correlation analysis on the vibration response feature sequence and the traffic load feature sequence to generate a structural damage sensitive feature matrix, wherein the structural damage sensitive feature matrix is ​​used to quantify the degree of abnormal fluctuation in the structural response under different load conditions; predicting the damage evolution trend based on the structural damage sensitive feature matrix to generate a damage development probability distribution of key bridge components, wherein the damage development probability distribution includes the distribution of damage locations and a quantitative index of damage degree in different monitoring periods; generating a dynamic traffic control instruction based on the damage development probability distribution, wherein the dynamic traffic control instruction includes vehicle passage parameters dynamically adjusted based on the quantitative index of damage degree, and sending the dynamic traffic control instruction to a traffic management system.

[0005] According to another aspect of the present invention, a computer system is provided, comprising: a processor; and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the methods described above.

[0006] The present invention has at least the following beneficial effects:

[0007] This invention provides a machine learning-based bridge health monitoring and intelligent control method. It generates a structural monitoring dataset by collecting distributed sensor data from the bridge structure. Through spatiotemporal feature extraction, vibration response feature sequences and traffic load feature sequences are obtained. Then, a structural damage-sensitive feature matrix is ​​generated through correlation analysis. Based on this matrix, damage evolution trend prediction is performed to obtain the probability distribution of damage development, and finally, dynamic traffic control instructions are generated. The distributed sensor data collection generates a structural monitoring dataset containing multi-channel time series, enabling simultaneous acquisition of dynamic response information of key bridge components within a preset monitoring period. This provides high-quality input data covering the overall structural stress characteristics for subsequent feature extraction. The vibration response feature sequences and traffic load feature sequences obtained from the spatiotemporal feature extraction of the structural monitoring dataset achieve targeted separation and quantification of the structural dynamic deformation law and vehicle traffic load distribution characteristics, laying the foundation for in-depth analysis of the intrinsic correlation between structural response and external loads. The structural damage-sensitive feature matrix generated through correlation analysis of the vibration response feature sequences and traffic load feature sequences allows for... This method effectively quantifies the degree of abnormal fluctuations in structural response under different load conditions, thereby accurately locating key features reflecting changes in structural stiffness. Based on the structural damage-sensitive feature matrix, a damage development probability distribution is generated to predict damage evolution trends, enabling a probabilistic representation of the damage location distribution and damage degree quantification indicators for key bridge components at different monitoring periods. This more accurately reflects the gradual evolution of structural damage from micro to macro levels. Furthermore, traffic control instructions containing dynamically adjusted vehicle passage parameters are generated based on the damage development probability distribution, ensuring dynamic matching between traffic control measures and the actual damage state of the bridge structure. This achieves a balance between refined protection of the bridge structure and optimized traffic efficiency. Attached Figure Description

[0008] Figure 1 A schematic diagram illustrating an application scenario of a machine learning-based bridge health monitoring and intelligent control method according to an embodiment of the present invention is shown.

[0009] Figure 2 A flowchart of a bridge health monitoring and intelligent control method based on machine learning according to an embodiment of the present invention is shown.

[0010] Figure 3A schematic diagram of the composition of a computer system according to an embodiment of the present invention is shown. Detailed Implementation

[0011] Figure 1 A schematic diagram of an application scenario provided by an embodiment of the present invention is shown. This application scenario includes one or more sensing devices 101, a computer system 120, and one or more networks 110 coupling the one or more sensing devices 101 to the computer system 120. The sensing devices 101, such as strain sensors, acceleration sensors, etc., are used to collect sensing data at locations on the bridge and transmit it via the network 110 to the computer system 120 for processing, so that the computer system 120 implements the method provided by the embodiment of the present invention.

[0012] exist Figure 1 In the illustrated configuration, computer system 120 may include one or more components that implement the functions performed by computer system 120. These components may include software components, hardware components, or combinations thereof that can be executed by one or more processors. Computer system 120 may include one or more virtual machines running a virtual operating system, or other computing architectures involving virtualization (e.g., one or more flexible pools of logical storage devices that can be virtualized to maintain virtual storage devices for servers). In various embodiments, computer system 120 may run one or more services or software applications that provide the functions described below.

[0013] Furthermore, in the application scenarios of this invention, one or more databases 130 may also be included. In some embodiments, these databases may be used to store data and other information. For example, one or more of the databases 130 may be used to store a set of bridge structure monitoring data collected by sensors. The databases 130 may reside locally on the computer system 120.

[0014] Please refer to Figure 2 This is a flowchart of a bridge health monitoring and intelligent control method based on machine learning, according to an embodiment of the present invention, including the following steps:

[0015] Step S100: Perform distributed sensor data acquisition on the bridge structure to generate a set of structural monitoring data containing multi-channel time series.

[0016] The structural monitoring data set is collected by sensing devices deployed on key bridge components within a preset monitoring period.

[0017] A structural monitoring dataset is a collection of data obtained from monitoring bridge structures. This data exists in the form of multi-channel time series, where "multi-channel" indicates that data is acquired from different monitoring dimensions, and "time series" reflects the changes in data over time. Key bridge components, such as piers and beams, are crucial to the overall structural stability and safety of the bridge. Sensing devices are instruments used to collect data related to the bridge structure, such as strain sensors and accelerometers. A preset monitoring cycle is a pre-defined time period during which data is continuously collected. For example, strain sensors, accelerometers, and other sensing devices are deployed on key bridge components. Data is collected according to a preset monitoring cycle, such as once per hour, for a continuous period. The collected data from different sensing devices are then integrated to generate a structural monitoring dataset containing multi-channel time series data.

[0018] Step S200: Extract spatiotemporal features from the structural monitoring data set to obtain the vibration response feature sequence and traffic load feature sequence of the bridge structure. The vibration response feature sequence is used to characterize the dynamic deformation law of the structure under different loads, and the traffic load feature sequence is used to characterize the load distribution characteristics during vehicle passage.

[0019] Spatiotemporal feature extraction involves extracting time- and space-related feature information from structural monitoring datasets. Vibration response feature sequences are data sequences reflecting the dynamic deformation patterns of bridge structures under different loads, such as the changes in vibration frequency and amplitude under different vehicle loads. Traffic load feature sequences describe the load distribution characteristics during vehicle traffic, such as the impact of vehicle weight, speed, and travel time on bridge load distribution. Specifically, the previously collected structural monitoring datasets are analyzed and processed, and features related to vibration response and traffic load are extracted using specific algorithms and methods. For example, signal processing algorithms are used to filter and denoise the data, and then the data changes at different time and space points are analyzed to obtain vibration response feature sequences and traffic load feature sequences.

[0020] In some embodiments, step S200 may include the following steps S210 to S260:

[0021] Step S210: Divide the structural monitoring data set into three-dimensional data blocks according to the monitoring location and timestamp, and divide the data into block data sequences containing spatial coordinate markers and time series markers. The spatial dimension of the block data sequence corresponds to the deployment density of the sensing devices, and the time dimension corresponds to the sampling interval of the preset monitoring period.

[0022] Three-dimensional data segmentation divides the structural monitoring data set into blocks based on two dimensions: monitoring location and timestamp, forming data blocks with three-dimensional characteristics. Spatial coordinate markers are used to specify the exact location of each data block within the bridge structure, while time series markers indicate the corresponding time point. The spatial dimension of the segmented data sequence is related to the deployment density of the sensing devices; the denser the deployment, the higher the spatial resolution. The temporal dimension is related to the sampling interval of the preset monitoring period; the smaller the sampling interval, the higher the accuracy of the temporal dimension. For example, the structural monitoring data set can be segmented according to the deployment location of the sensing devices and the preset sampling interval.

[0023] Step S220: Suppress environmental noise in the segmented data sequence. By analyzing the physical field characteristics of the background environment, identify interference signal components that are unrelated to the dynamic response of the structure, and filter out the interference signal components.

[0024] Environmental noise suppression aims to remove noise interference from background environmental factors in the segmented data sequence, enabling the data to more accurately reflect the true state of the bridge structure. Background environmental physical field characteristics refer to the influence of physical factors in the surrounding environment, such as temperature and wind speed, on the monitoring data. Interference signal components are signal components unrelated to the dynamic response of the bridge structure, such as signal fluctuations caused by external wind speed and temperature changes. For example, the background environmental physical field characteristics are first analyzed, and interference signal components are identified by establishing a physical model or using machine learning algorithms. For instance, temperature and wind speed data from the surrounding environment are acquired using temperature and wind speed sensors, and then the correlation between these data and the segmented data sequence is analyzed to identify interference signals unrelated to the structure's dynamic response. Next, filtering algorithms, such as low-pass and high-pass filters, are used to filter out the interference signal components, resulting in a cleaner segmented data sequence.

[0025] Step S230: Separate the vibration signal source from the block data sequence. Based on the physical field propagation characteristics of each component of the bridge structure, distinguish the vibration signal components generated by the elastic deformation of the structure. The frequency change trend of the vibration signal components is negatively correlated with the change of structural stiffness.

[0026] Isolating vibration signal sources involves identifying the signal sources related to bridge structural vibration from a segmented data sequence after environmental noise suppression. The physical field propagation characteristics of each component of the bridge structure represent the characteristics of different components during vibration transmission, such as the propagation speed and attenuation of vibration in the beams and piers. Vibration signal components generated by the elastic deformation of the structure are vibration signals produced due to the elastic deformation of the bridge structure under load. The frequency variation trend of vibration signal components is negatively correlated with the change in structural stiffness, indicating that the frequency of the vibration signal increases when the structural stiffness decreases. For example, based on the physical model of the bridge structure and the physical field propagation characteristics of each component, signal separation algorithms, such as independent component analysis (ICA), are used to separate the vibration signal sources from the segmented data sequence. Then, by performing spectral analysis on the separated vibration signals, the vibration signal components generated by the elastic deformation of the structure are determined. For example, in a bridge, when the stiffness of the beam decreases due to long-term use, processing the collected segmented data sequence and separating the vibration signal sources reveals an increase in the frequency of the vibration signal through spectral analysis.

[0027] Step S240: Construct multi-scale dynamic features for vibration signal components. By analyzing the energy transfer law and time domain waveform distortion characteristics of signals in different frequency bands, generate a vibration response feature sequence.

[0028] Multi-scale dynamic feature construction involves extracting and analyzing features from vibration signal components at different scales and dimensions. The energy transfer patterns of signals in different frequency bands refer to the changes and transfers of signal energy within different frequency ranges. Time-domain waveform distortion features describe the distortion of the vibration signal waveform in the time domain, such as changes in peak values, valleys, and slope. Specifically, the separated vibration signal components undergo multi-scale analysis, for example, using wavelet transform to decompose the signal into different frequency bands, and then analyzing the energy distribution and time-domain waveform characteristics within each band. By comprehensively analyzing the energy transfer patterns and time-domain waveform distortion characteristics of signals in different frequency bands, a feature sequence reflecting the vibration response of the bridge structure is generated. For example, performing wavelet transform on the vibration signal components of a bridge decomposes it into sub-signals of multiple frequency bands. Analyzing the energy distribution and time-domain waveform of each sub-signal reveals that energy transfer occurs in certain frequency bands at specific time points, and the time-domain waveform also shows distortion. Organizing this feature information forms the vibration response feature sequence.

[0029] In some embodiments, step S240 may include the following steps S241 to S245:

[0030] Step S241: Based on the theoretical frequency range of each key component of the bridge structure, determine the frequency band division interval of the vibration signal component, decompose the vibration signal component into multiple sub-signals covering different frequency ranges, and the frequency boundary of each sub-signal corresponds to the frequency interval of the structural modal.

[0031] The theoretical frequency range of each key component of a bridge structure is the inherent frequency range of each key component, derived from the bridge structure's design and mechanical analysis. Frequency band division divides the frequency range of the vibration signal components into different intervals. Multi-level sub-signals are signals with multiple different frequency ranges obtained after decomposing the vibration signal components. The structural modal frequency interval is the frequency difference between different modes of the bridge structure. For example, firstly, based on the bridge structure's design drawings and mechanical model, the theoretical frequency range of each key component is calculated. Then, based on these theoretical frequency ranges, the frequency band division intervals of the vibration signal components are determined, for example, dividing the frequency range into low-frequency, mid-frequency, and high-frequency bands. Next, a suitable signal decomposition algorithm, such as wavelet packet decomposition, is used to decompose the vibration signal components into multi-level sub-signals covering different frequency ranges, so that the frequency boundaries of each sub-signal correspond to the structural modal frequency interval. For example, for the vibration signal components of a bridge, based on the theoretical frequency ranges of its beams and piers, the signal is decomposed into three frequency band sub-signals, each corresponding to a different structural modal frequency range.

[0032] Step S242: Calculate the energy distribution of each sub-signal, count the proportion of signal energy in each frequency range and the frequency of time domain peak occurrence, and generate an energy feature vector.

[0033] Energy distribution calculation involves determining the energy distribution of each sub-signal within different frequency intervals. The signal energy percentage is the proportion of signal energy within each frequency interval to the total energy. The time-domain peak frequency is the number of times a signal peak occurs in the time domain. The energy feature vector is a vector containing information such as the signal energy percentage within each frequency interval and the time-domain peak frequency. For example, energy calculations are performed on each sub-signal obtained from the decomposition, such as using power spectral density estimation to calculate the signal energy within each frequency interval. Then, the proportion of signal energy within each frequency interval to the total energy is calculated, along with the frequency of signal peaks in the time domain. This information is then organized into a vector, thus generating the energy feature vector.

