Underground pipeline detection system and method based on ground penetrating radar and deep learning

Abstract

The invention belongs to the field of ground penetrating radars, and provides an underground pipeline detection system and method based on a ground penetrating radar and deep learning, and the method comprises the steps: obtaining a ground penetrating radar echo signal; the echo signals are preprocessed; inputting the preprocessed radar image into a deep neural network model, and executing recognition processing of the underground pipeline; in the model training stage, a loss function of structural continuity constraint is introduced, and the loss function comprises classification cross entropy loss used for pipeline identification, mean square error loss used for burial depth estimation and a structural smoothness constraint term used for constraining response point space continuity; and performing visual display on an identification result, wherein the display comprises trend tracking of the target pipeline, burial depth marking and visual mapping of identification confidence.

Description

Technical Field

[0002] This invention belongs to the field of ground penetrating radar, and specifically relates to an underground pipeline detection system and method based on ground penetrating radar and deep learning. Background Technology

[0004] With the continuous development of urban infrastructure, the types of underground pipelines are increasing, covering multiple systems such as water supply, drainage, electricity, communications, gas, and heating. As the "invisible lifeline" of urban operation, the location, direction, and depth information of underground pipelines are of great significance for planning, construction, operation and maintenance management, and disaster emergency response. To avoid construction accidents and ensure operational safety, high-precision and intelligent underground pipeline detection has become a key aspect of urban management.

[0005] Currently, commonly used underground pipeline detection technologies include electromagnetic induction, audio detection, lidar, and ground-penetrating radar (GPR). Among these, GPR has advantages such as being non-contact, non-destructive, and applicable to a wide range of materials, and has been widely used for identifying underground targets. However, the interpretation of traditional GPR data mainly relies on human experience, is greatly affected by the operator's subjective judgment, has low identification efficiency, and accurately extracting target information in complex environments such as noise interference, coupled signals, and overlapping adjacent pipelines remains a significant challenge.

[0006] In recent years, deep learning technology has demonstrated powerful performance in fields such as image recognition and speech recognition, and has gradually been introduced into the field of underground detection. However, existing methods often directly input radar images into general neural network models without fully considering the physical prior characteristics of underground pipelines in terms of spatial structure, such as directional consistency, positional coherence, and continuity. This results in insufficient robustness of the model to complex interference, weak ability to judge spatial coherence, and limited target positioning accuracy. Summary of the Invention

[0008] To address the problems in the prior art, this invention provides an underground pipeline detection system based on ground-penetrating radar and deep learning, comprising:

[0009] The signal processing module is used to acquire ground-penetrating radar echo signals and preprocess the echo signals.

[0010] A deep learning recognition module is used to identify underground pipelines based on the preprocessed data. The deep learning recognition module includes:

[0011] Multi-scale convolutional feature extraction unit is used to extract spatial texture features and reflection features from radar images at different scales;

[0012] The orientation-aware convolutional structure unit is used to perform orientation weighting on the convolution kernel based on the movement direction of the detection device and the linear arrangement characteristics of the target reflection signal in the radar image;

[0013] Spatial coherence modeling unit, used to model the connectivity of target response points based on their spatial positional relationships;

[0014] The output layer is used to output the identification results, including the probability of target presence and the estimated pipeline burial depth.

[0015] The model training and constraint module is used to introduce a structural continuity constraint loss function during the model training phase. The loss function includes: classification cross-entropy loss for pipeline identification, mean square error loss for burial depth regression, and structural smoothness constraint term for spatial response points.

[0016] The results visualization module is used to display the target pipeline's route tracking, burial depth marking, and confidence level indicator.

[0017] Furthermore, the multi-scale convolutional feature extraction unit includes multiple parallel convolutional paths, each using convolutional kernels of different sizes to perform convolution operations on the radar image. The convolutional kernel sizes include 3×3, 5×5, and 7×7, and the feature maps extracted from multiple scales are fused through channel stitching.

[0018] Furthermore, the orientation-aware convolutional structure unit includes an orientation encoding module and an orientation-weighted convolutional kernel generation module. The orientation encoding module is used to obtain the movement orientation vector of the detection device and assign orientation weighting coefficients to each sampling position of the convolutional kernel according to the orientation vector to enhance the response characteristics along the direction of the target pipeline.

[0019] Furthermore, the spatial coherence modeling unit is a graph convolutional network structure. The graph convolutional network uses high-confidence response points in the radar image as graph nodes, constructs adjacent edges based on the spatial positional relationship between nodes, and performs graph convolution operations to identify connected structures and suppress isolated noise responses.

