A tunnel detection self-calibration method and system based on intelligent targets and geological information

Abstract

This invention discloses a tunnel detection self-calibration method and system based on intelligent targets and geological information. Multiple intelligent targets are deployed within the tunnel, and the surface of each target is divided into three functional regions. The detection vehicle identifies these three regions and decodes the data to obtain the target ID, location mileage, geological information, and reference dimensions, thus generating target decoding data. Local accuracy metrics are used to correct the tunnel scanning detection data, resulting in calibrated data and related information, and a tunnel detection report is output. This method solves the problems of unassessable online model accuracy, poor generalization ability, and high maintenance costs in tunnel AI detection, achieving traceability of detection results, self-evolution of model performance, and reduction of total lifecycle costs.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering inspection technology, and in particular to a tunnel inspection self-calibration method and system based on intelligent targets and geological information. Background Technology

[0002] With the advancement of computer vision and artificial intelligence technologies, image-based automatic detection technology for tunnel surface defects has been widely applied. However, this technology faces three core bottlenecks in engineering implementation: First, the training of AI models relies on massive amounts of high-quality labeled data, and manual labeling is costly and inconsistent. Second, the trained models lack generalization ability under different tunnel, lighting, and lining conditions, and their performance drifts over time and with environmental changes. Third, there is a lack of means to conduct online, quantitative evaluation of the model's measurement accuracy in actual operation, leading to doubts about the reliability and traceability of the detection results. Currently, the targets used in tunnel detection are mostly simple checkerboard or circular targets, whose function is limited to providing reference points for spatial coordinate transformation in photogrammetry or laser scanning. These traditional targets are passive and static, unable to interact with AI detection models, and cannot solve the fundamental problems of model training, evaluation, and iteration, resulting in low efficiency and accuracy in tunnel detection. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a tunnel detection self-calibration method and system based on intelligent targets and geological information.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned tunnel detection self-calibration method based on intelligent targets and geological information: Multiple smart targets are deployed inside the tunnel. The surface of the smart targets is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated disease pattern area, and a machine-readable coding area. The detection vehicle identifies three functional areas of the smart target and decodes them to obtain the target ID, location mileage, geological information, and reference size data, thus obtaining the target decoding data. The target decoding data is geometrically calibrated using the high-precision benchmark geometric area of ​​the intelligent target, and the accuracy of the disease identification AI model is evaluated using the standardized simulated disease pattern area to obtain the accuracy evaluation error. The accuracy assessment error is corrected by weighting the geological information to obtain the corrected accuracy assessment error. Obtain the correction accuracy assessment error of adjacent smart targets and generate a continuous accuracy field distribution function for the entire tunnel. Based on the correction accuracy assessment error and the continuous accuracy field distribution function, a local accuracy characterization quantity is generated; The tunnel scanning and detection data are corrected using local accuracy characterization parameters to obtain calibrated data and related information, and a tunnel detection report is output.

[0005] Furthermore, in the aforementioned tunnel detection self-calibration method based on intelligent targets and geological information, multiple intelligent targets are deployed within the tunnel, and the surface of each intelligent target is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated defect pattern area, and a machine-readable coding area, comprising: The high-precision reference geometry area contains geometric shapes with known absolute dimensions and micron-level manufacturing precision, used for online calibration of the image acquisition equipment of the inspection vehicle, obtaining pixel-to-physical size conversion relationships, and evaluating image quality; The standardized simulated defect pattern area contains a set of patterns of physical dimensions used to simulate actual tunnel defects. It includes at least multiple sets of straight segments to simulate cracks and evaluate the accuracy of the defect identification AI model in measuring crack width; and multiple irregular shapes to simulate lining spalling and evaluate the accuracy of the defect identification AI model in identifying defect morphology and area. The machine-readable encoding area stores codes, including Data Matrix codes, which are used to store the target's unique identification code, the absolute mileage information of the tunnel at the deployment location, the geological and design parameter information of the mileage section, and the actual size data of the reference geometry and simulated defect patterns.

[0006] Furthermore, in the aforementioned tunnel detection self-calibration method based on intelligent targets and geological information, the high-precision reference geometric area of ​​the intelligent target is used to perform geometric calibration on the target decoding data, and the standardized simulated defect pattern area is used to evaluate the accuracy of the defect identification AI model, resulting in an accuracy evaluation error, including: The camera intrinsic and extrinsic parameters of the image are calculated online using a high-precision reference geometric area to complete geometric calibration and obtain the scale transformation factor; By controlling the disease identification AI model to identify and measure the standardized simulated disease pattern area, a set of measurement values ​​output by the model is obtained; The measured set is compared with the set of true values ​​obtained from the coding area to calculate the accuracy assessment error of the current model at the target location, including the crack width measurement error, the disease area measurement error, and the disease morphology recognition accuracy.

[0007] Furthermore, in the aforementioned tunnel detection self-calibration method based on intelligent targets and geological information, the step of weighting and correcting the accuracy assessment error based on geological information to obtain a corrected accuracy assessment error includes: Based on the geological information and accuracy assessment error of the intelligent target, query the geological-accuracy correlation model to obtain the historical error distribution characteristics of the model; If the deviation between the current error and the historical error distribution exceeds a preset threshold, a geological anomaly warning will be triggered, indicating that the geological conditions of the section have changed and there are unrecorded geological risks. Geological information is used to weight and correct the accuracy assessment error, generating a corrected accuracy assessment error.

