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

Abstract

The invention discloses a high-temporal-spatial-resolution refined flow field reconstruction method, device and equipment and a medium, and relates to the technical field of ocean current reconstruction, and the method comprises the steps: carrying out the normalization and temporal-spatial alignment of satellite remote sensing, buoy observation and numerical simulation data; based on the alignment data, performing rehearsal on the 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 combined framework, and outputting a refined flow field through physical constraint loss and double-discriminator adversarial training; and scheduling a calculation task by adopting a heterogeneous accelerator, verifying the reconstructed refined flow field in real time, and performing feedback optimization. Through generation of the background flow field and refinement reconstruction, the ocean current flow field with high temporal-spatial resolution can be reconstructed efficiently and accurately.

Description

Technical Field

[0001] The present application relates to the field of ocean current reconstruction technology, in particular to methods, devices, equipment and media for fine-grained flow field reconstruction with high temporal and spatial resolution. Background Art

[0002] Traditional interpolation and dimensionality reduction methods suffer from insufficient accuracy and loss of high-frequency features. Interpolation algorithms (such as bilinear and bicubic interpolation) only achieve resolution improvement through local pixel weighted averaging and are unable to recover high-frequency details in the flow field (such as shock waves and vortex structures). Errors are significant in complex nonlinear flows (such as transonic shock-wave interactions and turbulence), and the mean square error of pressure coefficient prediction is high. Dynamic mode decomposition (DMD / POD) relies on linear or weak nonlinear assumptions. For high-Reynolds number turbulence, 300+ modes are required to capture the majority of energy, leading to the computational curse of dimensionality. Interpolation methods ignore fluid dynamics conservation laws (such as mass and momentum conservation), resulting in reconstructed flow fields that do not conform to the NS equations and high deviations in shock wave position prediction. Dimensionality 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 approaches suffer from the following issues: limited model generalization, computational efficiency and real-time bottlenecks, and insufficient physical interpretability. Hardware and multi-source data fusion challenges are also associated with high costs. Summary of the Invention

[0004] In view of this, the embodiments of the present application provide a method, apparatus, device and medium for reconstructing a refined flow field with high temporal and spatial resolution, so as to efficiently and accurately reconstruct a refined flow field with high temporal and spatial resolution.

[0005] An embodiment of the present application provides a method for reconstructing a flow field with high temporal and spatial resolution, the method comprising the following steps:

[0006] Normalize and align satellite remote sensing, buoy observation and numerical simulation data in time and space;

[0007] Based on the aligned data, the FVCOM model is used to perform pre-calculations on the unstructured nested grid, and then the grid is dynamically refined according to the flow field gradient to generate the background flow field.

[0008] The background flow field is input into the PINN-GAN joint framework, and the refined flow field is output through physical constraint loss and dual discriminator adversarial training;

[0009] Heterogeneous accelerators are used to schedule computing tasks, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

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

[0011] Nested grid layering is performed using horizontal grids and vertical surface layers.

[0012] In some embodiments, the method 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 on the encrypted unstructured nested grid to generate the background flow field.

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

[0015] Implement physical law constraints by adding momentum equation and continuity equation residuals to the loss function;

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

[0017] In some embodiments, the method of scheduling computing tasks using a heterogeneous accelerator includes the following steps:

[0018] The FVCOM coarse grid parallel computation is run on a CPU cluster, PINN-GAN is deployed on GPU nodes to achieve a single-frame reconstruction delay of less than 50 ms, and the real-time ADCP data stream is processed by an FPGA coprocessor.

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

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

[0021] In some embodiments, the real-time verification of the reconstructed refined flow field and feedback optimization includes the following steps:

[0022] Quantitative evaluation indicators are calculated by comparing the reconstructed refined flow field with independent observation data sets, and the parameters of the FVCOM model are optimized based on the feedback of the indicator results.

