Deep learning-based underground pipeline detection method

Abstract

This invention relates to the field of underground pipeline detection technology and discloses a deep learning-based method for underground pipeline detection. This method transforms raw geophysical field data into an enhanced semantic field through a deep feature generation network, while simultaneously parsing pipeline attribute description data into structured constraint rules. A pipeline existence verification network couples these two methods for reasoning, identifying the potential voxel space and extracting the initial 3D morphology of the pipeline. A differentiable rendering module projects the 3D results onto 2D data, compares them with the original data, and collaboratively optimizes and iteratively updates the network parameters accordingly. This method achieves coupled reasoning guided by prior knowledge and cross-dimensional consistency optimization, improving the accuracy and physical consistency of pipeline detection under complex interference environments.

Description

Technical Field

[0001] This invention relates to the field of underground pipeline detection technology, specifically a deep learning-based method for underground pipeline detection. Background Technology

[0002] Traditional underground pipeline detection primarily relies on geophysical exploration methods combined with manual interpretation. In recent years, deep learning technology, especially image segmentation networks, has been introduced for automatic identification of pipeline reflection signals. Typically, geophysical field data is treated as images, and the network is trained to directly output the pipeline's location or mask. However, this approach heavily depends on the quality of the data and annotations; the model's learning stems entirely from the statistical characteristics within the data, failing to incorporate prior knowledge and constraints inherent in pipeline engineering. When geophysical data quality is poor or complex interference exists, data-driven models are prone to misjudgments or omissions, and their outputs lack interpretability and verification from both physical and engineering perspectives.

[0003] Existing technical solutions suffer from two main drawbacks. Firstly, descriptive attributes such as pipeline material, approximate burial depth, and diameter range are typically presented independently in text form, separate from the geophysical data processing flow. Current methods lack a mechanism to integrate this non-image, structured prior knowledge into the deep learning model's inference process. Secondly, the extraction and optimization of 3D pipeline structures are usually conducted in a closed 3D space, lacking a direct, differentiable feedback link between the optimization objective and the original 2D observation data. This results in the reconstructed 3D model potentially being inconsistent with the actual acquired physical field data at the observational level, making it difficult to correct and refine the model using the original data itself. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for detecting underground pipelines, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a deep learning-based method for detecting underground pipelines, the method comprising:

[0006] Acquire raw geophysical field data and pipeline attribute description data for the target area;

[0007] The original geophysical field data is input into the depth feature generation network to generate the corresponding initial semantic field of the underground space.

[0008] The initial semantic field of underground space is subjected to hierarchical attention-weighted fusion to form an enhanced semantic field with multi-scale contextual information;

[0009] Parse pipeline attribute description data to generate structured pipeline constraint rules;

[0010] A pipeline existence verification network is used to couple the enhanced semantic field with pipeline constraint rules to identify the potential voxel space where pipelines exist.

[0011] Based on the latent voxel space, three-dimensional morphological operations are performed on the enhanced semantic field to extract the initial skeleton and boundary contour of the pipeline structure.

[0012] The initial skeleton and boundary contour are projected onto a two-dimensional view using the differentiable rendering module, and then compared with the original geophysical field data at the pixel level.

[0013] Based on the results of pixel-level consistency comparison, the internal parameters of the pipeline existence verification network and the deep feature generation network are collaboratively optimized and adjusted.

[0014] The original geophysical field data is reprocessed using the optimized network, and the initial skeleton and boundary profile of the pipeline structure are iteratively updated until the convergence condition is met.

[0015] By integrating the finally converged pipeline skeleton, boundary contour, and pipeline attribute description data, a pipeline detection result map containing three-dimensional coordinates and attributes is generated.

[0016] Preferably, the initial semantic field of the underground space is subjected to hierarchical attention-weighted fusion to form an enhanced semantic field with multi-scale contextual information, including:

[0017] Extract multiple intermediate layer feature maps corresponding to different depths from a deep feature generation network;

[0018] For each intermediate layer feature map, channel attention and spatial attention mechanisms are applied respectively, and the corresponding channel weight matrix and spatial weight matrix are calculated.

[0019] Multiply the channel weight matrix and the spatial weight matrix element by element to obtain the attention mask corresponding to each intermediate layer feature map;

[0020] The attention mask is used to perform weighted modulation on the corresponding intermediate layer feature maps to obtain multiple weighted feature representations;

[0021] A feature pyramid structure is adopted to perform top-down and horizontal connection operations on multiple weighted feature representations to achieve multi-scale feature fusion;

[0022] The fused feature maps are batch normalized and nonlinearly activated to output an enhanced semantic field.

[0023] Preferably, the pipeline attribute description data is parsed to generate structured pipeline constraint rules, including:

[0024] The natural language description in the pipeline attribute description data is segmented into independent semantic segments describing material, pipe diameter, burial depth, and direction;

[0025] Each independent semantic segment is converted into a high-dimensional word vector representation sequence;

[0026] The word vector representation sequence is input into a bidirectional recurrent neural network to obtain the context encoding of each semantic segment;

[0027] Input the context encoding into the sequence labeling model to label the key entities and corresponding values ​​that describe the physical properties of the pipeline.

[0028] Based on the marked key entities and their corresponding values, the pipeline constraint rules are instantiated according to the predefined syntax template to generate computer-executable rules. The pipeline constraint rules include the definition of pipeline reflection intensity threshold, geometric continuity, and spatial location range.

[0029] Preferably, a pipeline existence verification network is used to couple the enhanced semantic field with pipeline constraint rules for reasoning, identifying the potential voxel space where pipelines exist, including:

[0030] The enhanced semantic field is sliced ​​in the depth dimension to generate a series of two-dimensional feature slices;

[0031] Each two-dimensional feature slice is spliced ​​together with the part of the pipeline constraint rules concerning reflection intensity and geometry in the feature channel dimension.

[0032] The concatenated features are input into a lightweight convolutional network, which outputs the probability that each spatial location belongs to the pipeline region.

[0033] The probabilities of all depth slices are stacked to reconstruct a three-dimensional probability field with the same spatial dimensions as the original geophysical field data.

[0034] In the three-dimensional probability field, based on the definition of spatial location range in the pipeline constraint rules, a dynamic probability threshold is set to perform binarization segmentation of the probability field;

[0035] The binarized segmentation results are subjected to the requirements of geometric continuity in pipeline constraint rules to perform connected component analysis and screening, and output the potential voxel space.

[0036] Preferably, based on the latent voxel space, three-dimensional morphological operations are performed on the enhanced semantic field to extract the initial skeleton and boundary contour of the pipeline structure, including:

[0037] Using the latent voxel space as a mask, the corresponding sub-region semantic field is clipped from the enhanced semantic field;

[0038] Three-dimensional Gaussian filtering is applied to the semantic field of the sub-region to smooth the local irregularities caused by data noise.