[0008] Furthermore, in the aforementioned tunnel detection self-calibration method based on intelligent targets and geological information, the step of obtaining the correction accuracy assessment error of adjacent intelligent targets and generating a continuous accuracy field distribution function for the entire tunnel includes: The accuracy assessment error set of adjacent targets before and after the current target is obtained. Based on the target spacing and error distribution, Kriging interpolation is used to generate a continuous accuracy field distribution function for the entire tunnel. If the error change rate between adjacent targets exceeds a preset threshold, temporary virtual targets are added in the section, and virtual accuracy assessment points are generated by interpolation to improve calibration accuracy.

[0009] Furthermore, in the aforementioned tunnel detection self-calibration method based on intelligent targets and geological information, the step of generating a local accuracy characterization quantity based on the correction accuracy assessment error and the continuous accuracy field distribution function includes: The system uses the baseline correction accuracy assessment error and the continuous accuracy field distribution function to generate a local accuracy characterization quantity to describe the reliability of the detection results in the tunnel section adjacent to the smart target. The local accuracy characterization quantity includes a confidence weight curve, which decays non-linearly with the distance from the target position; an error distribution function, which describes the expected error range at different mileage positions; and a geological risk index, which is the detection reliability risk level assessed based on geological information.

[0010] Furthermore, in the aforementioned tunnel detection self-calibration method based on intelligent targets and geological information, the step of correcting the tunnel scanning detection data using local accuracy characterization quantities to obtain calibrated data and related information, and outputting a tunnel detection report, includes: The tunnel inspection report includes basic information such as the location, type, and size of the defects; the confidence weight of each inspection result; maintenance priority recommendations based on the geological risk index; a summary of the model accuracy assessment and source information.

[0011] Furthermore, in a tunnel detection self-calibration system based on intelligent targets and geological information, the system includes the following modules: The intelligent target setting module is used to deploy multiple intelligent targets in the tunnel. The surface of the intelligent target is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated defect pattern area, and a machine-readable code area. The target data decoding module is used to detect the three functional areas of the intelligent target by the detection vehicle, and decode and obtain the target ID, location mileage, geological information and reference size data to obtain the target decoding data; The accuracy error assessment module is used to perform geometric calibration on the target decoding data using the high-precision benchmark geometric area of ​​the smart target, and to assess the accuracy of the disease identification AI model using the standardized simulated disease pattern area, thereby obtaining the accuracy assessment error. The accuracy assessment and correction module is used to perform weighted correction on the accuracy assessment error based on geological information to obtain the corrected accuracy assessment error. The distribution function generation module is used to obtain the correction accuracy evaluation error of adjacent smart targets and generate the continuous accuracy field distribution function for the entire tunnel. The characterization quantity generation module is used to generate local accuracy characterization quantities based on the correction accuracy assessment error and the continuous accuracy field distribution function; The inspection report generation module is used to correct the tunnel scanning inspection data using local accuracy characterization quantities, obtain calibrated data and related information, and output a tunnel inspection report.

[0012] Furthermore, in a tunnel detection self-calibration system based on intelligent targets and geological information, the distribution function generation module includes the following sub-modules: The acquisition submodule is used to acquire the accuracy assessment error set of adjacent targets before and after the current target. Based on the target spacing and error distribution, Kriging interpolation is used to generate a continuous accuracy field distribution function for the entire tunnel. If the error change rate between adjacent targets exceeds a preset threshold, temporary virtual targets are added in the section, and virtual accuracy assessment points are generated by interpolation to improve calibration accuracy.

[0013] Furthermore, in a tunnel detection self-calibration system based on intelligent targets and geological information, the characterization quantity generation module includes the following sub-modules: A generation submodule is used to evaluate the accuracy of the baseline correction and the continuous accuracy field distribution function, and to generate a local accuracy characterization quantity to describe the reliability of the detection results in the tunnel section near the smart target. The local accuracy characterization quantity includes a confidence weight curve, which decays non-linearly with the distance from the target position; an error distribution function, which describes the expected error range at different mileage positions; and a geological risk index, which is the detection reliability risk level assessed based on geological information.

