A method for identifying tunnel defects based on dielectric distribution maps

Abstract

This invention relates to the field of tunnel defect identification technology, and more particularly to a tunnel defect identification method based on dielectric distribution maps. This method utilizes AI geotechnical radar field monitoring, gprMax simulation, and laboratory electromagnetic data to generate a two-dimensional dielectric distribution map using variational Bayesian inversion fusion. Through multi-scale spectral clustering and expert knowledge, potential anomaly regions are automatically identified. Subsequently, graph morphological constraint propagation and an integrated Transformer-graph neural network model are employed to achieve accurate classification and identification of defect regions. Finally, the identification results are projected onto the original map to generate a visualized defect annotation layer and a structured report. This invention achieves high-precision, automated, visualized, and structured detection of tunnel structural defects, significantly improving the intelligent level and risk warning capabilities of tunnel safety operation and maintenance, and promoting the digitalization and intelligentization of tunnel operation and maintenance management.

Description

Technical Field

[0001] This invention relates to the field of tunnel defect identification technology, and more particularly to a tunnel defect identification method based on dielectric distribution maps. Background Technology

[0002] As a crucial component of transportation infrastructure, the safety and durability of tunnel structures directly impact the stable operation of transportation systems. However, due to factors such as complex geological environments, diverse construction techniques, and environmental changes during service life, tunnels are prone to various structural defects, including voids, water seepage, cracks, and cavities. Current technologies for identifying and assessing internal tunnel defects largely rely on manual inspections or image recognition and radar image analysis. Among these, ground-penetrating radar (GPR), with its non-contact, rapid detection and penetration capabilities, can acquire information on the distribution of dielectric materials within the tunnel structure, thereby revealing potential structural defects.

[0003] However, due to the low signal-to-noise ratio, complex image, and irregular texture of dielectric distribution maps, traditional manual interpretation is inefficient and inaccurate, failing to meet the needs of automatic identification and classification of structural defects in complex tunnel scenarios. Currently, the following problems also exist: The widespread use of manual analysis and judgment of dielectric images in current engineering projects relies on expert experience, resulting in inconsistent judgment standards, poor repeatability, and low efficiency, making it difficult to meet the needs of large-scale tunnel inspection; different types of defects (such as voids, cracks, and water damage) exhibit irregularities and blurred boundaries on dielectric distribution maps, and traditional image processing algorithms (such as threshold segmentation and edge detection) are insensitive to these complex morphologies, resulting in low recognition rates; there is a lack of end-to-end automated identification methods, especially in real-world scenarios with multiple defect types and significant image interference, where existing solutions struggle to quickly complete identification, classification, and localization, limiting their intelligent application in engineering. Existing methods rely only on a single type of detection data (such as single radar measurements or simulation data), lacking multi-source information fusion, leading to insufficient accuracy in representing the complex dielectric distribution within tunnel structures and failing to comprehensively and accurately reflect defect characteristics. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a tunnel defect identification method based on dielectric distribution maps. This method solves the problem of how to achieve automated, intelligent, and high-precision tunnel defect identification and classification under multi-source data fusion when existing tunnel defect detection methods face challenges such as low signal-to-noise ratio of dielectric distribution maps, complex defect types, and low efficiency of manual interpretation. This method meets the actual needs of large-scale tunnel structural health monitoring and improves the intelligent level and risk warning capabilities of tunnel safety operation and maintenance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A tunnel defect identification method based on dielectric distribution maps includes the following steps:

[0007] S1: The actual monitoring data inside the tunnel is obtained by AI geotechnical radar. Combined with simulation data based on gprMax and electromagnetic experimental data obtained in the laboratory, the data is fused and reconstructed using the variational Bayesian inversion method to generate a two-dimensional dielectric distribution image sequence of the tunnel structure.

[0008] S2: Based on the two-dimensional dielectric distribution image sequence, a multi-scale spectral clustering algorithm is used, and a dielectric constant-defect type mapping model is constructed by combining expert knowledge and statistical labels. The two-dimensional dielectric distribution image sequence is then scanned and compared regionally to extract potential abnormal regions under threshold interval conditions and generate an initial defect candidate layer.

[0009] S3: Based on the initial defect candidate layer, a spatial adjacency graph is constructed using a graph morphology constraint propagation algorithm to jointly model the spatial connectivity and boundary morphology of the defect region. An optimized structural defect mask layer is generated through edge weight evolution and morphology preservation mechanism.

[0010] S4: Based on the structural defect mask layer, construct a multi-class defect recognition model integrating Transformer structure and graph neural network algorithm, classify and identify defect regions, and output defect recognition results; the defect recognition results include defect type label, location coordinates and classification confidence score;

[0011] S5: Project the defect identification result onto the original frame of the two-dimensional dielectric distribution image sequence to generate a visual defect annotation layer containing defect type, spatial location and confidence level, and output a structured inspection report.

[0012] Furthermore, the field monitoring data includes multi-polarized echo signals of the tunnel inner wall collected by AI geotechnical radar, three-dimensional spatial coordinates of each measuring point, reflection intensity amplitude matrix, time-domain waveform, instantaneous energy distribution, time delay characteristic parameters, and acquisition timestamp information;

[0013] The simulation data includes electromagnetic wave propagation path data under different tunnel structure sections obtained from the gprMax platform simulation, including target structure model, medium property configuration, radar source parameters, echo response data, reflection coefficient data, and signal-to-noise ratio calculation results;

[0014] The electromagnetic experimental data includes the measured dielectric constant data of the laboratory-prepared standard specimens, the electromagnetic wave penetration loss values ​​of each specimen at different frequencies, echo waveform data, amplitude-frequency characteristic curves, typical defect signal templates, extreme points and time delay parameters, and the measured values ​​of reflectivity and transmittance in each frequency band.

[0015] Furthermore, step S2 includes the following steps:

[0016] The dielectric constant is extracted pixel by pixel from the two-dimensional dielectric distribution image sequence, and a dielectric constant evolution tensor is constructed based on the image frame order.

[0017] The spatiotemporal tensor of dielectric constant is divided into multi-scale sliding windows to obtain local region blocks at different scales. Based on the dielectric constant distribution of the pixels contained in each region block, statistical feature parameters are calculated, including mean, variance, skewness, kurtosis and local gradient entropy.

[0018] The statistical feature parameters of multi-scale regional blocks are summarized to construct a feature vector set. A multi-scale spectral clustering algorithm is used to perform joint feature embedding and Laplacian feature decomposition to generate a cluster label layer.

[0019] Based on the clustering label layer, a dielectric constant-defect type mapping model is constructed by combining expert knowledge and statistical labels. The central feature vector of each candidate region is matched with the typical defect template for similarity, and the corresponding dielectric constant threshold range is set according to the confidence interval boundary conditions.

[0020] Based on the threshold range, the clustered regions are scanned block by block to extract regions with dielectric constant offset, discontinuous structural texture, or template matching confidence exceeding the threshold. These regions are marked as potential abnormal regions, and an initial defect candidate layer is output.

[0021] Furthermore, the formula for the dielectric constant-defect type mapping model is as follows:

[0022]

[0023] in, Represents the feature vector of the region block Belongs to the i-th type of defect Probability score; This indicates the label for the i-th type of defect, including categories such as voids, cracks, water content, and corrosion; This represents the multidimensional feature vector of the current candidate region block; This represents the feature mean vector of the i-th type of defect; Indicates the defect type The characteristic covariance matrix; d represents the eigenvectors. dimensionality; This represents the squared Mahalanobis distance between the current feature of the region to be identified and the defect template. Denotes the inverse of the covariance matrix; This represents the transpose operation of a vector.

[0024] Furthermore, step S3 includes the following steps:

[0025] By extracting spatial features from potential anomalous regions in the initial defect candidate layer, a spatial adjacency graph is constructed based on the region's centroid coordinates, boundary geometric properties, and dielectric constant gradient distribution.