[0039] The three-dimensional Sobel gradient operator is applied to the smoothed sub-region semantic field to calculate the gradient intensity of each voxel in three orthogonal directions;

[0040] By synthesizing the gradient intensities in the three directions, the gradient magnitude field of the semantic field of the entire sub-region is obtained;

[0041] The evolution is performed using a three-dimensional level set method in the gradient magnitude field to capture interfaces with drastic gradient changes and form the boundary profile of the pipeline structure.

[0042] For the internal region enclosed by the boundary contour, a three-dimensional thinning algorithm is used for iterative erosion until the structural width is a single voxel, thus extracting the initial skeleton of the pipeline structure.

[0043] Preferably, the initial skeleton and boundary contour are projected onto a two-dimensional view using a differentiable rendering module, and a pixel-level consistency comparison is performed with the original geophysical field data, including:

[0044] Based on the acquisition perspective parameters of the original geophysical field data, a differentiable projection transformation model from three-dimensional space to two-dimensional imaging plane is established.

[0045] Using a differentiable projection transformation model, the initial skeleton and boundary contour of the pipeline structure are rendered into two-dimensional skeleton projection and contour projection, respectively.

[0046] On the two-dimensional imaging plane corresponding to the original geophysical field data, the original signal intensity profile corresponding to the positions of the skeleton projection map and the contour projection map is extracted;

[0047] The difference between the original signal intensity corresponding to each projection point in the skeleton projection map and the expected reflection intensity defined in the pipeline constraint rules is calculated to obtain the skeleton intensity consistency error.

[0048] The difference between the original signal intensity gradient on both sides of the projected contour in the contour projection map and the theoretical gradient model is calculated to obtain the contour gradient consistency error.

[0049] The weighted summation of the skeleton strength consistency error and the contour gradient consistency error yields the pixel-level consistency comparison result.

[0050] Preferably, based on the results of pixel-level consistency comparison, the internal parameters of the pipeline existence verification network and the deep feature generation network are collaboratively optimized and adjusted, including:

[0051] The result of pixel-level consistency comparison is used as a loss signal and is passed back to the differentiable rendering module.

[0052] The loss signal is further backpropagated through the projection transformation model in the differentiable rendering module to the three-dimensional morphological operation steps of the initial skeleton and boundary contour of the generated pipeline structure;

[0053] The loss signal continues to propagate back to the pipeline existence verification network and the depth feature generation network through a three-dimensional morphological operation step;

[0054] The gradient descent algorithm is used to update the convolutional kernel weight parameters of the pipeline existence verification network based on the loss gradient propagated to the pipeline existence verification network.

[0055] Simultaneously, based on the loss gradient propagated to the deep feature generation network, the parameters of the fully connected layers and convolutional layers of the deep feature generation network are updated.

[0056] When updating parameters, the gradients from skeleton strength consistency error and contour gradient consistency error are normalized to balance the optimization speed of skeleton strength consistency error and contour gradient consistency error.

[0057] Preferably, the original geophysical field data is reprocessed using the optimized network to iteratively update the initial skeleton and boundary contour of the pipeline structure until the convergence condition is met, including:

[0058] One complete process of co-optimizing and adjusting internal parameters is recorded as one iteration.

[0059] After one iteration, the original geophysical field data is reprocessed using a deep feature generation network with updated internal parameters to generate a new enhanced semantic field.

[0060] Using a pipeline existence verification network with updated internal parameters, the new enhanced semantic field is coupled with pipeline constraint rules to generate a new latent voxel space.

[0061] Based on the new potential voxel space, three-dimensional morphological operations are performed again on the new enhanced semantic field to generate the initial skeleton and boundary contour of the updated pipeline structure.

[0062] Calculate the structural similarity index between the initial skeleton and boundary profile of the updated pipeline structure and the initial skeleton and boundary profile of the pipeline structure generated in the previous iteration;

[0063] Determine if the structural similarity index is greater than the preset convergence threshold. If it is, the convergence condition is met and the iteration stops; otherwise, the next iteration begins.

[0064] Preferably, the original geophysical field data and pipeline attribute description data of the target area are obtained, including:

[0065] A sensor array is deployed to perform a wide-area scan of the target area, collecting raw field data including electromagnetic wave reflection signals and seismic wave signals;

[0066] The original field data were preprocessed by time synchronization alignment, background noise suppression, and amplitude recovery to obtain standardized original geophysical field data;

[0067] Retrieve textual and graphical data related to the target area from urban planning archives or pipeline as-built drawings;

[0068] Optical character recognition and graphic symbol recognition are performed on text and chart data to extract descriptive text and parameter tables about pipelines, which are used as pipeline attribute description data.

[0069] Preferably, the pipeline skeleton, boundary contour, and pipeline attribute description data that have been finally converged are integrated to generate a pipeline detection result map containing three-dimensional coordinates and attributes, including:

[0070] The voxel coordinates of the finally converged pipeline skeleton are transformed into a three-dimensional coordinate sequence in the geodetic coordinate system through spatial coordinate transformation.

[0071] The voxel coordinates of the boundary contour of the pipeline structure are also converted into a three-dimensional mesh surface in the geodetic coordinate system.

[0072] From the pipeline constraint rules parsed from the pipeline attribute description data, the material, pipe diameter, and burial depth attributes are extracted and associated with the corresponding line segments in the three-dimensional coordinate sequence;

[0073] The three-dimensional coordinate sequence with attributes is defined as the centerline vector of the pipeline, and the three-dimensional mesh surface is defined as the outer surface model of the pipeline.

[0074] The centerline vector of the pipeline, the outer surface model, and the associated attribute information are imported into a 3D geographic information system platform for visualization rendering to generate a pipeline detection result map.

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] Structured pipeline constraint rules are generated by parsing pipeline attribute description data, and a dedicated pipeline existence verification network is used to couple these rules with an enhanced semantic field extracted from the physical field for inference. This allows the model to rely not only on data features but also be guided and constrained by explicit prior rules when performing voxel space identification. In complex scenarios with ambiguous geophysical data, low signal-to-noise ratios, or interference similar to pipelines, these rules can suppress unreasonable predictions, focusing the identification range on spatial regions that conform to engineering physics common sense, improving the reliability and rationality of the detection results, and reducing the absurd outputs that might arise from purely data-driven models.

[0077] A differentiable rendering module is introduced to project the initial pipeline skeleton and boundary contour extracted in 3D space back into a 2D view. This view is then compared pixel-level with the original geophysical field data, and the gradient generated by the comparison differences is used to collaboratively optimize the parameters of the feature generation and validation networks. A closed-loop feedback path is established from the 3D reconstruction results to the original 2D observation data. Optimization forces the 3D semantic representations and inference results generated within the network to match the actual physical field observation data through rendering inversion, achieving cross-dimensional self-supervision. This improves the fidelity of the 3D reconstruction results to the original data, giving the extracted pipeline morphology higher interpretability at the observation level, and achieving calibration between the model output and the original data evidence. Attached Figure Description

[0078] Figure 1 This is a schematic diagram illustrating the working principle of the deep learning-based underground pipeline detection method described in this invention.

