High-temporal-and-spatial-resolution refined flow field reconstruction method, device, equipment and medium

Abstract

The application discloses a high-temporal-spatial-resolution refined flow field reconstruction method and device, equipment and medium, relates to the ocean current reconstruction technical field, and the method comprises the following steps: normalizing and spatiotemporal alignment of satellite remote sensing, buoy observation and numerical simulation data; based on the aligned data, performing a pre-calculation on an unstructured nested grid through an FVCOM model, and then dynamically encrypting the grid according to the flow field gradient and generating a background flow field; inputting the background flow field into a PINN-GAN joint framework, outputting a refined flow field through physical constraint loss and double-discriminator adversarial training; using a heterogeneous accelerator to schedule a computing task, verifying the reconstructed refined flow field in real time and feeding back optimization. Through the generation of the background flow field and the refined reconstruction, the application can efficiently and accurately reconstruct a high-temporal-spatial-resolution ocean current flow field.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ocean current reconstruction, and in particular to a high-temporal-and-spatial-resolution refined flow field reconstruction method, device, equipment and medium. BACKGROUND

[0002] Traditional interpolation and dimension reduction methods have insufficient accuracy and loss of high-frequency characteristics. Interpolation algorithms (such as bilinear / bicubic interpolation) only achieve resolution enhancement through local pixel weighted averaging, and cannot recover high-frequency details (such as shock waves and vortex structures) in the flow field. In complex nonlinear flow (such as transonic shock wave interaction and turbulence), the error is significant, and the mean square error of pressure coefficient prediction is high. Dynamic mode decomposition (DMD / POD) relies on linear or weak nonlinear assumptions, and requires 300+ modes to capture most of the energy in high Reynolds number turbulence, resulting in a dimension disaster. Interpolation methods ignore the conservation laws of fluid mechanics (such as mass / momentum conservation), resulting in reconstructed flow fields that do not meet the constraints of the N-S equation, with high prediction bias for shock wave location. Dimension reduction models (such as POD) cannot accurately describe the high-frequency attenuation characteristics of the turbulent kinetic energy spectrum under the assumption of eddy viscosity.

[0003] However, data-driven methods have the following problems: limited model generalization ability, computational efficiency and real-time bottleneck, and insufficient physical interpretability. The hardware and multi-source data fusion challenge has the problem of high cost. SUMMARY

[0004] Therefore, the embodiments of the present application provide a high-temporal-and-spatial-resolution refined flow field reconstruction method, device, equipment and medium to efficiently and accurately reconstruct a high-temporal-and-spatial-resolution refined flow field.

[0005] An aspect of the embodiments of the present application provides a high-temporal-and-spatial-resolution refined flow field reconstruction method, which includes the following steps:

[0006] Normalizing and spatiotemporally aligning satellite remote sensing, buoy observation and numerical simulation data;

[0007] Based on the aligned data, performing a pre-calculation on an unstructured nested grid through an FVCOM model, and then dynamically encrypting the grid according to the flow field gradient and generating a background flow field;

[0008] Inputting the background flow field into a PINN-GAN joint framework, and outputting a refined flow field through physical constraint loss and double-discriminator adversarial training;

[0009] Using a heterogeneous accelerator to schedule a computing task, verifying the reconstructed refined flow field in real time and feeding back optimization.

[0010] In some embodiments, the pre-calculation on the unstructured nested grid through the FVCOM model based on the aligned data includes the following steps:

[0011] The horizontal grid and the vertical surface layer are nested to perform grid layering.

[0012] In some embodiments, the step of dynamically encrypting the grid according to the flow field gradient and generating the background flow field comprises the following steps:

[0013] The grid encryption operation is triggered by the flow field gradient threshold, and the FVCOM model is used to generate the background flow field on the encrypted unstructured nested grid.

[0014] In some embodiments, the step of inputting the background flow field into the PINN-GAN joint framework and outputting the refined flow field through physical constraint loss and double discriminator adversarial training comprises the following steps:

[0015] The momentum equation and the continuity equation residual are added to the loss function to implement the physical law constraint.

[0016] The 3D convolution kernel is used to process the spatial vortex feature, and the Bi-LSTM network is used to process the tidal time series feature.

[0017] In some embodiments, the step of scheduling the computing task by using the heterogeneous accelerator comprises the following steps:

[0018] The FVCOM coarse grid parallel computing is run by the CPU cluster, the PINN-GAN is deployed by the GPU node to realize single-frame reconstruction delay < 50 ms, and the real-time ADCP data stream is processed by the FPGA coprocessor.

[0019] In some embodiments, the step of inputting the background flow field into the PINN-GAN joint framework and outputting the refined flow field through physical constraint loss and double discriminator adversarial training comprises the following steps:

[0020] The high-resolution flow field features are extracted by the teacher model, the knowledge distillation is compressed to the student model of the MobileNet-v3 architecture, and the uncertainty sampling is implemented by the Monte Carlo Dropout.

[0021] In some embodiments, the step of real-time verifying the reconstructed refined flow field and feeding back the optimization comprises the following steps:

[0022] The quantitative evaluation index is calculated by comparing the reconstructed refined flow field with the independent observation data set, and the parameters of the FVCOM model are optimized according to the index result.

[0023] Another aspect of the embodiments of the application also provides a high-spatial-temporal resolution refined flow field reconstruction device, which comprises:

[0024] A preprocessing unit is configured to normalize and spatio-temporally align satellite remote sensing, buoy observation and numerical simulation data.

[0025] a flow field generation unit configured to perform pre-calculation on an unstructured nested grid based on the aligned data by using an FVCOM model, and further generate a background flow field by dynamically refining the grid according to a flow field gradient;

[0026] a flow field reconstruction unit configured to input the background flow field into a PINN-GAN joint framework, and output a refined flow field by physical constraint loss and double discriminator adversarial training;

[0027] an optimization unit configured to schedule a computing task by using a heterogeneous accelerator, and verify the reconstructed refined flow field in real time and feed back optimization.

[0028] Another aspect of the embodiment of the present application further provides an electronic device, including a processor and a memory.

[0029] The memory is configured to store a program.

[0030] The processor executes the program to implement the method described in any of the above.

[0031] Another aspect of the embodiment of the present application further provides a computer readable storage medium, which stores a program, and the program is executed by a processor to implement the method described in any of the above.

