Intelligent detection method of asphalt pavement segregation based on multi-source data analysis

Abstract

The present invention proposes an intelligent detection method for asphalt pavement segregation based on multi-source data analysis, which relates to the technical field of road engineering detection, including: determining a road section to be detected, obtaining multi-source data of the road section to be detected; preprocessing the multi-source data; constructing a dual-branch neural network detection model, including a first branch detection model and a second branch detection model, inputting the preprocessed optical data and laser point cloud data into the pre-trained first branch detection model to obtain a gradation segregation distribution result; inputting the pre-processed thermal imaging data into the pre-trained second branch detection model to obtain a temperature segregation distribution result; jointly analyzing the gradation segregation distribution result and the temperature segregation distribution result to generate a comprehensive segregation index; outputting the detection result according to the comprehensive segregation index. The present invention can realize comprehensive, accurate and automated detection of the segregation degree of asphalt pavement, and provide a scientific basis for pavement quality evaluation and maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of road engineering detection, and in particular to an intelligent detection method for asphalt pavement segregation based on multi-source data analysis. Background Art

[0002] Asphalt pavement segregation is a common pavement disease, which mainly manifests in two forms: gradation segregation and temperature segregation. Gradation segregation refers to the uneven distribution of coarse and fine aggregates in the asphalt mixture, which leads to the separation of coarse and fine aggregates on the pavement; temperature segregation is caused by improper temperature control during construction, which causes uneven temperature distribution of the mixture, thereby affecting the compaction and density of the pavement. These segregation phenomena will cause uneven roughness, reduced anti-skid performance, and reduced bearing capacity of the pavement, which will seriously affect the service life of the pavement, increase maintenance costs, and may endanger driving safety.

[0003] At present, asphalt pavement segregation detection mainly relies on manual visual inspection and single detection equipment for evaluation. Traditional manual detection methods are highly subjective, with low detection efficiency and poor reliability of results; and single detection equipment can often only obtain segregation characteristics in a certain aspect. For example, it is difficult to accurately evaluate temperature segregation by relying solely on optical detection equipment, and the use of infrared detection alone cannot fully reflect the gradation segregation status. At the same time, the existing detection technology lacks a systematic data analysis method, making it difficult to conduct in-depth mining and comprehensive evaluation of the collected data, unable to accurately quantify the degree of segregation, and unable to achieve automated and intelligent detection of segregation. Summary of the invention

[0004] In view of this, the present invention proposes an intelligent detection method for asphalt pavement segregation based on multi-source data analysis. By integrating optical data, laser point cloud data and thermal imaging data collected by multi-spectral sensors, lidar and infrared sensors, a dual-branch neural network model is used to detect gradation segregation and temperature segregation respectively, and a comprehensive segregation index is generated through joint analysis, thereby achieving comprehensive, accurate and automated detection of the segregation degree of asphalt pavement, providing a scientific basis for pavement quality evaluation and maintenance.

[0005] The technical solution of the present invention is implemented as follows: The present invention provides an intelligent detection method for asphalt pavement segregation based on multi-source data analysis, comprising:

[0006] S1. Determine the road section to be detected, and obtain multi-source data of the road section to be detected, including optical data obtained by a multispectral sensor, laser point cloud data collected by a laser radar, and thermal imaging data obtained by an infrared sensor;

[0007] S2. Preprocessing of multi-source data, including: feature extraction of optical data to obtain surface particle information reflecting the distribution characteristics of coarse and fine materials on the pavement surface; screening and quality assessment of laser point cloud data to exclude abnormal points and incomplete point groups, and extracting point cloud quality information based on the density and depth distribution characteristics between point clouds; correcting thermal imaging data to obtain pavement temperature field distribution information;

[0008] S3, constructing a dual-branch neural network detection model, including a first branch detection model and a second branch detection model, and pre-training the dual-branch neural network detection model, inputting the pre-processed optical data and laser point cloud data into the pre-trained first branch detection model to obtain a gradation segregation distribution result; inputting the pre-processed thermal imaging data into the pre-trained second branch detection model to obtain a temperature segregation distribution result;

[0009] S4, jointly analyzing the gradation segregation distribution results and the temperature segregation distribution results to generate a comprehensive segregation index for characterizing the segregation degree of the asphalt pavement;

[0010] S5. Output the test results according to the comprehensive segregation index, including the overall segregation status of the asphalt pavement of the road section to be tested and a local test report.

[0011] Preferably, step S2 comprises:

[0012] S21, subjecting the optical data collected by the multispectral sensor to multi-scale filtering and texture enhancement processing, extracting the size distribution and spatial texture characteristics of the surface particles, and obtaining a particle feature map that can be used to identify the difference in the distribution of coarse and fine materials; applying a morphological operator to divide the particle feature map into regions, removing redundant background, and obtaining surface particle information reflecting the distribution of coarse and fine materials on the surface of the road surface;

[0013] S22, using a noise point removal algorithm to detect and remove noise points in the original laser point cloud data; performing quality assessment on the processed laser point cloud data to obtain quality information of the laser point cloud data;

[0014] S23. Automatically gain and background subtraction are performed on thermal imaging data through a radiation correction model to eliminate sensor noise and ambient temperature interference; a normalization correction method is used to compensate for the measurement deviation between hot spots and cold spots, and finally the temperature field distribution information of the road surface to be tested is formed.

[0015] Preferably, in step S22, the noise point removal algorithm includes:

[0016] At each point In the local neighborhood of , the least squares plane is fitted and the covariance matrix of the point cloud in the neighborhood is calculated. ,make and Represent the maximum and minimum eigenvalues ​​of the matrix respectively, then define the point Anisotropy factor:

[0017]

[0018] in, The larger the value, the stronger the directional distribution of the point in the local area, which is suspected to be a local anomaly. When is a suspected noise point. is the preset threshold;

[0019] Through Points initially judged as suspected noise points , by calculating its density flow vector:

[0020]

[0021] in, for Neighborhood points, k is the number of neighborhood points, is the local diffusion coefficient, represents the Euclidean distance;

[0022] like , is the preset density flow threshold, then If it is a noise point, it will be removed; otherwise, it will be retained.

