An asphalt pavement global disease monitoring method, system and medium

Abstract

This invention relates to a method, system, and medium for monitoring asphalt pavement defects across the entire area, belonging to the field of asphalt pavement monitoring technology. The method includes: acquiring surface images of the road section, ground-penetrating radar, and positioning data; inputting these into first and second feature networks respectively to extract macroscopic geometric parameters and hidden burial depth amplitude parameters of the defects, and generating external and internal defect objects with unique identifiers through dimensionality reduction; aligning the attribute fields of both in a digital twin model based on the positioning data to construct a unified multidimensional mathematical recombination structure; extracting the joint change rate of external and internal defect objects within the same spatial unit in the multi-period recombination structure, inputting it into a prediction model, and outputting evolution data. This invention abandons the traditional three-dimensional rigid coordinate registration in physical space, and eliminates scale conflicts and semantic mismatches during the fusion of macroscopic surface and internal defects through multidimensional mathematical tensor recombination or graph theory topological mapping, achieving efficient and low-latency cross-modal defect co-evolution early warning under limited on-board computing power.

Description

Technical Field

[0001] This invention belongs to the field of asphalt pavement monitoring, specifically relating to a method, system, and medium for monitoring the overall defects of asphalt pavement. Background Technology

[0002] Long-term service of highway infrastructure leads to various types and scales of structural and functional defects in asphalt pavements. Traditional single-surface inspection or single-internal structure detection can no longer comprehensively and objectively reflect the true health status of the road. Collaborative monitoring of defects across the entire area has become the direction of technological evolution in the field of highway non-destructive testing.

[0003] In terms of multi-source data fusion and comprehensive disease monitoring, existing technologies have made relevant explorations. For example, patent document CN121617240A discloses a "Multi-dimensional Integrated Detection Method and Equipment for Highways," which simultaneously acquires vehicle-mounted pavement appearance data and vehicle-mounted ground-penetrating radar data of the target highway, and uses this to identify abnormal target areas to calculate a comprehensive highway technical condition index. This method achieves the aggregation of multi-dimensional data at the macro-assessment level, but it is essentially still a mode of comprehensive weighted calculation after independent evaluation of surface and internal data, failing to establish a precise spatiotemporal dynamic coupling relationship between surface cracks and underlying hidden diseases in the underlying physical space. On the other hand, deep learning recognition technology for specific modalities is also becoming increasingly mature. For example, patent document CN116434059A discloses a "Road Hidden Disease Detection Method and System Based on Deep Learning and Ground-Penetrating Radar," which effectively improves the accuracy of ground-penetrating radar in identifying underground hidden diseases by utilizing deep learning algorithms. However, these techniques are limited to feature extraction at the microscopic level of a single modality and do not involve the fusion of cross-modal heterogeneous data.

[0004] When existing technologies attempt to perform deep fusion of high-resolution surface images with ground-penetrating radar (GPR) waveforms, they encounter significant physical and engineering bottlenecks: surface defects and hidden internal defects in asphalt pavements are separated by orders of magnitude on a spatial physical scale. Surface longitudinal cracks typically range in length from meters to tens of meters, exhibiting a slender and macroscopically continuous topological morphology; while voids or loose subgrade areas within the pavement structure often have physical diameters confined to narrow intervals of tens of centimeters, appearing as blurred, microwave-attenuated bands. If conventional graphics methods are used to construct a continuous digital elevation model in a global three-dimensional coordinate system, and surface visual pixels are subjected to forced, rigid registration and ray collision testing with underground radar reflection points space-by-space, it will not only lead to severe semantic mismatches and scale conflicts between macroscopic and microscopic features, but also trigger a computational crisis at the underlying level. The video memory and concurrent scheduling mechanism of vehicle-mounted edge computing devices have a hard physical limit. When processing huge three-dimensional continuous entities containing a large number of defect-free blank areas, the continuous geometry intersection algorithm will cause the computational complexity to explode non-linearly, which can easily trigger video memory overflow and system kernel crash, making it completely unsuitable for the data throughput requirements of routine highway inspection.

[0005] Therefore, how to eliminate the fundamental conflict between heterogeneous data of asphalt pavement surface and interior in terms of spatial scale and physical semantics, break through the computing power dead zone of graphic collision registration under the constraint of limited on-board hardware computing power, and realize high-fidelity collaborative evolution and deduction of cross-modal entity-level defects has become a core technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and medium for monitoring asphalt pavement defects across the entire area. This invention solves the problem of how to eliminate the conflict between heterogeneous data on the surface and interior of asphalt pavement in terms of spatial scale and physical semantics, and breaks through the bottleneck of graphic registration under the limited computing power of vehicle-mounted hardware, thereby realizing the problem of high-fidelity collaborative evolution and deduction of cross-modal entity-level defects.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for monitoring asphalt pavement distress across the entire surface. This includes acquiring visual image data, ground-penetrating radar data, and spatial positioning data of the target road segment; It also includes the following steps: The apparent image data is input into the first feature extraction network to extract the geometric parameters of road surface defects and assign a unique identifier to generate external defect objects. The ground-penetrating radar data is input into the second feature extraction network to extract the burial depth parameters and amplitude parameters of internal defects and assign a unique identifier to generate internal defect objects. Based on the spatial positioning data, the attribute fields of the external disease objects and the internal disease objects are aligned in the digital twin model to construct a unified multidimensional mathematical recombination structure. Obtain the unified multidimensional mathematical recombination structure of multiple periods within the historical time series, extract the joint change rate of the external disease object and the internal disease object within the same spatial unit, and input the joint change rate into the prediction model to output disease evolution data.

[0008] Furthermore, the construction of the unified multidimensional mathematical reorganization structure includes dividing the digital twin model into a set of three-dimensional voxel meshes with a preset resolution.