[0032] The present application at least has the following beneficial effects:

[0033] The present application can normalize and align in space and time satellite remote sensing, buoy observation and numerical simulation data; based on the aligned data, pre-calculation is performed on an unstructured nested grid by using an FVCOM model, and further a background flow field is generated by dynamically refining the grid according to a flow field gradient; the background flow field is input into a PINN-GAN joint framework, and a refined flow field is output by physical constraint loss and double discriminator adversarial training; a computing task is scheduled by using a heterogeneous accelerator, and the reconstructed refined flow field is verified in real time and feedback optimization is fed back. By generating a background flow field and reconstructing in detail, the present application can efficiently and accurately reconstruct a high spatio-temporal resolution ocean current flow field. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Figure 1 A flow chart of a high spatio-temporal resolution refined flow field reconstruction method provided by the embodiment of the present application is shown in the figure.

[0036] Figure 2Data input and preprocessing schematic diagram provided for the embodiments of the present application;

[0037] Figure 3 Dynamic verification and feedback optimization schematic diagram provided for the embodiments of the present application;

[0038] Figure 4 Structure block diagram of the high-spatiotemporal resolution refined flow field reconstruction device provided for the embodiments of the present application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0040] Before the embodiments of the present application are described in detail, first, some related technologies involved in the embodiments of the present application are described as follows:

[0041] The demand for high-spatiotemporal resolution flow field data is growing in fluid mechanics research, especially in the fields of aerospace, weather prediction, energy systems, etc. Traditional experimental measurement techniques (such as particle image velocimetry, PIV) are limited by hardware costs (such as high-speed camera storage space, laser energy) and physical constraints (such as sampling frequency, spatial resolution), making it difficult to meet the measurement needs of high temporal resolution (such as 5000 Hz or higher) and high spatial resolution (such as micron-level structure capture) at the same time. Numerical simulation (such as CFD) can provide full flow field information, but it consumes a lot of computing resources and relies on simplified assumptions, making it difficult to fully reproduce the complexity of real flow.

[0042] Existing technical solutions mainly fall into two categories: ① Methods based on traditional interpolation and dimensionality reduction: such as bilinear interpolation, proper orthogonal decomposition (POD), but rely on local data fitting, and cannot capture the global spatiotemporal correlation and nonlinear characteristics of the flow field (such as vortices, shock waves). ② Methods combining data-driven and physical constraints: In recent years, deep learning (such as CNN, LSTM, GAN) has been introduced into flow field reconstruction, which learns the mapping relationship between low-resolution and high-resolution data through end-to-end learning, and combines physical equations (such as Navier-Stokes equation) to constrain the model output.

[0043] Existing technical solutions:

[0044] (1) Traditional interpolation and dimensionality reduction methods.

[0045] Bilinear / bicubic interpolation: based on local pixel weighted average, simple to calculate but low reconstruction accuracy (PSNR<30dB), unable to recover high-frequency details.

[0046] Dynamic Mode Decomposition (DMD): Extracts dominant modes of flow field through matrix decomposition, but only suitable for linear or weakly nonlinear systems, poor applicability for multi-scale flows such as turbulence.

[0047] (2) Super-resolution reconstruction based on deep learning.

[0048] Spatio-temporal feature fusion model:

[0049] CNN-LSTM hybrid architecture: Utilize Convolutional Neural Network (CNN) to extract spatial features (such as vortex structure boundary layer), combined with Long Short-Term Memory Network (LSTM) to capture time evolution law (such as atmospheric circulation period). For example, SRCNN model realizes flow field resolution enhancement through 3 layers of convolution.

[0050] Attention mechanism enhancement: Introduce spatial / temporal attention weights in feature extraction module, focus on key areas (such as shock wave surface) and key time steps, reduce vortex detail reconstruction error.

[0051] Physics-Informed GAN:

[0052] Generator (such as ResNet) maps low-resolution input to high-resolution flow field, discriminator optimizes generated result authenticity through adversarial training.

[0053] Physics loss function: The residual of Navier-Stokes equation is used as a regularization term to constrain the velocity field to satisfy mass and momentum conservation.

[0054] (3) Multi-source data fusion and real-time interaction system.

[0055] Sensor network and deep learning fusion: BiGRU (Bidirectional Gated Recurrent Unit): Joint training of high spatial resolution PIV data (10 Hz) and local high-frequency sensors (5000 Hz), after dimensionality reduction by deep convolutional autoencoder, BiGRU reconstructs the high time resolution features of the full flow field, realizing the global reconstruction of cross-scale flow (such as wind tunnel test Re=2.7×10^4).

[0056] Three-dimensional tomographic PIV system: Integrates multiple high-speed cameras (such as Wolf PIV, 10 kHz frame rate) and laser light sources (200 mJ pulse energy), reconstructs three-dimensional particle distribution through GPU acceleration, spatial resolution up to 23.28 Pixel / mm45.

[0057] Real-time interactive visualization platform: Application integration of VR / AR, combined with gesture recognition, voice commands and flow field dynamic rendering, supports multi-view observation (such as aircraft flow vortex evolution).

[0058] (4) Spatial resolution optimization for time-series data.

[0059] Full convolutional time-series network: For single-point time-series signals (such as wind speed sensors), design one-dimensional convolutional layers to extract time-series features, and combine spatial interpolation layers to generate high-resolution flow field snapshots.

[0060] Technical limitations and challenges:

[0061] Traditional methods: Interpolation algorithms cannot handle nonlinear features, and dimension reduction models rely on high-resolution training data.

[0062] Deep learning models: High computational complexity (e.g., GAN training time-consuming), lack of physical rationality (e.g., pressure field prediction bias).

[0063] Hardware limitations: High-cost PIV systems, large-scale sensor deployment difficult to achieve.

[0064] Drawbacks of existing technology:

[0065] I. Traditional interpolation and dimension reduction methods:

[0066] Traditional interpolation and dimension reduction methods have insufficient accuracy and loss of high-frequency features. Interpolation algorithms (such as bilinear / trilinear interpolation) only achieve resolution enhancement through local pixel weighted averaging, and cannot recover high-frequency details in the flow field (such as shock waves and vortex structures). In complex nonlinear flow (such as transonic shock wave interaction and turbulence), the error is significant, and the mean square error of pressure coefficient prediction is high. Dynamic mode decomposition (DMD / POD) relies on linear or weak nonlinear assumptions, and requires 300+ modes to capture most of the energy in high Reynolds number turbulence, leading to a computational dimension disaster. Interpolation methods ignore fluid mechanics conservation laws (such as mass / momentum conservation), resulting in reconstructed flow fields that do not meet N-S equation constraints, with high shock wave position prediction bias. Dimension reduction models (such as POD) cannot accurately describe the high-frequency attenuation characteristics of turbulent kinetic energy spectrum under the assumption of eddy viscosity.

