Road disaster detection method and system based on air-ground cooperative unmanned inspection

Abstract

The invention provides a road disaster detection method and system based on air-ground cooperative unmanned inspection, and the method comprises the steps: obtaining a multi-source road monitoring data set, carrying out the time-space alignment processing of the multi-source road monitoring data set, generating a road monitoring data set integrated with a time-space label, and carrying out the recognition of a road disaster based on a preset multi-mode disaster recognition model. Disaster feature extraction processing is carried out on the road monitoring data set fused with the space-time labels to obtain a road disaster feature set, a disaster detection result set is generated according to the road disaster feature set, the disaster detection result set comprises a plurality of disaster detection units, and an inspection optimization strategy set is generated based on the disaster detection result set. According to the method, the automation degree and the response efficiency of road disaster detection are comprehensively improved under the condition of not depending on artificial experience intervention, and the problems of high misjudgment rate, large monitoring blind area and rigid resource allocation caused by data splitting and time-space asynchronization in a traditional method are solved.

Description

Technical Field

[0001] The present invention relates to the fields of data processing and highway detection, and in particular, to a highway disaster detection method and system based on collaborative unmanned aerial and ground patrols. Background Art

[0002] With the continuous expansion of the scale of highway traffic infrastructure, highway disaster detection technology has gradually become a core requirement in the field of road operation and maintenance. Highway disaster detection refers to the technology of identifying disaster risks such as landslides, pavement cracks, and bridge structure damage along highways through monitoring means. The current mainstream methods mainly adopt data acquisition modes of aerial remote sensing image analysis or ground sensor monitoring, and respectively evaluate environmental disaster or structure damage risks by independently processing the surface deformation characteristics in optical images or the abnormal signals in structural vibration sensing data. Due to the spatio-temporal benchmark fragmentation of air-ground data and the lack of association between environmental and structural characteristics, such methods make it difficult to capture the disaster coupling evolution mechanism. For example, the synergistic effect between slow surface deformation and abnormal bridge stress cannot be effectively identified, which is likely to cause misjudgment of disaster types or underestimation of the affected range. At the same time, the static patrol path and fixed parameter configuration are difficult to adapt to the dynamic diffusion characteristics of disasters, resulting in problems such as insufficient coverage of monitoring blind spots and lagging response in high-risk areas, which limit the accuracy and timeliness of highway disaster detection. Summary of the Invention

[0003] The present invention provides a highway disaster detection method and system based on collaborative unmanned aerial and ground patrols.

[0004] In a first aspect, an embodiment of the present invention provides a method for highway disaster detection based on air-ground collaborative unmanned inspection, including: obtaining a multi-source highway monitoring data set, where the multi-source highway monitoring data set includes a multi-dimensional remote sensing image data set collected by an aerial inspection device and a multi-dimensional structural sensing data set collected by a ground inspection device; the multi-dimensional remote sensing image data set contains image area coverage information with different timestamps, and the multi-dimensional structural sensing data set contains structural vibration response information with different spatial positions; performing spatio-temporal alignment processing on the multi-source highway monitoring data set to generate a highway monitoring data set with fused spatio-temporal tags; based on a preset multi-modal disaster recognition model, performing disaster feature extraction processing on the highway monitoring data set with fused spatio-temporal tags to obtain a highway disaster feature set; the highway disaster feature set includes environmental correlation features and structural damage features; generating a disaster detection result set according to the highway disaster feature set, where the disaster detection result set contains multiple disaster detection units, and each disaster detection unit corresponds to a disaster type identifier and a disaster influence range identifier for the highway area; generating an inspection optimization strategy set based on the disaster detection result set. In a second aspect, an embodiment of the present invention provides a highway disaster detection system, including: a memory in which a computer program is stored; a processor for loading the computer program to implement the above-mentioned method for highway disaster detection based on air-ground collaborative unmanned inspection.

[0005] The method for highway disaster detection based on air-ground collaborative unmanned inspection provided by the present invention realizes the deep fusion of multi-dimensional remote sensing image data and structural sensing data through the collaborative acquisition of multi-source data by aerial inspection devices and ground inspection devices. By combining spatio-temporal alignment processing, it eliminates the spatio-temporal reference differences between air-ground devices, ensuring the accurate mapping of environmental correlation features and structural damage features in a unified spatio-temporal framework. Further, based on the multi-modal disaster recognition model, it performs cross-modal feature coupling analysis on the fused data, effectively correlating macroscopic environmental changes such as surface deformation and vegetation anomalies with microscopic structural responses such as vibration frequency offset and abnormal stress distribution, breaking through the limitation of traditional data sources or single-modal analysis methods in characterizing complex disaster coupling mechanisms, and significantly improving the early recognition accuracy and type discrimination ability of disasters such as landslides, road surface cracks, and bridge structure failures. At the same time, through the closed-loop inspection optimization strategy driven by the disaster detection results, it dynamically adjusts the collaborative paths and monitoring parameters of air-ground devices, realizes the refined monitoring of high-risk areas and the real-time tracking of the disaster diffusion trend, and comprehensively improves the automation degree and response efficiency of highway disaster detection without relying on manual experience intervention, solving the problems of high false judgment rate, large monitoring blind areas, and rigid resource allocation caused by data fragmentation and spatio-temporal asynchrony in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1It is a flowchart of a road disaster detection method based on air-ground collaborative unmanned inspection provided by an embodiment of the present invention.

[0007] Figure 2 It is a schematic diagram of the composition of a road disaster detection system provided by an embodiment of the present invention. Detailed implementation manners

[0008] Please refer to Figure 1 , which is a flowchart of a road disaster detection method based on air-ground collaborative unmanned inspection provided by an embodiment of the present invention. This road disaster detection method based on air-ground collaborative unmanned inspection can be executed by a road disaster detection system, and may include the following steps: Step S100: Obtain a multi-source road monitoring data set, where the multi-source road monitoring data set includes a multi-dimensional remote sensing image data set collected by an aerial inspection device and a multi-dimensional structural sensing data set collected by a ground inspection device; among them, the multi-dimensional remote sensing image data set contains image area coverage information with different timestamps, and the multi-dimensional structural sensing data set contains structural vibration response information with different spatial positions.

[0009] In an embodiment of the present invention, the aerial inspection device is, for example, a drone, a satellite, etc. equipped with various remote sensing imaging devices, which can monitor a large area of the road and its surrounding areas from the air. The timestamp records the specific time of image acquisition, and images at different times can reflect the changes of the road and its surrounding environment over time. The image area coverage information specifies the geographical range covered by each image. By analyzing images of different regions, the terrain, land surface conditions, etc. along the road can be comprehensively understood. The ground inspection device is, for example, various sensors installed on or around the road structure, such as acceleration sensors, strain sensors, etc. These sensors can monitor the physical state of the road structure in real time, and the collected multi-dimensional structural sensing data set contains structural vibration response information with different spatial positions. Different spatial positions refer to that the sensors are installed at different positions on the road, such as the piers and girders of a bridge, different sections of the road surface, etc. The structural vibration response information is the vibration situation of the road structure when subjected to various external forces (such as vehicle driving, wind force, etc.), including parameters such as vibration frequency, amplitude, and phase.

[0010] Step S200: Perform spatio-temporal alignment processing on the multi-source highway monitoring data set to generate a highway monitoring data set integrated with spatio-temporal tags; the spatio-temporal alignment processing includes time synchronization matching and spatial coordinate transformation of the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set. After obtaining the multi-source highway monitoring data set, since the data from different sources may be inconsistent in time and space, in order to more effectively conduct comprehensive analysis on these data, it is necessary to perform spatio-temporal alignment processing on the multi-source highway monitoring data set to generate a highway monitoring data set integrated with spatio-temporal tags. The spatio-temporal alignment processing mainly includes two key steps: time synchronization matching and spatial coordinate transformation. Time synchronization matching is to ensure the consistency in time of the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set. Because the acquisition times of the aerial inspection equipment and the ground inspection equipment may not be synchronized. For example, the aerial inspection equipment may concentrate on collecting image data within a certain time period, while the ground inspection equipment continuously collects structural vibration response information. Through time synchronization matching, the overlapping interval of the two sets of data in time can be found, enabling the data within this overlapping interval to be effectively correlated and analyzed. Spatial coordinate transformation is used to solve the problem of the difference in the spatial coordinate systems of the two sets of data. The multi-dimensional remote sensing image data set usually uses a geographic coordinate system (such as longitude and latitude) to represent the position of the image, while the multi-dimensional structural sensing data set may use a relative coordinate system (such as a coordinate system established with a fixed point on the highway as the origin). In order to unify these two sets of data in space, it is necessary to transform their coordinates into the same spatial coordinate system. Through the spatio-temporal alignment processing, the generated highway monitoring data set integrated with spatio-temporal tags will have a unified time and space reference, enabling data from different sources to be comprehensively analyzed within the same spatio-temporal framework, and improving the accuracy and reliability of highway disaster detection.

[0011] As an implementation manner, the above step S200 may specifically include the following steps: Step S210: Extract the first spatio-temporal reference information from the multi-dimensional remote sensing image data set. The first spatio-temporal reference information includes the image acquisition time series and the image geographic coordinate range. The image acquisition time series records the acquisition time of each remote sensing image. Through this time series, the acquisition sequence and time interval of different images can be understood. For example, if changes in surface deformation are found in images acquired at different times, the time period during which such changes occurred can be inferred by combining the image acquisition time series. The image geographic coordinate range defines the geographic area covered by each image, which can be represented by longitude and latitude. The geographic coordinate range can accurately locate the actual geographical location corresponding to the remote sensing image, helping to accurately associate the remote sensing image with the actual ground situation. For example, when analyzing the vegetation coverage around a road, the specific road sections affected by the vegetation can be determined through the image geographic coordinate range. When extracting the first spatio-temporal reference information, for example, for the image acquisition time series, the acquisition time information of the remote sensing image is included in the metadata of the remote sensing image data. The acquisition time can be extracted from the image data file and sorted in the order of acquisition time to generate the image acquisition time series. For the image geographic coordinate range, it can be obtained through remote sensing image processing software (such as ENVI, ERDAS, etc.).

[0012] Step S220: Extract the second spatio-temporal reference information from the multi-dimensional structure sensing data set. The second spatio-temporal reference information includes a sensor timestamp sequence and a sensor positioning coordinate sequence. The sensor timestamp sequence records the specific time when each sensor on the ground inspection device collects data. It is similar to the image acquisition time sequence and is accurate to a specific moment. Through the sensor timestamp sequence, the acquisition time corresponding to each structural vibration response data can be determined. The sensor positioning coordinate sequence specifies the specific position of each sensor in space and can be represented using relative coordinates or geographical coordinates. If relative coordinates are used, a coordinate system can be established with a fixed point on the road as the origin; if geographical coordinates are used, the position of the sensor can be represented by longitude and latitude. The sensor positioning coordinate sequence helps to associate the structural vibration response data with a specific geographical location. For example, when abnormal vibration data collected by a certain sensor is found, the location of the sensor can be quickly located through the sensor positioning coordinate sequence, so as to further analyze whether there are problems with the road structure at this location. When extracting the second spatio-temporal reference information, for the sensor timestamp sequence, the ground inspection device will add timestamp information to each data point when collecting data. These timestamp information can be read out through a data acquisition instrument and arranged in the order of acquisition time to generate a sensor timestamp sequence. For the sensor positioning coordinate sequence, if a GPS positioning device is used during sensor installation, the geographical coordinates of the sensor can be directly obtained from the GPS device; if relative coordinates are used, the relative coordinates of the sensor can be calculated by measuring the distance and direction of the sensor relative to the origin. Then, the positioning coordinates of all sensors are arranged in order to generate a sensor positioning coordinate sequence.

[0013] Step S230: Perform time window matching on the first spatio-temporal reference information and the second spatio-temporal reference information to determine the time overlap interval between the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set. Time window matching can ensure that in subsequent analyses, the remote sensing image data and the structural sensing data used are collected within the same time period, thereby improving the accuracy of data correlation analysis. During the time window matching process, first, determine a suitable time window size according to the image acquisition time series in the first spatio-temporal reference information and the sensor time stamp series in the second spatio-temporal reference information. The size of the time window can be determined according to actual requirements and the characteristics of the data. For example, if the time resolution of the data is high, a smaller time window can be selected; if the time span of the data is large, a larger time window can be selected. Then, slide the time window on the image acquisition time series and the sensor time stamp series, and compare the intersection of the two sets of time series within each time window. During the comparison process, an intersection algorithm can be used to calculate the overlapping part of the two sets of time series within each time window. Specifically, if the time length of the overlapping part exceeds a preset threshold, it is considered that there is a valid time overlap interval within this time window. By continuously sliding the time window and making comparisons, all the time overlap intervals between the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set can be finally determined.

[0014] Step S240: Perform data slicing processing on the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set according to the time overlap interval to generate a time-synchronized sub-image data set and a sub-sensing data set. Data slicing processing is to divide the original multi-dimensional data according to the time overlap interval, so that the divided sub-data sets are consistent in time, facilitating subsequent analysis and processing. Specifically, for the multi-dimensional remote sensing image data set, according to the time overlap interval, filter out the image data collected within this time interval. It can be judged whether each image is within the time overlap interval through the image acquisition time series. If so, include this image data in the sub-image data set. For the multi-dimensional structural sensing data set, also according to the time overlap interval, filter out the structural vibration response data collected within this time interval. Judge whether each data point is within the time overlap interval through the sensor time stamp series. If so, include this data point in the sub-sensing data set. The time-synchronized sub-image data set and sub-sensing data set generated through data slicing processing make the remote sensing image data and the structural sensing data consistent in time.

