Methods and devices for detecting blockages in tunnel drainage pipes

Abstract

This application relates to the fields of non-destructive testing and artificial intelligence in tunnel engineering, and provides a method and apparatus for detecting blockages in tunnel drainage pipes. The method includes: acquiring ground-penetrating radar (GPR) data of the area to be inspected; constructing a parameterized physical model library of the tunnel drainage pipe and its surrounding medium, and performing electromagnetic wave forward modeling on the models in the library to generate prior physical radar data; constructing a physically constrained neural network, using the GPR data and prior physical radar data as training samples to train the physically constrained neural network, wherein the loss function used for training includes a physical consistency loss term to constrain the consistency between the network output and the electromagnetic wave propagation law; inputting the GPR data into the trained physically constrained neural network to obtain the detection result of blockage in the tunnel drainage pipe. The tunnel drainage pipe blockage detection method and apparatus provided in this application can achieve physically reliable, real-time, and accurate blockage detection under small sample conditions.

Description

Technical Field

[0001] This application relates to the fields of non-destructive testing and artificial intelligence technology in tunnel engineering, and in particular to a method and device for detecting blockages in tunnel drainage pipes. Background Technology

[0002] The drainage pipes of the tunnel through the Yellow River are constantly filled with water and the space is confined. Existing invasive inspection methods such as Closed-Circuit Television (CCTV) and Pipe Quick View Inspection (QV) require water shut-off and drainage, making them unsuitable for operating tunnels. Sonar inspection is affected by small pipe diameters and water flow interference, making it difficult to guarantee accuracy.

[0003] In recent years, ground-penetrating radar combined with deep learning has been attempted for pipeline inspection, but it has the following shortcomings: First, deep neural networks require a large amount of labeled data for training, while real samples of blockage in tunnel drainage pipes are extremely scarce, making them prone to overfitting and resulting in poor generalization ability under small sample conditions; Second, pure data-driven models only fit statistical correlations, and the output results may violate the basic physical laws of electromagnetic wave propagation (such as negative blockage thickness), resulting in low reliability; Third, the tunnel drainage pipes are surrounded by multiple layers of media such as concrete lining and surrounding rock, resulting in complex electromagnetic wave reflection, and existing methods have not specifically modeled this scenario, leading to high false alarm and false negative rates; Fourth, high-precision physical inversion methods involve extremely large computational loads, which cannot meet the needs of real-time on-site detection.

[0004] Therefore, there is an urgent need for a method that can achieve physical reliability and real-time accuracy in detecting blockages in tunnel drainage pipes under small sample conditions. Summary of the Invention

[0005] This application provides a method and apparatus for detecting blockages in tunnel drainage pipes, which solves the technical defects of the prior art, such as overfitting under small sample conditions due to reliance on a large number of labeled samples, low reliability of output results due to lack of physical law constraints, and difficulty in balancing detection accuracy and on-site real-time performance. It achieves physically reliable, real-time and accurate detection of blockages in tunnel drainage pipes under small sample conditions.

[0006] In a first aspect, this application provides a method for detecting blockage in tunnel drainage pipes, comprising the following steps: Acquire ground-penetrating radar data for the area to be detected; A parametric physical model library for tunnel drainage pipes and their surrounding media is constructed, and electromagnetic wave forward modeling is performed on the models in the library to generate physical prior radar data. A physical constraint neural network is constructed, and ground-penetrating radar data and physical prior radar data are used as training samples to train the physical constraint neural network. The loss function used for training includes a physical consistency loss term to constrain the consistency between the network output and the electromagnetic wave propagation law. Ground-penetrating radar data is input into a trained physical constraint neural network to obtain the detection results of blockage in the tunnel drainage pipe.

[0007] According to the tunnel drainage pipe siltation detection method provided in this application, the physical consistency loss items include: Response consistency loss is used to constrain the consistency between the radar response output by the physically constrained neural network and the radar response obtained from forward modeling. And / or, Feature Consistency Loss, used to constrain the consistency between deep features extracted from ground-penetrating radar data by the physical constraint neural network and deep features extracted from physical prior radar data.

[0008] According to the tunnel drainage pipe siltation detection method provided in this application, the loss function also includes spatial continuity loss and classification loss; Spatial continuity loss is used to constrain the continuity of detection results along the axial direction and / or depth direction of the tunnel drainage pipe; classification loss is used to constrain the correctness of the physical constraint neural network in classifying the blockage state.

[0009] According to the tunnel drainage pipe blockage detection method provided in this application, the physical constraint neural network includes: The first feature extraction branch is used to extract the measured features of the ground penetrating radar data; The second feature extraction branch is used to extract physical features from the physical prior radar data. The feature fusion module is used to fuse measured features with physical features.

[0010] According to the method for detecting blockage in tunnel drainage pipes provided in this application, the detection result includes a two-dimensional blockage probability field. Each element value in the two-dimensional blockage probability field is used to characterize the probability that blockage exists at the corresponding spatial location of the ground penetrating radar data profile.

[0011] According to the method for detecting blockage in tunnel drainage pipes provided in this application, the method further includes: The two-dimensional clogging probability field is corrected by a physical consistency modulation factor, which is determined based on the degree of matching between ground-penetrating radar data and radar response obtained from forward modeling.

[0012] According to the tunnel drainage pipe clogging detection method provided in this application, a parameterized physical model library of the tunnel drainage pipe and its surrounding medium is constructed, including: Establish an underground structural model that includes the lining layer, surrounding rock area, drainage pipe area, and backfill area around the drainage pipe; Within a preset parameter range, the dielectric constant, conductivity, and spatial anomaly distribution of each medium in the underground structure model are randomly disturbed to form a model covering various siltation conditions, including normal drainage, local siltation, continuous siltation, high water content siltation, and multi-regional siltation.

