Underground Pipeline Detection System and Method Based on Deep Learning for Ground Penetrating Radar Detection

Abstract

This invention relates to an underground pipeline detection system and method based on deep learning for ground-penetrating radar (GPR) detection in the field of geological exploration technology. The system includes a GPR data acquisition module, a data preprocessing module, a deep learning inference module, and a positioning output module. By deploying a multi-scale feature fusion U-Net++ network and an attention mechanism, the system can automatically and accurately segment weakly reflective, deeply buried, or nearby pipeline targets from complex GPR images, improving recognition accuracy and reducing misjudgments caused by manual interpretation or traditional edge detection algorithms. A combined loss function of Dice Loss + Focal Loss is used, making the model more sensitive to sparse positive samples and reducing the false negative rate. The system integrates GNSS / IMU location information and radar wave velocity inversion, outputting the three-dimensional spatial x, y, z coordinates of the pipeline. Under known medium conditions, the burial depth error can be controlled within ±5 cm. Hyperbolic fitting combined with the time-depth conversion formula is used for pre-excavation calculations to meet the detection requirements before engineering excavation.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically to an underground pipeline detection system and method based on deep learning for ground-penetrating radar. Background Technology

[0002] Ground-penetrating radar (GPR), also known as ground-penetrating radar, is a non-destructive, high-resolution geophysical exploration technology. It detects shallow underground structures and targets by transmitting high-frequency electromagnetic waves into the ground and receiving their reflected signals.

[0003] Currently, most methods in this field rely on manual judgment based on experience. However, this method has low accuracy, depends on human experience, and is prone to misjudgment or missed detection. It relies on manual visual interpretation of hyperbolic features in GPR B-scan images, which is highly subjective and requires a high level of operator experience. It is difficult to handle weak reflection signals such as plastic pipes and deeply buried pipelines or dense parallel pipelines, resulting in a high false detection rate and poor robustness. In addition, the positive and negative samples are extremely imbalanced during model training, and small targets are easily ignored. Underground pipelines usually account for less than 2% of GPR images, which are typical sparse small targets. If existing deep learning methods only use the standard cross-entropy loss function, the model will be biased towards predicting the background due to the dominance of negative samples, resulting in serious missed detection of real pipelines. There is a lack of customized loss function design for the characteristics of ground-penetrating radar data, making it difficult to optimize the segmentation boundary and recall rate. Even when combined with human experience and existing deep learning methods, insufficient positioning accuracy and lack of three-dimensional spatial judgment still exist on the basis of the above description. The failure to effectively integrate high-precision GNSS / IMU positioning data may lead to large errors in burial depth calculation, making it difficult to meet engineering requirements. To address these issues, the inventors have proposed an underground pipeline detection system and method based on deep learning for ground-penetrating radar detection. Summary of the Invention

[0004] To overcome the shortcomings mentioned above, the invention aims to provide a technical solution that can solve the above problems.

[0005] An underground pipeline detection system based on deep learning for ground-penetrating radar includes: The ground-penetrating radar data acquisition module is used to transmit electromagnetic waves and receive underground reflected signals to generate B-scan image data; The data preprocessing module performs denoising, gain compensation, and time-depth conversion on the B-scan image; The deep learning inference module is equipped with a pre-trained multi-scale feature fusion convolutional neural network model, which is used to automatically identify and segment underground pipeline targets from pre-processed images. The positioning output module calculates the spatial coordinates (x, y, z) of the pipeline based on the identification results and displays them visually. The deep learning inference module adopts a lightweight U-Net++ architecture and integrates an attention mechanism to enhance the extraction capability of weak reflection signals.

[0006] Furthermore, the original radar signal The formula for background removal is as follows: Where N is the number of horizontal sampling points, and exponential gain is used to compensate for attenuation, as shown in the following formula: in The medium attenuation coefficient is determined based on the speed of radar wave propagation in the medium. The formula for converting the time axis to depth is as follows: .

[0007] Furthermore, the deep learning inference module includes an encoder, which consists of four downsampling blocks, each containing two 3×3 convolutional layers and one max pooling layer. The decoder fuses multi-level features through dense skip connections and introduces a channel attention module, as shown in the following formula: Where F is the input feature map, GAP is global average pooling, and δ is ReLU. It is Sigmoid; The output layer uses the Softmax activation function to generate a pixel-level classification probability map.

[0008] Furthermore, the system also includes a model training module for end-to-end training of the neural network using the labeled GPR dataset. The loss function is a combination of weighted Dice Loss and Focal Loss, and its formula is as follows: in For real labels, To predict probabilities, FocalLoss is defined as a balance factor. This is used to alleviate the problem of imbalance between positive and negative samples.