[0079] Figure 2 A flowchart for generating multi-scale enhanced semantic fields;

[0080] Figure 3 A flowchart for identifying potential voxel spaces for pipelines;

[0081] Figure 4 A polygonal plot showing the gradient intensity distribution of three-dimensional morphological operations during underground pipeline detection;

[0082] Figure 5 This is a two-dimensional planar distribution diagram of underground pipelines. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] Please see Figure 1This invention provides a deep learning-based method for detecting underground pipelines. The method includes: acquiring raw geophysical field data and pipeline attribute description data of a target area; inputting the raw geophysical field data into a deep feature generation network to generate a corresponding initial semantic field of underground space; performing hierarchical attention-weighted fusion on the initial semantic field to form an enhanced semantic field with multi-scale contextual information; parsing the pipeline attribute description data to generate structured pipeline constraint rules; using a pipeline existence verification network to couple the enhanced semantic field and pipeline constraint rules for inference, identifying the potential voxel space of pipeline existence; performing three-dimensional morphological operations on the enhanced semantic field based on the potential voxel space to extract the initial skeleton and boundary contour of the pipeline structure; projecting the initial skeleton and boundary contour onto a two-dimensional view using a differentiable rendering module and performing pixel-level consistency comparison with the raw geophysical field data; co-optimizing and adjusting the internal parameters of the pipeline existence verification network and the deep feature generation network based on the pixel-level consistency comparison results; and reprocessing the raw geophysical field data using the optimized network to iteratively update the initial skeleton and boundary contour of the pipeline structure until convergence conditions are met. By integrating the finally converged pipeline skeleton, boundary contour, and pipeline attribute description data, a pipeline detection result map containing three-dimensional coordinates and attributes is generated.

[0085] Example 1: See Figure 2 The process of acquiring raw geophysical field data and pipeline attribute description data for the target area involves deploying a sensor array to perform a wide-area scan of the target area, collecting raw field data containing electromagnetic wave reflection signals and seismic wave signals. The raw field data undergoes time synchronization alignment, background noise suppression, and amplitude recovery preprocessing to obtain standardized raw geophysical field data. Text and graphic data related to the target area are retrieved from urban planning archives or pipeline as-built drawings. Optical character recognition and graphic symbol recognition are performed on the text and graphic data to extract descriptive text and parameter tables about the pipelines, which serve as pipeline attribute description data. After obtaining the raw geophysical field data, it is input into a depth feature generation network to generate an initial semantic field for the underground space. The specific process of performing hierarchical attention-weighted fusion on the initial semantic field of the underground space to form an enhanced semantic field with multi-scale contextual information involves extracting multiple intermediate layer feature maps corresponding to different depths in the depth feature generation network. Channel attention and spatial attention mechanisms are applied to each intermediate layer feature map, and the corresponding channel weight matrix and spatial weight matrix are calculated. The channel weight matrix and the spatial weight matrix are multiplied element-wise to obtain the attention mask corresponding to each intermediate layer feature map. The attention mask is then used to perform weighted modulation on the corresponding intermediate layer feature maps to obtain multiple weighted feature representations. A feature pyramid structure is employed to perform top-down and horizontal concatenation operations on the multiple weighted feature representations, achieving multi-scale feature fusion.

[0086] In practical implementation, taking the detection of pipelines beneath urban roads as an example, the acquisition of raw geophysical field data is achieved by deploying a sensor array to perform a wide-area scan of the target area. The sensor array includes multiple ground-penetrating radar antenna units arranged in a predetermined grid and seismic wave excitation-receiver pairs. These sensor units work synchronously to collect raw field data containing electromagnetic wave reflection signals and seismic wave signals, which are represented as a series of time-series waveforms. Subsequently, the processing system performs time synchronization alignment on the raw field data to eliminate temporal drift caused by minor differences in instrument startup. Then, a wavelet transform algorithm is applied to suppress background noise, and a gain recovery function is used for amplitude recovery preprocessing. Finally, standardized raw geophysical field data with spatial coordinate alignment, improved signal-to-noise ratio, and normalized amplitude characteristics is obtained, which constitutes a three-dimensional data volume. Simultaneously, pipeline attribute description data is acquired by retrieving text reports and digitized pipeline as-built drawings related to the target road from the electronic database of the urban planning archive. The processing system performs optical character recognition on the text report, converting the image text into machine-readable text. It also performs graphic symbol recognition on the as-built drawings, extracting pipeline annotations and parameter tables from the legends. All the descriptive text about the pipelines identified, together with the parameter tables, constitute the pipeline attribute description data. A typical description may include information such as "Φ600 reinforced concrete rainwater pipe, buried at a depth of 2.5 meters, running along the center line of Zhongshan North Road".

[0087] In the specific implementation, standardized raw geophysical field data is input into a deep feature generation network, which is a three-dimensional convolutional neural network with an encoder-decoder structure. The raw geophysical field data serves as the input tensor, undergoing layer-by-layer abstraction through multiple convolutional and pooling layers. Finally, the decoder outputs an initial semantic field of the subsurface space, corresponding to the input spatial size, where each voxel contains a high-level semantic feature vector. To enhance the representational power of the initial semantic field, a hierarchical attention-weighted fusion process is applied. This involves extracting multiple intermediate feature maps corresponding to different depths from the encoder portion of the deep feature generation network. These intermediate feature maps have different spatial resolutions and semantic abstraction levels. Channel attention and spatial attention mechanisms are applied to each intermediate feature map. The channel attention mechanism calculates a channel weight matrix reflecting the importance of each channel through global average pooling and fully connected layer operations. The spatial attention mechanism focuses on salient regions in space through convolutional operations, calculating a spatial weight matrix.

[0088] In some embodiments, the channel weight matrix and the spatial weight matrix are fused by element-wise multiplication to obtain the attention mask corresponding to each intermediate layer feature map. The calculation of the attention mask can be expressed as: ;

[0089] in: Representing the One intermediate layer feature map This indicates a global average pooling operation. and These represent the learnable parameters used for channel and spatial attention computation, respectively. and Represents a non-linear activation function. This represents element-wise multiplication. This represents the convolution operation. This is the obtained attention mask. The attention mask is then used to perform weighted modulation on the corresponding intermediate layer feature maps, i.e., element-wise multiplication. This results in multiple weighted feature representations, where pipeline-related features are enhanced and irrelevant background features are suppressed.

[0090] In some embodiments, after obtaining multiple weighted feature representations, a feature pyramid structure is used to fuse multi-scale contextual information. The feature pyramid structure uses a top-down upsampling path to progressively upsample the deepest, highly semantically weighted feature representations to the same spatial size as the shallower layers. Simultaneously, through lateral connections, the upsampled features are element-wise added to the corresponding shallow, high-resolution weighted feature representations. This process achieves complementary fusion of high-level semantic information and low-level detail information. The fused feature maps are batch normalized for stable training and transformed using a non-linear activation function, ultimately outputting an enhanced semantic field with rich multi-scale contextual information. In road detection scenarios, the enhanced semantic field can more clearly distinguish the reflection features of pipelines from the responses of surrounding soil, gravel, or other underground obstacles.