[0013] Secondly, this application also provides a tunnel drainage pipe clogging detection device, comprising the following modules: Ground-penetrating radar data acquisition module, used to acquire ground-penetrating radar data of the area to be detected; The physical model library construction and forward modeling module is used to build a parametric physical model library of tunnel drainage pipes and their surrounding media, and to perform electromagnetic wave forward modeling simulations on the models in the model library to generate physical prior radar data. The neural network construction and training module is used to construct a physically constrained neural network. Ground penetrating radar data and physical prior radar data are used as training samples to train the physically constrained neural network. The loss function used for training includes a physical consistency loss term to constrain the consistency between the network output and the electromagnetic wave propagation law. The blockage detection output module is used to input ground-penetrating radar data into a trained physical constraint neural network to obtain the detection results of blockage in the tunnel drainage pipe.

[0014] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the tunnel drainage pipe blockage detection methods described above.

[0015] Fourthly, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the tunnel drainage pipe blockage detection methods described above.

[0016] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the tunnel drainage pipe blockage detection methods described above.

[0017] This application provides a method and apparatus for detecting blockages in tunnel drainage pipes. It utilizes ground-penetrating radar for non-contact data acquisition, eliminating the need for water outages or contact with the pipe wall, thus solving the problem that invasive methods are unsuitable for operating tunnel drainage pipes. By constructing a parameterized physical model library and generating prior radar data that satisfies the laws of electromagnetic wave propagation through forward modeling, it can significantly expand the effective training data under conditions of scarce real-world blockage samples, effectively alleviating the dependence of deep neural networks on large-scale labeled samples and avoiding overfitting. Furthermore, by introducing a physical consistency loss term into the loss function, the consistency between the network output and the laws of electromagnetic wave propagation is constrained, ensuring the physical reliability of the blockage detection results and avoiding the problem of purely data-driven model outputs violating physical laws. Simultaneously, the parameterized model library fully simulates the multi-layered media structure of the lining and surrounding rock around the tunnel drainage pipe, and combined with the neural network, achieves real-time accurate detection with low false alarms and low false negatives in complex environments. Therefore, this application can achieve physically reliable, real-time, and accurate blockage detection under small sample conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the method for detecting blockages in tunnel drainage pipes provided in this application; Figure 2 A schematic diagram illustrating the construction process of the parametric physical model library provided in this application; Figure 3 This is a schematic diagram of the tunnel drainage pipe structure provided in this application; Figure 4 Schematic diagrams of 11 types of drainage pipe blockages and their degree of blockage provided in this application; Figure 5 A schematic diagram of the simulation results of the 11 types of drainage pipe blockages and their degree of blockage provided in this application; Figure 6 This is a schematic diagram of the drainage pipe siltation test results provided in this application; Figure 7 A schematic diagram of the tunnel drainage pipe blockage detection device provided in this application; Figure 8 A schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The execution entity of the tunnel drainage pipe siltation detection method provided in this application can be a computer device, server, cloud computing platform, or embedded field computing terminal. Specifically, the method can be executed collaboratively by a field detection device equipped with a ground-penetrating radar data acquisition system and a background computing unit. The ground-penetrating radar data acquisition system is responsible for acquiring radar data of the area to be detected, while the background computing unit runs a physical constraint neural network model to process the radar data and output the siltation detection results. In actual engineering deployments, the execution entity can also be an embedded processor integrated into a portable detection vehicle or handheld detection terminal to achieve real-time on-site detection.

[0022] Based on the aforementioned implementing entities, the method described in this application can be widely applied in the following scenarios: This application is primarily used for the detection of siltation in drainage pipelines in water conservancy projects, tunnel engineering, and underground construction projects. It is particularly suitable for health monitoring of drainage systems in large water conveyance tunnels such as the Yellow River Tunnel, and can also be extended to the detection of siltation defects in urban integrated pipe corridors, subway tunnels, highway tunnels, undersea tunnels, and various pressurized or unpressurized buried pipelines. The method described in this application is especially suitable for complex engineering environments where pipelines are constantly filled with water, cannot be drained, have limited space, and lack genuine siltation samples, enabling non-invasive, rapid, and physically reliable quantitative assessment of siltation.

[0023] The following is combined with Figures 1 to 8 This application describes a method and apparatus for detecting blockages in tunnel drainage pipes.

[0024] Figure 1 This is a flowchart illustrating the tunnel drainage pipe clogging detection method provided in this application, as shown below. Figure 1 As shown, the method includes the following: Step 101: Obtain ground-penetrating radar data of the area to be detected.

[0025] The area to be inspected refers to the two-dimensional or three-dimensional spatial range of the tunnel drainage pipe and its surrounding medium.

[0026] Ground-penetrating radar data refers to radar profile images or waveform signals collected by transmitting high-frequency electromagnetic waves into the ground using ground-penetrating radar equipment and receiving the reflected echoes.

[0027] In this application, the ground-penetrating radar data can be B-scan radar profile images continuously collected along the tunnel axis, or A-scan waveform data collected point by point at a preset survey line location.

[0028] In one possible implementation, a vehicle-mounted ground-penetrating radar (GPR) system is used to continuously acquire radar profile data above the drainage pipe along the tunnel's axial direction. The tunnel drainage pipe extends axially, and continuous acquisition allows for the acquisition of media distribution information along the entire length, forming a complete B-scan radar image and providing a spatially continuous data foundation for subsequent blockage identification.

[0029] Another possible implementation is to use a portable GPR host with a shielded antenna to collect data point-by-point at predetermined survey locations. Given the limited detection window and complex field environments, portable devices are easier to deploy. This improves the flexibility of on-site data collection and allows for more intensive measurements in key areas.

[0030] Step 102: Construct a parametric physical model library of the tunnel drainage pipe and its surrounding medium, and perform electromagnetic wave forward modeling on the models in the model library to generate physical prior radar data.