[0015] Step S250: Perform spatial coordinate transformation processing on the sub-image data set and the sub-sensing data set, map the image geographic coordinate range into the spatial coordinate system corresponding to the sensor positioning coordinate sequence, and generate a highway monitoring data set with fused spatio-temporal tags having a unified spatial reference. After completing the data slicing process to obtain the time-synchronized sub-image data set and sub-sensing data set, it is necessary to perform spatial coordinate transformation processing on the sub-image data set and the sub-sensing data set, map the image geographic coordinate range into the spatial coordinate system corresponding to the sensor positioning coordinate sequence, so as to generate a highway monitoring data set with fused spatio-temporal tags having a unified spatial reference. Since the multi-dimensional remote sensing image data set uses a geographic coordinate system (such as longitude and latitude) to represent the position of the image, while the multi-dimensional structured sensing data set may use a relative coordinate system (such as a coordinate system established with a fixed point on the highway as the origin), there are differences between these two coordinate systems and conversion is required.

[0016] The specific implementation method of the spatial coordinate transformation processing can adopt a coordinate transformation model. Taking the seven-parameter transformation model as an example, it includes three translation parameters (translation amounts in the X, Y, and Z directions), three rotation parameters (rotation angles around the X, Y, and Z axes), and one scale parameter. First, select a set number of control points, and these control points have accurate coordinate values in both the geographic coordinate system and the relative coordinate system. Then, based on the coordinate values of these control points, establish an error equation. Solve the error equation by the least squares method to obtain the seven transformation parameters. After obtaining the transformation parameters, for each image point in the sub-image data set, use the transformation parameters to convert its geographic coordinate into the coordinate in the spatial coordinate system corresponding to the sensor positioning coordinate sequence. Similarly, for each data point in the sub-sensing data set, its coordinate is already in the target coordinate system and no conversion is required. Through the spatial coordinate transformation processing, map the image geographic coordinate range into the spatial coordinate system corresponding to the sensor positioning coordinate sequence, so that the sub-image data set and the sub-sensing data set have a unified reference in space.

[0017] Step S300: Based on a preset multi-modal disaster recognition model, perform disaster feature extraction processing on the highway monitoring data set integrated with spatio-temporal tags to obtain a highway disaster feature set; the highway disaster feature set includes environment-related features and structural damage features. The preset multi-modal disaster recognition model combines the characteristics of remote sensing image data and structural sensing data, and can extract features related to highway disasters from the highway monitoring data set integrated with spatio-temporal tags. This model can adopt a deep learning model architecture, such as the combination of a convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN is used to process remote sensing image data and extract spatial features in the images; the RNN is used to process the temporal features in the structural sensing data. During the disaster feature extraction processing, the highway monitoring data set integrated with spatio-temporal tags is input into the preset multi-modal disaster recognition model, and the model performs hierarchical processing on the input data to extract environment-related features and structural damage features respectively. The environment-related features reflect the impact of highway surrounding environmental factors on highway disasters, including surface deformation features, abnormal vegetation coverage features, and water body distribution change features, etc. The structural damage features reflect the health status of the highway's own structure, including vibration frequency offset features, abnormal stress distribution features, and material fatigue accumulation features, etc.

[0018] As an implementation manner, step S300 may specifically include the following steps: Step S310: Perform environmental feature extraction processing on the sub-image data set in the highway monitoring data set integrated with spatio-temporal tags to generate an environmental association feature set; the environmental association feature set includes surface deformation features, abnormal vegetation coverage features, and water body distribution change features. The environmental association feature set contains information such as surface deformation features, abnormal vegetation coverage features, and water body distribution change features. The sub-image data set is a subset of time-synchronized remote sensing image data obtained after spatio-temporal alignment processing, which contains image data in different bands, such as visible light band, infrared band, and radar band, etc. Performing environmental feature extraction processing on these image data can obtain information on the highway surrounding environment from different perspectives. For the extraction of surface deformation features, it can be achieved by analyzing the elevation information and crack distribution in the sub-image data set. For example, using the digital elevation model (DEM) technology to process remote sensing image data at different times to generate high-precision DEM data. By comparing DEM data at different times, the surface elevation change rate can be calculated to reflect the vertical deformation of the surface. At the same time, image processing algorithms, such as edge detection algorithms and morphological processing algorithms, can be used to identify surface cracks in the image and statistically analyze the distribution characteristics of the cracks, such as the length, width, and density of the cracks. The extraction of abnormal vegetation coverage features can be assisted by the image data in the infrared band. The infrared band image can reflect the thermal radiation characteristics of vegetation. There are obvious differences in the thermal radiation intensity between healthy vegetation and vegetation affected by diseases, pests, fires, etc. By analyzing the thermal radiation intensity of the infrared band data subset, such as calculating statistical parameters such as the average thermal radiation intensity and the standard deviation of the thermal radiation intensity, the surface temperature anomaly area and the heat source distribution characteristics can be identified to determine whether there are abnormal vegetation coverage conditions. The extraction of water body distribution change features mainly relies on the image data in the radar band. The radar band image has the ability to penetrate clouds and can clearly reflect the water body distribution on the surface. Using the interferometric synthetic aperture radar (InSAR) technology to process the radar band data subset at different times, the vertical displacement and horizontal displacement vector of the surface can be measured to analyze the change of water body distribution.

[0019] As an implementation manner, step S310 may specifically include the following steps: Step S311: Perform multi-spectral band decomposition processing on the sub-image data set to generate a visible light band data subset, an infrared band data subset, and a radar band data subset. The multi-spectral band decomposition processing is based on the fact that image data in different bands have different physical characteristics and information expression capabilities, and separates the sub-image data set according to the bands, so as to perform more targeted analysis and feature extraction on the data in different bands in the follow-up. The sub-image data set is obtained by performing spatio-temporal alignment processing on multi-dimensional remote sensing image data collected by an aerial inspection device, and contains image information in multiple bands. The multi-spectral band decomposition processing method can adopt, for example, the principal component analysis (PCA) algorithm or the independent component analysis (ICA) algorithm. Taking the principal component analysis algorithm as an example, by performing a linear transformation on the original multi-spectral image data, it is converted into a new coordinate system, so that the components of the data in the new coordinate system are independent of each other and are sorted according to the variance size. Specifically, the original multi-spectral image data is standardized so that the data in each band has the same mean and variance. Then, the covariance matrix of the standardized data is calculated. Next, the covariance matrix is subjected to eigenvalue decomposition to obtain eigenvalues and eigenvectors. The eigenvectors are sorted according to the size of the eigenvalues, and the first N eigenvectors with larger eigenvalues are selected as the principal components, and the value of N is not limited. Finally, the original multi-spectral image data is projected onto these principal components to obtain principal component image data. According to the distribution of different bands in the principal component image data, it is separated into a visible light band data subset, an infrared band data subset, and a radar band data subset.

[0020] Step S312: Perform surface deformation analysis processing on the visible light band data subset to extract the surface elevation change rate and the surface crack distribution characteristics. After completing the multi-spectral band decomposition processing and obtaining the visible light band data subset, it is necessary to perform surface deformation analysis processing on this data subset to extract the surface elevation change rate and the surface crack distribution characteristics. Surface deformation is an indicator of the change of the highway surrounding environment, which may be caused by natural factors (such as earthquakes, landslides, etc.) or human factors (such as engineering construction, groundwater extraction, etc.). By analyzing the surface elevation change rate and the surface crack distribution characteristics, potential highway disaster hazards can be discovered in a timely manner.

[0021] The process of performing surface deformation analysis and processing on the visible light band data subset is, for example, to use a stereo matching algorithm to process visible light band images at different times to generate a digital elevation model (DEM). The principle of the stereo matching algorithm is to find the corresponding points of the same ground object in different images, calculate the disparity of these corresponding points, and then calculate the three-dimensional coordinates of the ground object based on the disparity and the internal and external parameters of the camera. Possible stereo matching algorithms include region-based matching algorithms, feature-based matching algorithms, and energy optimization-based matching algorithms, etc. After obtaining the DEMs at different times, by comparing these DEM data, the surface elevation change rate can be calculated. Specifically, for each pixel point, calculate the elevation difference at different times, and then divide it by the time interval to obtain the elevation change rate of this pixel point. Statistical analysis is performed on the elevation change rates of all pixel points, such as calculating the average value, standard deviation, etc., to obtain the surface elevation change rate of the entire area. For the extraction of the distribution characteristics of surface cracks, image processing algorithms can be used, such as edge detection algorithms and morphological processing algorithms. The edge detection algorithm can identify the edges of ground objects in the image, thereby detecting the locations of surface cracks. Possible edge detection algorithms include the Sobel operator, Canny operator, etc. Taking the Canny operator as an example, it includes Gaussian smoothing, gradient calculation, non-maximum suppression, and double-threshold processing. First, perform Gaussian smoothing on the visible light band image to reduce the influence of noise. Then, calculate the gradient amplitude and direction of the image. Next, perform non-maximum suppression to only retain the pixel points with the local maximum gradient amplitude. Finally, through double-threshold processing, the pixel points are divided into strong edge points, weak edge points, and non-edge points, and only the strong edge points and the weak edge points connected to the strong edge points are retained to obtain the edge information of the surface cracks. After obtaining the edge information of the surface cracks, morphological processing algorithms, such as dilation, erosion, opening operation, and closing operation, etc., are used to refine, connect, and remove noise from the cracks to obtain the complete distribution characteristics of the surface cracks. For example, through the dilation operation, small cracks can be connected, and through the erosion operation, the noise points around the cracks can be removed.

[0022] Step S313: Analyze and process the thermal radiation intensity of the infrared band data subset to extract the surface temperature anomaly regions and the heat source distribution characteristics. For example, perform radiometric calibration on the infrared band image to convert the gray value of the image into the actual thermal radiation intensity value. Radiometric calibration is to convert the digital quantization value (DN value) of the image into the radiance value or temperature value according to the calibration coefficient of the infrared sensor. The possible radiometric calibration methods are absolute calibration and relative calibration. Absolute calibration is to establish the relationship between the gray value of the image and the actual radiation intensity by measuring the radiation intensity of the known radiation source; relative calibration is to establish the relative radiation intensity relationship by comparing different regions in the same image. After completing the radiometric calibration, perform statistical analysis on the thermal radiation intensity values to calculate the average thermal radiation intensity and standard deviation of the image. Then, according to the average thermal radiation intensity and standard deviation, set a threshold range, and define the regions where the thermal radiation intensity values exceed this threshold range as the surface temperature anomaly regions. For example, if the thermal radiation intensity value of a certain region is higher than the average thermal radiation intensity plus a preset multiple of the standard deviation, then this region is considered a high-temperature anomaly region; if it is lower than the average thermal radiation intensity minus a preset multiple of the standard deviation, then this region is considered a low-temperature anomaly region. For the extraction of the heat source distribution characteristics, a clustering analysis algorithm such as the K-means clustering algorithm can be used. Specifically, first randomly select K data points as the initial clustering centers. Then, calculate the distance from each data point to each clustering center, and assign the data point to the cluster where the nearest clustering center is located. Next, recalculate the center position of each cluster. Repeat the above steps until the clustering centers no longer change or reach the preset number of iterations. After performing K-means clustering analysis on the thermal radiation intensity values in the infrared band data subset, different clustering results are obtained. Each cluster represents a heat source region. By analyzing the position, size, and thermal radiation intensity characteristics of the clusters, the heat source distribution characteristics can be obtained. For example, parameters such as the area, average thermal radiation intensity, and maximum thermal radiation intensity of each heat source region can be statistically analyzed to understand the distribution and intensity of the heat sources.

[0023] Step S314: Perform interferometric synthetic aperture radar processing on the radar band data subset to extract the surface vertical displacement and horizontal displacement vector. The process of performing InSAR processing on the radar band data subset is, for example, to first select two or more radar band images at different times. These images need to cover the same area and have an overlap. Then, perform registration processing on these images to make them precisely aligned spatially. The registration method can use feature-based registration algorithms, such as the scale-invariant feature transform (SIFT) algorithm or the speeded-up robust features (SURF) algorithm. These algorithms can extract feature points in the images and achieve image registration by matching these feature points. After completing the image registration, perform interferometric processing to generate an interferogram. The interferogram is obtained by multiplying two registered radar images in the complex domain and contains the phase information of the surface. The phase information is related to factors such as surface deformation, terrain, and atmospheric delay. To isolate the surface deformation information, phase unwrapping processing needs to be performed on the interferogram. Phase unwrapping is to restore the phase values in the interferogram from the range of [-π, π] to the true phase values. Possible phase unwrapping algorithms include the minimum cost flow algorithm, the branch cut method, etc. After completing the phase unwrapping, perform atmospheric correction and terrain correction processing. The purpose of atmospheric correction is to eliminate the influence of atmospheric delay on the phase. Possible atmospheric correction methods include correction methods based on meteorological data and statistical analysis methods based on interferograms. The purpose of terrain correction is to eliminate the influence of terrain factors on the phase, and usually a digital elevation model (DEM) is used for terrain correction. After atmospheric correction and terrain correction, a phase map containing only surface deformation information is obtained. According to the relationship between the phase and the surface displacement, the phase map can be converted into a surface vertical displacement map. The specific conversion formula can be: , where is the surface vertical displacement,[[]] is the radar wavelength,[[]] is the phase change amount. To extract the surface horizontal displacement vector, multi-track InSAR data or other measurement means can be used. For example, radar images on different tracks can be used for InSAR processing to obtain surface displacement information in different directions, and then the surface horizontal displacement vector can be obtained by vector synthesis.[[]]

[0024] Step S315: Generate an environmental correlation feature set by synthesizing the surface elevation change rate, surface crack distribution characteristics, surface temperature anomaly regions, heat source distribution characteristics, surface vertical displacement, and horizontal displacement vector. The process of synthesizing these features is, for example, to normalize numerical features such as the surface elevation change rate, surface vertical displacement, and horizontal displacement vector so that they have the same scale range. The normalization process can use the min-max normalization method, that is, subtract the minimum value of each feature from the feature value and then divide by the difference between the maximum and minimum values of the feature. For non-numerical features such as surface crack distribution characteristics, surface temperature anomaly regions, and heat source distribution characteristics, they can be converted into numerical features by encoding. For example, for surface crack distribution characteristics, information such as the length, width, and density of cracks can be quantified, and then these quantified values can be combined into a feature vector. For surface temperature anomaly regions and heat source distribution characteristics, information such as the area, location, and intensity of the anomaly regions can be encoded to obtain the corresponding feature vectors. After completing the normalization and encoding of the features, these feature vectors are concatenated to generate a comprehensive feature vector. This comprehensive feature vector contains all environmental feature information such as the surface elevation change rate, surface crack distribution characteristics, surface temperature anomaly regions, heat source distribution characteristics, surface vertical displacement, and horizontal displacement vector. Finally, this comprehensive feature vector is used as an element of the environmental correlation feature set, and the above steps are repeated to comprehensively process all the extracted environmental features to generate a complete environmental correlation feature set.

