Tunnel defect identification method based on dielectric distribution diagram

Abstract

The invention relates to the technical field of tunnel defect identification, in particular to a tunnel defect identification method based on a dielectric distribution diagram. According to the method, on the basis of AI rock-soil perspective radar field monitoring, gprMax simulation and laboratory electromagnetic data, a two-dimensional dielectric distribution diagram is generated through variational Bayesian inversion fusion. Through multi-scale spectral clustering and expert knowledge, a potential abnormal region is automatically identified, and then accurate classification and identification of a defect region are realized by adopting graph form constraint propagation and integrating a Transform-graph neural network model. And finally, projecting an identification result to an original image, and generating a visual defect labeling layer and a structured report. According to the invention, high-precision, automatic, visual and structured detection of tunnel structure defects is realized, the intelligent level and risk early warning capability of tunnel safety operation and maintenance are significantly improved, and the digitization and intelligent process of tunnel operation and maintenance management is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel defect identification, and in particular to a tunnel defect identification method based on a dielectric distribution map. Background Art

[0002] As a vital component of transportation infrastructure, the safety and durability of tunnel structures are directly linked to 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 such as voids, water seepage, cracks, and cavities. Existing technologies for identifying and assessing internal tunnel defects rely on manual inspections or image recognition and radar image analysis. Geological radar (GPR), with its non-contact, rapid detection, and penetrating capabilities, can obtain information on the dielectric distribution within tunnel structures, thereby revealing potential structural defects.

[0003] However, due to the low signal-to-noise ratio, complex images, and irregular textures of dielectric distribution maps, traditional manual interpretation is inefficient and inaccurate, making it unable to meet the needs of automatic identification and classification of structural defects in complex tunnel scenarios. Currently, the following problems remain: Manual analysis and judgment of dielectric images is widely used in current projects. This method relies on expert experience and suffers from 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) appear irregular and have fuzzy boundaries on dielectric distribution maps. Traditional image processing algorithms (such as threshold segmentation and edge detection) are insensitive to these complex morphologies and have low recognition rates. There is a lack of end-to-end automated recognition methods. Existing solutions struggle to quickly identify, classify, and locate defects, especially in real-world scenarios with multiple defect types and high image noise, limiting their intelligent engineering applications. Existing methods rely solely on a single type of inspection data (such as single radar measurements or simulation data) and lack multi-source information fusion. This results in insufficient accuracy in characterizing the complex dielectric distribution within tunnel structures and hinders comprehensive and accurate characterization of defect characteristics. Summary of the Invention

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

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

[0006] A method for identifying tunnel defects based on a dielectric distribution map comprises the following steps:

[0007] S1: Using AI geotechnical radar to obtain field monitoring data inside the tunnel, combined with simulation data built using gprMax and electromagnetic experimental data obtained in the laboratory, the variational Bayesian inversion method is used to perform data fusion and reconstruction 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 regionally scanned and compared, potential abnormal areas within threshold interval conditions are extracted, and an initial defect candidate layer is generated;

[0009] S3: Based on the initial defect candidate layer, a graph morphology constrained propagation algorithm is used to construct a spatial adjacency graph, and the spatial connectivity and boundary morphology of the defect area are jointly modeled. Through edge weight evolution and morphology preservation mechanism, an optimized structural defect mask layer is generated;

[0010] S4: Based on the structural defect mask layer, a multi-class defect recognition model integrating the Transformer structure and the graph neural network algorithm is constructed to classify and identify the defect area and output the defect recognition result; the defect recognition result includes the defect type label, location coordinates and classification confidence score;

[0011] S5: Projecting the defect recognition result onto the original frame image of the two-dimensional dielectric distribution image sequence, generating a visual defect annotation layer including the defect type, spatial location and confidence level, and outputting a structured inspection report.

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

[0013] The simulation data includes electromagnetic wave propagation path data under different tunnel structure sections obtained by simulation on the gprMax platform, including target structure model, medium property configuration, radar wave 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 standard specimens prepared in the laboratory, 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 delay parameters, and measured values ​​of reflectivity and transmittance in each frequency band.

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

[0016] Extracting the dielectric constant pixel by pixel from the two-dimensional dielectric distribution image sequence, and constructing a dielectric constant evolution tensor based on the image frame sequence;

[0017] Performing multi-scale sliding window division on the dielectric constant spatiotemporal tensor to obtain local area blocks at different scales, and calculating statistical characteristic parameters including mean, variance, skewness, kurtosis and local gradient entropy based on the dielectric constant distribution of pixels contained in each area block;

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

[0019] Based on the cluster 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 area 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] According to the threshold range, the cluster area is scanned block by block, and areas with dielectric constant offset, structural texture discontinuity or template matching confidence exceeding the threshold are extracted and marked as potential abnormal areas, and an initial defect candidate layer is output.

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

[0022]

[0023] in, Represents the region block feature vector Belongs to the i-th defect type probability score of ; Indicates the defect label of category i, including void, crack, water and corrosion; A multi-dimensional feature vector representing the current candidate region block; Represents the characteristic mean vector of the i-th type of defect; Indicates defect type The characteristic covariance matrix of Dimensionality; Represents the square of the Mahalanobis distance between the current feature of the area to be identified and the defect template; represents the inverse of the covariance matrix; Represents the transpose operation of a vector.

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

[0025] By extracting spatial features of potential abnormal areas 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 constrained propagation algorithm. Edge morphology preservation constraints, regional connectivity weight control factors, and structural texture priors are introduced into the propagation mechanism to drive the structured propagation of defect information in the graph structure and the adaptive evolution of edge weights.

