Underground pipe network micro-motion scattering imaging method and system based on deep learning and medium

Abstract

The invention discloses an underground pipe network micro-motion scattering imaging method and system based on deep learning and a medium, and relates to the technical field of geophysical exploration and urban underground space safety detection. Performing dynamic signal-to-noise ratio superposition on the initial pure scattered wave record and the background micro-motion record fragment to construct a semi-synthetic training data set; constructing a U-Net convolutional neural network, and training by using the semi-synthetic training data set to obtain a scattered wave feature extraction model; preprocessing an actually measured micro-motion record, inputting the preprocessed actually measured micro-motion record into the trained scattered wave feature extraction model, and outputting a predicted pure scattered wave field; migration imaging is carried out based on the predicted pure scattered wave field and a background speed model, and a high-resolution space image is generated; according to the method, intelligent separation and imaging of weak scattering signals in passive source micro-motion records are achieved, and high-resolution space images of urban underground pipe networks and diseases are obtained.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration and urban underground space safety detection technology, and in particular to a method, system and medium for micro-motion scattering imaging of underground pipe networks based on deep learning. Background Technology

[0002] Seismic exploration technology originated in the oil and gas industry. Early mainstream techniques included reflection wave methods, which utilized Snell's Law to detect continuous subsurface stratigraphic interfaces (such as oil and gas reservoir caprocks). However, as exploration targets shifted from structural features to detailed analysis, seismic scattering imaging, based on Huygens' principle, emerged. Scattering or diffraction occurs only when subsurface inhomogeneities (such as faults, fractures, caverns, and pipelines) exist, and their scale is smaller than or close to the seismic wavelength.

[0003] Current research in this field mainly focuses on three areas: environmental noise imaging (passive sources), small-scale scatterer detection, and deep learning-based seismic signal processing. Although significant progress has been made in each branch, there is still a clear technological gap in the fusion application for targeting ultra-shallow (0-50m) micro-pipeline targets in urban areas.

[0004] ① Passive source detection technology based on micro-motion; Current mainstream micro-motion techniques (SPAC / HVSR) are essentially based on the layered medium assumption, and their main output is a one-dimensional or two-dimensional velocity profile. They cannot focus, and for small-scale discrete bodies (diffractors) such as pipelines and cavities, the SPAC algorithm smooths out the resulting scattering effects as high-frequency perturbations during the statistical process.

[0005] ② Seismic scattered wave / diffraction wave imaging technology; Existing scattering / diffraction imaging techniques heavily rely on active source data with high signal-to-noise ratios (such as explosives or falling vibration sources). In areas where vibrations are strictly prohibited and roads are closed, such as urban main roads and airports, active sources cannot be used, rendering this technology impractical. Furthermore, when directly applied to passive source data, traditional mathematical methods are unable to extract effective scattering signals due to the randomness and weakness of traffic noise sources.

[0006] ③ The current status of deep learning applications in seismic signal processing; Currently, most artificial intelligence research focuses on active source reflection seismic data (oil industry) or natural earthquake data (earthquake prevention and disaster reduction). Research on the specific scenario of extracting ultra-shallow scattered waves in urban passive source (micro-motion) environments is extremely limited. Existing AI models are mostly trained based on regular reflected waves and cannot recognize the chaotic hyperbolic scattering characteristics in micro-motion records. Furthermore, the lack of publicly available passive source seismic scattering datasets specifically addressing urban underground pipe network defects further restricts development in this direction.

[0007] In summary, there is currently no mature technical solution, either domestically or internationally, that can simultaneously meet the three core requirements of zero interference (passive source), high resolution (scattering imaging), and strong noise resistance (AI extraction). Summary of the Invention

[0008] Based on the technical problems existing in the background technology, this invention proposes a method, system and medium for micro-motion scattering imaging of underground pipe networks based on deep learning, which realizes intelligent separation and imaging of weak scattering signals in passive source micro-motion recordings, and obtains high-resolution spatial images of urban underground pipe networks and their defects.

[0009] The deep learning-based micro-motion scattering imaging method for underground pipe networks proposed in this invention includes: Continuous collection and segmentation of urban environmental background micro-motion records; generation of initial pure scattered wave records containing underground scatterers based on wave equation forward modeling, and dynamic signal-to-noise ratio superposition of background micro-motion record segments to construct a semi-synthetic training dataset; A U-Net convolutional neural network with embedded residual learning units and attention gating mechanism is constructed and trained using the semi-synthetic training dataset to obtain a scattered wave feature extraction model for separating weak scattered wave signals. The measured micro-motion records are preprocessed and then input into the trained scattered wave feature extraction model. The pure scattered wave field is predicted by nonlinear mapping. Based on the predicted pure scattered wave field, and combined with the background velocity model obtained by inversion from the urban environmental background micro-motion record, a high-resolution spatial image of the underground pipe network and its defects is generated by migration imaging.

[0010] Furthermore, by utilizing a gravity-coupled, rollable geophone array deployed on the hardened road surface, continuous data collection of urban environmental background micro-motion records was acquired. The gravity-coupled crushable detector array includes multiple improved detectors, and adjacent improved detectors are connected in series by flat Kevlar cables to form a crushable drag cable. Each improved detector includes a pressure-bearing top cover, an aviation connector, a housing, and a base. The pressure-bearing top cover is located on the upper part of the housing, and the base is located on the lower part of the housing. The pressure-bearing top cover, housing, and base are assembled into a sealed housing with waterproof and dustproof functions. A three-component detector is encapsulated inside the sealed housing. The aviation connector passes through the housing and connects to the three-component detector inside for power supply and data transmission. The main body of the outer shell adopts a trapezoidal pressure-resistant structure, designed as a streamlined or flat low-profile wedge structure; the base is a high-density metal counterweight, which achieves gravity coupling with the road surface.