[0034] Step S243: Perform waveform morphology analysis on each sub-signal, calculate the waveform slope change rate, peak symmetry and half-wave width parameters within the sliding time window, and generate a time-domain morphological feature vector.

[0035] Waveform morphology analysis analyzes the shape and characteristics of a signal's waveform. A sliding time window is a fixed-length time interval that slides along the signal's time axis. The waveform slope change rate describes how the slope of the signal waveform changes within the sliding time window. Peak symmetry is the degree of symmetry between the left and right sides of the signal's peak. The half-wave width parameter is the width of the half-wave of the signal waveform. The time-domain morphological feature vector is a vector containing information such as the waveform slope change rate, peak symmetry, and half-wave width parameter. For example, a sliding time window can be set for each sub-signal layer, and the waveform slope change rate, peak symmetry, and half-wave width parameter can be calculated within the window.

[0036] Step S244: Concatenate the energy feature vector and the temporal morphological feature vector in dimensional order according to the time series to obtain a joint feature vector, and perform time axis calibration on the joint feature vector to keep the feature vectors of different frequency band sub-signals synchronized on the time mark.

[0037] Dimensional concatenation merges the energy feature vector and the temporal morphological feature vector in terms of dimensions. The joint feature vector is a vector containing more information obtained by concatenating the energy feature vector and the temporal morphological feature vector. Time axis calibration adjusts the time stamp of the joint feature vector to ensure that the feature vectors of sub-signals in different frequency bands are consistent in time. Specifically, the energy feature vector and temporal morphological feature vector of each sub-signal layer are concatenated in chronological order, for example, by connecting the energy feature vector and temporal morphological feature vector at the same time point. Then, a time synchronization algorithm, such as cross-correlation, is used to calibrate the time axis of the joint feature vector to ensure that the feature vectors of sub-signals in different frequency bands are synchronized in time stamp. For example, for sub-signals in different frequency bands, after concatenating their corresponding energy feature vectors and temporal morphological feature vectors, a cross-correlation algorithm is used to find the time offset between them, and adjustments are made to align them in time.

[0038] Step S245: Arrange the calibrated joint feature vectors in frequency band order to generate a vibration response feature sequence. Each time node of the vibration response feature sequence contains the combination of energy and morphological features of each frequency band at the corresponding time.

[0039] The calibrated joint eigenvectors are arranged in frequency band order, that is, the joint eigenvectors of different frequency bands are arranged together sequentially. The vibration response characteristic sequence is a sequence composed of the arranged joint eigenvectors, and each time node contains the combination of energy and morphological characteristics of different frequency bands at the corresponding time. For example, the time-axis calibrated joint eigenvectors can be arranged in order of frequency band from low to high or from high to low to form a sequence.

[0040] Step S250: Extract the load action signal from the segmented data sequence. Based on the spatiotemporal propagation law of physical field disturbance during vehicle passage, separate the stress and strain signal components caused by the vehicle load. The amplitude change of the stress and strain signal components is positively correlated with the vehicle weight and driving speed.

[0041] Load action signals are the signals generated by load factors such as vehicle traffic on the bridge structure. The spatiotemporal propagation law of physical field disturbances during vehicle traffic refers to the characteristics of the propagation of disturbances to the surrounding physical fields (such as stress and strain fields) caused by vehicles traveling on the bridge in time and space. Stress-strain signal components are the stress and strain changes in the bridge structure caused by vehicle loads. The amplitude changes of stress-strain signal components are positively correlated with vehicle weight and speed, indicating that the heavier the vehicle and the faster the speed, the larger the amplitude of the stress-strain signal. For example, based on the spatiotemporal propagation law of physical field disturbances during vehicle traffic, signal separation algorithms, such as physical model-based separation algorithms, can be used to separate the stress-strain signal components caused by vehicle loads from the block data sequence. For instance, by establishing a physical field model during vehicle traffic, analyzing the propagation of stress and strain during vehicle movement, stress-strain signal components related to vehicle loads can be identified from the block data sequence. Meanwhile, by analyzing the data collected under different vehicle weights and driving speeds, the positive correlation between the amplitude changes of stress-strain signal components and vehicle weight and driving speed can be verified.

[0042] Step S260: Quantify the spatiotemporal distribution characteristics of the stress-strain signal components. By analyzing the propagation delay and amplitude attenuation of signals at different spatial locations, generate a traffic load characteristic sequence.

[0043] Spatiotemporal distribution feature quantization is the process of quantifying the temporal and spatial distribution characteristics of stress-strain signal components. Propagation delay is the time required for a signal to travel from one spatial location to another. Amplitude attenuation law describes the change in signal amplitude during propagation. A traffic load feature sequence is a data sequence that reflects the load distribution characteristics during vehicle traffic. For example, by analyzing data on stress-strain signal components at different spatial locations and measuring the time difference of signal propagation from one location to another, the propagation delay can be obtained. Simultaneously, the amplitude attenuation during propagation can be analyzed to summarize the amplitude attenuation law. Based on these propagation delays and amplitude attenuation laws, appropriate quantization methods, such as statistical analysis, can be used to generate a traffic load feature sequence. For instance, stress sensors can be deployed at different locations on a bridge to collect stress-strain signal components. The propagation delay and amplitude attenuation between these locations can be analyzed, and this information can be processed to generate a traffic load feature sequence.

[0044] In some embodiments, step S260 may include the following steps S261 to S266:

[0045] Step S261: Perform sensor channel correlation analysis on the stress-strain signal components, calculate the synchronicity and phase difference of the amplitude changes of signals at different sensing positions, and construct a directed graph of the load propagation path. The nodes of the directed graph represent the sensing channel positions, and the edge weights represent the signal propagation intensity.

[0046] Sensor channel correlation analysis examines the correlation between stress and strain signal components acquired by different sensor channels. Amplitude change synchronicity refers to the degree of synchronization in signal amplitude changes at different sensor locations. Phase difference represents the phase difference between signals at different sensor locations. A directed load propagation path graph is a tool used to graphically represent the propagation path of a load in a bridge structure, where nodes represent the locations of sensor channels, and edge weights represent the propagation intensity of the signal between different sensor locations. For example, analyzing stress and strain signal components acquired at different sensor locations calculates their amplitude change synchronicity and phase difference. For instance, cross-correlation analysis is used to calculate the correlation between different signals, yielding amplitude change synchronicity and phase difference. Then, based on these calculation results, a directed load propagation path graph is constructed, using sensor channel locations as nodes and signal propagation intensity as edge weights.

[0047] Step S262: Determine the dominant propagation direction of the load based on the directed graph of the load propagation path, extract the signal arrival time difference along the propagation direction, and calculate the load propagation speed parameter by the ratio of the time difference to the sensor spacing.

[0048] The dominant load propagation direction is the direction in which the load primarily propagates in the directed load propagation path graph. The signal arrival time difference (OTD) is the time difference required for a signal to travel from one sensor location to another along the dominant load propagation direction. The load propagation velocity parameter is the speed at which the load propagates within the bridge structure. Specifically, the constructed directed load propagation path graph is analyzed to identify the path with the strongest signal propagation intensity, which is determined as the dominant load propagation direction. Then, the time difference for the signal to travel from one sensor location to another along this dominant propagation direction is measured. Finally, this time difference is divided by the distance between the two sensors to obtain the load propagation velocity parameter.

[0049] Step S263: Perform load event segmentation on the stress-strain signal components, use the adaptive threshold method to identify signal pulse events caused by individual vehicle loads, and record the start time, peak time and end time of each pulse event.

[0050] Load event segmentation distinguishes different load events within stress-strain signal components. Adaptive thresholding is a method that automatically adjusts the threshold based on signal characteristics. A signal pulse event caused by a single vehicle load is the pulse change in the stress-strain signal generated when a vehicle crosses a bridge. The start time, peak time, and end time are the times when the signal pulse event begins, reaches its peak, and ends, respectively. For example, using adaptive thresholding, a suitable threshold is dynamically determined based on the statistical characteristics of the stress-strain signal components, such as the mean and standard deviation. When the signal exceeds this threshold, it is considered the start of a signal pulse event, and the start time is recorded; when the signal reaches its maximum value, the peak time is recorded; when the signal falls below the threshold again, the end time is recorded. For instance, analyzing the acquired stress-strain signal components, an adaptive thresholding method is used to determine a threshold. Recording begins when the signal exceeds this threshold, the time corresponding to the maximum value of the signal is found as the peak time, and recording stops when the signal falls below the threshold again. This completes the segmentation of a signal pulse event and records its start time, peak time, and end time.

[0051] Step S264: Statistically analyze the frequency and peak amplitude distribution of pulse events within a unit time to generate a statistical feature vector of load time interval.

[0052] The frequency of a pulse event per unit time is the number of times a signal pulse event occurs within a fixed time period. The peak amplitude distribution is the distribution of the peak amplitude of the signal pulse events. The load time interval statistical feature vector is a vector containing information about the frequency of pulse events and the peak amplitude distribution. For example, given a unit time, such as one minute, the number of signal pulse events occurring within that minute is counted. Simultaneously, statistical analysis is performed on the peak amplitude of all signal pulse events to obtain the distribution of peak amplitudes, such as the mean and standard deviation. Combining the statistical information of the pulse event frequency and peak amplitude distribution into a vector generates the load time interval statistical feature vector.

[0053] Step S265: Based on the deployment location of the sensing devices on the key components of the bridge, assign a corresponding spatial coordinate index to each stress-strain signal component, and keep the spatial coordinate index corresponding to the component number of the bridge structure.

[0054] The deployment location of sensing devices on key bridge components refers to the specific installation positions of strain sensors and other equipment on critical components such as bridge piers and beams. Spatial coordinate indices are used to identify the spatial location of each stress-strain signal component within the bridge structure. The component numbers of the bridge structure are pre-assigned numbers to each component. For example, based on the installation location of the sensing devices, the spatial coordinates corresponding to each stress-strain signal component are determined. Then, these spatial coordinates are mapped to the component numbers of the bridge structure, assigning a unique spatial coordinate index to each stress-strain signal component. For instance, if multiple strain sensors are installed on the bridge beams, the spatial coordinates of the stress-strain signal components collected by each sensor are determined based on the sensor's installation location. These coordinates are then mapped to the component numbers of the beams, assigning a spatial coordinate index to each signal component.

[0055] Step S266: Combine the load propagation speed parameters, load time interval statistical feature vectors, and spatial coordinate indices in time series order to generate a traffic load feature sequence.

[0056] The previously calculated load propagation velocity parameters, load time interval statistical feature vectors, and spatial coordinate indices assigned to stress-strain signal components are combined in chronological order, meaning that the load propagation velocity parameters, load time interval statistical feature vectors, and spatial coordinate indices at the same time point are combined together. This combination generates a traffic load feature sequence that comprehensively reflects the load distribution characteristics during vehicle traffic. Specifically, the load propagation velocity parameters and load time interval statistical feature vectors are standardized, converting them into dimensionless scalar values. These standardized scalar values ​​are then combined with the spatial coordinate indices in chronological order to generate the traffic load feature sequence. For example, load propagation velocity parameters, load time interval statistical feature vectors, and spatial coordinate indices are calculated at different time points. Arranging and combining these data in chronological order forms the traffic load feature sequence, which clearly shows the distribution of vehicle loads on the bridge at different time points.

[0057] Step S300: Perform correlation analysis on the vibration response characteristic sequence and the traffic load characteristic sequence to generate a structural damage sensitive feature matrix. The structural damage sensitive feature matrix is ​​used to quantify the degree of abnormal fluctuation in the structural response under different load conditions.

[0058] Correlation analysis studies the relationship between vibration response characteristic sequences and traffic load characteristic sequences. The structural damage sensitivity matrix is ​​a matrix that reflects the damage sensitivity of a bridge structure under different load conditions, quantifying the degree of abnormal fluctuations in the structural response. For example, by employing appropriate correlation analysis methods, such as correlation analysis and causal analysis, the vibration response characteristic sequences and traffic load characteristic sequences can be analyzed to identify their correlations. Then, the structural damage sensitivity matrix is ​​constructed based on these correlations. For instance, the correlation coefficient between the vibration response characteristic sequences and traffic load characteristic sequences can be calculated, and matrix elements are constructed based on the magnitude and variation of the correlation coefficient to form the structural damage sensitivity matrix. This matrix provides a clear visual representation of the abnormal fluctuations in the bridge structure's response under different load conditions.

[0059] In some embodiments, step S300 may include the following steps S310 to S360:

[0060] Step S310: Align the vibration response feature sequence and the traffic load feature sequence according to the timestamp to generate a time-synchronized joint feature sequence.

[0061] A timestamp is a marker indicating the time of data acquisition. Aligning the vibration response feature sequence and the traffic load feature sequence according to their timestamps means matching data from the same time point in these two sequences. A time-synchronized joint feature sequence is a new sequence formed by combining the aligned vibration response feature sequence and the traffic load feature sequence. For example, first, examine the timestamps of the vibration response feature sequence and the traffic load feature sequence to find their time correspondence. Then, combine the vibration response feature and traffic load feature from the same time point to form a joint feature. Arranging all joint features from different time points in chronological order generates a time-synchronized joint feature sequence. For example, if both the vibration response feature sequence and the traffic load feature sequence have their own timestamps, by comparing the timestamps, the vibration response feature and traffic load feature from the same time point are merged to generate a new joint feature. Arranging the joint features from different time points sequentially yields a time-synchronized joint feature sequence.