[0067] II. Data-driven methods:

[0068] (1) Limited model generalization ability:

[0069] CNN / LSTM architecture: Strong dependence on training data, significant feature loss when migrating across scenarios (such as aircraft engine flow field to ocean circulation), and possible increase in root mean square error.

[0070] Generative adversarial network (GAN): Generator and discriminator training is difficult to converge, with high sensitivity to hyperparameters, and overfitting weights when migrating to hypersonic flow fields, resulting in decreased PSNR reconstruction.

[0071] (2) Bottlenecks in computational efficiency and real-time performance:

[0072] The single-frame delay of super-resolution models (e.g., SRGAN, DenseNet) is >100 ms when processing 1024x1024 flow fields, which cannot meet the real-time processing requirements of 5000 Hz high-frequency data. Physical constraint models (e.g., PINN) need to optimize data loss and equation residual at the same time, which increases the training time by several times compared to pure data-driven models.

[0073] (3) Lack of physical interpretability:

[0074] Pure data-driven models (e.g., CNN) learn features through black-box mapping and cannot guarantee that the output meets the quality conservation and vorticity evolution rules, resulting in large prediction bias of pressure field. Physics-informed neural networks (e.g., Physics-Informed GAN) introduce N-S equation constraints, but the boundary condition embedding is imperfect, leading to high prediction error of shear layer velocity.

[0075] Three, hardware and multi-source data fusion challenges:

[0076] (1) High cost of sensor deployment:

[0077] High-frequency PIV systems (e.g., 10 kHz camera of Wolf-Eye) require multiple high-speed cameras, which are expensive and have high complexity in three-dimensional calibration.

[0078] (2) Difficulty in handling heterogeneity of multi-source data:

[0079] Sparse sensors (e.g., 5000 Hz cobra sensor) have large differences in spatial and temporal resolution compared to PIV data, and need to be aligned after dimensionality reduction, resulting in loss of vortex structure phase information. In the reconstruction of multi-physical field (velocity / temperature / component) coupling, the feature distribution of heterogeneous data does not match, and the GAN discriminator needs to be processed in different paths, increasing the number of model parameters.

[0080] Four, real-time interaction and visualization bottlenecks:

[0081] Large-scale flow field rendering delay: When traditional WebGL platforms process 10^6 particle data, VR interaction delay is >100 ms, which cannot achieve multi-view dynamic response within 50 ms. Poor cross-platform compatibility: Existing systems (e.g., RFlow3D3C) have insufficient support for heterogeneous data formats (e.g., CFD mesh, PIV vector field), which require manual preprocessing and are time-consuming.

[0082] The core contradiction of the prior art is the "resolution-efficiency-mechanism" triangular constraint: increasing the model complexity (such as GAN multi-discriminator) is needed to improve the accuracy, but the real-time performance is sacrificed; embedding physical constraints can improve the mechanism rationality, but the training data requirement increases exponentially (such as PINN requires 10^5 DNS samples); hardware cost compression depends on sensor sparsification, but interpolation error accumulation is introduced. Future breakthroughs need to focus on: multi-scale physical prior knowledge encoding, lightweight heterogeneous computing architecture, and cross-modal data distillation technology, so as to realize mechanism rational high-resolution reconstruction without sacrificing real-time performance.

[0083] Reference Figure 1 The embodiment of the application provides a high spatio-temporal resolution refined flow field reconstruction method, which specifically comprises the following steps S100-S130:

[0084] S100: normalizing and spatio-temporal aligning satellite remote sensing, buoy observation and numerical simulation data;

[0085] S110: based on the aligned data, performing pre-calculation on an unstructured nested grid by an FVCOM model, and then dynamically encrypting the grid according to the flow field gradient and generating a background flow field;

[0086] S120: inputting the background flow field into a PINN-GAN joint framework, and outputting a refined flow field through physical constraint loss and double-discriminator adversarial training;

[0087] S130: scheduling a computing task by using a heterogeneous accelerator, verifying the reconstructed refined flow field in real time and feeding back optimization.

[0088] Optionally, the pre-calculation on the unstructured nested grid by the FVCOM model based on the aligned data comprises the following steps:

[0089] The nested grid is layered by using horizontal grid and vertical surface layer.

[0090] Optionally, the dynamically encrypting the grid according to the flow field gradient and generating the background flow field comprises the following steps:

[0091] The grid encryption operation is triggered by a flow field gradient threshold, and the FVCOM model is used to generate the background flow field on the encrypted unstructured nested grid.

[0092] Optionally, the inputting the background flow field into the PINN-GAN joint framework, and outputting the refined flow field through the physical constraint loss and the double-discriminator adversarial training comprises the following steps:

[0093] The physical law constraint is implemented by adding momentum equation and continuity equation residual to the loss function;

[0094] The spatial vortex feature is processed by using a 3D convolution kernel, and the tidal time series feature is processed by using a Bi-LSTM network.

[0095] Optionally, the scheduling of the computing task by using the heterogeneous accelerator comprises the following steps:

[0096] The FVCOM coarse grid parallel computing is run by the CPU cluster, the PINN-GAN is deployed on the GPU node to realize single-frame reconstruction delay of <50 ms, and the real-time ADCP data stream is processed by the FPGA coprocessor.

[0097] Optionally, the input of the background flow field into the PINN-GAN joint framework comprises the following steps:

[0098] The high-resolution flow field features are extracted by the teacher model, the knowledge distillation is compressed to the student model of the MobileNet-v3 architecture, and the uncertainty sampling is implemented by the Monte Carlo Dropout.

[0099] Optionally, the real-time verification of the reconstructed refined flow field and the feedback optimization comprise the following steps:

[0100] The quantitative evaluation index is calculated by comparing the reconstructed refined flow field with the independent observation data set, and the parameters of the FVCOM model are fed back according to the index results.

[0101] Next, specific application examples will be combined to introduce and illustrate the scheme of the embodiments of the present application in detail.

[0102] The system is based on a three-dimensional unstructured grid FVCOM model, combined with a physical information neural network (PINN) and a generative adversarial network (GAN) multi-discriminator architecture, and realizes the refined spatiotemporal reconstruction of the ocean flow field through multi-scale physical constraints, heterogeneous computing optimization and cross-modal data distillation technology. The system architecture includes the following core modules:

[0103] I. Multi-scale nested grid and FVCOM model preprocessing

[0104] (1) Horizontal grid nesting design:

[0105] Coarse-fine grid layering: triangular fine grid (resolution ≤500 m) is used in key areas such as the coast and straits, and coarse grid (resolution ≥5 km) is used in the open sea area, and seamless nesting is realized through the unstructured grid technology of FVCOM.