[0011] Furthermore, the generation process of the initial pure scattered wave record is as follows: Establish an underground geological model that includes scatterers of circular, rectangular, or irregular shapes; The total wavefield is obtained by performing forward modeling on a subsurface geological model containing scatterers; The background wavefield was obtained by performing a parametric forward modeling on an underground geological model after removing scatterers. The difference between the total wave field and the background wave field is calculated, and the difference is used as the initial pure scattered wave record. The initial pure scattered wave record contains only hyperbolic texture features generated by the pipeline, which is used as the target label of the U-Net convolutional neural network.

[0012] Furthermore, the semi-synthetic training dataset The construction formula is as follows: ; in, Records of the initial, pure scattered waves. Recording snippets of subtle movements against the backdrop of the urban environment. This is the noise adjustment factor.

[0013] Furthermore, an attention gating module is embedded at the skip connection between the encoder and decoder of the U-Net convolutional neural network to automatically suppress instantaneous high-energy pulse interference caused by vehicle rolling and to enhance the response to hyperbolic texture features. Residual learning units are introduced into the convolutional layers of the encoder to learn the residual mapping between the semi-synthetic training dataset and the initial pure scattered wave record.

[0014] Furthermore, the combined loss function of the U-Net convolutional neural network during training... The settings are as follows: : in, The predicted pure scattering wave field is the output of the U-Net convolutional neural network after the semi-synthetic training dataset is input into it; The initial pure scattered wave record, which serves as the target label, corresponds to the semi-synthetic training dataset. for and The L1 norm loss between them is used to constrain the difference in amplitude between the predicted pure scattered wave record output by the U-Net convolutional neural network and the initial pure scattered wave record used as the target label. for and A multi-scale structural similarity loss between the two is used to constrain their structural differences in hyperbolic texture features. These are the weighting coefficients.

[0015] Furthermore, the preprocessing procedure for the measured micro-motion records is as follows: After performing amplitude equalization or truncation on the measured micro-motion records, the records are extracted into segments according to a set time window. Standardization processing: Z-Score standardization or maximum value normalization is performed on each measured micro-motion recording segment to make its amplitude distribution consistent with the distribution characteristics of the training samples in the semi-synthetic training dataset; The standardized measured micro-motion recording fragments were used as input to the trained scattered wave feature extraction model.

[0016] Furthermore, the migration imaging employs Kirchhoff integral migration, reverse time migration, beamforming, least squares migration, or full waveform inversion algorithms, taking the predicted pure scattered wave field as input, calculating the travel time based on the background velocity model, and backtracking the scattered energy at each time point to its spatial location of origin. Scattered energy from the same pipeline is focused and returned to the true spatial location of the underground anomaly.

[0017] A computer system includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program in the manner described above.

[0018] A computer-readable storage medium having stored a plurality of computer programs thereon, the plurality of computer programs being invoked by a processor and executing the method described above.

[0019] The advantages of the deep learning-based method, system, and medium for micro-motion scattering imaging of underground pipe networks provided by this invention are as follows: by constructing a U-Net convolutional neural network and utilizing the powerful nonlinear feature extraction capability of deep learning, the hyperbolic texture features of scattered waves in the spatiotemporal domain are learned, and weak scattering signals are intelligently identified and recovered, thereby realizing intelligent separation and imaging of weak scattering signals in passive source micro-motion recordings, and obtaining high-resolution spatial images of urban underground pipe networks and their defects. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the U-Net convolutional neural network structure; Figure 3 Schematic diagram for constructing a semi-synthetic training dataset; Figure 4 This is a schematic cross-sectional view of a rollable passive source observation system. Figure 5 This is a planar schematic diagram of a rollable passive source observation system; Figure 6 The images are comparisons of the imaging results. (a) shows the actual measured micro-motion record, (b) shows the predicted pure scattered wave field output by the U-Net convolutional neural network, and (c) shows the high-resolution spatial result of the predicted pure scattered wave field after migration imaging. Figure 7 This is a schematic diagram of the improved detector in this embodiment; Figure 8 This is a schematic diagram of the structure for introducing residual blocks into the convolutional layer of an encoder. Among them, 1-pressure-bearing top cover, 2-aviation plug interface, 3-outer shell, 4-base. Detailed Implementation

[0021] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] To understand this embodiment, the following terms are defined: Urban environmental background micro-motion records refer to passive source seismic records continuously acquired using geophone arrays deployed on paved roads in urban traffic environments, without relying on artificial seismic sources. These records include various urban environmental noises such as vehicle movement, pedestrian activity, and mechanical vibrations, as well as the response signals of the underground medium to these noises.

[0023] Actual micro-motion recording: refers to the raw passive source seismic data collected in actual engineering sites using the gravity-coupled rollable geophone array described in this invention, as input data for the scattered wave feature extraction model.

[0024] Initial pure scattered wave record: refers to a seismic record obtained through forward modeling of the wave equation that contains only scattered wave components generated by underground scatterers (such as pipelines, cavities, etc.). This record contains no environmental noise and only presents hyperbolic texture features generated by the scatterers, serving as the target label in the neural network training process.

[0025] Predicted pure scattered wave field: This refers to the seismic wave field output by a pre-trained scattered wave feature extraction model after inputting measured micromotion records or semi-synthetic training data into the model. Environmental noise has been suppressed or filtered out in this wave field, mainly retaining the scattered wave energy associated with underground scatterers.