[0062] Step S320: Perform cross-correlation modeling on the joint feature sequence, analyze the interdependence between vibration response features and traffic load features at different time scales, and identify feature combination patterns that cause vibration response anomalies due to load changes.

[0063] Cross-correlation modeling establishes a correlation model between vibration response characteristics and traffic load characteristics to analyze their interdependencies at different time scales. These time scales can include short-term, medium-term, and long-term periods. Feature combinations that induce vibration response anomalies due to load changes are the combinations of features that lead to abnormal vibration responses when the load changes. For example, appropriate modeling methods, such as multiple linear regression models or neural network models, can be used to model the joint feature sequences. The interdependencies between vibration response characteristics and traffic load characteristics are analyzed at different time scales, such as calculating their conditional probabilities and mutual information. By analyzing these relationships, feature combinations that induce vibration response anomalies due to load changes can be identified. For example, a neural network model can be trained on the joint feature sequences to analyze the relationship between vibration response characteristics and traffic load characteristics at short-term, medium-term, and long-term time scales, identifying which feature combinations lead to abnormal vibration responses when the load changes.

[0064] In some embodiments, step S320 may include the following steps S321 to S326:

[0065] Step S321: Divide the joint feature sequence into analysis windows of different time scales. Each analysis window contains short-term, medium-term and long-term feature subsequences. The time length of each window is set according to the bridge structure damage evolution cycle.

[0066] Analysis windows at different time scales are obtained by dividing the joint feature sequence according to different time lengths. Short-term, medium-term, and long-term feature subsequences are feature sequences at different time scales, respectively. The bridge structure damage evolution cycle is the time elapsed from the initial appearance of damage to its development to a certain extent. For example, based on the bridge structure damage evolution cycle, the time length of analysis windows at different time scales is determined. For instance, the short-term analysis window is set to one day, the medium-term analysis window to one month, and the long-term analysis window to one year. Then, the joint feature sequence is divided according to these time lengths, with each window containing feature subsequences of the corresponding time scale. For example, for a joint feature sequence, dividing it according to the time lengths of one day, one month, and one year yields different analysis windows, each containing short-term, medium-term, and long-term feature subsequences.

[0067] Step S322: For each time scale feature subsequence, calculate the mutual information value between the vibration response feature and the traffic load feature, and quantify the statistical dependence strength of the two features at that time scale.

[0068] Mutual information is an indicator used to measure the statistical dependence between two random variables. For each feature subsequence at each time scale, calculating the mutual information value between vibration response features and traffic load features quantifies their statistical dependence strength at that time scale. Specifically, the probability distributions of vibration response features and traffic load features are first estimated based on the data in the feature subsequence. Then, the mutual information value is calculated according to the formula for mutual information. The magnitude of the mutual information value indicates the degree of dependence between the two features at that time scale; a larger value indicates a stronger dependence.

[0069] Step S323: Construct a correlation strength matrix based on mutual information values. The row dimension of the matrix corresponds to the vibration response feature dimension, the column dimension corresponds to the traffic load feature dimension, and the element value represents the correlation strength of the corresponding feature pair.

[0070] The correlation strength matrix is ​​used to represent the correlation strength between vibration response features and traffic load features. The row dimensions of the matrix correspond to the dimensions of the vibration response features, and the column dimensions correspond to the dimensions of the traffic load features. The element values ​​in the matrix are the correlation strength of the corresponding feature pairs calculated using mutual information values. Specifically, in its construction, the mutual information value of each vibration response feature and traffic load feature pair is used as an element of the matrix. For example, if the vibration response features have m dimensions and the traffic load features have n dimensions, then the correlation strength matrix is ​​an m×n matrix. The element in the i-th row and j-th column of the matrix represents the correlation strength between the i-th vibration response feature and the j-th traffic load feature, which is determined by their mutual information values.

[0071] Step S324: Analyze the variation pattern of the correlation strength matrix at different time scales, identify feature pairs whose correlation strength shows a monotonically increasing trend with the increase of time scale and whose rate of change exceeds the preset rate of change threshold at time scales exceeding the preset proportion threshold. The preset proportion threshold is determined by statistically analyzing the variation pattern of feature correlation strength in historical monitoring data.

[0072] Analyzing the variation of the correlation strength matrix across different time scales involves observing the changes in matrix element values ​​over short, medium, and long-term time scales. The correlation strength exhibits a monotonically increasing trend with increasing time scale, indicating that the correlation strength between feature pairs continuously increases as the time scale grows. The preset proportion threshold and preset rate of change threshold are pre-set thresholds used to filter feature pairs. The preset proportion threshold is obtained through statistical analysis of the variation patterns of feature correlation strength in historical monitoring data. For example, analyzing the correlation strength matrix at different time scales identifies feature pairs whose correlation strength monotonically increases with increasing time scale. Then, the rate of change for these feature pairs at different time scales is calculated, and the rate of change is compared with a preset rate of change threshold. Simultaneously, it is checked whether the rate of change condition is met on time scales exceeding the preset proportion threshold.

[0073] Step S325: Compare the dynamic response curves of the identified feature pairs, analyze the waveform differences of the vibration response characteristics before and after the load characteristic changes, and determine the feature combination mode that causes abnormal vibration response due to load changes.

[0074] Dynamic response curve comparison involves comparing the dynamic response curves of identified feature pairs. The waveform difference in vibration response characteristics before and after load characteristic changes refers to the changes in the waveform of the vibration response characteristics when the load characteristics change. By analyzing these waveform differences, the characteristic combination patterns that cause abnormal vibration response due to load changes can be determined. For example, the dynamic response curves of the identified feature pairs can be plotted, and the waveforms of the vibration response characteristics before and after load characteristic changes can be compared. For example, by observing changes in parameters such as peak value, trough value, and period of the waveform, the characteristic combination patterns that cause abnormal vibration response due to load changes can be identified. For example, when load characteristics change, a sudden increase in the peak value of the vibration response waveform and a change in the period are observed; by analyzing these waveform differences, the characteristic combination patterns leading to this anomaly can be determined.

[0075] Step S326: Through repeatability verification of feature combination patterns, select feature combination patterns that appear stably in multiple monitoring periods.

[0076] Repeatability verification of feature combination patterns involves checking whether these patterns repeat across multiple monitoring periods. Stably occurring feature combination patterns are those that remain consistent across multiple monitoring periods. For example, analyzing data from multiple monitoring periods can help check whether previously identified feature combination patterns that induce abnormal vibration responses due to load changes appear in every monitoring period.

[0077] Step S330: Calculate the sensitivity of different feature combination patterns to structural damage using the correlation strength quantification method, and select feature combination patterns that meet the preset correlation threshold conditions as strong correlation feature combinations. The preset correlation threshold is determined by the feature correlation strength distribution in historical damage data.

[0078] The correlation strength quantification method is used to calculate the correlation strength between feature combination patterns and structural damage. The sensitivity of different feature combination patterns to structural damage represents the degree of influence of changes in feature combination patterns on structural damage. A preset correlation threshold is a pre-defined threshold used to screen strongly correlated feature combinations. This threshold is obtained through statistical analysis of the feature correlation strength distribution in historical damage data. For example, appropriate correlation strength quantification methods, such as partial correlation coefficient method or regression coefficient method, are used to calculate the correlation strength between different feature combination patterns and structural damage. Then, the calculated correlation strength is compared with the preset correlation threshold, and feature combination patterns with a correlation strength greater than the preset correlation threshold are selected as strongly correlated feature combinations.

[0079] Step S340: Perform dynamic evolution analysis on the strongly correlated feature combination, track the trend of feature value changes in different monitoring periods, and extract abnormal fluctuation features where feature values ​​deviate from the healthy baseline state.

[0080] Dynamic evolution analysis studies the changes in strongly correlated feature combinations across different monitoring periods. The trend of feature value changes represents the direction and extent of these changes over time. The health baseline state is the range of feature values ​​for the strongly correlated feature combinations when the bridge structure is in a healthy state. Abnormal fluctuation characteristics are the unusual changes exhibited when feature values ​​deviate from the health baseline state. For example, by recording and analyzing the feature values ​​of strongly correlated feature combinations across different monitoring periods, and plotting curves of feature value changes over time, the trend can be observed. Then, the feature values ​​are compared with the health baseline state to identify deviations and extract abnormal fluctuation characteristics.

[0081] In some embodiments, step S340 may include the following steps S341 to S346:

[0082] Step S341: Based on historical monitoring data of the bridge's health status, construct a benchmark feature library of strongly correlated feature combinations. The benchmark feature library contains the normal fluctuation range of each feature combination under different load conditions.

[0083] Historical monitoring data on bridge health status is collected when the bridge structure is in a healthy state. The benchmark feature library is a database storing the normal fluctuation range of strongly correlated feature combinations under different load conditions. For example, historical monitoring data on bridge health status is collected and analyzed to determine the normal fluctuation range of strongly correlated feature combinations under different load conditions. For instance, through statistical analysis of historical data, statistical parameters such as the mean and standard deviation of feature combinations under different load conditions are calculated, and these parameters are used as the boundaries of the normal fluctuation range to construct the benchmark feature library. For example, under different vehicle load conditions, historical data on strongly correlated feature combinations are analyzed to obtain their mean and standard deviation under each load condition, and this data is stored in the benchmark feature library as a basis for subsequent comparisons.

[0084] Step S342: Compare the strongly correlated feature combinations of the current monitoring period with the benchmark feature library, calculate the deviation between the current feature value and the benchmark feature value of each feature combination, and use the dynamic threshold method to quantify the deviation. The threshold is adaptively adjusted as the load conditions change.

[0085] The strongly correlated feature combinations for the current monitoring period are the data of strongly correlated feature combinations collected within the current monitoring period. The baseline feature value is the normal feature value of the strongly correlated feature combinations stored in the baseline feature library under the corresponding load conditions. The deviation is the degree of difference between the current feature value and the baseline feature value. The dynamic threshold method is a method that can automatically adjust the threshold according to changes in load conditions. For example, the feature values ​​of the strongly correlated feature combinations for the current monitoring period are compared with the baseline feature values ​​under the corresponding load conditions in the baseline feature library, and the deviation between them is calculated. For example, the deviation can be calculated using the absolute value difference method. Simultaneously, the dynamic threshold method adaptively adjusts the deviation threshold according to changes in load conditions. For example, when load conditions change, the deviation threshold is adjusted based on historical data and the current load situation, so that the threshold can more accurately reflect the normal fluctuation range of the feature values.

[0086] Step S343: Track the trend of deviation within a continuous monitoring period. By calculating the first derivative of the deviation time series, identify the characteristic combination where the derivative is positive for multiple consecutive periods and the absolute value of the derivative exceeds the preset rate of change threshold. The preset rate of change threshold is determined based on the statistical analysis of the historical fluctuation derivatives of deviation in the health benchmark feature library.

[0087] The trend of deviation over a continuous monitoring period refers to the change in deviation over multiple consecutive monitoring periods. The first derivative of the deviation time series is the rate of change of deviation over time. The preset rate of change threshold is a pre-set threshold used to screen feature combinations. This threshold is obtained through statistical analysis of the historical fluctuation derivatives of deviation in the health benchmark feature library. For example, deviation data from consecutive monitoring periods are compiled into a time series, and the first derivative of this time series is calculated. Then, feature combinations where the derivative is positive for multiple consecutive periods and the absolute value of the derivative exceeds the preset rate of change threshold are identified.

[0088] Step S344: Detect abrupt change points in the deviation time series, calculate the difference between the deviation of the current period and the deviation of the previous period, and determine the deviation abrupt change feature combination when the absolute value of the difference exceeds the upper limit of the normal fluctuation range in the health benchmark feature library.

[0089] Abrupt change detection identifies points of sudden change in the deviation time series. The difference between the deviation of the current period and the deviation of the previous period is the difference between the deviations of two adjacent monitoring periods. The upper limit of the normal fluctuation range in the health benchmark feature library is the upper limit of the normal fluctuation range of deviations stored in the benchmark feature library. A deviation abrupt change feature combination is a combination of features indicating a sudden change in deviation. For example, analyzing the deviation time series, the difference between the deviation of the current period and the deviation of the previous period is calculated. The absolute value of this difference is compared with the upper limit of the normal fluctuation range in the health benchmark feature library. When the absolute value of the difference exceeds the upper limit, the feature combination is determined to be a deviation abrupt change feature combination.

[0090] Step S345: Perform fluctuation morphology analysis on the continuously increasing feature combination and the abrupt change feature combination, extract the peak offset, waveform distortion rate and periodic change rate in the feature value time series, and generate abnormal fluctuation morphology parameters.

[0091] Continuously increasing feature combinations are those where the deviation increases continuously over time, while abruptly changing feature combinations are those where the deviation changes suddenly. Waveform analysis involves analyzing the waveforms of the eigenvalue time series of these feature combinations. Peak offset is the shift in the peak position within the eigenvalue time series. Waveform distortion rate is the degree of difference between the waveform of the eigenvalue time series and a normal waveform. Periodicity rate of change is the periodic variation of the eigenvalue time series. Abnormal waveform parameters are parameters that include information such as peak offset, waveform distortion rate, and periodicity rate of change. For example, analyzing the eigenvalue time series of continuously increasing and abruptly changing feature combinations allows for the calculation of peak offset, waveform distortion rate, and periodicity rate of change. Specifically, peak offset is calculated by comparing the peak position of the eigenvalue time series with the normal peak position; waveform distortion rate is obtained by calculating waveform similarity; and periodicity rate of change is calculated by analyzing periodic variations. Combining these parameters generates abnormal waveform parameters.

