A multi-dimensional road defect detection method and system based on ground-penetrating radar

Abstract

This application relates to the field of road detection technology, and in particular to a multi-dimensional road defect detection method and system based on ground-penetrating radar (GPR). The method includes: acquiring raw GPR data; performing DC component removal, gain adjustment, and background denoising on the raw GPR data to locate suspected defect areas and define them as defect entities; extracting time-domain features, frequency-domain features, and spatial context features in parallel for each defect entity to form a multi-dimensional feature vector; constructing a road defect knowledge graph based on the multi-dimensional feature vector and assigning initial weights to each graph relationship; performing node matching operations based on the multi-dimensional feature vector and the knowledge graph; subsequently performing logical reasoning based on the graph relationship paths and fusing similarity and reasoning confidence to output diagnostic label signals and confidence signals. This application improves the accuracy of locating suspected defect areas, realizes the comprehensive utilization of multi-dimensional information, and provides a visualized causal link for diagnosis.

Description

Technical Field

[0001] This application relates to the field of road inspection technology, and in particular to a multi-dimensional road defect detection method and system based on ground-penetrating radar. Background Technology

[0002] Ground-penetrating radar (GPR), as a highly efficient non-destructive testing technology, has been widely used in road defect detection. Traditional detection methods mainly rely on manual visual interpretation or image analysis based on single-amplitude characteristic features, which is insufficient to meet the needs of accurate diagnosis for large-scale and complex road surfaces.

[0003] Existing road defect detection technologies based on ground-penetrating radar have low identification rates for defects with different physical mechanisms but similar images, and their model generalization ability is poor, making it difficult to make accurate judgments when faced with multidimensional and complex road conditions. Summary of the Invention

[0004] This application provides a multi-dimensional road defect detection method and system based on ground-penetrating radar to solve the above-mentioned problems.

[0005] Firstly, this application provides a multi-dimensional road defect detection method based on ground-penetrating radar, the method comprising:

[0006] The raw ground-penetrating radar data is acquired, and DC component removal, gain adjustment and background denoising are performed on the raw ground-penetrating radar data. Then, in the B-Scan image, the suspected disease area is located by energy cluster detection or hyperbola fitting and defined as disease entity.

[0007] For each diseased entity, time-domain features, frequency-domain features, and spatial context features are extracted in parallel to form a multi-dimensional feature vector;

[0008] Based on multidimensional feature vectors, a road disease knowledge graph is constructed. The graph nodes include disease type nodes and road structure nodes. The graph relationships include location relationships, inducing relationships, accompanying relationships, similarity relationships and exclusion relationships. An initial weight is assigned to each graph relationship.

[0009] Based on the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, the cosine similarity is calculated and a node matching operation based on the similarity is performed. Then, logical reasoning is performed based on the graph relationship path and the similarity and reasoning confidence are fused to output the diagnostic label signal and confidence signal.

[0010] The above technical solutions enable systematic preprocessing of raw radar data, improving the signal-to-noise ratio of B-Scan images; dual localization based on energy clusters and hyperbolas improves the accuracy of locating suspected disease areas; parallel extraction of multi-dimensional feature vectors enables comprehensive utilization of time, frequency, time-frequency, and spatial information; structured representation and relational reasoning of knowledge graphs provide a visualized causal link for diagnosis, enhancing interpretability; and weighted fusion of cosine similarity and relational confidence improves the robustness of diagnostic results.

[0011] Optionally, locating suspected disease areas in the B-Scan image by energy cluster detection or hyperbola fitting includes:

[0012] An adaptive thresholding algorithm is used to binarize the B-Scan image, and the adaptive threshold is dynamically calculated based on the signal intensity distribution of local regions of the image.

[0013] Connectivity analysis is performed on the binarized image, and continuous regions with signal energy exceeding a preset threshold are marked as energy clusters;

[0014] The outline of the energy cluster is fitted with a hyperbolic model. When the fitting error is lower than the set tolerance, the core area of ​​the energy cluster is determined as the diseased entity.

[0015] Through the above technical solutions, the adaptive threshold can dynamically adjust the binarization standard based on local statistical characteristics, the connected component analysis can capture continuous energy distribution, and the hyperbola fitting utilizes the typical geometric features of the road structure to achieve more accurate positioning. This enables the detection process to maintain the robustness of defect positioning under different road surface conditions, reduce the false detection rate, and improve the detection sensitivity for minor defects.

[0016] Optionally, the step of extracting time-domain features, frequency-domain features, and spatial context features in parallel for each diseased entity to form a multi-dimensional feature vector includes:

[0017] Extract time-domain features, including maximum amplitude, average energy, root mean square, waveform factor, and number of zero crossings;

[0018] Extracting frequency domain features includes performing a fast Fourier transform on the A-Scan signal corresponding to the diseased entity to obtain the dominant frequency, spectral centroid, spectral variance, and energy proportion of multiple frequency bands;

[0019] Extracting time-frequency domain features includes performing continuous wavelet transform on the A-Scan sequence of the region where the diseased entity is located to obtain the wavelet energy spectrum and wavelet entropy;

[0020] Extract spatial context features, including the burial depth, horizontal scale, hyperbolic opening size, spatial clustering degree, and relative positional relationship with other surrounding entities, from B-Scan and C-Scan data.

[0021] The above technical solution can significantly reduce the total time delay of feature extraction by replacing sequential extraction. Moreover, multi-dimensional features provide a more comprehensive description of the physical properties of the disease, improve the discrimination ability of subsequent similarity calculation, reduce the confusion probability between different diseases, and enhance the robustness of diagnosis.

[0022] Optionally, the construction of the road defect knowledge graph includes:

[0023] The disease type nodes include at least one of surface layer delamination, base layer cracks, and water-rich looseness;

[0024] The road structure node includes at least one of asphalt surface layer and cement-stabilized base layer;

[0025] The term "located in relation" indicates that the defect is located in a specific road structure layer.

[0026] The induced relationship indicates that one disease induces another disease;

[0027] The accompanying relationship indicates that multiple diseases often occur simultaneously;

[0028] The similarity relationship indicates that the two diseases are highly similar in characteristics;

[0029] The exclusion relationship indicates that the two diseases will not occur at the same time.

[0030] Through the above technical solutions, graph structures can unify the modeling of road defects and their interrelationships, facilitating confidence correction using relational information during the reasoning stage. Knowledge graphs provide a semantic network for defects, making the diagnostic process interpretable and providing support for subsequent correlation analysis and maintenance decisions.

