Bridge scour pit three-dimensional reconstruction method based on physical constraint generative adversarial network

Abstract

The invention relates to a bridge scouring pit three-dimensional reconstruction method based on a physical constraint generative adversarial network, and the method comprises the steps: collecting optical data and multi-beam sonar data, generating a joint calibration file, and carrying out the preprocessing; performing multi-source data fusion processing to obtain a training set and a test set; constructing a physical constraint generative adversarial network; and inputting the test set into the physical constraint generative adversarial network, obtaining point cloud data, and performing engineering verification. The method has the beneficial effects that through collaborative innovation of multi-modal sensing fusion and physical constraint deep learning, the problem of high-precision reconstruction of the scouring pit terrain in the complex underwater environment is effectively solved. Compared with a traditional single-mode detection or pure data driven reconstruction method, the method creatively introduces a Navier-Stokes equation as a physical constraint term of the generative adversarial network, and embeds a fluid mechanics mechanism into a deep learning framework, thereby remarkably improving the physical rationality of a reconstruction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater engineering detection, and in particular relates to a three-dimensional reconstruction method of a bridge scour pit based on a physical constraint generative adversarial network. Background Art

[0002] Bridge foundation scour is the primary risk factor threatening the safety of river-crossing structures. Traditional scour monitoring relies primarily on contact methods such as diver exploration and single-beam bathymetry, which suffer from low efficiency and insufficient spatial resolution. While recently developed 3D reconstruction technology enables non-contact detection, it still faces numerous technical challenges in underwater environments.

[0003] Underwater optical imaging is susceptible to interference from light attenuation and suspended matter scattering, especially in turbid waters, where reconstruction errors increase significantly. Although polarization imaging technology can partially suppress the scattering effect, its ability to solve the problem of occlusion by dynamic floating objects is limited. Sonar technology has strong water penetration capabilities, but its point cloud density is usually low, making it difficult to accurately depict the detailed features of scour pits, and it is easily affected by water flow noise, resulting in pseudo-topography data. Although existing deep learning-based three-dimensional reconstruction methods have made breakthroughs in geometric accuracy, most of them do not consider fluid mechanics constraints, resulting in large deviations between the reconstruction results and the actual water flow-sediment interaction mechanism. In addition, the differences in coordinate system, resolution, and acquisition timing between optical and sonar data make multi-source data fusion difficult, and traditional registration methods may cause obvious spatial misalignment problems. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a three-dimensional reconstruction method for bridge scour pits based on a physical constraint generative adversarial network.

[0005] First, a 3D reconstruction method for bridge scour pits based on a physical constraint generative adversarial network is provided, including:

[0006] S1, collect optical data and multi-beam sonar data, generate joint calibration files, and perform preprocessing;

[0007] S2: Based on the data obtained in S1, multi-source data fusion processing is performed to obtain training sets and test sets;

[0008] S3. Constructing a physical constraint generative adversarial network and training it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network;

[0009] S4. Input the test set into the physical constraint generative adversarial network to obtain point cloud data and perform engineering verification.

[0010] Preferably, S1 includes:

[0011] S101, collecting optical data and multi-beam sonar data; the optical data is an RGB-D video stream with polarization filtering, and the multi-beam sonar data is riverbed benchmark elevation data;

[0012] S102, aligning the riverbed benchmark elevation data and the optical data using UTC timestamps to generate a joint calibration file;

[0013] S103, performing water scattering compensation on each frame of image using a dark channel priori defogging algorithm;

[0014] S104, detect floating objects on the water surface based on the YOLOv5s model and generate a dynamic mask;

[0015] S105: Perform polarization fusion processing.

[0016] Preferably, S2 includes:

[0017] S201, obtaining a preset data set;

[0018] S202, performing coordinate unification processing on the optical data and multi-beam sonar data acquired in S1;

[0019] S203, based on the turbidity simulation noise generated by the Monte Carlo light propagation model, the spatial and optical data obtained in S1 are enhanced to ensure that the data used in the training set are enhanced data;

[0020] S204. Obtain true value labels through field measurements or high-precision simulations, and construct a training set containing optical data, sonar data, and true value labels;

[0021] S205. Construct a test set including CFD simulation parameters and turbulence intensity parameters.