[0031] Among them, the parametric physical model library refers to a set of underground electromagnetic models generated by preset parameter ranges and random parameter perturbations.

[0032] Electromagnetic wave forward modeling refers to the numerical solution of the propagation process of electromagnetic waves in underground media based on Maxwell's equations.

[0033] Among them, physical prior radar data refers to simulated radar response data that satisfies the laws of electromagnetic wave propagation.

[0034] In this application, a parameterized physical model library can be constructed based on the electromagnetic response change law of underground medium caused by drainage pipe blockage. Physical prior data that meets the electromagnetic propagation law can be generated through forward modeling of ground penetrating radar electromagnetic waves. The physical prior data can be used to constrain and train the convolutional neural network, thereby realizing intelligent detection and identification of drainage pipe blockage areas in complex environments.

[0035] In one possible implementation, the construction process of the parametric physical model library is as follows: First, an underground structure model is established based on the tunnel lining structure, surrounding rock conditions, and drainage pipe layout. The underground structure model includes the lining layer, surrounding rock area, drainage pipe area, and backfill area around the drainage pipes. The drainage pipes are installed behind the lining and continuously distributed along the tunnel axis. Structural parameters such as drainage pipe diameter, burial depth, lining thickness, and surrounding rock type are set according to different engineering conditions.

[0036] Since underground medium parameters are difficult to obtain directly in actual engineering projects, this application does not rely on the precise measurement values ​​of real underground dielectric parameters, but establishes a parametric physical model library based on the possible changes in underground medium under different clogging conditions.

[0037] Furthermore, the parametric physical model library randomly generates dielectric parameters, conductivity parameters, and spatial anomaly distributions within a preset range to simulate changes in underground electromagnetic characteristics under different drainage pipe blockage conditions. Under normal drainage conditions, the interior of the drainage pipe is mainly composed of air or low-water-content fluid, and its ground-penetrating radar response exhibits weak reflection and continuous waveform characteristics. When the drainage pipe becomes blocked, due to sediment deposition, increased water content, and localized water accumulation, the electromagnetic properties of the underground medium change, resulting in a significant abnormal radar response. High-water-content blockage areas typically exhibit characteristics such as stronger reflection, waveform tailing, enhanced high-frequency attenuation, and increased two-way travel time.

[0038] Therefore, by setting different siltation lengths, siltation locations, water-bearing anomaly ranges, and anomaly spatial distribution patterns, underground models under various siltation conditions are established. Within a reasonable parameter range, the dielectric parameters of the surrounding rock, the dielectric parameters of the lining, and the electromagnetic parameters of the siltation area are randomly disturbed, thereby forming a parametric physical model library covering different engineering conditions.

[0039] Subsequently, forward modeling of ground-penetrating radar electromagnetic waves was performed based on the established parametric physical model library. The finite-difference time-domain (FDTD) method was used to numerically calculate the propagation process of ground-penetrating radar electromagnetic waves, simulating the propagation, reflection, and attenuation of ground-penetrating radar electromagnetic waves underground, and obtaining radar response data corresponding to different siltation conditions. The radar response data is the physical prior radar data.

[0040] It is understandable that radar response data includes A-scan waveform, B-scan radar image, echo amplitude information, and wave field energy distribution information.

[0041] It should be noted that the propagation of electromagnetic waves from ground-penetrating radar satisfies the following formula: in, For curl operator, For electric field, Permeability, For electrical conductivity, is the dielectric constant.

[0042] The above formula, by numerically solving the propagation process of electric field E in the multi-layer medium of the tunnel, can simulate the radar echo response under different sludge conditions, and provide physical prior radar data for the physical constraint neural network.

[0043] Among them, the physical prior radar data includes radar response data under different operating conditions such as normal drainage, local siltation, continuous siltation, high water content siltation, and multi-area siltation.

[0044] Step 103: Construct a physical constraint neural network. Use ground-penetrating radar data and physical prior radar data as training samples to train the physical constraint neural network. The loss function used for training includes a physical consistency loss term to constrain the consistency between the network output and the electromagnetic wave propagation law.

[0045] Among them, physical constraint neural networks refer to deep learning models that embed physical laws as prior information into the network structure and training process.

[0046] In the embodiments of this application, the physical constraint neural network adopts a convolutional neural network structure. However, those skilled in the art will understand that the physical constraint neural network can also adopt other deep learning architectures capable of feature extraction and parameter learning, such as Transformer, Recurrent Neural Network (RNN), or Multilayer Perceptron (MLP), as long as it can implement the physical constraint training mechanism of this application.

[0047] Among them, the physical consistency loss term refers to the loss function component that measures the deviation between the network output and the propagation law of electromagnetic waves.

[0048] In one possible implementation, the physical consistency loss term includes a response consistency loss. The response consistency loss is used to constrain the consistency between the radar response output by the physically constrained neural network and the radar response obtained from forward modeling. The response consistency loss directly measures the difference between the radar waveform output by the network and the physically modeled waveform, enabling the network to learn realistic electromagnetic reflection characteristics. This effectively suppresses physically unreliable outputs that might arise from purely data-driven models, such as negative sludge thickness or anomalous dielectric constant values.

[0049] In another possible implementation, the physical consistency loss term can also include a feature consistency loss. The feature consistency loss is used to constrain the consistency between deep features extracted from ground-penetrating radar data and deep features extracted from prior physical radar data by the physical constraint neural network. The feature consistency loss aligns the deep feature space, making it more robust to noise and local deformations. It can improve the network's feature generalization ability in complex noisy environments and reduce false detections caused by field interference.

[0050] In one alternative implementation, the loss function used for training also includes a classification loss. The classification loss is used to constrain the correctness of the physical constraint neural network's classification of congestion states. The classification loss is a fundamental loss term in supervised learning, ensuring the network can distinguish between congested and non-congested regions. It enables the network to possess basic anomaly detection capabilities, providing a stable training basis for the physical consistency loss.