[0031] Optionally, the step of calculating the cosine similarity between the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, and performing node matching based on the similarity, followed by logical reasoning based on the graph relationship path and fusing the similarity and reasoning confidence, includes:

[0032] Calculate the cosine similarity between the multidimensional feature vector of the disease entity and the baseline feature vector of each disease type node in the knowledge graph;

[0033] Based on the cosine similarity, the preliminary matching results between the disease entity and the nodes of each disease type are initially determined;

[0034] Logical reasoning is performed on the preliminary matching results based on the relational paths in the knowledge graph, wherein the logical reasoning includes adjusting the confidence level based on induced relations, accompanying relations, similarity relations, or exclusion relations;

[0035] The adjusted confidence level is weighted and fused with the cosine similarity to generate and output the diagnostic label signal and the confidence level signal.

[0036] By introducing graph path reasoning through the above technical solution, the matching results can be corrected a second time using known relationship-based prior information, reducing the fluctuation of confidence in complex scenarios. By fusing feature similarity and structured relationship information, the diagnostic results are robustly improved, and the risk of misjudgment is reduced.

[0037] Optionally, the extraction of spatial clustering in the spatial context features includes:

[0038] Using the current diseased entity as the center, count the number of other diseased entities within a defined spatial neighborhood.

[0039] Calculate the average signal energy of all diseased entities within this neighborhood;

[0040] Based on the product of the quantity and the average signal energy, a quantitative index characterizing the degree of disease clustering is calculated, and the quantitative index is used as the spatial clustering degree.

[0041] By coupling the number of entities with energy intensity, the above technical solution can obtain a more physically meaningful aggregation measure, which can accurately reflect the aggregation characteristics of diseases at the energy level, improve the accuracy of identifying local high-risk areas, provide spatial correlation basis for subsequent confidence adjustment, and enhance the detection sensitivity of clustered diseases.

[0042] Optionally, adjusting the confidence level based on induced relationship, accompanying relationship, similarity relationship, or repulsion relationship includes:

[0043] When there is an induced or accompanying relationship in the knowledge graph from candidate matching node A to another confirmed node B, and node B has been diagnosed as existing in the adjacent spatial location of the current disease entity, the confidence of node A is increased according to the weight ratio of the relationship.

[0044] When there is an exclusion relationship from candidate matching node C to another high-confidence candidate node D in the knowledge graph, the confidence of the lower-confidence node between node C and node D is reduced according to the weight ratio of the relationship.

[0045] When there is a similarity relationship between two candidate matching nodes E and F in the knowledge graph, a discriminative feature comparison is performed on the multidimensional feature vectors of nodes E and F. The discriminative features are predefined feature subsets that can effectively distinguish between these two similar diseases. Based on the degree of difference in the comparison results, the confidence of the one with higher feature matching degree is increased proportionally, and the confidence of the other one is decreased accordingly.

[0046] By using the above technical solution to dynamically adjust the confidence level using relation weights, diagnostic results that are more consistent with actual physical relationships can be achieved. This mechanism improves the discriminative power of diagnostic confidence and reduces the risk of mismatches caused by similar features.

[0047] Optionally, the method further includes:

[0048] After outputting diagnostic label signals and confidence signals, it receives feedback signals from manual confirmation or high-confidence cross-validation.

[0049] The baseline feature vectors of the corresponding nodes and the weights of the relation edges are updated based on manual confirmation or high-confidence cross-validation results. An online learning process is then executed, and the adaptive evolution of the knowledge graph is achieved by updating the node feature vectors and relation edge weights.

[0050] By introducing the above technical solutions and forming a closed loop through manual or cross-validation feedback, the accuracy and timeliness of the map can be continuously improved. This online learning process enables the adaptive evolution of the system and ensures performance maintenance during long-term deployment.

[0051] Optionally, the adaptive evolution of the knowledge graph includes:

[0052] Node vector update: Using the multidimensional feature vector of the diseased entity confirmed by the feedback signal, the baseline feature vector of the corresponding disease type node is updated using the moving average algorithm;

[0053] Relationship weight update: The frequency of co-occurrence and causal relationship of diseases in the feedback signals is statistically analyzed, and the weight of the accompanying and induced relationship is dynamically adjusted based on the statistical frequency. The weight adjustment follows the recency effect principle, that is, the weight of recent feedback data is greater than that of earlier data.

[0054] Through the above technical solutions, the progressive learning of node features and the frequency-driven approach of relation weights work together to maintain the timeliness of the graph and continuously optimize the structured knowledge. This mechanism ensures the self-correction capability of the graph in long-term operation and improves the reliability of diagnosis.

[0055] Secondly, this application provides a multi-dimensional road defect detection system based on ground-penetrating radar, the system comprising:

[0056] The data analysis module is used to acquire raw ground-penetrating radar data, perform DC component removal, gain adjustment and background denoising on the raw ground-penetrating radar data, and then locate suspected disease areas in the B-Scan image by energy cluster detection or hyperbola fitting and define them as disease entities.

[0057] The vector fusion module is used to extract time-domain features, frequency-domain features, and spatial context features in parallel for each diseased entity, forming a multi-dimensional feature vector.

[0058] The graph construction module is used to construct a road disease knowledge graph based on multi-dimensional feature vectors. The graph nodes include disease type nodes and road structure nodes. The graph relationships include location, induction, companion, similarity and exclusion. An initial weight is assigned to each graph relationship.

[0059] The detection and evaluation module is used to calculate the cosine similarity between the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, and perform node matching operation based on the similarity. Then, it performs logical reasoning based on the graph relationship path and integrates the similarity and reasoning confidence to output diagnostic label signal and confidence signal.

[0060] Optionally, the data analysis module is specifically used for:

[0061] An adaptive thresholding algorithm is used to binarize the B-Scan image, and the adaptive threshold is dynamically calculated based on the signal intensity distribution of local regions of the image.

[0062] Connectivity analysis is performed on the binarized image, and continuous regions with signal energy exceeding a preset threshold are marked as energy clusters;

[0063] The outline of the energy cluster is fitted with a hyperbolic model. When the fitting error is lower than the set tolerance, the core area of ​​the energy cluster is determined as the diseased entity.