[0023] Another aspect of the present application further provides a high temporal and spatial resolution refined flow field reconstruction device, the device comprising:

[0024] Preprocessing unit, used to normalize and align satellite remote sensing, buoy observation and numerical simulation data in time and space;

[0025] The flow field generation unit is used to perform pre-calculations on the unstructured nested grid through the FVCOM model based on the aligned data, and then dynamically refine the grid according to the flow field gradient and generate the background flow field;

[0026] The flow field reconstruction unit is used to input the background flow field into the PINN-GAN joint framework and output the refined flow field through physical constraint loss and dual discriminator adversarial training;

[0027] The optimization unit is used to schedule computing tasks using heterogeneous accelerators, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

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

[0029] The memory is used to store programs;

[0030] The processor executes the program to implement any of the above methods.

[0031] Another aspect of the embodiments of the present application further provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement any of the above methods.

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

[0033] This application normalizes and spatially aligns satellite remote sensing, buoy observation, and numerical simulation data. Based on the aligned data, the FVCOM model performs pre-calculations on an unstructured nested grid, dynamically refining the grid based on the flow gradient and generating a background flow field. This background flow field is then fed into a joint PINN-GAN framework, where it outputs a refined flow field through physical constraint loss and dual-discriminator adversarial training. Using heterogeneous accelerators to schedule computational tasks, the reconstructed refined flow field is verified in real time and optimized through feedback. By generating a background flow field and refining its reconstruction, this application can efficiently and accurately reconstruct ocean currents with high spatiotemporal resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0035] Figure 1 A schematic diagram of the process of the high spatiotemporal resolution refined flow field reconstruction method provided in an embodiment of the present application;

[0036] Figure 2Schematic diagram of data input and preprocessing provided in the embodiment of the present application;

[0037] Figure 3 A schematic diagram of dynamic verification and feedback optimization provided in an embodiment of the present application;

[0038] Figure 4 This is a structural block diagram of the high temporal and spatial resolution refined flow field reconstruction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

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

[0041] The demand for high-resolution flow field data in both temporal and spatial order is growing in fluid mechanics research, particularly in aerospace, weather forecasting, and energy systems. Traditional experimental measurement techniques (such as particle image velocimetry (PIV)) are limited by hardware costs (e.g., high-speed camera storage space, laser energy) and physical constraints (e.g., sampling frequency, spatial resolution), making it difficult to simultaneously meet the measurement requirements of high temporal resolution (e.g., above 5000 Hz) and high spatial resolution (e.g., capturing micron-scale structures). While numerical simulations (e.g., CFD) can provide full flow field information, they consume significant computational resources and rely on simplifying assumptions, making it difficult to fully replicate the complexity of real flows.

[0042] Existing technical solutions fall into two main categories: ① Methods based on traditional interpolation and dimensionality reduction, such as bilinear interpolation and proper orthogonal decomposition (POD), rely on local data fitting and are unable to capture the global spatiotemporal correlations and nonlinear characteristics of the flow field (such as vortices and shock waves). ② Methods combining data-driven and physical constraints: In recent years, deep learning (such as CNN, LSTM, and GAN) has been introduced into flow field reconstruction. These methods learn the mapping relationship between low-resolution and high-resolution data end-to-end and incorporate physical equations (such as the Navier-Stokes equations) to constrain model output.

[0043] Existing technical solutions:

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

[0045] Bilinear / bicubic interpolation: Based on local pixel weighted averaging, it is computationally simple but has low reconstruction accuracy (PSNR < 30dB) and cannot restore high-frequency details.

[0046] Dynamic Mode Decomposition (DMD): Extracts the main modes of the flow field through matrix decomposition, but it is only applicable to linear or weakly nonlinear systems and has poor applicability to multi-scale flows such as turbulence.

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

[0048] Spatiotemporal feature fusion model:

[0049] CNN-LSTM hybrid architecture: Utilizes a convolutional neural network (CNN) to extract spatial features (such as vortex structure boundary layers) and combines it with a long short-term memory (LSTM) network to capture temporal evolution patterns (such as atmospheric circulation cycles). For example, the SRCNN model achieves improved flow field resolution through three layers of convolution.