[0009] Furthermore, the ground-penetrating radar data acquisition module includes: Pulse-type ground-penetrating radar main unit; Antenna arrays with interchangeable frequencies, center frequency range of 100 MHz–1 GHz; A high-precision GNSS / IMU positioning unit is used to record the geographic coordinates and attitude angles of each sampling point in real time; The antenna maintains a constant coupling distance with the ground and adapts to terrain undulations through a spring-buffered structure.

[0010] A method for detecting underground pipelines based on deep learning in ground-penetrating radar (GPR) detection is described. This method is used in GPR deep learning-based underground pipeline detection systems and includes the following steps: S1. Control the ground-penetrating radar to move at a constant speed along the survey line, and simultaneously collect electromagnetic reflection signals and location information; S2. Preprocess the raw data to generate standardized depth profile maps; S3. Input the preprocessed image into the deep learning model and output the pipeline mask map; S4. Perform morphological opening operations and connected component analysis on the mask image for post-processing to extract the pipeline center trajectory; S5. Combine radar wave velocity and two-way travel time to calculate pipeline burial depth and integrate GNSS coordinates to generate a three-dimensional pipeline model.

[0011] Furthermore, in step S4, the extracted trajectory points are smoothed using Hough transform or least squares fitting to eliminate interference from isolated noise points.

[0012] Furthermore, in step S5, the radar wave velocity Determined by the following formula: If the type of underground medium is known, the typical dielectric constant can be obtained by referring to a table. Calculation formula: Where c is the speed of light. If the type of underground medium is unknown, the wave velocity is determined on-site using the common center point (CMP) measurement method.

[0013] Furthermore, the positioning output module supports AR augmented reality display: the identified three-dimensional coordinates of the pipeline are superimposed onto the on-site camera image, and the location of the underground pipeline is presented in real time through a mobile terminal.

[0014] Furthermore, the system also includes an adaptive update mechanism. When the system detects a new pipeline reflection pattern such as composite material pipes and the confidence level is lower than the threshold, it automatically triggers a semi-supervised learning process to fine-tune the model parameters using user feedback samples, thereby achieving online data updates.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By deploying a multi-scale feature fusion U-Net++ network and attention mechanism, the system can automatically and accurately segment weakly reflective, deeply buried or nearby pipeline targets from complex GPR images, improving the recognition accuracy (IoU) and reducing misjudgments caused by manual interpretation or traditional edge detection algorithms; 2. Underground pipelines occupy less than 2% of pixels in B-scan images. The traditional cross-entropy loss method is prone to causing the model to ignore small targets. This solution adopts the Dice Loss + Focal Loss combined loss function, which makes the model more sensitive to sparse positive samples and reduces the false negative rate. 3. The system integrates GNSS / IMU location information and radar wave velocity inversion, and can output the three-dimensional spatial x, y, z coordinates of the pipeline. Under known medium conditions, the burial depth error can be controlled within ±5 cm. The system uses hyperbolic fitting combined with the time-depth conversion formula to perform pre-excavation calculations in order to meet the inspection requirements before engineering excavation. 4. By overlaying underground pipeline models onto real camera footage using mobile terminals such as tablets / AR glasses, construction personnel can intuitively grasp the pipeline route and burial depth, avoiding pipeline damage accidents caused by blind excavation and improving field efficiency. 5. An adaptive update mechanism can also be introduced. When encountering new pipelines and the model confidence is low, the system collects user-labeled samples and triggers semi-supervised fine-tuning without retraining the entire model, thus reducing the dependence on traditional labeled data. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to specific embodiments.