[0026] like Figures 1 to 8 As shown, the deep learning-based micro-motion scattering imaging method for underground pipe networks proposed in this invention includes: Step 1: Continuously collect background micro-motion records of the urban environment and extract them into background micro-motion record segments according to a set time window. Generate initial pure scattered wave records containing underground scatterers based on forward modeling of the wave equation, and perform dynamic signal-to-noise ratio superposition with the background micro-motion record segments to construct a semi-synthetic training dataset. Step 2: Construct a U-Net convolutional neural network with embedded residual learning units and attention gating mechanism, and train it using the semi-synthetic training dataset to obtain a scattered wave feature extraction model for separating weak scattered wave signals. Step 3: After preprocessing the measured micro-motion records, input them into the trained scattered wave feature extraction model, and output the predicted pure scattered wave field through nonlinear mapping; Step 4: Based on the predicted pure scattered wave field, and combined with the background velocity model obtained by inversion of urban environmental background micro-motion records, perform migration imaging to generate high-resolution spatial images of underground pipe networks and defects.

[0027] This embodiment constructs a U-Net convolutional neural network and utilizes the powerful nonlinear feature extraction capability of deep learning to learn the hyperbolic texture features of scattered waves in the spatiotemporal domain. It intelligently identifies and recovers weak scattered signals, thereby achieving intelligent separation and imaging of weak scattered signals in passive source micro-motion recordings, and obtaining high-resolution spatial images of urban underground pipe networks and defects.

[0028] In one embodiment, step one involves continuously collecting background motion records of the urban environment and segmenting them into background motion record segments according to a set time window. An initial clean scattered wave record containing underground scatterers is generated based on forward modeling of the wave equation, and dynamically superimposed with the background motion record segments to construct a semi-synthetic training dataset. Specifically: Unlike traditional micro-motion detection (large array, sparse acquisition), this embodiment must adopt a small-point-spacing, high-frequency acquisition strategy to capture small-scale scattering / diffraction signals. For special working conditions such as busy urban traffic, where it is impossible to drive into the ground or block traffic, this embodiment designs and employs a non-contact, gravity-coupled linear observation system, as shown in (a1) to (a3).

[0029] (a1) Rollable passive source observation system (gravity-coupled rollable detector array). Existing detectors are mostly single-point pin type or individually packaged, which cannot be directly deployed on urban hardened roads for long-term, high-density passive source acquisition.

[0030] The detector array in this embodiment consists of multiple improved detectors. Adjacent improved detectors are connected in series via flat Kevlar cables to form a runaway cable. This design allows vehicles to run directly over the measuring line without damaging the instrument or affecting driving safety, achieving "zero interference" deployment.

[0031] The specific structure of the improved detector in this embodiment is as follows: Figure 7 As shown, the pressure-bearing top cover 1 is a high-strength pressure-bearing structure. The pressure-bearing top cover 1 integrates a status indicator light, which is linked with the internal three-component detector signal. This is used to provide instantaneous visual feedback when the improved detector detects an effective vibration signal generated by vehicle rolling, so as to assist on-site operators in confirming the coupling status and data acquisition status of the detector in real time, and to provide on-site operators with immediate and visual data quality monitoring feedback.

[0032] The aviation connector 2 passes through the housing 3 and connects to the internal three-component detector for power supply and data transmission. The aviation connector 2 is an IP68 waterproof aviation connector, and its connection with the housing 3 is waterproofed by a sealing gasket or thread sealant.

[0033] The main body of the outer shell 3 adopts a trapezoidal compression-resistant structure and is designed as a streamlined or flat low-profile wedge structure (height less than 5cm). The base 4 is a high-density metal (such as steel or lead alloy) counterweight to achieve good gravity coupling with the hardened road surface (asphalt or concrete) without the need for coupling agent or pins.

[0034] The pressure-bearing top cover 1 is located on the upper part of the outer shell 3, and the base 4 is located on the lower part of the outer shell 3. The pressure-bearing top cover 1, the outer shell 3, and the base 4 are assembled by fasteners, and sealing rings or sealant are provided at each connection point to form a sealed shell with waterproof and dustproof functions. The overall protection level reaches IP68. The sealed shell encapsulates a three-component detector (an existing device). The sealed shell is equipped with a mounting bracket, and the three-component detector is fixed to the mounting bracket by shock-absorbing material; or, the sealed shell is filled with a potting layer, and the three-component detector is embedded in the potting layer; to ensure stable coupling even under the vibration environment generated by vehicle rolling.

[0035] The sealed housing also integrates an analog-to-digital converter circuit and a data transmission circuit. The analog-to-digital converter circuit has a sampling accuracy of no less than 24 bits, and the data transmission circuit adopts RS485 or Ethernet protocol to convert the analog signal collected by the three-component detector into a digital signal and transmit it to the outside through the aviation plug interface 2.

[0036] (a2) Setting observation parameters; Layout method: Lay the above-mentioned rollable tow cable directly along the edge of the road or the center line of the lane.

[0037] Channel spacing: In order to satisfy the spatial sampling theorem and prevent high-frequency scattered waves from generating spatial aliasing, the channel spacing of the improved detector can be set to 0.5m to 2.0m (depending on the depth of the target being detected; the shallower the target, the smaller the spacing).

[0038] Sampling rate: can be set to 2ms or 4ms (i.e. 500Hz or 250Hz) to retain high-frequency valid information; Data acquisition duration: 15 to 30 minutes (adjustable) of continuous data acquisition at each measurement point to obtain background micro-motion records of the urban environment. No artificial seismic source is required; it directly records environmental micro-motions generated by vehicle movement and pedestrian activity, while utilizing broadband vibrations generated by vehicles directly running over cables as a near-field excitation source.

[0039] (a3) Data preprocessing: Since direct vehicle crushing generates instantaneous high energy (large amplitude), amplitude equalization (AGC) or truncation processing is performed before segmenting the urban environmental background micro-motion recording data. Continuous recordings are then truncated into urban environmental background micro-motion recording segments according to time windows (e.g., 4 seconds) to establish the original dataset.