[0020] Furthermore, the structural smoothness constraint term employs a minimization constraint on the spatial second-order difference of the prediction result to guide the model to output recognition results with spatial continuity and smoothness. The structural smoothness constraint term is defined as follows:

[0021]

[0022] in, For structural smoothness constraints, This is the response output of the model at position i,j.

[0023] Furthermore, the result visualization module further includes a confidence mapping unit, which is used to perform color gradient encoding display on the pipeline area according to the target existence probability output by the deep learning recognition module, and reflect the recognition confidence level on the radar image in the form of color depth or transparency changes.

[0024] This invention also provides a method for detecting underground pipelines based on ground-penetrating radar and deep learning, comprising the following steps:

[0025] Acquire ground-penetrating radar echo signals;

[0026] The echo signal is preprocessed, including background removal, noise reduction, enhancement, normalization, or time window adjustment.

[0027] The preprocessed radar image is input into a deep neural network model to perform underground pipeline identification processing. The deep neural network model includes:

[0028] a) Extracting spatial texture and reflection features from radar images at different scales;

[0029] b) Based on the linear arrangement characteristics of the moving direction of the detection equipment and the target reflection signal, the convolution kernel is directionally weighted;

[0030] c) Model the connectivity of target response points based on their spatial relationships;

[0031] d) Output the identification results of the target existence probability and the estimated pipeline burial depth;

[0032] During the model training phase, a loss function with structural continuity constraints is introduced, which includes: classification cross-entropy loss for pipeline identification, mean square error loss for burial depth estimation, and structural smoothness constraint term for constraining the spatial continuity of response points.

[0033] The identification results are visualized, including the target pipeline's route tracking, burial depth marking, and a visual mapping of the identification confidence level.

[0034] Furthermore, extracting spatial texture features and reflection features from radar images at different scales includes setting up multiple parallel convolutional paths. Each convolutional path uses convolutional kernels of different sizes to perform convolution operations on the radar image. The convolutional kernel sizes include 3×3, 5×5, and 7×7. The feature maps extracted from multiple scales are then fused through channel stitching.

[0035] Furthermore, based on the linear arrangement characteristics of the moving direction of the detection device and the target reflection signal, the convolution kernel is directionally weighted, which includes using a directional encoding module and a directionally weighted convolution kernel generation module. The directional encoding module is used to obtain the moving direction vector of the detection device and assign directional weighting coefficients to each sampling position of the convolution kernel according to the directional vector, so as to enhance the response characteristics along the direction of the target pipeline.

[0036] Furthermore, modeling the connectivity of target response points based on their spatial positional relationships involves using high-confidence response points in radar images as graph nodes, constructing adjacent edges based on the spatial positional relationships between nodes, and performing graph convolution operations to identify connected structures and suppress isolated noise response loss.

[0037] This invention provides an underground pipeline detection system and method based on ground-penetrating radar and deep learning. It integrates radar non-destructive testing technology with a structure-aware neural network structure, effectively improving the automatic identification capability, interference suppression capability, and burial depth estimation accuracy of underground pipelines. It has the following beneficial effects:

[0038] By constructing a multi-scale convolutional feature extraction structure, target features with different burial depths and signal intensities can be effectively extracted, improving the ability to simultaneously identify shallow and deep buried pipelines and enhancing the model's ability to perceive multi-scale reflection features.

[0039] By introducing a direction-aware convolutional structure, the convolutional kernels are weighted according to the direction of device movement, highlighting the linearly arranged real target reflection structure, suppressing noise interference from random directions, and improving the model's ability to respond to pipeline alignment characteristics.

[0040] By modeling the connectivity between response points using spatial coherence modeling units (based on graph convolutional networks or attention mechanisms), the model can identify real pipeline structures with spatial extensibility and eliminate isolated interference points.

[0041] In model training, a structural continuity constraint loss function is introduced, including classification cross-entropy, regression mean square error, and structural smoothness regularization term, to jointly optimize the classification accuracy and spatial structure consistency of the model, thereby improving the stability and generalization ability of the system in complex scenarios.

[0042] In summary, this invention constructs a complete intelligent detection process for underground pipelines, from radar signal preprocessing, depth feature extraction, structural prior fusion, spatial modeling, burial depth estimation to visualization. It overcomes the problems of traditional methods, such as heavy reliance on human experience, low recognition accuracy, and weak understanding of spatial structures. It has high engineering application value and promising prospects for promotion. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a system block diagram of the present invention;

[0046] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0048] The invention will now be described in preferred form with reference to the accompanying drawings and specific embodiments.