[0051] In another alternative implementation, the loss function used for training also includes a spatial continuity loss. This spatial continuity loss constrains the continuity of detection results along the axial and depth directions of the tunnel drainage pipe. Clog anomalies in tunnel drainage pipes are typically continuously distributed, while random noise anomalies are discretely distributed. Therefore, utilizing the spatial continuity prior to suppress discrete artifacts can reduce the false alarm rate in complex environments.

[0052] In this application, prior physical data and field-measured ground-penetrating radar data are jointly input into a physically constrained neural network for training. During training, the physically constrained neural network not only learns the abnormal texture and waveform features in the radar images, but also learns abnormal response patterns that satisfy the laws of underground electromagnetic propagation.

[0053] The physical constraints in this application do not directly constrain the unmeasurable real underground dielectric parameters, but rather use forward modeling data that satisfies the laws of electromagnetic propagation to physically guide the neural network, so that the network output results conform to the actual ground-penetrating radar wave field response characteristics, thereby avoiding the false detection and false negative problems caused by traditional data-driven convolutional neural networks that rely solely on statistical feature learning.

[0054] Specifically, this application constructs a loss function that includes classification loss, response consistency loss, feature consistency loss, and spatial continuity loss. By jointly constraining the consistency between the output results of the network and the electromagnetic propagation law of ground-penetrating radar, the stability and physical consistency of drainage pipe blockage anomaly identification in complex environments are improved.

[0055] Specifically, the loss function satisfies the following formula: in, Classify losses based on siltation; This is due to the loss of consistency in the ground-penetrating radar response; This is the deep feature consistency loss; This results in a loss of spatial continuity. , , These are the corresponding weighting coefficients.

[0056] During network training, measured ground-penetrating radar data and physical prior forward modeling data are jointly input into the physically constrained neural network. The convolutional neural network extracts anomalous response features from radar images through multi-scale convolutional modules. For example, anomalous response features can be hyperbolic reflection features, local energy accumulation features, waveform tail features, or high-frequency attenuation features.

[0057] Because ground-penetrating radar responses differ significantly under different siltation conditions, this application uses physical prior modeling data to physically guide the network, enabling the convolutional neural network to learn abnormal response patterns that conform to the laws of underground electromagnetic propagation. Specifically, normal drainage areas typically exhibit continuous weak reflection characteristics, while high-water-content siltation areas exhibit abnormal response characteristics such as strong reflection, enhanced high-frequency attenuation, increased two-way travel time, and enhanced local shadow areas.

[0058] To improve the network's adaptability to complex noisy environments, this application further constructs a response consistency constraint mechanism. The outlier regions output by the convolutional neural network not only need to satisfy classification accuracy but also need to satisfy consistency with the physical forward modeling response. During network training, the measured radar response is compared with the corresponding physical prior forward modeling response, and the network output is constrained to satisfy the propagation law of the ground-penetrating radar wavefield through response consistency loss.

[0059] Specifically, the response consistency loss function is: in, This is the measured ground-penetrating radar response; The forward modeling simulation is used to determine the radar response. The measured radar response and the forward modeling simulation response include: amplitude distribution, waveform structure, two-way travel time, wave field energy distribution, and time-frequency characteristic distribution.

[0060] By responding to consistency constraints, the network output is made consistent with the radar response pattern that satisfies the laws of electromagnetic propagation, thereby reducing the false detection and false negative problems caused by the over-reliance on statistical features in traditional convolutional neural networks.

[0061] Because measured ground-penetrating radar data contains a large amount of random noise, scattering interference, and structural pseudo-anomalies, using only pixel-level response constraints can easily lead to unstable network training. Therefore, this application further introduces deep feature consistency constraints to constrain the distribution of measured data and physical prior data in the deep feature space. A convolutional neural network encoder extracts deep features from the measured data. Deep features of forward modeling data And constrain the difference between the two through feature consistency loss: in, This represents the deep features of the measured radar data; This involves identifying deep features in the forward modeling data. By constraining the consistency of deep features, the network learns stable physical anomalies under different clogging conditions, thereby improving its feature generalization ability in complex environments.

[0062] Meanwhile, since the tunnel drainage pipes are continuously distributed along the tunnel axis, their siltation anomalies typically exhibit spatial continuity, while random noise anomalies are usually discretely distributed. Therefore, this application further constructs a spatial continuity constraint mechanism to impose spatial structural constraints on the network output results. The spatial continuity loss function is: in, Indicates the position index of the ground-penetrating radar image along the survey line; Indicates the position index of a ground-penetrating radar image along the time axis or depth direction; As the continuity weight of the survey line direction, Weights for continuity in the depth direction; Indicates position The predicted probability value of blockage at the location.

[0063] Based on the aforementioned loss function, this application performs end-to-end training on a physically constrained neural network, enabling joint optimization of network parameters under the combined influence of physical prior constraints and measured data constraints, and applies this optimization to the detection of blockages in tunnel drainage pipes. During training, physical prior forward modeling data and measured ground-penetrating radar data are input into the neural network for feature extraction. The network parameters are then jointly optimized through response consistency constraints, feature consistency constraints, and spatial continuity constraints, ensuring that the network output simultaneously satisfies both data fit and electromagnetic propagation physical consistency requirements.

[0064] Step 104: Input the ground-penetrating radar data into the trained physical constraint neural network to obtain the detection results of the tunnel drainage pipe blockage.

[0065] The detection result refers to the set of information reflecting the blockage status of the drainage pipe output by the network forward inference.