[0106] Dynamic encryption strategy: based on the flow field gradient (such as vorticity, velocity shear), the grid density is adjusted in real time, for example, local encryption is triggered in the frontal or vortex area.

[0107] Design principles and operation steps:

[0108] Horizontal grid nesting realizes regional adaptive refinement through unstructured triangular mesh to balance computational efficiency and accuracy. The core idea is that coarse mesh covers the whole area and fine mesh focuses on key areas. The specific process is as follows:

[0109] ① Data preparation and boundary definition.

[0110] Topographic data: Import the coastline and water depth data in GIS format (such as ETOPO or measured data), and generate the initial mesh through Delaunay triangulation.

[0111] Key area demarcation: According to the research target (such as nearshore, strait, vortex area), demarcate the high-resolution area and generate the boundary point set.

[0112] ② Coarse-fine mesh layering implementation.

[0113] Global coarse mesh: Set the triangular mesh with a resolution of ≥5 km to cover the offshore area, reducing the amount of calculation.

[0114] Local fine mesh: Use high-resolution mesh with ≤500 m in nearshore, strait and other areas to realize seamless nesting using the unstructured mesh characteristics of FVCOM.

[0115] ③ Dynamic refinement strategy.

[0116] Gradient triggering mechanism: Real-time monitoring of flow field parameters (vorticity, velocity gradient), when the gradient exceeds the threshold (|∇u|>0.1s⁻¹), trigger local refinement.

[0117] Nested boundary processing: Set two rows of overlapping units (double-row node method) in the transition area of fine-coarse mesh to avoid flow velocity discontinuity.

[0118] Practical points and optimization.

[0119] Mesh quality check: Evaluate mesh quality through aspect ratio (<3), minimum internal angle (>30°) and other indicators, and optimize node position using GMS / SMS software.

[0120] Nested parameter configuration: Set NCNEST_ON=T in the namelist file of FVCOM, and control memory and disk write frequency through NCNEST_BLOCKSIZE.

[0121] (2) Vertical σ coordinate optimization.

[0122] Layered adaptive adjustment: Set high-resolution σ layers (10 layers) in the surface layer (0-50 m), and use exponentially decreasing layers in the deep layer (50 m-seabed), considering boundary layer capture and computational efficiency.

[0123] Terrain coupling mechanism: Import seabed DEM through GIS data, correct grid node elevation, ensure the accuracy of complex terrain (such as seamounts, trenches) flow field simulation.

[0124] Layering strategy and operation process:

[0125] Vertical sigma coordinate optimizes vertical resolution by following terrain characteristics, operation steps as follows:

[0126] ①σ layer definition and terrain coupling:

[0127] Layering formula: Adopt σ = (z - η) / (H + η), where z is the vertical coordinate, η is the free surface height, H is the water depth.

[0128] Adaptive layering design:

[0129] Surface high resolution: 0-50 m set 10 layers of σ coordinate (equal interval or exponential distribution), capture boundary layer and mixed layer process.

[0130] Deep optimization: 50 m to the seabed adopts exponential decreasing layering (such as σ=0.1, 0.3, 0.5, 0.7, 0.9), reduce calculation redundancy.

[0131] ②Terrain data processing:

[0132] Seafloor DEM fusion: Convert GIS terrain data to NetCDF format compatible with FVCOM, and correct grid node elevation through fvcom_prep tool.

[0133] ③Vertical parameterization configuration.

[0134] Mixing scheme selection: Enable Mellor-Yamada 2.5 order turbulence closure model in the surface layer, and adopt KPP scheme in the deep layer (need to set MELLOR_PROFILE=T in namelist).

[0135] Dry-wet grid processing: Dynamically mark dry-wet cells through critical water depth threshold (such as 0.2 m), avoid simulation distortion in intertidal zone.

[0136] Key optimization techniques:

[0137] Terrain roughness calibration: According to the measured flow rate data, the bottom friction coefficient (Manning coefficient n=0.02-0.05) is inverted, and the σ layer bottom boundary condition is corrected.

[0138] Vertical interpolation optimization: Adopt cubic spline interpolation for temperature and salinity initial field, reduce numerical diffusion caused by layering.

[0139] (3) Process integration and verification.

[0140] Full-process implementation steps:

[0141] Data input and preprocessing: integrate terrain, boundary forcing (tides, wind field) and initial field data.

[0142] Grid generation and nesting: generate nested grid through SMS or GMS, export as FVCOM's *.dat format.

[0143] Vertical layer configuration: set σ layer parameters and mixing scheme in fvcom_nml.

[0144] Model running and tuning: accelerate simulation through MPI parallel computing (recommended CPU+GPU heterogeneous architecture) and real-time monitoring of conservation error.

[0145] Verification indicators and cases:

[0146] Horizontal verification: compare the error of vortex center position between satellite altimeter and reconstructed flow field (target <2 km).

[0147] Vertical verification: use Argo float temperature and salinity profile data to evaluate layering accuracy (RMSE <0.5℃, salinity <0.2 PSU).

[0148] II. Embedding of physical constraints and PINN-GAN collaborative framework.

[0149] (1) Multi-physical field PINN model.

[0150] Control equation constraint: embed the residual term of FVCOM momentum equation and continuity equation in the loss function, forcing the velocity field to satisfy mass and momentum conservation.

[0151] Boundary condition fusion: use satellite altimeter and buoy observation data as Dirichlet boundary conditions, balance the contributions of data-driven and physical constraints through adaptive weights.

[0152] Multi-physical field PINN (Physics-Informed Neural Network) embeds fluid control equations (such as N-S equations) into the neural network loss function, forcing the model output to satisfy physical laws. Its core idea is to constrain both data-driven and physical mechanisms to solve the mechanism deviation problem of pure data models in sparse data areas. Operation steps and implementation process: (a) Multi-physical field control equation embedding Equation selection and discretization:

[0153] ;

[0154] Residual loss function construction, define physical residual term as regularization term added to total loss function:

[0155] ;

[0156] L data : Data-driven loss (e.g., velocity field MSE).

[0157] L phy : Mean square sum of equation residuals (momentum, continuity, temperature-salinity equations).

[0158] Adaptive weight adjustment: dynamically adjust λ phy during training phase, focusing on data fitting initially and strengthening physical constraints later.