[0049] This embodiment solves the above problems through the following steps:

[0050] In one embodiment, reference Figure 1 This invention provides an underground pipeline detection system based on ground-penetrating radar and deep learning. The system utilizes ground-penetrating radar to acquire electromagnetic echo signals from within the underground medium, and processes and analyzes these echo signals based on a constructed deep neural network model to achieve automatic identification and precise location of underground pipelines. The system integrates the principle of non-destructive radar detection with artificial intelligence feature extraction mechanisms, possessing strong environmental adaptability and interference suppression capabilities.

[0051] In this invention, ground-penetrating radar, also known as geological radar, refers to a non-destructive testing device that uses high-frequency electromagnetic waves to transmit pulse signals into underground media and receives the reflected echoes at different media interfaces. This device analyzes the time and intensity variations of the received electromagnetic wave signals to determine the existence, location, and shape of underground structures or anomalies. In the field of underground pipeline detection, ground-penetrating radar is mainly used to acquire electromagnetic reflection images generated by potentially existing metallic or non-metallic pipelines in underground media, thereby enabling the detection and imaging of target pipelines.

[0052] The system specifically includes:

[0053] The signal processing module is used to acquire ground-penetrating radar echo signals and preprocess the echo signals.

[0054] Specifically, the system acquires the echo signal emitted by a ground-penetrating radar antenna after penetrating the underground medium, and performs multi-stage preprocessing operations on the echo signal. The preprocessing operations include: data acquisition and buffering of the raw echo data; DC component removal processing to eliminate background level shift; suppression of low-frequency drift and high-frequency noise signals through high-pass or band-pass filtering; gain compensation of signal strength to unify the signal dynamic range; adaptive clipping of the time window based on the target detection depth; normalization processing of data from different sampling channels to improve signal consistency; further performing time-frequency transformation or wavelet decomposition to extract multi-scale spectral features; and constructing a two-dimensional B-Scan image or a three-dimensional radar data volume from the preprocessed data for subsequent identification processing. The preprocessing process can effectively enhance the texture features and temporal structure of the target reflection signal, suppress environmental interference signals, and improve the discriminability of the data and the input quality of the neural network model.

[0055] A deep learning recognition module is used to identify underground pipelines based on the preprocessed data. Based on the preprocessed radar data output by the signal processing module, feature extraction, target identification, and parameter estimation are performed on the underground pipelines. The deep learning recognition module includes a deep neural network model, which consists of multiple trainable layers. This model automatically learns the spatiotemporal distribution patterns and target reflection features in the radar image to achieve automatic identification of underground pipelines. The deep neural network model includes a multi-scale convolutional feature extraction unit for extracting local reflection features, a direction-aware convolutional structure unit for modeling the linear arrangement features of the pipeline along the detection direction, a spatial coherence modeling unit for analyzing the spatial relationships between response points and identifying connected structures, and an output layer for outputting the identification results. The identification results include the probability of the target pipeline's existence, the pipeline's center position, direction information, and estimated burial depth. The deep learning recognition module performs inference on the input radar image or signal sequence in an end-to-end manner and combines structural priors to enhance the response features of the target pipeline, suppressing misjudgments caused by non-target areas or interference signals, thereby achieving robust identification and location of underground pipeline targets.

[0056] Specifically, the deep learning recognition module includes:

[0057] The multi-scale convolutional feature extraction unit is used to extract spatial texture features and reflection features from radar images at different scales.

[0058] A multi-scale convolutional feature extraction unit is used to extract features from input radar images or radar signal data at multiple scales to fully acquire the spatial texture and electromagnetic reflection features of underground pipelines under different depths and reflection intensities. This feature extraction unit includes multiple parallel convolutional paths, each using convolutional kernels of different sizes to perform convolution operations on the input feature map. The kernel sizes include, but are not limited to, 1×1, 3×3, 5×5, and 7×7. Smaller kernels are used to extract local details and high-frequency texture information, while larger kernels are used to perceive continuous reflection structures and low-frequency signal distributions within a larger area.