[0066] In one possible implementation, the detection results include a two-dimensional clogging probability field. Each element value in the two-dimensional clogging probability field characterizes the probability that clogging exists at a corresponding spatial location in the ground-penetrating radar data profile. The probability field can quantitatively characterize the spatial distribution of clogging, overcoming the limitation of traditional methods that can only qualitatively determine the presence or absence of anomalies. It provides engineering maintenance personnel with an intuitive clogging probability distribution map, facilitating the location of clogging sections.

[0067] Specifically, assuming the original response output of the last layer of the network is Z, the corresponding probability field P is obtained by the Sigmoid function: Each spatial location For a sampling point in a ground-penetrating radar profile, its probability value is used to characterize the likelihood that the location belongs to a blockage anomaly, thus forming a probability field representation with spatial distribution characteristics.

[0068] In another possible implementation, the detection results also include estimates of the blockage thickness and the dielectric constant of the blockage area. Thickness and dielectric constant are key quantitative indicators for determining the severity and type of blockage.

[0069] The tunnel drainage pipe blockage detection method provided in this application utilizes ground-penetrating radar for non-contact data acquisition, eliminating the need for water outages or contact with the pipe wall, thus solving the problem that invasive methods cannot be applied to operating tunnel drainage pipes. By constructing a parameterized physical model library and generating prior radar data that satisfies the laws of electromagnetic wave propagation through forward modeling, the effective training data can be significantly expanded under conditions of scarce real blockage samples, effectively alleviating the dependence of deep neural networks on large-scale labeled samples and avoiding overfitting. Furthermore, by introducing a physical consistency loss term into the loss function, the consistency between the network output and the laws of electromagnetic wave propagation is constrained, ensuring the physical reliability of the blockage detection results and avoiding the problem of purely data-driven model outputs violating physical laws. At the same time, the parameterized model library fully simulates the multi-layered media structure of the lining and surrounding rock around the tunnel drainage pipe, and combined with the neural network, achieves real-time accurate detection with low false alarms and low false negatives in complex environments. Therefore, it is possible to achieve physically reliable, real-time, and accurate blockage detection under small sample conditions.

[0070] In some embodiments, the physically constrained neural network includes an input module, a multi-scale feature extraction module, a bi-branch feature encoding module, a feature fusion module, and a probability output module.

[0071] The input module is used to receive field-measured ground-penetrating radar data and forward modeling data generated by a parametric physical model library. The data formats include A-scan waveform data and B-scan radar profile images.

[0072] Among them, the multi-scale feature extraction module is used to extract ground-penetrating radar anomaly response features at different spatial scales to characterize the differences in electromagnetic response caused by blockage structures at different scales. For example, the anomaly response features can be local strong reflection features, hyperbolic scattering features, high-frequency attenuation features, or energy accumulation features.

[0073] The dual-branch feature encoding module includes a first feature extraction branch and a second feature extraction branch. The first feature extraction branch extracts the measured features from the ground-penetrating radar data, and the second feature extraction branch extracts the physical features from the prior physical radar data. The feature fusion module fuses the measured features with the physical features.

[0074] The probability output module is used to output the corresponding two-dimensional clogging probability field to characterize the probability distribution of each spatial location in the ground penetrating radar profile belonging to the clogging anomaly.

[0075] Therefore, by using a dual-branch structure to process real observation data and physical simulation data respectively, the fusion layer enables the collaborative learning of data-driven features and physical prior features. This allows the network to retain the expressive power of measured data while incorporating constraints from the laws of electromagnetic wave propagation, thus improving its stability and generalization performance in the multi-layered media environment of tunnels. During training, physical prior forward modeling data and measured ground-penetrating radar data are input into the corresponding branches for feature extraction, and joint representation learning is performed in the fusion layer. Simultaneously, the network parameters are jointly optimized through response consistency constraints, feature consistency constraints, and spatial continuity constraints, ensuring that the network output simultaneously meets the requirements of data fit and electromagnetic propagation physical consistency.

[0076] In this application embodiment, after obtaining the two-dimensional clogging probability field, this application also provides an optional correction mechanism to further improve the physical rationality of the probability field. Specifically, as follows: The two-dimensional clogging probability field is corrected by a physical consistency modulation factor, which is determined based on the degree of matching between ground-penetrating radar data and radar response obtained from forward modeling.

[0077] That is, the probability field can be further modified by introducing a physical consistency modulation factor to make the output result satisfy the law of electromagnetic propagation response magnitude: in, σ represents the final clogging probability value at radar profile position (i,j) after physical consistency correction; σ(·) represents the Sigmoid activation function, used to map the network output to the probability interval of 0~1; This represents the original output response of the physically constrained neural network at position (i,j); Represents the physical consistency modulation factor, which is a normalization function based on the consistency of electromagnetic response; This represents the measured ground-penetrating radar response collected on-site; This represents the theoretical physical radar response obtained through electromagnetic wave forward modeling.

[0078] It should be noted that when the measured radar response and the physical forward modeling response have a low degree of matching, the modulation factor... Decreasing the value reduces the reliability of the probability of blockage in the area, making the final output result more consistent with the physical laws of electromagnetic wave propagation.

[0079] Therefore, by using a physical consistency modulation factor to weight and correct the probability field, the probability of regions with high matching degree is enhanced and the probability of regions with low matching degree is weakened. This ensures that the output results conform to the attenuation and reflection laws of electromagnetic wave propagation in terms of spatial distribution, and significantly improves the physical authenticity and engineering credibility of the probability field.

[0080] In some embodiments, such as Figure 2 As shown, the construction of a parametric physical model library for the tunnel drainage pipe and its surrounding medium includes the following steps: Step 201: Establish an underground structure model that includes the lining layer, surrounding rock area, drainage pipe area, and backfill area around the drainage pipe.