[0026] The spatial adjacency graph is iteratively modeled using a graph morphology constraint propagation algorithm. The propagation mechanism incorporates edge morphology preservation constraints, regional connectivity weight adjustment factors, and structural texture priors to drive the structured propagation of defect information and adaptive evolution of edge weights within the graph structure.

[0027] During the propagation process, the weight coefficients of each graph edge are dynamically adjusted, and the local boundary curvature changes, regional morphological compactness, and defect template morphological similarity indicators are integrated.

[0028] Based on the output of the graph structure after propagation stabilization, the main defect region with high structural coherence and clear boundary morphology is extracted, edge fragments and isolated noise are removed, and an optimized structural defect mask layer is generated.

[0029] Furthermore, step S4 includes the following steps:

[0030] Based on the optimized structural defect mask layer, multi-scale feature extraction is performed on each defect candidate region, including dielectric constant distribution features, regional spatial geometric features, boundary contour morphological parameters, and difference indicators with the surrounding normal regions, and the above features are standardized.

[0031] The multi-scale features of each defect region are input into a multi-class defect recognition model that integrates Transformer structure and graph neural network algorithm. The Transformer module is used to extract the global contextual correlation between defect regions, while the graph neural network algorithm is used to capture the spatial connectivity and graph structure dependency of defect regions.

[0032] In the multi-class defect identification model, based on the multi-label supervision mechanism, the multi-class defect type is judged for each input defect region, and the spatial location coordinates and classification confidence score of each type of defect are output simultaneously.

[0033] The labels, coordinates, and confidence scores of the identified defects are mapped onto the original dielectric distribution map or tunnel structure diagram, automatically generating a visual defect labeling layer with multiple defect types, spatial distributions, and confidence levels, thus forming a structured defect identification result.

[0034] Furthermore, the formula for the multi-type defect identification model is as follows:

[0035]

[0036] in, This represents the probability that the m-th defect candidate region belongs to the n-th type of defect in the dielectric distribution map of the t-th frame, which is the final classification confidence score; This represents the Sigmoid normalization function; represents the normalization factor; K represents the number of adjacent regions of the m-th candidate region in the spatial adjacency graph; This represents the spatial adjacency weight between the m-th region and the k-th adjacent region in frame t. This represents the average dielectric constant of the m-th region within its k-th adjacent region; The average dielectric constant of the standard template region representing the k-th type of defect in frame t; This represents the standard deviation of the dielectric constant of the standard template region for the k-th type of defect in frame t. This represents the Euclidean distance between the spatial centroid coordinates of the m-th candidate region and the spatial center of the standard template for the n-th type of defect. This represents the average distance between all candidate regions and the center of the template; This represents the Hausdorff distance between the m-th region and the boundary of the n-th type of defect template in the t-th frame; This represents the average Hausdorff distance between all candidate regions and the boundaries of various templates; , and Represents the weight parameters for multi-feature fusion; This represents the bias term for the nth type of defect in the tth frame.

[0037] Furthermore, the spatial adjacency graph specifically uses the spatial centroid or main boundary point of the abnormal region as graph nodes, and edges are used to connect nodes that are spatially adjacent, have similar dielectric properties, or have continuous boundary shapes. The weight of the edges is jointly defined by the spatial distance between nodes, boundary continuity, and electromagnetic property similarity.

[0038] Furthermore, the structured inspection report includes an overview map of the distribution of multiple types of defects, a statistical chart of defect intensity levels classified by category, a confidence level table, a superimposed map of defects in key structural parts, and an evolution trend analysis chart.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention combines measured data from AI geotechnical radar, gprMax simulation data, and laboratory electromagnetic experimental data, employing a variational Bayesian inversion method to achieve efficient fusion and structural reconstruction of multi-source data. This effectively improves the realism and accuracy of dielectric distribution images, providing a more reliable data foundation for subsequent defect detection. A multi-scale spectral clustering algorithm is introduced, integrating expert knowledge and statistical labels to establish a mapping relationship between dielectric constant and defect type. This enables precise partitioning and identification of various potential defects in tunnel structures, effectively improving the detection rate of abnormal areas and the preliminary screening capability for multiple types of defects. By constructing a spatial adjacency graph and introducing a graph morphology constraint propagation mechanism, the spatial connectivity and boundary morphology of defect areas are jointly modeled, further improving the accuracy of defect segmentation and boundary discrimination, and significantly reducing the probability of false detection and missed detection. Integrating a Transformer structure and a graph neural network, deep feature mining and multimodal information fusion are performed on different types of defect areas, achieving intelligent classification and confidence assessment of multiple types of defects. The output includes structured detection results such as type, location, and confidence score, significantly enhancing the comprehensiveness and intelligence of defect identification. The identification results are projected onto the original dielectric diagram or tunnel structure diagram, automatically generating a visual annotation layer with defect type, spatial location and confidence level, and outputting a structured inspection report. This allows on-site engineers to intuitively understand the defect distribution and health status, improving the scientific nature and efficiency of subsequent operation, reinforcement and management decisions. Attached Figure Description

[0041] Figure 1 This is a schematic flowchart of a tunnel defect identification method based on dielectric distribution map according to the present invention.

[0042] Figure 2 This is a dielectric distribution comparison diagram provided in an embodiment of the present invention.

[0043] Figure 3 This is a flowchart illustrating step S2 provided in an embodiment of the present invention.

[0044] Figure 4 This is a flowchart illustrating step S4 provided in an embodiment of the present invention. Detailed Implementation

[0045] Please see Figure 1-4 As shown, this invention relates to a tunnel defect identification method based on dielectric distribution maps.

[0046] Example

[0047] A tunnel defect identification method based on dielectric distribution maps includes the following steps:

[0048] S1: The actual monitoring data inside the tunnel is obtained by AI geotechnical radar. Combined with simulation data based on gprMax and electromagnetic experimental data obtained in the laboratory, the data is fused and reconstructed using the variational Bayesian inversion method to generate a two-dimensional dielectric distribution image sequence of the tunnel structure.

[0049] The on-site monitoring data includes multi-polarized echo signals of the tunnel inner wall collected by AI geotechnical radar, three-dimensional spatial coordinates of each measuring point, reflection intensity amplitude matrix, time-domain waveform, instantaneous energy distribution, time delay characteristic parameters, and acquisition timestamp information;

[0050] The simulation data includes electromagnetic wave propagation path data under different tunnel structure sections obtained from the gprMax platform simulation, including target structure model, medium property configuration, radar source parameters, echo response data, reflection coefficient data, and signal-to-noise ratio calculation results;

[0051] The electromagnetic experimental data includes the measured dielectric constant data of the laboratory-prepared standard specimens, the electromagnetic wave penetration loss values ​​of each specimen at different frequencies, echo waveform data, amplitude-frequency characteristic curves, typical defect signal templates, extreme points and time delay parameters, and the measured values ​​of reflectivity and transmittance in each frequency band.

[0052] In one embodiment, the on-site monitoring data collection is specifically as follows;

[0053] (1) Measurement point layout: AI geotechnical radar measurement points are laid out in the tunnel at the designed interval (e.g., every 2 meters) to ensure full tunnel coverage.

[0054] (2) Radar parameter settings: Set the radar transmission frequency (e.g., 800MHz~2GHz), transmission polarization mode (vertical, horizontal, oblique) and waveform parameters.

[0055] (3) Multi-polarization data acquisition: The echo signal of the tunnel inner wall is acquired in sequence using different polarization modes to obtain the original waveform signal of each measuring point.

[0056] (4) Spatiotemporal calibration: All data are automatically associated with the three-dimensional coordinates of the measuring point (using IMU + odometer), timestamp, and structural information such as tunnel ring number and mileage, to achieve unified calibration of multi-source heterogeneous data.