[0059] In the specific implementation, the input radar image data is first processed in parallel through multi-scale convolutional layers. The multiple sets of feature maps output by each convolutional path are concatenated or weighted to form a unified high-dimensional feature representation, which is then input into subsequent deep neural networks for further processing. To improve feature selectivity and information preservation, batch normalization layers, non-linear activation functions (such as ReLU and LeakyReLU), and feature channel attention mechanism modules (such as SE or CBAM modules) can be optionally added after the convolutional paths to enhance key-scale responses and suppress redundant features.

[0060] The working principle of this multi-scale convolutional feature extraction unit is based on the multipath reflection characteristics of ground-penetrating radar signals during propagation in underground media. Because underground pipelines at different burial depths generate arc-shaped reflection echoes of different scales in radar images, and are influenced by factors such as the electromagnetic parameters, reflection angle, and material of the underground medium, their signal morphology manifests as a multi-scale texture structure in the image. Therefore, by constructing a multi-scale receptive field in the network, the reflection features of weak short arcs (shallow-buried targets) and wide arcs / long bands (deep-buried targets) can be captured simultaneously, thereby enhancing the model's ability to identify complex underground targets and its depth adaptability.

[0061] The beneficial effects of this feature extraction structure are twofold: firstly, it enhances the robustness of the network in recognizing underground pipelines at different scales and with different signal morphologies; secondly, it avoids the problem of feature information loss caused by a single convolution scale, thereby improving the detection rate of underground pipeline targets, the accuracy of burial depth estimation, and the generalization ability of the model. This structure is particularly suitable for typical application scenarios with complex underground environments, large differences in target burial depth, or weak reflected signals.

[0062] In one specific implementation method

[0063] The multi-scale convolutional feature extraction unit is used to process the preprocessed two-dimensional B-scan radar image data. The image size is 256×512 pixels, representing a radar reflection echo image collected in a typical urban road environment. The horizontal direction represents the moving distance of the detection device, and the vertical direction represents the electromagnetic wave propagation time (which can be converted into burial depth).

[0064] The multi-scale convolutional feature extraction unit includes three parallel convolutional paths, which use 3×3, 5×5 and 7×7 two-dimensional convolutional kernels for feature extraction, respectively. Its network structure is as follows:

[0065] Path 1 (Shallow Texture Extraction):

[0066] Kernel size: 3×3;

[0067] Number of convolution channels: 64;

[0068] Step size: 1;

[0069] Padding: 1 (preserve feature map size);

[0070] Activation function: ReLU;

[0071] Extracting the target: the narrow arc-shaped reflection features produced by the shallowly buried target.

[0072] Path Two (Mesoscale Structure Identification):

[0073] Kernel size: 5×5;

[0074] Number of convolution channels: 64;

[0075] Step size: 1;

[0076] Fill: 2;

[0077] Activation function: ReLU;

[0078] Target extraction: Reflected echoes from medium-depth targets to identify continuous textures.

[0079] Path 3 (Deep Reflection Structure Extraction):

[0080] Kernel size: 7×7;

[0081] Number of convolution channels: 64;

[0082] Step size: 1;

[0083] Fill: 3;

[0084] Activation function: ReLU;

[0085] Extracting the target: the wide arc-shaped, low-frequency reflected signal generated by the deeply buried target.

[0086] The three convolutional paths each output a feature map of the same size (256×512×64), which is then concatenated through channel dimensions to form a fused feature map (256×512×192), which is then input into the subsequent network structure for further processing.

[0087] The orientation-aware convolutional structure unit is used to apply orientation weights to the convolutional kernel based on the movement direction of the detection device and the linear arrangement characteristics of the target reflection signals in the radar image.

[0088] The orientation-aware convolutional structure unit is used in underground pipeline detection missions to combine the movement direction information of the detection equipment with the linear arrangement structure of the target reflection signal in the radar image to perform orientation-weighted processing on the response region of the convolutional kernel, thereby enhancing the network's ability to perceive the consistency characteristics of the target pipeline's orientation and suppressing the influence of spatially randomly distributed or discontinuous interference signals.

[0089] This structural unit includes the following components:

[0090] Direction encoding module:

[0091] Obtain the position change vector of the detection device at each moment. And calculate its unit direction vector. ;

[0092] Use the direction vector as the direction code Mapping to the angle domain or trigonometric function form, construct the direction guiding tensor. Its dimensions and features Figure 1 This is used for subsequent weight adjustments.

[0093] Directional weighted convolution kernel generation module:

[0094] Define the basic convolution kernel

[0095] According to direction coding Calculate the weight adjustment matrix

[0096]

[0097] Among them, among them, This represents the weight of the (i,j)th position of the total area kernel; This represents the angle between the (i,j)th position of the convolution kernel and the center direction.