[0009] Furthermore, based on the spatial positioning data and the road surface defect geometric parameters, the external defect objects are mapped to the surface voxel subset in the three-dimensional voxel mesh set.

[0010] Furthermore, based on the spatial positioning data and the internal disease burial depth parameters, the internal disease objects are mapped to the deep voxel subset in the three-dimensional voxel mesh set, and the spatial intersection attribute between the surface voxel subset and the deep voxel subset is extracted.

[0011] Furthermore, the construction of the unified multidimensional mathematical recombination structure includes constructing a heterogeneous graph topology model, defining the external disease objects and the internal disease objects as node sequences in the heterogeneous graph topology model.

[0012] Furthermore, when the spatial positioning data distance between the external disease object node and the internal disease object node is within a preset radius threshold, a connected topology edge is established between the external disease object node and the internal disease object node.

[0013] Furthermore, the step of inputting the apparent image data into the first feature extraction network to generate external disease objects includes mapping the apparent image features into a one-dimensional feature sequence, calling the state space equation to calculate the geometric feature points along the disease direction, and extracting discrete local width parameters and cumulative skeleton length parameters.

[0014] Furthermore, the step of inputting the ground-penetrating radar data into the second feature extraction network to generate internal disease objects includes extracting local radar wave reflection features and generating an adaptive reconstructed verification feature map based on the local radar wave reflection features, and then performing an upsampling operation.

[0015] A comprehensive asphalt pavement distress monitoring system includes: The acquisition module is configured to acquire the apparent image data, ground-penetrating radar data, and spatial positioning data of the target road segment; The extraction module is configured to input the apparent image data into a first feature extraction network to extract the geometric parameters of road surface defects and assign a unique identifier to generate external defect objects, and input the ground penetrating radar data into a second feature extraction network to extract the burial depth parameters and amplitude parameters of internal defects and assign a unique identifier to generate internal defect objects. The mapping module is configured to align the attribute fields of the external disease object and the internal disease object in the digital twin model based on the spatial positioning data to construct a unified multidimensional mathematical recombination structure. The extrapolation module is configured to obtain the unified multidimensional mathematical recombination structure of multiple periods within the historical time series, extract the joint rate of change, and input it into the prediction model to output disease evolution data.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the asphalt pavement whole-area distress monitoring method as described.

[0017] The beneficial effects of this invention are: 1. This invention uses an asymmetric feature extraction network to reduce the dimensionality of heterogeneous multi-source data, extracting external and internal disease objects carrying unique identifiers, and reconstructing them into a unified multi-dimensional mathematical recombination structure in a digital twin model. This mechanism completely abandons the traditional pixel-level image overlay or hard stitching of three-dimensional physical bounding boxes, transforming heterogeneous diseases with vastly different real-world geometric dimensions into numerical attribute operation nodes of equal status in the same coordinate domain, fundamentally eliminating the macro-micro scale conflicts and semantic mismatches during the fusion of multi-source detection data.

[0018] 2. Addressing the real-world constraints of limited edge computing hardware in vehicle-mounted mobile monitoring platforms, this invention reduces the complexity of the continuous three-dimensional absolute physical space into a set of three-dimensional voxel meshes or heterogeneous graph topology models at a preset resolution. This pure algebraic dimensional recombination (converting graph collisions into standard dense matrix multiplication and addition operations or graph network message passing) avoids unnecessary memory usage when processing massive amounts of blank areas in healthy road sections. It completely solves the memory overflow and computational crashes easily caused by traditional continuous spatial mapping, and stably compresses the single-frame fusion processing latency to the millisecond level, meeting the minimum real-time requirements of high-speed inspection.

[0019] 3. This invention overcomes the technical limitations of single-modal static status assessment. By extracting the joint change rate of surface and deep-seated defects within the same spatial unit and introducing traffic load intensity and pavement compaction as adaptive correction factors, it deeply quantifies the nonlinear coupling destructive effect of surface water seepage and accelerated expansion of deep cavities. This endows the system with the ability to accurately penetrate the surface phenomena and understand the causal relationships of defects, achieving high-fidelity spatiotemporal collaborative evolution early warning for cross-modal heterogeneous defects with extremely low hardware computing power. Attached Figure Description To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is the main flowchart of the asphalt pavement whole-area distress monitoring method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of cross-modal mapping in three-dimensional voxelized discrete space according to the present invention. Figure 3 This is a diagram illustrating the architecture for constructing and evolving the heterogeneous graph topology model of this invention. Figure 4 This is a simulation diagram of the combined rate of change time-series evolution and dynamic early warning of the present invention; Figure 5 This is a simulation diagram of the discretization mapping and intersection of the voxelized space in this invention; Figure 6 This is a simulation diagram of the cross-modal heterogeneous graph topology connectivity network of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for monitoring asphalt pavement defects across the entire area, including acquiring surface image data, ground-penetrating radar data, and spatial positioning data of a target road segment; the method includes the following steps: inputting the surface image data into a first feature extraction network to extract geometric parameters of road surface defects and assigning unique identifiers to generate external defect objects; inputting the ground-penetrating radar data into a second feature extraction network to extract internal defect burial depth parameters and amplitude parameters and assigning unique identifiers to generate internal defect objects; based on the spatial positioning data, aligning the attribute fields of the external defect objects and the internal defect objects in a digital twin model to construct a unified multidimensional mathematical recombination structure; acquiring the unified multidimensional mathematical recombination structure for multiple periods within a historical time series, extracting the joint change rate of the external defect objects and the internal defect objects within the same spatial unit, and inputting the joint change rate into a prediction model to output defect evolution data.

[0023] The step of inputting the apparent image data into the first feature extraction network to generate external disease objects includes mapping the apparent image features into a one-dimensional feature sequence, calling the state space equation to calculate the geometric feature points along the disease direction, and extracting discrete local width parameters and cumulative skeleton length parameters.