[0022] Preferably, S3 includes:

[0023] S301, initializing network parameters;

[0024] S302, constructing a generator network and a discriminator network;

[0025] S303. Utilize the Navier-Stokes equations to constrain the learning process of the generative adversarial network and configure a multi-constraint loss function; the multi-constraint loss function includes adversarial loss, geometric consistency loss, and fluid constraint loss.

[0026] Preferably, S4 includes:

[0027] S401, input the test set image to the generator to obtain point cloud data;

[0028] S402, perform CFD fluid verification;

[0029] S403. Use the Bootstrap resampling method to evaluate the error distribution.

[0030] In a second aspect, a bridge scour pit 3D reconstruction system based on a physical constraint generative adversarial network is provided, which is used in any of the methods described in the first aspect, including:

[0031] The acquisition module is used to collect optical data and multi-beam sonar data, generate joint calibration files, and perform pre-processing;

[0032] The fusion module is used to perform multi-source data fusion processing based on the data obtained by the acquisition module to obtain training sets and test sets;

[0033] A construction module is used to construct a physical constraint generative adversarial network and train it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network;

[0034] The verification module is used to input the test set into the physical constraint generative adversarial network, obtain point cloud data, and perform engineering verification.

[0035] According to a third aspect, a computer storage medium is provided, wherein a computer program is stored in the computer storage medium; when the computer program is executed on a computer, the computer executes any one of the methods described in the first aspect.

[0036] In a fourth aspect, an electronic device is provided, including:

[0037] Memory, used to store computer programs;

[0038] A processor is used to execute the computer program to implement any method as described in the first aspect.

[0039] The beneficial effects of the present invention are:

[0040] 1. This invention fuses optical data and multi-beam sonar data to achieve multimodal sensing fusion, which, combined with physical constraints, effectively solves the problem of high-precision reconstruction of scour pit topography in complex underwater environments.

[0041] 2. Compared with traditional single-modal detection or purely data-driven reconstruction methods, this method creatively introduces the Navier-Stokes equations as physical constraints for the generative adversarial network, embedding fluid mechanics mechanisms into the deep learning framework, significantly improving the physical rationality of the reconstruction results.

[0042] 3. This invention addresses pain points such as underwater optical attenuation and suspended object interference by using polarization fusion processing, combined with dark channel prior and dynamic masking technology, to maintain millimeter-level reconstruction accuracy even in a turbidity environment of 100 NTU.

[0043] 4. This invention utilizes a Monte Carlo ray propagation model to enhance data robustness, and employs CFD fluid validation and bootstrap resampling to assess error distribution. Compared to traditional sonar depth sounding and structured light scanning techniques, this method achieves a 3-5x improvement in spatial resolution (2000 points / m²) and shear stress prediction accuracy (error ≤ 12%), providing a highly accurate and robust solution for dynamic bridge scour monitoring.

[0044] 5. This invention aligns riverbed benchmark elevation data (sonar data) and optical data using UTC timestamps and performs coordinate unification processing. This can resolve the differences in coordinate system, resolution, and acquisition timing between optical and sonar data, thereby avoiding spatial misalignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a GAN architecture flow chart of the bridge scour pit 3D reconstruction method based on physical constraint generative adversarial network provided by the present invention;

[0046] Figure 2 This is the technical roadmap of the three-dimensional reconstruction method of bridge scour pits based on physical constraint generative adversarial network provided by the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described below with reference to the following examples. The following examples are provided only to facilitate understanding of the present invention. It should be noted that, without departing from the principles of the present invention, it is possible for a person skilled in the art to make various modifications to the present invention, and such improvements and modifications fall within the scope of the claims of the present invention.

[0048] Example 1:

[0049] To address the challenges of existing technologies, Example 1 of this application provides a 3D reconstruction method for bridge scour pits based on a physically constrained generative adversarial network. By leveraging multimodal data collaborative acquisition and fusion techniques, combined with advanced image processing algorithms and physical information neural networks, this method significantly improves the accuracy and reliability of underwater scour terrain reconstruction. Field measurements demonstrate that this method exhibits superior performance in complex underwater environments, providing effective technical support for bridge safety monitoring and digital twin applications.