[0014] Its beneficial effects are as follows: 1. It achieves online measurability and traceability of AI model accuracy. By embedding benchmark truths into the field environment, each model test serves as a "standard exam," ensuring the results have metrological credibility. 2. It constructs a data-driven model self-evolution loop. High-quality training data is automatically generated using intelligent targets, driving the model to continuously learn and optimize in real-world environments, fundamentally solving the model generalization and performance drift problems. 3. It enhances the engineering practical value of the detection results. By introducing geological information and local accuracy characterization, the detection report not only provides disease data but also assesses the reliability of that data, providing differentiated decision-making basis for maintenance in different geological sections. 4. It significantly reduces the overall lifecycle maintenance cost by automating model calibration, evaluation, and data annotation processes, reducing reliance on external manual annotation and frequent manual calibration. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a tunnel detection self-calibration method based on intelligent targets and geological information according to the present invention; Figure 2 This is a schematic diagram of the first embodiment of a tunnel detection self-calibration system based on intelligent targets and geological information according to the present invention; Figure 3 This is a schematic diagram of the planar structure of a tunnel detection self-calibration system based on intelligent targets and geological information in an embodiment of the present invention; Figure 4 This is a flowchart of a tunnel detection self-calibration system based on intelligent targets and geological information, as described in an embodiment of the present invention. Figure 5 This is a structural diagram of an AI model for defect identification in a tunnel detection self-calibration system based on intelligent targets and geological information, as described in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, a tunnel detection self-calibration method based on intelligent targets and geological information includes the following steps: Step 101: Deploy multiple smart targets inside the tunnel. Divide the surface of the smart targets into three functional areas, including at least a high-precision reference geometry area, a standardized simulated defect pattern area, and a machine-readable code area. Specifically, in this embodiment, multiple smart targets are arranged at preset intervals inside the tunnel. The smart targets are printed or etched on a high-stability substrate, and the surface is divided into three functional areas: High-precision reference geometry area: Contains micron-level manufacturing precision geometry with known absolute dimensions, used for online calibration of the image acquisition equipment of the inspection vehicle to obtain accurate pixel-physical size conversion relationship; Standardized simulated defect pattern area: contains a set of patterns with precise known physical dimensions to simulate actual tunnel defects, including at least multiple sets of straight segments with known widths (for simulating cracks) and multiple irregular shapes with known areas and shapes (for simulating lining spalling). Machine-readable coding area: stores the target's unique identification code, the absolute mileage information of the tunnel at the deployment location, the geological and design parameter information of that mileage section, and the actual size data of the benchmark geometry and simulated defect patterns; Establish a database that links targets with geological information, and bind and store the geological information of each target (including surrounding rock grade, lining type, groundwater condition, geostress state, and historical deformation data) with the target ID.

[0020] Step 102: The detection vehicle identifies the three functional areas of the smart target and decodes them to obtain the target ID, location mileage, geological information and reference size data, thus obtaining the target decoding data; Specifically, in this embodiment, when the detection vehicle passes over the intelligent target, it acquires the target image through the image acquisition device; identifies and locates three functional areas; decodes from the machine-readable encoding area to obtain the target's ID, location mileage L, geological information G, and all reference size data; and simultaneously reads the complete set of geological parameters corresponding to the target from the associated mapping database.

[0021] Step 103: Use the high-precision benchmark geometric area of ​​the smart target to perform geometric calibration on the target decoding data, and use the standardized simulated disease pattern area to evaluate the accuracy of the disease identification AI model and obtain the accuracy evaluation error. Specifically, in this embodiment, a high-precision reference geometric area is used to calculate the camera intrinsic and extrinsic parameters of the current image online, complete the geometric calibration, and obtain the scale transformation factor S; The AI ​​model for disease identification is controlled to identify and measure the standardized simulated disease pattern area, and the set of measured values ​​output by the model is obtained as M_model. M_model is compared with the set of true values ​​M_truth obtained from the encoding area, and the accuracy evaluation error E of the current model at the target position is calculated, including: crack width measurement error E_width; disease area measurement error E_area; and disease morphology recognition accuracy E_shape.

[0022] Step 104: Based on geological information, perform weighted correction on the accuracy assessment error to obtain the corrected accuracy assessment error; Specifically, in this embodiment, based on the geological information G and accuracy assessment error E of the current target, the geological-accuracy correlation model is queried to obtain the historical error distribution characteristics of the model under the geological conditions; If the deviation between the current error E and the historical error distribution exceeds the preset threshold, a geological anomaly warning will be triggered, indicating that the geological conditions in this section may have changed or there may be unrecorded geological risks. The error E is corrected by weighting the geological information G to generate the geologically corrected accuracy assessment error E_corrected = E × W_g, where W_g is the weighting coefficient calculated based on the geological conditions.

[0023] Step 105: Obtain the correction accuracy assessment error of adjacent smart targets and generate the continuous accuracy field distribution function for the entire tunnel. Specifically, in this embodiment, the accuracy evaluation error set {E_i} of the adjacent targets before and after the current target is obtained, i=1,2,...,n; Based on the target spacing and error distribution, Kriging interpolation or Gaussian process regression methods are used to generate the continuous accuracy field distribution function F(L) for the entire tunnel. For the detection result at any mileage location L, its confidence weight W(L) = f(F(L), G(L)), where G(L) is the geological information at that location; If the error rate between adjacent targets exceeds a preset threshold, a temporary virtual target is added to that section, and virtual accuracy evaluation points are generated through interpolation to improve calibration accuracy.

[0024] Step 106: Based on the correction accuracy assessment error and the continuous accuracy field distribution function, generate a local accuracy characterization quantity; Specifically, in this embodiment, based on the geologically corrected accuracy assessment error E_corrected and the continuous accuracy field distribution function F(L), a local accuracy characterization quantity is generated to describe the reliability of the detection results of the tunnel section adjacent to the target. Local accuracy characterization metrics include: confidence weight curve W(L): which decreases non-linearly with distance from the target location; error distribution function F_err(L): which describes the expected error range at different mileage locations; and geological risk index R_g(L): which is the risk level of detection reliability assessed based on geological information.