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

[0051] Physics-Informed GAN:

[0052] The generator (such as ResNet) maps low-resolution input to a high-resolution flow field, and the discriminator optimizes the authenticity of the generated results through adversarial training.

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

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

[0055] Fusion of sensor networks and deep learning: BiGRU (Bidirectional Gated Recurrent Unit): Utilizing joint training of high-spatial-resolution PIV data (10 Hz) and local high-frequency sensors (5000 Hz), and after dimensionality reduction through a deep convolutional autoencoder, the BiGRU reconstructs the high-temporal-resolution features of the entire flow field, enabling global reconstruction of cross-scale flows (e.g., wind tunnel test Re=2.7×10^4).

[0056] 3D tomographic PIV system: Integrates multiple high-speed cameras (such as Thousand Eyes Wolf PIV, 10 kHz frame rate) and a laser light source (200 mJ pulse energy), and uses GPU to accelerate 3D particle distribution reconstruction with a spatial resolution of 23.28 pixels / mm45.

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

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

[0059] Fully convolutional time series network: For single-point time series signals (such as wind speed sensors), a one-dimensional convolution layer is designed to extract time series features, and combined with a spatial interpolation layer to generate high-resolution flow field snapshots.

[0060] Technical limitations and challenges:

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

[0062] Deep learning models: high computational complexity (e.g., time-consuming GAN training) and insufficient physical rationality (e.g., bias in pressure field prediction).

[0063] Hardware limitations: High-frequency PIV systems are expensive, and large-scale sensor deployment is difficult to achieve.

[0064] Disadvantages of existing technology:

[0065] 1. Traditional interpolation and dimensionality reduction methods:

[0066] Traditional interpolation and dimensionality reduction methods suffer from insufficient accuracy and loss of high-frequency features. Interpolation algorithms (such as bilinear and bicubic interpolation) only achieve resolution improvement through local pixel weighted averaging and are unable to recover high-frequency details in the flow field (such as shock waves and vortex structures). Errors are significant in complex nonlinear flows (such as transonic shock-wave interactions and turbulence), and the mean square error of pressure coefficient prediction is high. Dynamic mode decomposition (DMD / POD) relies on linear or weak nonlinear assumptions. For high-Reynolds number turbulence, 300+ modes are required to capture the majority of energy, leading to the computational curse of dimensionality. Interpolation methods ignore fluid dynamics conservation laws (such as mass and momentum conservation), resulting in reconstructed flow fields that do not conform to the NS equations and high deviations in shock wave position prediction. Dimensionality 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.

[0067] 2. Data-driven approach:

[0068] (1) Limited model generalization ability:

[0069] CNN / LSTM architecture: It is highly dependent on training data, suffers from significant feature loss when migrating across scenarios (e.g., aircraft engine flow field to ocean circulation), and may increase the root mean square error.

[0070] Generative Adversarial Network (GAN): The adversarial training between the generator and the discriminator is difficult to converge, has high hyperparameter sensitivity, and overfitting weights when migrating to hypersonic flow fields leads to a decrease in the reconstruction PSNR.

[0071] (2) Computational efficiency and real-time bottlenecks:

[0072] Super-resolution models (such as SRGAN and DenseNet) experience single-frame latency exceeding 100 ms when processing 1024×1024 flow fields, making them unable to meet the real-time processing requirements of high-frequency data at 5000 Hz. Physically constrained models (such as PINN) require simultaneous optimization of data loss and equation residuals, making training several times more time-consuming than purely data-driven models.

[0073] (3) Insufficient physical interpretability:

[0074] Purely data-driven models (such as CNNs) learn features through black-box mapping, but cannot guarantee that the output conforms to mass conservation and vorticity evolution laws, resulting in large deviations in pressure field predictions. While physical-informed neural networks (such as Physics-Informed Generative Adversarial Networks) incorporate NS equation constraints, they incompletely embed boundary conditions, leading to high errors in shear layer velocity predictions.