[0024] The process of inputting the ground-penetrating radar data into the second feature extraction network to generate internal disease objects includes extracting local radar wave reflection features and generating an adaptive reconstructed verification feature map based on the local radar wave reflection features, followed by upsampling.

[0025] The acquisition module is configured to acquire apparent image data, ground-penetrating radar data, and spatial positioning data of the target road segment; the extraction module is configured to input the apparent image data into a first feature extraction network to extract the geometric parameters of road surface defects and assign unique identifiers to generate external defect objects, and input the ground-penetrating radar data into a second feature extraction network to extract the burial depth parameters and amplitude parameters of internal defects and assign unique identifiers to generate internal defect objects; the mapping module is configured to align the attribute fields of the external defect objects and the internal defect objects in a digital twin model based on the spatial positioning data to construct a unified multidimensional mathematical recombination structure; the inference module is configured to acquire the unified multidimensional mathematical recombination structure for multiple periods within a historical time series, extract the joint rate of change, and input it into a prediction model to output defect evolution data.

[0026] When the computer program is executed by the processor, it performs the steps of the method for monitoring the overall defects of asphalt pavement.

[0027] The linear array camera mounted on the inspection vehicle continuously captures surface image data of the road surface, and the synchronous transmission frequency of the three-dimensional multi-channel ground penetrating radar antenna array is limited to... to A continuous electromagnetic wave pattern is used to receive echo signals reflected from the interface of the subsurface medium to obtain deep structural detection data. The satellite navigation and positioning system and the inertial measurement unit operate according to a preset sampling period. The system synchronously records the absolute spatial positioning reference corresponding to each frame of image and each radar waveform. Multi-source heterogeneous data is transmitted via an onboard 100Mbps Ethernet bus to edge computing nodes with built-in multi-core computing units. Conventional non-destructive testing methods tend to construct a three-dimensional digital elevation model containing a global coordinate system, attempting to spatially map surface visual pixels to underground radar reflection points within a graphics processor. However, the scale of asphalt pavement damage in real physical environments is extremely wide, with the length parameter of apparent longitudinal cracks typically distributed across... The physical diameter of the hidden cavities or loose base areas is mostly concentrated in the meter range. Within a 1-meter interval. If a scheme is adopted that directly forces the construction of a full-dimensional bounding box in 3D space and performs voxel-level collision matching, the fusion rendering of the 3D point cloud generated for a single-kilometer lane with radar feature data will take more than [a certain amount of time / time]. The video memory resources are limited. The power supply of the vehicle-mounted mobile detection platform is constrained, and the peak available video memory of the edge computing nodes hovers around the hardware limit. Rendering and matching paths easily triggers video memory overflow errors and computational dead zones. This embodiment abandons direct stitching of the physical dimension of the graphics space, employing an asymmetric feature extraction algorithm to reduce the dimensionality of heterogeneous data, and logically transitioning to a multi-dimensional mathematical tensor reconstruction system.

[0028] The surface defects of the road surface exhibit a long, thin, and variable topological morphology. The road environment is mixed with dense environmental noise, including rough aggregate textures, remnants of traffic markings, and uneven lighting and shadows. When using a conventional convolutional stacked network with a fixed receptive field to process long, thin cracks spanning multiple receptive field windows, the feature map is prone to geometric connectivity breaks during the downsampling stage. The first feature extraction network uses a multi-layer feature aggregation backbone to extract shallow visual feature maps, flattening the two-dimensional structured feature map along the spatial reference dimension into a one-dimensional feature sequence. A state-space equation is introduced to iteratively calculate the one-dimensional feature sequence, capturing the global long-distance dependency of the crack along its direction. The discretized mathematical expression is set as follows: In the formula, The sequence position index is The input feature vector at the location represents the original pixel features of the local crack texture; For position The hidden state vector at the location stores and transmits the accumulated geometric topological connectivity information along the path of the disease; This is a state transition matrix, and the range of values ​​for its elements is limited to... This determines the specific attenuation ratio of the crack structure information from the previous spatial location to the current location; The input mapping matrix; To output the mapping matrix, the hidden state information is decoded into the predicted output feature response. The system applies a morphological topology thinning algorithm to extract the disease center skeleton with a single pixel width within the output connected region. It then calculates the actual tortuous length based on the cumulative Euclidean distance between adjacent pixels of the skeleton and outputs the cumulative skeleton length parameter. To accumulate skeleton length, This represents the total number of skeleton pixels. , For the first The system extracts discrete local width parameters by bidirectionally searching the contrast gradient boundary along the skeleton normal direction. The onboard computing node encapsulates the above geometric scale values ​​with automatically assigned unique identifiers and spatial coordinate anchors to generate external defect objects through dimensionality reduction.

[0029] Loose subgrade or interlayer voids within the road surface manifest as anomalous wave groups with abrupt amplitude changes and chaotic phases in electromagnetic wave reflection profiles, with strong diffuse reflection characteristics at the physical distribution boundaries. Traditional methods perform bilinear interpolation upsampling on the radar feature map during the fusion stage. The isotropic smoothing compensation introduced by the interpolation algorithm irreversibly destroys the edge sharpness of regions with abrupt changes in dielectric constant, resulting in the permanent loss of high-frequency boundary details carried in deep weak echo signals. The second feature extraction network replaces the static upsampling layer with a content-aware feature reconstruction module. The graphics processor utilizes a hollow spatial pyramid pooling structure to obtain multi-scale frequency domain response features of the radar profile matrix. Local radar wave reflection features are extracted, and an adaptive reconstruction kernel is dynamically generated. The calculation rule for the adaptive reconstruction kernel weights is set as follows: Compact feature variables after channel compression of the input local radar wave feature matrix; For linear mapping operators, calculate position. Local neighborhood support points The unnormalized response value; For The local electromagnetic sensing neighborhood space centered on; The content-aware reassembly weights variables. The network performs a pixel-by-pixel spatially weighted reassembly upsampling operation on the feature map: The reconstructed high-resolution radar characteristic response value. The original values ​​are low-resolution radar features. The reconstruction mechanism relies on the gradient of dielectric anomaly fluctuations in the neighborhood to perform irregular edge reconstruction, locking the physical depth range and amplitude nonlinear attenuation degree of deep-seated defects. The extracted hidden physical parameters are packaged with satellite positioning anchor points and system identification codes to generate structured internal defect objects.