[0091] Example 2: See Figure 3The process of parsing pipeline attribute description data and generating structured pipeline constraint rules involves segmenting the natural language descriptions in the pipeline attribute description data into independent semantic segments describing material, pipe diameter, burial depth, and orientation. Each independent semantic segment is converted into a high-dimensional word vector representation sequence. The word vector representation sequence is input into a bidirectional recurrent neural network to obtain the context encoding of each semantic segment. The context encoding is input into a sequence labeling model to label the key entities describing the pipeline's physical attributes and their corresponding values. Based on the labeled key entities and their corresponding values, a predefined syntax template is used to instantiate and generate computer-executable pipeline constraint rules. The pipeline constraint rules include definitions of pipeline reflection intensity threshold, geometric continuity, and spatial location range. A pipeline existence verification network is used to couple the enhanced semantic field with the pipeline constraint rules to identify the potential voxel space where the pipeline exists. This process involves slicing the enhanced semantic field in the depth dimension to generate a series of two-dimensional feature slices. Each two-dimensional feature slice is then concatenated with the parts of the pipeline constraint rules concerning reflection intensity and geometric shape in the feature channel dimension. The concatenated features are input into a lightweight convolutional network, which outputs the probability that each spatial location belongs to the pipeline region. The probabilities of all depth slices are stacked to reconstruct a 3D probability field with the same spatial dimensions as the original geophysical data. Within this 3D probability field, a dynamic probability threshold is set based on the definition of spatial location range in the pipeline constraint rules, and the probability field is binarized. The geometric continuity requirements of the pipeline constraint rules are applied to the binarized segmentation results, and connected component analysis and filtering are performed to output a latent voxel space.

[0092] In practice, pipeline attribute description data includes natural language descriptions such as "DN300 cast iron water supply pipe, buried at a depth of 1.8 meters, laid parallel to the south sidewalk of Renmin Road". The process of parsing pipeline attribute description data to generate structured pipeline constraint rules begins by inputting the entire natural language description into a text preprocessing module. Based on punctuation and domain keywords, the text preprocessing module segments the description into independent semantic fragments describing the material, pipe diameter, burial depth, and direction, such as "DN300 cast iron water supply pipe", "buried at a depth of 1.8 meters", and "laid parallel to the south sidewalk of Renmin Road". Subsequently, each independent semantic fragment is fed into a pre-trained word embedding layer, which converts each word in the fragment into a fixed-dimensional high-dimensional word vector, thus forming a sequence of word vector representations. These word vector representation sequences are sequentially input into a bidirectional recurrent neural network, which processes the sequence from both forward and backward directions, capturing the contextual information before and after each word and outputting a contextual encoding that incorporates the global context. After obtaining the contextual encoding, it is input into a sequence labeling model based on a conditional random field. The sequence labeling model assigns a label to each position in the context encoding, annotating the key entities and corresponding values ​​describing the pipeline's physical properties. The system instantiates the labeled entities and values ​​into corresponding templates, generating computer-executable pipeline constraint rules. These rules include definitions for pipeline reflection intensity thresholds, geometric continuity, and spatial location ranges.

[0093] In practical implementation, a pipeline existence verification network is used to couple the enhanced semantic field with pipeline constraint rules to identify the potential voxel space where pipelines exist. This process first uniformly slices the 3D enhanced semantic field along the depth direction, generating a series of 2D feature slices corresponding to different depths. Simultaneously, the numerical definitions of reflection intensity and geometry in the pipeline constraint rules, such as "reflection intensity threshold: 0.65" and "geometric continuity: rate of change of orientation angle < 5", are converted into scalar or vector forms. For each 2D feature slice, the processing system concatenates these rule-defined scalar or vector transformations at the end of its feature channel dimension, forming a new feature map with increased channel count. This concatenation operation ensures that the rule information and the original data features are associated in the same spatial location.

[0094] The process involves inputting the newly concatenated feature map into a lightweight convolutional network (CNN), which consists of several depthwise separable convolutional layers and point convolutional layers. The CNN performs nonlinear transformations and spatial information aggregation on the input features, ultimately outputting a two-dimensional probability map through a sigmoid activation function. Each pixel value in the probability map represents the probability that its corresponding spatial location belongs to the pipeline region, with a value range of [0,1]. This process is repeated for all depth slices, and the output two-dimensional probability maps are stacked according to their original depth order to reconstruct a three-dimensional probability field with spatial dimensions identical to the original geophysical field data. In the three-dimensional probability field, a dynamic probability threshold is set based on the definition of spatial location range in the pipeline constraint rules. For example, within the defined depth range of 1.6 meters to 2.0 meters, the probability threshold is set to 0.5; outside this range, the probability threshold is set to 0.8. The dynamic probability threshold is then used to binarize and segment the entire three-dimensional probability field, yielding preliminary three-dimensional voxel segmentation results.

[0095] Optionally, the requirements for geometric continuity in the pipeline constraint rules are applied to the binarized segmentation results to perform connected component analysis and filtering. Specifically, the system uses a 3D connected component labeling algorithm to identify all interconnected 3D voxel clumps, calculates the centerline orientation of each clumper, and calculates the angle change between adjacent line segments of the clumper centerline based on the condition of "angle change rate < 5 degrees / meter" in the pipeline constraint rules. If the average angle change of a clumper exceeds the threshold, the clumper is judged to be non-compliant with the geometric continuity requirement and is filtered out. The connected components retained after filtering occupy the 3D voxel space, which is the potential voxel space where the identified pipelines exist. The shape of the potential voxel space reflects the possible distribution of the pipelines in 3D space, and the spatial distribution of probability values ​​can be expressed in the following form: ;

[0096] in: Represents a three-dimensional probability field in coordinates The probability value at that location. Represents a two-dimensional feature slice. This represents the transformed feature vector of the relevant part in the pipeline constraint rules. This represents a concatenation operation along the feature channel dimension. The mapping function representing a lightweight convolutional network. This represents the Sigmoid normalization function.