[0075] 3. Challenges of Hardware and Multi-Source Data Fusion:

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

[0077] High-frequency PIV systems (such as the Thousand Eyes Wolf 10 kHz camera) require multiple high-speed cameras, which results in high cost per set of equipment and high complexity in three-dimensional calibration.

[0078] (2) Difficulty in processing heterogeneous multi-source data:

[0079] Sparse sensors (such as 5000 Hz Cobra sensors) have significantly different spatiotemporal resolutions than PIV data, requiring dimensionality reduction and alignment during fusion, which results in loss of phase information about vortex structures. Coupled reconstruction of multi-physics fields (velocity / temperature / composition) results in mismatched feature distributions in heterogeneous data, requiring separate processing paths in the GAN discriminator, increasing the number of model parameters.

[0080] 4. Bottlenecks of real-time interaction and visualization:

[0081] Large-scale flow field rendering latency: When traditional WebGL platforms process 10^6 particle data, VR interaction latency exceeds 100ms, making it impossible to achieve multi-view dynamic response within 50ms. Poor cross-platform compatibility: Existing systems (such as RFlow3D3C) lack support for heterogeneous data formats (such as CFD meshes and PIV vector fields), requiring time-consuming manual preprocessing.

[0082] The core contradiction of existing technologies lies in the "resolution-efficiency-mechanism" triangle: improving accuracy requires increasing model complexity (e.g., GAN multi-discriminator), but at the expense of real-time performance; embedding physical constraints can improve the rationality of the mechanism, but leads to an exponential increase in training data requirements (e.g., PINN requires 10^5 DNS samples); hardware cost reduction relies on sensor sparsification, but this introduces cumulative interpolation errors. Future breakthroughs should focus on: encoding multi-scale physical prior knowledge, lightweight heterogeneous computing architectures, and cross-modal data distillation techniques to achieve high-resolution reconstruction with rational mechanisms without sacrificing real-time performance.

[0083] Reference Figure 1 The embodiment of the present application provides a high temporal and spatial resolution refined flow field reconstruction method, which specifically includes the following steps S100 to S130:

[0084] S100: Normalize and align satellite remote sensing, buoy observation, and numerical simulation data in time and space;

[0085] S110: Based on the aligned data, the FVCOM model is used to perform pre-calculations on the unstructured nested grid, and then the grid is dynamically refined according to the flow field gradient to generate the background flow field;

[0086] S120: Input the background flow field into the PINN-GAN joint framework, and output the refined flow field through physical constraint loss and dual discriminator adversarial training;

[0087] S130: Use heterogeneous accelerators to schedule computing tasks, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

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

[0089] Nested grid layering is performed using horizontal grids and vertical surface layers.

[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 the flow field gradient threshold, and the FVCOM model is used on the encrypted unstructured nested grid to generate the background flow field.

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

[0093] Implement physical law constraints by adding momentum equation and continuity equation residuals to the loss function;

[0094] The 3D convolution kernel is used to process the spatial vortex features, and the Bi-LSTM network is used to process the tidal time series features.

[0095] Optionally, the scheduling of computing tasks using a heterogeneous accelerator includes the following steps:

[0096] The FVCOM coarse grid parallel computation is run on a CPU cluster, PINN-GAN is deployed on GPU nodes to achieve a single-frame reconstruction delay of less than 50 ms, and the real-time ADCP data stream is processed by an FPGA coprocessor.

[0097] Optionally, the step of inputting the background flow field into the PINN-GAN joint framework and outputting a refined flow field through physical constraint loss and dual discriminator adversarial training includes the following steps:

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

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

[0100] Quantitative evaluation indicators are calculated by comparing the reconstructed refined flow field with independent observation data sets, and the parameters of the FVCOM model are optimized based on the feedback of the indicator results.

[0101] Next, the solution of the embodiment of the present application will be introduced and explained in detail with reference to specific application examples.