[0064] Optionally, the vector fusion module is specifically used for:

[0065] Extract time-domain features, including maximum amplitude, average energy, root mean square, waveform factor, and number of zero crossings;

[0066] Extracting frequency domain features includes performing a fast Fourier transform on the A-Scan signal corresponding to the diseased entity to obtain the dominant frequency, spectral centroid, spectral variance, and energy proportion of multiple frequency bands;

[0067] Extracting time-frequency domain features includes performing continuous wavelet transform on the A-Scan sequence of the region where the diseased entity is located to obtain the wavelet energy spectrum and wavelet entropy;

[0068] Extract spatial context features, including the burial depth, horizontal scale, hyperbolic opening size, spatial clustering degree, and relative positional relationship with other surrounding entities, from B-Scan and C-Scan data.

[0069] Optionally, the map construction module is specifically used for:

[0070] The disease type nodes include at least one of surface layer delamination, base layer cracks, and water-rich looseness;

[0071] The road structure node includes at least one of asphalt surface layer and cement-stabilized base layer;

[0072] The term "located in relation" indicates that the defect is located in a specific road structure layer.

[0073] The induced relationship indicates that one disease induces another disease;

[0074] The accompanying relationship indicates that multiple diseases often occur simultaneously;

[0075] The similarity relationship indicates that the two diseases are highly similar in characteristics;

[0076] The exclusion relationship indicates that the two diseases will not occur at the same time.

[0077] Optionally, the detection and evaluation module is specifically used for:

[0078] Calculate the cosine similarity between the multidimensional feature vector of the disease entity and the baseline feature vector of each disease type node in the knowledge graph;

[0079] Based on the cosine similarity, the preliminary matching results between the disease entity and the nodes of each disease type are initially determined;

[0080] Logical reasoning is performed on the preliminary matching results based on the relational paths in the knowledge graph, wherein the logical reasoning includes adjusting the confidence level based on induced relations, accompanying relations, similarity relations, or exclusion relations;

[0081] The adjusted confidence level is weighted and fused with the cosine similarity to generate and output the diagnostic label signal and the confidence level signal.

[0082] Optionally, when extracting the spatial clustering degree in spatial context features, the vector fusion module is specifically used for:

[0083] Using the current diseased entity as the center, count the number of other diseased entities within a defined spatial neighborhood.

[0084] Calculate the average signal energy of all diseased entities within this neighborhood;

[0085] Based on the product of the quantity and the average signal energy, a quantitative index characterizing the degree of disease clustering is calculated, and the quantitative index is used as the spatial clustering degree.

[0086] Optionally, when adjusting confidence based on induced relationships, accompanying relationships, similarity relationships, or repulsion relationships, the graph construction module is specifically used for:

[0087] When there is an induced or accompanying relationship in the knowledge graph from candidate matching node A to another confirmed node B, and node B has been diagnosed as existing in the adjacent spatial location of the current disease entity, the confidence of node A is increased according to the weight ratio of the relationship.

[0088] When there is an exclusion relationship from candidate matching node C to another high-confidence candidate node D in the knowledge graph, the confidence of the lower-confidence node between node C and node D is reduced according to the weight ratio of the relationship.

[0089] When there is a similarity relationship between two candidate matching nodes E and F in the knowledge graph, a discriminative feature comparison is performed on the multidimensional feature vectors of nodes E and F. The discriminative features are predefined feature subsets that can effectively distinguish between these two similar diseases. Based on the degree of difference in the comparison results, the confidence of the one with higher feature matching degree is increased proportionally, and the confidence of the other one is decreased accordingly.

[0090] Optionally, the system further includes a loop iteration module, specifically used for:

[0091] After outputting diagnostic label signals and confidence signals, it receives feedback signals from manual confirmation or high-confidence cross-validation.

[0092] The baseline feature vectors of the corresponding nodes and the weights of the relation edges are updated based on manual confirmation or high-confidence cross-validation results. An online learning process is then executed, and the adaptive evolution of the knowledge graph is achieved by updating the node feature vectors and relation edge weights.

[0093] Optionally, the adaptive evolution of the knowledge graph in the iterative module is specifically used for:

[0094] Node vector update: Using the multidimensional feature vector of the diseased entity confirmed by the feedback signal, the baseline feature vector of the corresponding disease type node is updated using the moving average algorithm;

[0095] Relationship weight update: The frequency of co-occurrence and causal relationship of diseases in the feedback signals is statistically analyzed, and the weight of the accompanying and induced relationship is dynamically adjusted based on the statistical frequency. The weight adjustment follows the recency effect principle, that is, the weight of recent feedback data is greater than that of earlier data. Attached Figure Description

[0096] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0097] Figure 1 This is a schematic diagram of an application scenario provided in an embodiment of this application;

[0098] Figure 2 A flowchart illustrating a multi-dimensional road defect detection method based on ground-penetrating radar, provided as an embodiment of this application;

[0099] Figure 3 This is a schematic diagram of a multi-dimensional road defect detection system based on ground-penetrating radar, provided as an embodiment of this application. Detailed Implementation

[0100] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0101] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0102] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0103] Existing road defect detection technologies based on ground-penetrating radar have low identification rates for defects with different physical mechanisms but similar images, and their model generalization ability is poor, making it difficult to make accurate judgments when faced with multidimensional and complex road conditions.

[0104] Based on this, this application provides a multi-dimensional road defect detection method and system based on ground-penetrating radar, which realizes systematic preprocessing of raw radar data and improves the signal-to-noise ratio of B-Scan images; improves the positioning accuracy of suspected defect areas based on dual localization of energy clusters and hyperbolas; realizes the comprehensive utilization of time, frequency, time-frequency and spatial information through parallel extraction of multi-dimensional feature vectors; provides a visualized causal link for diagnosis through the structured representation and relational reasoning of knowledge graphs, and improves interpretability; and improves the robustness of diagnostic results through weighted fusion of cosine similarity and relational confidence.

[0105] Figure 1 This application provides an illustration of an application scenario. When conducting road defect detection, the method provided in this application can be used to improve the accuracy, interpretability, and robustness of the detection results.

[0106] Specifically, the method provided in this application is applied to any server. The server interacts with ground-penetrating radar (GPR), performing DC component removal, gain adjustment, and background denoising on the raw GPR data. Subsequently, suspected disease areas are located and defined as disease entities in the B-Scan image through energy cluster detection or hyperbolic fitting. For each disease entity, time-domain features, frequency-domain features, and spatial context features are extracted in parallel to form a multi-dimensional feature vector. Based on the multi-dimensional feature vector, a road disease knowledge graph is constructed. Cosine similarity is calculated between the multi-dimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, and node matching based on this similarity is performed. The system performs a matching operation, then performs logical reasoning based on the graph relationship path and fuses similarity and reasoning confidence to provide diagnostic label signals and confidence signals to the corresponding maintenance personnel. This achieves systematic preprocessing of raw radar data, improving the signal-to-noise ratio of B-Scan images. Dual localization based on energy clusters and hyperbolas improves the accuracy of locating suspected disease areas. Parallel extraction of multi-dimensional feature vectors enables comprehensive utilization of time, frequency, time-frequency, and spatial information. The structured representation and relational reasoning of the knowledge graph provide a visualized causal link for diagnosis, enhancing interpretability. The weighted fusion of cosine similarity and relational confidence improves the robustness of the diagnostic results.