[0023] Preferably, in step S22, the quality assessment steps are as follows:

[0024] For the point cloud data after noise removal, a sliding window method is used to divide the point cloud data into N x A local area adjacent to a space;

[0025] In each local area a, for multiple scales Calculating fractal dimension :

[0026]

[0027] in, Represents the cumulative count of points contained in the search radius r in the local area a; when When the value deviates from the preset range or shows abnormal jump, it indicates that there is a mutation or uneven distribution in area a;

[0028] Perform the least squares plane fitting on the local area a and calculate the residual between each point and the plane , calculate the curvature residual factor :

[0029]

[0030] in, represents the number of points in the local area a, for The gradient value of is the adjustment factor; The larger the value, the more uneven the surface is in the local area a;

[0031] Calculate the comprehensive factor of point cloud quality in local area a :

[0032]

[0033] in, is the average value of fractal dimension at multiple scales, , is the weighting coefficient;

[0034] The comprehensive factor of the point cloud quality of all local areas is calculated, and each local area is given a Q value. x The Q values ​​are summarized and a point cloud quality distribution map is generated as the quality information of the laser point cloud data.

[0035] Preferably, the first branch detection model includes:

[0036] The input layer receives the pre-processed optical data and laser point cloud data, and sends the optical data and laser point cloud data to two independent processing channels respectively;

[0037] The optical processing pipeline includes: multi-level convolutional network layers for step-by-step feature extraction of optical data; residual connections for preserving multi-scale information between convolutional layers; an attention module used in conjunction with the last convolutional layer; and a set of fully connected layers for transforming and encoding the extracted features within the pipeline.

[0038] The point cloud processing pipeline includes: a multi-layer perceptron module that performs preliminary mapping of the input point cloud data; a graph neural network layer that uses points in the local area as nodes and quality information as node features, and updates features based on the connection relationship between points; a global pooling layer that merges graph-level features; and a set of fully connected layers that perform the final transformation on the merged features.

[0039] The multimodal fusion layer performs multi-layer nonlinear fusion of the features output by the optical processing channel and the point cloud processing channel;

[0040] At the output layer, the Softmax classifier is used to output the classification information and classification probability distribution.

[0041] Preferably, the first branch detection model processes the optical data and the laser point cloud data as follows to generate a gradation segregation distribution result:

[0042] Receive the pre-processed optical data, divide it into local areas according to the laser point cloud data, process it uniformly, and transmit it to the optical processing channel through the input layer. At the same time, read the laser point cloud data of the corresponding local area and transmit it to the point cloud processing channel through the input layer.

[0043] In the multi-level convolutional network layer, the optical data is convolved and pooled layer by layer to extract deep representations including texture and particle distribution information; the attention module is connected after the last convolution layer to generate the attention and inhibition weight matrix; the features output by the attention module are input into the fully connected layer to complete the high-dimensional feature encoding of the optical processing channel;

[0044] In the multi-layer perceptron module, the salient features of the point cloud are initially mapped. The graph neural network layer iteratively updates the node features according to the topological relationship between points or grids to extract higher-order spatial structure information. In the global pooling layer, the features are reduced in dimension to obtain a compact point cloud representation. The fully connected layer is used to transform the dimension of the point cloud representation and map it into a fixed-size vector.

[0045] In the multimodal fusion layer, the output vectors of the optical processing channel and the point cloud processing channel are feature concatenated, and the fused vectors are compressed and reorganized through multiple layers of nonlinear transformation.

[0046] The reorganized vector is input into the Softmax classifier to calculate the probability that each local area belongs to a different gradation segregation level. The gradation segregation probabilities of all local areas are integrated to generate the gradation segregation distribution result.

[0047] Preferably, the second branch detection model includes:

[0048] The input layer receives the preprocessed thermal imaging data and divides the thermal imaging data into N spatial dimensions. x local regions, and input the thermal imaging data of each local region into the subsequent network in a sequence form;

[0049] The spatiotemporal feature extraction module includes: at least two 3D convolutional layers to represent the features of thermal imaging data in spatial and temporal dimensions; residual connections to retain multi-scale spatiotemporal information; and local pooling layers to aggregate the extracted features;

[0050] The multi-scale attention module sets several channel attentions of different scales on the acquired features and includes a learnable weight gating mechanism to assign attention weights to the features;

[0051] The dynamic analysis layer uses an LSTM network to process the features output by the multi-scale attention module in the same sequence frame and combine the time series outputs;

[0052] The output layer uses several fully connected layers to perform multi-level judgments on the output of the dynamic analysis layer and generate the temperature segregation distribution results of the thermal imaging data.

[0053] Preferably, the second branch detection model processes the thermal imaging data as follows to obtain the temperature segregation distribution result:

[0054] Read the preprocessed thermal imaging data, perform dimension correction on it, and divide it into N x The local areas are converted into sequence form; the thermal imaging data of each local area is taken as a minimum input unit and input into the subsequent network modules accordingly;

[0055] The thermal imaging sequence is input into the spatiotemporal feature extraction module, and its joint distribution features in height, width and time dimensions are captured through at least two layers of 3D convolutional layers; the spatiotemporal information of different scales extracted by multiple layers is retained by using residual connections; and the features are aggregated by using local pooling layers;

[0056] The multi-scale attention module is used to weight different spatial positions or time segments in the features, and the weighted features are input into the dynamic analysis layer to capture the evolution law of thermal imaging in time-continuous frames, and the output of the LSTM network is combined to form a comprehensive thermal feature vector;

[0057] The comprehensive thermal feature vector is sent to the output layer, and several fully connected layers are used to map the features to the multi-level category space of temperature segregation. The Sigmoid output function is used to obtain the segregation probability of each local area.

[0058] The segregation probability of each local area is integrated to generate the temperature segregation distribution result.

[0059] Preferably, step S4 comprises:

[0060] S41. Standardize the gradation segregation distribution results and map the gradation segregation probability of each local area to the [0,1] interval as the gradation segregation index of each local area. ; Perform the same standardization operation on the temperature segregation distribution results to obtain the temperature segregation index of each local area ;

[0061] S42. Calculate the local proportion factor of each local area :

[0062]

[0063] In the formula, a represents the local area index, is the adjustment coefficient; The larger the value, the higher the degree of segregation in the local area;

[0064] S43, calculating the local comprehensive segregation index of each local area: in, , is a hypercurve parameter greater than 0, used to control the nonlinear degree of segregation strength growth. The closer the value is to 1, the more severe the segregation is;

[0065] S44. Calculate the global comprehensive segregation index:

[0066] Will The spatial distribution of the segregation is plotted into a segregation heat map, and the global comprehensive segregation index is output , characterizing the overall segregation degree of the road section to be detected.