[0040] In actual urban engineering projects, it is difficult to obtain a large number of accurately labeled (i.e., the exact location of the pipelines is known and the signal is clean) underground pipeline scattering seismic records. This embodiment adopts a data construction strategy of forward modeling signal + real environmental noise to generate paired training samples, based on which a semi-synthetic training dataset is constructed, including (b1) to (b4).

[0041] (b1) Establish multi-morphological underground geological models; Background medium setting: Based on the characteristics of shallow urban surfaces, establish a homogeneous or layered medium model with a velocity range of 200m / s-800m / s.

[0042] Scatterer Implantation: Discrete scatterers with different properties are randomly implanted at various locations within the aforementioned background medium model (i.e., homogeneous or layered medium model) to construct a complete underground geological model for simulating underground pipe networks and defects. The parameters of the scatterer are as follows: Diverse geometric shapes: including circles (simulating various types of pipelines), rectangles (simulating box culverts and pipe corridors), and irregular shapes (simulating earthen caves and voids); Scale and depth: The diameter / side length of the scatterer can be set from 0.5m to 3.0m, and the burial depth can be set from 2m to 20m; Physical property filling: The interior of the scatterer is configured with low-velocity / low-density or high-velocity / high-density anomalies to generate scattered waves of different polarities; specifically: (1) When simulating low-speed / low-density anomalies such as empty pipes and soil caves, the interior can be filled with air (e.g., the longitudinal wave velocity Vp range can be set to 330m / s-350m / s, and the density can be set to 1.2kg / m³-1.3kg / m³) or filled with water (e.g., the longitudinal wave velocity Vp range can be set to 1450m / s-1550m / s, and the density can be set to 1000kg / m³). (2) When simulating high-speed / high-density anomalies such as abandoned concrete pipes or underground boulders after filling, their physical property parameters can be set to be higher (for example, the longitudinal wave velocity Vp range can be set to 2500m / s-6000m / s, and the density can be set to 2200kg / m³-2800kg / m³).

[0043] (b2) Forward modeling of pure scattered wave field based on wave equation; The equations for two-dimensional acoustic or elastic waves are solved using the high-order staggered grid finite difference method. To simulate the plane wave characteristics of micro-movement fields, randomly distributed weak seismic sources or plane wave sources are set up at the surface or deep underground.

[0044] Wave field separation for tag acquisition: First, a forward modeling simulation was performed on an underground geological model containing scatterers to obtain the total wavefield. ; Secondly, a parametric forward modeling was performed on the underground geological model after removing the scatterers to obtain the background wavefield. ; By subtraction Extract the initial pure scattered wave record The initial pure scattered wave record contains only hyperbolic texture features generated by the pipeline, which serve as the target label for the U-Net convolutional neural network.

[0045] (b3) Injection and fusion of micro-motion recording segments of urban environmental background; The set of urban environmental background micro-motion recording segments extracted by time window after preprocessing in step (a3) ​​contains instantaneous strong pulses from vehicle crushing, continuous traffic vibrations, and electrical noise from the instrument itself, and has the most realistic noise fingerprint.

[0046] like Figure 3 As shown, the initial pure scattered wave is recorded as... Recording with background micro-motion (i.e., real environmental noise) are linearly superimposed to generate a semi-synthetic training dataset, denoted as . : .

[0047] Dynamic signal-to-noise ratio adjustment: This represents the noise adjustment coefficient. To improve the adaptability of the U-Net convolutional neural network to extreme environments, it is dynamically adjusted during the generation process. This results in the signal-to-noise ratio (SNR) of the semi-synthetic training dataset being distributed between -20dB and -5dB (i.e., the noise energy is much greater than the signal energy), forcing the U-Net convolutional neural network to extract features under extremely low SNR conditions.

[0048] (b4) Data augmentation; For the generated paired samples Random time shifts, gather flips (simulating different detection directions), random slicing, and amplitude scaling are performed. Finally, a semi-synthetic training dataset containing tens of thousands of sample pairs is constructed.

[0049] This embodiment uses a semi-synthetic training dataset construction scheme based on physical forward modeling and real noise injection. Instead of directly using the extremely scarce real pipeline labeling data, it generates initial pure scattered wave records by solving the wave equation in forward modeling, which serve as the target label. These records are then superimposed with micro-motion records of the urban environment (including instantaneous vehicle pulses and random interference) collected from the field to generate the training input data. This solves the problem of scarce labeled data in the engineering field. Simultaneously, by injecting real noise fingerprints (i.e., urban environmental background micro-motion records), the U-Net convolutional neural network adapts to the complex interference environment of the city during the training phase, greatly improving the generalization ability and robustness of the scattered wave feature extraction model on measured micro-motion records.

[0050] In one embodiment, step two involves constructing a U-Net convolutional neural network embedding residual learning units and an attention gating mechanism, training it using the semi-synthetic training dataset, and obtaining a scattered wave feature extraction model for separating weak scattered wave signals; specifically: This step aims to build a high-precision U-Net convolutional neural network, which is trained using the semi-synthetic training dataset generated in step one, enabling it to intelligently identify and separate weak scattered wave signals from actual micro-motion records with extremely low signal-to-noise ratios (Low SNR).

[0051] (c1) Network architecture design: Design a U-Net convolutional neural network with embedded multi-scale attention mechanism; Backbone architecture: Employs a classic encoder-decoder U-shaped structure. The encoder (shrinkage path) consists of 4-5 convolutional layers. Each convolutional layer includes convolution (Conv2D), batch normalization, Leaky ReLU activation, and max pooling. Its function is to compress the data size layer by layer and extract abstract features of seismic waves at different spatiotemporal scales.

[0052] Decoder (expansion path): Consists of corresponding up-sampling and transposed convolutional layers, used to restore the spatiotemporal resolution of the data.