[0050] Specifically, such as Figure 2 As shown, the 3D reconstruction method of bridge scour pits based on physical constraint generative adversarial networks provided in this application includes:

[0051] S1. Collect optical data and multi-beam sonar data, generate joint calibration files, and perform preprocessing.

[0052] S1 includes:

[0053] S101. Collect optical data and multi-beam sonar data; the optical data is an RGB-D video stream with polarization filtering, and the multi-beam sonar data is riverbed benchmark elevation data.

[0054] For example, a circular array of six waterproof cameras (focal length 12 mm, baseline distance 1.5 m) was deployed to simultaneously capture polarization-filtered RGB-D video streams (resolution 3840 × 2160 @ 30 fps) and store them as H.265-encoded MP4 files. Furthermore, high-frequency multibeam sonar (1 MHz frequency, 120° beam angle) was used to obtain riverbed benchmark elevation data.

[0055] In addition, in S101, a camera attitude drift compensation model needs to be established. R represents the time interval The camera's attitude rotation change matrix within t, T represents the time interval The displacement of the camera in t:

[0056] ,

[0057] The angular velocity ω and linear velocity v are measured in real time by the IMU sensor (sampling rate 100 Hz), and the compensation period Δt = 1s.

[0058] S102: Align the riverbed benchmark elevation data and the optical data using UTC timestamps to generate a joint calibration file.

[0059] In S102, the format of the joint calibration file is JSON. Furthermore, time synchronization between sonar and optical data requires clock synchronization via the PTP protocol, with a master-slave clock offset compensation of δt ≤ 1ms. Spatial calibration uses the checkerboard target method. After initial extrinsic calibration in a dry environment, calibration validity is verified through underwater reprojection error optimization (RMSE ≤ 5cm).

[0060] S103: Using a dark channel prior defogging algorithm, water scattering compensation is performed on each frame of the image.

[0061] Specifically, guided filtering is used for transmittance estimation to reduce the influence of water scattering. This method effectively preserves image edge details while smoothing noise. The guided image is set as a weighted average of the RGB channels (weight coefficients [0.299, 0.587, 0.114]), with a filter radius of r = 7 pixels and a regularization parameter ε = 0.001. Furthermore, a dark channel prior dehazing algorithm is used to compensate for water scattering in each image frame. The atmospheric light value A is set to [0.8, 0.9, 0.95] × 255, and the transmittance map calculation window size is 15 × 15 pixels.

[0062] S104. Detect floating objects on the water surface based on the YOLOv5s model and generate a dynamic mask (binary image saved as PNG).

[0063] Specifically, the input size of the YOLOv5s model is 640×640 and the confidence threshold is 0.6.

[0064] S105: Perform polarization fusion processing.

[0065] Specifically, calculate the Stokes vector S=[I0,I45,I90,I135 ], extract the effective polarization component , saved in 16-bit TIFF format. In addition, the calculation of the effective polarization component needs to consider the Mueller matrix correction of the Stokes vector:

[0066]

[0067] The water body Mueller matrix Mwater is pre-calibrated through Monte Carlo simulation, and the simulation parameters include the water body absorption coefficient , scattering coefficient α and β are weight coefficients. Represents the low-frequency part of the image, Represents the low-frequency part of the Stokes vector, NSCT( , ) is the non-subsampled contourlet transform. Where: Represents the high-frequency part of the image, Represents the high frequency part of the Stokes vector.

[0068] In addition, polarization fusion processing adopts frequency domain fusion strategy, including: yield stress =2.5Pa, consistency coefficient K=0.8Pa·s^n, flow index n=0.6; the polarization component Ivalid is decomposed into low-frequency component ILF and high-frequency component IHF by Contourlet transform; the fusion formula is: , where the fusion coefficients α = 0.6, β = 1.2, and NSCT represents the non-subsampled contourlet transform fusion operator. The final fused image must meet the requirements of edge preservation index EPI ≥ 0.85 and structural similarity SSIM ≥ 0.92.