[0025] Step S320: Extract structural feature sets from the sub-sensing data sets in the highway monitoring data set fused with spatio-temporal tags to generate structural damage feature sets; the structural damage feature sets include vibration frequency offset features, stress distribution anomaly features, and material fatigue accumulation features. The sub-sensing data sets are subsets of time-synchronized structural sensing data obtained after spatio-temporal alignment processing, which contain vibration response information and stress distribution information of the highway structure, etc. Extracting structural feature sets from these data can understand the mechanical properties and damage conditions of the highway structure from different perspectives.

[0026] For the extraction of vibration frequency offset characteristics, it is mainly achieved by performing frequency-domain conversion processing on the structural vibration response information. The fast Fourier transform (FFT) algorithm can be used to convert the vibration response signal in the time domain into a set of vibration spectrum characteristics in the frequency domain. The FFT algorithm can convert a time-domain signal with a length of N into a frequency-domain signal with a length of N, and the computational complexity is O(NlogN). After obtaining the set of vibration spectrum characteristics, the fundamental frequency component, harmonic component, and noise component are extracted from it, and the fundamental frequency offset and the proportion of harmonic energy are calculated. The fundamental frequency offset is the difference between the current fundamental frequency and the initial fundamental frequency, which reflects the change in the stiffness of the highway structure. The proportion of harmonic energy is the proportion of the energy of the harmonic component in the total energy, which reflects the nonlinear vibration characteristics of the highway structure. The extraction of the abnormal stress distribution characteristics can be achieved by performing spatial interpolation processing on the stress distribution information in the sub-sensing data set. Spatial interpolation is a method of estimating the value of an unknown point based on the values of known points. Possible spatial interpolation methods include inverse distance weighted interpolation method, Kriging interpolation method, etc. Taking the inverse distance weighted interpolation method as an example, the value of the unknown point is determined according to the distance between the unknown point and the known points. The closer the known points are to the unknown point, the greater the influence on the unknown point. Through spatial interpolation processing, a stress distribution contour map can be generated, and the maximum stress value and the stress gradient change rate can be extracted from it. The maximum stress value reflects the maximum stress level borne by the highway structure, and the stress gradient change rate reflects the change of stress in space. The determination of the material fatigue accumulation characteristics needs to comprehensively consider parameters such as the fundamental frequency offset, the proportion of harmonic energy, the maximum stress value, and the stress gradient change rate. A material fatigue accumulation model can be established to evaluate the fatigue accumulation degree of the highway structure material according to the changes of these parameters. For example, the Miner linear cumulative damage criterion can be adopted. This criterion believes that the fatigue damage of materials is caused by the accumulation of damage caused by each stress cycle. When the damage accumulation reaches a certain level, the material will undergo fatigue failure.

[0027] As an implementation manner, step S320 may specifically include the following steps: Step S321: Perform frequency-domain conversion processing on the structural vibration response information in the sub-sensing data set to generate a set of vibration spectrum characteristics. The structural vibration response information is the vibration signal generated by the highway structure collected by the ground inspection equipment when it is subjected to external forces and exists in the form of a time-domain signal. The time-domain signal reflects the change of the vibration signal over time, but it is difficult to directly obtain the frequency information of the vibration from it. The frequency-domain conversion processing can convert the time-domain signal into a frequency-domain signal, thus clearly showing the frequency components and energy distribution of the vibration signal, providing a basis for subsequent extraction of structural damage characteristics such as vibration frequency offset characteristics. The frequency-domain conversion processing can adopt the fast Fourier transform (FFT) algorithm. By utilizing the symmetry and periodicity of complex numbers, the computational complexity of the DFT is reduced from O(N 2)Reduce to O(NlogN), where N is the length of the signal. Specifically, discretize the structural vibration response information according to the sampling frequency to obtain a discrete time-domain signal sequence. Then, perform zero-padding on the time-domain signal sequence to make its length a power of 2 to improve the computational efficiency of the FFT algorithm. Next, input the zero-padded time-domain signal sequence into the FFT algorithm for calculation to obtain a frequency-domain signal sequence. Finally, calculate the amplitude spectrum of the frequency-domain signal sequence to obtain a set of vibration spectrum features. The set of vibration spectrum features contains various frequency components of the structural vibration response signal and their corresponding energy values. By analyzing the set of vibration spectrum features, the vibration characteristics of the highway structure can be understood, such as the fundamental frequency, harmonic frequencies, etc. The fundamental frequency is the main frequency component of the structural vibration, and it is related to the stiffness and mass of the structure. When damage occurs to the highway structure, its stiffness changes, resulting in a shift in the fundamental frequency. Harmonic frequencies are integer multiples of the fundamental frequency, and they reflect the nonlinear vibration characteristics of the structure. By monitoring the changes in the fundamental frequency and harmonic frequencies, potential damage to the highway structure can be detected in a timely manner.

[0028] Step S322: Extract the fundamental frequency component, harmonic component, and noise component from the set of vibration spectrum features, and calculate the fundamental frequency offset and the harmonic energy ratio. The fundamental frequency component is the frequency component with the highest energy in the vibration spectrum, and it represents the main vibration frequency of the highway structure. The harmonic component is the frequency component that is an integer multiple of the fundamental frequency, and it reflects the nonlinear vibration characteristics of the structure. The noise component is the other frequency components in the vibration spectrum except for the fundamental frequency component and the harmonic component, which are caused by factors such as environmental interference and measurement errors.

[0029] The process of extracting the fundamental frequency component, harmonic component, and noise component is, for example, to perform peak detection on the vibration spectrum feature set to find the peak points in the spectrum. The frequencies corresponding to the peak points are the possible fundamental frequencies or harmonic frequencies. Then, based on the frequencies and energy values of the peak points, it is determined which peak points are the fundamental frequency components and which are the harmonic components. The frequency corresponding to the peak point with the highest energy is the fundamental frequency, and the frequencies corresponding to the other peak points that are integer multiples of the fundamental frequency are the harmonic components. Finally, the other frequency components in the spectrum except the fundamental frequency component and the harmonic component are taken as the noise component. The fundamental frequency offset is the difference between the current fundamental frequency and the initial fundamental frequency. The initial fundamental frequency is the fundamental frequency of the highway structure in a healthy state, which can be obtained by measuring within a period of time after the highway is built or repaired. The method for calculating the fundamental frequency offset is to subtract the initial fundamental frequency from the current fundamental frequency. The magnitude of the fundamental frequency offset reflects the change in the stiffness of the highway structure. When the fundamental frequency offset is large, it indicates that there may be damage to the highway structure. The harmonic energy ratio is the proportion of the energy of the harmonic component in the total energy. The method for calculating the harmonic energy ratio is to first calculate the total energy of the harmonic component, that is, add up the energy values of all harmonic components, then calculate the total energy, that is, add up the energy values of the fundamental frequency component, harmonic component, and noise component, and finally divide the total energy of the harmonic component by the total energy to obtain the harmonic energy ratio. The change in the harmonic energy ratio reflects the change in the nonlinear vibration characteristics of the highway structure. When the harmonic energy ratio increases, it indicates that there may be nonlinear damage to the highway structure.

[0030] Step S323: Perform spatial interpolation processing on the stress distribution information in the sub-sensing data set to generate a stress distribution contour map, and extract the maximum stress value and the stress gradient change rate. Various methods can be used for spatial interpolation processing, such as the inverse distance weighted interpolation method, the Kriging interpolation method, etc. Taking the inverse distance weighted interpolation method as an example, the principle of this method is to determine the stress value of the unknown point according to the distance between the unknown point and the known points. The closer the known point is to the unknown point, the greater the influence on the unknown point. Specifically, determine the position of the unknown point that needs to be interpolated. Then, find several known points that are closest to the unknown point. Next, calculate the weight of each known point according to the distance between it and the unknown point. The closer the distance, the greater the weight. Finally, perform weighted averaging on the stress values of the known points according to the weights to obtain the stress value of the unknown point. By performing the above interpolation processing on all unknown points, continuous stress distribution data is obtained.

[0031] After obtaining the continuous stress distribution data, it is visually displayed in the form of a contour map to generate a stress distribution contour map. The stress distribution contour map can intuitively show the stress distribution in the highway structure. Different colors represent different stress levels, and the darker the color, the greater the stress. The method of extracting the maximum stress value from the stress distribution contour map is to traverse all the pixel points in the contour map and find the pixel point with the maximum stress value. The stress value corresponding to this pixel point is the maximum stress value. The maximum stress value reflects the maximum stress level borne by the highway structure. When the maximum stress value exceeds the design bearing capacity of the highway structure, it may lead to structural damage. The stress gradient change rate is the change rate of stress in space. The method of calculating the stress gradient change rate is to select several adjacent pixel points in the stress distribution contour map, calculate the stress difference and distance difference between them, and then divide the stress difference by the distance difference to obtain the stress gradient change rate. By statistically analyzing the stress gradient change rates at different positions, the change of stress in space can be understood. When the stress gradient change rate is large, it indicates that the stress changes rapidly in this area, and there may be stress concentration, which also increases the risk of highway structure damage.

[0032] Step S324: Determine the vibration frequency offset feature, stress distribution anomaly feature, and material fatigue accumulation feature in the structural damage feature set according to the fundamental frequency offset, harmonic energy ratio, maximum stress value, and stress gradient change rate. The vibration frequency offset feature is mainly related to the fundamental frequency offset. The fundamental frequency offset reflects the change in the stiffness of the highway structure. When the fundamental frequency offset exceeds the threshold, it indicates that the highway structure may have been damaged, resulting in a change in its stiffness. Therefore, the fundamental frequency offset can be used as an important indicator of the vibration frequency offset feature. At the same time, the harmonic energy ratio can also be used as a supplementary indicator of the vibration frequency offset feature. When the harmonic energy ratio increases, it indicates that the nonlinear vibration characteristics of the highway structure have changed, which may also imply structural damage. Considering the fundamental frequency offset and the harmonic energy ratio comprehensively can more accurately determine the vibration frequency offset feature.

[0033] The abnormal characteristics of stress distribution are mainly related to the maximum stress value and the stress gradient change rate. The maximum stress value reflects the maximum stress level borne by the highway structure. When the maximum stress value exceeds the design bearing capacity of the highway structure, it indicates that there may be abnormal stress distribution in the highway structure. The stress gradient change rate reflects the change of stress in space. When the stress gradient change rate is large, it means that the stress changes rapidly in this area, and there may be stress concentration, which is also an indication of abnormal stress distribution. Therefore, the maximum stress value and the stress gradient change rate can be used as important indicators of abnormal stress distribution characteristics. By setting corresponding thresholds, it can be judged whether there is abnormal stress distribution in the highway structure. The determination of the material fatigue accumulation characteristics needs to comprehensively consider multiple parameters such as the fundamental frequency offset, the harmonic energy ratio, the maximum stress value, and the stress gradient change rate. A material fatigue accumulation model can be established to evaluate the fatigue accumulation degree of the highway structure materials according to the changes of these parameters. For example, the Miner linear cumulative damage criterion can be adopted. This criterion believes that the fatigue damage of materials is caused by the accumulation of damage caused by each stress cycle. When the damage accumulation reaches a certain level, the materials will undergo fatigue failure. When establishing the material fatigue accumulation model, parameters such as the fundamental frequency offset, the harmonic energy ratio, the maximum stress value, and the stress gradient change rate can be used as inputs, and the fatigue accumulation degree of the materials can be obtained through model calculation and used as an indicator of the material fatigue accumulation characteristics.

[0034] Step S330: Input the environmental correlation feature set and the structural damage feature set into the multi-modal disaster identification model for feature fusion processing to generate a fused feature vector set. The multi-modal disaster identification model can comprehensively consider the highway surrounding environmental factors and the damage conditions of the highway structure itself. Through feature fusion processing, the environmental correlation features and the structural damage features can be organically combined to extract more representative and discriminative features, thereby improving the accuracy and reliability of highway disaster identification. The multi-modal disaster identification model can adopt a deep learning model architecture, such as the combination of a convolutional neural network (CNN) and a recurrent neural network (RNN). CNN can be used to process the spatial features in the environmental correlation feature set, such as ground deformation features, abnormal vegetation coverage features, etc.; RNN can be used to process the temporal features in the structural damage feature set, such as vibration frequency offset features, abnormal stress distribution features, etc.