[0102] This system is based on the 3D unstructured grid FVCOM model, combined with the physical information neural network (PINN) and the generative adversarial network (GAN) multi-discriminator architecture, and achieves refined spatiotemporal reconstruction of ocean flow fields through multi-scale physical constraints, heterogeneous computing optimization, and cross-modal data distillation technology. The system architecture includes the following core modules:

[0103] 1. Multi-scale nested grid and FVCOM model preprocessing.

[0104] (1) Horizontal grid nesting design:

[0105] Coarse-fine grid stratification: triangular fine grids (resolution ≤ 500 m) are used in key areas such as nearshore and straits, and coarse grids (resolution ≥ 5 km) are used in offshore areas. Seamless nesting is achieved through FVCOM's unstructured grid technology.

[0106] Dynamic refinement strategy: adjust the grid density in real time based on flow field gradients (such as vorticity and velocity shear), for example, triggering local refinement at the front or vortex area.

[0107] Design principle and operation steps:

[0108] Horizontal mesh nesting achieves regional adaptive densification through unstructured triangular meshes to balance computational efficiency and accuracy. The core idea is: coarse meshes cover the entire area, while fine meshes focus on key areas. The specific process is as follows:

[0109] ①Data preparation and boundary definition.

[0110] Topographic data: Import shoreline 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: Delineate high-resolution areas based on research objectives (such as nearshore, strait, and eddy zones) and generate boundary point sets.

[0112] ② Coarse-fine grid layering implementation.

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

[0114] Local fine grid: A high-resolution grid of ≤500 m is used in nearshore and strait areas, and seamless nesting is achieved by utilizing the unstructured grid characteristics of FVCOM.

[0115] ③Dynamic encryption strategy.

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

[0117] Nested boundary treatment: Two rows of overlapping elements (double-row node method) are set in the fine-coarse grid transition area to avoid velocity discontinuity.

[0118] Practical points and optimization.

[0119] Mesh quality check: Mesh quality is assessed using indicators such as slenderness ratio (<3) and minimum internal angle (>30°), and node positions are optimized using GMS / SMS software.

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

[0121] (2) Vertical σ coordinate optimization.

[0122] Adaptive layering adjustment: a high-resolution σ layer (10 layers) is set for the surface layer (0-50 m), and an exponentially decreasing layer is used for the deep layer (50 m-seabed), taking into account both boundary layer capture and computational efficiency.

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

[0124] Layered strategy and operation process:

[0125] The vertical σ coordinate optimizes the vertical resolution by using the terrain following feature. The operation steps are as follows:

[0126] ① σ layer definition and terrain coupling:

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

[0128] Adaptive layered design:

[0129] High resolution of the surface layer: 10 layers of σ coordinates (equally spaced or exponentially distributed) are set from 0 to 50 m to capture boundary layer and mixing layer processes.

[0130] Deep layer optimization: From 50 m to the seabed, exponentially decreasing layers (e.g., σ = 0.1, 0.3, 0.5, 0.7, 0.9) are used to reduce computational redundancy.

[0131] ② Terrain data processing:

[0132] Seabed DEM fusion: Convert GIS terrain data into FVCOM-compatible NetCDF format and correct the grid node elevation using the fvcom_prep tool.

[0133] ③Vertical parameterized configuration.

[0134] Hybrid scheme selection: Enable the Mellor-Yamada 2.5-order turbulence closed model in the surface layer and use the KPP scheme in the deep layer (MELLOR_PROFILE=T needs to be set in the namelist).

[0135] Wet-dry grid processing: Dynamically mark wet-dry cells using a critical water depth threshold (e.g., 0.2 m) to avoid distortion in intertidal zone simulation.

[0136] Key optimization technologies:

[0137] Terrain roughness calibration: The bottom friction coefficient (Manning coefficient n = 0.02-0.05) was inverted based on the measured flow velocity data, and the bottom boundary condition of the σ layer was corrected.

[0138] Vertical interpolation optimization: Cubic spline interpolation is used for the initial temperature and salinity field to reduce numerical diffusion caused by stratification.

[0139] (3) Process integration and verification.