[0067] Preferably, the local proportion factor The adjustment coefficient Dynamic adjustment based on historical engineering data, including:

[0068] By analyzing the grading segregation index of different road sections and temperature segregation index Statistical analysis to determine The initial value of

[0069] Dynamically adjust the severity of road segment segregation according to its distribution pattern. The value range of The response to high segregation areas is more sensitive.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] The present invention realizes comprehensive, accurate and automated detection of asphalt pavement segregation by integrating multi-source data collected by multispectral sensors, laser radar and infrared sensors, and combining with a dual-branch neural network detection model. The method can simultaneously acquire and analyze the optical characteristics, point cloud characteristics and temperature characteristics of the pavement, generate a comprehensive segregation index, and comprehensively characterize the degree of pavement segregation, thereby effectively overcoming the limitation of the existing technology that a single detection method is difficult to fully reflect the segregation status, and provides a scientific basis for pavement quality evaluation and maintenance;

[0072] The present invention extracts surface particle information, point cloud quality information and temperature field distribution information respectively by preprocessing optical data, laser point cloud data and thermal imaging data, thus ensuring high-quality input of multi-source data. Through multimodal fusion analysis, it can comprehensively reflect the distribution of coarse and fine materials on the pavement surface, the spatial structure characteristics of the point cloud and the temperature distribution characteristics, thus improving the comprehensiveness and accuracy of segregation detection;

[0073] The first branch detection model of the present invention uses a multi-level convolutional network, residual connection and attention module to extract deep features from optical data and laser point cloud data, and realizes joint analysis of optical data and point cloud data through a multimodal fusion layer to generate gradation segregation distribution results. This method can accurately identify the distribution difference of coarse and fine materials in the road surface, quantify the degree of gradation segregation, and solve the problem of insufficient accuracy of gradation segregation detection in the prior art;

[0074] The second branch detection model of the present invention adopts a spatiotemporal feature extraction module, a multi-scale attention module and a dynamic analysis layer for thermal imaging data, which can capture the spatial distribution characteristics and time series characteristics of thermal imaging data and generate temperature segregation distribution results. Through dynamic analysis of thermal imaging data and multi-scale attention allocation, the accuracy and robustness of temperature segregation detection are significantly improved, overcoming the problem of insufficient accuracy of temperature segregation detection in the prior art;

[0075] The present invention generates a comprehensive segregation index by jointly analyzing the gradation segregation distribution results and the temperature segregation distribution results, and intuitively displays the segregation degree of the pavement through a segregation thermogram. The comprehensive segregation index can quantify the overall and local segregation of the pavement, provide accurate segregation evaluation results, and solve the problem of lack of means for quantifying the degree of segregation in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0077] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0078] Figure 2 A network structure diagram of a first branch detection model according to an embodiment of the present invention;

[0079] Figure 3 This is a network structure diagram of the second branch detection model of an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] like Figure 1 As shown, the present invention provides an intelligent detection method for asphalt pavement segregation based on multi-source data analysis, comprising:

[0082] S1. Determine the road section to be detected, and obtain multi-source data of the road section to be detected, including optical data obtained by a multispectral sensor, laser point cloud data collected by a laser radar, and thermal imaging data obtained by an infrared sensor;

[0083] S2. Preprocessing of multi-source data, including: feature extraction of optical data to obtain surface particle information reflecting the distribution characteristics of coarse and fine materials on the pavement surface; screening and quality assessment of laser point cloud data to exclude abnormal points and incomplete point groups, and extracting point cloud quality information based on the density and depth distribution characteristics between point clouds; correcting thermal imaging data to obtain pavement temperature field distribution information;

[0084] S3, constructing a dual-branch neural network detection model, including a first branch detection model and a second branch detection model, and pre-training the dual-branch neural network detection model, inputting the pre-processed optical data and laser point cloud data into the pre-trained first branch detection model to obtain a gradation segregation distribution result; inputting the pre-processed thermal imaging data into the pre-trained second branch detection model to obtain a temperature segregation distribution result;

[0085] S4, jointly analyzing the gradation segregation distribution results and the temperature segregation distribution results to generate a comprehensive segregation index for characterizing the segregation degree of the asphalt pavement;

[0086] S5. Output the test results according to the comprehensive segregation index, including the overall segregation status of the asphalt pavement of the road section to be tested and a local test report.

[0087] Specifically, in one embodiment of the present invention, step S1 includes:

[0088] Determine the road section to be tested: According to the construction plan or maintenance requirements, select the asphalt road section to be tested. The road section to be tested can be a newly paved asphalt road surface or an existing road surface that needs to be evaluated or repaired.

[0089] After determining the road section, the detection area is divided into a number of spatially adjacent local areas. In this embodiment, the road section can be divided into N local areas at this step.x In the subsequent preprocessing, model processing and comprehensive analysis, the local area is used as the processing unit.

[0090] In order to comprehensively evaluate the road segregation, a variety of sensor devices are used to collect data on the road section, including the following three types of data:

[0091] Optical data: Use a multispectral sensor to scan the road surface and collect optical data that reflects the distribution characteristics of the particles on the surface of the road surface. The multispectral sensor can capture light reflection information in different bands, especially visible light and near-infrared bands, to extract the distribution characteristics of coarse and fine particles on the surface of the road surface.

[0092] Laser point cloud data: Use laser radar equipment to scan the road surface and collect high-precision laser point cloud data. Laser point cloud data can reflect the texture depth and density distribution of the road surface, providing key three-dimensional spatial information for the detection of gradation segregation.

[0093] Thermal imaging data: Use infrared sensors to perform thermal imaging scans on the road surface and collect temperature distribution data on the road surface. Thermal imaging data can reflect the distribution of the road surface temperature field and is used to identify temperature segregation areas.

[0094] Arrangement and operation of data acquisition equipment:

[0095] Multispectral sensor: Install the multispectral sensor on the inspection vehicle or drone and drive or fly along the road section to ensure that the entire inspection area is covered. During the collection process, the sampling frequency and resolution of the sensor should be adjusted according to the length of the road section and the detection accuracy requirements.

[0096] LiDAR equipment: LiDAR equipment can also be installed on the inspection vehicle to collect point cloud data while driving along the road section. To ensure the integrity of the point cloud data, the scanning angle and sampling density of the LiDAR should be optimized according to the changes in road width and texture depth.

[0097] Infrared sensor: The infrared sensor is installed on the inspection vehicle or drone to collect thermal imaging data of the road in real time. To reduce the interference of ambient temperature on data collection, the sensor should be calibrated in real time during the collection process, and suitable weather conditions should be selected as much as possible (such as avoiding strong winds or direct sunlight).