[0081] For example, such as Figure 3 The diagram shows a cross-sectional view of the tunnel drainage pipe. From the inside out, the tunnel consists of a lining layer and a surrounding rock zone. The drainage pipe is located within the backfill area behind the lining and is distributed axially. The tunnel interior is marked with the direction indicated by the larger station number, representing the orientation during inspection. To achieve full coverage detection of the drainage pipe, three parallel radar survey lines (line 1, line 2, and line 3) are laid out above the backfill area at the bottom of the tunnel. Each survey line extends axially along the tunnel and is located directly above the drainage pipe. By moving the ground-penetrating radar equipment along the tunnel axis, B-scan radar profile data above the drainage pipe can be continuously collected, thus obtaining complete media distribution information and providing a data foundation for subsequent drainage pipe blockage identification.

[0082] Step 202: Randomly disturb the dielectric constant, conductivity and spatial anomaly distribution of each medium in the underground structure model within the preset parameter range to form a model covering various siltation conditions, including normal drainage, local siltation, continuous siltation, high water content siltation and multi-regional siltation.

[0083] For example, such as Figure 4 As shown, to accurately reflect the electromagnetic response characteristics of tunnel drainage pipes under different operating conditions, this application constructs a parameterized physical model encompassing 11 different states. This model covers normal drainage, partial blockage, continuous blockage, high water content blockage, and complete blockage, as detailed below: Normal drainage conditions: including Model 1, Model 2, and Model 3. Model 1 corresponds to an empty pipe filled with air, Model 2 corresponds to a full pipe filled with water, and Model 3 corresponds to a pipe with a mixture of water and air. These models are used to extract the ground-penetrating radar physical response characteristics of a baseline without solid blockage.

[0084] Localized clogging conditions: including Model 4, Model 5, Model 6, and Model 7. Model 4 corresponds to localized deformation of an empty pipe, Model 5 to localized deformation of a full-water pipe, Model 6 to clogging with concrete occupying 1 / 2 of the pipe diameter, and Model 7 to clogging with concrete occupying 3 / 4 of the pipe diameter. These models are used to simulate conditions where pipe wall deformation due to pressure or localized deposition of solid matter leads to a reduction in the water flow cross-section.

[0085] Continuous blockage state: including model eight, which corresponds to the state where the blockage concrete almost fills the pipe diameter, leaving only a very small gap at the top, and is used to simulate the severe disease condition of a large amount of solid blockage concrete continuously accumulating.

[0086] Completely blocked state: including model nine, corresponding to 100% of the pipe being blocked by concrete, used to simulate the extreme condition where the cross-section of the drainage pipe is completely blocked by concrete or hardened silt, in order to extract the wave field response law when strong reflection and electromagnetic wave penetration are blocked.

[0087] High water-content siltation state: including Model 10 and Model 11. Model 10 corresponds to siltation of 1 / 2 pipe diameter with the remaining space filled with water, while Model 11 corresponds to siltation of 1 / 2 pipe diameter with a sediment transition layer and the remaining space filled with water. These models are used to simulate the complex phase state of groundwater and siltation coexisting, constraining the neural network to learn the high-frequency attenuation and multiple reflection patterns at the interface of multiple complex media.

[0088] This application comprehensively covers various real-world scenarios of tunnel drainage pipes, from normal operation to extreme damage, through a model library, enriching the diversity of physical prior data and effectively overcoming the overfitting problem of deep learning models caused by the scarcity of real-world siltation samples in underground engineering. Secondly, by meticulously modeling multiphase media coupling conditions—that is, the complex environment where water, silt, and hardened materials coexist—it accurately recreates the multiple reflections and high-frequency attenuation characteristics of electromagnetic waves. This provides the physical constraint neural network with strong physical constraints that highly conform to the laws of electromagnetic wave propagation, significantly enhancing the model's generalization ability and detection accuracy in complex noisy environments.

[0089] After completing the construction of the parametric physical model library, this application performs electromagnetic wave forward modeling simulations on the aforementioned models. For example, as shown... Figure 5 As shown, based on the established parametric physical model library, this application uses the finite-difference time-domain method to numerically calculate the propagation process of electromagnetic waves from ground-penetrating radar, simulating the propagation, reflection, and attenuation processes of electromagnetic waves in different medium structures, thereby obtaining the results... Figure 4 The forward simulation radar response data corresponds one-to-one with various models in the system, and this data is the physical prior radar data.

[0090] from Figure 5 It can be seen that there are significant differences in radar response characteristics under different clogging conditions: Normal drainage corresponds to Models 1 to 3. Radar images show continuous and regularly shaped hyperbolic reflection characteristics. Due to the single or clearly layered medium inside the pipe, the wave field energy distribution is relatively uniform, and the reflected signal at the bottom of the pipe is clearly visible, with no obvious abnormal scattering or waveform distortion.

[0091] Localized blockage corresponds to models four through seven. The hyperbolic reflection pattern in the radar image undergoes localized distortion. Due to pipe wall deformation or the presence of localized blockages, an irregular diffracted wave and scattering interface is formed, leading to localized energy accumulation and disrupting the continuity of the reflected wave.

[0092] Model 8 corresponds to a continuous blockage state. Radar images show continuous and dense bands of strong reflection. Because the blockage almost occupies the entire pipe diameter, the travel time of electromagnetic waves changes significantly when penetrating the blockage, and the non-uniformity within the blockage leads to substantial interlayer scattering.

[0093] The completely blocked state corresponds to Model Nine. Radar images show extremely strong reflected signals at the top interface of the pipe, while the signals at the bottom and lower parts of the pipe exhibit severe energy attenuation and shadowed areas. This is because the dense blockage prevents electromagnetic waves from penetrating downwards, creating a strong electromagnetic shielding effect.

[0094] The high water-content siltation state corresponds to Model 10 and Model 11. The radar images exhibit extremely complex wavefield characteristics, including significant high-frequency signal attenuation, waveform tailing, and superposition of multiple diffraction hyperbolas. The complex water-solid multiphase interface induces strong multiple reflection interference, resulting in a significant increase in the duration of the radar signal on the time axis.