[0027] During the propagation process, the weight coefficient of each edge is dynamically adjusted to integrate the local boundary curvature change, regional morphological compactness and defect template morphological similarity indicators;

[0028] Based on the graph structure output after propagation stabilization, the main defect area 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 area, including dielectric constant distribution characteristics, regional spatial geometric characteristics, boundary contour morphological parameters, and difference indicators from surrounding normal areas, and the above features are standardized;

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

[0032] In the multi-class defect recognition model, based on the multi-label supervision mechanism, multiple defect types are identified for each input defect area, and the spatial location coordinates and classification confidence score of each type of defect are simultaneously output;

[0033] The labels, coordinates, and confidence scores of the identified defects are mapped to the original dielectric distribution map or tunnel structure diagram, and a visual defect annotation layer with multiple defect types, spatial distribution, and confidence levels is automatically generated to form a structured defect recognition result.

[0034] Furthermore, the formula of the multi-class defect recognition model is as follows:

[0035]

[0036] in, It represents the probability that the mth defect candidate area belongs to the nth type of defect in the dielectric distribution map of the tth frame, that is, the final classification confidence score; Represents the Sigmoid normalization function; represents the normalization factor; K represents the number of adjacent regions of the mth candidate region in the spatial adjacency graph; represents the spatial adjacency weight between the mth region and the kth adjacent region in the tth frame; represents the average dielectric constant of the mth region in the kth adjacent region; represents the average dielectric constant of the standard template area of ​​the k-th type of defect in the t-th frame; represents the standard deviation of the dielectric constant of the k-th type defect standard template area in the t-th frame; Represents the Euclidean distance between the coordinates of the spatial center of gravity of the mth candidate region and the spatial center of the nth type defect standard template; Represents the average distance between all candidate regions and the template center; represents the Hausdorff distance between the mth region and the boundary of the nth type defect template in the tth frame; Represents the average Hausdorff distance between all candidate regions and the boundaries of various templates; 、 and Represents the weight parameter of multi-feature fusion; It represents the bias term of the n-th type defect in the t-th frame.

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

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

[0039] The beneficial effects of the present invention are:

[0040] This invention combines AI geotechnical radar data, gprMax simulation data, and laboratory electromagnetic experimental data, and uses a variational Bayesian inversion method to achieve efficient fusion and structural reconstruction of multi-source data, effectively improving the authenticity and accuracy of dielectric distribution images and providing a more reliable data foundation for subsequent defect detection. A multi-scale spectral clustering algorithm is introduced, integrating expert knowledge with statistical labels to establish a mapping relationship between dielectric constant and defect type, enabling accurate zoning and identification of various potential defects in tunnel structures, effectively improving the detection rate of abnormal areas and the initial screening capability of 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. By integrating the Transformer structure with a graph neural network, deep feature mining and multimodal information fusion are carried out for different types of defect areas, achieving intelligent classification and confidence assessment of multiple types of defects, and outputting structured detection results including type, location, and confidence score, significantly enhancing the comprehensiveness and intelligence of defect identification. The identification results are projected onto the original dielectric map or tunnel structure schematic, 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 and maintenance, reinforcement, and management decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of a tunnel defect identification method based on a dielectric distribution map according to the present invention.

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

[0043] Figure 3 It is a flowchart of step S2 provided in one embodiment of the present invention.

[0044] Figure 4 It is a flowchart of step S4 provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0045] See also Figure 1-4 As shown, the present invention relates to a tunnel defect identification method based on dielectric distribution map.

[0046] Example

[0047] A method for identifying tunnel defects based on a dielectric distribution map comprises the following steps:

[0048] S1: Using AI geotechnical radar to obtain field monitoring data inside the tunnel, combined with simulation data built using gprMax and electromagnetic experimental data obtained in the laboratory, the variational Bayesian inversion method is used to perform data fusion and reconstruction to generate a two-dimensional dielectric distribution image sequence of the tunnel structure;

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

[0050] The simulation data includes electromagnetic wave propagation path data under different tunnel structure sections obtained by simulation on the gprMax platform, including target structure model, medium property configuration, radar wave 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 standard specimens prepared in the laboratory, 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 delay parameters, and measured values ​​of reflectivity and transmittance in each frequency band.

[0052] In one embodiment, field monitoring data collection is as follows;

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

[0054] (2) Radar parameter setting: 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 tunnel inner wall echo signal is collected in different polarization modes in turn 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 measurement 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 the reflection intensity amplitude matrix, time domain waveform, instantaneous energy distribution, main peak and trough points, main reflection delay and other parameters of the echo signal at each measuring point to construct the original data feature matrix.

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

[0059] The hardware details are as follows:

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

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

[0062] Receiving array: Multi-channel reception, real-time acquisition of different polarization echo signals, and support for high-precision A / D conversion (resolution is usually ≥12 bits, sampling rate ≥1GSps).

[0063] Auxiliary positioning: Equipped with an inertial measurement unit (IMU) and a high-precision odometer / laser rangefinder to achieve real-time calibration of the three-dimensional coordinates of the measuring point space.

[0064] Synchronous clock: supports synchronous acquisition and time stamp recording of multi-channel data to ensure spatial and temporal data alignment.

[0065] The software and intelligent algorithm are as follows:

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

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

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

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

[0070] The intelligent data collection and automated operation process includes the following steps: Based on tunnel structure modeling and on-site 3D point cloud information, the radar movement path and collection point distribution are automatically planned to achieve full coverage and high-density scanning of key areas. The system can automatically identify high-risk areas based on historical inspection data and dynamically adjust sampling intervals and measurement point density. It supports automatic switching between different polarization modes (such as vertical, horizontal, and oblique) at the same measurement point, capturing electromagnetic wave responses in different directions and improving sensitivity to multiple types of defects. It has variable frequency transmission capabilities and can automatically select the optimal frequency band for different structural depths and material types, enabling 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 transmit power, receive gain, and sampling rate. It provides intelligent alarms and automatic re-sampling for abnormal or interfering points (such as metal interference and signal obstruction), reducing data loss.