[0098] The collected multi-source data is stored in a unified data management system, and a preliminary check is performed on the data integrity and quality. The check includes: whether the resolution of the optical data meets the requirements, whether the coverage of the laser point cloud data is complete, and whether there is obvious noise or outliers in the thermal imaging data.

[0099] Specifically, in one embodiment of the present invention, step S2 includes:

[0100] S21. The optical data collected by the multispectral sensor are subjected to multi-scale filtering and texture enhancement processing to extract the size distribution and spatial texture characteristics of the surface particles, and obtain a particle feature map that can be used to identify the difference in the distribution of coarse and fine materials; the particle feature map is divided into regions using morphological operators to remove redundant background, and the surface particle information reflecting the distribution of coarse and fine materials on the road surface is obtained.

[0101] In this embodiment, the multi-scale filtering process uses a Gaussian filter or a bilateral filter to perform smoothing on texture features of different scales while retaining particle edge information. The filtered data can effectively reduce interference caused by factors such as ambient light and shadows in optical data.

[0102] After filtering, the surface texture features of the optical data are enhanced using a texture enhancement algorithm. Texture enhancement highlights the spatial texture features of the particles by calculating the gradient changes and frequency distribution of the local area, and generates a texture enhancement map that reflects the particle distribution. The texture enhancement map can more clearly show the spatial differences in the distribution of coarse and fine materials.

[0103] Based on the texture enhancement image, the image segmentation algorithm is used to segment the particle area and extract the size distribution characteristics of the particles. The segmentation algorithm can use a threshold-based segmentation method or a region growth-based segmentation method to ensure that the boundaries of the particle area are clear. The extracted particle size distribution characteristics include the average size of the particles, the variance of the particle size, and the distribution density of the particles.

[0104] The spatial texture features of particles are extracted by calculating the gray level co-occurrence matrix (GLCM) of the local area. The spatial texture features include indicators such as particle contrast, correlation, energy and uniformity, which are used to characterize the distribution pattern of particles and the difference between coarse and fine materials.

[0105] Morphological operators are applied to the particle feature map, including dilation, erosion, opening and closing operations, to remove redundant background information (such as shadows, cracks, etc.) in the optical data and retain significant areas related to the distribution of coarse and fine particles on the pavement surface. Based on the particle feature map after morphological processing, a regional partitioning algorithm is used to partition the pavement surface to generate a particle region map reflecting the distribution of coarse and fine particles.

[0106] The particle size distribution characteristics, spatial texture characteristics and regional division results are integrated to generate surface particle information. The surface particle information includes the average size of the particles, the distribution density of the particles, the spatial texture characteristics of the particles and the spatial position of the particle area.

[0107] S22. Use a noise point removal algorithm to detect and remove noise points in the original laser point cloud data; perform quality assessment on the processed laser point cloud data to obtain quality information of the laser point cloud data.

[0108] In this embodiment, the noise point removal algorithm includes:

[0109] At each point In the local neighborhood of , the least squares plane is fitted and the covariance matrix of the point cloud in the neighborhood is calculated. ,make and Represent the maximum and minimum eigenvalues ​​of the matrix respectively, then define the point Anisotropy factor:

[0110]

[0111] in, The larger the value, the stronger the directional distribution of the point in the local area, which is suspected to be a local anomaly. When is a suspected noise point. is the preset threshold;

[0112] Through Points initially judged as suspected noise points , by calculating its density flow vector:

[0113]

[0114] in, for Neighborhood points, k is the number of neighborhood points, is the local diffusion coefficient, represents the Euclidean distance;

[0115] like , is the preset density flow threshold, then If it is a noise point, it will be removed; otherwise, it will be retained.

[0116] Specifically, before removing noise points, some isolated points can be removed according to the local neighborhood density. The process is as follows: for each point p(i), count the number of neighboring points N(i) within the set neighborhood radius r, and calculate the local area density ρ(i)=N(i) / V(r), where V(r) is the volume of the neighborhood radius r. If ρ(i) is less than the preset density threshold, it is removed as an isolated point.

[0117] This embodiment uses local anisotropy factors and density flow vector verification to remove points with abnormal geometric shapes and ensure the geometric consistency of point cloud data. While removing noise points, points with practical significance in the point cloud are retained, avoiding the loss of valid data caused by excessive removal. The algorithm can complete the noise removal of large-scale point cloud data in a relatively short time through local neighborhood calculation and geometric feature analysis.

[0118] The steps of quality assessment are as follows:

[0119] For the point cloud data after noise removal, a sliding window method is used to divide the point cloud data into N x A local area adjacent to a space;

[0120] In each local area a, for multiple scales Calculating fractal dimension :

[0121]

[0122] in, Represents the cumulative count of points contained in the search radius r in the local area a; when When it deviates from the preset interval or shows abnormal jumps, it means that there is a mutation or uneven distribution in area a; the introduction of fractal dimension enhances the algorithm's ability to describe the complex geometric characteristics of point clouds.

[0123] Perform the least squares plane fitting on the local area a and calculate the residual between each point and the plane , calculate the curvature residual factor :

[0124]

[0125] in, represents the number of points in the local area a, for The gradient value of is the adjustment factor; The larger the value, the more uneven the surface is in the local area a. The curvature residual factor can quantify the degree of surface unevenness in the local area and reflect the flatness characteristics of the point cloud.

[0126] Calculate the comprehensive factor of point cloud quality in local area a :

[0127]

[0128] in, is the average value of fractal dimension at multiple scales, , is the weighting coefficient;

[0129] The comprehensive factor of the point cloud quality of all local areas is calculated, and each local area is given a Q value. x The Q values ​​are summarized and a point cloud quality distribution map is generated as the quality information of the laser point cloud data.

[0130] In this embodiment, the laser point cloud data is firstly partitioned locally using a sliding window to divide the entire point cloud into multiple spatially adjacent local areas. , a is regarded as its regional index, and each partition a has points, partition a can be regarded as a partition consisting of sampling points The size w and step size s of the sliding window are set according to the road surface characteristics (such as point cloud sampling density and detection accuracy) to ensure that each partition contains at least a sufficient number of points.

[0131] In each local area Perform multi-scale analysis on point cloud data: define different search radii , whose values ​​from minimum to maximum correspond to different scales from local area to global area. , based on different radii r, a neighborhood search algorithm is used to determine the number of points N(r) in the neighborhood. The point set is divided into multiple subsets according to different radii r, for example: .

