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

Abstract

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

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater engineering detection, and particularly relates to a bridge scour pit three-dimensional reconstruction method based on a physical constraint generative adversarial network. BACKGROUND

[0002] Bridge foundation scour is the primary risk factor threatening the safety of river-crossing structures. Traditional scour monitoring mainly relies on contact methods such as diver exploration and single-beam sounding, which have low efficiency and insufficient spatial resolution. Although the three-dimensional reconstruction technology developed in recent years can achieve non-contact detection, it still faces many technical challenges in underwater environments.

[0003] Underwater optical imaging is easily disturbed by light attenuation and suspended matter scattering, especially in turbid water, where the reconstruction error increases significantly. Although polarization imaging technology can partially suppress scattering effects, it has limited ability to solve the problem of dynamic floating object occlusion. Sonar technology has strong water penetration capability, but its point cloud density is usually low, making it difficult to accurately depict the detailed features of the scour pit, and it is easily affected by flow noise, resulting in false terrain data. Existing three-dimensional reconstruction methods based on deep learning have made breakthroughs in geometric accuracy, but most of them do not consider fluid mechanics constraints, leading to significant deviations between the reconstruction results and the actual flow-sediment interaction mechanism. In addition, the differences in coordinate systems, resolution, and acquisition timing between optical and sonar data make it difficult to fuse multi-source data, and traditional registration methods may cause significant spatial misalignment problems. SUMMARY

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

[0005] In a first aspect, a bridge scour pit three-dimensional reconstruction method based on a physical constraint generative adversarial network is provided, comprising:

[0006] S1, collecting optical data and multi-beam sonar data, generating a joint calibration file, and performing preprocessing;

[0007] S2, performing multi-source data fusion processing according to the data obtained in S1 to obtain a training set and a test set;

[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, inputting the test set into the physical constraint generative adversarial network to obtain point cloud data and performing engineering verification.

[0010] As a preferred embodiment, S1 comprises:

[0011] S101, collect optical data and multi-beam sonar data; the optical data is a polarized filter RGB-D video stream, and the multi-beam sonar data is riverbed datum elevation data;

[0012] S102, align the riverbed datum elevation data with the optical data using UTC timestamps to generate a joint calibration file;

[0013] S103, perform water body scattering compensation on each frame of image using a dark channel prior dehazing algorithm;

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

[0015] S105, perform polarization fusion processing.

[0016] Preferably, S2 comprises:

[0017] S201, obtain a preset data set;

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

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

[0020] S204, obtain true value labels through field measurement or high-precision simulation, and construct a training set containing optical data, sonar data and true value labels;

[0021] S205, construct a test set containing CFD simulation parameters and turbulence intensity parameters.

[0022] Preferably, S3 comprises:

[0023] S301, initialize network parameters;

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

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

[0026] Preferably, S4 comprises:

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

[0028] S402, perform CFD fluid verification;

[0029] S403, using Bootstrap resampling method to evaluate error distribution.

[0030] In a second aspect, a bridge scour pit three-dimensional reconstruction system based on a physically constrained generative adversarial network is provided for the method of any one of the first aspect, comprising:

[0031] The acquisition module is configured to acquire optical data and multi-beam sonar data, generate a joint calibration file, and perform preprocessing.

[0032] The fusion module is configured to perform multi-source data fusion processing according to the data obtained by the acquisition module, and obtain a training set and a test set.

[0033] The construction module is configured to construct a physically constrained generative adversarial network, and train the network according to the training set; the physically constrained generative adversarial network comprises a generator network and a discriminator network.

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

[0035] In a third aspect, a computer storage medium is provided, and the computer storage medium stores a computer program; when the computer program runs on a computer, the computer program causes the computer to execute the method of any one of the first aspect.

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

[0037] The memory is configured to save a computer program.

[0038] The processor is configured to execute the computer program to implement the method of any one of the first aspect.

[0039] The present application has the following advantages:

[0040] 1. The present application solves the problem of high-precision reconstruction of complex underwater environment scour pit terrain by fusing optical data and multi-beam sonar data, i.e., constructing multi-modal sensor fusion, and combining physical constraints.

[0041] 2. Compared with traditional single-mode detection or pure data-driven reconstruction method, the present application creatively introduces Navier-Stokes equation as a physically constrained term of generative adversarial network, embeds fluid mechanics mechanism into deep learning framework, and significantly improves the physical rationality of reconstruction results.