[0071] Combining IMU with odometry / laser scanning enables high-precision reconstruction of radar trajectory, aligning each set of echo data with the actual spatial location of the tunnel structure. It supports real-time integration with BIM / CAD tunnel structural models, automatically mapping data collection points to key structural locations. It can also be integrated with other monitoring systems (such as deformation monitoring and temperature and humidity sensing) to form a multi-dimensional structural health monitoring dataset.

[0072] The AI ​​geotechnical radar not only provides conventional radar echoes but also outputs multi-polarization, frequency-converting, multi-channel, high-dimensional raw data in real time, significantly improving the accuracy of media distribution imaging. On-site intelligent algorithms initially identify abnormal waveforms and remove noise, providing high-quality, structured basic data for subsequent Bayesian inversion and high-order data fusion. This intelligent, automated process simplifies manual intervention and implements an integrated, closed-loop process of "automated data acquisition - intelligent preliminary selection - high-precision calibration - multi-source fusion."

[0073] The gprMax simulation data is constructed as follows:

[0074] (1) Three-dimensional structural modeling: A structural geometric model consistent with the actual tunnel is established in gprMax, including lining thickness, internal and external contours, and the shape and location of typical defects (such as voids, water seepage, inclusions, etc.).

[0075] (2) Dielectric property configuration: Set the dielectric constant and conductivity of concrete, rock mass, soil, defect area, etc. respectively to ensure consistency with the actual material properties.

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

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

[0078] The specific acquisition of laboratory electromagnetic experimental data is as follows:

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

[0080] (2) Experimental testing: On the laboratory electromagnetic test platform, multi-frequency (such as 500MHz, 1GHz, 2GHz) transmission and reflection experiments are carried out on the specimens to collect the dielectric constant, penetration loss, reflection waveform, and amplitude-frequency characteristics at different frequencies.

[0081] (3) Feature summary: Organize and obtain typical signal templates for each defect type (such as extreme points, main time delays, echo amplitude ranges, frequency band responses of reflectivity and penetration, etc.), establish a standard database, and provide comparison samples for subsequent inversion and identification.

[0082] The variational Bayesian inversion data fusion is as follows:

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

[0084] (2) Prior modeling: Using experimental data as the prior distribution and combining it with 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 Bayesian method is used to dynamically adjust the weight of each data source through maximizing the a posteriori probability (MAP) inference 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, achieving data reconstruction with a spatial resolution higher than the original radar sampling.

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

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

[0090] (4) Structural annotation: Automatically associate the actual structural location (such as ring number, mileage, orientation, etc.) for each distribution map, providing an accurate coordinate basis for subsequent defect location and tracing.

[0091] In one embodiment, the Figure 2 The gray image above is a traditional radio radar scan, which is difficult to interpret. The color image below is the dielectric distribution map output by the AI ​​model. Concrete, steel bars, air, and other materials have different dielectric constants, which are reflected in the white, orange, and green colors in the image.

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

[0093] The green block in the center of the color image is the result of an abnormal inversion, representing an area with a significantly lower dielectric constant than the surrounding background. This is commonly seen in structural defects such as voids, cavities, water pockets, or localized loose layers. There is a clear boundary between it and the background, making it easy for intelligent algorithms to identify. The purple border surrounding the green area is the annotated outline of the defect candidate area identification result, indicating the potential structural anomaly area detected by the model. The narrow dark blue strips at the upper and lower edges of the color image are image boundary masks or positioning reference lines. They do not participate in structural judgment, but they help to align image areas and unify coordinates.

[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 regionally scanned and compared, potential abnormal areas within threshold interval conditions are extracted, and an initial defect candidate layer is generated;

[0095] Wherein, the step S2 includes the following steps:

[0096] Extracting the dielectric constant pixel by pixel from the two-dimensional dielectric distribution image sequence, and constructing a dielectric constant evolution tensor based on the image frame sequence;

[0097] Specifically, technicians use automated batch processing to extract the dielectric constant value of each frame at the pixel level from the acquired 2D dielectric distribution image sequence. All frames are sorted by time or acquisition sequence to form a 3D spatiotemporal data volume (evolution tensor), representing the three-dimensional correspondence between pixel position, frame number, and dielectric constant, reflecting the dynamic evolution of the tunnel structure's dielectric properties.

[0098] Performing multi-scale sliding window division on the dielectric constant spatiotemporal tensor to obtain local area blocks at different scales, and calculating statistical characteristic parameters including mean, variance, skewness, kurtosis and local gradient entropy based on the dielectric constant distribution of pixels contained in each area block;

[0099] Specifically, based on the actual size of the tunnel structure and the spatial scale of typical defects, multiple window sizes (such as 8×8, 16×16, and 32×32 pixels) and varying sliding step sizes are selected to ensure coverage of both minute defects and large-scale structural anomalies. The system automatically traverses the entire image, performing complete sliding coverage for each time frame and each window size, ensuring comprehensive detection.

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

[0101] Mean: The average value of the dielectric constants of all pixels in the area, reflecting the overall dielectric properties of the area.

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

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

[0104] Local gradient entropy: By calculating the entropy value of the dielectric constant gradient image in the area, the boundary complexity and texture mutation within the area are measured. It is an important indicator for identifying boundary defects such as cracks and voids.

[0105] All features are stored by triple index of window, frame, and scale to ensure traceability and multi-dimensional analysis.