[0069] S2: Based on the data obtained in S1, multi-source data fusion processing is performed to obtain training sets and test sets.

[0070] S2 has built a spatiotemporal compensation mechanism based on inertial navigation: the six-degree-of-freedom motion parameters are acquired in real time through the IMU module (sampling rate 200Hz) to establish a coordinate system dynamic compensation model:

[0071]

[0072] The gravitational acceleration g is estimated in real time by Kalman filtering, and the SO3 index mapping is used to update the rotation matrix. is the angular velocity vector measured by the IMU sensor.

[0073] Construct a motion distortion compensation network, input time series IMU data and point cloud sequence, and output deformation field (x, y, z, t), the network structure includes a spatiotemporal encoder and a deformation predictor. The spatiotemporal encoder is a 3-layer GRU network (hidden layer 128 dimensions), which is used to extract motion features; the deformation predictor is a fully connected layer, which is used to output 3D displacement. The constraints are The compensated data meets the requirements of ICP matching error of adjacent frame point clouds ≤ 3mm and IMU integral position residual ≤ 5cm.

[0074] S2 includes:

[0075] S201. Obtain a preset data set. This data set is mainly used to provide additional true value labels and reference information to assist in the construction of training and test sets.

[0076] For example, the preset dataset is the USGS Field Survey Archive dataset, and samples with a scanning integrity of <85% are removed (a total of 213 sets of valid data are retained).

[0077] S202: performing coordinate unification processing on the optical data and multi-beam sonar data acquired in S1.

[0078] Specifically, the local coordinate system was converted to the UTM Zone 50N coordinate system, and the elevation datum was unified to the EGM2008 geoid. During this process, an IMU module (sampling rate of 200Hz) was used to acquire six-degree-of-freedom motion parameters in real time. A coordinate system dynamic compensation model was established to reduce measurement errors caused by device movement. Furthermore, a motion distortion compensation network was constructed.

[0079] S203. Based on the turbidity simulation noise generated by the Monte Carlo light propagation model, the spatial and optical data obtained in S1 are enhanced to ensure that the data used in the training set is enhanced data.

[0080] S204: Using the data from S1 and the coordinate unification processing results in S202, a training set containing optical data, sonar data and true value labels is constructed.

[0081] Specifically, the training set contains optical data (RGB-D sequence resolution 1920×1080), sonar data (point cloud density ≥ 500 pts / m²), and true value labels (terrestrial laser scanning point cloud, accuracy 0.5mm).

[0082] S205 , based on the preset data set and CFD simulation parameters in S201 , construct a test set including CFD simulation parameters and turbulence intensity parameters.

[0083] Specifically, the test set includes CFD simulation parameters (Reynolds number Re∈[1×10^4,5×10^5]) and turbulence intensity parameters. The CFD simulation parameters include a non-Newtonian fluid model using the Herschel-Bulkley constitutive equation: .

[0084] In S205, in order to simulate the changes in the optical properties of water bodies in a real environment, the present invention uses a Monte Carlo light propagation model to generate turbidity simulation noise, enhances spatial and optical data, randomly rotates (angle range ±15°) and translates (offset ±0.3m), and adds turbidity simulation noise (scattering coefficient β = 0.6-1.2 corresponding to NTU values ​​of 30-100).

[0085] Specifically, the turbidity simulation noise is generated by the Monte Carlo light propagation model, and the number of photons N is set to 10. 6 , the scattering phase function adopts the Henyey-Greenstein model (anisotropy factor g = 0.85).

[0086] S3. Construct a physical constraint generative adversarial network and train it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network.

[0087] S4. Input the test set into the physical constraint generative adversarial network to obtain point cloud data and perform engineering verification.

[0088] Example 2:

[0089] Based on Example 1, Example 2 of the present application provides a more specific method for 3D reconstruction of bridge scour pits based on a physical constraint generative adversarial network, including:

[0090] S1. Collect optical data and multi-beam sonar data, generate joint calibration files, and perform preprocessing.

[0091] S2: Based on the data obtained in S1, multi-source data fusion processing is performed to obtain training sets and test sets.