[0092] Step S346: Combining the spatial location information corresponding to the feature combination, the abnormal fluctuation morphology parameters are associated with the physical location of the bridge structural components to determine the specific structural region corresponding to the abnormal fluctuation, and the abnormal fluctuation morphology parameters are converted into numerical features that characterize the degree of damage sensitivity to generate abnormal fluctuation features.

[0093] The spatial location information corresponding to the feature combination is the spatial location of the bridge structural component corresponding to the feature combination. Associating the abnormal fluctuation morphological parameters with the physical location of the bridge structural component means mapping the abnormal fluctuation morphological parameters to specific bridge structural components. The specific structural region corresponding to the abnormal fluctuation is the specific location of the bridge structure where the abnormal fluctuation occurs. Converting the abnormal fluctuation morphological parameters into numerical features characterizing damage sensitivity involves transforming the abnormal fluctuation morphological parameters into a value that reflects damage sensitivity using a certain method. For example, based on the spatial location information corresponding to the feature combination, the abnormal fluctuation morphological parameters are associated with the physical location of the bridge structural component. For example, using the bridge's BIM model, the abnormal fluctuation morphological parameters are mapped to specific component numbers. Then, using a suitable conversion method, such as a linear mapping method, the abnormal fluctuation morphological parameters are converted into numerical features characterizing damage sensitivity. For example, parameters such as peak offset, waveform distortion rate, and periodicity rate of change in the abnormal fluctuation morphological parameters are weighted and summed to obtain a value, which is used as the numerical feature characterizing damage sensitivity to generate the abnormal fluctuation feature.

[0094] Step S350: Arrange the abnormal fluctuation features in a matrix according to spatial location and time series to construct an initial feature matrix with spatial location as rows and time series as columns. The matrix element values ​​reflect the degree of abnormal fluctuation at the corresponding spatiotemporal point.

[0095] The abnormal fluctuation characteristics are arranged in a matrix according to their spatial location and time series. This means organizing the abnormal fluctuation characteristics according to the spatial location of the bridge structure and the monitoring time series. The initial feature matrix is ​​a matrix with spatial locations as rows and time series as columns. The element values ​​in the matrix represent the degree of abnormal fluctuation at the corresponding spatiotemporal point. For example, the abnormal fluctuation characteristic values ​​of each spatial location at different time points are organized into the elements of the matrix. For instance, if the bridge has m different spatial locations and n time points are monitored, then the initial feature matrix is ​​an m×n matrix, where the element in the i-th row and j-th column represents the degree of abnormal fluctuation at the i-th spatial location at the j-th time point. Through this matrix arrangement, the abnormal fluctuation situation at different spatiotemporal points can be visually observed.

[0096] Step S360: Enhance the spatiotemporal correlation of the initial feature matrix by analyzing the eigenvalue transmission patterns of adjacent spatial locations and continuous time series to generate a structural damage-sensitive feature matrix.

[0097] Spatiotemporal correlation enhancement strengthens the spatial and temporal correlations within the initial feature matrix. The eigenvalue transfer patterns between adjacent spatial locations and continuous time series represent the transmission and change of eigenvalues ​​between spatially adjacent locations and temporally continuous points. The structural damage sensitivity feature matrix, after spatiotemporal correlation enhancement, more accurately reflects the bridge structure's damage sensitivity. For example, the initial feature matrix is ​​analyzed to identify the eigenvalue transfer patterns between adjacent spatial locations and continuous time series. This can be achieved by calculating the correlation between eigenvalues ​​of adjacent spatial locations and the rate of change of eigenvalues ​​in continuous time series. Then, based on these patterns, the initial feature matrix is ​​adjusted and enhanced to generate the structural damage sensitivity feature matrix. For instance, spatial interpolation and temporal smoothing methods can be used to process the initial feature matrix, strengthening the spatiotemporal correlations between matrix elements, thus obtaining the structural damage sensitivity feature matrix.

[0098] In some embodiments, step S360 may include the following steps S361 to S366:

[0099] Step S361: Construct a physical connectivity topology diagram of key components based on the bridge structural design drawings. The physical connectivity topology diagram includes the connection relationships between components and the transmission path parameters. Establish a mapping relationship between the row dimensions of the initial feature matrix and the nodes of the topology diagram.

[0100] Bridge structural design drawings are drawings created during bridge design that contain detailed information about the bridge structure. A physical connectivity topology diagram of key components is a tool used to graphically represent the physical connections between key bridge components, including connection relationships and force transmission path parameters. These parameters can include information such as the direction and efficiency of force transmission. Mapping the rows of the initial feature matrix to nodes in the topology diagram involves assigning each row of the initial feature matrix to a node in the topology diagram. For example, based on the bridge structural design drawings, the connection relationships and force transmission path parameters between key components are analyzed, and a physical connectivity topology diagram is constructed. This includes determining the connection method and force transmission path between the beam and the pier. Then, the rows of the initial feature matrix are mapped to nodes in the topology diagram; for example, the first row of the initial feature matrix is ​​mapped to the first node in the topology diagram. This mapping relationship connects the initial feature matrix to the physical structure of the bridge.

[0101] Step S362: Traverse each spatial location element in the initial feature matrix, extract the feature value sequence of its adjacent spatial locations according to the connection relationship of the topology graph, calculate the attenuation coefficient and delay parameter of the feature value on the spatial transmission path, and construct the spatial transfer function.

[0102] Traversing each spatial element in the initial feature matrix involves processing each element sequentially. The eigenvalue sequence of adjacent spatial locations is a sequence of eigenvalues ​​from locations adjacent to the current spatial location. The attenuation coefficient and delay parameter along the spatial transmission path represent the degree of attenuation and transmission delay of the eigenvalue during spatial transmission, respectively; the delay parameter is a normalized parameter. The spatial transfer function describes the spatial transmission pattern of eigenvalues. For example, for each element in the initial feature matrix, its adjacent spatial locations are found based on the topological graph connections. For instance, nodes adjacent to the current node are located in the topological graph. Then, the eigenvalue sequences of adjacent spatial locations are extracted, and the attenuation coefficient and delay parameter along the spatial transmission path are calculated through analysis of these eigenvalues. For example, the attenuation coefficient is calculated by comparing the magnitudes of the eigenvalues ​​of adjacent spatial locations; the delay parameter is calculated by analyzing the time delay of the eigenvalues. Based on these coefficients and parameters, the spatial transfer function is constructed. For example, an exponential attenuation function and a delay operator can be used to construct the spatial transfer function.

[0103] Step S363: Apply the spatial transfer function to the initial feature matrix, perform neighborhood feature fusion on each element value, and dynamically determine the fusion weight by the attenuation coefficient of the spatial transfer path to generate a spatial correlation enhancement matrix.

[0104] Applying the spatial transfer function to the initial feature matrix means applying the spatial transfer function to each element of the initial feature matrix. Neighborhood feature fusion combines the eigenvalues ​​of the current element with those of its neighbors. The fusion weights are dynamically determined by the attenuation coefficient of the spatial transfer path, indicating that the fusion weights are adjusted according to the attenuation of the eigenvalues ​​along the spatial transfer path. The spatial correlation enhancement matrix is ​​a matrix whose spatial correlation is enhanced after neighborhood feature fusion. For example, for each element in the initial feature matrix, it is fused with the eigenvalues ​​of its neighbors according to the spatial transfer function. For example, a weighted average method can be used for fusion, with the fusion weights determined by the attenuation coefficient of the spatial transfer path. Through such fusion, a spatial correlation enhancement matrix is ​​generated. For example, for an element in the initial feature matrix, it is weighted and averaged with the eigenvalues ​​of its neighbors according to the spatial transfer function, with the fusion weights determined by the attenuation coefficient. After processing all elements, the spatial correlation enhancement matrix is ​​obtained.

[0105] Step S364: Perform time series segmentation on the spatial correlation enhancement matrix, divide continuous time windows according to the bridge monitoring cycle, calculate the rate of change and cumulative increment of feature values ​​in each window, and establish a time evolution rule base.

[0106] Time series segmentation of the spatial correlation enhancement matrix involves dividing the matrix according to the time dimension. Dividing the time series into continuous time windows based on the bridge monitoring cycle involves dividing the time series according to the length of the bridge monitoring cycle. The rate of change of eigenvalues ​​within each window is the speed at which eigenvalues ​​change within that time window, and the cumulative increment is the cumulative increase in eigenvalues ​​within that time window. The time evolution rule base is a database storing time evolution rules that can be used to describe the changing patterns of eigenvalues ​​over time. For example, based on the bridge monitoring cycle, the time series of the spatial correlation enhancement matrix is ​​divided into continuous time windows. For example, the time series can be divided into windows of one day. Then, the rate of change and cumulative increment of eigenvalues ​​within each window are calculated. For example, the rate of change is obtained by calculating the difference between eigenvalues ​​at adjacent time points; the cumulative increment is obtained by summing the differences between eigenvalues ​​within the window. Based on these rates of change and cumulative increments, a time evolution rule base is established. For example, the changing trends and patterns of eigenvalues ​​within different time windows are summarized and stored in the time evolution rule base.

[0107] Step S365: Adjust the spatial correlation enhancement matrix in terms of time dimension based on the time evolution rule base, strengthen the trend of the feature value sequence that conforms to the monotonic change rule, perform phase synchronization calibration on the sequence that conforms to the periodic rule, and generate a spatiotemporal coupling feature matrix.

[0108] Adjusting the spatial correlation enhancement matrix in the time dimension based on a time evolution rule base involves modifying the matrix according to the rules in the rule base. Eigenvalue sequences conforming to monotonically changing rules are sequences whose eigenvalues ​​monotonically increase or decrease over time. Strengthening the trend of eigenvalue sequences conforming to monotonically changing rules enhances this monotonically changing trend. Sequences conforming to periodic rules are sequences whose eigenvalues ​​exhibit periodic changes over time. Phase synchronization calibration of sequences conforming to periodic rules adjusts the phase of the sequence to synchronize it with other related sequences. The spatiotemporal coupling feature matrix is ​​a matrix that, after time dimension adjustment, better couples the spatial and temporal dimensions. For example, the spatial correlation enhancement matrix can be analyzed according to the rules in the time evolution rule base. For eigenvalue sequences conforming to monotonically changing rules, trend strengthening is performed using methods such as linear fitting. For sequences conforming to periodic rules, phase synchronization calibration is performed using methods such as calculating phase differences. For example, for a monotonically increasing sequence of eigenvalues, its trend can be fitted using linear regression to enhance this trend; for a periodically changing sequence, its phase difference with other related sequences can be calculated and adjusted to synchronize their phases. After such adjustments, a spatiotemporal coupling feature matrix is ​​generated.

[0109] Step S366: By analyzing the correlation between the element values ​​in the spatiotemporal coupling feature matrix and historical damage events, calculate the damage indication contribution of each element, retain elements with contribution values ​​higher than a set threshold, and generate a structural damage sensitive feature matrix.

[0110] The correlation between element values ​​in the spatiotemporal coupling feature matrix and historical damage events refers to the relationship between the element values ​​in the spatiotemporal coupling feature matrix and past bridge damage events. Damage indication contribution is the degree to which an element value contributes to the damage indication. A set threshold is a pre-defined threshold used to filter elements. For example, the damage indication contribution of each element is calculated by analyzing the element values ​​in the spatiotemporal coupling feature matrix and historical damage events, such as through correlation analysis. For instance, the correlation between element values ​​and the time and severity of historical damage events is calculated, and this correlation is used as the damage indication contribution. Then, the damage indication contribution of each element is compared with the set threshold; elements with a contribution higher than the set threshold are retained, while elements with a contribution lower than the set threshold are removed.

[0111] Step S400: Based on the structural damage sensitive feature matrix, predict the damage evolution trend and generate the damage development probability distribution of key bridge components. The damage development probability distribution includes the damage location distribution and damage degree quantitative indicators for different monitoring periods.

[0112] Damage evolution trend prediction involves forecasting the development trend of bridge structural damage based on a structural damage sensitivity feature matrix. The damage development probability distribution of key bridge components describes the distribution of the likelihood of damage development in these components across different monitoring periods. This distribution includes the distribution of damage locations and quantitative indicators of damage severity across different monitoring periods. The damage location distribution indicates the possible locations where damage may occur at different monitoring periods, while the damage severity indicator quantifies the degree of damage. For example, appropriate prediction methods, such as machine learning prediction models and time series prediction models, can be used to analyze the structural damage sensitivity feature matrix and predict the damage evolution trend. For instance, a neural network model can be used to train the structural damage sensitivity feature matrix to learn the patterns of damage development. Then, the damage development probability distribution of key bridge components can be generated based on the prediction results.

[0113] In some embodiments, step S400 may include the following steps S410 to S460:

[0114] Step S410: Input the structural damage sensitive feature matrix into the physical field linkage sensing network. The physical field linkage sensing network includes a spatial feature extraction layer, a time series prediction layer, and a probability output layer.