[0025] Step 107: Correct the tunnel scanning and detection data using local accuracy characterization parameters to obtain calibrated data and related information, and output the tunnel detection report.

[0026] Specifically, in this embodiment, the detection vehicle continuously scans and detects the current tunnel section to obtain the original defect data set X_raw; By using local precision characterization measures to correct or mark the confidence level of X_raw, a calibrated dataset X_calibrated is generated. The correction formula is as follows: X_calibrated = X_raw × W(L) + ΔX(E_corrected, G) Where ΔX is a systematic bias correction term based on error and geological information; The X_calibrated, the corresponding target ID, mileage L, geological information G, accuracy assessment error E, confidence weight W(L), and geological risk index R_g(L) are packaged together and uploaded to the cloud analysis platform.

[0027] The cloud platform collects calibrated data packets from multiple tunnels and multiple mileages to form a "structured dataset with accuracy labels and geological context"; This dataset can be used to incrementally train or fine-tune the disease identification AI model to generate an optimized new model version. Establish a geological-model parameter mapping library to record the optimal parameter configuration of the model under different geological conditions; When the inspection vehicle is about to enter a new geological section, the cloud platform retrieves the pre-adapted model parameters from the mapping library based on the geological information of the section and sends them to the inspection vehicle to achieve geological pre-adaptation of the model. The new model is deployed to the testing vehicle. In subsequent tests, the new model will undergo the accuracy evaluation of step S3 again at the next intelligent target. If the error E is less than the allowable threshold, the verification is successful. If it is greater than the threshold, the platform will use the target data to quickly and specifically optimize the model until the model reaches stable accuracy in this type of geological environment.

[0028] Time series analysis was performed on the accuracy assessment error of the same target at different times to establish a model performance drift monitoring model; When a systematic upward trend in model error is detected, a model retraining warning is triggered. Based on the drift rate prediction model performance degradation curve, a model maintenance time window suggestion is generated.

[0029] Output a calibrated disease detection report, which includes: basic information such as disease location, type, and size; confidence weight of each detection result; maintenance priority recommendations based on the geological risk index; model accuracy assessment summary and source information.

[0030] Its beneficial effects are as follows: 1. It achieves online measurability and traceability of AI model accuracy. By embedding benchmark truths into the field environment, each model test serves as a "standard exam," ensuring the results have metrological credibility. 2. It constructs a data-driven model self-evolution loop. High-quality training data is automatically generated using intelligent targets, driving the model to continuously learn and optimize in real-world environments, fundamentally solving the model generalization and performance drift problems. 3. It enhances the engineering practical value of the detection results. By introducing geological information and local accuracy characterization, the detection report not only provides disease data but also assesses the reliability of that data, providing differentiated decision-making basis for maintenance in different geological sections. 4. It significantly reduces the overall lifecycle maintenance cost by automating model calibration, evaluation, and data annotation processes, reducing reliance on external manual annotation and frequent manual calibration.

[0031] The above describes an embodiment of the tunnel detection self-calibration method based on intelligent targets and geological information according to the present invention. Please refer to [link / reference]. Figure 2 In a tunnel detection self-calibration system based on intelligent targets and geological information, the fiber optic-based telephone communication system includes the following modules: The intelligent target setting module is used to deploy multiple intelligent targets in the tunnel. The surface of the intelligent target is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated defect pattern area, and a machine-readable code area. The target data decoding module is used to detect the three functional areas of the intelligent target by the detection vehicle, and decode and obtain the target ID, location mileage, geological information and reference size data to obtain the target decoding data; The accuracy error assessment module is used to perform geometric calibration of the target decoding data using the high-precision benchmark geometric area of ​​the intelligent target, and to evaluate the accuracy of the disease identification AI model using the standardized simulated disease pattern area, thereby obtaining the accuracy assessment error. The accuracy assessment and correction module is used to perform weighted correction on the accuracy assessment error based on geological information to obtain the corrected accuracy assessment error. The distribution function generation module is used to obtain the correction accuracy evaluation error of adjacent smart targets and generate the continuous accuracy field distribution function for the entire tunnel. The characterization quantity generation module is used to generate local accuracy characterization quantities based on the correction accuracy assessment error and the continuous accuracy field distribution function; The inspection report generation module is used to correct the tunnel scanning inspection data using local accuracy characterization quantities, obtain calibrated data and related information, and output a tunnel inspection report.

[0032] Please see Figure 3 A schematic diagram of a planar structure in a tunnel detection self-calibration system based on intelligent targets and geological information; 100: Intelligent target; 101: High-precision reference geometry area; 102: Standardized simulated disease pattern area; 103: Machine-recognizable coding area.