[0132] In each local area For each scale r, extract the fractal dimension , fractal dimension Describes the complexity of point cloud distribution and reflects whether there is local point cloud sparseness or anomaly. Calculated at multiple scales , forming a sequence of fractal dimensions , which is used to evaluate the sparsity and uneven distribution of point clouds at different scales. At each scale r, the local area is calculated Point cloud density within : . is the number of points in the neighborhood with radius r; is the volume of the neighborhood with radius r. The changing trends at different scales r reflect the changes in the point cloud density of the region, providing a basis for noise analysis and flatness detection.

[0133] Fit a plane on the point cloud neighborhood and calculate the residual error of each point from the fitted plane: , Represents the depth function value of the fitted plane. The curvature residual factor is then calculated to quantify the flatness. High values ​​in the point cloud indicate that the surface is irregular or uneven.

[0134] Combined with the extracted features, the comprehensive factor of point cloud quality is calculated, and the Q values ​​of all partitions are summarized to generate the quality distribution map of the overall point cloud: high-quality areas are characterized by small Q(a), indicating that the point cloud is evenly distributed and the depth is stable. Low-quality areas are characterized by large Q(a), indicating that there are sparse point clouds, abnormal depth or uneven surfaces.

[0135] S23. Automatically gain and background subtraction are performed on thermal imaging data through a radiation correction model to eliminate sensor noise and ambient temperature interference; a normalization correction method is used to compensate for the measurement deviation between hot spots and cold spots, and finally the temperature field distribution information of the road surface to be tested is formed.

[0136] In this embodiment, the radiation correction model is used to automatically adjust the gain of the thermal imaging data to enhance the contrast between hot spots and cold spots in the image; the temperature value distribution of the thermal imaging data is made more uniform by dynamically adjusting the gain parameters. The radiation correction model is used to analyze the background temperature distribution in the thermal imaging data; the background subtraction algorithm is used to remove the influence of the ambient temperature from the thermal imaging data, retaining the effective information of the road surface temperature.

[0137] The normalization correction method is used to map the temperature values ​​of thermal imaging data to a unified standardized interval; the temperature values ​​of hot spots and cold spots are dynamically adjusted to compensate for measurement deviations and ensure the accuracy of temperature distribution. The temperature value of each pixel T(x,y) is normalized: T norm (x,y)=(T(x,y)-T min ) / (T max -T min ), where: T(x,y) is the temperature value of the pixel in the original thermal imaging data; T min 、T max : The lowest and highest temperature values ​​in the thermal imaging data; T norm : Normalized temperature value, ranging from [0,1].

[0138] In each local area The temperature distribution of the area is represented by a certain number of pixels. In the local area, the temperature features are extracted, including the average temperature :

[0139]

[0140] For Region The number of pixels in

[0141] Temperature variance :

[0142]

[0143] Temperature variance reflects the uniformity of temperature distribution in a region.

[0144] The temperature characteristics of all local areas are summarized to generate a temperature field distribution map: the temperature field distribution map intuitively displays the spatial distribution of road surface temperature in the form of a color gradient; high temperature areas (hot spots) and low temperature areas (cold spots) are clearly visible in the map.

[0145] Specifically, Figure 2 As shown, in one embodiment of the present invention, the first branch detection model includes:

[0146] The input layer receives the pre-processed optical data and laser point cloud data, and sends the optical data and laser point cloud data to two independent processing channels respectively;

[0147] The optical processing pipeline includes: multi-level convolutional network layers for step-by-step feature extraction of optical data; residual connections for preserving multi-scale information between convolutional layers; an attention module used in conjunction with the last convolutional layer; and a set of fully connected layers for transforming and encoding the extracted features within the pipeline.

[0148] The point cloud processing pipeline includes: a multi-layer perceptron module that performs preliminary mapping of the input point cloud data; a graph neural network layer that uses points in the local area as nodes and quality information as node features, and updates features based on the connection relationship between points; a global pooling layer that merges graph-level features; and a set of fully connected layers that perform the final transformation on the merged features.

[0149] The multimodal fusion layer performs multi-layer nonlinear fusion of the features output by the optical processing channel and the point cloud processing channel;

[0150] At the output layer, the Softmax classifier is used to output the classification information and classification probability distribution.

[0151] The first branch detection model processes the optical data and laser point cloud data as follows to generate the gradation segregation distribution results:

[0152] Receive the pre-processed optical data, divide it into local areas according to the laser point cloud data, process it uniformly, and transmit it to the optical processing channel through the input layer. At the same time, read the laser point cloud data of the corresponding local area and transmit it to the point cloud processing channel through the input layer.

[0153] In the multi-level convolutional network layer, the optical data is convolved and pooled layer by layer to extract deep representations including texture and particle distribution information; the attention module is connected after the last layer of convolution to generate the attention and inhibition weight matrix; the features output by the attention module are input into the fully connected layer to complete the high-dimensional feature encoding of the optical processing channel; Among them, the multi-level convolutional network layer includes 5 layers of 2D convolutional layers, that is, two-dimensional convolutional layers.

[0154] In the multi-layer perceptron module, the salient features of the point cloud are preliminarily mapped, and the graph neural network layer iteratively updates the node features according to the topological relationship between points or grids to extract higher-order spatial structure information; in the global pooling layer, the features are reduced in dimension to obtain a compact point cloud representation; the fully connected layer is used to transform the dimension of the point cloud representation and map it into a fixed-size vector; among them, the multi-layer perceptron module includes two multi-layer perceptrons, namely MLP1 and MLP2.

[0155] In the multimodal fusion layer, the output vectors of the optical processing channel and the point cloud processing channel are feature concatenated, and the fused vectors are compressed and reorganized through multiple layers of nonlinear transformation.

[0156] The reorganized vector is input into the Softmax classifier to calculate the probability that each local area belongs to a different gradation segregation level, and the gradation segregation probabilities of all local areas are integrated to generate the gradation segregation distribution results. The Softmax classifier uses a normalized exponential function (Softmax) for probability calculation.

[0157] In this embodiment, when the first branch detection model processes the optical data and the point cloud data, they are divided into N local areas. x partitions, for each local area After data processing, the gradation segregation probabilities of all local areas are summarized and corresponded to the spatial positions of the areas to form the gradation segregation distribution results.