[0097] Example 3: The process of extracting the initial skeleton and boundary contour of the pipeline structure by performing 3D morphological operations on the enhanced semantic field based on latent voxel space involves using the latent voxel space as a mask to crop out the corresponding sub-region semantic field from the enhanced semantic field. A 3D Gaussian filter is applied to the sub-region semantic field to smooth local irregularities caused by data noise. A 3D Sobel gradient operator is applied to the smoothed sub-region semantic field to calculate the gradient intensity of each voxel in three orthogonal directions. The gradient intensities in the three directions are synthesized to obtain the gradient magnitude field of the entire sub-region semantic field. A 3D level set method is used to evolve the gradient magnitude field, capturing interfaces with drastic gradient changes to form the boundary contour of the pipeline structure. The internal region enclosed by the boundary contour is iteratively eroded using a 3D thinning algorithm until the structural width is a single voxel, thus extracting the initial skeleton of the pipeline structure. The process of projecting the initial skeleton and boundary contour onto a 2D view using the differentiable rendering module and comparing it with the original geophysical field data at the pixel level involves establishing a differentiable projection transformation model from 3D space to a 2D imaging plane based on the acquisition perspective parameters of the original geophysical field data. Using this model, the initial skeleton and boundary contour of the pipeline structure are rendered as 2D skeleton projection and contour projection maps, respectively. On the 2D imaging plane corresponding to the original geophysical field data, the original signal intensity profiles corresponding to the positions in the skeleton and contour projection maps are extracted. The difference between the original signal intensity corresponding to each projection point in the skeleton projection map and the expected reflection intensity defined in the pipeline constraint rules is calculated to obtain the skeleton intensity consistency error. The difference between the original signal intensity gradient on both sides of the projected contour in the contour projection map and the theoretical gradient model is calculated to obtain the contour gradient consistency error.

[0098] In the specific implementation, the initial skeleton and boundary contour of the pipeline structure are extracted based on the latent voxel space. Using the latent voxel space as a spatial index and Boolean mask, the corresponding sub-region semantic field is cropped from the enhanced semantic field. The sub-region semantic field contains only voxel data related to the potential pipeline target. A three-dimensional Gaussian filter is applied to the sub-region semantic field. The three-dimensional Gaussian filter uses a three-dimensional Gaussian convolution kernel to perform convolution operations on the sub-region semantic field, smoothing local irregularities and small spikes caused by data noise. A three-dimensional Sobel gradient operator is applied to the smoothed sub-region semantic field. The three-dimensional Sobel gradient operator contains three sets of convolution kernels for gradient calculation in the X, Y, and Z directions, respectively. The gradient intensity of each voxel in the three orthogonal directions is calculated, and the gradient intensities in the three directions are synthesized by taking the square root of the sum of squares to obtain the gradient magnitude field characterizing the degree of drastic change at each point within the entire sub-region semantic field.

[0099] In some embodiments, a three-dimensional level set method is used for evolution in the gradient magnitude field to capture the boundary. The three-dimensional level set method represents the interface as a zero isosurface of a high-dimensional function, and evolves the zero isosurface into the region of drastic gradient change by solving partial differential equations based on the boundary stopping function defined by the gradient magnitude field. When the evolution converges, the closed surface described by the zero isosurface forms the boundary profile of the pipeline structure. For the internal region enclosed by the boundary profile, a three-dimensional thinning algorithm is used for iterative erosion. The three-dimensional thinning algorithm is based on the topological adjacency relationship of voxels, iteratively removing surface voxels that satisfy a specific template while maintaining the topological connectivity and center position of the structure. The iterative process continues until the width of the structure is a single voxel at all locations. The single voxel connectivity path extracted at this point is the initial skeleton of the pipeline structure.

[0100] In practice, a differentiable rendering module projects the initial skeleton and boundary contour of the pipeline structure onto a two-dimensional view, and performs pixel-level consistency comparison with the original geophysical field data. This process begins by establishing a projection transformation model based on the acquisition perspective parameters of the original geophysical field data. These parameters include the spatial coordinates of the sensor array, the emission angle of the probe beam, and the geometric relationship with the imaging plane. Based on these parameters, a differentiable projection transformation model is established from three-dimensional spatial coordinates to two-dimensional imaging plane coordinates. The rotation, translation, and perspective transformation operations in the model are all implemented as operations with computable gradients. Using the differentiable projection transformation model, the coordinates of the center point of each voxel contained in the initial skeleton of the pipeline structure, and the coordinates of the vertices of the triangular facets contained in the boundary contour, are projected and transformed to generate two-dimensional skeleton projection maps and contour projection maps. The skeleton projection map consists of a series of two-dimensional point sets, and the contour projection map consists of a series of two-dimensional line segments.

[0101] It can be understood that, on the two-dimensional imaging plane corresponding to the original geophysical field data, based on the coordinates of each projection point in the skeleton projection map, the geophysical signal intensity value recorded at that coordinate location by the original geophysical field data is extracted, typically the electromagnetic wave reflection amplitude. The absolute difference between the original signal intensity corresponding to each projection point in the skeleton projection map and the expected reflection intensity defined in the pipeline constraint rules and derived from the pipeline material is calculated. The average of the differences for all projection points is then used to obtain the skeleton intensity consistency error. Simultaneously, in the contour projection map, along the normal direction of the projection contour, a series of point pairs are sampled at symmetrical positions on both sides. The signal intensity values ​​corresponding to these point pairs in the original geophysical field data are extracted, and the intensity gradient of each point pair is calculated. The calculation of the intensity gradient can be expressed as: ;

[0102] in: Indicates the first The original signal intensity gradient is calculated for each sampling point. and These represent the signal intensities of the sampling points on the outer and inner sides of the contour, respectively. This represents the sampling interval of the point pair. Then, the original signal intensity gradient for all point pairs is calculated. The theoretical gradient calculated based on pipeline physics models (such as the ideal cylinder scattering model) The mean square error between the two is used to obtain the contour gradient consistency error.

[0103] Optionally, the skeleton strength consistency error and the contour gradient consistency error can be weighted and summed. The weighting coefficients can be preset according to the response characteristics of different geophysical methods. The result of the summation is the final quantization result of the pixel-level consistency comparison. This result serves as a scalar loss value, used to measure the degree of matching between the currently inferred 3D pipeline structure and the original observation data. This loss value will be used in the subsequent network parameter optimization process, driving the pipeline existence verification network and the depth feature generation network to adjust towards generating a 3D structure that is more consistent with the original data.

[0104] See Figure 4 This is a line graph showing the gradient intensity distribution of 3D morphological operations during underground pipeline detection. The gradient magnitude (red line) is consistently higher than the gradient intensity in a single direction, reflecting the combined effect of gradients from multiple directions. Significant peaks in the gradient magnitude appear at voxel points 2.5, 7.5, and 15.0, indicating these locations are candidate regions for pipeline boundaries. The gradient intensity in the Z direction (green line) is generally higher than in the X and Y directions, possibly reflecting more significant structural changes in the pipeline in the vertical direction. This graph is used to assist in locating pipeline boundary contours: by tracking the peak positions of gradient magnitudes, potential boundary regions of the pipeline can be quickly identified, providing data for subsequent 3D level set methods to extract boundary contours.