[0057] (5) Parameter feature extraction: Automatically extract parameters such as reflection intensity amplitude matrix, time domain waveform, instantaneous energy distribution, main peak and valley points, and main reflection delay of the echo signal at each measuring point to construct the original data feature matrix.

[0058] It should be noted that this application uses an independently developed AI-based geotechnical imaging radar platform, the core of which is an integrated multi-polarization ground-penetrating radar unit (supporting HH, VV, and HV polarization signal acquisition), installed on a track-mounted inspection robot. The radar frequency is 1.5 GHz, the spatial resolution is 0.2 meters, and the maximum penetration depth can reach 1.2 meters.

[0059] The hardware components are as follows:

[0060] Main control terminal: integrates high-performance embedded computing units (such as FPGA+ARM) and is responsible for signal acquisition and control, data caching and preprocessing.

[0061] Transmission module: It has broadband / multi-frequency transmission capability (typical bandwidth 400MHz-3GHz), supports multi-polarization switching, and adapts to the needs of different structures and defect detection.

[0062] Receiver array: Multi-channel reception, real-time acquisition of echo signals with different polarizations, supporting high-precision A / D conversion (resolution typically ≥12 bits, sampling rate ≥1GSps).

[0063] Assisted positioning: Equipped with an inertial measurement unit (IMU) and a high-precision odometer / laser rangefinder, it enables real-time calibration of the three-dimensional coordinates of the measurement point.

[0064] Synchronous clock: Supports multi-channel data synchronous acquisition and timestamp recording to ensure spatial and temporal data alignment.

[0065] The software and intelligent algorithm section is detailed below:

[0066] Data acquisition and control software: automatically completes sampling parameter settings (such as polarization, frequency, sampling rate), multi-channel synchronous management, anomaly detection, and acquisition process monitoring.

[0067] Waveform feature extraction module: Extracts key physical features from raw echo data in real time, including main reflection peak, time delay, energy envelope, attenuation rate, etc.

[0068] AI intelligent recognition engine: Built-in shallow neural network / decision tree model performs preliminary anomaly screening and noise suppression on the collected waveform signals, realizing intelligent preliminary selection on-site "collection and judgment".

[0069] Self-learning function: The system can automatically optimize waveform processing parameters based on historical samples and feedback results, improving its adaptability and sensitivity to complex defects.

[0070] The intelligent data acquisition and automated operation process includes the following steps: Based on tunnel structure modeling and on-site 3D point cloud information, the system automatically plans the radar movement path and acquisition point distribution to achieve full coverage and high-density scanning of key areas. The system can automatically identify high-risk areas based on historical detection data and dynamically adjust the sampling interval and measurement point density. It supports automatic switching of different polarization modes (such as vertical-horizontal-oblique) at the same measurement point to capture electromagnetic wave responses in different directions, improving sensitivity to various types of defects. It has frequency conversion transmission capability, automatically selecting the optimal frequency band for different structural depths and material types, achieving compatible detection of deep and shallow defects. The radar system uses AI algorithms to evaluate the signal quality (such as signal-to-noise ratio and echo integrity) of each sampling point in real time, automatically adjusting the transmission power, receiving gain, and sampling rate. It provides intelligent alarms and automatic re-sampling for abnormal or interference points (such as interference from metal components or signal obstruction), reducing data loss.

[0071] By combining IMU with odometry / laser scanning, high-precision reconstruction of radar movement trajectories is achieved, mapping each set of echo data to the actual spatial location of the tunnel structure. Real-time integration with BIM / CAD tunnel structural models is supported, automatically mapping data acquisition points to key structural components. It can also link with other monitoring systems (such as deformation monitoring and temperature / humidity sensing) to create multi-dimensional structural health monitoring datasets.

[0072] AI-powered geotechnical imaging radar not only provides conventional radar echoes but also outputs real-time multi-polarization, frequency-converted, and multi-channel high-dimensional raw data, significantly improving the imaging accuracy of media distribution. On-site intelligent algorithms preliminarily identify abnormal waveforms and remove noise, providing high-quality, structured foundational data for subsequent Bayesian inversion and high-order data fusion. Intelligent and automated processes simplify manual intervention, achieving an integrated closed-loop process of "automated data acquisition - intelligent initial selection - high-precision calibration - multi-source fusion."

[0073] The specific construction of gprMax simulation data is as follows:

[0074] (1) Three-dimensional structural modeling: In gprMax, establish a structural geometry model consistent with the actual tunnel, including lining thickness, inner and outer contours, and the shape and location of typical defects (such as voids, seepage, inclusions, etc.).

[0075] (2) Medium property configuration: Set the dielectric constant and conductivity of concrete, rock, soil, defect area, etc. respectively to ensure that they are consistent with the actual material properties.

[0076] (3) Radar parameter simulation: Set the parameters of the transmitting source (frequency, polarization, pulse width, and transmitting position) and perform full-wave electromagnetic simulation to obtain the path data and echo response of radar waves propagating, reflecting, and scattering in different cross sections.

[0077] (4) Feature extraction: Record the echo signal, reflection coefficient, signal-to-noise ratio (SNR), standard response waveforms and spatial distribution characteristics of various defects obtained from the simulation.

[0078] The specific data acquisition for laboratory electromagnetic experiments is as follows:

[0079] (1) Preparation of standard specimens: Prepare a batch of concrete specimens containing known defect types (such as voids, cracks, steel corrosion, etc.) according to the on-site material ratio.

[0080] (2) Experimental testing: On the laboratory electromagnetic testing platform, conduct multi-frequency (e.g., 500MHz, 1GHz, 2GHz) transmission and reflection experiments on the specimen, and collect the dielectric constant, penetration loss, reflection waveform, and amplitude-frequency characteristics at different frequencies.

[0081] (3) Feature summarization: Typical signal templates for each type of defect (such as extreme points, main time delays, echo amplitude ranges, frequency band responses of reflectivity and transmittance, etc.) are compiled and a standard database is established to provide comparative samples for subsequent inversion and identification.

[0082] The variational Bayesian inversion data fusion is detailed below:

[0083] (1) Preliminary normalization: The three types of data (field, simulation and experiment) are processed and normalized with unified dimensions to eliminate differences in amplitude, resolution and sampling interval.

[0084] (2) Prior modeling: Using experimental data as the prior distribution and combining simulation data to supplement the structural change pattern, a Bayesian probability model of dielectric distribution is constructed.

[0085] (3) Observation modeling: The field radar echo is used as the observation data. The variational Bayes method is adopted to infer by maximizing the posterior probability (MAP) and dynamically adjusting the weight of each data source to achieve adaptive information fusion.

[0086] The two-dimensional dielectric distribution image sequence is generated as follows:

[0087] (1) Pixel inversion: Based on the fused data, the dielectric constant distribution of the tunnel cross section is calculated pixel by pixel through the inversion algorithm, so as to realize data reconstruction with a spatial resolution higher than that of the original radar sampling.

[0088] (2) Image output: Based on the longitudinal profile of the tunnel, the output is a two-dimensional dielectric distribution image sequence with a resolution of up to millimeters, and each frame corresponds to a sampling section.

[0089] (3) Quality control: Interpolate or remove outliers (such as low signal-to-noise ratio or data loss areas) to ensure the continuity and usability of the reconstructed image.

[0090] (4) Structural labeling: Automatically associates the actual structural location (such as ring number, mileage, orientation, etc.) with each distribution map, providing a precise coordinate basis for subsequent defect location and tracing.

[0091] In one embodiment, combined with the appendix Figure 2 The content is as follows: the gray image above is a traditional radio radar scan, which is not easy to interpret. The colored image below is a dielectric distribution map output by the AI ​​model. Materials such as concrete, steel bars, and air correspond to different dielectric constants, which are reflected in the image as white, orange, and green, respectively.

[0092] The orange background area in the color image represents the main region of the dielectric distribution map generated after AI geotechnical radar and multi-source inversion processing. This region is a concrete structure with a relatively uniform dielectric constant distribution, stable dielectric properties, strong signal reflection continuity, and is generally a normal, defect-free background area.