[0107] For specific implementation details, please refer to the following examples.

[0108] Figure 2 This is a flowchart illustrating a multi-dimensional road defect detection method based on ground-penetrating radar, provided as an embodiment of this application. The method of this embodiment can be applied to the server in the above scenario. For example... Figure 2 As shown, the method includes:

[0109] S201. Acquire raw ground-penetrating radar data, perform DC component removal, gain adjustment and background denoising on the raw ground-penetrating radar data, and then locate suspected disease areas in the B-Scan image by energy cluster detection or hyperbola fitting and define them as disease entities.

[0110] S202. Extract time-domain features, frequency-domain features, and spatial context features in parallel for each diseased entity to form a multi-dimensional feature vector;

[0111] S203. Construct a road disease knowledge graph based on multidimensional feature vectors. The graph nodes include disease type nodes and road structure nodes. The graph relationships include location relationships, inducing relationships, accompanying relationships, similarity relationships and exclusion relationships. Assign an initial weight to each graph relationship.

[0112] S204. Calculate the cosine similarity based on the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, and perform node matching operation based on the similarity. Then, perform logical reasoning based on the graph relationship path and fuse the similarity and reasoning confidence to output the diagnostic label signal and confidence signal.

[0113] Ground-penetrating radar (GPR) can acquire the electromagnetic response of the road surface medium, but the raw echo signal often contains DC bias, gain unevenness, and background noise, resulting in unclear visualization of defects in B-Scan images. To achieve accurate identification of various defects, it is necessary to locate potential defects using spatial distribution features (energy clusters) or geometric features (hyperboles) after signal preprocessing, and further achieve semantic representation of defects through multi-dimensional feature description. Knowledge graphs provide a structured representation of nodes and relationships, forming a closed loop between feature and graph reasoning.

[0114] The system comprises a data analysis module (performing DC removal, gain adjustment, and background denoising), a vector fusion module (extracting time-domain, frequency-domain, time-frequency, and spatial context features in parallel), a knowledge graph construction module (building a knowledge graph containing disease type nodes, road structure nodes, and five types of relationships), a detection and evaluation module (calculating cosine similarity, performing node matching, performing logical reasoning based on relationship paths, and fusing confidence scores), and an output interface (outputting diagnostic labels and confidence scores). Each module is connected via a high-speed data bus, supporting real-time processing.

[0115] Through the above technical solutions, the following were achieved: (1) systematic preprocessing of the original radar data, which improved the signal-to-noise ratio of B-Scan images; (2) dual localization based on energy clusters and hyperbolas, which improved the localization accuracy of suspected disease areas; (3) parallel extraction of multidimensional feature vectors, which enabled the comprehensive utilization of time, frequency, time-frequency and spatial information; (4) structured representation and relational reasoning of knowledge graphs provided a visualized causal link for diagnosis, which improved interpretability; and (5) weighted fusion of cosine similarity and relational confidence improved the robustness of the diagnostic results.

[0116] Replaceable or modified implementation methods include: (1) using other preprocessing algorithms (such as wavelet denoising, empirical mode decomposition) to replace DC removal and gain adjustment; (2) using a deep learning-based target detection network to replace energy cluster detection or hyperbola fitting in the localization stage; (3) time-domain features can be expanded to peak slope and envelope features; frequency-domain features can be added to phase spectrum or power spectral density; time-frequency features can be different mother wavelets of short-time Fourier transform or continuous wavelet transform; (4) knowledge graphs can be implemented using RDF / OWL semantic models or graph neural network embedding; (5) cosine similarity can be replaced with Euclidean distance or Mahalanobis distance, and confidence fusion can be implemented using Bayesian inference or fuzzy logic.

[0117] In some embodiments, an adaptive thresholding algorithm is used to binarize the B-Scan image. The adaptive threshold is dynamically calculated based on the signal intensity distribution of the local region of the image. Connectivity analysis is performed on the binarized image, and continuous regions with signal energy exceeding a preset threshold are marked as energy clusters. The contours of the energy clusters are fitted with a hyperbolic model. When the fitting error is lower than a set tolerance, the core region of the energy cluster is determined as the disease entity.

[0118] Existing GPR image processing methods often employ fixed thresholds or single geometric models, making it difficult to adapt to variations in road surface materials, humidity, and noise levels. Adaptive thresholding dynamically adjusts the binarization standard based on local statistical characteristics, connected component analysis captures continuous energy distribution, and hyperbolic fitting utilizes typical geometric features of the road structure to achieve more accurate localization.

[0119] Key components of this implementation include: an adaptive thresholding module (for calculating local mean and variance), a binarization unit (for binarizing the B-Scan image and outputting a binary image), a connected component analysis unit (for marking all connected white pixel blocks using the 8-adjacency algorithm, calculating the total pixel energy of each block, and marking the corresponding pixel block as an energy cluster if the total energy is greater than the corresponding energy threshold, and outputting the energy cluster set), a hyperbola fitting unit (for performing least-squares hyperbola fitting on the edge point set P of each energy cluster, and defining the disease entity by taking the fitting center and the size of the curve opening if the residual is less than the corresponding residual threshold tolerance), and a threshold and tolerance parameter storage unit. Intermediate results are transferred between modules via shared memory or a high-speed bus.

[0120] This implementation method can maintain the robustness of disease location under different road surface conditions, reduce the false detection rate, and improve the detection sensitivity of minor diseases.

[0121] Alternative or modified implementations include: (1) using Otsu, local maximum entropy, or deep learning-based segmentation networks instead of adaptive thresholds; (2) replacing connected component analysis with region extraction based on morphological dilation / erosion; (3) using nonlinear least squares based on Levenberg-Marquardt or parameter space search based on Hough transform for hyperbola fitting; and (4) setting the threshold tolerance as an adaptive function that updates in real time with environmental noise.