[0030] Semantic conflict resolution and scale tensor reconstruction of cross-modal feature objects are core engineering bottlenecks in establishing a seamless full-domain monitoring link. The primary constraint faced by onboard computing environments is the processing latency of the detection data stream. Vehicles... During normal-speed inspections, the data throughput generated per unit time is extremely high. Building a 3D bounding box for the entire point cloud and performing collision testing in a 3D digital twin engine requires dense ray intersection tests on tens of millions of coordinate points. Conventional ray tracing algorithms or bounding box hierarchies, when processing coordinate points in extremely open, disease-free, healthy road sections, require the computational unit to traverse a large number of invalid blank areas. The computational complexity of the graphics-based processing logic is thus significantly increased. The growth trend indicates that the latency of single-frame fusion processing typically exceeds [a certain threshold]. The scope is completely contrary to the requirements of the vehicle hardware system. Real-time feedback of the bottom line. Replacing 3D collision matching with tensor operations in algebraic space becomes the only way to avoid this computational bottleneck. The system maps the continuously mapped digital twin road surface model according to a preset physical distance. The data is partitioned and reduced to standard 3D spatial geographic grid matrix units. For each spatial geographic grid unit, the independent attribute fields of external and internal disease objects falling within that coordinate range are extracted and reconstructed into a unified multidimensional mathematical reorganization structure that can be directly executed with high concurrency by a computer calling the matrix processing core. The mathematical definition is set as follows: To construct a unified multidimensional mathematical recombination structure matrix; The index vector of the discretized spatial grid cell solidifies the absolute spatial coordinates of the disease occurrence. The timestamp sequence dimension variable identifies the disease collection batch and the iterative monitoring cycle; The attribute dimension set variable for external disease objects includes a scalarized representation of apparent state quantities; This is a set of attribute dimensions for internal disease objects, encompassing a scalarized representation of hidden physical state quantities. The physical significance of this tensor-based dimensionality reduction and reconstruction mapping process lies in stripping away the visual space mapping requirements dependent on graphics processor shader rendering. In the new algebraic coordinate system, the length of external cracks spanning the road surface is represented by the tensor... The dimension collapses into a concrete scalar, and the amplitude boundary of the underground cavity lies in the tensor. Dimensional synchronization is mapped to an independent scalar. Heterogeneous defects with vastly different real-world geometric dimensions are transformed into numerical computation nodes of equal status within the same multidimensional tensor coordinate domain. Collision operations, which originally involved interferometry testing of massive numbers of triangular facets, are transformed into large-scale dense matrix multiplication and addition operations, which are best suited for execution by the underlying hardware. The algorithm complexity is reduced to Magnitude, parameters Represents the total number of diseased objects extracted within a single grid; the data fusion latency of a single frame is stably compressed to [amount missing]. Within the threshold, the risk of system crashes due to computational overflow is eliminated.

[0031] The database contains continuous multi-period, multi-dimensional tensor data based on historical time series. The system retrieves data from specific spatial grid cells with overlapping spatiotemporal attributes, extracting the joint rate of change characteristics between external and internal diseased objects within the same spatial cell. This parameter quantifies the nonlinear coupling destructive effect between surface damage propagation and internal structural degradation. The core partial differential equation for calculation is set as follows: It is a high-order comprehensive joint change rate index; Measure the magnitude of the rate of expansion of the core attributes of external diseased objects over time; Quantify the degradation gradient of the physical properties of internal diseased objects over time; and The system dynamically adjusts the weighting coefficients, subject to basic constraints. Limited. The system adaptively adjusts values ​​based on the physical and dynamic feedback of surface water seepage rate and subgrade loosening expansion in the input time-series data. The combined rate of change variable is input into the spatial-temporal prediction model, outputting disease evolution assessment data for future time intervals. The system sets an adaptive early warning threshold; when the quantified assessment coefficient exceeds the dynamic threshold, a tiered active early warning mechanism is triggered. The calculation rule for this is: This is the lower limit of the warning threshold. The dynamic evaluation coefficient of external vehicle load intensity connected to the traffic perception subsystem; It is the environmental constant of the basic mechanical resistance of pavement materials, which is strictly constrained by ambient temperature and material composition. This is a penalty factor for material fatigue sensitivity. It is caused by increased frequency of heavy traffic flow or axle loads from heavy trucks on the monitored road section. As the threshold increases, the warning threshold decreases exponentially. The dynamic threshold correction mechanism based on the traffic environmental load compensation algorithm enables the prediction model to balance the superimposed effects of pavement distress deterioration and external heavy load environment.

[0032] like Figure 4 As shown, this is the time-series simulation curve of the joint rate of change and the dynamic early warning threshold generated in this embodiment. When the higher-order joint rate of change increases nonlinearly with time and breaks through the dynamically adjusted early warning threshold, the system accurately outputs the early intervention point.