[0033] Please see Figure 4 A flowchart of a tunnel detection self-calibration system based on intelligent targets and geological information; Specifically, in this embodiment, the smart target is printed or etched on a high-stability substrate, characterized in that the target surface is divided into three functional regions: High-precision reference geometry area: This area contains geometry with known absolute dimensions and micron-level manufacturing precision, used for online calibration of the image acquisition equipment of the inspection vehicle, obtaining accurate pixel-to-physical size conversion relationships, and evaluating image quality.

[0034] Standardized Simulated Defect Pattern Area: This area contains a set of patterns with precisely known physical dimensions used to simulate actual tunnel defects. It includes at least: ① multiple sets of straight segments with known widths, used to simulate cracks and evaluate the model's accuracy in measuring crack width; ② multiple irregular shapes with known areas and shapes, used to simulate lining spalling and evaluate the model's accuracy in identifying defect morphology and area. The dimensions of these patterns serve as the "benchmark truth" for evaluating model performance.

[0035] Machine-readable coding area: This area stores coding, preferably Data Matrix code or QR code, for storing the target's unique identification code (ID), the absolute mileage information of the tunnel at the deployment location, the geological and design parameter information of the mileage section, and the actual size data of the reference geometry and simulated defect patterns.

[0036] Step S1: Target identification and information decoding.

[0037] When the detection vehicle passes over the intelligent target, it acquires the target image through the image acquisition device; identifies and locates three functional areas; and decodes the machine-readable encoding area to obtain the target's ID, location mileage L, geological information G, and all reference size data.

[0038] Step S2: Online calibration and model accuracy evaluation.

[0039] Using a high-precision benchmark geometric region, the camera intrinsic and extrinsic parameters of the current image are calculated online to complete the geometric calibration. Subsequently, the disease identification AI model is controlled to identify and measure the standardized simulated disease pattern area to obtain the set of measurement values ​​M_model output by the model. M_model is compared with the set of true values ​​M_truth obtained from the encoding region to calculate the accuracy evaluation error E of the current model at the target position.

[0040] Step S3: Generate local precision representation.

[0041] Based on the accuracy assessment error E, a local accuracy characterization quantity is generated to describe the confidence level of the detection results in the tunnel section adjacent to the target. This characterization quantity is not a simple constant offset, but rather a weighting coefficient W or error distribution function F(L) generated according to the model error properties, geological information G, and target spacing. For example, for a target at mileage L, a confidence weight that is inversely proportional to the distance can be assigned to the detection points within a certain range before and after it.

[0042] Step S4: Detection data correction and upload.

[0043] The inspection vehicle continuously scans and inspects the current tunnel section, obtaining the raw defect data set X_raw. X_raw is then corrected or its reliability is marked using local accuracy metrics, generating a calibrated data set X_calibrated. X_calibrated, the corresponding target ID, mileage L, geological information G, and accuracy assessment error E are packaged together and uploaded to a cloud analysis platform.

[0044] Step S5: Iterative evolution and validation of the model.

[0045] The cloud platform collects calibrated data packets from multiple tunnels and mileages, forming a "structured dataset with accuracy labels and geological context." This dataset is used to incrementally train or fine-tune the disease identification AI model, generating an optimized new model version. The new model is then deployed to the inspection vehicle. In subsequent inspections, the new model will undergo another accuracy evaluation (step S2) at the next intelligent target. If the error E is less than the allowable threshold, the verification is successful; if it exceeds the threshold, the platform can use the target data to quickly and specifically optimize the model until it reaches stable accuracy in this type of geological environment.

[0046] Specifically, this embodiment also includes The smart target in this embodiment uses a high-precision ceramic substrate and measures 30cm × 30cm.

[0047] (1) Region 1 (high-precision reference geometry region) is located at the center of the target and consists of a series of chromium wire gratings with a line width of 10 μm and a known spacing.

[0048] (2) Area 2 (standardized simulated disease pattern area) surrounds Area 1 and includes four sets of straight line segments with widths of 0.1mm, 0.2mm, 0.5mm and 1.0mm respectively, and three irregular polygons with areas of 5mm², 20mm² and 50mm² respectively.

[0049] (3) Area 3 (machine-readable coding area) is located in one corner of the target and is a Data Matrix code. The coding information includes: target ID T-2023-005, mileage K105+730, geological information {surrounding rock grade: IV, lining type: reinforced concrete}, and the design dimension truth table of all graphics in Area 1 and Area 2.

[0050] (4) The inspection vehicle (equipped with a 20-megapixel industrial camera and an edge computing unit) conducts inspection tests at a speed of 20 km / h. When it passes target T-2023-005, the system executes the following procedure: ① Capture the image, identify and decode the Data Matrix code to determine that the current location is a Class IV surrounding rock section.

[0051] ② Use the raster pattern in region 1 for online calibration and calculate the distortion coefficient of the current image.

[0052] ③ The current crack identification model (such as DeepCrack) is invoked to identify the 0.5mm wide line segment in region 2. The model outputs a width of 0.48mm. The system calculates the width measurement error as E_width = 0.02mm.

[0053] ④ Based on E_width and geological information (Class IV surrounding rock is easily deformable), the system generates a linearly decaying confidence weight curve for the 50-meter sections before and after the target. All cracks detected within this section are reported with this confidence weight (e.g., a weight of 0.98 at the target and 0.85 at 50 meters).