[0098] Orientation-aware convolution operation:

[0099] For the input feature map Perform weighted convolution:

[0100]

[0101] in, This represents a weighted convolution, which enhances the response in regions with consistent detection directions (such as the actual pipeline direction) and automatically attenuates it in inconsistent directions (such as coupling interference).

[0102] The working principle of the direction-aware convolutional structure unit is as follows: When a ground-penetrating radar device moves, it will present continuously buried metal or non-metal pipelines as "a continuous, highly reflective arc extending linearly along the direction of movement" in the image. Environmental noise, coupled signals, or local abnormal reflections are mostly isolated, transient, and inconsistent in direction.

[0103] By constructing orientation-aware convolutional kernels, the neural network is guided to "focus" on linear structures aligned with the device's forward direction, thereby achieving:

[0104] Enhance the signal of the real target;

[0105] Suppress unstructured noise;

[0106] Enhance the ability to learn continuously.

[0107] In a specific implementation, the detection device is set to move along the X-axis, and the coordinates of the device's movement trajectory are recorded every 0.1 meters. The direction encoding vector is... The input radar image is a B-scan image with a size of 256×512 pixels. After processing by the pre-module, it enters the orientation-aware convolutional structure unit. This unit is configured as follows:

[0108] Kernel size: 5×5;

[0109] Number of convolution channels: 64;

[0110] The position of each convolutional kernel relative to the center point. Compare and generate a directional weighted matrix. ;

[0111] The weight center position (0°) is 1, and the weights on both sides decay exponentially (e.g., ...). );

[0112] All convolution operations are executed synchronously with the orientation weighting matrix.

[0113] For example, in one experiment, this structure significantly enhanced the "continuous high reflectivity feature along the X-axis" of the target pipeline in the image, while suppressing random noise interference from the Y-axis.

[0114] Spatial coherence modeling unit, used to model the connectivity of target response points based on their spatial positional relationships.

[0115] The spatial coherence modeling unit is used to model the connectivity, extensibility and structural continuity of target response points in radar images based on their spatial positional relationships. This enhances the overall recognition capability of deep learning models for targets with linear spatial structures, such as underground pipelines, and suppresses discontinuous, isolated or random noise responses.

[0116] Specifically, the spatial coherence modeling unit is implemented based on graph convolutional networks.

[0117] Each high-confidence response point in the feature map (output by the orientation-aware convolutional structure) is treated as a graph node; based on the relative spatial position of the response point in the image, edges are established according to Euclidean distance or orientation angle to form an adjacency matrix. ; Use graph convolutional layers to perform feature aggregation and propagation:

[0118]

[0119] in, For adjacency matrices with self-loops, For the features of the l-th layer nodes, These are trainable weights.

[0120] The network ultimately outputs a structural connectivity score for each response point, which is used to determine whether it belongs to a real pipeline.

[0121] The working principle of the spatial coherence modeling unit is as follows: Underground metal or non-metal pipelines are typically represented in images as continuous high-reflectivity structures arranged along a certain direction (such as the direction of movement of the detection equipment), and have the following spatial characteristics:

[0122] Locational coherence: adjacent response points are close to each other;

[0123] Consistent direction: small angle between response points;

[0124] Similar response intensity: small difference in reflection amplitude.

[0125] Environmental noise, interfering cables, and local anomalies often appear only in specific locations, or exhibit broken or discontinuous responses. Spatial coherence modeling utilizes these "spatial structural priors" to establish spatial constraints within the neural network through graph structures or attention mechanisms, thereby enhancing the ability to recognize linear structures.

[0126] In a specific example, in the aforementioned embodiment, the ground-penetrating radar device moves along the X-axis, and the obtained radar image (B-scan) is processed by a multi-scale convolution and orientation-aware convolution structure to output a feature map with a size of 256×512×64.

[0127] This embodiment introduces a spatial coherence modeling unit based on the feature map. The specific steps are as follows:

[0128] Pixels with response values ​​greater than 0.6 (normalized confidence) from the feature map are selected as candidate nodes, denoted as . ;

[0129] For any two nodes If its Euclidean distance Pixels, and the angle of orientation If so, then establish an edge connection and record the similarity weight;

[0130] A two-layer graph convolutional network is used to fuse features of the above nodes and extract the "structural connectivity score" of the nodes;

[0131] Set a structural connectivity threshold of 0.4, and only retain nodes with scores higher than the threshold as true target points;

[0132] The filtered results are backmapped to the original image to obtain continuous underground pipeline route markers, which are then transmitted to the visualization module for display.