[0105] Example 4: The process of collaboratively optimizing and adjusting the internal parameters of the pipeline existence verification network and the deep feature generation network based on the pixel-level consistency comparison results involves using the pixel-level consistency comparison results as a loss signal, which is then backpropagated to the differentiable rendering module. The loss signal is further backpropagated through the projection transformation model in the differentiable rendering module to the 3D morphological operation steps for generating the initial skeleton and boundary contours of the pipeline structure. The loss signal continues to backpropagate through the 3D morphological operation steps to the pipeline existence verification network and the deep feature generation network. Using the gradient descent algorithm, the convolutional kernel weight parameters of the pipeline existence verification network are updated based on the loss gradient propagated to the pipeline existence verification network. Simultaneously, the fully connected layer and convolutional layer parameters of the deep feature generation network are updated based on the loss gradient propagated to the deep feature generation network. During parameter updates, the gradients from the skeleton strength consistency error and the contour gradient consistency error are normalized to balance the optimization speed of the skeleton strength consistency error and the contour gradient consistency error. The process of reprocessing the original geophysical field data using an optimized network and iteratively updating the initial skeleton and boundary contours of the pipeline structure until convergence is met is recorded as one iteration. After one iteration, a depth feature generation network with updated internal parameters is used to reprocess the original geophysical field data to generate a new enhanced semantic field. A pipeline existence verification network with updated internal parameters is used to couple the new enhanced semantic field with pipeline constraint rules to generate a new latent voxel space. Based on the new latent voxel space, three-dimensional morphological operations are performed again on the new enhanced semantic field to generate updated initial skeleton and boundary contours of the pipeline structure. The structural similarity index between the updated initial skeleton and boundary contours of the pipeline structure and the initial skeleton and boundary contours of the pipeline structure generated in the previous iteration is calculated. It is determined whether the structural similarity index is greater than a preset convergence threshold. If it is greater, the convergence condition is met, and the iteration stops. If it is not greater, the next iteration begins.

[0106] In practice, the internal parameters of the pipeline existence verification network and the deep feature generation network are collaboratively optimized and adjusted based on the results of pixel-level consistency comparison. This process begins by using the scalar loss value calculated from the pixel-level consistency comparison as the loss signal. The loss signal is then propagated backward through the computational graph to the differentiable rendering module. The matrix multiplication and perspective division operations included in the projection transformation model of the differentiable rendering module are predefined as differentiable operations, allowing the loss signal to successfully calculate the gradient with respect to the 3D coordinates. The loss signal and its gradient are further backpropagated to the 3D morphological operation steps for generating the initial skeleton and boundary contours of the pipeline structure. Although the refinement and level set evolution in the 3D morphological operations are discrete processes, a gradient path is established in practice using a differentiable distance field approximation and a soft thresholding function, enabling the gradient to continue backpropagating.

[0107] In some embodiments, the loss signal, after passing through the differentiable approximation module of the 3D morphological operation step, continues to propagate back to the pipeline existence verification network and the deep feature generation network. In the pipeline existence verification network, the gradient flows through the depthwise separable convolutional layers and point convolutional layers of the lightweight convolutional network, eventually reaching the convolutional kernel weight parameters of the network. In the deep feature generation network, the gradient propagates back along the path of the decoder and encoder, passing through the feature pyramid fusion layer, attention weighting layer, and basic convolutional layers, reaching the fully connected layers and convolutional layer parameters of the deep feature generation network. A gradient descent algorithm based on adaptive moment estimation is used to update the convolutional kernel weight parameters of the pipeline existence verification network according to the loss gradient propagated to the pipeline existence verification network, with the update direction aimed at reducing the loss signal. Simultaneously, the fully connected layers and convolutional layer parameters of the deep feature generation network are updated according to the loss gradient propagated to the deep feature generation network. The general form of parameter update can be expressed as:

[0108] ;

[0109] in: This represents the network parameters to be updated (the convolutional kernel weights of the pipeline existence verification network or the relevant parameters of the depth feature generation network). and These represent the parameters before and after the update, respectively. Represents the learning rate. Represents loss signal Regarding parameters gradient, This represents the gradient normalization and momentum adjustment functions used by the adaptive moment estimator.

[0110] Optionally, when updating parameters, the gradients from the skeleton strength consistency error and the contour gradient consistency error need to be normalized. Specifically, the gradient norms of the skeleton strength consistency error component and the contour gradient consistency error component in the loss signal are calculated separately. Then, the gradient of each component is divided by its norm, and finally, the normalized gradients are re-weighted and summed using the original weight coefficients. This process balances the impact of the skeleton strength consistency error and the contour gradient consistency error on the network parameter optimization speed, preventing a single error term from dominating the optimization process.

[0111] It is understandable that the optimized network is used to reprocess the original geophysical field data and iteratively update the initial skeleton and boundary contour of the pipeline structure until the convergence condition is met. One complete process of co-optimizing and adjusting the intrinsic parameters is recorded as one iteration. After one iteration, a depth feature generation network with updated intrinsic parameters is used to reprocess the original geophysical field data, generating a new enhanced semantic field different from the previous one. A pipeline existence verification network with updated intrinsic parameters is then used to couple the new enhanced semantic field with pipeline constraint rules to generate a new latent voxel space. Based on the new latent voxel space, three-dimensional morphological operations are performed again on the new enhanced semantic field, including clipping, filtering, gradient calculation, level set evolution, and refinement, thereby generating an updated initial skeleton and boundary contour of the pipeline structure.

[0112] In practice, the structural similarity index between the updated pipeline structure's initial skeleton and boundary contour and the initial skeleton and boundary contour generated in the previous iteration is calculated. The structural similarity index is calculated by comprehensively comparing the overlap in spatial occupancy, the similarity in centerline orientation, and the consistency of surface curvature between the two 3D structures. The structural similarity index is then checked against a preset convergence threshold, which is pre-set based on the required detection accuracy. If the structural similarity index exceeds the preset convergence threshold, the convergence condition is met, the iteration stops, and the current structure is output as the final result. If it does not exceed the threshold, the next iteration begins, involving parameter optimization and structural updates again. The convergence status during the iteration process can be recorded, as shown in Table 1.

[0113] Number of iterations Structural Similarity Index (SSI) Does it meet the convergence condition (threshold = 0.98)? 1 0.75 no 2 0.89 no 3 0.94 no 4 0.97 no 5 0.985 yes

[0114] Example 5: The process of integrating the finally converged pipeline skeleton, boundary contour, and pipeline attribute description data to generate a pipeline detection result map containing 3D coordinates and attributes involves converting the voxel coordinates of the finally converged pipeline skeleton into a 3D coordinate sequence in a geodetic coordinate system through spatial coordinate transformation. Similarly, the voxel coordinates of the pipeline structure's boundary contour are converted into a 3D mesh surface in a geodetic coordinate system. Material, diameter, and burial depth attributes are extracted from the pipeline constraint rules parsed from the pipeline attribute description data and associated with the corresponding line segments in the 3D coordinate sequence. The attributed 3D coordinate sequence is defined as the pipeline's centerline vector, and the 3D mesh surface is defined as the pipeline's outer surface model. The pipeline's centerline vector, outer surface model, and associated attribute information are imported into a 3D geographic information system platform for visualization rendering to generate the pipeline detection result map.