[0053] This embodiment introduces a residual learning unit (ResidualBlock) and an attention gate based on the encoder-decoder structure described above.

[0054] Residual learning units: Residual blocks (ResNet Blocks) are introduced into the convolutional layers of the encoder. Specifically, such as... Figure 8 As shown, the residual block contains a main path and a shortcut branch. The main path includes, in sequence, a first convolutional layer (Conv2D), a first batch normalization layer (Batch Normalization), an activation function layer (Leaky ReLU), a second convolutional layer, and a second batch normalization layer. The shortcut branch directly passes the input feature map of the residual block across layers, adds it element-wise with the output of the second batch normalization layer in the main path, and then outputs it through the activation function.

[0055] By introducing this residual learning unit, the U-Net convolutional neural network no longer directly learns complex scattering waveforms under extremely low signal-to-noise ratios, but instead learns the residual mapping during the feature map transition process. This design effectively prevents gradient vanishing in deep networks and significantly improves the ability to capture weak local features such as scattering signals from tiny pipelines.

[0056] The U-Net convolutional neural network embeds an attention gating module at the skip connections between the encoder and decoder to automatically suppress transient high-energy pulse interference caused by vehicle rolling and enhance the response to hyperbolic texture features. Specifically, the attention gating module automatically calculates the weight distribution of the feature map. Regions containing hyperbolic texture features are assigned high weights, while regions containing irregular vehicle noise or random white noise are assigned zero weights. For the transient high-energy pulse interference caused by vehicle rolling mentioned in step one, the attention mechanism plays a crucial role in visual focusing, suppressing the destructive effect of high-energy interference on imaging.

[0057] Specifically, the basic computational logic of the Attention Gate (AG) module in this embodiment references the existing Attention U-Net architecture, and adaptive improvements have been made for the micro-motion recording of noisy urban environments in this embodiment. The specific soft attention weight generation process is as follows: A feature map from the deep layers of the decoder, rich in global context information, is used as a gating signal and added element-wise with a feature map from the same level of skip connections in the encoder (containing high-frequency spatial details but mixed with a large amount of environmental noise); then, features are fused sequentially through 1×1 convolution and the ReLU activation function, and then the Sigmoid activation function is used to non-linearly compress the feature response value of each pixel to between 0 and 1, thereby generating a soft attention weight matrix; finally, this soft attention weight matrix is ​​multiplied element-wise with the original skip connection feature map.

[0058] This embodiment differs from conventional medical image segmentation. In this embodiment, the attention gating mechanism works in conjunction with the encoder's residual learning unit. The residual learning unit ensures that the features of weak scattered signals in the deep network are preserved, while the attention gating module selectively enhances and suppresses the feature map. Together, they significantly improve the model's ability to extract scattered waves at extremely low signal-to-noise ratios. The attention gating mechanism and the encoder's residual learning unit are specifically driven to learn the coherence of the wave field in the spatiotemporal domain. Under this mechanism, the soft attention weight matrix output by the sigmoid activation function of the attention gating module has a clear geophysical meaning: for hyperbolic texture feature regions with strong coherence, the network learns to assign them high weights approaching 1; while for random white noise, especially transient high-energy pulse interference lacking hyperbolic spatial coherence caused by vehicle rolling, the sigmoid activation function forces convergence to produce extremely low weights approaching 0, thus physically blocking the transmission of high-energy noise to the decoder before feature stitching, achieving extreme focusing on weak effective scattered energy.

[0059] In this embodiment, an attention gating mechanism is embedded at the skip connections in the classic U-Net architecture, combined with residual learning units. For instantaneous high-energy pulse interference caused by vehicles directly running over people on urban roads, the attention gating mechanism can automatically learn and suppress these non-hyperbolic strong noises, forcing the network to focus on weak scattered wave textures. The residual learning units effectively prevent the gradient vanishing of weak signals in deep networks, significantly improving the signal-to-noise ratio (SNR) after denoising.

[0060] (c2) Construction of the loss function; To balance the accuracy of waveform values ​​and the integrity of geometric structure, this embodiment abandons the single mean square error (MSE) and adopts a combined loss function. : ; Among them, among them, The predicted pure scattering wave field is the output of the U-Net convolutional neural network after the semi-synthetic training dataset is input into it; The initial pure scattered wave record, which serves as the target label, corresponds to the semi-synthetic training dataset; for and L1 norm loss between; for and Multi-scale structural similarity loss between them These are weighting coefficients; a value of 1 is recommended for all of them.

[0061] For multi-scale structural similarity loss: seismic scattered waves mainly exhibit texture features (hyperbole). The loss focuses on the brightness, contrast, and structural information of the image, which strongly constrains the network to recover a clear and continuous hyperbolic trajectory, avoiding image blurring.

[0062] (c3) Training and optimization of U-Net convolutional neural network; Input / Output: The input is the single-channel or three-channel (time-window slice) semi-synthetic training dataset generated in step one. The output is the corresponding predicted value of the pure scattered wave record. .

[0063] Optimizer: The Adam optimizer is used, with an initial learning rate set to... Furthermore, it employs a cosine annealing strategy to dynamically adjust the learning rate in order to avoid getting trapped in local optima.

[0064] Training strategy: Set the batch size to 32 or 64 (configurable) and perform iterative training on a GPU workstation. When the loss function on the validation set... When the value stops decreasing and tends to stabilize, stop training, save the model parameter file, and you will obtain the final scattered wave feature extraction model.

[0065] In one embodiment, step three involves inputting the preprocessed measured micro-motion record into the trained scattering wave feature extraction model, and outputting a predicted pure scattering wave field through nonlinear mapping; specifically: The measured micro-motion records were collected using the gravity-coupled crushable detector array from step one and input into the trained scattered wave feature extraction model. The model's nonlinear mapping capability was used to extract the weak scattered wave signal from the strong environmental noise.