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

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

[0108] It should be noted that the specific process of multi-scale spectral clustering involves aggregating the feature vectors of all windows and frames according to spatial coordinates and scale indices to form a high-dimensional feature matrix. A feature similarity graph (e.g., based on Euclidean distance or cosine similarity) is constructed between the regional blocks. Each regional block is a node in the graph, and nodes with high similarity are connected by edges. A spectral clustering algorithm is used to perform feature decomposition on the data by calculating the Laplacian matrix of the feature similarity graph, automatically finding the optimal segmentation (e.g., distinguishing between normal background, boundary transition zones, and high-risk anomaly areas). Based on the decomposition results, the system clusters and annotates the original image. Each pixel is assigned a cluster label, and the number of labels is automatically adjusted based on the actual anomaly type and clustering algorithm parameters.

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

[0110] Based on the cluster 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 area 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 testing and laboratory research, the engineering team has established a "feature template library" that includes typical cracks, voids, water seepage, honeycomb surfaces, foreign body inclusions, and other types of defects. Each template contains:

[0112] Statistical feature vectors of typical defect areas (combinations of the above-mentioned mean, variance, skewness, kurtosis, entropy, and other parameters);

[0113] Description of the common dielectric constant value range and distribution morphology when defects occur;

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

[0115] Expert annotated confidence intervals and defect grade assessment criteria for historical inspection samples.

[0116] For each cluster area center feature, the system compares the similarity with the center features of each defect type in the template library in turn (such as calculating Euclidean distance, Mahalanobis distance, feature intersection probability, etc.), and combines expert experience and historical data to set the "dielectric constant threshold interval" and "similarity confidence limit" for each type of defect.

[0117] If the central feature of a 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 determined as a high-suspect region of this type of defect.

[0118] For difficult and complicated cases or incompletely matched areas, the system can automatically report them as "requires manual review" or "low confidence defect areas" for subsequent manual screening and confirmation.

[0119] The system supports experts in reviewing and annotating identification results. Newly confirmed defect areas can be fed back in real time, and the template library can be automatically expanded and threshold and confidence settings adjusted to achieve self-learning and continuous optimization of the model.

[0120] According to the threshold range, the cluster area is scanned block by block, and areas with dielectric constant offset, structural texture discontinuity or template matching confidence exceeding the threshold are extracted and marked as potential abnormal areas, and an initial defect candidate layer is output.

[0121] Specifically, for the cluster label layer, each cluster area 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 area block mean or maximum / minimum dielectric constant exceeds the threshold of the corresponding defect type;

[0123] Obvious discontinuities in local texture (detected by sudden changes in local gradient entropy, kurtosis, and variance);

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

[0125] Areas that meet any of the above abnormality criteria are automatically marked as potential abnormal areas and summarized to generate an "initial defect candidate layer" (a binary or multi-class label matrix).

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

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

[0128]

[0129] in, Represents the region block feature vector Belongs to the i-th defect type probability score of ; Indicates the defect label of category i, including void, crack, water and corrosion; The multidimensional feature vector representing the current candidate region block contains the following five statistical parameters: the mean, variance, skewness, kurtosis, and local gradient entropy of the dielectric constant within the region. These characteristic parameters are derived from the sliding window statistical extraction process based on the two-dimensional dielectric distribution image sequence in step S2. The original data includes the reflection intensity amplitude matrix obtained by the AI ​​geotechnical perspective radar and the dielectric property configuration in the gprMax simulation data. The characteristic 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 in the gprMax platform simulation data, radar wave source response data, and the penetration loss values ​​of laboratory-prepared specimens at different frequencies; Indicates defect type The characteristic covariance matrix of Dimensionality; Represents the square of the Mahalanobis distance between the current feature of the area to be identified and the defect template; represents the inverse of the covariance matrix; Represents the transpose operation of a vector.

[0130] Indicates defect type The characteristic covariance matrix of Represents the covariance between the e-th and l-th feature dimensions. The element calculation formula is as follows:

[0131]

[0132] in, Indicates defect type The covariance between the e-th feature dimension and the l-th feature dimension; represents the total number of training samples belonging to the i-th type of defects; Indicates that the qth item belongs to defect type The value of the sample on the kth statistical feature dimension; Indicates defect type The statistical mean on the e-th feature dimension; and Represents the value of the qth sample on the lth feature dimension and the mean value of this type of defect on this dimension.

[0133] S3: Based on the initial defect candidate layer, a graph morphology constrained propagation algorithm is used to construct a spatial adjacency graph, and the spatial connectivity and boundary morphology of the defect area are jointly modeled. Through edge weight evolution and morphology preservation mechanism, an optimized structural defect mask layer is generated;

[0134] Wherein, the step S3 includes the following steps:

[0135] By extracting spatial features from potential abnormal 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 uses the spatial centroid or main boundary points of the abnormal region as graph nodes, and edges are used to connect nodes with adjacent spatial locations, similar dielectric properties, or continuous boundary morphology. The edge weights are jointly defined by the spatial distance, boundary continuity, and electromagnetic property similarity between nodes.

[0136] Specifically, the system automatically traverses the initial defect candidate layer, labels all connected domains (i.e., abnormal blocks of connected pixels), and assigns a unique ID to each potential anomaly. For each abnormal region, the spatial centroid (i.e., the average of all pixel coordinates) is calculated based on the (x, y) coordinates of all pixels. This determines the region's center position within the entire layer and is used for subsequent spatial relationship modeling. The boundary pixels of each region are automatically extracted to form a boundary point sequence. The perimeter, principal direction (principal axis), maximum / minimum enclosing rectangle, local boundary curvature (e.g., the change in tangent angle every 10 pixels to reflect boundary curvature), and boundary continuity (e.g., a boundary without breakpoints is considered highly continuous) are calculated for each abnormal region. The dielectric constant gradient (i.e., the rate of change of dielectric constant between adjacent pixels) is calculated within and along the boundary of the region. Boundaries with high gradient values ​​are marked to highlight clearly demarcated defects and assist in subsequent spatial relationship analysis.