[0115] The Physical Field Interconnected Sensing Network (PNSIN) is a neural network used to process structural damage-sensitive feature matrices and predict damage evolution trends. The spatial feature extraction layer is a layer within the PNSIN used to extract spatial features from the structural damage-sensitive feature matrix. The time series prediction layer performs time series analysis and prediction on the extracted spatial features. The probability output layer outputs the probability distribution of damage development. For example, the structural damage-sensitive feature matrix is ​​input into the PNSIN. Alternatively, the structural damage-sensitive feature matrix can be adjusted according to the network's input format before being input to the spatial feature extraction layer. The spatial feature extraction layer can employ a Convolutional Neural Network (CNN) structure, extracting spatial features from the structural damage-sensitive feature matrix through convolutional kernels. The time series prediction layer can employ a Recurrent Neural Network (RNN) or a Long Short-Term Memory (LSTM) network to perform time series analysis and prediction on the extracted spatial features. The probability output layer can be a fully connected layer that converts the prediction results into a probability distribution of damage development. For example, the features extracted by the spatial feature extraction layer are input into an LSTM network for time series prediction, and finally, the probability distribution of damage development is output through a fully connected layer.

[0116] Step S420: Perform spatial pattern recognition on the structural damage-sensitive feature matrix through the spatial feature extraction layer, extract the spatial distribution pattern of damage-sensitive features in different regions of the bridge structure, and generate a spatial distribution feature map.

[0117] Spatial pattern recognition identifies spatial patterns within a structural damage-sensitive feature matrix. The spatial distribution patterns of damage-sensitive features across different regions of a bridge structure represent their distribution at different locations. A spatial distribution feature map is a tool that graphically represents these patterns. For example, the spatial feature extraction layer uses methods such as convolutional neural networks to process the structural damage-sensitive feature matrix. For instance, convolutional kernels are used to perform convolution operations on the matrix to extract features from different regions. Then, methods such as cluster analysis are used to identify the spatial distribution patterns of the damage-sensitive features. For example, regions with similar damage-sensitive features are clustered together. Finally, these spatial distribution patterns are displayed graphically, generating a spatial distribution feature map.

[0118] In some embodiments, step S420 may include the following steps S421 to S426:

[0119] Step S421: Construct a three-dimensional component connection map based on the bridge structural design drawings, and establish a spatial correspondence between the row dimension of the structural damage sensitive feature matrix and the component nodes in the map and the physical structure.

[0120] Bridge structural design drawings are detailed drawings used in bridge design, containing various information about the bridge structure. A 3D component connection atlas is a graphical representation of the connection relationships between bridge components. Establishing a spatial correspondence between the row dimensions of the structural damage sensitivity feature matrix and the component nodes in the atlas means mapping each row of the structural damage sensitivity feature matrix to a component node in the atlas. For example, based on the bridge structural design drawings, the connection relationships between bridge components are analyzed, and a 3D component connection atlas is constructed. This includes determining the connection methods and positional relationships between components such as beams, piers, and supports. Then, the row dimensions of the structural damage sensitivity feature matrix are mapped to the component nodes in the atlas; for example, the first row of the matrix is ​​mapped to the first component node in the atlas.

[0121] Step S422: Expand the adjacency matrix of the structural damage sensitive feature matrix. Based on the connection strength of the nodes in the three-dimensional component connection map, add an associated feature column to the matrix to represent the mechanical transmission path between components. The column dimensions of the expanded matrix include the original features and associated features.

[0122] The adjacency matrix expansion is an extension of the structural damage-sensitive feature matrix. The connection strength of nodes in the 3D component connection map represents the tightness of connections between component nodes. The association feature columns characterizing the mechanical transfer paths between components are newly added feature columns in the matrix to represent these paths. The expanded matrix column dimensions include both original and association features, indicating that new association features have been added to the column dimensions. For example, the association features of the mechanical transfer paths between components are determined based on the connection strength of nodes in the 3D component connection map. For example, the association features are determined based on the connection method and force transmission between the beam and the pier. Then, these association feature columns are added to the structural damage-sensitive feature matrix to obtain the expanded matrix. For example, by adding several columns to the original matrix to represent the association features of the mechanical transfer paths between components, the expanded matrix column dimensions increase, including both original and association features.

[0123] Step S423: Use a graph convolutional network to learn spatial features of the expanded matrix. By aggregating the feature values ​​of each component node and its adjacent nodes, calculate the spatial correlation feature vector considering the mechanical transmission path. The convolution kernel parameters of the graph convolutional network are positively correlated with the elastic modulus of the component material.

[0124] Graph convolutional networks (GCNNs) are convolutional networks used to process graph-structured data. Spatial feature learning is performed on the expanded matrix, meaning that spatial features are extracted from the expanded matrix using a GCNN. Aggregating the feature values ​​of each component node and its neighboring nodes involves comprehensively considering the feature values ​​of each component node and its neighbors. The spatial correlation feature vector considering the force transmission path is a vector containing spatial correlation features that take into account the force transmission path factor. The convolution kernel parameters of the GCNN are positively correlated with the elastic modulus of the component material, indicating that the larger the elastic modulus of the component material, the larger the convolution kernel parameters. For example, the expanded matrix is ​​input into the GCNN. The GCNN aggregates the feature values ​​of each component node and its neighboring nodes through convolution operations. For example, the feature values ​​of a component node and its neighboring nodes are weighted and summed. Then, the convolution kernel parameters are determined based on the elastic modulus of the component material, and the spatial correlation feature vector considering the force transmission path is calculated.

[0125] Step S424: Perform hierarchical clustering on the spatially associated feature vectors, and divide the component nodes into different spatial feature clusters according to feature similarity. Each feature cluster corresponds to a region in the bridge structure with similar damage-sensitive characteristics.

[0126] Hierarchical clustering is a clustering method that groups data objects hierarchically. The feature similarity of spatially related feature vectors represents the degree of similarity between vectors. Spatial feature clusters are groups of component nodes divided based on feature similarity. Each feature cluster corresponds to a region in the bridge structure with similar damage sensitivity characteristics, indicating that component nodes within the same feature cluster have similar damage sensitivity characteristics. For example, hierarchical clustering algorithms, such as agglomerative hierarchical clustering, can be used to cluster spatially related feature vectors. The similarity between spatially related feature vectors is calculated, for example, using Euclidean distance or cosine similarity. Then, component nodes are divided into different spatial feature clusters based on similarity. For example, component nodes with high similarity are clustered together. Through such clustering, the bridge structure is divided into different regions, each with similar damage sensitivity characteristics. For example, the bridge beams can be divided into several different feature clusters, with beam components within each feature cluster having similar damage sensitivity characteristics.

[0127] Step S425: Assign a damage-sensitive identifier to each spatial feature cluster, calculate the mean and variance of the feature values ​​of all nodes within the cluster, and generate a spatial distribution descriptor containing the cluster center coordinates, feature mean, and coverage area.

[0128] Damage sensitivity identifiers are used to identify the degree of damage sensitivity of spatial feature clusters. The mean eigenvalue of all nodes within a cluster is the average of the eigenvalues ​​of all component nodes within a spatial feature cluster. Variance is the dispersion of the eigenvalues. A spatial distribution descriptor is a descriptor containing information such as cluster center coordinates, eigenvalue mean, and coverage. For example, a damage sensitivity identifier is assigned to each spatial feature cluster, using numbers or letters to represent different levels of damage sensitivity. Then, the mean and variance of the eigenvalues ​​of all nodes within the cluster are calculated. For example, the mean eigenvalue is obtained by summing the eigenvalues ​​of all nodes within the cluster and dividing by the number of nodes; the variance is obtained by averaging the sum of the squares of the differences between the eigenvalues ​​and the mean. Simultaneously, the cluster center coordinates and coverage are determined, for example, by calculating the average coordinates of the nodes within the cluster to obtain the cluster center coordinates, and by statistically analyzing the distribution range of the nodes within the cluster to obtain the coverage. This information is combined to generate the spatial distribution descriptor.

[0129] Step S426: Render the spatial distribution descriptor according to the coordinate position of the three-dimensional component connection map to generate a spatial distribution feature map with color gradient representing the damage sensitivity intensity and contour lines representing the feature cluster boundaries. The resolution of the feature map is consistent with the mesh density of the bridge structure finite element model.

[0130] Rendering the spatial distribution descriptor according to the coordinates of the 3D component connection map means displaying the information in the spatial distribution descriptor according to the coordinates of the 3D component connection map. Damage sensitivity intensity is represented by color gradients, with different colors representing different levels of damage sensitivity; darker colors indicate higher damage sensitivity. Feature cluster boundaries are represented by contour lines, outlining the boundaries of spatial feature clusters. The resolution of the feature map is consistent with the mesh density of the bridge structure finite element model, indicating that the accuracy of the feature map is the same as that of the bridge structure finite element model. For example, the information in the spatial distribution descriptor is rendered based on the coordinates of the 3D component connection map. For example, in 3D space, the position of the spatial feature cluster is determined based on the cluster center coordinates and coverage area. Then, based on the damage sensitivity identifier and the feature mean, damage sensitivity intensity is represented by color gradients. For example, red represents high damage sensitivity intensity, and blue represents low damage sensitivity intensity. Simultaneously, contour lines are used to represent the boundaries of the feature clusters. Finally, the resolution of the feature map is adjusted to be consistent with the mesh density of the bridge structure finite element model. For example, based on the mesh size of the finite element model, the pixel size of the feature map is determined, and a spatially distributed feature map is generated.

[0131] Step S430: Model the spatial distribution feature map in the time dimension through the time series prediction layer, analyze the evolution law of damage sensitivity features over time, and construct a time series prediction model for damage development.

[0132] The time series prediction layer is a layer in the physical field linkage sensing network used to model and analyze the spatial distribution feature map over time. The evolution of damage-sensitive features over time represents their changes over time. The time series prediction model for damage development is used to predict the trend of damage development. For example, the spatial distribution feature map is input into the time series prediction layer. The time series prediction layer can use models such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to model the spatial distribution feature map over time. For example, the spatial distribution feature maps at different time points are used as input sequences and input into the LSTM network. By learning from the input sequences, the evolution of damage-sensitive features over time is analyzed. For example, the changing trends and periodicity of damage-sensitive features are observed. Then, based on these evolutionary patterns, a time series prediction model for damage development is constructed. For example, by training an LSTM network, a model capable of predicting future damage development trends is obtained.

[0133] Step S440: Use a time series prediction model to perform multi-step prediction of the spatial distribution feature map within a preset time period in the future, and generate a damage sensitivity feature prediction matrix for each time node in the future.

[0134] A future preset time period is a pre-defined time frame, such as the next month or year. Multi-step prediction involves making multiple predictions over a time series to forecast the situation at multiple future time points. The damage-sensitive feature prediction matrix is ​​a matrix containing the predicted damage-sensitive feature values ​​for each future time point. For example, using a pre-built time series prediction model for damage development, multi-step predictions can be made on the spatial distribution feature map within the future preset time period. For instance, using the current spatial distribution feature map as input, the time series prediction model can predict the spatial distribution feature map for each day of the next month. The predicted damage-sensitive features for each time point are then organized into a matrix to generate the damage-sensitive feature prediction matrix.

[0135] Step S450: Convert the damage-sensitive feature prediction matrix into the probability of damage occurrence, and map the matrix element values ​​to the probability of damage occurrence through a probability transformation function to obtain a preliminary probability distribution matrix.

[0136] A probability transformation function is used to convert damage-sensitive feature values ​​into damage occurrence probabilities. Damage occurrence probability represents the likelihood of a bridge structure being damaged at different locations and times. The preliminary probability distribution matrix is ​​the matrix obtained after converting the damage-sensitive feature prediction matrix into damage occurrence probabilities. For example, by using a suitable probability transformation function, such as a logistic regression function or a Gaussian distribution function, the element values ​​of the damage-sensitive feature prediction matrix are mapped to the probability of damage occurring. For instance, using the logistic regression function P = 1 / (1+e^(-1 / 2))... -z (where z is the damage-sensitive feature value and P is the probability of damage occurrence), substituting each element value in the matrix into the function yields the corresponding probability of damage occurrence. Arranging these probabilities according to the spatial position and time series order of the original matrix yields a preliminary probability distribution matrix. For example, converting each element in the damage-sensitive feature prediction matrix into a probability value using a logistic regression function creates a new matrix, which is the preliminary probability distribution matrix.

[0137] In some embodiments, step S450 may include the following steps S451 to S456:

[0138] Step S451: Construct a damage probability mapping dictionary. The dictionary contains the correspondence between damage-sensitive feature value intervals and damage level probability vectors. It is generated by training feature value-damage level sample pairs in historical monitoring data.

[0139] The damage probability mapping dictionary is a dictionary that stores the correspondence between damage-sensitive feature value intervals and damage level probability vectors. The damage level probability vector is a vector containing the probability of different damage levels occurring, such as the probability values ​​for four levels: no damage, minor damage, moderate damage, and severe damage. It is generated through training on feature value-damage level sample pairs from historical monitoring data, indicating that the damage probability mapping dictionary is obtained by analyzing and learning the damage-sensitive feature values ​​and corresponding damage levels in historical monitoring data. For example, damage-sensitive feature values ​​and corresponding damage level information are collected from historical monitoring data, and the damage-sensitive feature values ​​are divided into different intervals. For example, the damage-sensitive feature values ​​are divided into three intervals: low, medium, and high. Then, the probability of different damage levels occurring within each interval is calculated to form a damage level probability vector. The correspondence between damage-sensitive feature value intervals and damage level probability vectors is stored in the dictionary, thus constructing the damage probability mapping dictionary.