[0042] 3. The present application adopts polarization fusion processing, combines dark channel prior and dynamic mask technology, and still maintains millimeter-level reconstruction accuracy in a turbidity of 100 NTU, aiming at the pain points such as underwater optical attenuation and suspended matter interference.

[0043] 4. The application uses the Monte Carlo light propagation model to enhance the robustness of the data, and adopts the CFD fluid verification and Bootstrap resampling method to evaluate the error distribution. Compared with traditional sonar depth measurement and structured light scanning technology, the method provided by the application realizes 3-5 times improvement in spatial resolution (2000 points / m²) and shear stress prediction accuracy (error ≤12%), and provides a solution with high precision and strong generalization ability for bridge scour dynamic monitoring.

[0044] 5. The application aligns the riverbed datum elevation data (sonar data) with the optical data using UTC time stamp, and performs coordinate unification processing, which can solve the differences in coordinate system, resolution and acquisition time sequence between optical and sonar data, and further avoid spatial misplacement problems. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the GAN architecture flow chart of the bridge scour pit three-dimensional reconstruction method based on the physical constraint generative adversarial network provided by the application;

[0046] Figure 2 is the technical route map of the bridge scour pit three-dimensional reconstruction method based on the physical constraint generative adversarial network provided by the application. DETAILED DESCRIPTION

[0047] The application will be further described below in combination with examples. The following examples are only used to help understand the application. It should be noted that for ordinary people in the technical field, some modifications can be made to the application without departing from the principles of the application, and these improvements and modifications also fall within the protection scope of the claims of the application.

[0048] Example 1

[0049] To solve the problems of the prior art, Example 1 of the application provides a bridge scour pit three-dimensional reconstruction method based on a physical constraint generative adversarial network. Through multi-modal data cooperative acquisition and fusion technology, combined with advanced image processing algorithms and physical information neural networks, the reconstruction accuracy and reliability of underwater scour topography are significantly improved. The actual measurement results show that the method has superior performance in complex underwater environments, and provides effective technical support for bridge safety monitoring and digital twin application.

[0050] Specifically, as shown in Figure 2 The bridge scour pit three-dimensional reconstruction method based on the physical constraint generative adversarial network provided by the application comprises:

[0051] S1, collecting optical data and multi-beam sonar data, generating a joint calibration file, and performing preprocessing.

[0052] S1 comprises:

[0053] S101, collect optical data and multi-beam sonar data; the optical data is a polarized filter RGB-D video stream, and the multi-beam sonar data is riverbed datum elevation data.

[0054] For example, 6 waterproof cameras are deployed to build a ring array (focal length 12 mm, baseline distance 1.5 m), and a polarized filter RGB-D video stream (resolution 3840x2160@30fps) is synchronously collected and stored as an H.265 encoded MP4 file. In addition, riverbed datum elevation data is obtained by a high-frequency multi-beam sonar (frequency 1 MHz, beam opening angle 120°).

[0055] In addition, in S101, a camera pose drift compensation model needs to be established, R represents the pose rotation change matrix of the camera in the time interval t, T represents the displacement of the camera in the time interval t:

[0056] ,

[0057] where the angular velocity ω and the 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 datum elevation data with the optical data using UTC timestamps to generate a joint calibration file.

[0059] In S102, the format of the joint calibration file is JSON. In addition, time synchronization of the sonar and optical data needs to be achieved by PTP protocol clock synchronization, and the master-slave clock deviation compensation amount δt≤1ms is set; spatial calibration adopts a checkerboard target method, and after initial external parameter calibration is completed in a dry environment, underwater reprojection error optimization (RMSE≤5cm) is used to verify the validity of the calibration.

[0060] S103, perform water body scattering compensation on each frame of image using a dark channel prior dehazing (Dark Channel Prior) algorithm.

[0061] Specifically, guided filtering is used for transmittance estimation to reduce the influence of water scattering, which can effectively preserve image edge details while smoothing noise, the guide image is set as the weighted average image of the RGB channel (weight coefficient [0.299, 0.587, 0.114]), the filtering radius r = 7 pixels, and the regularization parameter ε = 0.001. In addition, the dark channel prior dehazing algorithm is used to perform water scattering compensation on each frame of image, the atmospheric light value A = [0.8, 0.9, 0.95] × 255, and the transmittance map calculation window size is 15 × 15 pixels.

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

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

[0064] S105, performing polarization fusion processing.

[0065] Specifically, the Stokes vector S = [I0, I45, I90, I135 is calculated, and the effective polarization component is extracted and saved as a 16-bit TIFF format. In addition, the effective polarization component calculation needs to consider the Mueller matrix correction of the Stokes vector:

[0066]

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

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

[0069] S2, according to the data obtained in S1, performing multi-source data fusion processing to obtain a training set and a test set.