[0093] The green blocky area in the center of the color image represents an anomaly inversion result, indicating a region with a significantly lower dielectric constant than the surrounding background. This is commonly found in structural defects such as voids, cavities, water pockets, or locally loose layers. It has a clear boundary with the background, making it easily identifiable by intelligent algorithms. The purple border surrounding the green area is the labeled outline of the defect candidate region identification results, representing potential structural anomalies detected by the model. The dark blue narrow bars at the top and bottom edges of the color image are image boundary occlusion or positioning reference lines; they do not participate in structural judgment but help with image region alignment and coordinate unification.

[0094] S2: Based on the two-dimensional dielectric distribution image sequence, a multi-scale spectral clustering algorithm is used, and a dielectric constant-defect type mapping model is constructed by combining expert knowledge and statistical labels. The two-dimensional dielectric distribution image sequence is then scanned and compared regionally to extract potential abnormal regions under threshold interval conditions and generate an initial defect candidate layer.

[0095] Step S2 includes the following steps:

[0096] The dielectric constant is extracted pixel by pixel from the two-dimensional dielectric distribution image sequence, and a dielectric constant evolution tensor is constructed based on the image frame order.

[0097] Specifically, in the obtained two-dimensional dielectric distribution image sequence, technicians used an automated batch processing method to extract the dielectric constant value of each frame at the pixel level. All frames were sorted according to time or acquisition order to form a three-dimensional spatiotemporal data volume (evolution tensor), that is, a three-dimensional correspondence of "pixel position - frame number - dielectric constant", reflecting the dynamic evolution of the dielectric properties of the tunnel structure.

[0098] The spatiotemporal tensor of dielectric constant is divided into multi-scale sliding windows to obtain local region blocks at different scales. Based on the dielectric constant distribution of the pixels contained in each region block, statistical feature parameters are calculated, including mean, variance, skewness, kurtosis and local gradient entropy.

[0099] Specifically, based on the actual dimensions of the tunnel structure and the spatial scale of typical defects, multiple window sizes (such as 8×8, 16×16, 32×32 pixels, etc.) and different sliding step sizes are selected to ensure that both minute defects and large-scale structural anomalies can be covered. The system automatically traverses the entire image, performing complete sliding coverage for each time frame and each set of window scales, ensuring no blind spots in the detection.

[0100] For each region block covered by the window, the system calculates the following statistical parameters one by one:

[0101] Mean: The average dielectric constant of all pixels in this region, reflecting the overall dielectric properties of the region.

[0102] Variance: The degree of fluctuation in the dielectric constant within this region, which helps to detect structural inhomogeneities or abnormal dielectric disturbances.

[0103] Skewness and kurtosis: used to describe the symmetry of distribution within a region and the central tendency of outliers, and are very sensitive to distinguishing normal and abnormal regions.

[0104] Local gradient entropy: By calculating the entropy value of the dielectric constant gradient image within a region, it measures the boundary complexity and texture abrupt changes within the region, and is an important indicator for identifying boundary-type defects such as cracks and voids.

[0105] All features are stored using a triple index of window, frame, and scale to ensure traceability and multidimensional analysis.

[0106] The feature parameters of each region block are packaged into a feature vector, along with its time frame, spatial location, and window scale label, providing a data foundation for subsequent clustering and comparison.

[0107] The statistical feature parameters of multi-scale regional blocks are summarized to construct a feature vector set. A multi-scale spectral clustering algorithm is used to perform joint feature embedding and Laplacian feature decomposition to generate a cluster label layer.

[0108] It should be noted that the specific process of multi-scale spectral clustering includes: summarizing the feature vectors of all windows and all frames according to spatial coordinates and scale indexes to form a high-dimensional feature matrix; constructing a feature similarity map between regions (e.g., based on Euclidean distance or cosine similarity), where each region is a node in the graph, and nodes with high similarity are connected by edges; using a spectral clustering algorithm, performing feature decomposition on the data by calculating the Laplacian matrix of the feature similarity map, and automatically finding the optimal segmentation (e.g., distinguishing normal background, boundary transition areas, and high-risk abnormal areas); and the system performing cluster labeling on the original image based on the decomposition results, assigning a cluster label to each pixel, with the number of labels automatically adjusted according to the actual anomaly type and clustering algorithm parameters.

[0109] After clustering, the system generates a "cluster label layer" of the same size as the original distribution map, with each cluster encoded by a different color or grayscale value. This layer provides a regional basis for subsequent defect type mapping and threshold screening based on expert knowledge.

[0110] Based on the clustering label layer, a dielectric constant-defect type mapping model is constructed by combining expert knowledge and statistical labels. The central feature vector of each candidate region is matched with the typical defect template for similarity, and the corresponding dielectric constant threshold range is set according to the confidence interval boundary conditions.

[0111] Specifically, based on years of tunnel inspection and laboratory research, the engineering team has established a "feature template library" that includes various types of defects such as typical cracks, voids, water seepage, honeycomb pitting, and foreign object encapsulation. Each template contains:

[0112] Statistical feature vector of typical defect areas (combination of multiple parameters such as mean, variance, skewness, kurtosis, and entropy).

[0113] Description of common dielectric constant ranges and distribution patterns when defects occur;

[0114] Typical spatial shapes, areas, and boundary attributes (e.g., elongated, blocky, point-like, etc.);

[0115] The confidence intervals and defect level assessment criteria for historical test samples are marked by experts.

[0116] For the central features of each cluster region, the system sequentially compares the similarity with the central features of each defect type in the template library (such as calculating Euclidean distance, Mahalanobis distance, feature crossover probability, etc.). Combining expert experience and historical data, the system sets the "dielectric constant threshold range" and "similarity confidence limit" for each type of defect.

[0117] If the central feature of a certain region falls within the threshold range of a certain type of defect, and its similarity with the template library is higher than the empirical threshold, it will be automatically identified as a high-suspected region of that type of defect.

[0118] For complex or incompletely matched areas, the system can automatically report them as "requiring manual review" or "low-confidence defect area" for subsequent manual screening and confirmation.

[0119] The system supports experts in reviewing and annotating the identification results. Newly identified defect areas can be reported in real time. The system can automatically expand the template library and adjust the threshold and confidence settings to achieve self-learning and continuous optimization of the model.

[0120] Based on the threshold range, the clustered regions are scanned block by block to extract regions with dielectric constant offset, discontinuous structural texture, or template matching confidence exceeding the threshold. These regions are marked as potential abnormal regions, and an initial defect candidate layer is output.

[0121] Specifically, for the cluster label layer, each cluster region is scanned and compared block by block according to the dielectric constant threshold range of the defect type to screen for the following situations:

[0122] The average or maximum / minimum dielectric constant of the region block exceeds the threshold for the corresponding defect type;

[0123] Significant discontinuities are observed in local textures (detected by abrupt changes in local gradient entropy, kurtosis, and variance).

[0124] The similarity score of the matched template is higher than the confidence threshold (e.g., similarity > 0.85);

[0125] Regions that meet any of the above anomaly criteria are automatically marked as potential anomaly regions, and are aggregated to generate an "initial defect candidate layer" (which is a binary or multi-class label matrix).

[0126] The final output is an initial defect candidate layer that corresponds directly to the original two-dimensional dielectric distribution map in space, and provides high-quality suspected defect input for subsequent steps such as morphological modeling and depth recognition.