[0066] (d1) Standardization processing of measured data; Measured micro-motion records were acquired using a gravity-coupled crushable detector array. After performing amplitude equalization or truncation on the measured micro-motion records, the records are extracted into segments according to a set time window. Standardization processing: Z-Score standardization or maximum value normalization is performed on each measured micro-motion recording segment to make its amplitude distribution consistent with the distribution characteristics of the training samples in the semi-synthetic training dataset.

[0067] The advantages of standardizing the measured data are: it effectively eliminates the difference in absolute physical dimensions between the semi-synthetic training data generated by the physical forward modeling of the wave equation and the measured data from the field instruments, alleviates the domain offset problem caused by data distribution mismatch, and ensures that the corresponding feature channels in the U-Net convolutional neural network can be accurately activated when the measured data is input, thereby significantly improving the generalization ability and extraction accuracy of the scattered wave feature extraction model in real complex urban environments.

[0068] (d2) Intelligent Inference The standardized measured micro-motion recording fragments were used as input. The data is then fed in batches into the trained scattered wave feature extraction model.

[0069] The scattered wave feature extraction model automatically identifies hyperbolic texture features hidden in cluttered environmental noise through forward propagation and outputs a predicted pure scattered wave field. .

[0070] Noise reduction effect self-check: The output at this time In the process, the surface waves that originally dominated, the instantaneous strong pulses generated by vehicle rolling, and random electrical noise have all been suppressed or filtered out by the scattered wave feature extraction model, and only the scattered / diffraction energy related to underground inhomogeneities (pipelines, cavities) is retained.

[0071] This embodiment utilizes a scattered wave feature extraction model to replace traditional linear filters (such as FK filtering and Radon transform). Without the need for artificial seismic sources, it separates scattered waves from continuous micromotion records and combines this with reverse time migration (RTM) or Kirchhoff migration for imaging. This overcomes the resolution limitations of traditional micromotion technology (SPAC), which can only layer and cannot locate pipes, achieving decimeter-level fine imaging of small-scale discrete underground bodies (pipelines, cavities). Simultaneously, it overcomes the construction limitations of active source technology (SSP), which requires road closure and tapping, achieving zero-interference detection.

[0072] (d3) Signal reconstruction and gain recovery; Since the data output by the scattered wave feature extraction model is usually a normalized value, it needs to be denormalized to restore the true magnitude of the earthquake amplitude in order to perform subsequent physical imaging.

[0073] Spherical diffusion compensation: Considering that the energy of scattered waves decays rapidly with distance during propagation, the predicted pure scattered wave field output by the scattered wave feature extraction model is calculated. application or A form of time gain compensation is used to balance the energy difference between deep (10-20 meters) and shallow (3-5 meters) layers, wherein, For time, This is the attenuation compensation parameter, which is the compensation intensity for the attenuation of seismic wave energy absorption by the underground medium.

[0074] In one embodiment, step four involves performing migration imaging based on the predicted pure scattered wavefield and background velocity model to generate a high-resolution spatial image of the underground pipe network and its defects; specifically: This step utilizes the separated high signal-to-noise ratio scattered wave data, combined with a background velocity model, to generate high-resolution spatial images of underground pipe networks and defects.

[0075] (e1) Establish a background velocity model; Using the urban environmental background micro-motion records collected in step one, the surface wave dispersion curves are extracted by passive-source multichannel surface wave analysis (Passive-MASW) or extended spatial autocorrelation method (L-SPAC / ESPAC) suitable for linear arrays, and the one-dimensional or two-dimensional background S-wave velocity profile below the survey line is obtained by inversion.

[0076] Specifically, the passive source MASW can directly utilize road traffic noise parallel to the linear seismic line as an effective plane wave source for cross-correlation calculation; L-SPAC, based on the spatially stationary random field assumption, obtains the spatial autocorrelation coefficient by calculating the cross-correlation function of improved detector pairs with different channel spacings in the linear array. The background velocity model obtained here is no longer used to directly locate micro-pipelines, but only serves as a macroscopic velocity background for subsequent imaging, used to calculate the travel time of scattered waves.

[0077] (e2) Migration Imaging: The Kirchhoff Integral Migration or Reverse Time Migration (RTM) algorithms are employed.

[0078] Imaging principle: The predicted pure scattered wave field Pscatt output from step three is used as input. The travel time is calculated based on the background velocity model, and the scattered energy at each time point is extrapolated back to its spatial location (i.e., the location of the scattering point source).

[0079] Energy focusing: Due to the removal of surface wave interference, scattered energy from the same pipeline will be perfectly focused into a strong energy cluster on the imaging profile, thereby obtaining the true spatial location of the underground anomaly.

[0080] (e3) Image interpretation and disease identification: Generate depth-range domain imaging profiles; Pipeline identification: In the cross-sectional view, anomalies that appear as hyperbolic apex shapes or local strong amplitude clusters are identified as underground pipeline locations.

[0081] Disease identification: If a low-velocity loose halo or strong scattering energy dispersion appears above a known pipeline location, it is identified as loose soil or potential leakage. If irregular strong scattering clusters appear in non-pipeline areas, it is identified as underground cavities or isolated boulders.