[0115] In practice, the pipeline detection results map is generated by integrating the finally converged pipeline skeleton, boundary contour, and pipeline attribute description data. This begins with the transformation of the pipeline skeleton coordinate system. The voxel coordinates of the finally converged pipeline skeleton are converted into a three-dimensional coordinate sequence in the geodetic coordinate system through spatial coordinate transformation. The voxel coordinates exist in a local image coordinate system with a point in the detection area as the origin. The parameters of the spatial coordinate transformation originate from the positioning and attitude determination system records during geophysical field data acquisition. The transformation process includes scaling, rotation, and translation operations. The scaling operation converts the voxel unit index to the actual metric unit, the rotation operation corrects the angle between the detection area and true north, and the translation operation maps the origin of the local coordinate system to the real-world geodetic coordinates. The same spatial coordinate transformation is performed on the voxel coordinates of the pipeline structure's boundary contour, converting them into a set of polygonal patches composed of a series of three-dimensional spatial points. A continuous three-dimensional mesh surface is then generated through a surface reconstruction algorithm.

[0116] The mathematical expression for coordinate transformation is: ;

[0117] in: Represents the coordinate vector in the voxel coordinate system , It is a 3x3 diagonal scaling matrix, where the diagonal elements represent the actual physical dimensions of the voxels in the X, Y, and Z directions, respectively. It is a 3x3 rotation matrix used to align the local coordinate system with the geodetic coordinate system. It is a 3x1 translation vector, representing the coordinates of the local coordinate system origin in the geodetic coordinate system. This refers to the geodetic coordinates obtained after the transformation. This transformation applies simultaneously to the coordinates of the center point of each voxel in the pipeline skeleton and the coordinates of each vertex of the boundary profile.

[0118] In some embodiments, extracting material, diameter, and burial depth attributes from pipeline constraint rules parsed from pipeline attribute description data requires associating the textual or numerical descriptions in the rules with corresponding line segments in the three-dimensional coordinate sequence. For example, a pipeline constraint rule might include "Material: Polyethylene", "Diameter: 0.3 meters", and "Burial depth range: 1.2-1.5 meters". The system establishes a one-to-one correspondence between these attributes and the three-dimensional coordinate sequence of the pipeline skeleton passing through the spatial range defined by the rule. The association operation is implemented by assigning attribute values ​​as metadata fields to the corresponding line segment data structure in the three-dimensional coordinate sequence. Optionally, the attributed three-dimensional coordinate sequence is defined as the pipeline's centerline vector. The centerline vector not only contains a series of ordered geodetic coordinate points, but each point or line segment also includes attribute fields such as material, diameter, and estimated burial depth. The three-dimensional mesh surface is defined as the pipeline's outer surface model, stored as a set of triangular facets, with each facet associated with its corresponding pipeline material information.

[0119] The process involves importing the pipeline's centerline vector, outer surface model, and associated attribute information into a 3D geographic information system (GIS) platform for visualization and rendering. The GIS platform reads the centerline vector data and assigns different colors and line types to different pipelines based on material properties. Simultaneously, it reads the outer surface model data and renders it in semi-transparent or wireframe mode. The system automatically generates annotations next to the pipeline model based on the attribute information. These annotations can include pipeline type, diameter, and burial depth, ultimately generating a pipeline detection map that integrates 3D geometry and attribute information. This map can be rotated, scaled, displayed in layers, and its attributes queried within the GIS platform.

[0120] See Figure 5 This is a two-dimensional planar schematic diagram of underground pipeline distribution. Pipelines of different materials and diameters are intersecting and distributed in the same area, reflecting the complex layout of underground pipelines. PVC pipes (0.6m diameter) have a wider spatial coverage, while polyethylene pipes (0.2m diameter) are relatively concentrated. Multiple pipelines intersect densely around the X / Y coordinates, which is a key area for pipeline management. This diagram is used for spatial layout visualization of underground pipelines: by intuitively showing the location and distribution of different types of pipelines, it can provide spatial reference for pipeline maintenance, construction avoidance, and fault diagnosis, and is one of the core display forms of underground pipeline detection results.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish an entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting underground pipelines based on deep learning, characterized in that, include: Acquire raw geophysical field data and pipeline attribute description data for the target area; The original geophysical field data is input into a depth feature generation network to generate the corresponding initial semantic field of the underground space. The initial semantic field of the underground space is subjected to hierarchical attention-weighted fusion to form an enhanced semantic field with multi-scale contextual information; Parse the pipeline attribute description data to generate structured pipeline constraint rules; A pipeline existence verification network is used to couple the enhanced semantic field with the pipeline constraint rules to identify the potential voxel space where the pipeline exists. Based on the potential voxel space, three-dimensional morphological operations are performed on the enhanced semantic field to extract the initial skeleton and boundary contour of the pipeline structure. The initial skeleton and boundary contour are projected onto a two-dimensional view using a differentiable rendering module, and then compared with the original geophysical field data at the pixel level. Based on the results of the pixel-level consistency comparison, the internal parameters of the pipeline existence verification network and the deep feature generation network are collaboratively optimized and adjusted. The original geophysical field data is reprocessed using the optimized network, and the initial skeleton and boundary contour of the pipeline structure are iteratively updated until the convergence condition is met. By integrating the finally converged pipeline skeleton, boundary contour, and pipeline attribute description data, a pipeline detection result map containing three-dimensional coordinates and attributes is generated.

2. The deep learning-based underground pipeline detection method according to claim 1, characterized in that, The initial semantic field of the underground space is subjected to hierarchical attention-weighted fusion to form an enhanced semantic field with multi-scale contextual information, including: Extract multiple intermediate layer feature maps corresponding to different depths from the deep feature generation network; For each of the intermediate layer feature maps, channel attention mechanism and spatial attention mechanism are applied respectively, and the corresponding channel weight matrix and spatial weight matrix are calculated. The channel weight matrix is ​​multiplied element-wise by the spatial weight matrix to obtain the attention mask corresponding to each intermediate layer feature map; The attention mask is used to perform weighted modulation on the corresponding intermediate layer feature map to obtain multiple weighted feature representations; A feature pyramid structure is adopted to perform top-down and horizontal connection operations on the multiple weighted feature representations to achieve multi-scale feature fusion; The fused feature maps are batch normalized and nonlinearly activated to output the enhanced semantic field.

3. The deep learning-based underground pipeline detection method according to claim 2, characterized in that, Parse the pipeline attribute description data to generate structured pipeline constraint rules, including: The natural language description in the pipeline attribute description data is segmented into independent semantic segments describing the material, pipe diameter, burial depth, and orientation. Each of the independent semantic segments is converted into a high-dimensional word vector representation sequence; The word vector representation sequence is input into a bidirectional recurrent neural network to obtain the context encoding of each semantic segment; The context encoding is input into the sequence labeling model to label the key entities and corresponding values ​​that describe the physical properties of the pipeline. Based on the marked key entities and their corresponding values, the pipeline constraint rules are instantiated according to a predefined syntax template to generate computer-executable rules; the pipeline constraint rules include the definition of pipeline reflection intensity threshold, geometric continuity, and spatial location range.