[0017] In this embodiment, please refer to the specific implementation of the underground pipeline detection system based on deep learning for ground-penetrating radar, which includes: The ground-penetrating radar data acquisition module is used to transmit electromagnetic waves and receive underground reflected signals to generate B-scan image data; The data preprocessing module performs denoising, gain compensation, and time-depth conversion on the B-scan image; The deep learning inference module is equipped with a pre-trained multi-scale feature fusion convolutional neural network model, which is used to automatically identify and segment underground pipeline targets from pre-processed images. This module employs an improved multi-scale attention U-Net++ network, specifically designed for the typical hyperbolic reflectance features in ground-penetrating radar (GPR) B-scan images, including: Multi-scale feature extraction encoder, the first The input to the layer decoding block is: Where ⊕ represents channel splicing, this formula is used to improve the adaptability to pipelines with different burial depths and diameters. To enhance the strong and weak signal response of deeply buried plastic pipes, an attention module is inserted after the jump connection. This module is a channel-space dual attention mechanism. The channel attention formula is as follows: The formula for spatial attention is as follows: in convolution, For channel stitching, this function aims to suppress background clutter such as soil inhomogeneity and steel mesh, and enhance the features of the hyperbola's vertices and wings. The output and loss function utilize a network output as a binary segmentation map. The combined loss function is used to address the problem of extreme imbalance between positive and negative samples, and its formula is as follows: in: ,set up =2. =0.75, complete the formula calculation; The equipment operates as follows: Step 1: Data Collection The operator pushes a trolley equipped with a 1 GHz shielded ground-penetrating radar antenna along the survey line at a constant speed. The radar host emits electromagnetic pulses at a density of 100–500 channels per meter, with a center frequency of 100 MHz–2.6 GHz. During this process, the reflected signals, GNSS / IMU position and attitude, and wheel encoder distance are recorded simultaneously. Step 2: Real-time preprocessing of edge computing units DC offset removal: Subtract the mean value of each signal; Background removal: Sliding window average elimination of horizontal stripes; Gain compensation: Applying exponential gain Compensation for attenuation; Time-depth conversion: If the dielectric constant of the medium is known Then depth Output a normalized B-scan image of 256×512 pixels; Step 3: Deep learning inference deployed on embedded AI chips The preprocessed image is fed into the trained MSA-U-Net++ model. The model outputs the pipeline probability mask within <100 ms. This step requires post-processing. The binarized image with a threshold of 0.5 is binarized and morphological opening is used to remove isolated noise points. Connectivity analysis is performed to extract the contour of each pipeline target. Step 4: Pipeline parameter inversion and location A hyperbolic model is fitted to each connected component, and the formula is as follows: in For the vertex time, The horizontal position For wave speed The burial depth is determined as By combining GNSS coordinates, a 3D pipeline point cloud is generated: ; Step 5: Visualization and Output The system overlays and displays ground trajectory lines, pipeline depth profiles, and 3D pipeline models that support color-coded material types on a tablet or AR glasses. It supports exporting Shapefile, DXF, or KML formats for use by GIS systems. The following analysis table is obtained through the above 5 steps. The positioning output module calculates the spatial coordinates (x, y, z) of the pipeline based on the identification results and displays them visually. The deep learning inference module adopts a lightweight U-Net++ architecture and integrates an attention mechanism to enhance the extraction capability of weak reflection signals.

[0018] The original radar signal The formula for background removal is as follows: Where N is the number of horizontal sampling points, and exponential gain is used to compensate for attenuation, as shown in the following formula: in The medium attenuation coefficient is determined based on the speed of radar wave propagation in the medium. The formula for converting the time axis to depth is as follows: .

[0019] The deep learning inference module includes an encoder, which consists of 4 levels of downsampling blocks, each containing two 3×3 convolutional layers and one max pooling layer. The decoder fuses multi-level features through dense skip connections and introduces a channel attention module, as shown in the following formula: Where F is the input feature map, GAP is global average pooling, and δ is ReLU. It is Sigmoid; The output layer uses the Softmax activation function to generate a pixel-level classification probability map.

[0020] The system also includes a model training module for end-to-end training of the neural network using the labeled GPR dataset. The loss function is a combination of weighted Dice Loss and Focal Loss, and its formula is as follows: in For real labels, To predict probabilities, FocalLoss is defined as a balance factor. This is used to alleviate the problem of imbalance between positive and negative samples.

[0021] The ground-penetrating radar data acquisition module includes: Pulse-type ground-penetrating radar main unit; Antenna arrays with interchangeable frequencies, center frequency range of 100 MHz–1 GHz; A high-precision GNSS / IMU positioning unit is used to record the geographic coordinates and attitude angles of each sampling point in real time; The antenna maintains a constant coupling distance with the ground and adapts to terrain undulations through a spring-buffered structure.

[0022] A method for detecting underground pipelines based on deep learning in ground-penetrating radar (GPR) detection is described. This method is used in GPR deep learning-based underground pipeline detection systems and includes the following steps: S1. Control the ground-penetrating radar to move at a constant speed along the survey line, and simultaneously collect electromagnetic reflection signals and location information; S2. Preprocess the raw data to generate standardized depth profile maps; S3. Input the preprocessed image into the deep learning model and output the pipeline mask map; S4. Perform morphological opening operations and connected component analysis on the mask image for post-processing to extract the pipeline center trajectory; S5. Combine radar wave velocity and two-way travel time to calculate pipeline burial depth and integrate GNSS coordinates to generate a three-dimensional pipeline model.

[0023] In step S4, the extracted trajectory points are smoothed using Hough transform or least squares fitting to eliminate interference from isolated noise points.

[0024] In step S5, the radar wave velocity Determined by the following formula: If the type of underground medium is known, the typical dielectric constant can be obtained by referring to a table. Calculation formula: Where c is the speed of light. If the type of underground medium is unknown, the wave velocity is determined on-site using the common center point (CMP) measurement method.