[0158] In this embodiment, the pre-training process of the first branch detection model is as follows:

[0159] The optical data and point cloud data are divided into local areas to generate training sets and validation sets. The label of each local area is the segregation level (such as normal, slight segregation, severe segregation).

[0160] Forward propagation: Optical data and point cloud data are processed through their own processing channels to extract deep features. In the multimodal fusion layer, the feature vectors of the two data are concatenated and fused. Finally, the probability distribution of the separation level is output through the Softmax classifier.

[0161] The loss function uses the cross entropy loss function L CE The Adam optimizer is used, the initial learning rate is 0.001, and the weight decay coefficient is 0.0001. The cosine annealing learning rate scheduler is used to dynamically adjust the learning rate according to the training round.

[0162] Calculate the classification accuracy and F1 score on the validation set to evaluate the model performance. Adjust the model parameters based on the performance of the validation set.

[0163] In this embodiment, the first branch detection model uses the complementarity of optical data and point cloud data through joint processing, thereby improving the detection accuracy of gradation segregation. The introduction of an attention module in the optical processing channel can dynamically focus on the key areas of coarse and fine material distribution, improving the model's ability to extract texture and particle features. The use of a graph neural network layer in the point cloud processing channel can capture the high-order spatial relationship of the point cloud and improve the ability to understand the point cloud data.

[0164] Specifically, Figure 3 As shown, in one embodiment of the present invention, the second branch detection model includes:

[0165] The input layer receives the preprocessed thermal imaging data and divides the thermal imaging data into N spatial dimensions. x local regions, and input the thermal imaging data of each local region into the subsequent network in a sequence form;

[0166] The spatiotemporal feature extraction module includes: at least two 3D convolution layers, i.e., three-dimensional convolution layers, which represent the features of thermal imaging data in spatial and temporal dimensions; residual connections to retain multi-scale spatiotemporal information; and local pooling layers to aggregate the extracted features;

[0167] The multi-scale attention module sets several channel attentions of different scales on the acquired features and includes a learnable weight gating mechanism to assign attention weights to the features;

[0168] The dynamic analysis layer uses the LSTM network, i.e. the long short-term memory network, to process the features output by the multi-scale attention module in the same sequence frame and combine the time series outputs;

[0169] The output layer uses several fully connected layers to perform multi-level judgments on the output of the dynamic analysis layer and generate the temperature segregation distribution results of the thermal imaging data.

[0170] The second branch detection model processes the thermal imaging data as follows to obtain the temperature segregation distribution results:

[0171] Read the preprocessed thermal imaging data, perform dimension correction on it, and divide it into N x The local areas are converted into sequence form; the thermal imaging data of each local area is taken as a minimum input unit and input into the subsequent network modules accordingly;

[0172] The thermal imaging sequence is input into the spatiotemporal feature extraction module, and its joint distribution features in height, width and time dimensions are captured through at least two layers of 3D convolutional layers; the spatiotemporal information of different scales extracted by multiple layers is retained by using residual connections; and the features are aggregated by using local pooling layers;

[0173] The multi-scale attention module is used to weight different spatial positions or time segments in the features, and the weighted features are input into the dynamic analysis layer to capture the evolution law of thermal imaging in time-continuous frames, and the output of the LSTM network is combined to form a comprehensive thermal feature vector;

[0174] The comprehensive thermal feature vector is sent to the output layer, and several fully connected layers are used to map the features to the multi-level category space of temperature segregation. The Sigmoid output function is used to obtain the segregation probability of each local area.

[0175] The segregation probability of each local area is integrated to generate the temperature segregation distribution result.

[0176] In this embodiment, the pre-training process of the second branch detection model is as follows:

[0177] The thermal imaging data is divided into local areas to generate training sets and validation sets. The label of each local area is the temperature segregation level (such as normal, slight segregation, severe segregation).

[0178] Forward propagation: The thermal imaging data is extracted through the 3D convolution layer to extract spatiotemporal features. The dynamic changes of the thermal imaging data are captured through the LSTM network. Finally, the probability distribution of the temperature segregation level is output through the fully connected layer.

[0179] The loss function uses the weighted cross entropy loss function L WCE The RMSprop optimizer is used, the initial learning rate is 0.0005, and the momentum coefficient is 0.9. The exponential decay learning rate scheduler is used to decay the learning rate every 10 epochs.

[0180] Calculate the classification accuracy, F1 score, and AUC value on the validation set to evaluate the model performance. Adjust the model parameters based on the performance of the validation set.

[0181] In this embodiment, the second branch detection model captures the joint features of spatial and temporal dimensions through 3D convolution, and then captures the dynamic changes of time series through LSTM, which improves the detection ability of temperature segregation. Dynamically allocating attention weights can better focus on key spatial positions and time segments, improving the feature extraction ability of the model. Introducing residual connections in the spatiotemporal feature extraction module avoids information loss in deep networks and improves the stability of the model.

[0182] Specifically, in another embodiment of the present invention, the two branch detection models can be jointly trained. During training, the joint loss function is defined as: The AdamW optimizer is used, with an initial learning rate of 0.0001 and a weight decay coefficient of 0.01. Joint training process: The two models are trained alternately to ensure that the feature extraction capabilities of both models are improved simultaneously. The independent performance and joint performance of the two models (such as the accuracy of the comprehensive segregation index) are evaluated on the validation set.

[0183] Specifically, in one embodiment of the present invention, step S4 includes:

[0184] S41. Standardize the gradation segregation distribution results and map the gradation segregation probability of each local area to the [0,1] interval as the gradation segregation index of each local area. ,in, is the probability of gradation segregation in local area a, , is the minimum and maximum value of the gradation segregation probability; the same standardization operation is performed on the temperature segregation distribution results to obtain the temperature segregation index of each local area , is the temperature segregation probability of local area a, , is the minimum and maximum value of temperature segregation probability; , The closer the value is to 1, the more severe the segregation is.

[0185] S42. Calculate the local proportion factor of each local area :

[0186]

[0187] In the formula, a represents the local area index, is the adjustment coefficient; The larger the value, the higher the segregation degree of the local area, and the greater its contribution to the overall segregation analysis. The adjustment coefficient Dynamic adjustment based on historical engineering data, including:

[0188] By analyzing the grading segregation index of different road sections and temperature segregation index Statistical analysis to determine The initial value of

[0189] Dynamically adjust the severity of road segment segregation according to its distribution pattern. The value range of The response to high segregation areas is more sensitive.