[0127] Furthermore, the formula for the dielectric constant-defect type mapping model is as follows:

[0128]

[0129] in, Represents the feature vector of the region block Belongs to the i-th type of defect Probability score; This indicates the label for the i-th type of defect, including categories such as voids, cracks, water content, and corrosion; The multidimensional feature vector representing the current candidate region block includes the following five statistical dimensions: mean dielectric constant, variance, skewness, kurtosis, and local gradient entropy within the region. The feature parameters are derived from the sliding window statistical extraction processing based on the two-dimensional dielectric distribution image sequence in step S2. The original data includes the reflection intensity amplitude matrix obtained by AI geotechnical radar and the dielectric property configuration in the gprMax simulation data. The mean vector of the i-th type of defect is used to characterize the statistical center position of this type of defect in the above five-dimensional feature space. The mean vector is obtained based on the typical defect model constructed from the simulation data of the gprMax platform, the radar wave source response data, and the penetration loss values ​​of the laboratory-prepared specimens at different frequencies. Indicates the defect type The characteristic covariance matrix; d represents the eigenvectors. dimensionality; This represents the squared Mahalanobis distance between the current feature of the region to be identified and the defect template. Denotes the inverse of the covariance matrix; This represents the transpose operation of a vector.

[0130] Indicates the defect type The eigencovariance matrix, whose elements The covariance between the e-th and l-th feature dimensions is represented by the following formula:

[0131]

[0132] in, Indicates the defect type The covariance between the e-th feature dimension and the l-th feature dimension; This represents the total number of training samples belonging to the i-th type of defect; This indicates that the q-th defect belongs to the defect type. The value of the sample in the k-th statistical feature dimension; Indicates the defect type The statistical mean on the e-th feature dimension; and Let represent the value of the q-th sample on the l-th feature dimension and the mean of the defect type on that dimension, respectively.

[0133] S3: Based on the initial defect candidate layer, a spatial adjacency graph is constructed using a graph morphology constraint propagation algorithm to jointly model the spatial connectivity and boundary morphology of the defect region. An optimized structural defect mask layer is generated through edge weight evolution and morphology preservation mechanism.

[0134] Step S3 includes the following steps:

[0135] By extracting spatial features from potential anomalous regions in the initial defect candidate layer, a spatial adjacency graph is constructed based on the region's centroid coordinates, boundary geometric properties, and dielectric constant gradient distribution. Specifically, the spatial adjacency graph uses the spatial centroid or main boundary points of the anomalous region as graph nodes, and edges are used to connect nodes that are spatially adjacent, have similar dielectric properties, or have continuous boundary shapes. The weight of the edges is jointly defined by the spatial distance between nodes, boundary continuity, and electromagnetic property similarity.

[0136] Specifically, the system automatically traverses the initial defect candidate layer, labels all connected components (i.e., pixel-connected abnormal blocks), and assigns a unique ID to each potential abnormality. For each abnormal region, the system calculates the spatial centroid (i.e., the average of all pixel coordinates) based on the (x, y) coordinates of all pixels, obtaining the center position of the region in the entire layer for subsequent spatial relationship modeling. The system automatically extracts the boundary pixels of each region, forming a boundary point sequence. It calculates the perimeter, principal direction (principal axis), maximum / minimum enclosing rectangle, local boundary curvature (e.g., calculating the change in tangent angle every 10 pixels to reflect boundary bending), and boundary continuity (e.g., boundaries without breaks have high continuity). Within each abnormal region, the system calculates the dielectric constant gradient (i.e., the rate of change of dielectric constant between adjacent pixels) inside and on the boundary. Boundaries with high gradient values ​​are marked to highlight clearly demarcated defects, aiding in subsequent spatial relationship judgment.

[0137] Node definition: The centroid, key points of the main boundary direction, or points with obvious characteristics on the boundary (such as points with maximum curvature) of each abnormal region are defined as nodes of the adjacency graph. Node attributes include spatial location, area, dielectric constant distribution, morphological type, etc.

[0138] The spatial adjacency determination and edge construction are as follows:

[0139] The system automatically traverses all node pairs and determines whether the following conditions are met: the spatial distance is within a set threshold (e.g., 30 pixels); the difference in the mean / variance of the dielectric constant of the corresponding regions of the node is lower than an empirical threshold (e.g., the difference is less than 0.15); the angle between the principal directions of the two regions is small, the boundary lines are close, and there is local boundary overlap; there is a spatial connection channel between the regions (e.g., the pixel distance between the two regions is less than 2 pixels). If any of the above conditions are met, the system automatically creates an edge for the node pair and assigns an initial weight.

[0140] The initial weights of edges are determined by the following criteria: the closer the spatial distance, the higher the weight; the more similar the electromagnetic properties, the higher the weight; and the more continuous / smooth the boundary line shape, the higher the weight. Conversely, for node pairs that are far apart in space, have large differences in properties, or have abrupt boundary changes, the initial weights are low.

[0141] Nodes, edges, and weights are recorded using adjacency lists or adjacency matrices, which facilitates efficient algorithm iteration and attribute queries.

[0142] The spatial adjacency graph is iteratively modeled using a graph morphology constraint propagation algorithm. The propagation mechanism incorporates edge morphology preservation constraints, regional connectivity weight adjustment factors, and structural texture priors to drive the structured propagation of defect information and adaptive evolution of edge weights within the graph structure.

[0143] Specifically, using each node (anomaly region) as an information source, the system simulates the propagation of "information flow" in the spatial adjacency graph. Initially, each node possesses its own spatial characteristics, morphological attributes, and dielectric constant distribution, transmitting "defect attribution probability" information to neighboring nodes. During propagation, information flow is preferentially transmitted along edges with similar boundary morphology and gentle changes in boundary curvature, ensuring the continuity of the overall outline and the true shape as the defect region expands. For node pairs with abrupt changes in boundary curvature or repulsive boundary morphologies, the propagation intensity is weakened or interrupted to avoid false clustering caused by crossing physical / material boundaries.

[0144] Nodes with large and compact connected regions are prioritized for enhanced propagation, while nodes with scattered shapes, small areas, or isolated from major defect areas are gradually weakened in propagation weight, thus gradually filtering out isolated noise. If adjacent nodes show convergence in their belonging probabilities after the first few rounds of propagation, their connectivity weight is automatically increased to enhance the formation of the backbone structure.

[0145] During propagation, each node is continuously compared with the standard defect template in terms of texture, spatial distribution, and boundary attributes. If the morphological characteristics of certain nodes are found to highly match the known template during propagation, their priority for being absorbed into the main defect region is significantly increased, driving the entire graph structure to gradually converge towards the true defect morphology.

[0146] After each iteration, the system updates the "defect attribution confidence" of all nodes based on the information transmitted between nodes and the results of edge weight adjustment. The edge weights are dynamically increased or decreased based on the local performance after propagation, forming self-reinforcement in high-confidence defect areas and self-weakening in low-confidence noise areas.

[0147] When the change in the ownership probability of most nodes is lower than a set threshold (e.g., 0.01), or the number of propagation rounds reaches the upper limit (e.g., 10 rounds), the propagation is considered to have converged. After convergence, the system automatically outputs the connected main regions with ownership probabilities higher than the threshold as the main body of the optimization defects, removes edges, isolated points, etc., and generates a final defect mask layer with a clear structure and coherent boundaries.

[0148] During the propagation process, the weight coefficients of each graph edge are dynamically adjusted, and the local boundary curvature changes, regional morphological compactness, and defect template morphological similarity indicators are integrated.

[0149] Specifically, after each round of propagation, the system automatically analyzes the local boundary curvature changes, regional compactness (such as area / perimeter ratio), and template morphological similarity of each edge-connected node, and dynamically fine-tunes the weight coefficients of each edge:

[0150] If adjacent nodes are more consistent in morphology, spatial and electromagnetic characteristics, the edge weight is further increased, which facilitates information aggregation.

[0151] If morphological breaks, attribute mutations, or template mismatches occur, the edge weights are automatically reduced to minimize the impact of "pseudo-connectivity".

[0152] After multiple iterations, the edge weights stabilize into a set of states that truly reflect the spatial coherence and boundary morphology of the region.

[0153] Based on the output of the graph structure after propagation stabilization, the main defect region with high structural coherence and clear boundary morphology is extracted, edge fragments and isolated noise are removed, and an optimized structural defect mask layer is generated.