[0122] In some embodiments, time-domain features are extracted, including maximum amplitude, average energy, root mean square, waveform factor, and number of zero crossings; frequency-domain features are extracted, including performing a fast Fourier transform on the A-Scan signal corresponding to the diseased entity to obtain the dominant frequency, spectral centroid, spectral variance, and energy proportion of multiple frequency bands; time-frequency-domain features are extracted, including performing a continuous wavelet transform on the A-Scan sequence of the region where the diseased entity is located to obtain the wavelet energy spectrum and wavelet entropy; spatial context features are extracted, including obtaining the burial depth, horizontal scale, hyperbola opening size, spatial clustering degree, and relative positional relationship with other surrounding entities of the diseased entity from B-Scan and C-Scan data.

[0123] Traditional GPR (Germ-Related Particle Particle) disease identification often only utilizes time-domain amplitude information, ignoring frequency, time-frequency, and spatial correlation features, making it difficult to distinguish between diseases with similar structures but different physical mechanisms. Parallel feature extraction, instead of sequential extraction, can significantly reduce the total latency of feature extraction, and multi-dimensional features provide a more comprehensive description of the physical properties of the disease.

[0124] The system implementation includes: a time-domain feature extraction unit (performing peak detection, mean, root mean square, waveform factor (peak / root mean square), and zero-crossing count for each A-Scan waveform in a multi-core CPU environment); a frequency-domain feature extraction unit (calling the FFT library to perform an N-point FFT on the A-Scan signal length N, calculating the amplitude spectrum, and thus obtaining the dominant frequency, spectral centroid, spectral variance, and energy proportion of predefined frequency bands); a time-frequency feature extraction unit (using continuous wavelet transform (Morlet is selected as the mother wavelet) to calculate wavelet coefficient energy at multiple scales, forming a wavelet energy spectrum and calculating entropy); a spatial context feature extraction unit (geometric analysis module); and a feature vector assembly unit (reading the B-Scan depth coordinates and C-Scan lateral coordinates, calculating the center burial depth and horizontal scale (lateral projection length) of the diseased entity, obtaining the opening size based on hyperbola fitting parameters; calculating the number of neighboring entities by setting a neighborhood radius, calculating the average neighborhood energy, and thus obtaining the aggregation index; the relative positional relationship is described by vectors relative to the nearest neighbor entity's coordinates). All units execute in parallel on the same computing node and use shared memory for data exchange.

[0125] The introduction of multidimensional feature vectors improves the discriminative ability of subsequent similarity calculations, reduces the probability of confusion between different diseases, and enhances the robustness of diagnosis.

[0126] Replaceable or modified implementations include: (1) adding peak slope and envelope area to time-domain features; (2) using power spectral density or phase spectrum to frequency-domain features; (3) using short-time Fourier transform or multi-resolution analysis to time-frequency features; (4) using density estimation or kernel-based clustering measures to spatially contextual features; and (5) replacing the normalization method of feature vectors with z-score or max-min normalization.

[0127] In some embodiments, the defect type node includes at least one of surface layer voids, base layer cracks, and water-rich looseness; the road structure node includes at least one of asphalt surface layer and cement-stabilized base layer; the location relationship indicates that the defect is located in a specific road structure layer; the induced relationship indicates that one defect induces another defect; the accompanying relationship indicates that multiple defects often occur simultaneously; the similarity relationship indicates that two defects are highly similar in characteristics; and the exclusion relationship indicates that two defects will not occur simultaneously.

[0128] Existing GPR diagnostic systems lack a structured representation of the spatial and causal relationships between defects, leading to subsequent decisions based solely on single-point features. A graph structure can unify the modeling of defects and their interactions with road levels, facilitating confidence correction using relational information during the reasoning phase.

[0129] The implementation includes: a node definition library (creating node labels for each type of defect (e.g., surface voids) and each road structural layer (e.g., asphalt surface layer) in the graph database), relation type enumeration (located in, induced, associated, similar to, and excluded), and a relation weight initialization module (edges for induced, associated, similar to, and excluded are added based on historical co-occurrence statistics, with weights initialized to empirical values ​​(0.3-0.8), which can be dynamically updated in subsequent learning phases; weights are stored using attribute fields (weight), supporting batch updates). Graph storage uses Neo4j or an equivalent attribute graph database, with nodes and edges indexed by unique identifiers.

[0130] This map provides a semantic network of diseases, making the diagnostic process interpretable and supporting subsequent correlation analysis and maintenance decisions.

[0131] Replaceable or modified implementation methods include: (1) nodes can be expanded to more types of diseases (such as frost heave, settlement) or road materials (such as crushed stone base); (2) relation weights can be initialized using learning (such as Bayesian prior based on historical data); (3) the graph database can be replaced with RDF / OWL semantic web or distributed graph computing platform; (4) custom edges such as "residual after repair" and "temporal evolution" can be added to relation types.

[0132] In some embodiments, the cosine similarity between the multidimensional feature vector of the disease entity and the baseline feature vector of each disease type node in the knowledge graph is calculated; based on the cosine similarity, the preliminary matching result between the disease entity and each disease type node is initially determined; according to the relational path in the knowledge graph, logical reasoning is performed on the preliminary matching result, wherein the logical reasoning includes adjusting the confidence based on induced relations, accompanying relations, similarity relations, or exclusion relations; the adjusted confidence and the cosine similarity are weighted and fused to generate and output a diagnostic label signal and a confidence signal.

[0133] Traditional methods rely solely on similarity for matching, neglecting the structured relationships between diseases, leading to significant confidence fluctuations in complex scenarios. By introducing graph path reasoning, known relationship-based prior information can be used to perform secondary correction on the matching results.

[0134] The system implementation includes: a similarity calculation engine (vector dot product and modulus normalization), a node matching scheduler (outputting a candidate matching set based on a threshold), a relation inference engine (traversing graph edges and weighting confidence scores according to preset weights), a confidence score fusion module (using linear weighting or weighted averaging), and a result output interface. All modules are executed serially or in parallel under a unified scheduling framework to ensure real-time performance.

[0135] This solution achieves a robust improvement in diagnostic results and reduces the risk of misjudgment by integrating feature similarity and structured relationship information.

[0136] Replaceable or modified implementation methods include: (1) similarity measure can be replaced by Euclidean distance, Manhattan distance or Mahalanobis distance; (2) inference mechanism can be Bayesian network, graph convolutional network or rule-based forward chain inference; (3) confidence fusion method can be changed to multiplicative model or fuzzy comprehensive; (4) relation weights can be dynamically updated through online learning or manually adjusted by experts.