[0190] S43, calculating the local comprehensive segregation index of each local area:

[0191] in, , is a hypercurve parameter greater than 0, used to control the nonlinear degree of segregation strength growth. The closer the value is to 1, the more severe the segregation is; , The value of is adjusted experimentally to ensure Sensitivity to areas of high segregation.

[0192] S44. Calculate the global comprehensive segregation index:

[0193] Will The spatial distribution of the segregation is plotted into a segregation heat map, and the global comprehensive segregation index is output , characterizing the overall segregation degree of the road section to be detected. The closer the value is to 1, the more serious the overall segregation of the road section to be detected is.

[0194] Local Composite Segregation Index The spatial distribution of the road is plotted into a segregation heat map to intuitively display the segregation distribution of the road section. Output The value of is used as a quantitative indicator of the overall segregation degree of the road section to be detected.

[0195] The present invention generates a comprehensive segregation index by jointly analyzing the distribution results of gradation segregation and temperature segregation, which can comprehensively characterize the segregation degree of the road section. The calculation of the local comprehensive segregation index and the local proportion factor can accurately locate the area with severe segregation, providing a basis for subsequent maintenance and repair. The output of the segregation heat map and the global comprehensive segregation index intuitively displays the segregation distribution of the road section, which is convenient for rapid evaluation and decision-making.

[0196] Specifically, in one embodiment of the present invention, step S5 includes:

[0197] according to The value of is used to classify the overall segregation of the road section to be detected into different levels (for example: normal, slight segregation, moderate segregation, severe segregation).

[0198] Generate an overall segregation assessment report, which includes: basic information of the road section to be tested (such as road section number, test time, test length, etc.). Global comprehensive segregation index The specific value of. The determination result of the overall segregation level of the road section.

[0199] For each local area a, the following detection information is output: gradation segregation index :Indicates the degree of gradation segregation in a local area. Temperature segregation index : Indicates the degree of temperature segregation in a local area. Local comprehensive segregation index :Comprehensively reflects the segregation degree of local area. Local proportion factor : Represents the weight of the local area in the overall segregation analysis.

[0200] According to the local comprehensive segregation index and local proportion factor , mark the local areas with higher segregation degree. Provide the specific location information of key areas (such as area number, spatial coordinates, etc.) and description of segregation characteristics.

[0201] The local inspection report is presented in the form of a table or list, including: the number of each local area. The gradation segregation index of each local area , Temperature Segregation Index , Local comprehensive segregation index and proportion factor . Priority treatment recommendations for areas with severe segregation.

[0202] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent detection method for asphalt pavement segregation based on multi-source data analysis, characterized in that: include: S1. Determine the road section to be detected, and obtain multi-source data of the road section to be detected, including optical data obtained by a multispectral sensor, laser point cloud data collected by a laser radar, and thermal imaging data obtained by an infrared sensor; S2. Preprocessing of multi-source data, including: feature extraction of optical data to obtain surface particle information reflecting the distribution characteristics of coarse and fine materials on the pavement surface; screening and quality assessment of laser point cloud data to exclude abnormal points and incomplete point groups, and extracting point cloud quality information based on the density and depth distribution characteristics between point clouds; correcting thermal imaging data to obtain pavement temperature field distribution information; S3, constructing a dual-branch neural network detection model, including a first branch detection model and a second branch detection model, and pre-training the dual-branch neural network detection model, inputting the pre-processed optical data and laser point cloud data into the pre-trained first branch detection model to obtain a gradation segregation distribution result; inputting the pre-processed thermal imaging data into the pre-trained second branch detection model to obtain a temperature segregation distribution result; The first branch detection model includes: The input layer receives the pre-processed optical data and laser point cloud data, and sends the optical data and laser point cloud data to two independent processing channels respectively; The optical processing pipeline includes: multi-level convolutional network layers for step-by-step feature extraction of optical data; residual connections for preserving multi-scale information between convolutional layers; an attention module used in conjunction with the last convolutional layer; and a set of fully connected layers for transforming and encoding the extracted features within the pipeline. The point cloud processing pipeline includes: a multi-layer perceptron module that performs preliminary mapping of the input point cloud data; a graph neural network layer that uses points in the local area as nodes and quality information as node features, and updates features based on the connection relationship between points; a global pooling layer that merges graph-level features; and a set of fully connected layers that perform the final transformation on the merged features. The multimodal fusion layer performs multi-layer nonlinear fusion of the features output by the optical processing channel and the point cloud processing channel; The output layer uses Softmax classifier to output gradation separation classification information and classification probability distribution; The second branch detection model includes: The input layer receives the preprocessed thermal imaging data and divides the thermal imaging data into N spatial dimensions. x local regions, and input the thermal imaging data of each local region into the subsequent network in a sequence form; The spatiotemporal feature extraction module includes: at least two 3D convolutional layers to represent the features of thermal imaging data in spatial and temporal dimensions; residual connections to retain multi-scale spatiotemporal information; and local pooling layers to aggregate the extracted features; The multi-scale attention module sets several channel attentions of different scales on the acquired features and includes a learnable weight gating mechanism to assign attention weights to the features; The dynamic analysis layer uses an LSTM network to process the features output by the multi-scale attention module in the same sequence frame and combine the time series outputs; The output layer uses several fully connected layers to perform multi-level judgments on the output of the dynamic analysis layer to generate the temperature segregation distribution results of the thermal imaging data; S4, jointly analyzing the gradation segregation distribution results and the temperature segregation distribution results to generate a comprehensive segregation index for characterizing the segregation degree of the asphalt pavement; S5. Output the test results according to the comprehensive segregation index, including the overall segregation status of the asphalt pavement of the road section to be tested and a local test report.

2. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis according to claim 1, characterized in that: Step S2 includes: S21, subjecting the optical data collected by the multispectral sensor to multi-scale filtering and texture enhancement processing, extracting the size distribution and spatial texture characteristics of the surface particles, and obtaining a particle feature map that can be used to identify the difference in the distribution of coarse and fine materials; applying a morphological operator to divide the particle feature map into regions, removing redundant background, and obtaining surface particle information reflecting the distribution of coarse and fine materials on the surface of the road surface; S22, using a noise point removal algorithm to detect and remove noise points in the original laser point cloud data; performing quality assessment on the processed laser point cloud data to obtain quality information of the laser point cloud data; S23. Automatically gain and background subtraction are performed on thermal imaging data through a radiation correction model to eliminate sensor noise and ambient temperature interference; a normalization correction method is used to compensate for the measurement deviation between hot spots and cold spots, and finally the temperature field distribution information of the road surface to be tested is formed.

3. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis as claimed in claim 2, characterized in that: In step S22, the noise point removal algorithm includes: In the local neighborhood of each point p(i1), the least squares plane is fitted and the covariance matrix C(i1) of the point cloud in the neighborhood is calculated. Let λ max (i1) and λ min (i1) represent the maximum and minimum eigenvalues ​​of the matrix respectively, and the anisotropy factor of the point p(i1) is defined as: Among them, the larger the LAF(i1) value, the stronger the directional distribution of the point in the local area, and the suspected local anomaly. anis When , the point p(i1) is determined to be a suspected noise point, T anis is the preset threshold; For the point p(i2) initially judged as a suspected noise point by LAF(i1), its density flow vector is calculated: Where p(j) is the neighborhood point of p(i2), k is the number of neighborhood points, σ ​​is the local diffusion coefficient, and ‖·‖ represents the Euclidean distance; If ‖d f (i2)‖ <T d , T d is the preset density flow threshold, then p(i2) is determined to be a noise point and removed, otherwise it is retained.

4. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis as claimed in claim 2, characterized in that: In step S22, the quality assessment steps are as follows: For the point cloud data after noise removal, a sliding window method is used to divide the point cloud data into N x A local area adjacent to a space; In each local area a, for multiple scales r = {r1, r2, ..., r n }Calculate the fractal dimension D(a,r): Wherein, C(a,r) represents the cumulative count of points contained in the search radius r in the local area a; when D(a,r) deviates from the preset interval or shows abnormal jump, it indicates that there is a mutation or uneven distribution in the area a; Perform the least squares plane fitting on the local area a and calculate the residual between each point and the plane Calculate the curvature residual factor CR(a): Among them, |Ω(a)| represents the number of points in the local area a, for The gradient value of , δ is the adjustment coefficient; the larger the CR(a) value, the more uneven the surface in the local area a is; Calculate the comprehensive factor Q(a) of the point cloud quality of the local area a: Q(a)=w1·F(a)+w2·CR(a) Where F(a) = ∑ r D(a,r) / n is the average value of fractal dimension at multiple scales, w1 and w2 are weighting coefficients; The comprehensive factor of the point cloud quality of all local areas is calculated, and each local area is given a Q value. x The Q values ​​are summarized and a point cloud quality distribution map is generated as the quality information of the laser point cloud data.

5. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis according to claim 4 is characterized in that: The first branch detection model processes the optical data and laser point cloud data as follows to generate the gradation segregation distribution results: Receive the pre-processed optical data, divide it into local areas according to the laser point cloud data, process it uniformly, and transmit it to the optical processing channel through the input layer. At the same time, read the laser point cloud data of the corresponding local area and transmit it to the point cloud processing channel through the input layer. In the multi-level convolutional network layer, the optical data is convolved and pooled layer by layer to extract deep representations including texture and particle distribution information; the attention module is connected after the last convolution layer to generate the attention and inhibition weight matrix; the features output by the attention module are input into the fully connected layer to complete the high-dimensional feature encoding of the optical processing channel; In the multi-layer perceptron module, the salient features of the point cloud are initially mapped. The graph neural network layer iteratively updates the node features according to the topological relationship between points or grids to extract higher-order spatial structure information. In the global pooling layer, the features are reduced in dimension to obtain a compact point cloud representation. The fully connected layer is used to transform the dimension of the point cloud representation and map it into a fixed-size vector. In the multimodal fusion layer, the output vectors of the optical processing channel and the point cloud processing channel are feature concatenated, and the fused vectors are compressed and reorganized through multiple layers of nonlinear transformation. The reorganized vector is input into the Softmax classifier to calculate the probability that each local area belongs to a different gradation segregation level. The gradation segregation probabilities of all local areas are integrated to generate the gradation segregation distribution result.

6. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis according to claim 4, characterized in that: The second branch detection model processes the thermal imaging data as follows to obtain the temperature segregation distribution results: Read the preprocessed thermal imaging data, perform dimension correction on it, and divide it into N x The local areas are converted into sequence form; the thermal imaging data of each local area is taken as a minimum input unit and input into the subsequent network modules accordingly; The thermal imaging sequence is input into the spatiotemporal feature extraction module, and its joint distribution features in height, width and time dimensions are captured through at least two layers of 3D convolutional layers; the spatiotemporal information of different scales extracted by multiple layers is retained by using residual connections; and the features are aggregated by using local pooling layers; The multi-scale attention module is used to weight different spatial positions or time segments in the features, and the weighted features are input into the dynamic analysis layer to capture the evolution law of thermal imaging in time-continuous frames, and the output of the LSTM network is combined to form a comprehensive thermal feature vector; The comprehensive thermal feature vector is sent to the output layer, and several fully connected layers are used to map the features to the multi-level category space of temperature segregation. The Sigmoid output function is used to obtain the segregation probability of each local area. The segregation probability of each local area is integrated to generate the temperature segregation distribution result.

7. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis according to claim 4, characterized in that: Step S4 includes: S41, standardize the gradation segregation distribution results, map the gradation segregation probability of each local area to the [0,1] interval, and use it as the gradation segregation index I of each local area. g (a); Perform the same standardization operation on the temperature segregation distribution results to obtain the temperature segregation index I of each local area t (a); S42, calculate the local proportion factor G(a) of each local area: In the formula, a represents the local area index, and η is the adjustment coefficient; the larger G(a) is, the higher the degree of segregation in the local area is; S43, calculating the local comprehensive segregation index of each local area: Among them, γ and μ are hypercurve parameters greater than 0, which are used to control the nonlinear degree of segregation strength growth. com (a) The closer the value is to 1, the more serious the segregation is; S44. Calculate the global comprehensive segregation index: Will I com The spatial distribution of (a) is plotted as a segregation heat map, and the global comprehensive segregation index I is output overall , characterizing the overall segregation degree of the road section to be detected.

8. The intelligent detection method for asphalt pavement segregation based on multi-source data analysis according to claim 7, characterized in that: The adjustment coefficient η of the local proportion factor G(a) is dynamically adjusted according to historical engineering data, including: By analyzing the gradation segregation index I of different road sections g (a) and temperature segregation index I t (a) Statistical analysis to determine the initial value of η; According to the distribution law of the severity of road section segregation, the value range of η is dynamically adjusted to make G(a) more sensitive to high segregation areas.

Images