[0033] Example 2 The construction of a unified multidimensional mathematical reorganization structure includes dividing the digital twin model into a set of three-dimensional voxel grids with a preset resolution, mapping the external disease objects to a surface voxel subset in the three-dimensional voxel grid set based on the spatial positioning data and the road surface defect geometric parameters, mapping the internal disease objects to a deep voxel subset in the three-dimensional voxel grid set based on the spatial positioning data and the internal defect burial depth parameters, and extracting the spatial intersection attribute between the surface voxel subset and the deep voxel subset.

[0034] The physical memory (typically 16GB or 32GB maximum) and concurrent thread scheduling mechanism of the edge computing nodes on the vehicle-mounted mobile monitoring platform are limited by the vehicle's power supply and heat dissipation troubleshooting. Conventional cross-modal spatial fusion tends to retain all continuous geometric features in the global three-dimensional coordinate system, using Delaunay triangulation or Poisson surface reconstruction algorithms to construct discrete coordinate patches. Surface defects and internal hidden defects in asphalt pavements are absolutely separated at the physical scale. The meandering span of longitudinal surface cracks is typically... On the order of meters, the anomaly in the dielectric constant of the internal concealed radar is concentrated in the region. Within a confined space of meters, rigid registration and 3D ray collision detection of raw continuous geometric entities are performed inside the graphics processing core. The rendering of the 3D point cloud generated for a single kilometer of lane often consumes over 32GB of video memory. Massive amounts of blank areas in healthy road sections without defects continuously encroach on the buffer zone, leading to a significant increase in computational complexity due to the intersection algorithm for continuous geometry based on hierarchical bounding boxes. Even non-linear explosions. The vehicle was... At normal inspection speeds, hundreds of megabytes of heterogeneous data are generated per second. Continuous 3D mapping inevitably leads to memory overflow errors and system crashes. This embodiment adopts a 3D voxel-based discrete dimensionality reduction spatial mapping architecture, which forcibly converts the continuously distributed unstructured disease spatial coordinates into a standard physical 3D mesh matrix with a unified discrete resolution, bypassing the hardware computing power dead zone encountered when processing heterogeneous scale geometries.

[0035] like Figure 2 As shown, the digital twin model is defined within a computing node as a set of discretized three-dimensional voxel meshes with absolute spatial boundaries. The central processing unit uses the absolute geographic coordinates of the origin and destination points output by the satellite positioning module. Based on the construction drawings Standard driving lane width and The roadbed design depth constructs a global orthogonal three-dimensional physical space encompassing the monitoring area. This global space is divided along the longitudinal driving direction, the transverse lane direction, and the vertical depth direction, with side lengths defined by preset resolution parameters. The cube-shaped mesh element. The lower limit is locked to the maximum nominal particle size of the pavement asphalt mixture (typically 100 μm). or The upper limit is constrained by the vertical resolution limit of ground-penetrating radar in asphalt media (usually set to ). \ The grid resolution is too low (e.g., set to a specific value). This will cause tiny holes to be smoothly erased by the surrounding healthy medium mesh; excessively high resolution (such as setting it to 0) will cause these holes to be smoothly erased by the surrounding healthy medium mesh; This will cause the number of meshes to increase cubically, triggering a memory overflow crisis again. The topology of a 3D voxel mesh set is defined as a structured set of discrete coordinates: For a voxel mesh set matrix, For discrete integer space index coordinates, This represents the maximum number of discrete cuts in the mesh along the three-dimensional direction. The graphics processing unit reads the external defect object and projects the two-dimensional visual features of the continuous tortuous length and discrete local width onto... The topmost horizontal slice layer. A mapping algorithm extracts the skeleton feature coordinates and extended width boundaries, calculates the surface mesh index of the envelope 2D apparent contour, assigns activation state labels and damage intensity feature values, and generates a surface voxel subset through dimensionality reduction. Activation determination is controlled by the boundary occupancy model. The Boolean value representing the activation state of a road surface voxel unit with a depth level of zero. The anchor points are the projections of the continuous topological skeleton onto the grid coordinate system. The dynamic projection search coverage radius (unit: grid number) is modulated by the width of the disease expansion. Through activation function mapping, continuous surface cracks are visualized as a set of discrete active grid blocks on the surface, eliminating the geometric edge uncertainty caused by pixel jaggedness and complex lighting burrs.

[0036] The deep mapping reconstruction of internal defects and the quantitative extraction of cross-modal spatial intersection attributes constitute the core computational power consumption link that determines the early warning accuracy in the voxel mapping model. In real physical scenarios, the interlayer delamination and water damage evolution inside the asphalt pavement structure are by no means isolated columnar entities falling vertically. When natural precipitation or surface water seeps into the lower layer or base layer along longitudinal cracks in the surface layer, the fluid at the interface of interlayer materials with different compaction degrees (such as the interface between the asphalt surface layer and the semi-rigid water-stabilized base layer) will undergo non-uniform lateral capillary diffusion due to the influence of surface tension and material porosity characteristics. The physical boundary of the hidden loose area appears as a gradually blurred attenuation transition zone with a gradually blurred dielectric constant in the B-Scan electromagnetic wave reflection profile received by ground penetrating radar, which does not have the rigid characteristics of a cut surface of a hard material at all. Mechanically forcibly stuffing internal defects into a single fixed depth (such as the center burial depth parameter and the absolute amplitude of the radar) according to the center burial depth parameter and the absolute amplitude of the radar Within a single voxel mesh, the three-dimensional diffusion trajectory of fluid erosion and the spherical transmission and dissipation law of vehicle axle load compressive stress are severely violated. The mapping mechanism from internal defect objects to deep voxel subsets incorporates a soft-enclosure volume expansion algorithm based on a nonlinear spatial diffusion factor. The central processing unit retrieves the internal burial depth parameters as the starting anchoring slice level index for the vertical dimension. The electromagnetic wave amplitude parameters are converted into an initial internal damage intensity scalar of the central grid. The algorithm uses the anchored voxel mesh as the physical origin and radiates dielectric anomalous damage states to neighboring mesh cells in a 3D space in an ellipsoidal shape, generating a deep voxel subset with continuous gradient properties. The rule for calculating the extended damage intensity in 3D space is set as follows: For spatial damage intensity parameters of deep, specific coordinate grids, The initial amplitude scalar for the core anomaly region, , , A three-dimensional mesh index for the hidden void absolute physical center. , , This represents the diffusion damping coefficient of the material structure along three dimensions. The diffusion damping coefficient is inversely proportional to the foundation compaction obtained from core sampling and directly proportional to the porosity. The base porosity parameter retrieved from radar waves surges to [value missing]. At the above, the lateral diffusion damping coefficient , The degradation is precipitous, and the algorithm automatically generates a large subset of deep water bladder damage voxels containing hundreds of adjacent grids.