[0159] (b) Boundary conditions and multi-source data fusion.

[0160] Boundary condition processing:

[0161] Dirichlet boundary: satellite altimeter data (sea surface height) as boundary conditions, directly constrain output layer node values.

[0162] Neumann boundary: gradient information (e.g., temperature vertical gradient) of buoy observations is calculated through automatic differentiation and embedded into the loss function.

[0163] Multi-source data alignment and dimensionality reduction:

[0164] Temporal and spatial interpolation: align low-frequency satellite data (1 Hz) and high-frequency buoy data (5000 Hz) to a unified time step through cubic spline interpolation.

[0165] Feature fusion layer: add a fully connected layer after the neural network input layer to map heterogeneous data (velocity, temperature, salinity) to a unified hidden space.

[0166] (c) Network structure and training optimization.

[0167] Network architecture design:

[0168] Main network: use ResNet-50 as the basic structure to alleviate gradient vanishing using residual connections.

[0169] Multi-task branch: velocity field branch: output three-dimensional velocity components (u, v, w).

[0170] Temperature and salinity branch: output temperature (T) and salinity (S) profiles.

[0171] Adaptive activation function: use Swish function (x⋅σ(x)) to enhance non-linear expression ability.

[0172] Training strategy:

[0173] Stage training:

[0174] Pre-training stage: only Ldata is used for training, and the observed data is quickly fitted.

[0175] Physical fine-tuning stage: Lphy is added, and λphy is gradually increased to 0.5.

[0176] Optimizer selection: AdamW (weight decay optimization) combined with cosine annealing learning rate scheduling.

[0177] (2) GAN multi-discriminator enhancement.

[0178] Space-time dual-path discriminator:

[0179] Spatial discriminator (D1): 3D convolution kernel (7x7x7) extracts spatial features such as vortices and fronts, and supervises the consistency of the generator output with high-resolution FVCOM simulation results.

[0180] Temporal discriminator (D2): LSTM network captures time series patterns such as tidal period and mesoscale vortex evolution, and constrains the physical reasonableness of flow field evolution.

[0181] Adversarial loss optimization: the generator (G) adopts U-Net structure, and improves the detail fidelity of the generated flow field through Wasserstein distance loss.

[0182] Traditional GAN single discriminator is difficult to supervise space-time features simultaneously, and the system adopts space-time dual discriminator architecture, focusing on spatial details and temporal evolution rules of flow field respectively.

[0183] The implementation process and key technologies of space-time dual-path discriminator are as follows:

[0184] (a) Generator (Generator) design U-Net structure optimization.

[0185] Encoder: 5 layers of convolution (kernel size 3x3x3), step 2, extract multi-scale features.

[0186] Decoder: 5 layers of transpose convolution, add jump connection to transfer bottom details.

[0187] Attention gate: introduce spatial attention module in jump connection, focus on key areas such as vortices and fronts.

[0188] Multi-resolution input fusion:

[0189] Low-resolution FVCOM field: as the main input (512x512x32).

[0190] High-resolution local observation: compressed by 1x1 convolution and spliced with the main input.

[0191] (b) Discriminator design.

[0192] Spatial discriminator (D1):

[0193] 3D convolutional network: 5 layers of convolution (kernel size 7x7x7), output spatial feature confidence map.

[0194] Spectral normalization: constrain discriminator Lipschitz continuity, improve training stability.

[0195] Temporal discriminator (D2):

[0196] Bi-LSTM architecture: bidirectional LSTM captures temporal evolution, outputs time series plausibility score. Window sliding mechanism: 30 frames as time window, assesses physical consistency of flow field evolution. Adversarial loss function: Wasserstein GAN-GP loss:

[0197] ;

[0198] Multi-discriminator collaborative optimization:

[0199] ;

[0200] (c) Training process and tuning:

[0201] Alternating training strategy:

[0202] Discriminator update: fix generator, jointly train D1, D2 with real and generated data.

[0203] Generator update: fix discriminator, deceive D1, D2 with generated data, while optimizing PINN physical constraints.

[0204] Gradient penalty (GP) application:

[0205] Randomly interpolate sampling x^ between real and generated data, calculate gradient penalty term to prevent mode collapse.

[0206] Dynamic weight adjustment:

[0207] Adjust the loss weights of D1, D2 according to the discriminator confidence, avoid the dominance of a certain discriminator in training.

[0208] Collaborative framework integration and verification.

[0209] ① Joint training process.

[0210] Data preprocessing: align FVCOM simulation data, satellite remote sensing, buoy observations to a unified spatio-temporal grid.

[0211] Initial pre-training: Train PINN model alone until data loss converges (about 200 epochs).

[0212] GAN adversarial training: Freeze PINN encoder, jointly optimize generator and discriminator (500+ epochs).

[0213] End-to-end fine-tuning: Unfreeze all parameters, fine-tune model with Ltotal = Ldata + Lphy + LGAN.

[0214] ② Validation metrics.

[0215] Quantitative metrics: PSNR (Peak Signal-to-Noise Ratio), vortex capture rate (compared to satellite altimeter vortex database).

[0216] Qualitative evaluation: Frontal details: Generated fields can resolve 1 km scale frontal structures (traditional methods blur).

[0217] Temporal evolution: Tidal cycle phase error (verified with observed data).

[0218] ③ Performance optimization techniques.

[0219] Mixed precision training: Use FP16 / FP32 mixed precision to reduce GPU memory usage.

[0220] Knowledge distillation: Compress the full model into a lightweight version for edge device inference delay.

[0221] Key challenges and solutions:

[0222] Physical constraints and data conflicts: The problem is that the equation residual loss may deviate from the observed data. The solution is to introduce uncertainty quantification (Bayesian PINN) and dynamically adjust the strength of physical constraints.

[0223] Multi-discriminator training instability: The problem is that D1 and D2 have inconsistent convergence speeds. The solution is to use Two-Time-Scale Update Rule (TTUR) to set different learning rates for D1 and D2.

[0224] Large computational resource consumption: The problem is that 3D convolution requires high memory. The solution is to use gradient checkpointing technology to sacrifice speed for memory savings.

[0225] Three, lightweight heterogeneous computing and data distillation.

[0226] (1) Task-level heterogeneous acceleration architecture.

[0227] Division of computing units:

[0228] CPU cluster: Run FVCOM coarse grid global simulation, handle open boundary conditions and tidal forcing through MPI parallelization.

[0229] GPU node: Deploy PINN-GAN model, utilize CUDA to accelerate tensor operations, single-frame flow field reconstruction delay <50 ms.