[0137] In some embodiments, taking the current diseased entity as the center, the number of other diseased entities within a set spatial neighborhood is counted; the average signal energy of all diseased entities within the neighborhood is calculated; based on the product of the number and the average signal energy, a quantitative index characterizing the degree of disease aggregation is calculated, and the quantitative index is used as the degree of spatial aggregation.

[0138] Simple geometric distance cannot reflect the energy-level aggregation characteristics of disease, leading to insufficient identification of local high-risk areas. By coupling the number of entities with energy intensity, a more physically meaningful aggregation metric can be obtained.

[0139] The implementation includes: a neighborhood definition unit (setting a spherical or elliptical search range based on burial depth and horizontal distance), an entity counting unit (traversing the knowledge graph or spatial index), an energy statistics unit (averaging the time-domain or frequency-domain energy of entities within the neighborhood), and a clustering calculation unit (performing product operations and normalizing).

[0140] This indicator provides a spatial correlation basis for subsequent confidence level adjustments and improves the detection sensitivity of clustered diseases.

[0141] The alternative or modified implementation methods include: (1) the neighborhood shape can be changed to a rectangle, hexagon or a strip area based on the road direction; (2) the clustering degree calculation can be carried out by weighted sum (weights are attenuated by distance) or information entropy; (3) the average energy can be changed to peak energy or frequency domain energy ratio.

[0142] In some embodiments, when there is an induced or accompanying relationship from candidate matching node A to another confirmed node B in the knowledge graph, and node B has been diagnosed as existing in the vicinity of the current disease entity, the confidence of node A is increased according to the weight ratio of the relationship; when there is an exclusion relationship from candidate matching node C to another high-confidence candidate node D in the knowledge graph, the confidence of the lower-confidence node between node C and node D is decreased according to the weight ratio of the relationship; when there is a similarity relationship between two candidate matching nodes E and F in the knowledge graph, a discriminative feature comparison is performed on the multidimensional feature vectors of nodes E and F, wherein the discriminative features are predefined feature subsets that can effectively distinguish these two similar diseases; based on the degree of difference in the comparison results, the confidence of the node with higher feature matching degree is increased proportionally, and the confidence of the other node is decreased accordingly.

[0143] Matching based solely on similarity ignores known causal or exclusionary information, potentially leading to an unreasonable distribution of confidence scores. Dynamically adjusting confidence scores through relational weights can achieve diagnostic results that more accurately reflect actual physical relationships.

[0144] Key implementation units include: a relationship detector (which queries the map to obtain AB, CD, and EF relationships), a spatial proximity determination unit (based on burial depth and horizontal distance thresholds), a confidence adjustment unit (which performs multiplication or addition correction based on weight coefficients), and a discriminative feature comparison unit (which performs feature difference calculation on predefined region sets).

[0145] This mechanism improves the discriminative power of diagnostic confidence and reduces the risk of false matches due to similar features.

[0146] Replaceable or modified implementation methods include: (1) relation weights can be set by adaptive learning or by experts manually; (2) discriminative features similar to relations can be replaced by a machine learning-based discriminative model; (3) confidence adjustment strategies can be changed to threshold pruning or fuzzy rules.

[0147] In some embodiments, after outputting diagnostic label signals and confidence signals, feedback signals from manual confirmation or high-confidence cross-validation are received; the baseline feature vectors of the corresponding nodes and the weights of the relation edges are updated based on the results of manual confirmation or high-confidence cross-validation, and an online learning process is performed to achieve adaptive evolution of the knowledge graph by updating the node feature vectors and relation edge weights.

[0148] The one-way output of diagnostic results lacks closed-loop learning and is difficult to adapt to environmental changes. Introducing manual or cross-validation feedback to form a closed loop can continuously improve the accuracy and timeliness of diagnostic maps.

[0149] The implementation includes: a feedback receiving interface (supporting manual annotation or automatic cross-validation), a feature vector update unit (using moving average or exponential weighting), a relation weight update unit (increasing or decreasing based on the frequency of feedback statistics), and a graph persistence module (writing updates back to the graph database).

[0150] This online learning process enables the system to evolve adaptively, ensuring performance maintenance during long-term deployment.

[0151] Replaceable or modified implementations include: (1) the feedback source can be extended to a remote expert system or an automatic label generation model; (2) the feature update algorithm can employ Kalman filtering or incremental learning; and (3) the weight update strategy can employ reinforcement learning or confidence-based adaptive adjustment.

[0152] In some embodiments, node vector update: using the multidimensional feature vector of the disease entity confirmed by the feedback signal, the baseline feature vector of the corresponding disease type node is updated by a moving average algorithm; relation weight update: the frequency of disease co-occurrence and causal association in the feedback signal is statistically analyzed, and the weight of accompanying and induced relationships is dynamically adjusted based on the statistical frequency. The weight adjustment follows the recency effect principle, that is, the influence of recent feedback data on the weight is greater than that of earlier data.

[0153] There are two existing update mechanisms: progressive learning of node features and frequency-driven relation weights. The two, working together, can maintain the timeliness of the graph and continuously optimize structured knowledge.

[0154] The system implementation includes: a sliding average updater (setting the window size or decay coefficient), a frequency statistician (real-time cumulative co-occurrence / causal events), a weight adjuster (calculating the increase or decrease based on the recency effect function), and a version management unit (recording the timestamp and parameters of each update).

[0155] This mechanism ensures the self-correction capability of the spectrum during long-term operation, improving the reliability of diagnosis.

[0156] Replaceable or modified implementation methods include: (1) the moving average can be replaced by the exponentially weighted moving average or recursive least squares; (2) frequency statistics can be replaced by Bayesian update or weighted counting; (3) the recency effect function can be replaced by the exponential decay or linear decay model.

[0157] Figure 3 A schematic diagram of a multi-dimensional road defect detection system based on ground-penetrating radar provided in an embodiment of this application is shown below. Figure 3 As shown, a multi-dimensional road defect detection system 300 based on ground-penetrating radar in this embodiment includes: a data analysis module 301, a vector fusion module 302, a map construction module 303, and a detection and evaluation module 304.

[0158] Data analysis module 301 is used to acquire raw ground-penetrating radar data, perform DC component removal, gain adjustment and background denoising on the raw ground-penetrating radar data, and then locate suspected disease areas in the B-Scan image by energy cluster detection or hyperbola fitting and define them as disease entities.

[0159] The vector fusion module 302 is used to extract time-domain features, frequency-domain features and spatial context features in parallel for each diseased entity to form a multi-dimensional feature vector.