[0133] The output layer is used to output the identification results, including the probability of target presence and the estimated pipeline burial depth.

[0134] The output layer is used to output the recognition results such as the existence probability, spatial location and burial depth estimate of the target pipeline based on the comprehensive feature map of the previous stage of the deep learning recognition module. The output layer includes a classification sub-branch for judging the probability of the target and a regression sub-branch for performing numerical regression of burial depth, which correspond to the output requirements of the binary classification task and the continuous variable prediction task, respectively.

[0135] Specifically, the output layer includes the following components:

[0136] Feature compression module:

[0137] The multi-channel feature map output by the previous network module (such as the spatial coherence modeling unit) Compress it into a fixed-dimensional vector representation through 1×1 convolution or fully connected operations.

[0138] Optionally, a global average pooling (GAP) operation can be added to decouple the output from the spatial dimension.

[0139] Classification branch (probability of target existence):

[0140] It contains a set of fully connected (FC) layers, and finally outputs the probability of the target's existence through a sigmoid or softmax activation function. Output a single value and set a threshold (e.g., 0.6) for the decision.

[0141] Regression branch (buried depth estimation):

[0142] It contains a set of fully connected (FC) layers, which output a real value through linear activation. This indicates the predicted pipeline burial depth;

[0143] Furthermore, to improve prediction stability, regularization modules such as Dropout and LayerNorm can be introduced;

[0144] Furthermore, this branch can be trained as a component of the multi-task loss function.

[0145] The spatial texture, directional consistency, and coherence features extracted by the pre-level feature extraction module characterize the presence of reflective structures with linear extension properties; the network then performs classification learning based on these features and outputs the probability of whether a real underground pipeline exists.

[0146] The reflection time in ground-penetrating radar echo signals is approximately linearly related to the target depth.

[0147]

[0148] in For burial depth, The speed at which radar waves propagate in underground media. This refers to the round-trip time.

[0149] The network learns from a large number of reflection images of known depths to establish a mapping function from the spatial structure of radar images to the numerical value of burial depth.

[0150] The model training and constraint module is used to introduce a structural continuity constraint loss function during the model training phase. The loss function includes: classification cross-entropy loss for pipeline identification, mean square error loss for burial depth regression, and structural smoothness constraint term for spatial response points.

[0151] The model training and constraint module is used to introduce multiple types of loss functions during the training phase of the neural network model. This guides the model to simultaneously learn the spatial classification features of underground pipelines, the numerical regression features of burial depth, and the spatial structural continuity features of the target in the image, thereby improving the model's recognition accuracy, prediction precision, and spatial consistency.

[0152] The model training and constraint module is configured to jointly optimize the following three types of loss functions:

[0153] Pipeline identification classification cross-entropy loss:

[0154] Used to constrain the consistency between the target existence probability output by the model and the true annotation.

[0155] The classification cross-entropy loss function for pipeline identification is expressed as:

[0156]

[0157] in, The probability of the target's existence is output by the model (Sigmoid function output); This is a real label; This represents the number of samples.

[0158] This loss function guides the model to approach the correct label in the probability space by maximizing the logarithm of the true class probability. It is sensitive to small probability errors and is beneficial for the model to converge quickly in the early training stage.

[0159] Mean squared error loss for burial depth regression:

[0160] This is used to constrain the difference between the target pipeline burial depth estimate output by the model and the actual burial depth.

[0161] The expression for the mean squared error loss in depth regression is:

[0162]

[0163] in, The burial depth value predicted by the model for the i-th sample; The depth of the sample is the actual burial depth; N is the number of samples.

[0164] This loss function averages the squared error between the predicted and the true values, thus penalizing predictions that deviate too much.

[0165] Structural smoothness constraint:

[0166] Used to impose continuity constraints on the response map output by the model in the image space, penalizing discontinuities, jumps, or nonlinear changes between adjacent spatial response points, reflecting the physical prior of underground pipelines as having "strong spatial extensibility and smooth structure".

[0167] The mathematical expression for the structural smoothness constraint term is:

[0168]

[0169] in, This is the response output of the model at position i,j.