[0230] FPGA co-processor: Process real-time sensor data (e.g., ADCP profile flow velocity), implement low-power edge computing.

[0231] Task-level heterogeneous acceleration architecture design.

[0232] Design principles and hardware division:

[0233] Task-level heterogeneous acceleration matches task requirements with computing unit characteristics, maximizing resource utilization efficiency. The core idea is as follows:

[0234] CPU cluster: Process global simulation tasks with high parallelism and low computational density (e.g., FVCOM coarse mesh solution).

[0235] GPU node: Accelerate computationally intensive tasks (e.g., PINN-GAN model training, 3D flow field rendering).

[0236] FPGA co-processor: Process real-time tasks with low latency and high determinism (e.g., sensor data preprocessing, edge inference.

[0237] Implementation steps and key technologies:

[0238] (a) Task partitioning and scheduling strategy.

[0239] Task classification:

[0240] Computationally intensive: GAN generator forward propagation, PINN automatic differentiation (allocated to GPU).

[0241] Communication-intensive: MPI parallel computing of FVCOM (allocated to CPU cluster).

[0242] Real-time sensitive: ADCP flow velocity analysis, data denoising (allocated to FPGA).

[0243] Dynamic resource scheduling:

[0244] Kubernetes orchestration: Define computing task types (e.g., job-type: gpu-inference) through custom resources (CRD), automatically allocate to corresponding nodes.

[0245] Priority queue: Real-time data streams (e.g., 5000 Hz sensor) have priority in occupying GPU resources, ensuring delay <50 ms.

[0246] (b) Hardware co-optimization technology.

[0247] CPU-GPU heterogeneous communication:

[0248] Zero-Copy Memory: CUDA Unified Memory enables zero-copy data transfer between CPU and GPU, reducing data migration overhead between FVCOM and PINN.

[0249] Pipeline parallelism: The FVCOM simulation is broken down into time-step pipelines, where the GPU processes the current step while the CPU calculates the boundary conditions for the next step.

[0250] FPGA edge acceleration:

[0251] Customized IP cores: Design dedicated preprocessing pipelines (such as Kalman filtering and outlier detection) for sensor data, reducing latency to 5 μs.

[0252] Low-precision quantization: Converts the floating-point model to an 8-bit fixed-point number (INT8), reducing power consumption while maintaining low inference precision error.

[0253] (2) Cross-modal data distillation technology.

[0254] Feature decoupling and transfer:

[0255] Teacher model: A deep residual network is trained based on high-resolution FVCOM output to extract multi-scale flow field features (global circulation and local turbulence).

[0256] Student model: The features of the teacher model are compressed into the lightweight MobileNet-v3 through knowledge distillation, reducing the number of parameters while maintaining PSNR>35 dB.

[0257] Active learning strategy: Based on uncertainty quantification (such as Monte Carlo Dropout), key areas are dynamically selected to supplement observations, reducing data acquisition costs.

[0258] Technical principles and core ideas:

[0259] Cross-modal data distillation, through knowledge transfer and feature decoupling, compresses the knowledge of a high-precision model (teacher) into a lightweight model (student), while fusing multi-source heterogeneous data (satellites, buoys, simulations). The core challenge lies in the differences in feature distribution and semantic alignment between modalities.

[0260] Implementation process and key technologies.

[0261] (a) Teacher-student model construction:

[0262] Teacher model design:

[0263] Multi-modal fusion backbone: DenseNet-201 as the base architecture, input layer is processed differently for different modalities, satellite remote sensing: 3D convolution processing of height-time series (e.g. sea surface temperature SST).

[0264] Buoy data: LSTM processing time series flow rate.

[0265] FVCOM simulation: graph convolutional network (GCN) processing unstructured grid data.

[0266] Feature alignment loss: constraint consistency of different modal features in the hidden space through contrastive learning (Contrastive Loss).

[0267] Student model design:

[0268] Lightweight architecture: based on MobileNet-v3, channel number reduced to 1 / 4 of the teacher model.

[0269] Multi-scale attention: introduce lightweight CBAM module (channel + spatial attention), focus on key areas (e.g. vortex center).

[0270] (b) Knowledge distillation and transfer.

[0271] Feature distillation strategy:

[0272] Hidden layer feature matching: minimize the KL divergence of the intermediate layer features of the teacher and student models:

[0273]

[0274] where T l , S l are the teacher and student features of the l-th layer, respectively.

[0275] Relationship distillation: guide the student model to learn feature correlation through the relationship matrix (Gram matrix) of the teacher model feature map.

[0276] Cross-modal data alignment:

[0277] Temporal and spatial interpolation alignment: use cubic spline interpolation to unify satellite data (1 km resolution) and buoy point data to FVCOM grid.

[0278] Adversarial domain adaptation: add a domain classifier in the teacher model, eliminate the distribution difference between modalities through a gradient reversal layer (GRL).

[0279] (c) Active learning and incremental training.

[0280] Uncertainty sampling:

[0281] Monte Carlo Dropout: Calculate prediction variance by multiple forward propagation, and select the region with the largest variance to supplement the observation.

[0282] Boundary-sensitive sampling: Preferentially collect data points with large flow field gradients (such as frontal surfaces and vortex edges).

[0283] Incremental training strategy:

[0284] Elastic weight solidification: Protect important parameters of existing knowledge, and only fine-tune the network part related to new data.

[0285] Online distillation: Use new data inference results as pseudo labels to dynamically update student models.

[0286] Verification and performance optimization.

[0287] (a) Performance indicators:

[0288] Model compression rate and accuracy retention: PSNR decrease <0.5 dB, vortex center positioning error increase <0.1 km.

[0289] Inference speed: Single-frame processing time <20 ms on edge devices (such as Jetson Xavier).

[0290] (b) Optimization techniques.

[0291] Dynamic channel pruning: Dynamically close redundant channels according to neuron activation intensity, and accelerate inference.

[0292] Quantization-aware training: Simulate INT8 quantization errors during training to improve model robustness after deployment.

[0293] Full-process integration and challenge response:

[0294] Collaborative workflow:

[0295] Data preprocessing: Temporal and spatial alignment of multi-source data → feature encoding → storage in distributed database (such as Redis).

[0296] Heterogeneous computing scheduling: Kubernetes allocates resources according to task labels, priority: real-time tasks > training tasks > simulation tasks.

[0297] Distillation and inference: Teacher model offline training → knowledge transfer to student model → edge device deployment.