[0160] The graph construction module 303 is used to construct a road disease knowledge graph based on multi-dimensional feature vectors. The graph nodes include disease type nodes and road structure nodes. The graph relationships include location, induction, accompaniment, similarity and exclusion. An initial weight is assigned to each graph relationship.

[0161] The detection and evaluation module 304 is used to calculate the cosine similarity based on the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, and perform node matching operation based on the similarity. Then, it performs logical reasoning based on the graph relationship path and integrates the similarity and reasoning confidence to output the diagnostic label signal and confidence signal.

[0162] Optionally, the data analysis module 301 is specifically used for:

[0163] An adaptive thresholding algorithm is used to binarize the B-Scan image, and the adaptive threshold is dynamically calculated based on the signal intensity distribution of local regions of the image.

[0164] Connectivity analysis is performed on the binarized image, and continuous regions with signal energy exceeding a preset threshold are marked as energy clusters;

[0165] The outline of the energy cluster is fitted with a hyperbolic model. When the fitting error is lower than the set tolerance, the core area of ​​the energy cluster is determined as the diseased entity.

[0166] Optionally, the vector fusion module 302 is specifically used for:

[0167] Extract time-domain features, including maximum amplitude, average energy, root mean square, waveform factor, and number of zero crossings;

[0168] Extracting frequency domain features includes performing a fast Fourier transform on the A-Scan signal corresponding to the diseased entity to obtain the dominant frequency, spectral centroid, spectral variance, and energy proportion of multiple frequency bands;

[0169] Extracting time-frequency domain features includes performing continuous wavelet transform on the A-Scan sequence of the region where the diseased entity is located to obtain the wavelet energy spectrum and wavelet entropy;

[0170] Extract spatial context features, including the burial depth, horizontal scale, hyperbolic opening size, spatial clustering degree, and relative positional relationship with other surrounding entities, from B-Scan and C-Scan data.

[0171] Optionally, the map construction module 303 is specifically used for:

[0172] The disease type nodes include at least one of surface layer delamination, base layer cracks, and water-rich looseness;

[0173] The road structure node includes at least one of asphalt surface layer and cement-stabilized base layer;

[0174] The term "located in relation" indicates that the defect is located in a specific road structure layer.

[0175] The induced relationship indicates that one disease induces another disease;

[0176] The accompanying relationship indicates that multiple diseases often occur simultaneously;

[0177] The similarity relationship indicates that the two diseases are highly similar in characteristics;

[0178] The exclusion relationship indicates that the two diseases will not occur at the same time.

[0179] Optionally, the detection and evaluation module 304 is specifically used for:

[0180] Calculate the cosine similarity between the multidimensional feature vector of the disease entity and the baseline feature vector of each disease type node in the knowledge graph;

[0181] Based on the cosine similarity, the preliminary matching results between the disease entity and the nodes of each disease type are initially determined;

[0182] Logical reasoning is performed on the preliminary matching results based on the relational paths in the knowledge graph, wherein the logical reasoning includes adjusting the confidence level based on induced relations, accompanying relations, similarity relations, or exclusion relations;

[0183] The adjusted confidence level is weighted and fused with the cosine similarity to generate and output the diagnostic label signal and the confidence level signal.

[0184] Optionally, when extracting the spatial clustering degree in the spatial context features, the vector fusion module 302 is specifically used for:

[0185] Using the current diseased entity as the center, count the number of other diseased entities within a defined spatial neighborhood.

[0186] Calculate the average signal energy of all diseased entities within this neighborhood;

[0187] Based on the product of the quantity and the average signal energy, a quantitative index characterizing the degree of disease clustering is calculated, and the quantitative index is used as the spatial clustering degree.

[0188] Optionally, when adjusting the confidence level based on induced relationships, accompanying relationships, similarity relationships, or repulsion relationships, the map construction module 303 is specifically used for:

[0189] When there is an induced or accompanying relationship in the knowledge graph from candidate matching node A to another confirmed node B, and node B has been diagnosed as existing in the adjacent spatial location of the current disease entity, the confidence of node A is increased according to the weight ratio of the relationship.

[0190] When there is an exclusion relationship from candidate matching node C to another high-confidence candidate node D in the knowledge graph, the confidence of the lower-confidence node between node C and node D is reduced according to the weight ratio of the relationship.

[0191] When there is a similarity relationship between two candidate matching nodes E and F in the knowledge graph, a discriminative feature comparison is performed on the multidimensional feature vectors of nodes E and F. The discriminative features are predefined feature subsets that can effectively distinguish between these two similar diseases. Based on the degree of difference in the comparison results, the confidence of the one with higher feature matching degree is increased proportionally, and the confidence of the other one is decreased accordingly.

[0192] Optionally, the multi-dimensional road defect detection system 300 based on ground-penetrating radar further includes a loop iteration module 305, specifically used for:

[0193] After outputting diagnostic label signals and confidence signals, it receives feedback signals from manual confirmation or high-confidence cross-validation.

[0194] The baseline feature vectors of the corresponding nodes and the weights of the relation edges are updated based on manual confirmation or high-confidence cross-validation results. An online learning process is then executed, and the adaptive evolution of the knowledge graph is achieved by updating the node feature vectors and relation edge weights.

[0195] Optionally, the adaptive evolution of the knowledge graph in the iterative loop module 305 is specifically used for:

[0196] Node vector update: Using the multidimensional feature vector of the diseased entity confirmed by the feedback signal, the baseline feature vector of the corresponding disease type node is updated using the moving average algorithm;

[0197] Relationship weight update: The frequency of co-occurrence and causal relationship of diseases in the feedback signals is statistically analyzed, and the weight of the accompanying and induced relationship is dynamically adjusted based on the statistical frequency. The weight adjustment follows the recency effect principle, that is, the weight of recent feedback data is greater than that of earlier data.