[0035] As an implementation manner, step S330 may specifically include the following steps: Step S331: Normalize each environmental association feature in the environmental association feature set to generate a standardized environmental feature vector. The environmental association feature set includes various different types of features such as ground deformation features, abnormal vegetation coverage features, and water body distribution change features. The value ranges and dimensions of these features may be different. Directly performing feature fusion processing may cause the influence of some features to be too large and the influence of other features to be too small, thus affecting the performance of the model. Therefore, it is necessary to normalize these features so that they have the same scale range for subsequent feature fusion processing. The normalization processing can adopt various methods, such as the min-max normalization method, the Z-score normalization method, etc., and no specific limitation is made. It can be understood that in other data calculations involving different dimensions or dimensions in the present invention, those skilled in the art can perform preprocessing such as data dimension elimination and dimension unification according to general knowledge and then perform corresponding data processing. By performing min-max normalization processing on each environmental association feature in the environmental association feature set, each feature value is mapped to the range of [0,1], eliminating the scale difference between the features. Combining all the normalized environmental association features into a vector, the standardized environmental feature vector is obtained. Each element in the standardized environmental feature vector has the same scale range, which enables different environmental association features to be treated fairly in subsequent feature fusion processing, improving the model's learning ability for environmental association features and the effect of feature fusion.

[0036] Step S332: Normalize each structural damage feature in the structural damage feature set to generate a standardized structural feature vector. Similarly, the structural damage feature set includes various different types of features such as vibration frequency offset features, abnormal stress distribution features, and material fatigue accumulation features. The value ranges and dimensions of these features may also be different, and normalization processing is also required to eliminate the scale difference between the features and improve the effect of feature fusion. The normalization processing can also adopt the min-max normalization method.

[0037] Step S333: Align and splice the standardized environmental feature vector and the standardized structural feature vector according to timestamps and spatial coordinates to generate a spatio-temporal aligned multi-modal feature matrix. Timestamps and spatial coordinates are key information in the highway monitoring data set that integrates spatio-temporal labels, and they can clarify the specific time and geographical location corresponding to environmental correlation features and structural damage features. By aligning and splicing according to timestamps and spatial coordinates, it can be ensured that the environmental correlation features and structural damage features are consistent in the spatio-temporal dimension, so that the multi-modal disaster recognition model can more accurately capture the correlation between environmental factors and structural damage. Specifically, first match each feature element in the standardized environmental feature vector and the standardized structural feature vector according to its corresponding timestamp and spatial coordinate. During the matching process, ensure that the environmental correlation features and structural damage features with the same timestamp and spatial coordinate can correspond. Then, splice the matched standardized environmental feature vector and standardized structural feature vector. Finally, arrange the spliced vectors corresponding to all different combinations of timestamps and spatial coordinates in order to form a matrix, which is the spatio-temporal aligned multi-modal feature matrix. Each row of the matrix represents a combination of environmental correlation features and structural damage features at a certain time and spatial position, and each column represents a feature dimension. By generating the spatio-temporal aligned multi-modal feature matrix, a structured data input is provided for the multi-modal disaster recognition model, enabling the model to more effectively perform cross-modal feature analysis and learning.

[0038] Step S334: Through the feature fusion layer in the multi-modal disaster recognition model, cross-modal correlation analysis is performed on the multi-modal feature matrix to generate a set of fused feature vectors; among them, the cross-modal correlation analysis includes the calculation of the correlation weight between the environmental feature and the structural feature and the feature dimension reduction process. For the calculation of the correlation weight between the environmental feature and the structural feature, an attention mechanism can be adopted. The attention mechanism can automatically learn the correlation between different features, assign different weights to different features, and make the model pay more attention to important features. Specifically, the spatio-temporally aligned multi-modal feature matrix is first input into the attention layer. The attention layer calculates the correlation score between each feature and other features. The higher the correlation score, the closer the association between the feature and other features. Then, according to the correlation score, a weight is assigned to each feature. The larger the weight, the more important the feature is in the cross-modal correlation analysis. Finally, each feature is multiplied by its corresponding weight to obtain the weighted feature. The feature dimension reduction process can adopt the principal component analysis (PCA) algorithm or the linear discriminant analysis (LDA) algorithm. The principles have been introduced above and will not be elaborated here. Through the feature fusion layer, cross-modal correlation analysis is performed on the multi-modal feature matrix, the environmental feature and the structural feature are organically combined, and a set of fused feature vectors is generated. Each vector in the set of fused feature vectors synthesizes the information of the environmental correlation feature and the structural damage feature, and can more comprehensively and accurately reflect the characteristics of highway disasters.

[0039] Step S340: Perform disaster pattern matching on the set of fused feature vectors to determine a set of highway disaster features that match the disaster patterns in the preset disaster pattern library. When performing disaster pattern matching, specifically, obtain multiple historical disaster pattern feature vectors from the preset disaster pattern library, and each historical disaster pattern feature vector corresponds to a standard feature template for a disaster type. These standard feature templates are obtained through the analysis and summary of a large amount of historical disaster data and are representative and typical. Then, calculate the similarity scores between each fused feature vector in the set of fused feature vectors and each historical disaster pattern feature vector. The similarity scores can be calculated using various methods, such as Euclidean distance, cosine similarity, etc., and are not specifically limited. Determine the optimal matching disaster type corresponding to each fused feature vector according to the similarity scores, and generate a preliminary disaster feature set containing disaster type identifiers and matching confidence levels. The matching confidence level is the degree of matching between the fused feature vector and the corresponding disaster type and can be represented by the similarity score. The higher the similarity score, the higher the matching confidence level. Finally, perform spatial continuity verification processing on the preliminary disaster feature set, eliminate disaster features with matching confidence levels lower than the preset threshold and spatially isolated distributions, and generate a set of highway disaster features. The spatial continuity verification processing is to exclude some false disaster features caused by data noise or accidental factors. The preset threshold can be set according to the actual situation. Generally speaking, when the matching confidence level is lower than this threshold, the matching result of this disaster feature is considered unreliable. Spatially isolated disaster features are features that are not adjacent or are far away from other disaster features in space. These features may be caused by local anomalies and do not necessarily represent real highway disasters. Through the spatial continuity verification processing, the accuracy and reliability of the set of highway disaster features can be improved.

[0040] As an implementation manner, step S340 may specifically include the following steps: Step S341: Obtain multiple historical disaster mode feature vectors from a preset disaster mode library, and each historical disaster mode feature vector corresponds to a standard feature template of a disaster type. The preset disaster mode library is established by collecting, organizing, and analyzing a large amount of historical highway disaster data. These historical disaster data include various information such as environmental characteristics and structural damage characteristics when different types of highway disasters occur. To construct the preset disaster mode library, it is necessary to first collect various highway disaster cases that occurred in different regions and at different times, including landslides, pavement cracks, bridge structure failures, etc. For each disaster case, collect relevant environmental data (such as ground deformation, vegetation coverage, water body distribution, etc.) and structural data (such as vibration frequency, stress distribution, etc.). Then, preprocess the collected data, including operations such as data cleaning and feature extraction, to remove noise and redundant information and extract the key information that can represent the disaster characteristics. Next, for each disaster type, analyze and summarize the corresponding feature data and extract the representative feature patterns. These feature patterns are the standard feature templates of the disaster type. For example, for landslide disasters, the standard feature template may include ground deformation characteristics (such as a large change rate of ground elevation, dense distribution of ground cracks, etc.), abnormal vegetation coverage characteristics (such as large-area lodging of vegetation, etc.), and structural damage characteristics (such as a large deviation of the vibration frequency of the structure near the highway, etc.). Finally, convert the standard feature template of each disaster type into a historical disaster mode feature vector and store it in the preset disaster mode library. The historical disaster mode feature vector is a multi-dimensional vector, and each dimension corresponds to a feature index. By obtaining the historical disaster mode feature vector from the preset disaster mode library, a reference standard can be provided for subsequent disaster mode matching.

[0041] Step S342: Calculate the similarity scores between each fusion feature vector in the fusion feature vector set and each historical disaster mode feature vector. Multiple methods can be used to calculate the similarity scores, such as Euclidean distance, cosine similarity, Manhattan distance, etc. For each fusion feature vector in the fusion feature vector set, it is necessary to calculate the similarity with all historical disaster mode feature vectors in the preset disaster mode library. For example, assume that there are m fusion feature vectors in the fusion feature vector set and k historical disaster mode feature vectors in the preset disaster mode library, then a total of m×k similarity calculations need to be performed. Through these similarity scores, the matching degree of each fusion feature vector with different disaster types can be initially judged.

[0042] Step S343: Determine the optimal matching disaster type corresponding to each fused feature vector based on the similarity score, and generate a preliminary disaster feature set containing a disaster type identifier and a matching confidence. For each fused feature vector, find the maximum value from all similarity scores. The disaster type represented by the historical disaster pattern feature vector corresponding to the maximum value is the optimal matching disaster type for the fused feature vector. The matching confidence can be expressed by the similarity score, that is, the similarity score corresponding to the maximum value. The matching confidence reflects the degree of matching between the fused feature vector and the corresponding disaster type. The higher the confidence, the more reliable the matching result. The disaster type identifier and matching confidence corresponding to each fused feature vector are combined to generate a preliminary disaster feature set. Each element in the preliminary disaster feature set contains the disaster type information and matching reliability corresponding to a fused feature vector, providing basic data for subsequent spatial continuity verification processing.

[0043] Step S344: Perform spatial continuity verification on the preliminary disaster feature set, remove disaster features with matching confidence lower than a preset threshold and isolated spatial distribution, and generate a highway disaster feature set. The preset threshold is a similarity scoring standard pre-set according to the actual situation. When the matching confidence is lower than the threshold, the matching result of the disaster feature is considered unreliable. Disaster features with isolated spatial distribution are features that are not adjacent to or far away from other disaster features in space. These features may be caused by local abnormal conditions and do not necessarily represent real highway disasters. The process of spatial continuity verification is, for example, to determine the spatial position of each disaster feature based on the spatial coordinate information in the highway monitoring data set with integrated spatiotemporal labels. Then, for each disaster feature, check whether its matching confidence is lower than the preset threshold. If it is lower than the preset threshold, further check its spatial distribution. The neighborhood analysis method can be used to determine whether the spatial distribution of the disaster feature is isolated. For example, set a neighborhood radius, and for each disaster feature, check whether there are other disaster features with matching confidence higher than the preset threshold within the neighborhood radius. If it does not exist, the disaster feature is considered to be spatially isolated and is removed from the preliminary disaster feature set.

[0044] Step S400: Generate a disaster detection result set based on the highway disaster feature set. The disaster detection result set includes multiple disaster detection units, each of which corresponds to a disaster type identifier and a disaster impact range identifier of a highway area. By generating a disaster detection result set, the types of disasters that may exist in different highway areas and the impact range of the disasters can be clarified, providing important information for early warning and handling of highway disasters.

[0045] As an implementation manner, step S400 may specifically include the following steps: Step S410: Perform disaster type classification processing on each highway disaster feature in the highway disaster feature set to determine the disaster type identifier corresponding to each highway disaster feature; the disaster type identifier is a landslide risk identifier, a road surface crack identifier, or a bridge structure failure identifier. In the process of generating the disaster detection result set according to the highway disaster feature set, it is first necessary to perform disaster type classification processing on each highway disaster feature in the highway disaster feature set to determine the disaster type identifier corresponding to each highway disaster feature. The disaster type identifier may be a landslide risk identifier, a road surface crack identifier, or a bridge structure failure identifier, etc. These identifiers can clearly indicate the disaster types that the highway may face and provide important basis for subsequent disaster response and management. The disaster type classification processing may adopt machine learning classification algorithms, such as decision tree classification algorithms, support vector machine classification algorithms, etc. Taking the decision tree classification algorithm as an example, by constructing a decision tree model, classification decisions are made according to different attribute values of highway disaster features. Specifically, first select a part of highway disaster features with known disaster types as training data, input it into the decision tree classification algorithm for training, and construct a decision tree model. The training data contains various attribute information of highway disaster features (such as surface deformation features, vibration frequency offset features, etc.) and the corresponding disaster type identifiers. During the training process, the decision tree classification algorithm will continuously perform node division and learning of decision rules according to the attribute values and disaster type identifiers of the training data. For example, for the attribute of surface elevation change rate, if its value is greater than a certain threshold, the highway disaster feature may be classified as a landslide risk identifier; if its value is less than the threshold, continue to classify according to other attributes. After training, input each highway disaster feature in the highway disaster feature set into the trained decision tree model, and classify according to the decision rules of the decision tree to determine the disaster type identifier corresponding to each highway disaster feature.

[0046] Step S420: Determine the disaster impact range identifier corresponding to each highway disaster feature according to the spatial distribution information of the environmental correlation features and structural damage features in the highway disaster feature set. The process of determining the disaster impact range identifier is, for example, to obtain the spatial coordinate sequences of the surface deformation features, abnormal vegetation coverage features, and water body distribution change features in the environmental correlation feature set, and generate the first spatial distribution information. The spatial coordinate sequences of these environmental correlation features can be obtained by fusing the geographic coordinate information in the highway monitoring data set with spatio-temporal tags. For example, for the surface deformation feature, the longitude and latitude coordinates of its occurrence location can be obtained; for the abnormal vegetation coverage feature, the boundary coordinates of the abnormal area can be obtained, etc. Then, obtain the spatial positioning sequences of the vibration frequency offset features, stress distribution abnormal features, and material fatigue accumulation features in the structural damage feature set, and generate the second spatial distribution information. The spatial positioning sequences of the structural damage features can also be obtained from the highway monitoring data set fused with spatio-temporal tags, and these information can reflect the specific locations of highway structural damage.