[0298] Core challenges and solutions:

[0299] Heterogeneous hardware synchronization scheme: Use NCCL communication library to realize high-speed communication between GPUs, and use RDMA to reduce delay between CPU and GPU.

[0300] Inter-modal semantic gap solution: Introduce cross-modal contrastive learning (e.g. CLIP style) to construct a unified semantic space.

[0301] Edge device resource limitation solution: Use TensorRT optimized inference engine combined with INT8 quantization and layer fusion techniques.

[0302] Four, system implementation process and verification.

[0303] (1) Full process steps:

[0304] a. Data input: Fusion satellite remote sensing, buoys, numerical simulation and other multi-source data, normalization and spatio-temporal alignment. Figure 2 For data input and preprocessing diagram.

[0305] b. Coarse grid FVCOM pre-algorithm: Generate background flow field, identify high gradient area to trigger grid encryption.

[0306] c. PINN-GAN joint training: Alternately optimize physical loss and adversarial loss, iterate to convergence.

[0307] d. Heterogeneous computing task scheduling: Through Kubernetes dynamic allocation of CPU / GPU resources, preferentially process real-time data streams.

[0308] e. Dynamic verification and feedback: Compare the reconstruction results with independent observation data sets (such as Argo buoys), calculate RMSE, vortex capture rate and other indicators. Figure 3 For dynamic verification and feedback optimization diagram.

[0309] (2) Performance advantages:

[0310] Accuracy improvement: In the South China Sea vortex case, compared with traditional interpolation methods, this system reduces the speed field RMSE and vortex center positioning error.

[0311] Efficiency optimization: Through heterogeneous computing and model lightweight, the global reconstruction time of 1080x1080 grid is shortened.

[0312] Five, key technology innovation points.

[0313] Physical-data dual-driven fusion: PINN embedded FVCOM control equation solves the mechanism deviation problem of pure data-driven models in data sparse areas.

[0314] Multi-discriminator GAN architecture: Spatial-temporal dual-path supervision significantly improves the capture ability of mesoscale vortex, internal wave and other transient processes.

[0315] Grid-computing collaborative optimization: Dynamic nested grid combined with heterogeneous computing achieves Pareto optimization of resolution and efficiency.

[0316] The scheme includes the following key technical solutions:

[0317] (1) Spatiotemporal dynamic alignment and fusion method of multi-source heterogeneous data.

[0318] A cross-scale spatiotemporal benchmark unification method is proposed to solve the adaptive alignment problem of satellite (low frequency / large range), buoy (high frequency / point) and numerical simulation field (mesoscale / gridded) data. Unlike traditional linear interpolation, a cubic spline time interpolation + inverse distance weighted spatial fusion (IDW) joint algorithm is used to realize the continuous mapping of satellite data (1 day period) and buoy data (5000 Hz sampling rate) at a unified time step (1 second), and dynamic weight distribution is realized. At the same time, Z-score normalization and isolation forest anomaly detection (threshold 3σ) are introduced to reduce data fusion error.

[0319] (2) Dynamic nested grid generation technology based on flow field gradient detection.

[0320] A grid refinement mechanism triggered by flow field gradient (velocity gradient, vorticity) is proposed to break through the limitations of traditional fixed multi-scale grid. The gradient threshold trigger condition is defined, combined with the vertical layering strategy, to realize the adaptive switching of grid resolution from kilometer level (background field) to meter level (vortex area), and the calculation efficiency is improved. Unlike existing methods, a maximum edge length ratio constraint for adjacent grids is proposed to avoid topological distortion of the encrypted area and unstructured grid.

[0321] (3) Joint training architecture of physical information neural network and generative adversarial network (PINN-GAN).

[0322] The Navier-Stokes equation is embedded into the GAN generator to build a PINN-GAN dual-driven model. The physical loss term and dynamic weight adjustment strategy are designed to solve the physical inconsistency problem of traditional data-driven models. The generator uses the U-Net structure, and the discriminator introduces a spatiotemporal 3D convolution kernel (5x5x5) to reduce spatiotemporal discrimination error.

[0323] (4) Dynamic task scheduling method of heterogeneous computing resources (CPU-GPU-FPGA).

[0324] A real-time scheduling algorithm based on task priority and hardware characteristics is proposed. The task allocation rules are defined: CPU processes FVCOM coarse grid simulation (priority 1), GPU executes PINN-GAN inference (priority 2), and FPGA is responsible for ADCP data analysis. A pipeline timing synchronization protocol is developed to realize end-to-end flow field generation with multi-device collaboration, which is faster than traditional CPU clusters.

[0325] (5) Dynamic feedback-driven active sampling and model incremental update mechanism.

[0326] Construct the prediction uncertainty (based on Monte Carlo Dropout) guided buoy path planning algorithm. Define the active sampling radius formula and set the model update trigger condition. Propose a lightweight teacher-student model distillation method to reduce the time consumption of incremental update.

[0327] (6) Cross-domain feature alignment technology of multi-modal contrast learning.

[0328] Design a cross-modal (flow rate, temperature, salinity) contrast loss function to solve the feature space difference between satellite, buoy and simulation data. Use Triplet Loss (anchor = satellite data, positive sample = buoy data, negative sample = simulation background field) to constrain the cosine similarity of feature vectors, realize efficient mapping of multi-source data in a unified hidden space, and improve the feature alignment accuracy.

[0329] (7) Edge-cloud collaborative flow field real-time generation system architecture.

[0330] Propose a collaborative framework of FPGA edge preprocessing and GPU cloud inference. Define ADCP signal analysis dedicated logic unit to support real-time flow rate analysis; deploy multi-GPU parallel inference engine on the cloud, and realize data synchronization through gRPC protocol.

[0331] (8) High-precision vortex and front capture physical constraint enhancement method.

[0332] For mesoscale vortex and temperature front, propose a vorticity conservation constraint term and front gradient enhancement loss. Combine PINN-GAN generator to reduce vortex center positioning error and improve temperature front resolution.

[0333] (9) Bidirectional conversion algorithm of unstructured grid-structured tensor.

[0334] Develop a lossless conversion method between grid and tensor data. Define a sparse mapping matrix from grid nodes to tensors to support efficient conversion between FVCOM unstructured grid and PINN-GAN input tensors, reducing data preprocessing time.

[0335] (10) High-frequency PIV-RANS fusion diagnosis technology.

[0336] Expand the method to industrial scenarios and propose a fusion scheme of high-frequency PIV (10 kHz) and RANS simulation. Design a dynamic turbulent viscosity correction model combined with GAN super-resolution reconstruction to realize high-precision capture of special flow fields.