[0070] In S2, a space-time compensation mechanism based on inertial navigation is constructed: real-time six-degree-of-freedom motion parameters are obtained through an IMU module (sampling rate 200 Hz), and a coordinate system dynamic compensation model is established:

[0071]

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

[0073] A motion distortion compensation network is constructed, which inputs time-series IMU data and point cloud sequences and outputs deformation field (x,y,z,t), the network structure includes a space-time encoder and a deformation predictor. The space-time encoder is a 3-layer GRU network (hidden layer 128 dimensions) for extracting motion features; the deformation predictor is a fully connected layer for outputting 3D displacement, and the constraint condition is After compensation, the data satisfies the ICP matching error of adjacent frames of point clouds ≤3mm and the IMU integral position residual error ≤5cm.

[0074] S2 includes:

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

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

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

[0078] Specifically, the local coordinate system is converted to the UTM Zone 50N coordinate system, and the height reference is unified to the EGM2008 geoid. In this process, real-time six-degree-of-freedom motion parameters are obtained through an IMU module (sampling rate 200 Hz), and a coordinate system dynamic compensation model is established to reduce measurement errors caused by device movement, and a motion distortion compensation network is further 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 by the training set is enhanced data.

[0080] S204, using the data from S1, combining the coordinate normalization processing result in S202, a training set containing optical data, sonar data and true value label is constructed.

[0081] Specifically, the training set includes optical data (RGB-D sequence resolution 1920x1080), sonar data (point cloud density >=500pts / m2), and true value label (ground laser scanning point cloud, accuracy 0.5mm).

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

[0083] Specifically, the test set includes CFD simulation parameters (Reynolds number Re [1x10^4, 5x10^5]) and turbulence intensity parameters. The CFD simulation parameters include a non-Newtonian fluid model, which uses the Herschel-Bulkley constitutive equation: .

[0084] In S205, in order to simulate the change of the optical properties of the water body in the real environment, the application generates turbidity simulation noise by using the Monte Carlo light propagation model to enhance the spatial and optical data, randomly rotates (angle range ±15°) and translates (offset ±0.3m), and adds turbidity simulation noise (NTU value 30-100 corresponding to scattering coefficient β=0.6-1.2).

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

[0086] S3, a physically constrained generative adversarial network is constructed, and the training set is trained according to the training set; the physically constrained generative adversarial network includes a generator network and a discriminator network.

[0087] S4, input the test set into the physically constrained generative adversarial network, obtain the point cloud data, and perform engineering verification.

[0088] Embodiment 2:

[0089] Based on embodiment 1, the application embodiment 2 provides a more specific bridge scour pit three-dimensional reconstruction method based on the physically constrained generative adversarial network, which includes:

[0090] S1, collect optical data and multi-beam sonar data, generate a joint calibration file, and perform preprocessing.

[0091] S2, according to the data obtained in S1, multi-source data fusion processing is performed to obtain a training set and a test set.

[0092] S3, a physically constrained generative adversarial network is constructed, and the training set is used for training; the physically constrained generative adversarial network includes a generator network and a discriminator network.

[0093] S3 includes:

[0094] S301, the network parameters are initialized.

[0095] In S301, the generator input size is set to 256x256x32 (corresponding to the actual space 2m x 2m x 0.5m), and the discriminator adopts the PatchGAN architecture (5 layers of three-dimensional convolution, kernel size 4x4x4).

[0096] S302, the generator network and the discriminator network are constructed.

[0097] Specifically, as shown in Figure 1 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 is composed of the following modules:

[0098] Generator: responsible for generating preliminary point cloud data from input noise point cloud.

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

[0100] 3D convolution network: multi-scale feature extraction is performed on the feature vector, including 4-level down-sampling 3DResNet blocks (channel number [64, 128, 256, 512] per level), each residual block contains a bottleneck structure, which helps to reduce the calculation complexity and speed up the training process.

[0101] MLP generator: directly generates more accurate point cloud details from the latent space vector, so that the final output is closer to the actual terrain features.

[0102] The discriminator network is composed of the following modules:

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

[0104] Residual block: introduces a skip connection to alleviate the gradient vanishing problem in deep networks, further optimizing feature extraction.

[0105] Fully connected layer: integrates feature information of different levels and outputs final discrimination result.