[0154] S4: Based on the structural defect mask layer, construct a multi-class defect recognition model integrating Transformer structure and graph neural network algorithm, classify and identify defect regions, and output defect recognition results; the defect recognition results include defect type label, location coordinates and classification confidence score;

[0155] Step S4 includes the following steps:

[0156] Based on the optimized structural defect mask layer, multi-scale feature extraction is performed on each defect candidate region, including dielectric constant distribution features, regional spatial geometric features, boundary contour morphological parameters, and difference indicators with the surrounding normal regions, and the above features are standardized.

[0157] Specifically, using the optimized structural defect mask layer obtained from S3, all connected defect regions are automatically identified, and each region is automatically assigned a unique ID. For each defect region, the system automatically extracts rich features at multiple spatial scales, including:

[0158] Dielectric constant distribution characteristics: statistical mean, variance, kurtosis, distribution of local extrema, etc., reflect the stability and anomalies of regional electromagnetic properties.

[0159] Spatial geometric features: area, perimeter, principal axis direction, compactness, shape factor, and centroid coordinates, which help identify different types of structural anomalies (such as slender cracks and honeycomb-like structures).

[0160] Boundary contour parameters, such as boundary length, curvature variation, boundary smoothness, and complexity, are used to distinguish the morphological characteristics of different defects.

[0161] Difference index: A measure of the difference between the defect area and its adjacent normal area in terms of dielectric constant, boundary morphology, etc., to help the model learn the boundary between the defect and the background.

[0162] Normalize and standardize the feature data of all regions (such as Z-score or Min-Max scaling) to eliminate the influence of scale, measurement conditions and distribution bias, so as to make the model input data structure consistent and improve the robustness and generalization ability of subsequent model discrimination.

[0163] The multi-scale features of each defect region are input into a multi-class defect recognition model that integrates Transformer structure and graph neural network (GNN) algorithm. The Transformer module is used to extract the global contextual correlation between defect regions, while the graph neural network algorithm is used to capture the spatial connectivity and graph structure dependency of defect regions.

[0164] Specifically, for each candidate defect region, the system first encodes its multi-scale features (such as dielectric constant statistics, spatial morphology, boundary parameters, and differences from surrounding regions) into a high-dimensional feature vector, and records the region ID, spatial coordinates, and the mask region number to which it belongs. All defect regions are arranged according to spatial distribution or topological relationship to form a feature input sequence, facilitating the Transformer module's processing of global context information between regions. Simultaneously, the adjacency relationships between regions (such as adjacent, overlapping, and distance less than a set threshold) are automatically constructed into an attribute graph, with regions as nodes, and edges established between spatially and attribute-proximity regions, preparing for subsequent modeling in the GNN algorithm.

[0165] Each feature vector incorporates location information (such as absolute spatial coordinates or relative position to the tunnel's main axis) to ensure the model can recognize spatial distribution patterns. The sequence of feature vectors from all regions is then input into the Transformer's encoder layer.

[0166] The Transformer uses a self-attention layer to calculate the feature correlation score between each region and all other regions. This allows it to detect phenomena such as multiple cracks arranged in bands or spatial correlations among different types of defects. The model can automatically identify large-scale, non-local structural defect features, effectively compensating for the blind spots of local features in single regions.

[0167] Using each defect region as a node, the system automatically determines whether to connect edges based on spatial distance, boundary contact, and attribute similarity, and assigns attribute weights to each edge (e.g., higher scores for closer spatial proximity and smaller attribute differences). Through graph convolution or message passing mechanisms, GNNs ensure that each node, after multiple rounds of computation, not only contains its own features but also incorporates spatial and attribute information from its neighborhood. This helps the model capture complex spatial dependencies such as regional clustering, multi-level structural defects, and abnormal distribution trends. For defects that overlap, connect in series, or cluster (such as honeycomb-like pitted areas or intersecting crack areas), GNNs can effectively integrate their overall morphological features, enhancing classification accuracy.

[0168] Finally, the model concatenates or weights the global feature embeddings output by the Transformer with the spatial structure embeddings output by the GNN to form the final representation of each region. This representation will simultaneously reflect the region's "global distribution features + spatial clustering trends + local detailed information," providing a solid data foundation for subsequent classification.

[0169] In the multi-class defect identification model, based on the multi-label supervision mechanism, the multi-class defect type is judged for each input defect region, and the spatial location coordinates and classification confidence score of each type of defect are output simultaneously.

[0170] Specifically, during the model training phase, engineers manually labeled each sample region with all real-world defect labels (e.g., if the same region has both "crack" and "water seepage" labels), and then combined the category labels of all regions into a multi-label binary vector. The model employs multi-label loss functions (such as binary cross-entropy, Focal Loss, etc.) to optimize the discriminator to support multi-class output for single regions, thereby improving the ability to identify complex defects.

[0171] During the actual inference phase, the model outputs probability scores for all categories in parallel for each region (e.g., "crack" 0.95, "seepage" 0.18, "honeycomb" 0.03, etc.), allowing multiple labels to coexist. The system determines which defect type the region belongs to based on a preset threshold (e.g., 0.5), supporting automatic determination of high-confidence regions and optional "to be reviewed" labeling for low-confidence regions.

[0172] The system automatically outputs the spatial coordinates of each identified defect area, including the region's centroid, a list of boundary points, or the minimum bounding box, ensuring accurate location on the original distribution map or structural diagram. Each defect type corresponds to a confidence probability score, reflecting the model's confidence level in the judgment. The system supports tiered display of classification confidence levels (e.g., above 0.8 for high confidence, 0.5-0.8 for moderate confidence, and below 0.5 for doubtful confidence), which can be used for subsequent maintenance prioritization or review process formulation.

[0173] All identification results are automatically generated into structured tables, including fields such as region number, type label, probability of various defects, spatial coordinates, area, and confidence level, facilitating one-click export and data archiving by engineers. The system directly overlays all classification, spatial, and confidence level results onto the original dielectric distribution map or tunnel structure diagram. It supports displaying different defect types with different colors, shapes, and transparency, highlighting high-confidence areas and distinguishing low-confidence intervals with dashed lines or light colors, aiding on-site operation and maintenance and decision-making.

[0174] The labels, coordinates, and confidence scores of the identified defects are mapped onto the original dielectric distribution map or tunnel structure diagram, automatically generating a visual defect labeling layer with multiple defect types, spatial distributions, and confidence levels, thus forming a structured defect identification result.

[0175] Specifically, the system automatically maps all identified defect types, spatial coordinates, confidence scores, and other results back onto the original two-dimensional dielectric distribution map or three-dimensional tunnel structure schematic diagram.

[0176] Different defect types are visually coded using color, shape, and transparency. Areas with high confidence are highlighted, while low-confidence areas are marked with dashed lines or low saturation. Users can view detailed attributes and links to the original data for each defect area through an interactive interface.

[0177] The system automatically generates structured output, including: an overview map of the distribution of multiple types of defects, a spatial distribution heat map of each type of defect, and structured table files (such as Excel, CSV, database, etc.) containing defect type, coordinates, and confidence level. These can be used for subsequent inspection report generation, risk assessment, and operation and maintenance decisions.

[0178] Furthermore, the formula for the multi-type defect identification model is as follows:

[0179]

[0180] in, This represents the probability that the m-th defect candidate region belongs to the n-th type of defect in the dielectric distribution map of the t-th frame, which is the final classification confidence score; This represents the Sigmoid normalization function; represents the normalization factor; K represents the number of adjacent regions of the m-th candidate region in the spatial adjacency graph; This represents the spatial adjacency weight between the m-th region and the k-th adjacent region in frame t. This represents the average dielectric constant of the m-th region within its k-th adjacent region; The average dielectric constant of the standard template region representing the k-th type of defect in frame t; This represents the standard deviation of the dielectric constant of the standard template region for the k-th type of defect in frame t. This represents the Euclidean distance between the spatial centroid coordinates of the m-th candidate region and the spatial center of the standard template for the n-th type of defect. This represents the average distance between all candidate regions and the center of the template; This represents the Hausdorff distance between the m-th region and the boundary of the n-th type of defect template in the t-th frame; This represents the average Hausdorff distance between all candidate regions and the boundaries of various templates; , and Represents the weight parameters for multi-feature fusion; This represents the bias term for the nth type of defect in the tth frame.