[0037] After completing the mapping and assignment between the surface and deep meshes, the system initiates spatial intersection attribute extraction within a unified mesh coordinate domain. In engineering practice, the distance between the bottom of the surface crack and the top of the deep void is usually significant. In healthy structural interlayers, multimodal defects have absolutely no possibility of direct intersection or collision at the continuous geometric entity level. The physical essence of searching and extracting spatial intersection attributes is not to find simple three-dimensional volume overlap, but to quantify the potential of surface water seepage channels and the accelerated expansion trend of deep cavities, as well as the stress coupling damage intensity generated within the same vertical columnar bearing space. The system constructs a topological search mapping matrix that runs through the entire depth of the pavement structure in a discrete voxel matrix, and calculates the numerical values ​​of cooperative damage attributes for vertically aligned grid columns in the same latitude and longitude coordinate plane.

[0038] like Figure 5As shown in the three-dimensional voxel mapping space simulation diagram, the system retains only the blue surface voxels and red deep damage voxels in the active state without performing full point cloud rendering. By extracting the spatial intersection attribute of the two in the vertical projection column (shown by dashed lines), stress field coupling analysis with low computational overhead is achieved. The core operator mechanism of spatial intersection properties is defined as follows: This is the sum of the cross-modal space intersection attributes extracted from the grid coordinates of the current projection plane. and The dynamic weighting correction coefficient is adaptively allocated by the system based on the local average annual rainfall and the proportion of heavy traffic, constraining the contribution ratio of surface water damage and loss of internal bearing capacity. For physical depth indexing The vertical spatial stress attenuation transfer function exhibits nonlinear negative correlation ( (For attenuation rate parameter), simulating the Boussinesq compressive stress bubble diffusion effect of vehicle-mounted dynamic loads in a deep semi-infinite space. When the surface crack voxel ( The area directly below the projection contains deeply hidden, highly damaged pathogenic elements. And the vertical depth distance falls within the effective stress transfer range ( When the mesh multiply-accumulate operation outputs results in an exponential jump in the intersection attribute values, exceeding the safety monitoring threshold. Past industry attempts to establish 3D finite element models containing fluid-structure interaction equations for real-time stress field inference have resulted in single-step partial differential equation solutions taking several minutes, completely unsuitable for vehicle-mounted inspection requirements. By replacing the finite element field solution process with low-computational-cost matrix discrete multiply-accumulate operations, the single-step inference latency at vehicle nodes is compressed to [amount missing]. The system accurately identifies the coordinate grid of overlapping surface and internal defects that are extremely prone to causing large-scale collapse accidents.

[0039] The spatial intersection attribute matrix of the quantized output serves as a feature dimension vector characterizing the degree of coupling of diseases across the entire domain. The vehicle-mounted edge computing node directly encapsulates the sequence of three-dimensional voxel mesh sets, which includes the degree of topological location co-overlap and the gradient of vertical damage correlation, into the vehicle-mounted NVMe solid-state storage array. The discretized and dimensionality-reduced three-dimensional voxel matrix avoids the semantic mismatch contradiction of heterogeneous detection signals at the microscopic spatial scale, reducing the complex interlayer damage tracing evaluation to standard array parallel indexing and matrix multiplication and addition matching calculations. During the inference phase, the prediction model retrieves the gradient parameters of the dynamic changes of intersection attributes in the time series. When the value of a certain grid column approaches the warning threshold, it outputs a disease evolution intervention command containing precise three-dimensional grid coordinates. The voxelization recombination mechanism avoids dependence on the computing power of large workstations and achieves low-latency multi-source disease spatial coupling risk determination by relying on the limited hardware resources of the vehicle.

[0040] Example 3 The method for monitoring asphalt pavement defects across the entire area involves constructing a unified multidimensional mathematical reorganization structure, including building a heterogeneous graph topology model. External and internal defect objects are defined as node sequences in the heterogeneous graph topology model. When the spatial positioning data distance between external and internal defect object nodes is within a preset radius threshold, a connected topology edge is established between the external and internal defect object nodes.

[0041] When processing multi-source detection data from a provincial highway network spanning hundreds of kilometers, the vehicle-mounted edge computing nodes... to Physical video memory becomes a rigid constraint. Conventional solutions attempt to build an elevation rendering model containing all discrete coordinate points in a continuous 3D coordinate system. Rendering a single-kilometer dual-lane road using 3D point cloud data stitched with radar feature data consumes over [amount missing]. Video memory. If a uniform voxelization matrix is ​​used for spatial discretization, it is difficult to handle situations where... The above continuous inspection tasks, due to their occupation The road surface, which dominates the absolute space mentioned above, is in a healthy, non-destructive state. The system must allocate massive zero-value data structures to these blank areas. When the graphics processing core performs sparse matrix multiplication, invalid traversal operations cause severe memory bandwidth constraints, and the continuous influx of data can easily trigger memory overflow errors and system-wide crashes. Limited by... The power consumption limit of the vehicle battery and the heat dissipation bottleneck of the enclosed chassis prevent the improvement of computing power by directly stacking multiple graphics accelerator cards with large video memory. This embodiment abandons the geometric mapping and volume occupation of the three-dimensional absolute physical geometric space, abstracts the damaged entities in the physical world, constructs a heterogeneous graph topology model in the algebraic dimension, and reconstructs the physical and mechanical relationship between the surface and the interior of the disease by only using the topological connection relationship of points and lines.