[0054] ⑤ All detection data, confidence weights, and target information for this section are packaged and uploaded to the cloud.

[0055] ⑥ The cloud platform detected a slight upward trend in the errors of multiple targets from Class IV surrounding rock sections recently. Therefore, the platform automatically selected all calibration data packages tagged "Class IV surrounding rock" and performed an incremental training on the existing model, focusing on optimizing the model's robustness to edge detection under such geological conditions.

[0056] ⑦ After the new model was issued, it was verified at the target location of the next Class IV surrounding rock section. The width measurement error converged to 0.005 mm, and the system returned to a high confidence state.

[0057] Through the above closed loop, the system achieves adaptation to environmental changes and autonomous evolution of the model.

[0058] Please see Figure 4 A structural diagram of an AI model for defect identification in a tunnel detection self-calibration system based on intelligent targets and geological information; Specifically, the AI ​​model used—the "Dual-Path Fusion Network Guided by Physical Benchmarks"—is an end-to-end deep learning model designed specifically for collaboration with the "Intelligent Target System." Its core objective is to achieve accurate tunnel defect detection with online traceability of measurement results. This model adds a "stitching perception and adaptive weighting module" to the "Dual-Path Feature Fusion Network (DFF-Net)." Its core innovation lies in utilizing the absolute spatial coordinates of intelligent targets deployed in the tunnel to solve the inherent and long-neglected engineering challenge of inconsistent defect detection within the physical stitching band of images in multi-camera tunnel detection systems.

[0059] The network adopts an overall architecture of "dual-path parallel input, hierarchical feature fusion, and physical information guidance," and mainly consists of four core modules. 1. Global Disease Sensing Network (Path A) (1) Localization: The backbone feature extractor of the network, responsible for understanding the global context of the tunnel scene.

[0060] (2) Input: The pre-processed and initially stitched panoramic image of the tunnel ring, with a fixed size of [Batch, 3, 1024, 4096] (height x width).

[0061] (3) Structure: An encoder-decoder structure is adopted, and an axial attention mechanism is introduced in the encoder.

[0062] ① Encoder: Using the lightweight MixTransformer as the backbone, it extracts multi-scale features through four levels of downsampling, and effectively models the long-range dependencies of the longitudinal strip structure of the tunnel using its axial attention module, which is beneficial for identifying extended cracks.

[0063] ② Decoder: It adopts progressive upsampling and skips connections with the features of the corresponding level of the encoder to gradually restore spatial details and finally outputs a feature map with the same resolution as the input image.

[0064] (4) Output: High-dimensional disease feature map F_a, with size [Batch, 256, 1024, 4096]. F_a contains semantic information and preliminary geometric information of each pixel belonging to various diseases (cracks, peeling, etc.).

[0065] 2. Local reference calibration network (Path B) (1) Positioning: The network’s “physical ruler” and “precision sensor” provides a traceable benchmark for the system.

[0066] (2) Input: The intelligent target region image is cropped from the panoramic image after preliminary target localization through target detection and precise affine transformation. The size is uniformly [Batch, 3, 256, 256].

[0067] (3) Structure: It consists of two parallel light quantum networks.

[0068] ① Baseline Image Analysis Branch: A small CNN (e.g., 4 convolutional layers) processes a high-precision raster for region 1. The output is a scale factor S (a floating-point number) representing the actual physical length (in millimeters) corresponding to 1 pixel in the current image. This branch is trained by directly regressing the known true dimensions using mean squared error loss.

[0069] ② Simulated Disease Measurement Branch: This branch uses another small CNN with the same structure to process simulated diseases in region 2. Its output is a precision confidence vector C, for example, [c_0.1mm, c_0.2mm, c_0.5mm, c_1.0mm, c_spalling], where each component represents the model's confidence in identifying diseases at the corresponding scale in the current environment (between 0 and 1). This branch is jointly trained on the classification and regression tasks of simulated diseases.

[0070] (4) Output: Calibration vector V_cal = Concatenate(S, C). S provides the geometric scale benchmark, and C provides a self-evaluation of the model performance.

[0071] 3. Adaptive Feature Fusion Module (1) Positioning: core information hub, realizing real-time guidance of perception by physical benchmark.

[0072] (2) Input: F_a and V_cal.

[0073] (3) Structure and workflow: ① Feature modulator generation: V_cal generates two sets of modulation parameters through a multilayer perceptron: channel weight vector Gamma and spatial attention map W. Gamma is used to emphasize or suppress different feature channels of F_a; W is used to spatially focus on the region that may be most affected by the current calibration state (such as regions with complex textures or uneven illumination).

[0074] ② Feature modulation: Channel modulation: F_a' = F_a * Gamma Spatial modulation: F_a'' = F_a' * Sigmoid(W) ③ Output: Modulated feature map F_fused_pre. At this point, the disease features have been optimized based on the current real-time camera calibration status and model confidence.

[0075] 4. Stitching perception and adaptive weighting module (1) Positioning: An enhancement module for multi-camera systems to ensure the continuity of physical space.

[0076] (2) Triggering conditions: Activated only when the system is configured with multiple cameras and the currently processed area is calculated as a physical stitching strip.