[0181] The calculation formula is as follows:

[0182]

[0183] in, This represents the spatial centroid coordinates of region m in frame t. This represents the spatial centroid coordinates of the k-th adjacent region in frame t. This represents the mean distance between the centroids of all candidate regions; This represents the average dielectric constant of the m-th region; This represents the average dielectric constant of the k-th adjacent region; This represents the average Hausdorff distance between the boundaries of all candidate regions; The Hausdorff distance represents the boundary contours of the m-th region and the k-th adjacent region. , and This represents the weight parameters for feature fusion.

[0184] The calculation formula is as follows:

[0185]

[0186] in, This represents the set of boundary points of the m-th region in the t-th frame; Represents the set of boundary points of the k-th adjacent region; It represents the Euclidean distance between boundary points x and y, with coordinates directly derived from the pixel space of the dielectric distribution map; It represents the smallest upper bound of a set, that is, the smallest value of the largest set; It represents the maximum lower bound of the set, that is, the minimum maximum value.

[0187] S5: Project the defect identification results onto the original frame of the two-dimensional dielectric distribution image sequence to generate a visual defect annotation layer containing defect type, spatial location and confidence level, and output a structured inspection report; the structured inspection report includes an overview map of the distribution of multiple types of defects, a statistical chart of defect intensity levels classified by category, a confidence level table, a defect overlay map of key structural parts and an evolution trend analysis map.

[0188] It should be noted that the original frame refers to the original image frame in the two-dimensional dielectric distribution image sequence that has not undergone defect annotation or post-processing.

[0189] The recognition results can also be projected onto a tunnel structure diagram; the tunnel structure diagram refers to a two-dimensional or three-dimensional structure diagram generated based on a tunnel structure design model or surveying data, which is used for spatial registration and visualization overlay with the recognition results.

[0190] Specifically, the defect type, spatial coordinates, and confidence level information output by S4 are first standardized. The center points, bounding boxes, and contours of all defect areas are then spatially registered with the original two-dimensional dielectric distribution map and the tunnel structure CAD diagram. Image registration algorithms (such as affine transformation, feature point matching, and nearest neighbor interpolation) are used to ensure accurate overlay of the defect identification results on the structural diagram. For data from different inspection batches or monitoring cycles, spatiotemporal synchronization and evolution alignment of multi-phase results are also required.

[0191] Based on the registered spatial data, a visual defect annotation layer is automatically generated. The following methods are used to enhance readability and usability: different types of defects are marked with highly distinguishable colors and shapes (e.g., cracks are red curves, voids are blue polygons, and seepage is green areas). Each defect area is overlaid with its type label (text), confidence score (numerical / level), and spatial coordinates. Defect areas with high confidence are highlighted with thick borders; areas with low confidence are rendered with dashed lines or transparency, allowing engineers to focus on high-risk areas.

[0192] Supports hierarchical display: It can display a specific type of defect, the overall distribution of all defects, or filter by confidence level or regional attributes.

[0193] It provides an interactive interface that allows users to click, zoom, and query detailed attribute data for each defect (such as detection time, historical evolution, and relationships with adjacent defects).

[0194] Based on all detection and identification data, the system automatically generates structured detection reports. Typical outputs include:

[0195] Defect Distribution Overview Map: Displays the spatial distribution of all identified defects throughout the entire tunnel section. Combined with location information such as tunnel number, mileage marker, and cross-section number, it helps the maintenance team quickly grasp the overall health status.

[0196] Classification of Defect Intensity Statistics: The number, distribution density, and intensity level (e.g., minor / moderate / severe) of different types of defects are statistically analyzed using bar charts, pie charts, heat maps, etc., and the corresponding spatial range is marked.

[0197] Confidence Level Table: All defects are sorted from highest to lowest confidence level, and the table provides the category, location, area, confidence score, and whether a review is recommended. Each record can be traced back to the original identification layer and collected information.

[0198] Overlay diagram of defects in key structural components: For special key structures in the tunnel (such as arch, sidewalls, floor, joints, secondary lining, etc.), the diagram highlights all types of defects and their changing trends detected in the past and present in a zoned manner, which facilitates accurate location and customized repair recommendations.

[0199] Evolutionary trend analysis chart: When multiple periods of detection data are available, a trend analysis curve is generated based on the time series data of the defect area, including indicators such as the change of defect area over time, type conversion rate, and recurrence rate after repair, to help judge the process of structural deterioration and the evolution of hidden dangers.

[0200] In summary, this application fully integrates field monitoring data from AI geotechnical imaging radar, gprMax simulation data, and laboratory electromagnetic experimental data. Through variational Bayesian inversion, it achieves deep fusion and high-precision reconstruction of multi-source data, significantly improving the characterization accuracy of the dielectric constant distribution inside the tunnel. Leveraging the advantages of AI geotechnical imaging radar—multi-polarization, frequency conversion, and high spatial resolution—it can acquire rich raw physical signals and multi-dimensional feature data, providing a solid data foundation for subsequent defect identification. Compared with traditional single-radar scanning techniques, this method not only obtains more comprehensive and detailed electromagnetic feature information but also overcomes monitoring challenges under actual working conditions such as data noise and spatial obstruction.

[0201] This application achieves dynamic and fine-grained analysis of the dielectric constant of tunnel structures by generating two-dimensional dielectric distribution image sequences and extracting statistical features at the pixel, frame, and spatiotemporal tensor levels. The combination of a multi-scale spectral clustering algorithm and an expert knowledge base effectively identifies various potential anomaly regions and achieves accurate classification through a dielectric constant-defect type mapping model, significantly improving the accuracy and sensitivity of defect identification. For typical defects such as voids, cracks, seepage, and honeycombing, this method automatically identifies high-suspection defect areas through statistical templates and threshold filtering, greatly reducing the subjectivity of manual interpretation and the risk of missed detections.

[0202] This application employs a graph morphology constraint propagation algorithm to construct a spatial adjacency graph, jointly modeling the spatial connectivity, boundary morphology, and attribute similarity of defect regions. This effectively suppresses noise and isolated points, maintains the true morphology and coherent structure of defect regions, and optimizes the defect mask layer. Subsequently, a multi-class defect recognition model integrating a Transformer structure and a graph neural network organically fuses global contextual information with spatial dependencies. This model can not only identify single defects but also accurately handle scenarios with multiple coexisting defects and complex structural interactions, significantly improving the model's generalization ability and adaptability to complex scenarios.

[0203] The defect labels, spatial coordinates, and classification confidence scores identified in this application are automatically mapped back to the original distribution map, enabling one-click generation of multi-type defect visualization annotation layers and structured inspection reports. The system supports multi-batch, multi-temporal data registration and evolution trend analysis, providing visualized, quantitative, and intelligent support for tunnel structure health monitoring, risk warning, and maintenance decision-making. The structured output greatly facilitates risk assessment and maintenance planning for the operation and maintenance team, improving inspection efficiency and the scientific nature of decision-making.