[0042] The graph processing core receives external disease objects output by the first feature extraction network and internal disease objects output by the second feature extraction network, defining them as a set of discrete mathematical nodes. The model is mathematically defined as a graph-structured topology containing two types of heterogeneous entities. Represents a global heterogeneous graph topology model; This represents a sequence of external disease object nodes transformed from dimensionality reduction of apparent image data. These nodes encapsulate discrete local width parameters (within a defined range) characterizing the crack orientation. ) and cumulative skeleton length parameter (setting range) ); This represents a sequence of internal defect object nodes generated from ground-penetrating radar data inversion, carrying the burial depth parameters (set range) of the loose underground area. ) and amplitude attenuation parameters; set and The complete set of nodes that constitute the entire monitoring space replaces the three-dimensional grid. It represents the set of connected topological edges established across nodes of different modalities; This represents the set of feature weights assigned to each connected topological edge. Latitude and longitude location data of external and internal diseased objects ( The coordinate system and vertical depth data are converted into numerical attribute vectors within the nodes. A road spanning half a lane... Long surface cracks and a diameter lurking deep in the base layer The microscopic voids are compressed into two isolated mathematical nodes in the heterogeneous graph topology model. This dimensionality reduction and stripping strategy eliminates the data description of blank areas in healthy road sections, and the memory space only records the actual disease entities that have occurred, enabling the on-board hardware to load the full amount of disease detection data of ultra-long-distance highway networks under the constraint of limited physical memory.

[0043] like Figure 3 As shown, the dynamic determination and feature weight assignment mechanism of the connected topology edge set is the core computational hub for filtering redundant associations in heterogeneous graph topology models, and its implementation logic determines the actual availability of cross-modal collaborative early warning. The asphalt pavement structure exhibits anisotropic mechanical characteristics; the gravity-driven resistance of surface water infiltration in the vertical direction is much smaller than the resistance of capillary lateral diffusion of water within the same horizontal structural layer. The transmission path of the deviatoric stress generated by vehicle-loaded dynamic loads within the pavement structure follows the Boussinesq spherical stress bubble attenuation and diffusion boundary. An attempt was made to determine whether physical coupling failure occurs between surface cracks and the underlying void solely based on traditional three-dimensional Euclidean distance. Test data shows that when a fixed Euclidean distance threshold is set (e.g., ...), ... When this was done, a large number of defects that were only similar in the horizontal direction but completely isolated in the vertical direction by the impermeable asphalt sublayer were incorrectly associated, causing more than [number missing] [issues / problems]. The false alarm rate; if this empirical threshold is forcibly reduced, it will miss the true water damage correlation that seeps into the deep base layer along vertical micro-cracks. When the central processing unit calculates the spatial location data distance between external and internal defect object nodes, a modified metric equation designed for the anisotropic characteristics of layered pavements must be introduced: Represents external nodes With internal nodes Spatial positioning data distance between them after road surface anisotropy correction; The longitudinal and transverse absolute coordinates of the two nodes in the road surface horizontal projection coordinate system; and The absolute depth positioning parameters of two nodes in the vertical direction of gravity, and the apparent external disease object nodes. Constantly set to zero; This is the anisotropic depth penalty factor. A small increase in vertical depth has a physical effect on blocking water damage seepage and stress transmission that is equivalent to a multiplier effect on horizontal distance; the penalty factor... The introduction of this will amplify the effect of vertical isolation at the algebraic level.

[0044] The spatial distance calculated by the corrected equation needs to be compared with the dynamically preset radius threshold set by the system. The pavement's impermeability and fatigue life differ significantly depending on its service life and asphalt mixture ratio (e.g., AC-20 or SMA-13). Using static, fixed values ​​as the judgment threshold cannot adapt to the changing material properties of the road network. The system introduces a dynamic adaptive preset radius threshold calculation mechanism linked to the node's own damage intensity and environmental load: This represents a dynamically preset radius threshold. Material constants for calibrating basic hydrological diffusion capacity; This represents a normalized comprehensive damage intensity scalar that combines the discrete local width of external cracks with the amplitude of internal dielectric anomalies. The search radius is set as a sensitive amplification parameter for damage intensity, resulting in an exponential expansion of severe cracks. The compaction parameters of the local structural layers are obtained by inversion based on the high-frequency reflection spectrum characteristics of ground-penetrating radar. This represents the material density damping coefficient. The calculated corrected spatial distance... At that time, the core of graphics processing lies in the nodes. With nodes Instantiate a physically existing connected topology edge between them. This step, at the mathematical logic level, confirms the existence of a clear hydraulic seepage channel or stress concentration transmission path between surface macroscopic damage and deep microscopic hazards. To quantify the accessibility of the coupled damage path, the system calculates feature weights for the connected topological edges: The feature weight coefficients representing the topological edges; This is the normalized equilibrium constant; To prevent a safe smoothing term with a denominator of zero, the topological edge weights of node pairs that are spatially closer and have higher overall damage intensity tend to approach the saturation upper limit. This characterizes the physical risk of localized pavement structural collapse caused by the synergistic effect of surface and internal defects.

[0045] like Figure 6 As shown in the simulation diagram of the heterogeneous graph topology model, external apparent defects (square nodes) and internal hidden defects (circular nodes) are completely extracted from the physical space and connected by topological edges with different feature weights (different thicknesses), forming a pure mathematical foundation for the graph convolutional network to perform message passing.