[0095] Through the aforementioned forward modeling, this application uses radar response features containing specific physical meanings as prior physical knowledge. These features include amplitude distribution, waveform structure, two-way travel time, and energy attenuation patterns. In subsequent training of the physically constrained neural network, these prior radar data are used to calculate the response consistency loss, thereby guiding the network to learn and output detection results that conform to the aforementioned physical laws of electromagnetic wave propagation.

[0096] Furthermore, such as Figure 6 As shown, this application inputs the ground-penetrating radar data to be detected into a trained physical constraint neural network, which outputs a two-dimensional clogging probability field. After obtaining the two-dimensional clogging probability field, this application determines the spatial boundary of high-probability anomaly areas by segmenting and extracting the probability field; and combines the classification results corresponding to the classification loss to generate corresponding bounding boxes and status labels on the original ground-penetrating radar profile image, thereby intuitively displaying the final detection results to engineers.

[0097] in, Figure 6 A, Figure 6 B. Figure 6C corresponds to radar echoes under the conditions of no blockage, water in the pipeline, and blockage, respectively, clearly demonstrating the differences in signal characteristics under different conditions.

[0098] from Figure 6 It can be seen that the method of this application exhibits extremely high recognition accuracy and robustness under complex tunnel background noise, specifically as follows: Precise positioning and multi-state differentiation: such as Figure 6 As shown in A and 6B, the bounding boxes extracted based on the two-dimensional probability field can not only accurately define the target of a non-clogging drainage pipe exhibiting typical hyperbolic features, but also keenly capture and locate the pipe segments where anomalies have occurred. In particular, the results in 6B show that, combined with the classification results output by the classification loss, the network can effectively distinguish between the water-filled pipe state exhibiting multiple reflection tails and the solid-phase clogged pipe state exhibiting chaotic scattering features in 6A, achieving fine-grained identification of disease types.

[0099] Robust identification under complex working conditions: such as Figure 6 As shown in C, even when faced with a large-scale continuous blockage, where the wave field exhibits large-area chaotic strong reflections and lower signal obstruction, the bounding box extracted based on the two-dimensional probability field can still accurately delineate the entire blockage pipeline area without any missed detections or misclassification as normal geological interfaces.

[0100] In summary, based on the expansion of forward modeling prior data from the parametric physical model library and the joint constraint of the physical consistency loss term, this application effectively overcomes the inadequacy of pure data-driven models in the complex scattering environment of multi-layered media in tunnels, and realizes non-invasive, automated, and high-precision detection and spatial positioning of drainage pipe blockage status.

[0101] The tunnel drainage pipe blockage detection device provided in this application is described below. The tunnel drainage pipe blockage detection device described below can be referred to in correspondence with the tunnel drainage pipe blockage detection method described above.

[0102] Figure 7 This is a schematic diagram of the tunnel drainage pipe blockage detection device provided in this application. Figure 7 As shown, this application provides a tunnel drainage pipe blockage detection device, which may include: Ground penetrating radar data acquisition module 701 is used to acquire ground penetrating radar data of the area to be detected; The physical model library construction and forward modeling module 702 is used to construct a parametric physical model library of the tunnel drainage pipe and its surrounding medium, and to perform electromagnetic wave forward modeling simulation on the models in the model library to generate physical prior radar data. The neural network construction and training module 703 is used to construct a physical constraint neural network. Ground penetrating radar data and physical prior radar data are used as training samples to train the physical constraint neural network. The loss function used for training includes a physical consistency loss term to constrain the consistency between the network output and the electromagnetic wave propagation law. The blockage detection output module 704 is used to input ground penetrating radar data into a trained physical constraint neural network to obtain the detection results of blockage in the tunnel drainage pipe.

[0103] In some embodiments, the physical consistency loss term includes: Response consistency loss is used to constrain the consistency between the radar response output by the physically constrained neural network and the radar response obtained from forward modeling. And / or, Feature Consistency Loss, used to constrain the consistency between deep features extracted from ground-penetrating radar data by the physical constraint neural network and deep features extracted from physical prior radar data.

[0104] In some other embodiments, the loss function also includes spatial continuity loss and classification loss; Spatial continuity loss is used to constrain the continuity of detection results along the axial direction and / or depth direction of the tunnel drainage pipe; classification loss is used to constrain the correctness of the physical constraint neural network in classifying the blockage state.

[0105] In yet other embodiments, the physically constrained neural network includes: The first feature extraction branch is used to extract the measured features of the ground penetrating radar data; The second feature extraction branch is used to extract physical features from the physical prior radar data. The feature fusion module is used to fuse measured features with physical features.

[0106] In some other embodiments, the detection results include a two-dimensional clogging probability field, where each element value is used to characterize the probability that clogging exists at a corresponding spatial location in the ground-penetrating radar data profile.

[0107] In some other embodiments, the above module further includes: a correction module for correcting the two-dimensional clogging probability field by means of a physical consistency modulation factor, wherein the physical consistency modulation factor is determined based on the degree of matching between ground penetrating radar data and radar response obtained by forward modeling.

[0108] In some other embodiments, the physical model library construction and forward modeling module 702 is specifically used for: Establish an underground structural model that includes the lining layer, surrounding rock area, drainage pipe area, and backfill area around the drainage pipe; Within a preset parameter range, the dielectric constant, conductivity, and spatial anomaly distribution of each medium in the underground structure model are randomly disturbed to form a model covering various siltation conditions, including normal drainage, local siltation, continuous siltation, high water content siltation, and multi-regional siltation.