[0106] In addition, the 4-level down-sampling 3DResNet block (channel number of each level [64, 128, 256, 512]) described above is used as an encoder-decoder with a cross-modal attention module, and its operation formula is as follows:

[0107]

[0108] wherein, represents the final feature representation after processing by the cross-modal attention module, is a 3x3 convolution operation, used for spatial convolution operation on the input feature, represents the feature representation extracted from the image data, is an activation function, used to scale the attention score to the range of [0, 1], and are weight matrices, which are query matrix and key matrix respectively, represents the feature representation extracted from the point cloud data.

[0109] In addition, the 3DResNet block adopts a bottleneck structure, and each residual block contains a 1x1x1 down-sampling convolution, a 3x3x3 main convolution, and a 1x1x1 up-sampling convolution. The skip connection adds a channel attention mechanism.

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

[0111] Dynamic weight matrix calculation:

[0112] Feature recombination formula:

[0113] wherein the dynamic adjustment factor is: ,

[0114] is the output feature, is the image feature, is the point cloud feature, and σ represents the Sigmoid function. The depth separable convolution DWConv kernel size is 5x5x5, and the expansion rate d=2.

[0115] S303, to ensure physical authenticity, the learning process of the generative adversarial network is constrained by using the Navier-Stokes equation, and a multi-constraint loss function is configured; the multi-constraint loss function includes an adversarial loss, a geometric consistency loss, and a fluid constraint loss.

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

[0117]

[0118] where is the magnitude of the fluid velocity, c is the speed of sound, representing the speed at which sound waves propagate in the fluid, and Δx = 0.01 m is the spatial discretization step, from which the time step Δt = 0.001 s is determined.

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

[0120]

[0121] where 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 gravitational acceleration; is the pressure gradient term; represents the spatial second-order derivative of the velocity field.

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

[0123]

[0124] . Where, The adversarial loss weight = 1.0, indicating the relative importance of the adversarial loss in the total loss function; is the adversarial loss; is the geometric consistency loss weight = 0.5; is the geometric consistency loss; is the fluid constraint loss weight = 2.0; is the Navier-Stokes equation residual, representing 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 to the physical constraint generative adversarial network, obtain point cloud data, and perform engineering verification.

[0126] S4 includes:

[0127] S401, input the test set image to the generator to obtain point cloud data.

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

[0129] S402, CFD fluid verification is performed.

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

[0131] S403, the error distribution is evaluated using the Bootstrap resampling method.

[0132] Specifically, the Bootstrap resampling method is introduced to evaluate the error distribution, and a large number of repeated sampling (such as 1000 times with replacement) is used to estimate the error confidence interval: the 95% CI of MAE should meet [1.8 cm, 2.2 cm], and the normality of the error distribution is verified (Kolmogorov-Smirnov test p>0.05). For example, the precision evaluation adopts the calculation of the average point cloud error (compared with the ground laser scanning data) and the relative error of shear stress (compared with the CFD simulation benchmark value), which meets the index: MAE≤2 cm, ≤12%.

[0133] In addition, the shear stress calculation adopts the wall function method, and the height of the first layer of near-wall grid meets y+≈30, and the calculation formula is:

[0134]

[0135] 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. the position of y=0).

[0136] To verify the effectiveness of the method, comparative experiments (see Table 1) were conducted between the method and representative prior art under the same test environment (turbidity 100 NTU, flow rate 3 m / s, sediment particle size 0.2 mm), and the data in Table 1 are the average values of 5 independent experiments.

[0137] Table 1: Comparison of scour pit reconstruction performance of different methods under turbidity 100 NTU (test conditions: flow rate 3 m / s, sediment particle size 0.2 mm)

[0138]

[0139] Compared with the traditional sonar reconstruction method, the method reduces the mean absolute error (MAE) by 78% (1.8 cm vs 8.2 cm) and the prediction error of the key hydraulic parameter shear stress by 77% (9.5% vs 42.1%). This is mainly due to the fusion of multi-modal data (high-resolution RGB-D and polarization information) and the powerful feature extraction capability of the deep learning model, which significantly overcomes the limitations of sparse sonar point clouds and optical scattering interference. Compared with the generative adversarial network (GAN) without introducing physical constraints, the MAE of the method is reduced by 66% (1.8 cm vs 5.3 cm), and by 70% (9.5% vs 31.7%). This fully proves the effectiveness of introducing the Navier-Stokes equation as a physical constraint loss term (weight = 2.0), which significantly improves the physical reasonableness of the reconstructed terrain and the prediction accuracy of the flow field, avoiding the artifacts that may be produced by purely data-driven methods that violate the laws of fluid mechanics. At the same time, the method successfully outputs high-density point clouds (≥2000 pts / m²), providing sufficient data basis for depicting the subtle topographic features of the scour pit (such as local vortex pits and steep slopes), which is much higher than the traditional sonar method and ensures physical authenticity. In summary, the method significantly surpasses the existing technology in reconstruction accuracy (MAE), key engineering parameter prediction accuracy (shear stress), and detail restoration capability (point cloud density) (3-5 times improvement), fully verifying its ability to perform high-precision and high-reliability three-dimensional reconstruction of bridge scour pits in high-turbidity complex underwater environments.