[0137] Node definition: The center of gravity of each abnormal area, the key point of the main direction of the boundary, or the point with obvious characteristics on the boundary (such as the maximum curvature point) is defined as a node of the adjacency graph. The node attributes include spatial position, area, dielectric constant distribution, morphological type, etc.

[0138] The details of spatial adjacency judgment and edge construction are as follows:

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

[0140] The basis for initializing edge weights is: the closer the spatial distance, the higher the weight; the closer the electromagnetic properties, the higher the weight; the more continuous / smooth the boundary line shape, the higher the weight; conversely, node pairs that are far apart in space, have large attribute differences, and have sudden boundary changes have low initial edge weights.

[0141] Nodes, edges, and weights are recorded in adjacency lists or adjacency matrices to facilitate subsequent efficient algorithm iterative processing and attribute query.

[0142] The spatial adjacency graph is iteratively modeled using a graph morphology constrained propagation algorithm. Edge morphology preservation constraints, regional connectivity weight control factors, and structural texture priors are introduced into the propagation mechanism to drive the structured propagation of defect information in the graph structure and the adaptive evolution of edge weights.

[0143] Specifically, with each node (abnormal region) as the information source, the system simulates the "information flow" propagation process within a spatial adjacency graph. Initially, each node carries its own spatial characteristics, morphological properties, and dielectric constant distribution, transmitting "defect attribution probability" information to adjacent nodes. During propagation, information flow is preferentially transmitted along edges with similar boundary morphology and gently varying boundary curvature, ensuring the continuity and true shape of the overall outline as the defect region expands. For node pairs with sudden changes in boundary curvature or repulsive boundary morphologies, the propagation intensity is weakened or interrupted to avoid misclustering caused by crossing physical / material boundaries.

[0144] Nodes with large, compact connected areas are prioritized for propagation, while nodes with scattered shapes, small areas, or isolated from the main defect area are gradually weakened to gradually filter out isolated noise. If adjacent nodes show similar probability of belonging after the first few rounds of propagation, their connectivity weight is automatically increased to strengthen the formation of the backbone structure.

[0145] During propagation, each node is continuously compared with the standard defect template in terms of texture morphology, spatial distribution, and boundary attributes. If the morphological characteristics of certain nodes are found to be highly consistent with the known template during the propagation process, their priority for being absorbed into the main defect area is significantly increased, driving the entire graph structure to gradually converge to 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 adjustments. Edge weights are dynamically increased or decreased based on the local performance after propagation, resulting in self-enhancement of high-confidence defect areas and self-weakening of low-confidence noise areas.

[0147] When the change in the attribution probability of most nodes falls below a set threshold (e.g., 0.01) or the number of propagation rounds reaches an upper limit (e.g., 10 rounds), the propagation is considered converged. After convergence, the system automatically outputs the connected main regions with attribution probabilities above the threshold as the optimized defect bodies, removing edges and isolated points to generate a final defect mask layer with a clear structure and coherent boundaries.

[0148] During the propagation process, the weight coefficient of each edge is dynamically adjusted to integrate the local boundary curvature change, regional morphological compactness and defect template morphological similarity indicators;

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

[0150] If adjacent nodes are more consistent in morphological, spatial, and electromagnetic characteristics, the edge weight will be further improved, facilitating information aggregation.

[0151] If there is a morphological break, attribute mutation or template mismatch, the edge weight will be automatically reduced to reduce the impact of "pseudo connectivity".

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

[0153] Based on the graph structure output after propagation stabilization, the main defect area 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, a multi-class defect recognition model integrating the Transformer structure and the graph neural network algorithm is constructed to classify and identify the defect area and output the defect recognition result; the defect recognition result includes the defect type label, location coordinates and classification confidence score;

[0155] Wherein, the 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 area, including dielectric constant distribution characteristics, regional spatial geometric characteristics, boundary contour morphological parameters, and difference indicators from surrounding normal areas, and the above features are standardized;

[0157] Specifically, the optimized structural defect mask layer obtained by S3 is used to automatically identify all connected defect areas and automatically assign a unique ID to each area. For each defect area, the system automatically extracts rich features at multiple spatial scales, including:

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

[0159] Spatial geometric features: area, perimeter, principal axis direction, compactness, shape factor, and center of gravity coordinates, which assist in identifying different types of structural anomalies (such as slender cracks and honeycomb blocks).

[0160] Boundary profile parameters: boundary length, curvature change, boundary smoothness, complexity, etc., are used to distinguish the morphological characteristics of different defects.

[0161] Difference index: Compare the difference between the defect area and its adjacent normal area in terms of dielectric constant, boundary morphology and other characteristics to assist the model in learning 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 deviation, 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 area are input into a multi-class defect recognition model that integrates the Transformer structure and the graph neural network (GNN) algorithm. The Transformer module is used to extract the global contextual correlation between the defect areas, while the GNN algorithm is used to capture the spatial connectivity and graph structure dependencies of the defect areas.

[0164] Specifically, for each defect candidate region, the system first encodes its multi-scale features (such as dielectric constant statistics, spatial morphology, boundary parameters, and indicators of difference from the surrounding area) 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 order or topological relationship to form a feature input sequence, which facilitates the Transformer module to process global contextual information between regions. At the same time, 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 regions with similar space and attributes, preparing for subsequent modeling of the GNN algorithm.

[0165] Each feature vector is assigned position information (such as absolute spatial coordinates or relative position to the tunnel axis) to ensure the model can identify spatial distribution patterns. The sequence of all regional feature vectors is input into the encoder layer of the Transformer.