[0337] Referring to Figure 4 Another aspect of the embodiments of the present application also provides a high spatiotemporal resolution refined flow field reconstruction device, which comprises:

[0338] a preprocessing unit configured to normalize and spatiotemporally align satellite remote sensing, buoy observation and numerical simulation data;

[0339] a flow field generation unit configured to perform a pre-calculation on an unstructured nested grid based on the aligned data by using a FVCOM model, and further dynamically refine the grid according to a flow field gradient and generate a background flow field;

[0340] a flow field reconstruction unit configured to input the background flow field into a PINN-GAN joint framework, and output a refined flow field through physical constraint loss and double discriminator adversarial training;

[0341] an optimization unit configured to schedule a computing task by using a heterogeneous accelerator, verify the reconstructed refined flow field in real time, and feed back optimization.

[0342] It can be understood that the contents in the above method embodiments are all applicable to the device embodiments, the device embodiments specifically implement the functions of the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0343] In some alternative embodiments, the functions / operations mentioned in the block diagram can not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two blocks shown in succession can actually be executed substantially simultaneously or the blocks can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example, with the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of a larger operation are independently executed.

[0344] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features can be integrated in a single physical device and / or software module, or one or more functions and / or features can be implemented in separate physical devices or software modules. It can also be understood that a detailed discussion of the actual implementation of each module is unnecessary for an understanding of the present application. Rather, given the properties, functions and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be within the routine skill of the engineer, given the teachings of the present application. Thus, the present application, as set forth in the claims, can be realized without undue experimentation using ordinary skill in the art. It can also be understood that the disclosed specific concepts are merely illustrative and are not intended to limit the scope of the present application, which is defined by the full scope of the appended claims and their equivalents.

[0345] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0346] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logic functions, which can be embodied in any computer readable medium for use by or in connection with an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, apparatus or device and execute the instructions, or in conjunction with these instruction execution systems, apparatus or devices. For the purpose of this specification, "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by or in connection with an instruction execution system, apparatus or device, or in conjunction with these instruction execution systems, apparatus or devices.

[0347] More specific examples (non-exhaustive list) of the computer readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CD ROM). In addition, the computer readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be obtained electronically, for example, by optical scanning of the paper or other medium, followed by editing, interpreting or otherwise processing, if necessary, in other suitable ways, to be electronically obtained and then stored in the computer memory.

[0348] It should be understood that portions of the present application can be realized with hardware, software, firmware or a combination thereof. In the foregoing embodiments, a number of steps or methods can be realized as software or firmware to be executed by a suitable instruction-executing system. For example, if realized with hardware, and as in another embodiment, any one or a combination of the following technologies known in the art can be used: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0349] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0350] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the claims and their equivalents.

[0351] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the described embodiments, and those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present application, and these equivalent modifications or substitutions are included in the scope defined by the claims of the present application.

Claims

1. A method for high spatiotemporal resolution refined flow field reconstruction, characterized in that, The method includes the following steps: Normalize and align satellite remote sensing, buoy observation, and numerical simulation data in a spatiotemporal manner; Based on the aligned data, pre-calculation is performed on the unstructured nested mesh using the FVCOM model, and then the mesh is dynamically refined and the background flow field is generated according to the flow field gradient. The background flow field is input into the PINN-GAN joint framework, and the refined flow field is output through adversarial training with physical constraint loss and dual discriminators. Heterogeneous accelerators are used to schedule computational tasks, and the reconstructed refined flow field is verified in real time and feedback is provided for optimization. The method of scheduling computing tasks using heterogeneous accelerators includes the following steps: FVCOM coarse-grid parallel computation is run through a CPU cluster, PINN-GAN is deployed through GPU nodes to achieve a single-frame reconstruction latency of <50 ms, and real-time ADCP data streams are processed through an FPGA coprocessor. Define task allocation rules: CPUs with priority 1 handle FVCOM coarse mesh simulation, GPUs with priority 2 perform PINN-GAN inference, and FPGAs are responsible for ADCP data parsing; develop pipeline timing synchronization protocols to achieve end-to-end flow field generation through multi-device collaboration; The real-time verification and reconstruction of the refined flow field and its feedback optimization include the following steps: Quantitative evaluation metrics are calculated by comparing the reconstructed refined flow field with independent observation datasets, and the parameters of the FVCOM model are optimized based on the metric results.

2. The high spatiotemporal resolution refined flow field reconstruction method according to claim 1, characterized in that, The pre-calculation based on aligned data and using the FVCOM model on an unstructured nested mesh includes the following steps: Nested mesh layering is performed using horizontal meshes and vertical surface layers.

3. The high spatiotemporal resolution refined flow field reconstruction method according to claim 2, characterized in that, The process of dynamically refining the mesh and generating a background flow field based on the flow field gradient includes the following steps: The mesh refinement operation is triggered by the flow field gradient threshold, and the background flow field is generated on the refined unstructured nested mesh using the FVCOM model.

4. The high spatiotemporal resolution refined flow field reconstruction method according to claim 1, characterized in that, The process of inputting the background flow field into the PINN-GAN joint framework and training a refined flow field through physical constraint loss and adversarial training with dual discriminators includes the following steps: Physical constraints are implemented by adding momentum equations and continuity equations to the residuals of the loss function; 3D convolutional kernels are used to process spatial vortex features, and Bi-LSTM networks are used to process tidal time series features.

5. The high spatiotemporal resolution refined flow field reconstruction method according to claim 1, characterized in that, The process of inputting the background flow field into the PINN-GAN joint framework and training a refined flow field through physical constraint loss and adversarial training with dual discriminators includes the following steps: High-resolution flow field features are extracted using the teacher model, compressed into the student model of the MobileNet-v3 architecture using knowledge distillation, and uncertainty sampling is implemented using Monte Carlo Dropout.

6. A device for high spatiotemporal resolution refined flow field reconstruction, characterized in that, The apparatus is applied to the high spatiotemporal resolution refined flow field reconstruction method as described in claim 1, and the apparatus comprises: The preprocessing unit is used to normalize and spatiotemporally align satellite remote sensing, buoy observation, and numerical simulation data; The flow field generation unit is used to perform pre-calculation on an unstructured nested mesh based on aligned data and the FVCOM model, and then dynamically refine the mesh and generate the background flow field according to the flow field gradient. The flow field reconstruction unit is used to input the background flow field into the PINN-GAN joint framework and output a refined flow field through adversarial training with physical constraint loss and dual discriminators. The optimization unit is used to schedule computational tasks using heterogeneous accelerators, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 5.

Images