[0204] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying tunnel defects based on dielectric distribution maps, characterized in that, Includes the following steps: S1: The actual monitoring data inside the tunnel is obtained by AI geotechnical radar. Combined with simulation data based on gprMax and electromagnetic experimental data obtained in the laboratory, the data is fused and reconstructed using the variational Bayesian inversion method to generate a two-dimensional dielectric distribution image sequence of the tunnel structure. S2: Based on the two-dimensional dielectric distribution image sequence, a multi-scale spectral clustering algorithm is used, and a dielectric constant-defect type mapping model is constructed by combining expert knowledge and statistical labels. The two-dimensional dielectric distribution image sequence is then scanned and compared regionally to extract potential abnormal regions under threshold interval conditions and generate an initial defect candidate layer. S3: Based on the initial defect candidate layer, a spatial adjacency graph is constructed using a graph morphology constraint propagation algorithm to jointly model the spatial connectivity and boundary morphology of the defect region. An optimized structural defect mask layer is generated through edge weight evolution and morphology preservation mechanism. S4: Based on the structural defect mask layer, construct a multi-class defect recognition model integrating Transformer structure and graph neural network algorithm, classify and identify defect regions, and output defect recognition results; The defect identification results include the defect type label, location coordinates, and classification confidence score; S5: Project the defect identification result onto the original frame of the two-dimensional dielectric distribution image sequence to generate a visual defect annotation layer containing defect type, spatial location and confidence level, and output a structured inspection report.

2. The tunnel defect identification method based on dielectric distribution map according to claim 1, characterized in that, The field monitoring data includes multi-polarized echo signals of the tunnel inner wall collected by AI geotechnical radar, three-dimensional spatial coordinates of each measuring point, reflection intensity amplitude matrix, time-domain waveform, instantaneous energy distribution, time delay characteristic parameters, and acquisition timestamp information; The simulation data includes electromagnetic wave propagation path data under different tunnel structure sections obtained from the gprMax platform simulation, including target structure model, medium property configuration, radar source parameters, echo response data, reflection coefficient data, and signal-to-noise ratio calculation results; The electromagnetic experimental data includes the measured dielectric constant data of the laboratory-prepared standard specimens, the electromagnetic wave penetration loss values ​​of each specimen at different frequencies, echo waveform data, amplitude-frequency characteristic curves, typical defect signal templates, extreme points and time delay parameters, and the measured values ​​of reflectivity and transmittance in each frequency band.

3. The tunnel defect identification method based on dielectric distribution map according to claim 1, characterized in that, Step S2 includes the following steps: The dielectric constant is extracted pixel by pixel from the two-dimensional dielectric distribution image sequence, and a dielectric constant evolution tensor is constructed based on the image frame order. The spatiotemporal tensor of dielectric constant is divided into multi-scale sliding windows to obtain local region blocks at different scales. Based on the dielectric constant distribution of the pixels contained in each region block, statistical feature parameters are calculated, including mean, variance, skewness, kurtosis and local gradient entropy. The statistical feature parameters of multi-scale regional blocks are summarized to construct a feature vector set. A multi-scale spectral clustering algorithm is used to perform joint feature embedding and Laplacian feature decomposition to generate a cluster label layer. Based on the clustering label layer, a dielectric constant-defect type mapping model is constructed by combining expert knowledge and statistical labels. The central feature vector of each candidate region is matched with the typical defect template for similarity, and the corresponding dielectric constant threshold range is set according to the confidence interval boundary conditions. Based on the threshold range, the clustered regions are scanned block by block to extract regions with dielectric constant offset, discontinuous structural texture, or template matching confidence exceeding the threshold. These regions are marked as potential abnormal regions, and an initial defect candidate layer is output.

4. The tunnel defect identification method based on dielectric distribution map according to claim 3, characterized in that, The formula for the dielectric constant-defect type mapping model is as follows: in, Represents the feature vector of the region block Belongs to the i-th type of defect Probability score; This indicates the label for the i-th type of defect, including categories such as voids, cracks, water content, and corrosion; This represents the multidimensional feature vector of the current candidate region block; This represents the feature mean vector of the i-th type of defect; Indicates the defect type The characteristic covariance matrix; d represents the eigenvectors. dimensionality; This represents the squared Mahalanobis distance between the current feature of the region to be identified and the defect template. Denotes the inverse of the covariance matrix; This represents the transpose operation of a vector.

5. The tunnel defect identification method based on dielectric distribution map according to claim 1, characterized in that, Step S3 includes the following steps: By extracting spatial features from potential anomalous regions in the initial defect candidate layer, a spatial adjacency graph is constructed based on the region's centroid coordinates, boundary geometric properties, and dielectric constant gradient distribution. The spatial adjacency graph is iteratively modeled using a graph morphology constraint propagation algorithm. The propagation mechanism incorporates edge morphology preservation constraints, regional connectivity weight adjustment factors, and structural texture priors to drive the structured propagation of defect information and adaptive evolution of edge weights within the graph structure. During the propagation process, the weight coefficients of each graph edge are dynamically adjusted, and the local boundary curvature changes, regional morphological compactness, and defect template morphological similarity indicators are integrated. Based on the output of the graph structure after propagation stabilization, the main defect region with high structural coherence and clear boundary morphology is extracted, edge fragments and isolated noise are removed, and an optimized structural defect mask layer is generated.

6. The tunnel defect identification method based on dielectric distribution map according to claim 1, characterized in that, Step S4 includes the following steps: Based on the optimized structural defect mask layer, multi-scale feature extraction is performed on each defect candidate region, including dielectric constant distribution features, regional spatial geometric features, boundary contour morphological parameters, and difference indicators with the surrounding normal regions, and the above features are standardized. The multi-scale features of each defect region are input into a multi-class defect recognition model that integrates Transformer structure and graph neural network algorithm. The Transformer module is used to extract the global contextual correlation between defect regions, while the graph neural network algorithm is used to capture the spatial connectivity and graph structure dependency of defect regions. In the multi-class defect identification model, based on the multi-label supervision mechanism, the multi-class defect type is judged for each input defect region, and the spatial location coordinates and classification confidence score of each type of defect are output simultaneously. The labels, coordinates, and confidence scores of the identified defects are mapped onto the original dielectric distribution map or tunnel structure diagram, automatically generating a visual defect labeling layer with multiple defect types, spatial distributions, and confidence levels, thus forming a structured defect identification result.

7. The tunnel defect identification method based on dielectric distribution map according to claim 6, characterized in that, The formula for the multi-type defect identification model is as follows: in, This represents the probability that the m-th defect candidate region belongs to the n-th type of defect in the dielectric distribution map of the t-th frame, which is the final classification confidence score; This represents the Sigmoid normalization function; represents the normalization factor; K represents the number of adjacent regions of the m-th candidate region in the spatial adjacency graph; This represents the spatial adjacency weight between the m-th region and the k-th adjacent region in frame t. This represents the average dielectric constant of the m-th region within its k-th adjacent region; The average dielectric constant of the standard template region representing the k-th type of defect in frame t; This represents the standard deviation of the dielectric constant of the standard template region for the k-th type of defect in frame t. This represents the Euclidean distance between the spatial centroid coordinates of the m-th candidate region and the spatial center of the standard template for the n-th type of defect. This represents the average distance between all candidate regions and the center of the template; This represents the Hausdorff distance between the m-th region and the boundary of the n-th type of defect template in the t-th frame; This represents the average Hausdorff distance between all candidate regions and the boundaries of various templates; , and Represents the weight parameters for multi-feature fusion; This represents the bias term for the nth type of defect in the tth frame.

8. The tunnel defect identification method based on dielectric distribution map according to claim 5, characterized in that, The spatial adjacency graph specifically uses the spatial centroid or main boundary points of the abnormal region as graph nodes. Edges are used to connect nodes that are spatially adjacent, have similar dielectric properties, or have continuous boundary shapes. The weight of the edge is jointly defined by the spatial distance between nodes, boundary continuity, and electromagnetic property similarity.

9. The tunnel defect identification method based on dielectric distribution map according to claim 1, characterized in that, The structured inspection report includes an overview map of the distribution of various defects, a statistical chart of defect intensity levels classified by category, a confidence level table, a superimposed diagram of defects in key structural parts, and an evolution trend analysis diagram.

Images