4. The deep learning-based underground pipeline detection method according to claim 3, characterized in that, A pipeline existence verification network is used to couple the enhanced semantic field with the pipeline constraint rules for reasoning, identifying the potential voxel space where pipelines exist, including: The enhanced semantic field is sliced ​​in the depth dimension to generate a series of two-dimensional feature slices; Each of the two-dimensional feature slices is spliced ​​together with the part of the pipeline constraint rule concerning reflection intensity and geometry in the feature channel dimension. The concatenated features are input into a lightweight convolutional network, which outputs the probability that each spatial location belongs to the pipeline region. The probabilities of all depth slices are stacked to reconstruct a three-dimensional probability field with the same spatial dimensions as the original geophysical field data. In the three-dimensional probability field, based on the definition of spatial location range in the pipeline constraint rules, a dynamic probability threshold is set to perform binarization segmentation of the probability field; The requirements for geometric continuity in the pipeline constraint rules are applied to the binarized segmentation results to perform connected component analysis and screening, and the potential voxel space is output.

5. The deep learning-based underground pipeline detection method according to claim 4, characterized in that, Based on the latent voxel space, three-dimensional morphological operations are performed on the enhanced semantic field to extract the initial skeleton and boundary contour of the pipeline structure, including: Using the potential voxel space as a mask, the corresponding sub-region semantic field is clipped from the enhanced semantic field; A three-dimensional Gaussian filter is applied to the semantic field of the sub-region to smooth the local irregularities caused by data noise. The three-dimensional Sobel gradient operator is applied to the smoothed sub-region semantic field to calculate the gradient intensity of each voxel in three orthogonal directions; By synthesizing the gradient intensities in the three directions, the gradient magnitude field of the semantic field of the entire sub-region is obtained; The three-dimensional level set method is used to evolve in the gradient magnitude field to capture the interface with drastic gradient changes and form the boundary profile of the pipeline structure. The internal region enclosed by the boundary contour is iteratively eroded using a three-dimensional thinning algorithm until the structural width is a single voxel, thereby extracting the initial skeleton of the pipeline structure.

6. The deep learning-based underground pipeline detection method according to claim 5, characterized in that, The initial skeleton and boundary contour are projected onto a two-dimensional view using a differentiable rendering module, and a pixel-level consistency comparison is performed with the original geophysical field data, including: Based on the acquisition perspective parameters of the original geophysical field data, a differentiable projection transformation model from three-dimensional space to two-dimensional imaging plane is established. Using the differentiable projection transformation model, the initial skeleton and boundary contour of the pipeline structure are rendered as two-dimensional skeleton projection map and contour projection map, respectively. On the two-dimensional imaging plane corresponding to the original geophysical field data, the original signal intensity profile corresponding to the positions of the skeleton projection map and the contour projection map is extracted; The difference between the original signal strength corresponding to each projection point in the skeleton projection map and the expected reflection intensity defined in the pipeline constraint rules is calculated to obtain the skeleton strength consistency error. The difference between the original signal intensity gradient on both sides of the projected contour in the contour projection map and the theoretical gradient model is calculated to obtain the contour gradient consistency error. The weighted summation of the skeleton strength consistency error and the contour gradient consistency error yields the pixel-level consistency comparison result.

7. The deep learning-based underground pipeline detection method according to claim 6, characterized in that, Based on the results of the pixel-level consistency comparison, the internal parameters of the pipeline existence verification network and the deep feature generation network are collaboratively optimized and adjusted, including: The result of the pixel-level consistency comparison is used as a loss signal and transmitted in reverse to the differentiable rendering module. The loss signal is further backpropagated through the projection transformation model in the differentiable rendering module to the three-dimensional morphological operation steps that generate the initial skeleton and boundary contour of the pipeline structure. The loss signal continues to propagate back to the pipeline existence verification network and the depth feature generation network through the three-dimensional morphological operation steps; The gradient descent algorithm is used to update the convolutional kernel weight parameters of the pipeline existence verification network based on the loss gradient propagated to the pipeline existence verification network. Simultaneously, based on the loss gradient propagated to the deep feature generation network, the parameters of the fully connected layers and convolutional layers of the deep feature generation network are updated; When updating parameters, the gradients from the skeleton strength consistency error and the contour gradient consistency error are normalized to balance the optimization speed of the skeleton strength consistency error and the contour gradient consistency error.

8. The deep learning-based underground pipeline detection method according to claim 7, characterized in that, The original geophysical field data is reprocessed using the optimized network to iteratively update the initial skeleton and boundary contour of the pipeline structure until convergence conditions are met, including: One complete process of co-optimizing and adjusting internal parameters is recorded as one iteration. After one iteration, the original geophysical field data is reprocessed using the depth feature generation network with updated internal parameters to generate a new enhanced semantic field. Using the pipeline existence verification network with updated internal parameters, the new enhanced semantic field is coupled with the pipeline constraint rules to generate a new latent voxel space. Based on the new potential voxel space, three-dimensional morphological operations are performed again on the new enhanced semantic field to generate the initial skeleton and boundary contour of the updated pipeline structure. Calculate the structural similarity index between the initial skeleton and boundary profile of the updated pipeline structure and the initial skeleton and boundary profile of the pipeline structure generated in the previous iteration; Determine whether the structural similarity index is greater than a preset convergence threshold. If it is greater, the convergence condition is met and the iteration stops; if it is not greater, the next iteration begins.

9. The deep learning-based underground pipeline detection method according to claim 1, characterized in that, Obtain the raw geophysical field data and pipeline attribute description data of the target area, including: A sensor array is deployed to perform a wide-area scan of the target area, collecting raw field data including electromagnetic wave reflection signals and seismic wave signals; The original field data is preprocessed by time synchronization alignment, background noise suppression, and amplitude recovery to obtain standardized original geophysical field data; Retrieve textual and graphical data related to the target area from urban planning archives or pipeline as-built drawings; Optical character recognition and graphic symbol recognition are performed on the text and chart data to extract descriptive text and parameter tables about the pipeline, which are used as pipeline attribute description data.

10. The deep learning-based underground pipeline detection method according to claim 1, characterized in that, Integrate the finally converged pipeline skeleton, boundary contour, and pipeline attribute description data to generate a pipeline detection result map containing three-dimensional coordinates and attributes, including: The voxel coordinates of the finally converged pipeline skeleton are transformed into a three-dimensional coordinate sequence in the geodetic coordinate system through spatial coordinate transformation. The voxel coordinates of the boundary contour of the pipeline structure are also converted into a three-dimensional mesh surface in the geodetic coordinate system. From the pipeline constraint rules parsed from the pipeline attribute description data, material, pipe diameter, and burial depth attributes are extracted and associated with the corresponding line segments in the three-dimensional coordinate sequence; The attributed three-dimensional coordinate sequence is defined as the centerline vector of the pipeline, and the three-dimensional mesh surface is defined as the outer surface model of the pipeline. The centerline vector of the pipeline, the outer surface model, and the associated attribute information are imported into a 3D geographic information system platform for visualization rendering to generate the pipeline detection result map.

Images