[0092] S3. Construct a physical constraint generative adversarial network and train it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network.

[0093] S3 includes:

[0094] S301: Initialize network parameters.

[0095] In S301, the generator input size is set to 256×256×32 (corresponding to 2m×2m×0.5m in actual space), and the discriminator adopts the PatchGAN architecture (5 layers of 3D convolution, kernel size 4×4×4).

[0096] S302. Construct a generator network and a discriminator network.

[0097] Specifically, such as Figure 1 As shown, the output of the discriminator network provides optimization feedback to the generator network to guide the generator to continuously optimize the generated point cloud data, and the generator network consists of the following modules:

[0098] Generator: Responsible for generating preliminary point cloud data from the input noisy point cloud.

[0099] Vector module: Converts the preliminary point cloud output by the generator into a feature vector.

[0100] 3D Convolutional Network: Performs multi-scale feature extraction on feature vectors, including 3DResNet blocks with 4 levels of downsampling (number of channels per level [64, 128, 256, 512]). Each residual block contains a bottleneck structure, which helps reduce computational complexity and accelerate the training process.

[0101] MLP generator: Generates more accurate point cloud details directly from latent space vectors, making the final output closer to the actual terrain features.

[0102] The discriminator network consists of the following modules:

[0103] Discriminator: Receives the SDF field as input and evaluates it.

[0104] Residual block: By introducing skip connections, it alleviates the gradient vanishing problem in deep networks and further optimizes feature extraction.

[0105] Fully connected layer: Integrates feature information at different levels and outputs the final discrimination result.

[0106] In addition, the above 4-level downsampling 3DResNet block (number of channels per level [64, 128, 256, 512]) uses a cross-modal attention module as the encoder-decoder, and its operation formula is:

[0107]

[0108] in, represents the final feature representation after processing by the cross-modal attention module, It is a 3x3 convolution operation, which is used to perform spatial convolution operations on the input features. represents the feature representation extracted from the image data, An activation function to scale the attention scores to the range [0, 1], and The weight matrices are the query matrix and the key matrix, Represents the feature representation extracted from point cloud data.

[0109] In addition, the 3DResNet block adopts a bottleneck structure. Each residual block contains a 1×1×1 dimensionality reduction convolution, a 3×3×3 main convolution, and a 1×1×1 dimensionality increase convolution. The jump connection adds a channel attention mechanism.

[0110] In addition, the cross-modal attention module adopts a dual-stream adaptive gating mechanism. The specific implementation includes:

[0111] Dynamic weight matrix calculation:

[0112] Feature recombination formula:

[0113] The dynamic adjustment factor is: ,

[0114] is the output feature, is the image feature, is the point cloud feature, σ represents the Sigmoid function, the depthwise separable convolution DWConv kernel size is 5×5×5, and the expansion rate d=2.

[0115] S303. To ensure physical authenticity, the Navier-Stokes equations are used to constrain the learning process of the generative adversarial network, and a multi-constraint loss function is configured; the multi-constraint loss function includes adversarial loss, geometric consistency loss, and fluid constraint loss.

[0116] Specifically, the residual calculation of the Navier-Stokes equation must satisfy the CFL stability condition:

[0117]

[0118] in is the magnitude of the fluid velocity, c is the speed of sound, which indicates the speed at which sound waves propagate in the fluid, and Δx = 0.01 m is the spatial discrete step length, thereby determining the time step Δt = 0.001 s.

[0119] The residual calculation of the Navier-Stokes equations uses finite volume discretization, and the governing equation is expressed as:

[0120]

[0121] Wherein, the sediment volume force fsediment=α(ρs-ρf)g, α is the sediment volume fraction, is the sediment density; is the velocity field; is the fluid density; is the dynamic viscosity; is the acceleration due to gravity; is the pressure gradient term; Represents the spatial second-order derivative of the velocity field.