[0082] Experiments were conducted using the scheme described in steps one through four above, and the results were obtained. Figure 6 The image effect diagram, Figure 6 (a) shows a fragment of the original measured micro-motion recordings acquired using a gravity-coupled crushable geophone array in an urban traffic environment. It can be seen that the effective weak geological signals are completely overwhelmed by the extremely low signal-to-noise ratio of traffic environment noise (such as continuous traffic flow and instantaneous strong pulses generated by vehicle rolling), resulting in a chaotic data structure. Traditional linear filtering methods (such as bandpass filtering, FK filtering, or Radon transform) are ineffective against this type of extremely low signal-to-noise ratio data. The underlying technical reason for their failure is twofold: firstly, urban broadband traffic noise (especially multi-directional surface waves) and weak pipeline scattered waves exhibit severe aliasing in both the frequency and apparent velocity domains, making it impossible for traditional linear filters based on fixed thresholds to remove noise without damaging the effective signal; secondly, the transient high-energy pulses generated by vehicles directly rolling over the geophone array induce severe Gibbs phenomenon (i.e., ringing effect) when processed by traditional linear filtering, and the resulting artifacts further mask the surrounding weak true hyperbolic signals.

[0083] Figure 6 (b) is to Figure 6 (a) After the data is input into the trained scattered wave feature extraction model, the output is the predicted pure scattered wave field. Through automatic suppression via nonlinear mapping and attention mechanisms, the dominant strong noise was completely stripped away, clearly and continuously restoring the multiple hyperbolic texture features generated by underground pipelines or discrete anomalies.

[0084] Figure 6 (c) is to Figure 6 (b) shows the high-resolution spatial result obtained after migration imaging of the predicted pure scattered wavefield. Figure 6 In (c), the hyperbolic energy that was originally diverging in the time domain is precisely back-calculated and perfectly focused into clear, strong energy clusters. Figure 6 (The black solid dots in (c)). These focused energy clusters precisely correspond to the actual physical spatial locations of underground pipe networks and minor defects.

[0085] contrast Figure 6As can be seen from (a), (b), and (c), this embodiment not only successfully overcomes the industry challenge of extracting weak discrete scattering signals under strong traffic noise (achieving a leap from disordered noise to ordered hyperbolic curves), but also breaks through the bottleneck of traditional passive source micro-motion technology, which can only perform macroscopic stratification (cannot focus), and achieves decimeter-level precise spatial positioning of small-scale underground pipelines and diseased bodies in cities.

[0086] Compared with the most widely used passive source micromotion detection techniques (SPAC / HVSR method) and active source seismic scattering (SSP) techniques in the industry, this embodiment has the following advantages by introducing deep learning feature extraction and a rollable passive source observation system: (f1) It breaks through the resolution limit of traditional micro-motion detection and realizes fine imaging of small-scale pipelines; The corresponding problem to be solved is that the existing passive source micro-motion technology (SPAC) is based on the layered medium assumption, which can only invert the velocity structure of the strata and cannot identify discrete small targets such as horizontally distributed underground pipelines and cavities.

[0087] Technical derivation and causal logic: Existing technology has shortcomings: SPAC technology uses environmental noise to calculate the spatial autocorrelation coefficient. Its physical essence is statistical averaging, which treats the local scattering signals generated by pipelines as interference and suppresses them. Therefore, its lateral resolution is usually only 1 / 2 of the array aperture (usually on the order of ten meters), presenting a blurry stratum slice.

[0088] Technical solution of this embodiment: This embodiment utilizes the U-Net convolutional neural network to directly extract scattered / diffracted waves from micro-motion records. According to Huygens' principle, scattered waves are the direct response of underground non-uniform bodies (pipelines) and carry high-frequency boundary information of the target body.

[0089] Results: By combining offset imaging algorithms, this embodiment can accurately locate the extracted scattered energy, clearly outlining the geometric contours of underground pipelines and local cavities with diameters of 0.5m-1.0m. This improves the lateral resolution of micro-motion detection from the ten-meter level to the decimeter level, filling the gap in passive source technology for refined urban health checks.

[0090] (f2) It has overcome the problem of weak signal extraction in noisy environments and significantly improved the imaging signal-to-noise ratio; The corresponding problem to be solved is that in busy urban areas, weak underground scattered signals (often with energy two orders of magnitude lower than environmental noise) are submerged, and traditional linear filters (FK filter, Radon transform) fail.

[0091] Technical derivation and causal logic: Current technology has limitations: Traditional mathematical filtering is based on the linear assumption that signal and noise are separated at apparent speed or frequency. However, urban traffic noise has an extremely wide frequency band and complex apparent speed variations, often covering the effective signal area. Forced filtering can lead to the loss of effective signals or the generation of a large number of Gibbs effect artifacts.

[0092] Technical solution of this embodiment: This embodiment employs a residual U-Net network with embedded attention gating. This network does not rely on simple frequency differences, but rather identifies signals by learning the unique hyperbolic texture characteristics of scattered waves in the spatiotemporal domain. In particular, the attention gating mechanism can automatically identify and suppress transient incoherent strong noise generated by vehicle rolling.

[0093] Results: This embodiment can still accurately recover the complete scattering waveform even in extreme environments where the signal-to-noise ratio is as low as -15dB (i.e., the noise energy is more than 30 times that of the signal). The effective signal extraction rate is more than 40% higher than that of traditional methods, and false anomalies in the imaging image are eliminated, reducing the engineering misjudgment rate.

[0094] (f3) It solves the pain point of construction being impossible in the core urban area and achieves rapid survey with zero interference; The corresponding problem to be solved: The traditional high-precision scattering technology (SSP) relies on an active seismic source (hammering / falling weight), which makes it impossible to implement in sensitive areas such as urban main roads and airports due to strict road closures and vibrations.

[0095] Technical derivation and causal logic: Current technology has shortcomings: Active source SSP technology requires closed roads for targeted tapping, resulting in low construction efficiency and significant administrative approval obstacles.

[0096] Technical solution of this embodiment: This embodiment designs a gravity-coupled crushable detector array, combined with a fully passive source acquisition mode. It is directly laid on the roadway, using passing vehicles as random vibration sources and the vibration of vehicles running over the cable as a near-field excitation source.