[0025] The positioning output module supports AR augmented reality display: it overlays the identified pipeline's three-dimensional coordinates onto the on-site camera image, and presents the location of the underground pipeline in real time through a mobile terminal.

[0026] The system also includes an adaptive update mechanism. When the system detects a new pipeline reflection pattern such as composite material pipes and the confidence level is lower than the threshold, it automatically triggers a semi-supervised learning process to fine-tune the model parameters using user feedback samples and achieve online data updates.

[0027] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An underground pipeline detection system based on deep learning for ground-penetrating radar, characterized in that, include: The ground-penetrating radar data acquisition module is used to transmit electromagnetic waves and receive underground reflected signals to generate B-scan image data; The data preprocessing module performs denoising, gain compensation, and time-depth conversion on the B-scan image; The deep learning inference module is equipped with a pre-trained multi-scale feature fusion convolutional neural network model, which is used to automatically identify and segment underground pipeline targets from pre-processed images. The positioning output module calculates the spatial coordinates (x, y, z) of the pipeline based on the identification results and displays them visually. The deep learning inference module adopts a lightweight U-Net++ architecture and integrates an attention mechanism to enhance the extraction capability of weak reflection signals.

2. The underground pipeline detection system based on deep learning for ground-penetrating radar as described in claim 1, characterized in that, The original radar signal The formula for background removal is as follows: Where N is the number of horizontal sampling points, and exponential gain is used to compensate for attenuation, as shown in the following formula: in The medium attenuation coefficient is determined based on the speed of radar wave propagation in the medium. The formula for converting the time axis to depth is as follows: 。 3. The underground pipeline detection system based on deep learning for ground-penetrating radar as described in claim 1, characterized in that, The deep learning inference module includes an encoder, which consists of 4 levels of downsampling blocks, each containing two 3×3 convolutional layers and one max pooling layer. The decoder fuses multi-level features through dense skip connections and introduces a channel attention module, as shown in the following formula: Where F is the input feature map, GAP is global average pooling, and δ is ReLU. It is Sigmoid; The output layer uses the Softmax activation function to generate a pixel-level classification probability map.

4. The underground pipeline detection system based on deep learning for ground-penetrating radar as described in claim 1, characterized in that, The system also includes a model training module for end-to-end training of the neural network using the labeled GPR dataset. The loss function is a combination of weighted Dice Loss and Focal Loss, and its formula is as follows: in For real labels, To predict probabilities, FocalLoss is defined as a balance factor. This is used to alleviate the problem of imbalance between positive and negative samples.

5. The underground pipeline detection system based on deep learning for ground-penetrating radar as described in claim 1, characterized in that, The ground-penetrating radar data acquisition module includes: Pulse-type ground-penetrating radar main unit; Antenna arrays with interchangeable frequencies, center frequency range of 100 MHz–1 GHz; A high-precision GNSS / IMU positioning unit is used to record the geographic coordinates and attitude angles of each sampling point in real time; The antenna maintains a constant coupling distance with the ground and adapts to terrain undulations through a spring-buffered structure.

6. A method for detecting underground pipelines based on deep learning in ground-penetrating radar, characterized in that: This detection method is used in the underground pipeline detection system based on deep learning for ground-penetrating radar as described in any one of claims 1-5, and the method includes the following steps: S1. Control the ground-penetrating radar to move at a constant speed along the survey line, and simultaneously collect electromagnetic reflection signals and location information; S2. Preprocess the raw data to generate standardized depth profile maps; S3. Input the preprocessed image into the deep learning model and output the pipeline mask map; S4. Perform morphological opening operations and connected component analysis on the mask image for post-processing to extract the pipeline center trajectory; S5. Combine radar wave velocity and two-way travel time to calculate pipeline burial depth and integrate GNSS coordinates to generate a three-dimensional pipeline model.

7. The underground pipeline detection method based on deep learning for ground-penetrating radar according to claim 6, characterized in that, In step S4, the extracted trajectory points are smoothed using Hough transform or least squares fitting to eliminate interference from isolated noise points.

8. The underground pipeline detection method based on deep learning for ground-penetrating radar according to claim 7, characterized in that, In step S5, the radar wave velocity Determined by the following formula: If the type of underground medium is known, the typical dielectric constant can be obtained by referring to a table. Calculation formula: Where c is the speed of light. If the type of underground medium is unknown, the wave velocity is determined on-site using the common center point (CMP) measurement method.

9. The underground pipeline detection system based on deep learning for ground-penetrating radar according to claim 1, characterized in that, The positioning output module supports AR augmented reality display: it overlays the identified pipeline's three-dimensional coordinates onto the on-site camera image, and presents the location of the underground pipeline in real time through a mobile terminal.

10. The underground pipeline detection system based on deep learning for ground-penetrating radar according to claim 1, characterized in that, The system also includes an adaptive update mechanism.

Images