[0077] (3) Input: For the same physical point within the stitching strip, F_fused_pre_left, V_cal_left and F_fused_pre_right, V_cal_right from different camera paths.

[0078] (4) Core Algorithm: ① Dynamic weight calculation: Calculate the fusion weight based on the calibration vector of each path.

[0079] Weight_left = (1.0 / (|ΔS_left| + ε)) * Mean(C_left) Weight_right = (1.0 / (|ΔS_right| + ε)) * Mean(C_right) Here, ΔS represents the deviation between the current scaling factor and the standard calibration value. This means that the path with more accurate scaling and higher confidence will receive greater weight.

[0080] ② Adaptive weighted fusion: F_final_pixel =(Weight_left * F_fused_pre_left + Weight_right * F_fused_pre_right) / (Weight_left + Weight_right) This operation is performed at the feature level, allowing gradients to propagate back along the camera path, thereby training the network to actively produce features that are more conducive to consistent fusion.

[0081] (5) Output: The final feature F_final after seamless fusion. For non-stitched regions, F_final is directly equal to F_fused_pre of the current camera.

[0082] End-to-end workflow: 1. Data forward propagation: (1) Input the tunnel image Path A and output the global feature F_a.

[0083] (2) Simultaneously, the smart target image is cropped and input into Path B, and the calibration vector V_cal is output.

[0084] (3) F_a and V_cal are combined in the adaptive feature fusion module to generate the preliminary optimized feature F_fused_pre.

[0085] (4) The system determines whether the spatial point corresponding to F_fused_pre is located in the multi-camera stitching strip. If yes, the stitching perception module is called to perform weighted fusion with the features of adjacent cameras to generate F_final; if no, then F_final = F_fused_pre.

[0086] (5) F_final uses two parallel lightweight prediction heads: ①Segmentation Header (a 1x1 convolution + Softmax): Outputs the disease category for each pixel.

[0087] ② Regression Head (a 1x1 convolution): Outputs the physical dimensions of the disease (such as crack width).

[0088] 2. Online calibration and feedback loop (patented core): (1) During the network inference process, the system will automatically compare the identification results of the “simulated disease” on the target by Path B with the “real value” stored in the coding area.

[0089] (2) The generated error signal will be fed back to the system's monitoring unit in real time to assess the overall confidence level of the detection and mark the sections that may require manual review. At the same time, the error data is uploaded for subsequent model iteration training.

[0090] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A tunnel detection self-calibration method based on intelligent targets and geological information, characterized in that, The tunnel detection self-calibration method based on intelligent targets and geological information includes the following steps: Multiple smart targets are deployed inside the tunnel. The surface of the smart targets is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated disease pattern area, and a machine-readable coding area. The detection vehicle identifies three functional areas of the smart target and decodes them to obtain the target ID, location mileage, geological information, and reference size data, thus obtaining the target decoding data. The target decoding data is geometrically calibrated using the high-precision benchmark geometric area of ​​the intelligent target, and the accuracy of the disease identification AI model is evaluated using the standardized simulated disease pattern area to obtain the accuracy evaluation error. The accuracy assessment error is corrected by weighting the geological information to obtain the corrected accuracy assessment error. Obtain the correction accuracy assessment error of adjacent smart targets and generate a continuous accuracy field distribution function for the entire tunnel. Based on the correction accuracy assessment error and the continuous accuracy field distribution function, a local accuracy characterization quantity is generated; The tunnel scanning and detection data are corrected using local accuracy characterization parameters to obtain calibrated data and related information, and a tunnel detection report is output.

2. The tunnel detection self-calibration method based on intelligent targets and geological information as described in claim 1, characterized in that, The method involves deploying multiple intelligent targets within the tunnel. The surface of each intelligent target is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated defect pattern area, and a machine-readable coding area. The high-precision reference geometry area contains geometric shapes with known absolute dimensions and micron-level manufacturing precision, used for online calibration of the image acquisition equipment of the inspection vehicle, obtaining pixel-to-physical size conversion relationships, and evaluating image quality; The standardized simulated defect pattern area contains a set of patterns of physical dimensions used to simulate actual tunnel defects. It includes at least multiple sets of straight segments to simulate cracks and evaluate the accuracy of the defect identification AI model in measuring crack width; and multiple irregular shapes to simulate lining spalling and evaluate the accuracy of the defect identification AI model in identifying defect morphology and area. The machine-readable encoding area stores codes, including Data Matrix codes, which are used to store the target's unique identification code, the absolute mileage information of the tunnel at the deployment location, the geological and design parameter information of the mileage section, and the actual size data of the reference geometry and simulated defect patterns.

3. The tunnel detection self-calibration method based on intelligent targets and geological information as described in claim 1, characterized in that, The high-precision reference geometric region of the intelligent target is used to perform geometric calibration on the target decoding data, and the standardized simulated disease pattern region is used to evaluate the accuracy of the disease identification AI model, resulting in an accuracy evaluation error, including: The camera intrinsic and extrinsic parameters of the image are calculated online using a high-precision reference geometric area to complete geometric calibration and obtain the scale transformation factor; By controlling the disease identification AI model to identify and measure the standardized simulated disease pattern area, a set of measurement values ​​output by the model is obtained; The measured set is compared with the set of true values ​​obtained from the coding area to calculate the accuracy assessment error of the current model at the target location, including the crack width measurement error, the disease area measurement error, and the disease morphology recognition accuracy.