[0140] Full process implementation steps:

[0141] Data input and preprocessing: integration of topography, boundary forcing (tide, wind field) and initial field data.

[0142] Mesh generation and nesting: Generate nested meshes through SMS or GMS and export them to FVCOM *.dat format.

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

[0144] Model operation and tuning: Accelerate simulation through MPI parallel computing (CPU+GPU heterogeneous architecture is recommended) and monitor quality conservation errors in real time.

[0145] Verification indicators and cases:

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

[0147] Vertical verification: Argo float temperature and salinity profile data were used to evaluate the stratification accuracy (RMSE < 0.5°C, salinity < 0.2 PSU).

[0148] 2. Physical constraint embedding and PINN-GAN collaborative framework.

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

[0150] Control equation constraints: The residual terms of the FVCOM momentum equation and continuity equation are embedded in the loss function to force the velocity field to satisfy mass and momentum conservation.

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

[0152] The multi-physics PINN (Physics-Informed Neural Network) embeds fluid control equations (such as the NS equations) into the neural network loss function, forcing the model output to conform to physical laws. Its core concept is the dual constraints of data-driven and physical mechanisms, addressing the mechanism deviation problem of pure data models in sparse data areas. Operational steps and implementation process: (a) Embedding multi-physics control equations, equation selection, and discretization:

[0153] ;

[0154] The residual loss function is constructed, and the physical residual term is defined as a regular term and added to the total loss function:

[0155] ;

[0156] L data : Data-driven loss (such as velocity field MSE).

[0157] L phy : The mean square sum of the residuals of the equations (momentum, continuity, and temperature-haline equations).

[0158] Adaptive weight adjustment: Dynamically adjust λ according to the training stage phy , initially focusing on data fitting, and later strengthening physical constraints.

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

[0160] Boundary condition processing:

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

[0162] Neumann bound: The gradient information of the buoy observation (such as the vertical temperature gradient) is embedded in the loss function through automatic differentiation.

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

[0164] Spatiotemporal interpolation: Low-frequency satellite data (1 Hz) and high-frequency buoy data (5000 Hz) are aligned to the same time step through cubic spline interpolation.

[0165] Feature fusion layer: A fully connected layer is added after the neural network input layer to map heterogeneous data (speed, temperature, salinity) into a unified latent space.

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

[0167] Network architecture design:

[0168] Backbone network: ResNet-50 is used as the basic structure, and residual connections are used to alleviate gradient disappearance.

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

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

[0171] Adaptive activation function: Use the Swish function (x⋅σ(x)) to enhance nonlinear expression capabilities.

[0172] Training strategy:

[0173] Training in stages:

[0174] Pre-training stage: Only use Ldata for training and quickly fit the observed data.

[0175] Physical fine-tuning stage: add Lphy and gradually increase λphy to 0.5.

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

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

[0178] Spatial-temporal dual-path discriminator:

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

[0180] Temporal Discriminator (D2): The LSTM network captures temporal patterns such as tidal cycles and mesoscale eddy evolution, constraining the physical rationality of flow field evolution.

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

[0182] The traditional GAN ​​single discriminator is difficult to supervise the spatiotemporal features simultaneously. This system adopts a spatial-temporal dual discriminator architecture, focusing on the spatial details and temporal evolution laws of the flow field respectively.

[0183] The implementation process and key technologies of the spatial-temporal dual-path discriminator are as follows:

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

[0185] Encoder: 5 convolutional layers (kernel size 3×3×3), stride 2, extracting multi-scale features.

[0186] Decoder: 5 layers of transposed convolution, with skip connections added to pass low-level details.

[0187] Attention gating: Introducing a spatial attention module in skip connections to focus on key areas such as vortices and fronts.

[0188] Multi-resolution input fusion:

[0189] Low-resolution FVCOM field: used as the main input (512×512×32).

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

[0191] (b) Discriminator design.

[0192] Spatial Discriminator (D1):

[0193] 3D convolutional network: 5 layers of convolution (kernel size 7×7×7), outputting spatial feature confidence maps.