[0047] Next, perform a spatial overlay process on the first spatial distribution information and the second spatial distribution information to generate spatial distribution overlay data that includes the environmental feature density distribution and the structural feature density distribution. The spatial overlay process can use Geographic Information System (GIS) technology to overlay the spatial distribution information of the environmental correlation features and the structural damage features in the same geographic coordinate system, visually showing the comprehensive spatial distribution of environmental factors and structural damage. Perform a spatial grid division process on the spatial distribution overlay data to generate multiple spatial grid units, with each spatial grid unit containing an environmental feature density value and a structural feature density value. The spatial grid division can be set according to actual needs and the size of the geographic area. For example, a highway area can be divided into several square grids of equal size. Then, count the quantity or intensity of the environmental features and structural features within each grid unit, and calculate the environmental feature density value and the structural feature density value. Perform a feature density fusion process on each spatial grid unit to calculate the weighted overlay value of the environmental feature density value and the structural feature density value of each spatial grid unit, generating a spatial grid density distribution map. The calculation of the weighted overlay value can determine the weights according to the influence degree of the environmental features and structural features on highway disasters. For example, if it is considered that the structural damage features have a greater impact on highway disasters, a greater weight can be assigned to the structural feature density value. According to the weighted overlay values of each spatial grid unit in the spatial grid density distribution map, extract the continuous spatial grid regions that exceed the preset density threshold to generate a preliminary disaster impact area. The preset density threshold can be set according to historical disaster data and actual experience. When the weighted overlay value of a certain spatial grid unit exceeds this threshold, it is considered that this area may be affected by disasters. Perform a spatial morphology analysis process on the preliminary disaster impact area to extract the boundary contour coordinates and the spatial diffusion direction vector of the preliminary disaster impact area. The spatial morphology analysis can use image processing algorithms, such as edge detection algorithms and morphological processing algorithms, to identify the boundary of the preliminary disaster impact area and calculate its diffusion direction. Finally, determine the core area of the disaster impact range according to the boundary contour coordinates, and generate a disaster impact range diffusion buffer zone in combination with the spatial diffusion direction vector. Perform a spatial merge process on the core area and the diffusion buffer zone to generate a disaster impact range identifier that includes the boundary coordinates of the disaster impact range and the diffusion trend parameters.

[0048] As an implementation manner, step S420 may specifically include the following steps: Step S421: Obtain the spatial coordinate sequences of the surface deformation features, vegetation coverage anomaly features, and water body distribution change features in the environmental correlation feature set, and generate the first spatial distribution information. The surface deformation features, vegetation coverage anomaly features, and water body distribution change features are key features in the environmental correlation feature set, and their spatial coordinate sequences can accurately reflect the distribution of these environmental factors in the geographical space. The specific method for obtaining the spatial coordinate sequences of these features is, for example, for the surface deformation features, when performing surface deformation analysis and processing before, a digital elevation model (DEM) is generated through a stereo matching algorithm, and the surface elevation change rate and surface crack distribution features are calculated. In this process, the geographical coordinates of the locations where the surface deformation occurs can be recorded to form the spatial coordinate sequence of the surface deformation features. For the vegetation coverage anomaly features, when performing thermal radiation intensity analysis and processing on the infrared band data subset, the surface temperature anomaly regions and heat source distribution features are identified. The boundary coordinates or central point coordinates of these anomaly regions can be obtained to form the spatial coordinate sequence of the vegetation coverage anomaly features. For the water body distribution change features, when performing interferometric synthetic aperture radar (InSAR) processing on the radar band data subset, the surface vertical displacement and horizontal displacement vector are extracted to analyze the change in the water body distribution. The boundary coordinates of the regions where the water body distribution changes can be recorded to generate the spatial coordinate sequence of the water body distribution change features. Integrating the spatial coordinate sequences of the surface deformation features, vegetation coverage anomaly features, and water body distribution change features generates the first spatial distribution information.

[0049] Step S422: Obtain the spatial positioning sequences of the vibration frequency offset features, stress distribution anomaly features, and material fatigue accumulation features in the structural damage feature set, and generate the second spatial distribution information. The vibration frequency offset features, stress distribution anomaly features, and material fatigue accumulation features are key features in the structural damage feature set, and their spatial positioning sequences can accurately reflect the distribution of highway structural damage in the geographical space. The process of obtaining the spatial positioning sequences of these features is, for example, for the vibration frequency offset features, after performing frequency domain conversion processing on the structural vibration response information in the sub-sensor data set, the fundamental frequency component, harmonic component, and noise component are extracted, and the fundamental frequency offset and harmonic energy ratio are calculated. In this process, the position coordinates of each sensor can be recorded. Since parameters such as the fundamental frequency offset are corresponding to the sensor positions, these sensor position coordinates constitute the spatial positioning sequence of the vibration frequency offset features.

[0050] For the abnormal characteristics of stress distribution, after performing spatial interpolation processing on the stress distribution information in the sub-sensor data set, a stress distribution cloud map is generated, and the maximum stress value and the stress gradient change rate are extracted. The central point coordinates or boundary coordinates of the stress abnormal area in the stress distribution cloud map can be obtained to form the spatial positioning sequence of the stress distribution abnormal characteristics. For the material fatigue accumulation characteristics, their values are determined according to parameters such as the fundamental frequency offset, the harmonic energy ratio, the maximum stress value, and the stress gradient change rate. Similarly, the sensor position coordinates corresponding to these parameters or the coordinates of the stress abnormal area can be used as the spatial positioning sequence of the material fatigue accumulation characteristics. By integrating the spatial positioning sequences of the vibration frequency offset characteristics, the stress distribution abnormal characteristics, and the material fatigue accumulation characteristics, the second spatial distribution information is generated.

[0051] Step S423: Perform spatial overlay processing on the first spatial distribution information and the second spatial distribution information to generate spatial distribution overlay data including the environmental feature density distribution and the structural feature density distribution. The spatial overlay processing can adopt Geographic Information System (GIS) technology, which can overlay and analyze spatial data from different sources. The specific process of implementing the spatial overlay processing is as follows: First, import the first spatial distribution information and the second spatial distribution information into the GIS software to ensure that they are in the same geographic coordinate system. For example, a common geographic coordinate system such as the WGS84 coordinate system can be selected. Then, in the GIS software, use the spatial overlay analysis tool to perform an overlay operation on the first spatial distribution information and the second spatial distribution information. The overlay operation can merge the geographic features in the two spatial distribution information to generate a new spatial data set. In this new data set, each geographic unit (such as a grid, a polygon, etc.) contains the relevant information of environmental features and structural features. Next, count the number of environmental features and the number of structural features in each geographic unit, and calculate the environmental feature density and the structural feature density. The environmental feature density can be defined as the number of environmental features (such as surface deformation features, abnormal vegetation cover features, etc.) in each geographic unit divided by the area of the geographic unit; the structural feature density can be defined as the number of structural features (such as vibration frequency offset features, stress distribution abnormal features, etc.) in each geographic unit divided by the area of the geographic unit. Finally, add the environmental feature density and structural feature density information to the new spatial data set to generate spatial distribution overlay data including the environmental feature density distribution and the structural feature density distribution.

[0052] Step S424: Perform spatial grid division processing on the spatially distributed superimposed data to generate multiple spatial grid cells, each spatial grid cell containing an environmental feature density value and a structural feature density value. The process of spatial grid division processing is, for example, to determine the size and shape of the spatial grid according to the scope and actual requirements of the highway area. The size of the spatial grid can be set according to the scale of the highway and the accuracy requirements of the data. The shape of the spatial grid can be selected as square, rectangular, hexagonal, etc. Then, in the Geographic Information System (GIS) software, use the grid generation tool to divide the spatially distributed superimposed data according to the determined grid size and shape. The grid generation tool will divide the highway area into multiple non-overlapping grid cells and assign a unique identifier to each grid cell. Next, for each spatial grid cell, count the number of environmental features and structural features inside it. The number of environmental features can be determined by counting the number of surface deformation features, vegetation cover anomaly features, and water body distribution change features contained in the grid cell; the number of structural features can be determined by counting the number of vibration frequency offset features, stress distribution anomaly features, and material fatigue accumulation features contained in the grid cell. Finally, according to the area of the grid cell and the counted number of environmental features and structural features, calculate the environmental feature density value and the structural feature density value of each spatial grid cell. The environmental feature density value can be expressed as the number of environmental features divided by the area of the grid cell, and the structural feature density value can be expressed as the number of structural features divided by the area of the grid cell. Adding the environmental feature density value and the structural feature density value to the attribute information of each spatial grid cell generates a spatial grid cell containing the environmental feature density value and the structural feature density value.

[0053] Step S425: Perform feature density fusion processing on each spatial grid cell, calculate the weighted superposition value of the environmental feature density value and the structural feature density value of each spatial grid cell, and generate a spatial grid density distribution map. The process of feature density fusion processing is, for example, to determine the weights of the environmental feature density value and the structural feature density value. The determination of the weights is evaluated according to the influence degrees of environmental factors and structural damage on highway disasters. For example, if in a certain area, the geological conditions around the highway are relatively complex and environmental factors such as ground deformation have a greater impact on highway disasters, then a larger weight can be assigned to the environmental feature density value; if the structural design of the highway is more critical and the impact of structural damage on highway disasters is more prominent, then a larger weight can be assigned to the structural feature density value. The value range of the weights is between [0, 1], and the sum of the weight of the environmental feature density value and the weight of the structural feature density value is 1. Then, for each spatial grid cell, calculate the weighted superposition value of its environmental feature density value and structural feature density value according to the determined weights. Finally, add the weighted superposition value of each spatial grid cell to its attribute information and perform visual display in a Geographic Information System (GIS) software to generate a spatial grid density distribution map. The spatial grid density distribution map can use different colors or grayscales to represent different weighted superposition values. The darker the color or the higher the grayscale, the greater the comprehensive density value of the grid cell, that is, the higher the potential risk of being affected by highway disasters in this area.

[0054] Step S426: According to the weighted superposition values of each spatial grid unit in the spatial grid density distribution map, extract continuous spatial grid regions that exceed a preset density threshold to generate a preliminary disaster impact area. The preset density threshold is a comprehensive density value standard preset based on historical disaster data and actual experience. When the weighted superposition value of a certain spatial grid unit exceeds this threshold, it is considered that this area may be affected by disasters. The process of extracting the preliminary disaster impact area is, for example, to determine the preset density threshold. The determination of the preset density threshold can be achieved by analyzing a large amount of historical disaster data, statistically analyzing the distribution of comprehensive density values under different disaster levels, and thus determining a suitable preset density threshold. For example, if historical data shows that when the comprehensive density value exceeds 0.8, the probability of highway disasters occurring in this area is relatively high, then the preset density threshold can be set to 0.8. Then, in the spatial grid density distribution map, traverse each spatial grid unit and check whether its weighted superposition value exceeds the preset density threshold. For spatial grid units that exceed the preset density threshold, mark them as units that may be affected by disasters. Next, use a connected region analysis algorithm, such as the flood filling algorithm or the four-neighborhood / eight-neighborhood search algorithm, to find all continuous spatial grid regions marked as possibly affected by disasters. A continuous spatial grid region is an area composed of adjacent marked units, and the definition of adjacent can be selected as the four-neighborhood (adjacent up, down, left, and right) or the eight-neighborhood (adjacent up, down, left, right, and the four diagonals) according to actual needs. Finally, merge these continuous spatial grid regions to generate a preliminary disaster impact area. The preliminary disaster impact area is a polygon area composed of multiple continuous spatial grid units, which can reflect the geographical range that may be affected by highway disasters.

[0055] Step S427: Conduct spatial form analysis and processing on the preliminary disaster impact area, and extract the boundary contour coordinates and spatial diffusion direction vectors of the preliminary disaster impact area. The process of spatial form analysis and processing is, for example, to extract the boundary of the preliminary disaster impact area. Edge detection algorithms in image processing, such as the Canny edge detection algorithm or the Sobel edge detection algorithm, can be used to separate the boundary of the preliminary disaster impact area from the background. These algorithms calculate the gradient values of the pixels in the image, find the pixels with larger gradient values, and connect them to form the boundary. After extracting the boundary of the preliminary disaster impact area, obtain the coordinates of each point on the boundary, and these coordinates are the boundary contour coordinates of the preliminary disaster impact area. The boundary contour coordinates can accurately describe the geometric shape and position of the preliminary disaster impact area. Then, extract the spatial diffusion direction vectors. The principal component analysis (PCA) algorithm can be used to analyze the spatial grid cells within the preliminary disaster impact area. Specifically, take the coordinates of the spatial grid cells within the preliminary disaster impact area as input data and calculate its covariance matrix. Then, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors. Sort the eigenvectors according to the magnitudes of the eigenvalues, and select the eigenvector with the largest eigenvalue as the principal component direction. The principal component direction represents the main diffusion direction of the preliminary disaster impact area, and take the unit vector in this direction as the spatial diffusion direction vector.

[0056] Step S428: Determine the core area of the disaster impact range according to the boundary contour coordinates, and generate a disaster impact range diffusion buffer in combination with the spatial diffusion direction vector. The core area of the disaster impact range is the area where the disaster impact is the most severe, usually located inside the preliminary disaster impact area; the disaster impact range diffusion buffer is the area where the disaster may further spread, and it surrounds the core area. The process of determining the core area of the disaster impact range is, for example, to calculate the geometric center of the preliminary disaster impact area according to the boundary contour coordinates. For example, use the centroid calculation method to perform weighted averaging on the coordinates of each point on the boundary contour to obtain the centroid coordinates of the preliminary disaster impact area. Then, with the centroid as the center, determine the range of the core area according to the set rules. For example, according to the weighted superposition values of each spatial grid cell in the spatial grid density distribution map, select the area with a higher weighted superposition value as the core area. It is also possible to set a suitable radius with the centroid as the center of the circle and take the area within this radius as the core area. The process of generating the disaster impact range diffusion buffer is, for example, to determine the possible diffusion direction of the disaster in combination with the spatial diffusion direction vector. The spatial diffusion direction vector represents the main diffusion direction of the preliminary disaster impact area. According to the direction and magnitude of this vector, the possible diffusion range of the disaster can be predicted. Then, based on the boundary of the core area, expand along the direction of the spatial diffusion direction vector according to a preset distance to generate the diffusion buffer. The expanded distance can be set according to historical disaster data and actual experience.