[0166] The Transformer uses a self-attention layer to calculate the feature correlation score of each region relative to all other regions. This can reveal phenomena such as the banded arrangement of multiple cracks and the spatial correlation of multiple defect types. The model can automatically identify large-scale, non-localized structural defect characteristics, effectively addressing the blind spots of single-region, localized features.

[0167] Using each defect region as a node, the system automatically determines whether an edge is connected based on spatial distance, boundary contact, and attribute similarity, and assigns attribute weights to each edge (e.g., high scores for close spatial proximity and small attribute differences). Through graph convolution or message passing, GNNs ensure that, after multiple rounds of computation, each node incorporates not only its own characteristics but also the spatial and attribute information of its neighbors. This helps the model capture complex spatial dependencies such as regional clustering, multi-level structural defects, and anomalous distribution trends. For overlapping, tandem, or clustered defects (e.g., honeycombed areas and interlaced cracks), GNNs effectively integrate their overall morphological features, enhancing classification accuracy.

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

[0169] In the multi-class defect recognition model, based on the multi-label supervision mechanism, multiple defect types are identified for each input defect area, and the spatial location coordinates and classification confidence score of each type of defect are simultaneously output;

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

[0171] During the actual inference phase, the model outputs probability scores for all categories for each region (e.g., 0.95 for "crack," 0.18 for "water seepage," 0.03 for "honeycomb," etc.), allowing for the coexistence of multiple labels. The system determines the defect type of the region based on a preset threshold (e.g., 0.5), automatically identifying high-confidence regions and optionally marking low-confidence regions as "pending review."

[0172] The system automatically outputs the spatial coordinates of each identified defect area, including the area's centroid, a list of boundary points, or a minimum bounding box, ensuring precise location within the original distribution map or structural diagram. Each defect type is assigned a confidence probability score, reflecting the model's confidence in the identification. Classification confidence levels are displayed (e.g., above 0.8 for high confidence, 0.5-0.8 for medium confidence, and below 0.5 for doubtful), which can be used for subsequent maintenance prioritization and review process development.

[0173] All identification results are automatically generated into a structured table, including fields such as region number, type label, probability of each defect type, spatial coordinates, area, and confidence level, allowing engineers to easily export and archive data with a single click. The system directly overlays all classification, spatial, and confidence level results onto the original dielectric distribution map or tunnel structure diagram. Different defect types are displayed with varying colors, shapes, and transparencies, with high-confidence areas highlighted and low-confidence intervals demarcated by dashed lines or light colors, aiding on-site maintenance and decision-making.

[0174] The labels, coordinates, and confidence scores of the identified defects are mapped to the original dielectric distribution map or tunnel structure diagram, and a visual defect annotation layer with multiple defect types, spatial distribution, and confidence levels is automatically generated to form a structured defect recognition result.

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

[0176] Different defect types are visually coded using color, shape, and transparency. High-confidence areas 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 outputs, including: overview maps of multiple defect distributions, heat maps of the spatial distribution of each defect type; and structured table files (such as Excel, CSV, and database formats) showing defect type, coordinates, and confidence levels. These can be used for subsequent inspection report generation, risk assessment, and operation and maintenance decision-making.

[0178] Furthermore, the formula of the multi-class defect recognition model is as follows:

[0179]

[0180] in, It represents the probability that the mth defect candidate area belongs to the nth type of defect in the dielectric distribution map of the tth frame, that is, the final classification confidence score; Represents the Sigmoid normalization function; represents the normalization factor; K represents the number of adjacent regions of the mth candidate region in the spatial adjacency graph; represents the spatial adjacency weight between the mth region and the kth adjacent region in the tth frame; represents the average dielectric constant of the mth region in the kth adjacent region; represents the average dielectric constant of the standard template area of ​​the k-th type of defect in the t-th frame; represents the standard deviation of the dielectric constant of the k-th type defect standard template area in the t-th frame; Represents the Euclidean distance between the coordinates of the spatial center of gravity of the mth candidate region and the spatial center of the nth type defect standard template; Represents the average distance between all candidate regions and the template center; represents the Hausdorff distance between the mth region and the boundary of the nth type defect template in the tth frame; Represents the average Hausdorff distance between all candidate regions and the boundaries of various templates; 、 and Represents the weight parameter of multi-feature fusion; It represents the bias term of the n-th type defect in the t-th frame.

[0181] The calculation formula is as follows:

[0182]

[0183] in, Represents the coordinates of the center of gravity of the mth region in the tth frame; represents the coordinates of the center of gravity of the kth adjacent area in the tth frame; Represents the mean of the centroid distances of all candidate regions; represents the mean dielectric constant of the mth region; represents the mean dielectric constant of the kth adjacent region; Represents the average Hausdorff distance between all candidate regions; represents the Hausdorff distance between the boundary contours of the mth region and the kth adjacent region; 、 and Represents the weight parameter of each feature fusion.

[0184] The calculation formula is as follows:

[0185]

[0186] in, represents the set of boundary points of the mth region in the tth frame; represents the set of boundary points of the kth adjacent region; Represents the Euclidean distance between boundary points x and y, and the coordinates come directly from the pixel space of the dielectric distribution map; Represents the minimum upper bound of a set, that is, the largest minimum value; Represents the maximum lower bound of a set, that is, the smallest maximum value.

[0187] S5: Projecting the defect identification results onto the original frame image of the two-dimensional dielectric distribution image sequence, generating a visual defect annotation layer including the defect type, spatial location and confidence level, and outputting a structured detection report; the structured detection report includes an overview map of multiple defect distributions, a statistical map of defect intensity levels classified by category, a confidence level table, an overlay map of defects in key structural parts, and an evolution trend analysis map.