[0122] In addition, the adversarial loss uses HingeLoss (weight 1.0), the geometric consistency loss uses ChamferDistance (weight 0.5), and the fluid constraint loss uses the Navier-Stokes equation residual L2 norm (weight 2.0). The total loss function expression is:

[0123]

[0124] .in, Adversarial loss weight = 1.0, indicating the relative importance of adversarial loss in the total loss function; To combat losses; The geometric consistency loss weight is 0.5; is the geometric consistency loss; The fluid constraint loss weight is 2.0; is the residual of the Navier-Stokes equation, which represents the residual value of the fluid dynamics equation under the current predicted velocity field u and pressure field p.

[0125] S4. Input the test set into the physical constraint generative adversarial network to obtain point cloud data and perform engineering verification.

[0126] S4 includes:

[0127] S401: Input the test set image into the generator to obtain point cloud data.

[0128] Specifically, the density of point cloud data is 2000 pts / m².

[0129] S402. Perform CFD fluid verification.

[0130] Specifically, the reconstructed model was imported into ANSYS Fluent (turbulence model k-ω SST, boundary conditions: flow velocity at the inlet 3 m / s, sediment particle size 0.2 mm, density 2650 kg / m³).

[0131] S403. Use the Bootstrap resampling method to evaluate the error distribution.

[0132] Specifically, the Bootstrap resampling method is introduced to evaluate the error distribution. A large number of repeated sampling (e.g., 1000 samplings with replacement) is used to estimate the error confidence interval: the 95% CI of the MAE should meet [1.8cm, 2.2cm], and the normality of the error distribution is verified (Kolmogorov-Smirnov test p>0.05). For example, the accuracy assessment is performed by calculating the average point cloud error (compared to terrestrial laser scanning data) and the relative error of shear stress (compared to the CFD simulation benchmark value), meeting the indicator: MAE ≤ 2cm, ≤12%.

[0133] In addition, the shear stress is calculated using the wall function method, and the height of the first layer of grid near the wall satisfies y+≈30. The calculation formula is:

[0134] where u τ is the friction velocity, which is calculated by the turbulent kinetic energy k and the dissipation rate ε, is the dynamic viscosity, is the fluid density, is the velocity gradient at the wall (i.e., at y = 0).

[0135] To verify the effectiveness of this method, a comparative experiment was conducted between this method and representative existing technologies under the same test environment (turbidity 100 NTU, flow velocity 3 m / s, sediment particle size 0.2 mm) (results are shown in Table 1). The data in Table 1 are the average values ​​of five independent experiments.

[0136] Table 1: Comparison of scour pit reconstruction performance of different methods in a turbidity environment of 100 NTU (test conditions: flow velocity 3 m / s, sediment particle size 0.2 mm)

[0137]

[0138] Compared with the traditional sonar reconstruction method, this method reduces the mean absolute error (MAE) by 78% (1.8cm vs 8.2cm) and the prediction error of the key hydraulic parameter shear stress ( ) by 77% (9.5% vs. 42.1%). This is primarily due to the fusion of multimodal data (high-resolution RGB-D and polarization information) and the powerful feature extraction capabilities of the deep learning model, which significantly overcomes the limitations of sparse sonar point clouds and optical scattering interference. Compared to a generative adversarial network (GAN) without physical constraints, this method achieves a 66% reduction in MAE (1.8cm vs. 5.3cm). This fully demonstrates the introduction of the Navier-Stokes equation as a physical constraint loss term (weight =2.0), significantly improving the physical rationality of the reconstructed terrain and the accuracy of flow field prediction, and avoiding artifacts that may violate the laws of fluid mechanics that may be produced by purely data-driven methods. At the same time, this method successfully outputs a high-density point cloud (≥2000 pts / m²), providing a sufficient data basis for depicting the subtle terrain features of the scour pit (such as local vortex pits and steep slopes), which is much higher than traditional sonar methods and ensures physical authenticity. In summary, this method has achieved excellent results in reconstruction accuracy (MAE), key engineering parameter prediction accuracy ( ) and detail restoration capabilities (point cloud density) have significantly surpassed existing technologies (increased by 3-5 times), fully verifying its ability to perform high-precision and high-reliability 3D reconstruction of bridge scour pits in complex underwater environments with high turbidity.