[0140] Step S452: Traverse each element of the damage-sensitive feature prediction matrix, query the damage probability mapping dictionary based on the element value, and obtain the corresponding damage level probability vector. The probability vector contains the probability values ​​of four levels: no damage, minor damage, moderate damage, and severe damage.

[0141] Traversing each element of the damage-sensitive feature prediction matrix involves processing each element sequentially. Looking up the damage probability mapping dictionary based on the element value involves searching the dictionary for the damage level probability vector corresponding to the interval containing the current element value. The probability vector contains probability values ​​for four levels: no damage, minor damage, moderate damage, and severe damage, representing the likelihood of different damage levels occurring. For example, for each element in the damage-sensitive feature prediction matrix, its damage-sensitive feature value interval is determined. Specifically, it's determined whether the element value falls within the low, medium, or high interval. Then, the damage level probability vector corresponding to that interval is searched in the damage probability mapping dictionary.

[0142] Step S453: Arrange the obtained damage level probability vectors according to the spatial position and time series order of the original matrix to construct an initial probability matrix. The row dimension of the matrix corresponds to the node index of the three-dimensional component connection map, and the column dimension corresponds to the prediction timestamp.

[0143] The retrieved damage level probability vectors are arranged according to the spatial location and temporal sequence of the original matrix. This means arranging the retrieved probability vectors according to the spatial location and temporal sequence of the damage-sensitive feature prediction matrix. The initial probability matrix is ​​the matrix obtained after arranging the damage level probability vectors. The row dimension of the matrix corresponds to the node index of the 3D component connection map, indicating that each row of the matrix corresponds to a component node in the bridge structure; the column dimension corresponds to the prediction timestamp, indicating that each column of the matrix corresponds to a prediction time point. For example, the retrieved damage level probability vectors are arranged according to the spatial location and temporal sequence of the damage-sensitive feature prediction matrix. For example, the damage level probability vector with the first spatial location at the first time point is placed in the first row and first column of the matrix. After arranging all the probability vectors, the initial probability matrix is ​​constructed.

[0144] Step S454: Introduce the material fatigue accumulation coefficient to adjust the initial probability matrix, calculate the cumulative fatigue damage value under different stress levels based on the SN curve of the component material, and sum the cumulative value with the severe damage probability in the probability vector.

[0145] The material fatigue accumulation factor is a coefficient that considers the fatigue effect of materials. The S-N curve of a component material describes the fatigue life of the material under alternating stress, where S represents the stress level and N represents the fatigue life. The cumulative fatigue damage value is the degree of cumulative fatigue damage of the material under different stress levels. The cumulative value is weighted and summed with the probability of severe damage in the probability vector; that is, the cumulative fatigue damage value and the probability of severe damage are added according to certain weights. For example, based on the S-N curve of the component material, the cumulative fatigue damage value under different stress levels can be calculated. For example, the cumulative fatigue damage value of the material over a period of time can be calculated by using the relationship between stress level and fatigue life. Then, the material fatigue accumulation factor is introduced to adjust the initial probability matrix. For example, the cumulative fatigue damage value is weighted and summed with the probability of severe damage in the probability vector to adjust the value of the probability of severe damage.

[0146] Step S455: Calculate the entropy value of each probability vector, and perform neighborhood probability fusion on elements whose entropy value exceeds a preset threshold. The fusion weight is dynamically determined based on the component connection strength and time decay factor to reduce prediction uncertainty.

[0147] The entropy of a probability vector is an indicator used to measure its uncertainty; a higher entropy indicates greater uncertainty. A preset threshold is a pre-defined threshold used to filter elements. Neighborhood probability fusion combines the probabilities of elements whose entropy exceeds the preset threshold with those of their neighbors. The fusion weights are dynamically determined based on component connection strength and time decay factors, meaning the fusion weights adjust according to the tightness of the connections between components and the time elapsed. Reducing prediction uncertainty is achieved through neighborhood probability fusion. For example, for each probability vector in the initial probability matrix, its entropy is calculated. The entropy is compared to a preset threshold, and neighborhood probability fusion is performed on elements whose entropy exceeds the threshold. For example, neighbors are determined based on a 3D component connection map, and fusion weights are determined based on component connection strength and time decay factors. For example, a stronger component connection results in a larger fusion weight, and a more distant time elapsed results in a smaller fusion weight. The probabilities of these elements are then weighted and averaged to reduce prediction uncertainty. For example, if the entropy of an element's probability vector exceeds the preset threshold, it is weighted and averaged with the probability vectors of its neighbors according to the fusion weights to obtain a new probability vector.

[0148] Step S456: Perform spatial consistency verification on the fused probability matrix. Use a 3D component connection map to check whether the probability gradients of adjacent nodes conform to the structural damage propagation law. Perform probability smoothing on gradient anomaly regions to generate a preliminary probability distribution matrix. Spatial consistency verification checks whether the fused probability matrix spatially conforms to the structural damage propagation law. The probability gradient of adjacent nodes represents the degree of change between the damage occurrence probabilities of adjacent component nodes. The structural damage propagation law describes the pattern of damage propagating from one component to adjacent components in a bridge structure; typically, damage propagation exhibits a certain degree of continuity and gradualness. The 3D component connection map clarifies the relationships between adjacent nodes, thereby checking whether the probability gradients of adjacent nodes are reasonable.

[0149] For example, based on the probability matrix obtained by traversing and fusing the 3D component connection map, for each node, the probability gradient between it and its neighboring nodes is checked. For instance, the difference in damage probability between neighboring nodes is calculated, and it is determined whether this difference is within a reasonable range. If the probability gradient of neighboring nodes does not conform to the structural damage propagation law, i.e., gradient anomalies occur (e.g., one neighboring node shows a high damage probability while the other shows an extremely low damage probability, which does not conform to the actual damage propagation logic), probability smoothing processing is required for that region.

[0150] Probability smoothing can be achieved using methods such as simple averaging and weighted averaging. Taking weighted averaging as an example, weights are determined based on the spatial distance and connection strength between adjacent nodes and the anomalous node. The probability values ​​of the anomalous node and its adjacent nodes are then weighted and averaged, resulting in a smoother and more realistic change in probability values. For instance, adjacent nodes that are spatially close to the anomalous node and have a strong connection have higher weights.

[0151] After probabilistic smoothing of all gradient anomaly regions, a preliminary probability distribution matrix is ​​generated. This matrix spatially conforms more closely to the structural damage propagation pattern and can more accurately reflect the probability distribution of damage occurrence in key bridge components at different monitoring periods.

[0152] Step S460: Perform spatial location calibration and time series smoothing on the preliminary probability distribution matrix, and generate a damage development probability distribution by combining the physical location information of key bridge components.

[0153] Spatial location calibration involves precisely mapping and adjusting the probability values ​​in the preliminary probability distribution matrix to the actual physical locations of key bridge components. Since previous processing may introduce errors or inaccuracies in coordinate mapping, spatial location calibration ensures that the probability values ​​accurately reflect the damage probability of the corresponding components. For example, based on the bridge's BIM model or precise structural design drawings, each element in the preliminary probability distribution matrix is ​​accurately matched to a specific component, adjusting for any potential positional deviations.

[0154] Time series smoothing involves processing the initial probability distribution matrix along the time dimension to eliminate potential noise and fluctuations, making the change in damage probability over time smoother and more reasonable. Methods such as moving averages and exponential smoothing can be used for time series smoothing. For example, the moving average method calculates the average of the probability values ​​within a certain time window and uses this average to replace the probability value at the center of the window, thereby reducing the impact of short-term fluctuations.

[0155] By incorporating the physical location information of key bridge components, the matrix, after spatial location calibration and time series smoothing, is further refined. For example, detailed information about the components, such as their type, size, and importance, is taken into account, and the probability values ​​are adjusted accordingly. For important critical components, greater attention may be paid to their damage probability, and appropriate corrections may be made.

[0156] The resulting damage development probability distribution includes the distribution of damage locations and quantitative indicators of damage severity across different monitoring periods. The damage location distribution clarifies which critical bridge components are more likely to be damaged at different times, while the damage severity quantitative indicators represent the severity of the damage with specific numerical values ​​or levels. For example, some piers and beams of the bridge may experience minor damage within the next month, while some beams may develop moderate damage within the next three months. This information is clearly presented in the damage development probability distribution, providing a comprehensive and accurate basis for bridge maintenance and management.

[0157] Step S500: Generate dynamic traffic control instructions based on the probability distribution of damage development. The dynamic traffic control instructions include vehicle passage parameters that are dynamically adjusted based on the quantitative indicators of damage degree, and send the dynamic traffic control instructions to the traffic management system.

[0158] The probability distribution of damage development provides a crucial basis for generating dynamic traffic control instructions. This distribution includes information such as the location and degree of damage to key bridge components during different monitoring periods. Dynamic traffic control instructions are dynamically adjusted based on the real-time damage status of the bridge, aiming to ensure the safety and normal use of the bridge while rationally controlling traffic flow.

[0159] In some embodiments, step S500 may include the following steps S510 to S560:

[0160] Step S510: Analyze the probability distribution of damage development, extract the quantitative indicators of the maximum damage degree of each key component of the bridge in the future preset time period and the corresponding spatial coordinates. The spatial coordinates are identified by the component code of the bridge BIM model.

[0161] Analyzing the probability distribution of damage development involves a detailed analysis and interpretation of this distribution to obtain the necessary key information. The future preset time period is a pre-defined time frame, such as the next week or month, during which the bridge's damage status is assessed. The maximum damage quantification index is a quantitative representation of the maximum damage that each key component of the bridge may reach within the preset time period, expressed as a numerical value or a grade indicating the severity of the damage.

[0162] Spatial location coordinates are identified through component codes in the bridge BIM model. The bridge BIM model is a digital 3D model that contains detailed structural information and component codes of the bridge. Each component has a unique code in the BIM model, which allows for accurate location of the component's spatial coordinates.

[0163] For example, by traversing the damage development probability distribution, the maximum damage level quantification index for each key component within a preset future time period is identified. Simultaneously, based on the correspondence between components in the damage development probability distribution, their spatial coordinates are obtained through the component codes in the bridge BIM model. For instance, in the damage development probability distribution of a bridge, if the maximum damage level quantification index for a certain beam component within the next month is moderate damage, its specific spatial coordinates, such as (x, y, z) coordinates in three-dimensional space, are determined through the component's code in the BIM model. This completes the extraction of the maximum damage level quantification index and corresponding spatial coordinates for each key bridge component within a preset future time period, providing an important foundation for subsequently generating vehicle traffic parameters.

[0164] Step S520: Query the preset damage-control mapping matrix based on the maximum damage level quantification index. The row dimension of the mapping matrix is ​​the damage level, the column dimension is the vehicle passage parameter type, and the matrix element value is the baseline parameter value under the corresponding damage level.

[0165] The preset damage-control mapping matrix is ​​a pre-defined matrix used to establish the mapping relationship between damage severity levels and vehicle passage parameters. The row dimension of the matrix represents the damage severity level, which can be divided into different levels such as no damage, minor damage, moderate damage, and severe damage; the column dimension represents the vehicle passage parameter type, including recommended load limits, recommended speed limits, and open status parameters; the matrix element values ​​are the baseline parameter values ​​for the corresponding damage level.

[0166] For example, the maximum damage level quantification index extracted in step S510 is matched with the row dimensions of the damage-control mapping matrix. For instance, if the maximum damage level quantification index of a critical component corresponds to moderate damage, the "moderate damage" row is found in the mapping matrix. Then, based on the required vehicle passage parameter type, the corresponding baseline parameter value is extracted from that row. For instance, if a load limit recommendation value is needed, the column element corresponding to the load limit recommendation value is found in the "moderate damage" row, and the value of that element is used as the load limit baseline parameter value for that component under moderate damage conditions. In this way, the corresponding vehicle passage parameter baseline value is determined for each critical component based on its maximum damage level quantification index, providing a basis for subsequent parameter adjustments.

[0167] Step S530: Based on the spatial topological relationship between the damage location coordinates and the bridge lanes, calculate the spatial influence coefficient of each lane on the damage location. The spatial influence coefficient is calculated by the ratio of the vertical distance from the damage location to the lane centerline to the lane width.

[0168] The damage location coordinates are the spatial location coordinates of the key bridge components extracted in step S510. The spatial topology of the bridge lanes describes the layout and relative positions of the lanes on the bridge. The spatial influence coefficient measures the degree of influence of each lane on the damage location; the larger the coefficient, the greater the influence of that lane on the damage location.

[0169] In the specific calculation, the first step is to determine the vertical distance from the damaged location to the centerline of each lane. This can be done using geometric calculation methods based on the coordinates of the damaged location and the lane centerline. For example, on a two-dimensional plane, the vertical distance from the damaged location to the lane centerline can be calculated using the distance formula between two points. Then, this vertical distance is divided by the lane width to obtain the spatial influence coefficient. By calculating the spatial influence coefficient of each lane on the damaged location, we can understand the degree of influence of different lanes on the damaged area, providing a more accurate basis for subsequent adjustments to vehicle traffic parameters. For example, lanes with a larger spatial influence coefficient have a more significant impact on the damaged location, requiring stricter restrictions when adjusting vehicle traffic parameters.