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

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

[0190] Specifically, the defect type, spatial coordinates, and confidence information output by the S4 are first standardized. Parameters such as the center point, bounding box, and outline of all defect areas are spatially registered with the original 2D dielectric distribution map and the CAD schematic of the tunnel structure. Image registration algorithms (such as affine transformation, feature point matching, and nearest neighbor interpolation) are used to ensure the accurate overlay of defect identification results on the structural schematic. For data from different inspection batches or monitoring cycles, spatiotemporal synchronization and evolutionary alignment of multi-temporal results are also required.

[0191] Based on the registered spatial data, a visual defect annotation layer is automatically generated. This approach enhances readability and usability: Different defect types are marked with distinctive colors and shapes (e.g., cracks are marked as red curves, voids as blue polygons, and water seepage as green areas). Each defect area is overlaid with its type label (text), confidence score (value / level), and spatial coordinates. High-confidence defect areas are highlighted with bold borders, while low-confidence areas are marked with dashed lines or transparency, allowing engineers to focus on high-risk areas.

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

[0193] Provides an interactive interface that allows users to click, zoom, and query detailed attribute data of each defect (such as detection time, historical evolution, adjacent defect relationships, etc.).

[0194] The system automatically compiles a structured inspection report based on all inspection and identification data. Typical output includes:

[0195] Defect Distribution Overview: This diagram displays the spatial distribution of all identified defects across the entire tunnel. Combined with location information such as tunnel number, mileage pile number, and cross-section sequence number, it helps the operation and maintenance team quickly understand the overall health status.

[0196] Classification defect intensity statistics: Use bar charts, pie charts, heat maps, etc. to count the number, distribution density, and intensity level (such as mild / moderate / severe) of defects of different categories, and mark the corresponding spatial range.

[0197] Confidence level table: Sorts all defects by confidence from high to low, and gives the category, location, area, confidence score and whether to recommend review. Each record can be traced back to the original identification layer and collection information.

[0198] Defect overlay map of key structural parts: For special key structures in the tunnel (such as vaults, side walls, floors, joints, secondary linings, etc.), all historical and currently detected defect types and change trends of the part are highlighted in a partitioned manner, facilitating precise positioning and customized repair suggestions.

[0199] Evolution trend analysis chart: When multiple periods of inspection 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 in defect area over time, type conversion rate, and recurrence rate after repair, to assist in judging the structural degradation process and the evolution of hidden dangers.

[0200] In summary, this application fully integrates the field monitoring data of AI geotechnical perspective radar, gprMax simulation data and laboratory electromagnetic experimental data, and realizes the deep fusion and high-precision reconstruction of multi-source data through the variational Bayesian inversion method, which significantly improves the characterization accuracy of the dielectric constant distribution inside the tunnel. By taking advantage of the multi-polarization, frequency conversion and spatial high resolution of AI geotechnical perspective radar, rich original physical signals and multi-dimensional feature data can be obtained, providing a solid data foundation for subsequent defect identification. Compared with traditional single radar scanning technology, this method can not only obtain more comprehensive and detailed electromagnetic feature information, but also overcome monitoring difficulties under actual working conditions such as data noise and spatial occlusion.

[0201] This application achieves dynamic, fine-grained analysis of the dielectric constant of tunnel structures by generating a two-dimensional dielectric distribution image sequence and extracting statistical features at the pixel, frame, and spatiotemporal tensors. The combination of a multi-scale spectral clustering algorithm and an expert knowledge base effectively identifies various potential abnormal areas and achieves precise 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, water seepage, and honeycombs, this method can automatically identify highly suspected defect areas through statistical templates and threshold screening, greatly reducing the subjectivity and risk of missed detection in manual interpretation.

[0202] This application uses a graph-based constrained 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 defect mask layers. A subsequent multi-class defect recognition model integrating the Transformer structure with a graph neural network organically fuses global contextual information with spatial dependencies. This model can not only identify single defects, but also accurately respond to scenarios where multiple defects coexist and complex structural interactions occur, significantly improving the model's generalization 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-category defect visualization annotation layers and structured inspection reports. The system supports multi-batch, multi-temporal data registration and evolution trend analysis, providing visual, quantitative, and intelligent support for tunnel structural 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 scientific 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. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A tunnel defect identification method based on dielectric distribution map, characterized in that: The following steps are involved: S1: Using AI geotechnical radar to obtain field monitoring data inside the tunnel, combined with simulation data built using gprMax and electromagnetic experimental data obtained in the laboratory, the variational Bayesian inversion method is used to perform data fusion and reconstruction 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 regionally scanned and compared, potential abnormal areas within threshold interval conditions are extracted, and an initial defect candidate layer is generated; S3: Based on the initial defect candidate layer, a graph morphology constrained propagation algorithm is used to construct a spatial adjacency graph, and the spatial connectivity and boundary morphology of the defect area are jointly modeled. Through edge weight evolution and morphology preservation mechanism, an optimized structural defect mask layer is generated; S4: Based on the structural defect mask layer, a multi-class defect recognition model integrating the Transformer structure and the graph neural network algorithm is constructed to classify and identify the defect areas and output the defect recognition results; The defect recognition result includes the defect type label, location coordinates and classification confidence score; S5: Projecting the defect recognition result onto the original frame image of the two-dimensional dielectric distribution image sequence, generating a visual defect annotation layer including the defect type, spatial location and confidence level, and outputting a structured inspection report.