[0057] Step S429: Perform a spatial merging process on the core area and the diffusion buffer to generate a disaster impact range identifier that includes the boundary coordinates of the disaster impact range and the diffusion trend parameters. Specifically, first import the geographical data (such as the boundary coordinates of the polygon) of the core area and the diffusion buffer into GIS software to ensure that they are in the same geographical coordinate system. Then, use the spatial merging tool in the GIS software to perform a merging operation on the core area and the diffusion buffer. The merging operation will integrate the boundaries of the two areas to generate a new polygon area, which is the final boundary of the disaster impact range. Obtain the boundary coordinates of the merged polygon area, and these coordinates are the boundary coordinates of the disaster impact range. The boundary coordinates of the disaster impact range can accurately describe the geometric shape and location of the disaster impact range. The diffusion trend parameters can be determined based on the previously extracted spatial diffusion direction vector and the expansion distance of the diffusion buffer. For example, the direction and magnitude of the spatial diffusion direction vector can be used as the diffusion direction parameter, and the expansion distance of the diffusion buffer can be used as the diffusion distance parameter. Combine these diffusion trend parameters with the boundary coordinates of the disaster impact range to form the disaster impact range identifier.

[0058] Step S430: Associate and combine the disaster type identifier and the disaster impact scope identifier to generate a disaster detection result set containing multiple disaster detection units; each disaster detection unit includes the disaster occurrence location coordinates, the disaster type priority score, and the disaster impact diffusion trend prediction result. The process of association and combination is, for example, for each highway disaster feature, associate its corresponding disaster type identifier and disaster impact scope identifier. This can be achieved by adding corresponding fields in the data record to bind the disaster type identifier and the disaster impact scope identifier to this highway disaster feature. Then, determine the disaster occurrence location coordinates. The disaster occurrence location coordinates can be determined based on the geometric center coordinates of the core area in the disaster impact scope identifier. For example, the centroid coordinates of the core area boundary contour coordinates can be calculated and used as the disaster occurrence location coordinates. Next, conduct the disaster type priority scoring. The disaster type priority scoring needs to comprehensively consider the harm degree of the disaster type and the size of the disaster impact scope. Different initial priority scores can be set for different disaster types (such as landslide risk, road surface cracks, bridge structure failure, etc.) according to historical disaster data and expert experience. For example, the landslide risk may cause greater damage to the highway, and its initial priority score can be set to a relatively high value; the harm degree of road surface cracks is relatively small, and its initial priority score can be set to a relatively low value. Then, adjust the initial priority score according to the size of the disaster impact scope. The larger the disaster impact scope, the higher the severity of the disaster, and the priority score should also be increased accordingly. A linear weighting method can be used for adjustment. For example, set an impact scope adjustment coefficient, multiply the area of the disaster impact scope by this coefficient, and then add the initial priority score to obtain the final disaster type priority score. Finally, determine the disaster impact diffusion trend prediction result. The disaster impact diffusion trend prediction result can be determined based on the diffusion trend parameters in the disaster impact scope identifier. The diffusion trend parameters include information such as the diffusion direction and diffusion distance, and the possible diffusion range and direction of the disaster in the future can be predicted based on this information. Associate the disaster occurrence location coordinates, the disaster type priority score, and the disaster impact diffusion trend prediction result with the disaster type identifier and the disaster impact scope identifier to form a disaster detection unit. Repeat the above steps for all highway disaster features to generate multiple disaster detection units. Integrate these disaster detection units, that is, generate a disaster detection result set containing multiple disaster detection units.

[0059] As an implementation manner, the above step S430 may specifically include the following steps: Step S431: Obtain the disaster type classification result corresponding to the disaster type identifier and the disaster impact range boundary coordinates corresponding to the disaster impact range identifier. The process of obtaining the disaster type classification result corresponding to the disaster type identifier is, for example, that in the previous disaster type classification process, the disaster type identifier corresponding to each highway disaster feature has been determined. The specific disaster type corresponding to each disaster type identifier can be obtained by querying the mapping table between the disaster type identifier and the disaster type classification result. For example, the mapping table may stipulate that the disaster type identifier "1" corresponds to the landslide risk, and the disaster type identifier "2" corresponds to the road surface crack, etc. By querying this mapping table, the disaster type classification result corresponding to each disaster type identifier can be obtained. The process of obtaining the disaster impact range boundary coordinates corresponding to the disaster impact range identifier is, for example, that in the previous disaster impact range identifier generation process, the disaster impact range identifier including the disaster impact range boundary coordinates has been obtained through the spatial merging process. The disaster impact range boundary coordinates can be directly extracted from the data record of the disaster impact range identifier. The disaster impact range boundary coordinates are usually represented in the form of a polygon, and the coordinates of each vertex can be represented by the longitude and latitude or the coordinate values in other geographical coordinate systems.

[0060] Step S432: Extract the change rate of the surface deformation feature in the environmental correlation feature set and the change rate of the stress gradient in the structural damage feature set within the boundary coordinates of the disaster impact range, and generate the coordinates of the disaster occurrence location. The process of extracting the change rate of the surface deformation feature and the change rate of the stress gradient is, for example, according to the boundary coordinates of the disaster impact range, to screen out the feature data in the environmental correlation feature set and the structural damage feature set located within the boundary coordinates. The spatial query function of Geographic Information System (GIS) software can be used to compare the geographical coordinates of the environmental correlation feature set and the structural damage feature set with the boundary coordinates of the disaster impact range, and screen out the feature data located within the boundary coordinates. For the surface deformation features in the screened environmental correlation feature set, calculate their change rates. The change rate of the surface deformation feature can be calculated by comparing the change rates of the surface elevation at different times. For example, in the previous surface deformation analysis and processing, digital elevation models (DEMs) at different times have been obtained. The change rate of the surface deformation feature can be obtained by calculating the elevation difference between two adjacent DEMs and then dividing it by the time interval. For the stress distribution information in the screened structural damage feature set, calculate its stress gradient change rate. The stress gradient change rate can be obtained by selecting several adjacent pixel points in the stress distribution contour map, calculating the stress difference and the distance difference between them, and then dividing the stress difference by the distance difference. After obtaining the change rate of the surface deformation feature and the change rate of the stress gradient, use them as elements of the feature vector. A clustering analysis algorithm, such as the K-means clustering algorithm, can be used to perform clustering analysis on these feature vectors. The principle of the K-means clustering algorithm is to divide the data points into K clusters, so that the distance between the data points within each cluster is the smallest, while the distance between the data points in different clusters is the largest. Through clustering analysis, find the geographical coordinates corresponding to the cluster center, and use this geographical coordinate as the coordinates of the disaster occurrence location.

[0061] Step S433: According to the diffusion rate parameter and the influence range growth rate parameter in the historical disaster pattern library of each disaster type in the disaster type classification result, perform diffusion trend prediction processing on the current disaster impact range boundary coordinates, and generate the disaster impact diffusion trend prediction result. The historical disaster pattern library is established in advance and contains standard feature templates and related parameter information of various disaster types, such as diffusion rate parameters and influence range growth rate parameters. The diffusion rate parameter represents the distance that the disaster spreads per unit time, and the influence range growth rate parameter represents the growth ratio of the disaster influence range per unit time. The process of the diffusion trend prediction processing is, for example, according to the disaster type classification result, obtain the diffusion rate parameter and the influence range growth rate parameter of the corresponding disaster type from the historical disaster pattern library. For example, if the disaster type classification result is landslide risk, then find the diffusion rate parameter and the influence range growth rate parameter of the landslide disaster from the historical disaster pattern library.

[0062] Then, based on the diffusion rate parameter and the growth rate parameter of the influence range, diffusion simulation is performed on the boundary coordinates of the current disaster influence range. A vector-based diffusion model can be used, such as a buffer analysis model. The principle of the buffer analysis model is to expand based on the boundary of the current disaster influence range according to the diffusion rate parameter and the growth rate parameter of the influence range. Specifically, the boundary coordinates of the current disaster influence range are converted into vector data, such as polygons. Then, according to the diffusion rate parameter and the growth rate parameter of the influence range, the expansion distance and direction are determined. The expansion direction can be determined based on the spatial diffusion direction vector extracted previously. Finally, based on the boundary of the current disaster influence range, expansion is carried out according to the determined distance and direction to generate the boundary of the disaster influence range at a future time point.

[0063] Repeat the above steps to simulate the boundaries of the disaster influence range at different time points and obtain the change of the disaster influence range over time. Organize these simulation results into the prediction results of the disaster influence diffusion trend, including information such as the boundary coordinates, diffusion direction, and diffusion speed of the disaster influence range at different time points.

[0064] Step S434: Based on the preset disaster hazard level mapping table corresponding to the disaster type classification result, assign an initial disaster type priority score to each disaster type identifier. The establishment process of the preset disaster hazard level mapping table is as follows: For example, evaluate the hazard levels of various highway disaster types (such as landslide risks, pavement cracks, bridge structure failures, etc.). The hazard level evaluation needs to consider multiple factors, such as the degree of impact of the disaster on highway traffic, the degree of damage to the surrounding environment, and the economic losses caused. An evaluation team can be invited, consisting of highway engineering experts, geological experts, environmental experts, etc., to score the hazard levels of each disaster type based on historical disaster data and practical experience. Then, according to the scoring results of the hazard levels, the disaster types are classified into different hazard levels, such as low hazard level, medium hazard level, and high hazard level, etc. Set corresponding priority score ranges for each hazard level. For example, the priority score range for the low hazard level can be set as 1 - 3 points, the priority score range for the medium hazard level can be set as 4 - 6 points, and the priority score range for the high hazard level can be set as 7 - 10 points. Finally, associate each disaster type with the corresponding hazard level and priority score to form the preset disaster hazard level mapping table.

[0065] Step S435: Dynamically modify the initial disaster type priority score based on the diffusion rate and the growth rate of the influence range in the disaster impact diffusion trend prediction result to generate a disaster type priority score. For example, the process of dynamic modification is to determine the influence coefficients of the diffusion rate and the growth rate of the influence range on the priority score. The influence coefficients can be set according to historical disaster data and practical experience. For example, the influence coefficient of the diffusion rate can be set as k1, and the influence coefficient of the growth rate of the influence range can be set as k2. The larger the value of the influence coefficient, the greater the influence of this factor on the priority score. Then, calculate the correction value according to the diffusion rate and the growth rate of the influence range in the disaster impact diffusion trend prediction result. For example, the correction value = k1 × diffusion rate + k2 × growth rate of the influence range. Finally, add the correction value to the initial disaster type priority score to obtain the disaster type priority score. For example, assume that the initial disaster type priority score is P0 and the correction value is , then the disaster type priority score P = P0 + .

[0066] Step S436: Correlate and map the coordinates of the disaster occurrence location, the disaster type priority score, and the disaster impact diffusion trend prediction result according to the spatial coordinates to generate a disaster detection unit corresponding to each disaster impact range identifier.

[0067] The process of association mapping, for example, ensures that the coordinates of the disaster occurrence location, the priority score of the disaster type, and the prediction results of the disaster impact diffusion trend all have clear spatial coordinate information. The coordinates of the disaster occurrence location are inherently based on geographical spatial positioning, and the boundary coordinates of the disaster impact scope at different time points in the prediction results of the disaster impact diffusion trend also contain spatial location information. For the priority score of the disaster type, it can be associated with the corresponding coordinates of the disaster occurrence location to endow it with spatial attributes. Then, based on the spatial coordinates of the disaster impact scope identifier, the coordinates of the disaster occurrence location are matched with it. If the coordinates of the disaster occurrence location are within the geographical area defined by the disaster impact scope identifier, the priority score of the disaster type corresponding to the coordinates of the disaster occurrence location and the prediction results of the disaster impact diffusion trend are associated with the disaster impact scope identifier. During the matching process, the spatial query function of the Geographic Information System (GIS) can be used to complete the matching operation by determining whether the point (coordinates of the disaster occurrence location) is within the polygon (the area formed by the boundary coordinates of the disaster impact scope identifier). For each successfully matched combination, the coordinates of the disaster occurrence location, the priority score of the disaster type, and the prediction results of the disaster impact diffusion trend are integrated together to form a complete disaster detection unit. This unit clearly shows the occurrence location, severity, and possible future diffusion of the disaster within the disaster impact scope. For example, a disaster detection unit may include specific longitude and latitude coordinates representing the disaster occurrence location, a specific numerical value representing the priority score of the disaster type, and a series of boundary coordinates of the disaster impact scope and diffusion direction vectors at different time points to represent the prediction results of the disaster impact diffusion trend.

[0068] Step S437: Aggregate the spatial distribution information of all disaster detection units to generate a disaster detection result set containing a set of coordinates of disaster occurrence locations, a sequence of priority scores of disaster types, and a set of prediction results of disaster impact diffusion trends. The process of aggregating the spatial distribution information of all disaster detection units, for example, is to establish three different data structures, which are respectively used to store the set of coordinates of disaster occurrence locations, the sequence of priority scores of disaster types, and the set of prediction results of disaster impact diffusion trends. For the set of coordinates of disaster occurrence locations, a list or an array can be used to store the coordinates of the disaster occurrence location in each disaster detection unit. For example, if the coordinates of the disaster occurrence location are represented by longitude and latitude, each element in the list can be a binary tuple containing longitude and latitude values. For the sequence of priority scores of disaster types, a list or an array can also be used to store the priority scores of the disaster type in each disaster detection unit. The elements in the list are arranged in the order of the disaster detection units, so as to ensure that each priority score corresponds to the corresponding coordinates of the disaster occurrence location and the prediction results of the disaster impact diffusion trend.

[0069] For the prediction result set of the disaster impact diffusion trend, a nested data structure can be used for storage. Since each prediction result of the disaster impact diffusion trend contains information such as the boundary coordinates of the disaster impact scope, the diffusion direction, and the diffusion speed at different time points, the prediction results of the disaster impact diffusion trend for each disaster detection unit can be stored as a sub-list or sub-dictionary, and then all sub-lists or sub-dictionaries are stored in a main list. After establishing the data structure, traverse all disaster detection units, add the coordinates of the disaster occurrence location in each disaster detection unit to the set of disaster occurrence location coordinates, add the disaster type priority score to the sequence of disaster type priority scores, and add the prediction results of the disaster impact diffusion trend to the prediction result set of the disaster impact diffusion trend. By aggregating the spatial distribution information of all disaster detection units, the generated disaster detection result set contains comprehensive information about highway disasters, including the location, severity, and future diffusion trend of the disasters.