[0170] This constraint utilizes the local smoothing properties of response values ​​in the image to guide the model to predict continuous trends and gradually changing structures, which conforms to the spatial structure patterns of underground pipelines.

[0171] The results visualization module is used to display the target pipeline's route tracking, burial depth marking, and confidence level indicator.

[0172] The results visualization module is used to graphically display the underground pipeline identification results output by the deep learning recognition module, including visual annotations of the target pipeline's trajectory, burial depth information, and recognition confidence level, to assist users in making intuitive judgments and subsequent engineering decisions.

[0173] The result visualization module includes an image overlay unit, a burial depth annotation unit, a confidence mapping unit, and a geographic coordinate mapping unit, with the following specific functions:

[0174] Image overlay unit

[0175] This is used to overlay the identified pipeline target area onto the original radar image (B-scan or C-scan) in the form of pseudo-color, outline, or connected lines, so that users can intuitively observe the spatial distribution characteristics of the identified area and its corresponding reflection intensity.

[0176] Walking towards the tracking unit

[0177] This tool is used to extract and draw the direction curve of underground pipelines based on the spatial relationship between continuous response points in the identification results. The direction can be displayed as vector line segments on radar images or two-dimensional plan maps (such as GIS maps), and supports parallel or intersecting display of multiple pipeline segments.

[0178] Burial depth marking unit

[0179] This is used to mark the predicted burial depth at selected key points of the traverse line segment. The burial depth information is represented in the form of floating labels, contour lines, color depth, or numerical annotations, and can be automatically thinned out according to the interval set by the user.

[0180] Confidence mapping unit

[0181] Based on the probability of target presence or classification confidence level output by the recognition module, the confidence level of the target region is configured, and the region is visualized and encoded through color gradient (such as red-yellow-green) or transparency changes, so that users can distinguish between high-reliability and low-reliability recognition regions.

[0182] Geographic coordinate mapping unit (optional)

[0183] When the detection device is integrated with GPS, IMU or ranging device, the module can map the identification results to a geographic coordinate system or CAD drawing to realize the geographic marking and route tracking of the target pipeline on the actual map.

[0184] User interaction interface

[0185] It supports users to zoom in, zoom out, rotate, export data, or perform partial queries on visualized content, improving the interactive experience and the flexibility of data use.

[0186] See Figure 2 In another embodiment, the present invention also provides a method for detecting underground pipelines based on ground-penetrating radar and deep learning, comprising:

[0187] Acquire ground-penetrating radar echo signals;

[0188] The echo signal is preprocessed, including operations such as background removal, noise reduction, enhancement, normalization, or time window adjustment.

[0189] The preprocessed radar image is input into a deep neural network model to perform underground pipeline identification processing. The deep neural network model includes:

[0190] a) Extracting spatial texture and reflection features from radar images at different scales;

[0191] b) Based on the linear arrangement characteristics of the moving direction of the detection equipment and the target reflection signal, the convolution kernel is directionally weighted;

[0192] c) Model the connectivity of target response points based on their spatial relationships;

[0193] d) Output the identification results of the target existence probability and the estimated pipeline burial depth;

[0194] During the model training phase, a loss function with structural continuity constraints is introduced, which includes: classification cross-entropy loss for pipeline identification, mean square error loss for burial depth estimation, and structural smoothness constraint term for constraining the spatial continuity of response points.

[0195] The identification results are visualized, including the target pipeline's route tracking, burial depth marking, and a visual mapping of the identification confidence level.

[0196] It should be noted that the explanations and descriptions of the aforementioned underground pipeline detection system based on ground penetrating radar and deep learning also apply to the methods of the embodiments of this application, and will not be repeated here.

[0197] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0198] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0199] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, 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.

[0200] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. For some module structures not specifically defined in this invention, the content described in the prior art shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered as part of this invention and used to understand the meaning of some technical features or parameters.