[0140] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be mutually referenced, and will not be described again in this application.

[0141] Embodiment 3

[0142] Based on Embodiment 2, the bridge scour pit three-dimensional reconstruction system based on a physically constrained generative adversarial network is provided in Embodiment 3 of the present application, which comprises:

[0143] The acquisition module is configured to acquire optical data and multi-beam sonar data, generate a joint calibration file, and perform preprocessing.

[0144] The fusion module is configured to perform multi-source data fusion processing according to the data obtained by the acquisition module to obtain a training set and a test set.

[0145] The construction module is configured to construct a physically constrained generative adversarial network and train the network according to the training set. The physically constrained generative adversarial network comprises a generator network and a discriminator network.

[0146] The verification module is configured to input the test set into the physical constraint generative adversarial network, obtain point cloud data, and perform engineering verification.

[0147] It should be noted that the system provided in this embodiment corresponds to the method provided in Embodiment 2, and therefore, in this embodiment, the same or similar parts as Embodiment 2 can be cross-referenced, and will not be described herein again.

Claims

1. A method for 3D reconstruction of bridge scour craters based on physical constraint generative adversarial networks, characterized in that, include: S1. Acquire optical data and multibeam sonar data, generate a joint calibration file, and perform preprocessing; S2. Based on the data obtained in S1, perform multi-source data fusion processing to obtain the training set and test set; S3. Construct a physical constraint generative adversarial network and train it based on the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network; S3 includes: S301. Initialize network parameters; S302. Construct the generator network and the discriminator network; S303. The learning process of the generative adversarial network is constrained by the Navier-Stokes equations, and a multi-constraint loss function is configured; the multi-constraint loss function includes adversarial loss, geometric consistency loss and fluid constraint loss. S4. Input the test set into the physical constraints to generate an adversarial network, obtain point cloud data, and perform engineering verification.

2. The method for three-dimensional reconstruction of bridge scour craters based on physical constraint generative adversarial networks according to claim 1, characterized in that, S1 includes: S101, Acquire optical data and multibeam sonar data; The optical data is an RGB-D video stream with polarization filtering, and the multibeam sonar data is riverbed reference elevation data; S102. Align the riverbed baseline elevation data and optical data using UTC timestamps to generate a joint calibration file; S103. Use the dark channel prior dehazing algorithm to perform water scattering compensation on each frame of the image; 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 method for three-dimensional reconstruction of bridge scour craters based on physical constraint generative adversarial networks according to claim 2, characterized in that, S2 include: S201. Obtain the preset dataset; S202. Perform coordinate unification processing on the optical data and multibeam sonar data acquired in S1. S203. Turbidity simulation noise generated based on the Monte Carlo ray propagation model is used to enhance the optical data and multibeam sonar data obtained in S1. S204. Obtain ground truth labels through field measurements or high-precision simulations, and construct a training set containing optical data, sonar data, and ground truth labels; S205. Construct a test set that includes CFD simulation parameters and turbulence intensity parameters.

4. The method for three-dimensional reconstruction of bridge scour craters based on physical constraint generative adversarial networks according to claim 3, characterized in that, S4 include: S401. Input the test set image into the generator to obtain point cloud data; S402, Perform CFD fluid validation; S403. Use the Bootstrap resampling method to evaluate the error distribution.

5. A 3D reconstruction system for bridge scour craters based on physical constraint generative adversarial networks, characterized in that, For performing the method according to any one of claims 1 to 4, comprising: The acquisition module is used to acquire optical data and multibeam sonar data, generate joint calibration files, and perform preprocessing. The fusion module is used to perform multi-source data fusion processing based on the data obtained from the acquisition module to obtain training and test sets; A construction module is used to construct a physical constraint generative adversarial network and train it based on the training set; the physical constraint generative adversarial network includes a generator network and a discriminator network; The verification module is used to generate an adversarial network by inputting the test set with physical constraints, obtain point cloud data, and perform engineering verification.

6. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 4.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 4.

Images