2. The method for identifying tunnel defects based on dielectric distribution maps according to claim 1, characterized in that: The field monitoring data includes the multi-polarization echo signal of the tunnel inner wall collected by the AI ​​geotechnical perspective radar, the three-dimensional coordinates of each measuring point, the reflection intensity amplitude matrix, the time domain waveform, the instantaneous energy distribution, the time delay characteristic parameters and the acquisition timestamp information; The simulation data includes electromagnetic wave propagation path data under different tunnel structure sections obtained by simulation on the gprMax platform, including target structure model, medium property configuration, radar wave 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 standard specimens prepared in the laboratory, 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 delay parameters, and measured values ​​of reflectivity and transmittance in each frequency band.

3. The method for identifying tunnel defects based on dielectric distribution maps according to claim 1, characterized in that: The step S2 comprises the following steps: Extracting the dielectric constant pixel by pixel from the two-dimensional dielectric distribution image sequence, and constructing a dielectric constant evolution tensor based on the image frame sequence; Performing multi-scale sliding window division on the dielectric constant spatiotemporal tensor to obtain local area blocks at different scales, and calculating statistical characteristic parameters including mean, variance, skewness, kurtosis and local gradient entropy based on the dielectric constant distribution of pixels contained in each area block; The statistical feature parameters of multi-scale regional blocks are aggregated to construct a feature vector set. A multi-scale spectral clustering algorithm is used to perform joint feature embedding and Laplace feature decomposition to generate a cluster label layer. Based on the cluster 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 area 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; According to the threshold range, the cluster area is scanned block by block, and areas with dielectric constant offset, structural texture discontinuity or template matching confidence exceeding the threshold are extracted and marked as potential abnormal areas, and an initial defect candidate layer is output.

4. The method for identifying tunnel defects based on dielectric distribution maps according to claim 3, characterized in that: The formula of the dielectric constant-defect type mapping model is as follows: in, Represents the region block feature vector Belongs to the i-th defect type probability score of ; Indicates the defect label of category i, including void, crack, water and corrosion; A multi-dimensional feature vector representing the current candidate region block; Represents the characteristic mean vector of the i-th type of defect; Indicates defect type The characteristic covariance matrix of Dimensionality; Represents the square of the Mahalanobis distance between the current feature of the area to be identified and the defect template; represents the inverse of the covariance matrix; Represents the transpose operation of a vector.

5. The method for identifying tunnel defects based on dielectric distribution maps according to claim 1, characterized in that: The step S3 comprises the following steps: By extracting spatial features of potential abnormal areas 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 constrained propagation algorithm. Edge morphology preservation constraints, regional connectivity weight control factors, and structural texture priors are introduced into the propagation mechanism to drive the structured propagation of defect information in the graph structure and the adaptive evolution of edge weights. During the propagation process, the weight coefficient of each edge is dynamically adjusted to integrate the local boundary curvature change, regional morphological compactness and defect template morphological similarity indicators; Based on the graph structure output after propagation stabilization, the main defect area 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 method for identifying tunnel defects based on dielectric distribution maps according to claim 1, characterized in that: The step S4 comprises the following steps: Based on the optimized structural defect mask layer, multi-scale feature extraction is performed on each defect candidate area, including dielectric constant distribution characteristics, regional spatial geometric characteristics, boundary contour morphological parameters, and difference indicators from surrounding normal areas, and the above features are standardized; The multi-scale features of each defect area are input into a multi-class defect recognition model that integrates the Transformer structure and the graph neural network algorithm. The Transformer module is used to extract the global contextual correlation between the defect areas, while the graph neural network algorithm is used to capture the spatial connectivity and graph structure dependencies of the defect areas. In the multi-class defect recognition model, based on the multi-label supervision mechanism, multiple defect types are identified for each input defect area, and the spatial location coordinates and classification confidence score of each type of defect are simultaneously output; The labels, coordinates, and confidence scores of the identified defects are mapped to the original dielectric distribution map or tunnel structure diagram, and a visual defect annotation layer with multiple defect types, spatial distribution, and confidence levels is automatically generated to form a structured defect recognition result.

7. The method for identifying tunnel defects based on dielectric distribution maps according to claim 6, characterized in that: The formula of the multi-class defect recognition model is as follows: in, It represents the probability that the mth defect candidate area belongs to the nth type of defect in the dielectric distribution map of the tth frame, that is, the final classification confidence score; Represents the Sigmoid normalization function; represents the normalization factor; K represents the number of adjacent regions of the mth candidate region in the spatial adjacency graph; represents the spatial adjacency weight between the mth region and the kth adjacent region in the tth frame; represents the average dielectric constant of the mth region in the kth adjacent region; represents the average dielectric constant of the standard template area of ​​the k-th type of defect in the t-th frame; represents the standard deviation of the dielectric constant of the k-th type defect standard template area in the t-th frame; Represents the Euclidean distance between the coordinates of the spatial center of gravity of the mth candidate region and the spatial center of the nth type defect standard template; Represents the average distance between all candidate regions and the template center; represents the Hausdorff distance between the mth region and the boundary of the nth type defect template in the tth frame; Represents the average Hausdorff distance between all candidate regions and the boundaries of various templates; 、 and Represents the weight parameter of multi-feature fusion; It represents the bias term of the n-th type defect in the t-th frame.

8. The method for identifying tunnel defects based on dielectric distribution maps according to claim 5, characterized in that: The spatial adjacency graph specifically uses the spatial centroid or main boundary points of the abnormal area as the graph nodes, and the edges are used to connect nodes with adjacent spatial positions, similar dielectric properties or continuous boundary shapes. The edge weights are jointly defined by the spatial distance, boundary continuity and electromagnetic property similarity between nodes.

9. The method for identifying tunnel defects based on dielectric distribution maps according to claim 1, characterized in that: The structured inspection report includes an overview diagram of the distribution of multiple types of defects, a statistical diagram of defect intensity levels classified by category, a confidence level table, an overlay diagram of defects in key structural parts, and an evolution trend analysis diagram.

Images