[0170] Step S540: Adjust the model by combining the baseline parameter values ​​obtained from the mapping matrix query with the spatial influence coefficient input parameters, and calculate the dynamic adjustment coefficient of each lane through a nonlinear coupling formula. The input terms of the coupling formula include the baseline parameter values, the spatial influence coefficient, and the damage development rate.

[0171] The parameter adjustment model is used to adjust vehicle traffic parameters based on bridge damage and lane influencing factors. The nonlinear coupling formula is the core calculation method of this model, which comprehensively considers the complex relationships between multiple factors.

[0172] The baseline parameter values ​​are the dimensionless vehicle passage parameter values ​​obtained from the damage-control mapping matrix in step S520 under the corresponding damage level. The spatial influence coefficient is the influence coefficient of each lane on the damage location calculated in step S530. The damage development rate is the rate of change of the bridge damage degree over time, which can be calculated using the time derivative of the damage degree quantification index.

[0173] For example, the baseline parameter values, spatial influence coefficient, and damage development rate can be substituted into the nonlinear coupling formula as inputs. The nonlinear coupling formula can be expressed as C = f(P, S, R), where C is the dynamic adjustment coefficient, P is the baseline parameter value, S is the spatial influence coefficient, R is the damage development rate, and f is the nonlinear function. This function can be designed according to the actual situation, for example, it can take the form of a polynomial function, exponential function, etc.

[0174] In some embodiments, step S540 may include the following steps S541 to S546:

[0175] Step S541: Construct a lane-component association network, establish a topological connection relationship between the key components of the bridge and each lane, and the weight of each connection edge represents the load transfer efficiency of the lane to the corresponding component. The weight value is obtained through structural mechanics simulation calculation.

[0176] The lane-component relationship network is a network model used to describe the relationships between bridge lanes and key components. By constructing this network, the impact of each lane on different key components can be clearly shown. The topological connections clarify the connection methods and interactions between lanes and components.

[0177] The weight of each connecting edge represents the load transfer efficiency of that lane to the corresponding component, that is, the proportion of load transferred to the corresponding component when a vehicle travels in that lane. The weight values ​​are obtained through structural mechanics simulation calculations, which can simulate the stress conditions of the bridge structure when vehicles travel in the lanes. For example, a mechanical model of the bridge can be built using finite element analysis software to simulate the load distribution when vehicles travel in different lanes, calculate the load transfer ratio of each lane to different components, and use this ratio as the weight of the connecting edge.

[0178] For example, first, the location information of the key components and lanes of the bridge is determined. Then, based on the structural mechanics simulation results, connection edges are established between each lane and the key components, and corresponding weights are assigned. For example, for a certain bridge with n key components and m lanes, a network containing n component nodes and m lane nodes is constructed, with each lane node connected to a component node by a connection edge, and the weights of the edges are calculated based on the simulation.

[0179] Step S542: Calculate the load contribution of each lane based on the lane-component association network. Traverse all lane connection edges in the network that are connected to the damaged component, and sum the products of the edge weights and the current lane traffic flow to obtain the load contribution of each lane to the damaged location.

[0180] Lane load contribution refers to the degree to which each lane contributes to the load on the damaged component. Based on the lane-component relationship network, this contribution is calculated using the weights of the connecting edges and the traffic flow of the lanes.

[0181] For example, first, find all lane connections to the damaged component in the lane-component association network. Then, obtain the current traffic flow information for each lane, which can be obtained in real time through traffic monitoring equipment such as vehicle counters and sensors. For each connection edge, multiply its weight by the traffic flow of the corresponding lane. Finally, sum the products of all connections to the damaged component to obtain the load contribution of that lane to the damaged location.

[0182] Step S543: Construct a coupling matrix between the spatial influence coefficient and the lane load contribution. The row dimension of the matrix is ​​the lane number, the column dimension is the type of influencing factor, and the matrix element value is the normalized spatial influence coefficient or load contribution.

[0183] The coupling matrix is ​​a matrix used to comprehensively consider the spatial influence coefficient and the lane load contribution. By integrating these two factors, the comprehensive impact of each lane on the damage location can be analyzed in subsequent analysis.

[0184] The row dimension of the matrix represents the lane number, meaning each row corresponds to one lane. The column dimension represents the type of influencing factor, which includes two factors: spatial influence coefficient and lane load contribution. The matrix element values ​​are normalized spatial influence coefficients or load contributions. Normalization is used to unify data with different dimensions and ranges to the same scale, facilitating comparison and calculation.

[0185] For example, the spatial influence coefficient calculated in step S530 and the lane load contribution calculated in step S542 are first normalized. The normalization method can be the min-max normalization method, and a similar normalization method is used for the lane load contribution L_i. Then, the normalized spatial influence coefficient and lane load contribution are filled into the coupling matrix. For example, the element in the first row and first column of the matrix is ​​the normalized spatial influence coefficient of the first lane, and the element in the first row and second column is the normalized lane load contribution of the first lane.

[0186] Step S544: Perform eigenvalue decomposition on the coupling matrix and extract the eigenvector corresponding to the largest eigenvalue as the main influence factor. The main influence factor reflects the combined effect intensity of spatial location and load contribution.

[0187] Eigenvalue decomposition is a matrix analysis method used to decompose a matrix into eigenvalues ​​and eigenvectors. For a coupled matrix, eigenvalue decomposition yields its eigenvalues ​​and corresponding eigenvectors.

[0188] The main influencing factor is an important factor extracted from the coupling matrix, reflecting the combined effect of spatial location and load contribution. For example, eigenvalue decomposition is performed on the coupling matrix A, i.e., solving the equation Av = λv, where λ is the eigenvalue and v is the eigenvector. Numerical calculation methods, such as the Jacobi method and the QR algorithm, are used to obtain all eigenvalues ​​and eigenvectors of the coupling matrix.

[0189] Then, find the largest eigenvalue λ. max And extract its corresponding feature vector v max As the main influencing factor, each element of the main influencing factor corresponds to a lane, and the value of the element represents the combined effect intensity of that lane in terms of spatial location and load contribution.

[0190] Step S545: Nonlinearly fuse the main influencing factor and the damage development rate, and map the fusion result to the base value of the dynamic adjustment coefficient through the hyperbolic tangent function. The damage development rate is calculated by the time derivative of the damage degree quantification index.

[0191] Nonlinear fusion integrates the main influencing factors and the damage development rate, considering their nonlinear relationship. The damage development rate is calculated using the time derivative of the damage severity quantification index, as mentioned earlier, by calculating the rate of change of the damage severity quantification index over a period of time. For example, the main influencing factors and the damage development rate are first fused, perhaps using a weighted summation method. Then, the fusion result is substituted into the hyperbolic tangent function to obtain the base value of the dynamic adjustment coefficient. Through the mapping of the hyperbolic tangent function, the fusion result can be transformed into a suitable range, facilitating subsequent adjustments to vehicle traffic parameters.

[0192] Step S546: Introduce a traffic flow fluctuation coefficient to correct the base value. The traffic flow fluctuation coefficient is obtained by analyzing the ratio of the standard deviation to the mean of traffic flow in historical traffic data. The larger the coefficient, the more unstable the traffic flow, and the greater the correction range of the adjustment coefficient.

[0193] The traffic flow fluctuation coefficient is a measure of traffic flow stability, reflecting the volatility of traffic volume. By introducing this coefficient, the baseline value of the dynamic adjustment coefficient can be corrected, making the adjustment coefficient more consistent with actual traffic conditions. The traffic flow fluctuation coefficient is derived by analyzing the ratio of the standard deviation to the mean of traffic flow in historical traffic data. A larger coefficient indicates greater traffic flow volatility and less stable traffic flow.

[0194] For the base value of the dynamic adjustment coefficient A traffic flow fluctuation coefficient is introduced for correction. For example, the correction formula can be: Here, α is a correction coefficient, set according to the actual situation. When the traffic flow fluctuation coefficient K is large, the correction range is also large, meaning the change in the adjustment coefficient is more significant. For example, in cases of unstable traffic flow, to better protect the bridge, the recommended load limit or speed limit may be further reduced. Through this correction, the dynamic adjustment coefficient can more accurately reflect the actual situation of the bridge and traffic conditions, providing a more reliable basis for generating reasonable vehicle traffic parameters.

[0195] Step S550: Correct the baseline parameter values ​​based on the dynamic adjustment coefficient to generate suggested load limits, suggested speed limits, and open status parameters for each lane.

[0196] The dynamic adjustment coefficient is a coefficient reflecting the comprehensive influence of each lane, which is obtained through the previous steps. The benchmark parameter value is the vehicle passing parameter value corresponding to the damage level queried from the damage-control mapping matrix. By correcting the benchmark parameter value with the dynamic adjustment coefficient, a vehicle passing parameter more in line with the actual situation can be obtained.

[0197] For the recommended load limit value, assume that the benchmark recommended load limit value is If the dynamic adjustment coefficient is C, the corrected recommended load limit value For example, if the dynamic adjustment coefficient is less than 1, it means that the recommended load limit value needs to be reduced to reduce the load of vehicles on the bridge; if the dynamic adjustment coefficient is greater than 1, the recommended load limit value may be appropriately increased according to the actual situation, but within a safe range.

[0198] For the recommended speed limit value, similarly assume that the benchmark recommended speed limit value is The corrected recommended speed limit value By adjusting the recommended speed limit value, the impact force of vehicles on the bridge can be controlled and the damage risk can be reduced.

[0199] The open state parameter is determined according to the dynamic adjustment coefficient and the actual situation. When the dynamic adjustment coefficient is too low, it indicates that the lane has a greater impact on the damaged position of the bridge, and the lane may be closed; when the dynamic adjustment coefficient is within a certain range, the lane remains open, but traffic management may be carried out according to the recommended load limit and speed limit values. For example, set a threshold T. When C < T, the lane is closed; when C ≥ T, the lane is open and managed according to the corrected recommended load limit and speed limit values.

[0200] By correcting and calculating the benchmark parameter values of each lane, the recommended load limit value, recommended speed limit value and open state parameter of each lane are finally generated, providing specific parameter information for generating dynamic traffic control instructions in the follow-up.

[0201] Step S560: Package the recommended load limit value, recommended speed limit value and open state parameter of each lane into a dynamic traffic control instruction according to a preset data protocol. The instruction contains the effective timestamp and update period of each parameter.

[0202] The preset data protocol is a pre-specified protocol for packaging and transmitting dynamic traffic control instructions, ensuring that the instructions can be accurately received and parsed by the traffic management system. The protocol stipulates the format, coding method, data structure, etc. of the instructions.

[0203] The recommended load limits, speed limits, and lane opening status parameters for each lane are the specific parameter information calculated in the previous steps. These parameters are then encapsulated according to a preset data protocol. First, the basic structure of the instruction is determined; for example, fixed-length fields can be used to represent different parameters. For instance, one field can represent the recommended load limit, one field the recommended speed limit, and one field the lane opening status parameter.

[0204] The instruction includes the effective timestamp and update cycle for each parameter. The effective timestamp specifies the exact time when each parameter in the instruction takes effect, such as down to the specific date and time. The update cycle specifies the time interval for updating the instruction, such as updating the instruction once per hour, to ensure that traffic control measures can be adjusted promptly based on the real-time damage to the bridge and traffic conditions. For example, the parameter information, effective timestamps, and update cycles for each lane are encoded and organized according to the requirements of a preset data protocol. For example, recommended load limits and speed limits are converted into specific encoding formats, and open status parameters are represented using binary code. Then, these encoded data are combined into a complete instruction package. Finally, the packaged dynamic traffic control instruction is sent to the traffic management system, which adjusts the relevant traffic control equipment according to the instruction content to achieve dynamic control of bridge traffic.

[0205] The general algorithms listed in the above descriptions of the embodiments of the present invention can be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. When implementing the embodiments of the present invention, omissions can be made based on common knowledge in the art. For example, based on common knowledge in the art, the dimensional conflicts before feature fusion can be eliminated by normalization; interpolation can be used to eliminate dimensional differences; thresholds can be reasonably set based on historical data, experience, or business scenario requirements; the model can be trained based on general model training methods; the number of layers in the model structure and the selection of activation functions can be set according to actual needs, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes.

[0206] According to an embodiment of the present invention, a computer system 120 is also provided.

[0207] Please refer to Figure 3The diagram below shows the structural block diagram of the computer system 120 of the present invention. The computer system 120 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in a ROM 1002 (read-only memory) or a computer program loaded from a storage unit 1008 into a RAM 1003 (random access memory). The RAM 1003 may also store various programs and data required for the operation of the computer system 120. The computing unit 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An I / O interface 1005 (input / output) is also connected to the bus 1004.

[0208] Multiple components in the computer system 120 are connected to the I / O interface 1005, including: an input unit 1006, an output unit 1007, a storage unit 1008, and a communication unit 1009. The input unit 1006 can be any type of device capable of inputting information into the computer system 120. The input unit 1006 can receive input numeric or character information and generate key signal inputs related to user settings and / or function control of the server. The output unit 1007 can be any type of device capable of presenting information. The communication unit 1009 allows the computer system 120 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0209] The computing unit 1001 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 performs the various methods and processes described above, such as a machine learning-based bridge health monitoring and intelligent control method. For example, in some embodiments, the machine learning-based bridge health monitoring and intelligent control method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1008.