[0070] Step S500: Generate a set of inspection optimization strategies based on the disaster detection result set. The set of inspection optimization strategies is used to adjust the collaborative inspection path and monitoring parameter configuration of aerial inspection devices and ground inspection devices. The process of generating the set of inspection optimization strategies is, for example, to determine the resource allocation weights of aerial inspection devices and ground inspection devices according to the disaster type priority scores in the disaster detection result set.

[0071] Then, based on the prediction results of the disaster impact diffusion trend, generate dynamic path adjustment instructions for aerial inspection devices and monitoring parameter update instructions for ground inspection devices. The prediction results of the disaster impact diffusion trend provide the possible diffusion scope and direction of the disaster in the future. According to this information, the inspection path of the aerial inspection device can be dynamically adjusted to cover the area where the disaster may spread; at the same time, update the monitoring parameters of the ground inspection device, such as the sampling frequency of the sensor, the monitoring range, etc., to better monitor the development of the disaster. Next, according to the spatial distribution density of the disaster occurrence location coordinates, adjust the collaborative inspection frequency and monitoring data acquisition resolution of the aerial inspection device and the ground inspection device. If the spatial distribution density of the disaster occurrence location coordinates is high, it indicates that the possibility of disasters occurring in this area is relatively large, and it is necessary to increase the collaborative inspection frequency and monitoring data acquisition resolution to more timely and accurately detect potential disaster hazards. Finally, integrate the resource allocation weights, dynamic path adjustment instructions, monitoring parameter update instructions, and collaborative inspection frequency into a set of inspection optimization strategies and send them to the corresponding inspection device control terminal. After receiving the set of inspection optimization strategies, the inspection device control terminal adjusts the working modes of the aerial inspection device and the ground inspection device according to the instructions and parameters therein to achieve the optimal configuration of the collaborative inspection path and monitoring parameters.

[0072] As an implementation manner, step S500 may specifically include the following steps: Step S510: Determine the resource allocation weights of the aerial inspection equipment and the ground inspection equipment according to the disaster type priority scores in the disaster detection result set. The disaster type priority scores are obtained by comprehensively considering factors such as the harm degree and influence scope of the disasters, and reflect the severity and urgency of different disasters. Reasonable resource allocation weights can ensure that limited inspection resources are preferentially allocated to areas with higher disaster risks, improving the efficiency and effectiveness of highway disaster detection.

[0073] As an implementation manner, step S510 may specifically include the following steps: Step S511: Allocate initial priority scores to each disaster detection unit according to the disaster harm level corresponding to the disaster type identifier. To establish the correspondence between the disaster type identifier and the disaster harm level, a disaster harm level mapping table can be pre-constructed. This mapping table is formulated by analyzing and summarizing a large amount of historical highway disaster data and combining the opinions of highway engineering experts, geological experts, and environmental experts. For example, for the disaster type of landslide risk, since it may cause extensive damage to the highway, traffic interruption, and even endanger the lives and property of surrounding residents, the corresponding disaster harm level in the mapping table is usually set to a high level; while for the disaster type of road surface cracks, although it will also affect the normal use of the highway, the harm degree is relatively small, and its corresponding disaster harm level may be set to a medium level.

[0074] When allocating initial priority scores to each disaster detection unit, find the corresponding disaster harm level from the disaster harm level mapping table according to the disaster type identifier. By allocating initial priority scores to each disaster detection unit according to the disaster harm level corresponding to the disaster type identifier, the severity of different disasters can be initially quantified, providing a basis for subsequent dynamic weighted adjustment in combination with the disaster impact diffusion trend prediction results.

[0075] Step S512: Dynamically adjust the initial priority scores in combination with the diffusion rate and the growth rate of the influence scope in the disaster impact diffusion trend prediction results to generate disaster type priority scores. The diffusion rate represents the distance that the disaster spreads per unit time, and the growth rate of the influence scope represents the growth ratio of the disaster influence scope per unit time. The faster the diffusion rate and the larger the growth rate of the influence scope, the more rapid the development of the disaster and the greater the impact on the highway and the surrounding environment. Therefore, the priority score of the disaster needs to be increased accordingly.

[0076] The process of dynamic weighted adjustment is, for example, to determine the influence coefficients of the diffusion rate and the growth rate of the influence range on the priority score. The determination of the influence coefficients needs to consider multiple factors, such as the type of disaster, the importance of the highway, the sensitivity of the surrounding environment, etc. By analyzing a large amount of historical disaster data and combining expert experience, the influence coefficient of the diffusion rate can be set as k1, and the influence coefficient of the growth rate of the influence range can be set as k2. Then, according to the diffusion rate and the growth rate of the influence range in the prediction result of the disaster influence diffusion trend, the correction value is calculated. Finally, the correction value is added to the initial priority score to obtain the priority score of the disaster type. For example, assume that the initial priority score of a certain disaster detection unit is P0, the diffusion rate is v, and the growth rate of the influence range is r. Then the priority score of the disaster type P = P0 + k1×v + k2×r. By combining the diffusion rate and the growth rate of the influence range in the prediction result of the disaster influence diffusion trend and dynamically weighting and adjusting the initial priority score, the generated priority score of the disaster type can more accurately reflect the actual severity of the disaster, providing a more accurate basis for calculating the resource allocation weights of aerial inspection equipment and ground inspection equipment according to the spatial distribution density of the priority score of the disaster type in the subsequent process.

[0077] Step S513: Calculate the image acquisition frequency weight required for the aerial inspection equipment and the sensor deployment density weight required for the ground inspection equipment according to the spatial distribution density of the priority score of the disaster type. The process of calculating the image acquisition frequency weight and the sensor deployment density weight is, for example, to divide the highway area into several grid units, and each grid unit has an area. The grid division can be carried out using geographic information system (GIS) technology. According to the actual range and accuracy requirements of the highway, the size of the grid unit is determined. For example, the highway area can be divided into square grid units with a side length of 100 meters. Then, the sum of the priority scores of the disaster types in each grid unit is counted. For each disaster detection unit, its belonging grid unit is determined according to its disaster occurrence location coordinates, and the priority score of the disaster type of this disaster detection unit is accumulated into the score sum of the belonging grid unit.

[0078] Next, calculate the spatial distribution density of the disaster type priority scores for each grid cell. The spatial distribution density can be obtained by dividing the sum of the disaster type priority scores within the grid cell by the area of the grid cell. Based on the spatial distribution density of the disaster type priority scores, calculate the image acquisition frequency weights required for the aerial inspection equipment and the sensor deployment density weights required for the ground inspection equipment. A proportionality coefficient can be set, which is used to calculate the image acquisition frequency weights and the sensor deployment density weights respectively. For example, set the proportionality coefficient for the image acquisition frequency weights as m1 and the proportionality coefficient for the sensor deployment density weights as m2. For each grid cell, its image acquisition frequency weight can be obtained by multiplying the spatial distribution density of the disaster type priority scores of this grid cell by m1, and the sensor deployment density weight can be obtained by multiplying the spatial distribution density of the disaster type priority scores of this grid cell by m2. Finally, normalize the image acquisition frequency weights and the sensor deployment density weights of all grid cells to ensure that their sum is 1. The normalization process can be achieved by dividing the weight of each grid cell by the sum of the weights of all grid cells.

[0079] Step S514: Generate resource allocation weights based on the image acquisition frequency weights and the sensor deployment density weights, and dynamically adjust the task scheduling queues of the aerial inspection equipment and the ground inspection equipment. The resource allocation weights can clarify the proportion of resource allocation between the aerial inspection equipment and the ground inspection equipment, and dynamically adjusting the task scheduling queue can ensure that the inspection resources can be allocated to the areas with higher disaster risks in a timely and accurate manner.

[0080] The process of generating the resource allocation weights is, for example, to integrate the image acquisition frequency weights and the sensor deployment density weights. The image acquisition frequency weights and the sensor deployment density weights can be added according to a preset ratio to obtain the resource allocation weight corresponding to each grid cell. For example, assume the image acquisition frequency weight is w1 and the sensor deployment density weight is w2, and a integration coefficient α (0 ≤ α ≤ 1) can be set, then the resource allocation weight w = α × w1 + (1 - α) × w2. The value of the integration coefficient α can be adjusted according to the actual situation. If more emphasis is placed on the image acquisition function of the aerial inspection equipment, the value of α can be set larger; if more emphasis is placed on the sensor monitoring function of the ground inspection equipment, the value of α can be set smaller. Then, normalize the resource allocation weights of all grid cells to ensure that their sum is 1. The normalization process can be achieved by dividing the resource allocation weight of each grid cell by the sum of the resource allocation weights of all grid cells.

[0081] After generating the resource allocation weights, dynamically adjust the task scheduling queues of the aerial inspection equipment and the ground inspection equipment. The task scheduling queues record the inspection task arrangements of the aerial inspection equipment and the ground inspection equipment, including information such as the inspection areas, times, and frequencies. According to the resource allocation weights, increase the inspection tasks in areas with higher disaster risks and reduce the inspection tasks in areas with lower disaster risks. For example, for grid cells with large resource allocation weights, the image acquisition tasks of the aerial inspection equipment and the sensor monitoring tasks of the ground inspection equipment can be increased to improve the inspection frequency; for grid cells with small resource allocation weights, the inspection tasks can be appropriately reduced. At the same time, the working capabilities and resource limitations of the aerial inspection equipment and the ground inspection equipment need to be considered. If the resource allocation weight of a certain area is very large, but the working capabilities of the aerial inspection equipment and the ground inspection equipment are limited and cannot meet the inspection requirements of this area, reasonable task allocation and coordination are required. By optimizing the inspection routes, adjusting the inspection times, etc., the working efficiency of the inspection equipment can be improved to ensure that, under resource limitations, areas with higher disaster risks can be covered as much as possible.

[0082] Step S520: Based on the prediction results of the disaster impact diffusion trend, generate a dynamic path adjustment instruction for the aerial inspection equipment and a monitoring parameter update instruction for the ground inspection equipment. The process of generating the dynamic path adjustment instruction for the aerial inspection equipment is, for example, to determine the areas where the disaster may spread according to the boundary coordinates and diffusion directions of the disaster impact scope at different time points in the prediction results of the disaster impact diffusion trend. The boundary coordinates of the disaster impact scope at different time points can be visually displayed in a Geographic Information System (GIS), and by analyzing the changes in the boundaries and the diffusion directions, the main areas where the disaster may spread can be determined.

[0083] Then, combined with the current position and working capabilities of the aerial inspection equipment, plan a new inspection route. The new inspection route should try to cover the areas where the disaster may spread, and at the same time, factors such as the flight time and endurance of the aerial inspection equipment need to be considered. Path planning algorithms such as the A* algorithm and the Dijkstra algorithm can be used to calculate the optimal inspection route according to the disaster impact scope and the constraint conditions of the inspection equipment. Finally, convert the new inspection route into a dynamic path adjustment instruction, including parameters such as flight direction, flight altitude, and flight speed. Send these instructions to the control terminal of the aerial inspection equipment, and the aerial inspection equipment adjusts the flight path according to the instructions to inspect the areas where the disaster may spread.

[0084] The process of generating the monitoring parameter update instruction for the ground inspection equipment is, for example, to analyze the impact of disaster diffusion on the highway structure and the surrounding environment based on the prediction result of the disaster impact diffusion trend. For example, if the disaster diffusion may lead to situations such as increased ground vibration and greater stress changes, the monitoring parameters of the ground inspection equipment need to be adjusted accordingly. For the sensors of the ground inspection equipment, adjust their sampling frequency and monitoring range. If the risk in the disaster diffusion area increases, increase the sampling frequency of the sensors to capture information such as structural vibration and stress changes more timely; expand the monitoring range to ensure that the areas that the disaster may spread to can be covered. Convert the adjusted monitoring parameters into monitoring parameter update instructions and send them to the control terminal of the ground inspection equipment. The ground inspection equipment updates the monitoring parameters according to the instructions and conducts more effective monitoring of the disaster diffusion area. By generating the dynamic path adjustment instruction for the aerial inspection equipment and the monitoring parameter update instruction for the ground inspection equipment based on the prediction result of the disaster impact diffusion trend, the inspection equipment can better adapt to the development and changes of the disaster, improving the timeliness and accuracy of highway disaster detection.

[0085] Step S530: Adjust the collaborative inspection frequency and the monitoring data acquisition resolution of the aerial inspection equipment and the ground inspection equipment according to the spatial distribution density of the disaster occurrence location coordinates. The spatial distribution density of the disaster occurrence location coordinates reflects the concentration degree of disasters occurring in different regions of the highway. By reasonably adjusting the collaborative inspection frequency and the monitoring data acquisition resolution, the inspection efficiency can be improved, and highway disasters can be detected more accurately.

[0086] The process of adjusting the collaborative inspection frequency is, for example, to divide the highway area into several grid units, and each grid unit has an area. The grid division can be carried out using Geographic Information System (GIS) technology. According to the actual range and accuracy requirements of the highway, determine the size of the grid unit. Then, count the number of disaster occurrence location coordinates within each grid unit. For each disaster detection unit, determine the grid unit to which it belongs according to its disaster occurrence location coordinates, and accumulate the disaster occurrence location coordinates into the coordinate number of the belonging grid unit. Next, calculate the spatial distribution density of the disaster occurrence location coordinates for each grid unit. The spatial distribution density can be obtained by dividing the number of disaster occurrence location coordinates within the grid unit by the area of the grid unit. According to the spatial distribution density of the disaster occurrence location coordinates, adjust the collaborative inspection frequency of the aerial inspection equipment and the ground inspection equipment. For the grid units with a higher spatial distribution density, it indicates that the possibility of disasters occurring in this area is greater, and the collaborative inspection frequency needs to be increased to increase the inspection times; for the grid units with a lower spatial distribution density, the collaborative inspection frequency can be appropriately reduced to decrease the inspection times. A frequency adjustment coefficient can be set, and the collaborative inspection frequency is adjusted according to the product of the spatial distribution density and the frequency adjustment coefficient.