[0198] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A multi-dimensional road defect detection method based on ground-penetrating radar, characterized in that, include: Acquire raw ground-penetrating radar (GPR) data, perform DC component removal, gain adjustment, and background denoising on the raw GPR data, and then... In the Scan image, suspected disease areas are located and defined as disease entities by energy cluster detection or hyperbola fitting. For each diseased entity, time-domain features, frequency-domain features, and spatial context features are extracted in parallel to form a multi-dimensional feature vector; Based on multidimensional feature vectors, a road disease knowledge graph is constructed. The graph nodes include disease type nodes and road structure nodes. The graph relationships include location relationships, inducing relationships, accompanying relationships, similarity relationships and exclusion relationships. An initial weight is assigned to each graph relationship. Based on the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, the cosine similarity is calculated and a node matching operation based on the similarity is performed. Then, logical reasoning is performed based on the graph relationship path and the similarity and reasoning confidence are fused to output the diagnostic label signal and confidence signal. The method of locating suspected disease areas in B-Scan images by energy cluster detection or hyperbola fitting includes: An adaptive thresholding algorithm is used to binarize the B-Scan image, and the adaptive threshold is dynamically calculated based on the signal intensity distribution of local regions of the image. Connectivity analysis is performed on the binarized image, and continuous regions with signal energy exceeding a preset threshold are marked as energy clusters; The outline of the energy cluster is fitted with a hyperbolic model. When the fitting error is lower than the set tolerance, the core area of ​​the energy cluster is determined as the disease entity.

2. The method according to claim 1, characterized in that, The process involves extracting temporal, frequency, and spatial context features in parallel for each diseased entity to form a multi-dimensional feature vector, including: Extract time-domain features, including maximum amplitude, average energy, root mean square, waveform factor, and number of zero crossings; Extracting frequency domain features includes performing a fast Fourier transform on the A-Scan signal corresponding to the diseased entity to obtain the dominant frequency, spectral centroid, spectral variance, and energy proportion of multiple frequency bands; Extracting time-frequency domain features includes performing continuous wavelet transform on the A-Scan sequence of the region where the diseased entity is located to obtain the wavelet energy spectrum and wavelet entropy; Extract spatial context features, including the burial depth, horizontal scale, hyperbolic opening size, spatial clustering degree, and relative positional relationship with other surrounding entities, from B-Scan and C-Scan data.

3. The method according to claim 2, characterized in that, The construction of the road defect knowledge graph includes: The disease type nodes include at least one of surface layer delamination, base layer cracks, and water-rich looseness; The road structure node includes at least one of asphalt surface layer and cement-stabilized base layer; The term "located in relation" indicates that the defect is located in a specific road structure layer. The induced relationship indicates that one disease induces another disease; The accompanying relationship indicates that multiple diseases often occur simultaneously; The similarity relationship indicates that the two diseases are highly similar in characteristics; The exclusion relationship indicates that the two diseases will not occur at the same time.

4. The method according to claim 3, characterized in that, The process involves calculating the cosine similarity between the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, performing node matching based on this similarity, and then performing logical reasoning based on the graph relationship path and fusing the similarity and reasoning confidence, including: Calculate the cosine similarity between the multidimensional feature vector of the disease entity and the baseline feature vector of each disease type node in the knowledge graph; Based on the cosine similarity, the preliminary matching results between the disease entity and the nodes of each disease type are initially determined; Logical reasoning is performed on the preliminary matching results based on the relational paths in the knowledge graph, wherein the logical reasoning includes adjusting the confidence level based on induced relations, accompanying relations, similarity relations, or exclusion relations; The adjusted confidence level is weighted and fused with the cosine similarity to generate and output the diagnostic label signal and the confidence level signal.

5. The method according to claim 4, characterized in that, The degree of spatial clustering in the extracted spatial context features includes: Using the current diseased entity as the center, count the number of other diseased entities within a defined spatial neighborhood. Calculate the average signal energy of all diseased entities within this neighborhood; Based on the product of the quantity and the average signal energy, a quantitative index characterizing the degree of disease clustering is calculated, and the quantitative index is used as the spatial clustering degree.

6. The method according to claim 5, characterized in that, The confidence level adjustment based on induced relationship, accompanying relationship, similarity relationship, or exclusion relationship includes: When there is an induced or accompanying relationship in the knowledge graph from candidate matching node A to another confirmed node B, and node B has been diagnosed as existing in the adjacent spatial location of the current disease entity, the confidence of node A is increased according to the weight ratio of the relationship. When there is an exclusion relationship from candidate matching node C to another high-confidence candidate node D in the knowledge graph, the confidence of the lower-confidence node between node C and node D is reduced according to the weight ratio of the relationship. When there is a similarity relationship between two candidate matching nodes E and F in the knowledge graph, a discriminative feature comparison is performed on the multidimensional feature vectors of nodes E and F. The discriminative features are predefined feature subsets that can effectively distinguish between these two similar diseases. Based on the degree of difference in the comparison results, the confidence of the one with higher feature matching degree is increased proportionally, and the confidence of the other one is decreased accordingly.

7. The method according to claim 6, characterized in that, The method further includes: After outputting diagnostic label signals and confidence signals, it receives feedback signals from manual confirmation or high-confidence cross-validation. The baseline feature vectors of the corresponding nodes and the weights of the relation edges are updated based on manual confirmation or high-confidence cross-validation results. An online learning process is then executed, and the adaptive evolution of the knowledge graph is achieved by updating the node feature vectors and relation edge weights.

8. The method according to claim 7, characterized in that, The adaptive evolution of the knowledge graph includes: Node vector update: Using the multidimensional feature vector of the diseased entity confirmed by the feedback signal, the baseline feature vector of the corresponding disease type node is updated using the moving average algorithm; Relationship weight update: The frequency of co-occurrence and causal relationship of diseases in the feedback signals is statistically analyzed, and the weight of the accompanying and induced relationship is dynamically adjusted based on the statistical frequency. The weight adjustment follows the recency effect principle, that is, the weight of recent feedback data is greater than that of earlier data.

9. A multi-dimensional road defect detection system based on ground-penetrating radar, characterized in that, The method applied to any one of claims 1-8 includes: The data analysis module is used to acquire raw ground-penetrating radar (GPR) data, perform DC component removal, gain adjustment, and background denoising on the raw GPR data, and then... In the Scan image, suspected disease areas are located and defined as disease entities by energy cluster detection or hyperbola fitting. The vector fusion module is used to extract time-domain features, frequency-domain features, and spatial context features in parallel for each diseased entity, forming a multi-dimensional feature vector. The graph construction module is used to construct a road disease knowledge graph based on multi-dimensional feature vectors. The graph nodes include disease type nodes and road structure nodes. The graph relationships include location, induction, companion, similarity and exclusion. An initial weight is assigned to each graph relationship. The detection and evaluation module is used to calculate the cosine similarity between the multidimensional feature vector and the baseline feature vector of each disease node in the knowledge graph, and perform node matching operation based on the similarity. Then, it performs logical reasoning based on the graph relationship path and integrates the similarity and reasoning confidence to output diagnostic label signal and confidence signal.

Images