Claims

1.A ground penetrating radar and deep learning based underground pipeline detection system, characterized in that, The system comprises: a signal processing module for acquiring ground penetrating radar echo signals and pre-processing the echo signals; a deep learning identification module for identifying underground pipelines based on the pre-processed data, the deep learning identification module comprising: a multi-scale convolution feature extraction unit for extracting spatial texture features and reflection features in the radar image from different scales; a direction-aware convolution structure unit for directionally weighting the convolution kernel according to the movement direction of the detection device and the linear arrangement characteristics of the target reflection signals in the radar image; a spatial coherence modeling unit for modeling the connectivity of the target response points based on their spatial position relationship; an output layer for outputting the identification results of the target existence probability and the pipeline burial depth estimate; a model training and constraint module for introducing a structural continuity constraint loss function in the model training stage, the loss function comprising: a classification cross-entropy loss of pipeline identification, a mean square error loss of burial depth regression, and a structural smoothness constraint term for spatial response points; a result visualization module for displaying the target pipeline trend tracking, burial depth labeling and confidence identification. 2.The ground penetrating radar and deep learning based underground pipeline detection system of claim 1, wherein, The multi-scale convolution feature extraction unit comprises a plurality of parallel convolution paths, each convolution path using a convolution kernel of different size to perform convolution operation on the radar image, the convolution kernel size including 3x3, 5x5 and 7x7, and the features extracted at multiple scales are fused by channel concatenation. 3.The ground penetrating radar and deep learning based underground pipeline detection system of claim 1, wherein, The direction-aware convolution structure unit comprises a direction encoding module and a directionally weighted convolution kernel generation module, the direction encoding module being used to obtain the movement direction vector of the detection device and assign direction weighting coefficients to each sampling position of the convolution kernel according to the direction vector to enhance the response features along the direction of the target pipeline trend. 4.The ground penetrating radar and deep learning based underground pipeline detection system of claim 1, wherein, The spatial coherence modeling unit is a graph convolution network structure, the graph convolution network taking high-confidence response points in the radar image as graph nodes, constructing adjacent edges according to the spatial position relationship between the nodes, and performing graph convolution operation to identify the connected structure and suppress isolated noise responses. 5.The ground penetrating radar and deep learning based underground pipeline detection system of claim 1, wherein, The structural smoothness constraint term minimizes the spatial second-order difference of the prediction results to guide the model to output identification results with spatial continuity and smoothness, and the structural smoothness constraint term is defined as: 6.The ground penetrating radar and deep learning based underground pipeline detection system of claim 1, wherein, The result visualization module further comprises a confidence mapping unit for color gradient coding and displaying the pipeline area according to the target existence probability output by the deep learning identification module, and reflecting the identification confidence level in the form of color depth or transparency change on the radar image. 7.A method for detecting underground pipelines based on ground penetrating radar and deep learning, characterized by, The method comprises the following steps: acquiring ground penetrating radar echo signals; pre-processing the echo signals, the pre-processing including background removal, denoising, enhancement, normalization or time window adjustment operation; inputting the pre-processed radar image into a deep neural network model to perform underground pipeline identification processing, the deep neural network model comprising: a) extracting spatial texture features and reflection features of the radar image from different scales; b) directionally weighting the convolution kernel according to the moving direction of the detection device and the linear arrangement characteristic of the target reflection signal; c) modeling the connectivity of the target response points based on the spatial position relationship between the target response points; d) outputting the identification result of the target existence probability and the pipeline buried depth estimation value; In the model training stage, a loss function with structure continuity constraint is introduced, which includes: a classification cross-entropy loss for pipeline identification, a mean square error loss for buried depth estimation, and a structure smoothness constraint term for constraining the spatial continuity of response points; The identification result is visually displayed, which includes the trend tracking of the target pipeline, the buried depth annotation, and the visual mapping of the identification confidence. 8.The ground penetrating radar and deep learning based underground pipeline detection method of claim 7, wherein, The spatial texture features and reflection features of the radar image are extracted from different scales, including setting multiple parallel convolution paths, each convolution path using a convolution kernel of different size to perform convolution operation on the radar image, the convolution kernel size including 3x3, 5x5 and 7x7, and fusing the feature maps extracted under multiple scales through channel splicing. 9.The ground penetrating radar and deep learning based underground pipeline detection method of claim 7, wherein, Directionally weighting the convolution kernel according to the moving direction of the detection device and the linear arrangement characteristic of the target reflection signal includes using a direction encoding module and a directionally weighted convolution kernel generation module, the direction encoding module is used to obtain the moving direction vector of the detection device, and according to the direction vector, direction weighting coefficients are assigned to each sampling position of the convolution kernel to enhance the response features along the direction of the target pipeline trend. 10.The ground penetrating radar and deep learning based underground pipeline detection method of claim 7, wherein, Modeling the connectivity of the target response points based on the spatial position relationship between the target response points includes taking high confidence response points in the radar image as graph nodes, constructing adjacent edges according to the spatial position relationship between the nodes, performing graph convolution operation to identify the connected structure and suppress the loss of isolated noise response.

Images