[0087] The process of adjusting the monitoring data acquisition resolution is, for example, also based on the spatial distribution density of the coordinates of the disaster occurrence location. For grid cells with a higher spatial distribution density, the image acquisition resolution of the aerial inspection equipment and the sensor data acquisition resolution of the ground inspection equipment are increased. High-resolution images and data can more clearly reflect the structure and environmental information of the highway, helping to detect potential disaster hazards. For grid cells with a lower spatial distribution density, the monitoring data acquisition resolution can be appropriately reduced to reduce the pressure of data processing and storage. By adjusting the collaborative inspection frequency of the aerial inspection equipment and the ground inspection equipment and the monitoring data acquisition resolution according to the spatial distribution density of the coordinates of the disaster occurrence location, the inspection resources can be reasonably allocated, and the efficiency and accuracy of highway disaster detection can be improved.

[0088] Step S540: Integrate the resource allocation weight, the dynamic path adjustment instruction, the monitoring parameter update instruction, and the collaborative inspection frequency into an inspection optimization strategy set and send it to the corresponding inspection equipment control terminal. The inspection optimization strategy set is a set of instructions for comprehensively optimizing the working modes of the aerial inspection equipment and the ground inspection equipment. By sending it to the inspection equipment control terminal, the inspection equipment can work according to the optimized strategy, improving the efficiency and accuracy of highway disaster detection. The process of integrating the inspection optimization strategy set is, for example, to establish a data structure to store the inspection optimization strategy set. A dictionary or a list can be used for storage, where each element corresponds to an optimization strategy information. For example, a dictionary can be used for storage. The keys of the dictionary can be "resource allocation weight", "dynamic path adjustment instruction", "monitoring parameter update instruction", and "collaborative inspection frequency", etc., and the corresponding values are the respective optimization strategy information.

[0089] Add the resource allocation weight information to the inspection optimization strategy set. The resource allocation weight information can be stored in the form of a list or a matrix, and each element corresponds to the resource allocation weight of a grid cell. Take the resource allocation weight information as the value of the "resource allocation weight" key in the dictionary. Add the dynamic path adjustment instruction to the inspection optimization strategy set. The dynamic path adjustment instruction can include parameters such as flight direction, flight altitude, and flight speed, and is stored in the form of a list or a dictionary. Take the dynamic path adjustment instruction as the value of the "dynamic path adjustment instruction" key in the dictionary. Add the monitoring parameter update instruction to the inspection optimization strategy set. The monitoring parameter update instruction can include parameters such as the sampling frequency and monitoring range of the sensor, and is also stored in the form of a list or a dictionary. Take the monitoring parameter update instruction as the value of the "monitoring parameter update instruction" key in the dictionary. Add the collaborative inspection frequency information to the inspection optimization strategy set. The collaborative inspection frequency information can be stored in the form of a list, and each element corresponds to the collaborative inspection frequency of a grid cell. Take the collaborative inspection frequency information as the value of the "collaborative inspection frequency" key in the dictionary. After completing the integration of the inspection optimization strategy set, send it to the corresponding inspection equipment control terminal. After receiving the inspection optimization strategy set, the inspection equipment control terminal parses the instructions and parameters therein, and adjusts the working mode of the inspection equipment according to this information to achieve the optimal configuration of the collaborative inspection path and monitoring parameters.

[0090] Please refer to Figure 2 , Figure 2Schematic structural diagram of a highway disaster detection system provided by an embodiment of the present invention. The highway disaster detection system is, for example, a server or a computer device in a monitoring and command center, and at least includes a processor 101, a communication interface 102, and a memory 103. Among them, the processor 101, the communication interface 102, and the memory 103 can be connected through a bus or other means. Among them, the processor 101 (or central processing unit (CPU)) is the computing core and control core of the highway disaster detection system, which can parse various instructions in the highway disaster detection system and process various data of the highway disaster detection system. The communication interface 102 may optionally include a standard wired interface, a wireless interface (such as WI-FI, a mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for the transmission and interaction of internal data of the highway disaster detection system. The memory 103 (Memory) is a memory device in the highway disaster detection system, used to store programs and data. It can be understood that the memory 103 here can include both the built-in memory of the highway disaster detection system and, of course, the extended memory supported by the highway disaster detection system. The memory 103 provides a storage space, and the operating system of the highway disaster detection system is stored in this storage space, and the present invention does not limit this. In one embodiment, the processor 101 executes the highway disaster detection method based on air-ground collaborative unmanned patrol provided above in the embodiments of the present invention by running the computer program in the memory 103.

Claims

1. A method for highway disaster detection based on air-ground collaborative unmanned inspection, characterized in that, Including: Obtain a multi-source highway monitoring data set, where the multi-source highway monitoring data set includes a multi-dimensional remote sensing image data set collected by aerial inspection equipment and a multi-dimensional structural sensing data set collected by ground inspection equipment; the multi-dimensional remote sensing image data set contains image area coverage information with different timestamps, and the multi-dimensional structural sensing data set contains structural vibration response information with different spatial positions; Perform spatio-temporal alignment processing on the multi-source highway monitoring data set to generate a highway monitoring data set with fused spatio-temporal tags; Based on a preset multi-modal disaster recognition model, perform disaster feature extraction processing on the highway monitoring data set with fused spatio-temporal tags to obtain a highway disaster feature set; the highway disaster feature set includes environment-related features and structural damage features; Generate a disaster detection result set according to the highway disaster feature set, where the disaster detection result set contains multiple disaster detection units, and each disaster detection unit corresponds to a disaster type identifier and a disaster impact range identifier for the highway area; Generate an inspection optimization strategy set based on the disaster detection result set.

2. The method according to claim 1, characterized in that, The performing spatio-temporal alignment processing on the multi-source highway monitoring data set to generate a highway monitoring data set with fused spatio-temporal tags includes: Extract first spatio-temporal reference information from the multi-dimensional remote sensing image data set, where the first spatio-temporal reference information includes an image acquisition time series and an image geographic coordinate range; Extract second spatio-temporal reference information from the multi-dimensional structural sensing data set, where the second spatio-temporal reference information includes a sensor timestamp series and a sensor positioning coordinate series; Perform time window matching on the first spatio-temporal reference information and the second spatio-temporal reference information to determine the time overlap interval between the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set; Perform data slicing processing on the multi-dimensional remote sensing image data set and the multi-dimensional structural sensing data set according to the time overlap interval to generate a time-synchronized sub-image data set and a sub-sensing data set; Perform spatial coordinate conversion processing on the sub-image data set and the sub-sensing data set, and map the image geographic coordinate range to the spatial coordinate system corresponding to the sensor positioning coordinate series to generate a highway monitoring data set with fused spatio-temporal tags having a unified spatial reference.

3. The method according to claim 2, wherein The performing disaster feature extraction processing on the highway monitoring data set with fused spatio-temporal tags based on a preset multi-modal disaster recognition model to obtain a highway disaster feature set includes: Perform environmental feature extraction processing on the sub-image data set in the highway monitoring data set with fused spatio-temporal tags to generate an environment-related feature set; the environment-related feature set includes surface deformation features, abnormal vegetation coverage features, and water body distribution change features; Perform structural feature extraction processing on the sub-sensing data set in the highway monitoring data set with fused spatio-temporal tags to generate a structural damage feature set; the structural damage feature set includes vibration frequency offset features, abnormal stress distribution features, and material fatigue accumulation features; Input the environmental correlation feature set and the structural damage feature set into the multi-modal disaster recognition model for feature fusion processing to generate a fused feature vector set; Perform disaster pattern matching on the fused feature vector set to determine a highway disaster feature set that matches the disaster patterns in the preset disaster pattern library.

4. The method according to claim 3, wherein Generating a disaster detection result set according to the highway disaster feature set includes: Perform disaster type classification processing on each highway disaster feature in the highway disaster feature set to determine the disaster type identifier corresponding to each highway disaster feature; the disaster type identifier is a landslide risk identifier, a road surface crack identifier, or a bridge structure failure identifier; Determine the disaster impact range identifier corresponding to each highway disaster feature according to the spatial distribution information of the environmental correlation features and structural damage features in the highway disaster feature set; Associate and combine the disaster type identifier and the disaster impact range identifier to generate a disaster detection result set including multiple disaster detection units; Wherein, each disaster detection unit includes the coordinates of the disaster occurrence location, the disaster type priority score, and the prediction result of the disaster impact diffusion trend; Generating an inspection optimization strategy set based on the disaster detection result set includes: Determine the resource allocation weights of the aerial inspection equipment and the ground inspection equipment according to the disaster type priority scores in the disaster detection result set; Generate a dynamic path adjustment instruction for the aerial inspection equipment and a monitoring parameter update instruction for the ground inspection equipment based on the disaster impact diffusion trend prediction result; Adjust the collaborative inspection frequency and the monitoring data acquisition resolution of the aerial inspection equipment and the ground inspection equipment according to the spatial distribution density of the disaster occurrence location coordinates; Integrate the resource allocation weights, the dynamic path adjustment instruction, the monitoring parameter update instruction, and the collaborative inspection frequency into an inspection optimization strategy set and send it to the corresponding inspection equipment control terminal.

5. The method according to claim 4, characterized in that, Performing environmental feature extraction processing on the sub-image data set in the highway monitoring data set with fused spatio-temporal tags to generate an environmental correlation feature set includes: Perform multi-spectral band decomposition processing on the sub-image data set to generate a visible light band data subset, an infrared band data subset, and a radar band data subset; Perform surface deformation analysis processing on the visible light band data subset to extract the surface elevation change rate and the surface crack distribution feature; Perform thermal radiation intensity analysis processing on the infrared band data subset to extract the surface temperature anomaly area and the heat source distribution feature; Perform interferometric synthetic aperture radar processing on the radar band data subset to extract the surface vertical displacement amount and the horizontal displacement vector; Integrate the surface elevation change rate, the surface crack distribution feature, the surface temperature anomaly area, the heat source distribution feature, the surface vertical displacement amount, and the horizontal displacement vector to generate an environmental correlation feature set.

6. The method according to claim 5, wherein Performing structural feature extraction processing on the sub-sensor data set in the highway monitoring data set with fused spatio-temporal tags to generate a structural damage feature set includes: Performing frequency domain conversion processing on the structural vibration response information in the sub-sensor data set to generate a vibration spectrum feature set; Extracting fundamental frequency components, harmonic components and noise components from the vibration spectrum feature set, and calculating fundamental frequency offset and harmonic energy ratio; Performing spatial interpolation processing on the stress distribution information in the sub-sensor data set to generate a stress distribution cloud map, and extracting the maximum stress value and the stress gradient change rate; According to the fundamental frequency offset, harmonic energy proportion, maximum stress value and stress gradient change rate, the vibration frequency offset feature, stress distribution abnormality feature and material fatigue accumulation feature in the structural damage feature set are determined.

7. The method according to claim 6, wherein The step of inputting the environment-related feature set and the structural damage feature set into the multimodal disaster identification model for feature fusion processing to generate a fusion feature vector set includes: Normalizing each environment-related feature in the environment-related feature set to generate a standardized environment feature vector; Normalizing each structural damage feature in the structural damage feature set to generate a standardized structural feature vector; Aligning and splicing the standardized environment feature vector and the standardized structure feature vector according to timestamps and spatial coordinates to generate a spatiotemporally aligned multimodal feature matrix; Through the feature fusion layer in the multimodal disaster identification model, cross-modal correlation analysis is performed on the multimodal feature matrix to generate a fusion feature vector set; The cross-modal correlation analysis includes the calculation of correlation weights between environmental features and structural features and feature dimension reduction processing.

8. The method according to claim 7, wherein The performing disaster pattern matching on the fused feature vector set to determine a highway disaster feature set that matches a disaster pattern in a preset disaster pattern library includes: Acquire multiple historical disaster pattern feature vectors from the preset disaster pattern library, each historical disaster pattern feature vector corresponds to a standard feature template of a disaster type; Calculating a similarity score between each fused feature vector in the fused feature vector set and each historical disaster pattern feature vector; Determine the best matching disaster type corresponding to each fused feature vector according to the similarity score, and generate a preliminary disaster feature set including a disaster type identifier and a matching confidence; The preliminary disaster feature set is subjected to spatial continuity verification processing, disaster features with matching confidence lower than a preset threshold and isolated spatial distribution are eliminated, and a highway disaster feature set is generated.

9. The method according to claim 8, wherein The determining the resource allocation weights of the aerial inspection equipment and the ground inspection equipment according to the disaster type priority scores in the disaster detection result set includes: assigning an initial priority score to each disaster detection unit according to the disaster hazard level corresponding to the disaster type identifier; Combined with the diffusion rate and the growth rate of the impact range in the disaster impact diffusion trend prediction result, the initial priority score is dynamically weighted and adjusted to generate a disaster type priority score; Calculate the image acquisition frequency weight required for the aerial inspection equipment and the sensor deployment density weight required for the ground inspection equipment according to the spatial distribution density of the disaster type priority score; Generate a resource allocation weight based on the image acquisition frequency weight and the sensor deployment density weight, and dynamically adjust the task scheduling queues of the aerial inspection device and the ground inspection device.

10. A highway disaster detection system, characterized in that, Including: A memory in which a computer program is stored; A processor for loading the computer program to implement the road disaster detection method based on collaborative unmanned aerial and ground inspection according to any one of claims 1-9.

Images