[0139] It should be noted that the parts in this embodiment that are the same or similar to those in Example 1 can be referenced to each other and will not be described in detail in this application.

[0140] Example 3:

[0141] Based on Example 2, Example 3 of the present application provides a bridge scour pit 3D reconstruction system based on a physical constraint generative adversarial network, including:

[0142] The acquisition module is used to collect optical data and multi-beam sonar data, generate joint calibration files, and perform pre-processing;

[0143] The fusion module is used to perform multi-source data fusion processing based on the data obtained by the acquisition module to obtain training sets and test sets;

[0144] A construction module is used to construct a physical constraint generative adversarial network and train it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network;

[0145] The verification module is used to input the test set into the physical constraint generative adversarial network, obtain point cloud data, and perform engineering verification.

[0146] It should be noted that the system provided in this embodiment is a system corresponding to the method provided in Example 2. Therefore, the parts in this embodiment that are the same or similar to those in Example 2 can be referenced to each other and will not be repeated in this application.

Claims

1. A 3D reconstruction method for bridge scour pits based on physical constraint generative adversarial networks, characterized by: include: S1, collect optical data and multi-beam sonar data, generate joint calibration files, and perform preprocessing; S2: Based on the data obtained in S1, multi-source data fusion processing is performed to obtain training sets and test sets; S3. Constructing a physical constraint generative adversarial network and training it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network; S4. Input the test set into the physical constraint generative adversarial network to obtain point cloud data and perform engineering verification.

2. The method for 3D reconstruction of bridge scour pits based on physical constraint generative adversarial networks according to claim 1 is characterized in that: S1 includes: S101, collect optical data and multi-beam sonar data; The optical data is an RGB-D video stream with polarization filtering, and the multi-beam sonar data is riverbed benchmark elevation data; S102, aligning the riverbed benchmark elevation data and the optical data using UTC timestamps to generate a joint calibration file; S103, performing water scattering compensation on each frame of image using a dark channel priori defogging algorithm; S104, detect floating objects on the water surface based on the YOLOv5s model and generate a dynamic mask; S105: Perform polarization fusion processing.

3. The bridge scour pit 3D reconstruction method based on physical constraint generative adversarial network according to claim 2 is characterized in that S2 include: S201, obtaining a preset data set; S202, performing coordinate unification processing on the optical data and multi-beam sonar data acquired in S1; S203, enhancing the spatial and optical data acquired in S1 based on turbidity simulation noise generated by the Monte Carlo light propagation model; S204. Obtain true value labels through field measurements or high-precision simulations, and construct a training set containing optical data, sonar data, and true value labels; S205. Construct a test set including CFD simulation parameters and turbulence intensity parameters.

4. The method for 3D reconstruction of bridge scour pits based on physical constraint generative adversarial networks according to claim 3 is characterized in that S3 include: S301, initializing network parameters; S302, constructing a generator network and a discriminator network; S303. Utilize the Navier-Stokes equations to constrain the learning process of the generative adversarial network and configure a multi-constraint loss function; the multi-constraint loss function includes adversarial loss, geometric consistency loss, and fluid constraint loss.

5. The method for 3D reconstruction of bridge scour pits based on physical constraint generative adversarial networks according to claim 4 is characterized in that S4 include: S401, input the test set image to the generator to obtain point cloud data; S402, perform CFD fluid verification; S403. Use the Bootstrap resampling method to evaluate the error distribution.

6. A 3D reconstruction system for bridge scour pits based on physical constraint generative adversarial networks, characterized by: Used to perform the method according to any one of claims 1 to 5, comprising: The acquisition module is used to collect optical data and multi-beam sonar data, generate joint calibration files, and perform pre-processing; The fusion module is used to perform multi-source data fusion processing based on the data obtained by the acquisition module to obtain training sets and test sets; A construction module is used to construct a physical constraint generative adversarial network and train it according to the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network; The verification module is used to input the test set into the physical constraint generative adversarial network, obtain point cloud data, and perform engineering verification.

7. A computer storage medium, characterized in that The computer storage medium stores a computer program; when the computer program is run on a computer, the computer executes the method according to any one of claims 1 to 5.

8. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 5.

Images