[0210] While embodiments or examples of the invention have been described with reference to the accompanying drawings, it should be understood that the methods, systems, and devices described above are merely exemplary embodiments or examples, and the scope of the invention is not limited by these embodiments or examples, but only by the granted claims and their equivalents. Various elements in the embodiments or examples may be omitted or replaced by their equivalents. Furthermore, the steps may be performed in a different order than that described in the invention. Further, various elements in the embodiments or examples may be combined in various ways. Importantly, as the technology evolves, many elements described herein can be replaced by equivalents that appear later in the invention.

Claims

1. A machine learning based method for bridge health monitoring and intelligent control, characterized in that, The method comprises: Distributed sensing data acquisition is performed on the bridge structure to generate a structure monitoring data set containing a multi-channel time sequence; Temporal and spatial feature extraction is performed on the structure monitoring data set to obtain a vibration response feature sequence and a traffic load feature sequence of the bridge structure; Correlation analysis is performed on the vibration response feature sequence and the traffic load feature sequence to generate a structure damage sensitive feature matrix, which is used to quantify the abnormal fluctuation degree of the structure response under different load conditions; Damage evolution trend prediction is performed based on the structure damage sensitive feature matrix to generate a damage development probability distribution of the key components of the bridge, which contains a damage location distribution and a damage degree quantification index of different monitoring periods; Dynamic traffic control instructions are generated according to the damage development probability distribution, which contain vehicle passing parameters dynamically adjusted based on the damage degree quantification index, and the dynamic traffic control instructions are sent to a traffic management system.

2. The method of claim 1, wherein, The temporal and spatial feature extraction on the structure monitoring data set to obtain the vibration response feature sequence and the traffic load feature sequence of the bridge structure comprises: The structure monitoring data set is divided into three-dimensional data blocks according to the monitoring location and the time stamp to divide block data sequences containing spatial coordinate markers and time sequence markers, the spatial dimension of the block data sequences corresponds to the deployment density of the sensing device, and the time dimension corresponds to the sampling interval of the preset monitoring period; Environmental noise suppression is performed on the block data sequences, interference signal components unrelated to the structure dynamic response are identified by analyzing the background environmental physical field characteristics, and the interference signal components are filtered out; Vibration signal sources are separated from the block data sequences, vibration signal components generated by structure elastic deformation are distinguished according to the physical field propagation characteristics of the components of the bridge structure, and the frequency variation trend of the vibration signal components is negatively correlated with the structure stiffness change; Multi-scale dynamic feature construction is performed on the vibration signal components to generate a vibration response feature sequence by analyzing the energy transfer law and time domain waveform distortion characteristics of signals in different frequency bands; Load action signals are extracted from the block data sequences, stress and strain signal components caused by vehicle loads are separated according to the spatio-temporal propagation law of physical field disturbance during vehicle passing, and the amplitude variation of the stress and strain signal components is positively correlated with the vehicle weight and driving speed; Temporal and spatial distribution characteristics of the stress and strain signal components are quantified to generate a traffic load feature sequence by analyzing the propagation time delay and amplitude attenuation law of signals at different spatial positions.

3. The method of claim 2, wherein, The multi-scale dynamic feature construction on the vibration signal components to generate a vibration response feature sequence by analyzing the energy transfer law and time domain waveform distortion characteristics of signals in different frequency bands comprises: The frequency band division interval of the vibration signal components is determined according to the theoretical frequency range of each key component of the bridge structure, the vibration signal components are decomposed into multi-layer sub-signals covering different frequency ranges, and the frequency boundaries of each layer of sub-signals correspond to the structure modal frequency intervals; Energy distribution calculation is performed on each layer of sub-signals, the signal energy proportion in each frequency interval and the time domain peak value occurrence frequency are counted, and an energy feature vector is generated; Waveform morphology analysis is performed on each layer of sub-signals, the waveform slope change rate, peak symmetry and half-wave width parameters of the signals in a sliding time window are calculated, and a time domain morphology feature vector is generated; The energy feature vector and the time domain morphology feature vector are dimensionally spliced in time sequence order to obtain a joint feature vector, and the joint feature vector is time axis calibrated to keep the feature vectors of sub-signals of different frequency bands synchronized in time marking; The calibrated joint feature vector is arranged in frequency band order to generate a vibration response feature sequence, and each time node of the vibration response feature sequence contains the energy and morphology feature combination of each frequency band at the corresponding time.

4. The method of claim 2, wherein, The space-time distribution feature quantization of the stress-strain signal components generates a traffic load feature sequence by analyzing the propagation time delay and amplitude attenuation law of signals at different spatial positions, including: The correlation analysis of the stress-strain signal components is performed on the sensing channel, the amplitude change synchronization and phase difference of signals at different sensing positions are calculated, the load propagation path directed graph is constructed, the nodes of the directed graph represent the sensing channel positions, and the edge weight represents the signal propagation strength; The dominant load propagation direction is determined according to the load propagation path directed graph, the signal arrival time difference along the propagation direction is extracted, and the load propagation speed parameter is calculated by the ratio of the time difference to the sensor spacing; The load event segmentation of the stress-strain signal components is performed, the adaptive threshold method is used to identify the signal pulse events caused by single vehicle load, and the start time, peak time and end time of each pulse event are recorded; The occurrence frequency and peak value amplitude distribution of the pulse events in unit time are counted to generate a load time interval statistical feature vector; According to the deployment position of the sensing device on the key components of the bridge, a corresponding spatial coordinate index is assigned to each stress-strain signal component, and the spatial coordinate index corresponds to the component number of the bridge structure; The load propagation speed parameter, load time interval statistical feature vector and spatial coordinate index are combined in time sequence order to generate a traffic load feature sequence.

5. The method of claim 1, wherein, The correlation analysis of the vibration response feature sequence and the traffic load feature sequence generates a structure damage sensitive feature matrix, including: The vibration response feature sequence and the traffic load feature sequence are aligned according to the time stamp to generate a time-synchronized joint feature sequence; Cross-correlation modeling is performed on the joint feature sequence to analyze the mutual dependence of vibration response features and traffic load features at different time scales, and to identify the feature combination mode of load change triggering vibration response anomaly; Through the correlation strength quantification method, the sensitivity of different feature combination modes to structure damage is calculated, and the feature combination modes meeting the preset correlation threshold condition are selected as strong correlation feature combinations, and the preset correlation threshold is determined by the feature correlation strength distribution in the historical damage data. The strong correlation feature combination is subjected to dynamic evolution analysis, and the feature value change trend thereof in different monitoring periods is tracked, and abnormal fluctuation features of feature values deviating from the healthy benchmark state are extracted; The abnormal fluctuation features are arranged in a matrix according to spatial positions and time sequences, an initial feature matrix with spatial positions as rows and time sequences as columns is constructed, and the matrix element values reflect the abnormal fluctuation degrees of the corresponding space-time points; The initial feature matrix is subjected to space-time correlation enhancement, and a structural damage sensitive feature matrix is generated by analyzing the feature value transmission law of adjacent spatial positions and continuous time sequences.

6. The method of claim 5, wherein, The joint feature sequence is subjected to cross-correlation modeling, the mutual dependence of the vibration response features and the traffic load features in different time scales is analyzed, the feature combination mode of the load change causing the vibration response abnormality is identified, including: The joint feature sequence is divided into analysis windows of different time scales, and the time length of each window is set according to the bridge structure damage evolution period; For each feature subsequence of a time scale, the mutual information value of the vibration response features and the traffic load features is calculated, and the statistical dependence strength of the two features in the time scale is quantified; An association strength matrix is constructed based on the mutual information value, the row dimension of the matrix corresponds to the vibration response feature dimension, the column dimension corresponds to the traffic load feature dimension, and the element value represents the association strength of the corresponding feature pair; The change law of the association strength matrix in different time scales is analyzed, the feature pairs whose association strength increases monotonously with the increase of the time scale and whose change rate exceeds a preset change rate threshold at a time scale exceeding a preset proportion threshold are identified, and the preset proportion threshold is determined by statistical analysis of the change law of the feature association strength in historical monitoring data; The identified feature pairs are subjected to dynamic response curve comparison, the waveform difference of the vibration response features before and after the load feature change is analyzed, and the feature combination mode of the load change causing the vibration response abnormality is determined; Through repetitive verification of the feature combination mode, the feature combination mode that stably appears in multiple monitoring periods is screened out.

7. The method of claim 5, wherein, The strong correlation feature combination is subjected to dynamic evolution analysis, and the feature value change trend thereof in different monitoring periods is tracked, and abnormal fluctuation features of feature values deviating from the healthy benchmark state are extracted, including: Based on the historical monitoring data of the bridge in the healthy state, a benchmark feature library of the strong correlation feature combination is constructed, and the benchmark feature library includes the normal fluctuation range of each feature combination under different load conditions; The strong correlation feature combination of the current monitoring period is compared with the benchmark feature library, and the deviation degree of the current feature value of each feature combination from the benchmark feature value is calculated; The change trend of the deviation degree in continuous monitoring periods is tracked, the first derivative of the deviation degree time sequence is calculated, the feature combinations whose derivative is positive for continuous multiple periods and whose derivative absolute value exceeds a preset change rate threshold are identified, and the preset change rate threshold is determined based on the historical fluctuation derivative of the deviation degree in the healthy benchmark feature library; The mutation point detection is performed on the deviation degree time sequence, a difference between the current period deviation degree and the previous period deviation degree is calculated, and when an absolute value of the difference exceeds an upper limit value of a normal fluctuation range in the health benchmark feature library, it is determined that the deviation degree mutation feature combination is determined; The fluctuation pattern analysis is performed on the continuously increasing feature combination and the mutation feature combination, a peak value offset, a waveform distortion rate and a period change rate in the feature value time sequence are extracted, and an abnormal fluctuation pattern parameter is generated; In combination with spatial position information corresponding to the feature combination, the abnormal fluctuation pattern parameter is associated with a physical position of the bridge structure component, a specific structure region corresponding to the abnormal fluctuation is determined, the abnormal fluctuation pattern parameter is converted into a numerical feature representing a damage sensitivity degree, and an abnormal fluctuation feature is generated.

8. The method of claim 5, wherein, The spatiotemporal correlation enhancement is performed on the initial feature matrix, a structure damage sensitive feature matrix is generated by analyzing feature value transmission rules of adjacent spatial positions and continuous time sequences, and the structure damage sensitive feature matrix includes: A physical connectivity topology graph of the key component is constructed, the physical connectivity topology graph includes connection relationships and transmission path parameters between components, and a mapping relationship between a row dimension of the initial feature matrix and nodes of the topology graph is established; Each spatial position element in the initial feature matrix is traversed, feature value sequences of adjacent spatial positions of the element are extracted according to the connection relationship of the topology graph, an attenuation coefficient and a delay parameter of the feature value on a spatial transmission path are calculated, and a spatial transmission function is constructed; The spatial transmission function is applied to the initial feature matrix, neighborhood feature fusion is performed on each element value, a fusion weight is dynamically determined according to the attenuation coefficient of the spatial transmission path, and a spatial correlation enhanced matrix is generated; The spatial correlation enhanced matrix is segmented in a time sequence, continuous time windows are divided according to bridge monitoring periods, a change rate and an accumulated increment of the feature value in each window are calculated, and a time evolution rule library is established; The spatial correlation enhanced matrix is adjusted in a time dimension based on the time evolution rule library, a trend of a feature value sequence that meets a monotonic change rule is strengthened, and a phase is synchronized and calibrated for a sequence that meets a periodicity rule, and a spatiotemporal coupling feature matrix is generated; By analyzing the correlation between element values in the spatiotemporal coupling feature matrix and historical damage events, a damage indication contribution degree of each element is calculated, elements with a contribution degree higher than a set threshold are retained, and a structure damage sensitive feature matrix is generated.

9. The method of claim 1, wherein, The damage evolution trend prediction is performed based on the structure damage sensitive feature matrix, and a damage development probability distribution of the bridge key component is generated, including: The structure damage sensitive feature matrix is input into a physical field linkage perception network, the physical field linkage perception network includes a spatial feature extraction layer, a time sequence prediction layer and a probability output layer; The spatial feature extraction layer is used for performing spatial pattern recognition on the structure damage sensitive feature matrix, a spatial distribution pattern of the damage sensitive feature in different regions of the bridge structure is extracted, and a spatial distribution feature map is generated; The time sequence prediction layer is used for performing time dimension modeling on the spatial distribution feature map, an evolution law of the damage sensitive feature with time is analyzed, and a time sequence prediction model of damage development is constructed; and The probability output layer is used for performing probability distribution prediction on the time sequence prediction model of damage development, and the damage development probability distribution of the bridge key component is generated. The time series prediction model is used to perform multi-step prediction on the spatial distribution feature map in a future preset time period, to generate a damage sensitive feature prediction matrix of each time node in the future; The damage sensitive feature prediction matrix is converted into a damage occurrence probability, and the matrix element value is mapped into the possibility of damage occurrence through a probability conversion function, to obtain a preliminary probability distribution matrix; The preliminary probability distribution matrix is subjected to spatial position calibration and time series smoothing, and the physical position information of the key components of the bridge is combined, to generate a damage development probability distribution.

10. A computer system, characterized by Comprise: a processor; and a memory connected in communication with the processor; wherein the memory stores instructions executable by the processor, and the instructions are executed by the processor to enable the processor to perform the method of any one of claims 1-9.

Images