[0046] Heterogeneous graph structures containing node attribute sequences and topologically connected edge matrices are incorporated into spatiotemporal heterogeneous graph convolutional prediction models. Conventional convolutional kernels, which perform local sliding window scanning on traditional pixel matrices, rely on regular spatial grid arrangements and are completely incapable of handling discrete topological coordinates. The graph convolutional prediction model performs message passing and neighborhood feature aggregation operations along the connected topological edge network, simulating the physical evolution dynamics of surface precipitation seeping into deep, hidden cavities through cracks in digital space. The aggregation and update rules for the implicit temporal state features of nodes in the graph neural network are governed by an algebraic partial differential framework. Representative node After a single simulation time step (e.g.) The updated predicted feature vector obtained after the (day) deduction; The activation function represents the nonlinear fatigue damage characteristics introduced into the material degradation process; This represents the current initial feature extraction state of the node; Represents the relationship between topological edges and nodes. The set of all directly connected heterogeneous neighboring nodes; To retain factors for its own state, Neighborhood collaborative destruction penetration factor, network training constraints limit The network uses this equation to extract joint features from external and internal disease objects bound together within the same spatial unit, outputting a quantitative representation of the joint rate of change of their coupled evolution intensity: To obtain the higher-order joint rate of change tensor parameters, Represents the Kronecker product operation, capturing the surface crack propagation rate vector (unit: The deep, loosely coupled gradient vectors exhibit a full-dimensional orthogonal synergistic effect. A point with a width of only... The surface microcracks at the nodes themselves exhibit extremely low rates of change, but through topological edge aggregation, they converge directly beneath them, resulting in a porosity surge gradient. When the water bladder node characteristics are observed, the norm of the joint rate of change tensor exhibits a nonlinear jump. The monitoring core quantifies and compares this tensor norm with the safety evolution benchmark. When the value exceeds the set warning line, it outputs data instructions containing latitude and longitude anchor points and evolution prediction trends to the maintenance management terminal.

[0047] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0048] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for monitoring asphalt pavement global diseases, comprising obtaining apparent image data, ground penetrating radar data and spatial positioning data of a target road section; characterized in that, further comprising the following steps: inputting the apparent image data into a first feature extraction network to extract road surface disease geometric parameters and assign unique identification codes to generate external disease objects; inputting the ground penetrating radar data into a second feature extraction network to extract internal disease depth parameters and amplitude parameters and assign unique identification codes to generate internal disease objects; based on the spatial positioning data, aligning the attribute fields of the external disease objects and the internal disease objects in a digital twin model to construct a unified multi-dimensional mathematical reorganization structure; obtaining multiple periods of the unified multi-dimensional mathematical reorganization structure in the historical time series, extracting the joint change rate of the external disease objects and the internal disease objects in the same spatial unit, and inputting the joint change rate into a prediction model to output disease evolution data.

2. The asphalt pavement global distress monitoring method of claim 1, wherein, The construction of the unified multi-dimensional mathematical reorganization structure includes dividing the digital twin model into a set of three-dimensional voxel grids with a preset resolution.

3. The asphalt pavement global distress monitoring method of claim 2, wherein, Based on the spatial positioning data and the road surface disease geometric parameters, the external disease objects are mapped to a subset of surface voxels in the set of three-dimensional voxel grids.

4. The asphalt pavement global distress monitoring method of claim 3, wherein, Based on the spatial positioning data and the internal disease depth parameters, the internal disease objects are mapped to a subset of deep voxels in the set of three-dimensional voxel grids, and the spatial intersection attributes between the subset of surface voxels and the subset of deep voxels are extracted.

5. The asphalt pavement global distress monitoring method of claim 1, wherein, The construction of the unified multi-dimensional mathematical reorganization structure includes constructing a heterogeneous graph topology model, and defining the external disease objects and the internal disease objects as a sequence of nodes in the heterogeneous graph topology model.

6. The asphalt pavement global distress monitoring method of claim 5, wherein, When the spatial positioning data distance between an external disease object node and an internal disease object node is within a preset radius threshold, a connected topology edge is established between the external disease object node and the internal disease object node.

7. The asphalt pavement global distress monitoring method of claim 1, wherein, The inputting of the apparent image data into the first feature extraction network to generate external disease objects includes mapping apparent image features to a one-dimensional feature sequence, calling a state space equation to calculate geometric feature points along the disease trend direction, and extracting discrete local width parameters and cumulative skeleton length parameters.

8. The asphalt pavement global distress monitoring method of claim 1, wherein, The inputting of the ground penetrating radar data into the second feature extraction network to generate internal disease objects includes extracting local radar wave reflection feature content, generating an adaptive reorganization kernel for feature map upsampling operation based on the local radar wave reflection feature content.

9. A comprehensive asphalt pavement distress monitoring system, characterized in that, comprising: an acquisition module configured to obtain apparent image data, ground penetrating radar data and spatial positioning data of a target road section; an extraction module configured to input the apparent image data into a first feature extraction network to extract road surface disease geometric parameters and assign unique identification codes to generate external disease objects, and input the ground penetrating radar data into a second feature extraction network to extract internal disease depth parameters and amplitude parameters and assign unique identification codes to generate internal disease objects; a mapping module configured to align the attribute fields of the external disease objects and the internal disease objects in a digital twin model based on the spatial positioning data to construct a unified multi-dimensional mathematical reorganization structure; The deduction module is configured to obtain multiple periods of the unified multidimensional mathematical restructuring structure in the historical time sequence, extract a combined change rate, and input a prediction model to output disease evolution data.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to perform the steps of the asphalt pavement global disease monitoring method according to any one of claims 1 to 8.

Images