[0109] Figure 8 A schematic diagram of the structure of the electronic device provided in this application, such as... Figure 8 As shown, the electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 840. The processor 810, communication interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a method for detecting blockages in tunnel drainage pipes. This method includes: acquiring ground-penetrating radar data of the area to be detected; constructing a parameterized physical model library of the tunnel drainage pipe and its surrounding medium, and performing electromagnetic wave forward modeling on the models in the model library to generate prior physical radar data; constructing a physically constrained neural network, using the ground-penetrating radar data and the prior physical radar data as training samples to train the physically constrained neural network, wherein the loss function used for training includes a physical consistency loss term used to constrain the consistency between the network output and the electromagnetic wave propagation law; and inputting the ground-penetrating radar data into the trained physically constrained neural network to obtain the detection result of the tunnel drainage pipe blockage.

[0110] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the tunnel drainage pipe blockage detection method provided by the above methods. The method includes: acquiring ground-penetrating radar data of the area to be detected; constructing a parameterized physical model library of the tunnel drainage pipe and its surrounding medium, and performing electromagnetic wave forward modeling on the models in the model library to generate physical prior radar data; constructing a physical constraint neural network, using the ground-penetrating radar data and the physical prior radar data as training samples to train the physical constraint neural network, wherein the loss function used for training includes a physical consistency loss term used to constrain the consistency between the network output and the electromagnetic wave propagation law; and inputting the ground-penetrating radar data into the trained physical constraint neural network to obtain the detection result of tunnel drainage pipe blockage.

[0112] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is implemented to perform the tunnel drainage pipe blockage detection method provided by the above methods. The method includes: acquiring ground-penetrating radar data of the area to be detected; constructing a parameterized physical model library of the tunnel drainage pipe and its surrounding medium, and performing electromagnetic wave forward modeling on the models in the model library to generate physical prior radar data; constructing a physical constraint neural network, using the ground-penetrating radar data and the physical prior radar data as training samples to train the physical constraint neural network, wherein the loss function used for training includes a physical consistency loss term used to constrain the consistency between the network output and the electromagnetic wave propagation law; and inputting the ground-penetrating radar data into the trained physical constraint neural network to obtain the detection result of tunnel drainage pipe blockage.

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting blockage in tunnel drainage pipes, characterized in that, The method includes: Acquire ground-penetrating radar data for the area to be detected; A parametric physical model library for tunnel drainage pipes and their surrounding media is constructed, and electromagnetic wave forward modeling is performed on the models in the model library to generate physical prior radar data. A physical constraint neural network is constructed, and the ground-penetrating radar data and the physical prior radar data are used as training samples to train the physical constraint neural network; wherein, the loss function used for training includes a physical consistency loss term to constrain the consistency between the network output and the electromagnetic wave propagation law. The ground-penetrating radar data is input into the trained physical constraint neural network to obtain the detection results of tunnel drainage pipe blockage.

2. The method for detecting blockage in tunnel drainage pipes according to claim 1, characterized in that, The physical consistency loss term includes: Response consistency loss is used to constrain the consistency between the radar response output by the physical constraint neural network and the radar response obtained by the forward modeling. And / or, feature consistency loss, used to constrain the consistency between deep features extracted by the physical constraint neural network from the ground-penetrating radar data and deep features extracted from the physical prior radar data.

3. The method for detecting blockage in tunnel drainage pipes according to claim 1, characterized in that, The loss function also includes spatial continuity loss and classification loss; The spatial continuity loss is used to constrain the continuity of the detection results along the axial direction and / or depth direction of the tunnel drainage pipe; The classification loss is used to constrain the correctness of the classification of the clogging state by the physical constraint neural network.

4. The method for detecting blockage in tunnel drainage pipes according to claim 1, characterized in that, The physical constraint neural network includes: The first feature extraction branch is used to extract the measured features of the ground penetrating radar data; The second feature extraction branch is used to extract the physical features of the physical prior radar data; The feature fusion module is used to fuse the measured features with the physical features.

5. The method for detecting blockage in tunnel drainage pipes according to claim 1, characterized in that, The detection results include a two-dimensional clogging probability field, where each element value is used to characterize the probability that clogging exists at the corresponding spatial location of the ground-penetrating radar data profile.

6. The method for detecting blockage in tunnel drainage pipes according to claim 5, characterized in that, The method further includes: The two-dimensional clogging probability field is corrected by a physical consistency modulation factor, which is determined based on the degree of matching between the ground-penetrating radar data and the radar response obtained from the forward modeling.

7. The method for detecting blockage in tunnel drainage pipes according to claim 1, characterized in that, The parametric physical model library for constructing tunnel drainage pipes and their surrounding media includes: Establish an underground structural model that includes the lining layer, surrounding rock area, drainage pipe area, and backfill area around the drainage pipe; Within a preset parameter range, the dielectric constant, conductivity, and spatial anomaly distribution of each medium in the underground structure model are randomly disturbed to form a model covering various siltation conditions, including normal drainage, local siltation, continuous siltation, high water content siltation, and multi-regional siltation.

8. A device for detecting blockage in tunnel drainage pipes, characterized in that, include: Ground-penetrating radar data acquisition module, used to acquire ground-penetrating radar data of the area to be detected; The physical model library construction and forward modeling module is used to construct a parametric physical model library of the tunnel drainage pipe and its surrounding medium, and to perform electromagnetic wave forward modeling simulation on the models in the model library to generate physical prior radar data. The neural network construction and training module is used to construct a physically constrained neural network, using the ground-penetrating radar data and the physical prior radar data as training samples to train the physically constrained neural network; wherein, the loss function used for training includes a physical consistency loss term used to constrain the consistency between the network output and the electromagnetic wave propagation law; The blockage detection output module is used to input the ground penetrating radar data into the trained physical constraint neural network to obtain the detection result of blockage in the tunnel drainage pipe.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the tunnel drainage pipe blockage detection method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the tunnel drainage pipe blockage detection method as described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the tunnel drainage pipe blockage detection method as described in any one of claims 1 to 7.

Images