4. The tunnel detection self-calibration method based on intelligent targets and geological information as described in claim 1, characterized in that, The weighted correction of the accuracy assessment error based on geological information to obtain the corrected accuracy assessment error includes: Based on the geological information and accuracy assessment error of the intelligent target, query the geological-accuracy correlation model to obtain the historical error distribution characteristics of the model; If the deviation between the current error and the historical error distribution exceeds a preset threshold, a geological anomaly warning will be triggered, indicating that the geological conditions of the section have changed and there are unrecorded geological risks. Geological information is used to weight and correct the accuracy assessment error, generating a corrected accuracy assessment error.

5. The tunnel detection self-calibration method based on intelligent targets and geological information as described in claim 1, characterized in that, The step of obtaining the correction accuracy assessment error of adjacent smart targets and generating a continuous accuracy field distribution function for the entire tunnel includes: The accuracy assessment error set of adjacent targets before and after the current target is obtained. Based on the target spacing and error distribution, Kriging interpolation is used to generate a continuous accuracy field distribution function for the entire tunnel. If the error change rate between adjacent targets exceeds a preset threshold, temporary virtual targets are added in the section, and virtual accuracy assessment points are generated by interpolation to improve calibration accuracy.

6. The tunnel detection self-calibration method based on intelligent targets and geological information as described in claim 1, characterized in that, The generation of local accuracy characterization quantities based on the correction accuracy assessment error and the continuous accuracy field distribution function includes: The baseline correction accuracy assessment error and continuous accuracy field distribution function are used to generate a local accuracy characterization quantity to describe the reliability of the detection results in the tunnel section adjacent to the smart target. The local accuracy characterization quantity includes a confidence weight curve, which decays non-linearly with the distance from the target position; an error distribution function, which describes the expected error range at different mileage positions; and a geological risk index, which is the detection reliability risk level assessed based on geological information.

7. The tunnel detection self-calibration method based on intelligent targets and geological information as described in claim 1, characterized in that, The process involves correcting the tunnel scanning and detection data using local precision characterization parameters to obtain calibrated data and related information, and outputting a tunnel detection report, including: The tunnel inspection report includes basic information such as the location, type, and size of the defects; the confidence weight of each inspection result; maintenance priority recommendations based on the geological risk index; a summary of the model accuracy assessment and source information.

8. A tunnel detection self-calibration system based on intelligent targets and geological information, characterized in that, The tunnel detection self-calibration system based on intelligent targets and geological information includes the following modules: The intelligent target setting module is used to deploy multiple intelligent targets in the tunnel. The surface of the intelligent target is divided into three functional areas, including at least a high-precision reference geometry area, a standardized simulated defect pattern area, and a machine-readable code area. The target data decoding module is used to detect the three functional areas of the intelligent target by the detection vehicle, and decode and obtain the target ID, location mileage, geological information and reference size data to obtain the target decoding data; The accuracy error assessment module is used to perform geometric calibration on the target decoding data using the high-precision benchmark geometric area of ​​the smart target, and to assess the accuracy of the disease identification AI model using the standardized simulated disease pattern area, thereby obtaining the accuracy assessment error. The accuracy assessment and correction module is used to perform weighted correction on the accuracy assessment error based on geological information to obtain the corrected accuracy assessment error. The distribution function generation module is used to obtain the correction accuracy evaluation error of adjacent smart targets and generate the continuous accuracy field distribution function for the entire tunnel. The characterization quantity generation module is used to generate local accuracy characterization quantities based on the correction accuracy assessment error and the continuous accuracy field distribution function; The inspection report generation module is used to correct the tunnel scanning inspection data using local accuracy characterization quantities, obtain calibrated data and related information, and output a tunnel inspection report.

9. The tunnel detection self-calibration system based on intelligent targets and geological information as described in claim 8, characterized in that, The distribution function generation module includes the following sub-modules: The acquisition submodule is used to acquire the accuracy assessment error set of adjacent targets before and after the current target. Based on the target spacing and error distribution, Kriging interpolation is used to generate a continuous accuracy field distribution function for the entire tunnel. If the error change rate between adjacent targets exceeds a preset threshold, temporary virtual targets are added in the section, and virtual accuracy assessment points are generated by interpolation to improve calibration accuracy.

10. The tunnel detection self-calibration system based on intelligent targets and geological information as described in claim 8, characterized in that, The characterization quantity generation module includes the following sub-modules: A generation submodule is used to evaluate the accuracy of the baseline correction and the continuous accuracy field distribution function, and to generate a local accuracy characterization quantity to describe the confidence of the detection results in the tunnel section near the smart target. The local accuracy characterization quantity includes a confidence weight curve, which decays non-linearly with the distance from the target position; and an error distribution function, which describes the expected error range at different mileage positions. Geological risk index is a risk level of detection reliability based on geological information assessment.

Images