[0097] Results: It enables towed operation that allows for measurement on the go, without the need for road closures or manual activation. It reduces the operation time for a single measurement point from the traditional hours to 10-15 minutes, improves overall construction efficiency by more than 3 times, and achieves true zero interference with urban traffic and the surrounding environment.

[0098] (f4) solves the problem of lacking labeled data for the application of the model in the engineering field; The corresponding problem to be solved is to resolve the contradiction that deep learning models usually require massive amounts of real labeled data, while underground pipe networks are difficult to excavate and verify, and there is a lack of pure signal samples.

[0099] Technical derivation and causal logic: The technical solution of this embodiment proposes a semi-synthetic training dataset construction strategy that combines physical forward modeling (providing accurate labels) with real-world noise injection (providing environmental adaptability).

[0100] Results: This approach ensures that the pipeline wavefield characteristics learned by the model conform to the laws of physics (wavefield accuracy) and also makes it immune to the complex noise of real cities (environmental robustness). This strategy allows this embodiment to train a high-precision detection model without relying on expensive on-site excavation verification, greatly reducing the threshold and cost of technology implementation.

[0101] Based on the above description of the embodiments, those skilled in the art will understand that the deep learning-based underground pipe network micro-motion scattering imaging method, system, and medium described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0102] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for micro-motion scattering imaging of underground pipe networks, characterized in that, include: Continuous collection and segmentation of urban environmental background micro-motion records were performed. Initial pure scattered wave records containing underground scatterers were generated based on forward modeling of the wave equation. The records were then dynamically superimposed with the background micro-motion record segments to construct a semi-synthetic training dataset. A U-Net convolutional neural network with embedded residual learning units and attention gating mechanism is constructed and trained using the semi-synthetic training dataset to obtain a scattered wave feature extraction model for separating weak scattered wave signals. The measured micro-motion records are preprocessed and then input into the trained scattered wave feature extraction model. The pure scattered wave field is predicted by nonlinear mapping. Based on the predicted pure scattered wave field, and combined with the background velocity model obtained by inversion from the urban environmental background micro-motion record, a high-resolution spatial image of the underground pipe network and its defects is generated by migration imaging.

2. The method according to claim 1, characterized in that, Using a gravity-coupled, rollable geophone array deployed on a hardened road surface, continuous data collection of background micro-motions in the urban environment is achieved. The gravity-coupled crushable detector array includes multiple improved detectors, and adjacent improved detectors are connected in series by flat Kevlar cables to form a crushable drag cable. Each improved detector includes a pressure-bearing top cover (1), an aviation connector (2), a housing (3), and a base (4). The pressure-bearing top cover (1) is located on the upper part of the housing (3), and the base (4) is located on the lower part of the housing (3). The pressure-bearing top cover (1), the housing (3), and the base (4) are assembled into a sealed housing with waterproof and dustproof functions. A three-component detector is encapsulated inside the sealed housing. The aviation connector (2) passes through the housing (3) and connects to the three-component detector inside for power supply and data transmission. The main body of the outer shell (3) adopts a trapezoidal pressure-resistant structure and is designed as a streamlined or flat low-profile wedge structure; the base (4) is a high-density metal counterweight that achieves gravity coupling with the road surface.

3. The method according to claim 1, characterized in that, The process of generating the initial pure scattered wave record is as follows: Establish an underground geological model that includes scatterers of circular, rectangular, or irregular shapes; The total wavefield is obtained by performing forward modeling on a subsurface geological model containing scatterers; The background wavefield was obtained by performing a parametric forward modeling on an underground geological model after removing scatterers. The difference between the total wave field and the background wave field is calculated, and the difference is used as the initial pure scattered wave record. The initial pure scattered wave record contains only hyperbolic texture features generated by the pipeline, which is used as the target label of the U-Net convolutional neural network.

4. The method according to claim 1, characterized in that, The semi-synthetic training dataset The construction formula is as follows: ； in, Records of the initial, pure scattered waves. Recording snippets of subtle movements against the backdrop of the urban environment. This is the noise adjustment factor.

5. The method according to claim 1, characterized in that, The U-Net convolutional neural network embeds an attention gating module at the skip connection between the encoder and decoder to automatically suppress instantaneous high-energy pulse interference caused by vehicle rolling and to enhance the response to hyperbolic texture features. Residual learning units are introduced into the convolutional layers of the encoder to learn the residual mapping between the semi-synthetic training dataset and the initial pure scattered wave record.

6. The method according to claim 1, characterized in that, The combined loss function during training of the U-Net convolutional neural network The settings are as follows: ； in, The predicted pure scattering wave field is the output of the U-Net convolutional neural network after the semi-synthetic training dataset is input into it; The initial pure scattered wave record, which serves as the target label, corresponds to the semi-synthetic training dataset; for and L1 norm loss between; for and Multi-scale structural similarity loss between them These are the weighting coefficients.

7. The method according to claim 1, characterized in that, The preprocessing process for the measured micro-motion records is as follows: After performing amplitude equalization or truncation on the measured micro-motion records, the records are extracted into segments according to a set time window. Standardization processing: Z-Score standardization or maximum value normalization is performed on each measured micro-motion recording segment to make its amplitude distribution consistent with the distribution characteristics of the training samples in the semi-synthetic training dataset; The standardized measured micro-motion recording fragments were used as input to the trained scattered wave feature extraction model.

8. The method according to claim 1, characterized in that, The migration imaging employs Kirchhoff integral migration, reverse time migration, beamforming, least squares migration, or full waveform inversion algorithms. The predicted pure scattered wave field is used as input, and the travel time is calculated based on the background velocity model. The scattered energy at each time point is then backtracked to its spatial location. Scattered energy from the same pipeline is focused and returned to the true spatial location of the underground anomaly.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of computer programs, which are used to be invoked by a processor and to execute the method as described in any one of claims 1-8.

Images