[0194] Spectral normalization: constrains the Lipschitz continuity of the discriminator and improves training stability.

[0195] Temporal Discriminator (D2):

[0196] Bi-LSTM architecture: Bidirectional LSTM captures time series evolution and outputs a time series rationality score. Window sliding mechanism: Uses a 30-frame time window to evaluate the 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 the generator and jointly train D1 and D2 with real data and generated data.

[0203] Generator update: fix the discriminator and deceive D1 and D2 by generating data while optimizing the PINN physical constraints.

[0204] Gradient Penalty (GP) application:

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

[0206] Dynamic weight adjustment:

[0207] The loss weights of D1 and D2 are dynamically adjusted according to the discriminator confidence to avoid a certain discriminator dominating the training.

[0208] Collaborative framework integration and verification.

[0209] ① Joint training process.

[0210] Data preprocessing: Align FVCOM simulation data, satellite remote sensing, and buoy observations to a unified spatiotemporal grid.

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

[0212] GAN adversarial training: freeze the PINN encoder and jointly optimize the generator and discriminator (500+ rounds).

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

[0214] ②Verification indicators.

[0215] Quantitative indicators: PSNR (peak signal-to-noise ratio), vortex capture rate (compared with satellite altimeter vortex database).

[0216] Qualitative assessment: Frontal details: The generated field can resolve 1 km scale frontal structures (which are blurred by traditional methods).

[0217] Time series evolution: tidal cycle phase error (verified by measured data).

[0218] ③Performance optimization technology.

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

[0220] Knowledge distillation: compressing the full model into a lightweight version, reducing inference latency on edge devices.

[0221] Key challenges and solutions:

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

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

[0224] High computational resource consumption: The problem is that 3D convolution leads to high video memory requirements. The solution is to use gradient checkpointing technology, which sacrifices speed in exchange for video memory savings.

[0225] 3. Lightweight heterogeneous computing and data distillation.

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

[0227] Computing unit division of labor:

[0228] CPU cluster: Runs the FVCOM coarse-grid global simulation, handling open boundary conditions and tidal forcing via MPI parallelization.

[0229] GPU nodes: Deploy the PINN-GAN model and use CUDA to accelerate tensor operations, with a single-frame flow field reconstruction delay of less than 50 ms.

[0230] FPGA coprocessor: Processes real-time sensor data (such as ADCP profile flow velocity) to achieve low-power edge computing.

[0231] Task-level heterogeneous acceleration architecture design.

[0232] Design principles and hardware division of labor:

[0233] Task-level heterogeneous acceleration maximizes resource utilization efficiency by matching computing unit characteristics to task requirements. The core ideas are as follows:

[0234] CPU cluster: handles high-parallelism, low-computational-intensity global simulation tasks (such as FVCOM coarse grid solution).

[0235] GPU nodes: Accelerate computationally intensive tasks (such as PINN-GAN model training and 3D flow field rendering).

[0236] FPGA coprocessor: handles low-latency, highly deterministic real-time tasks (such as sensor data preprocessing and edge reasoning).

[0237] Implementation steps and key technologies:

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

[0239] Task classification:

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

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

[0242] Real-time sensitive: ADCP flow rate analysis and data noise reduction (distributed to FPGA).

[0243] Dynamic resource scheduling:

[0244] Kubernetes orchestration: Define computing task types (such as job-type: gpu-inference) through custom resource definitions (CRDs) and automatically assign them to corresponding nodes.

[0245] Priority queue: Real-time data streams (such as 5000 Hz sensors) are given priority to preempt GPU resources, ensuring latency < 50 ms.

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

[0247] CPU-GPU heterogeneous communication:

[0248] Zero-Copy Memory: Use CUDA Unified Memory to achieve zero-copy data transmission between CPU and GPU, reducing the data migration overhead between FVCOM and PINN.

[0249] Pipeline parallelism: Split the FVCOM simulation into a pipeline of time steps, with the GPU processing the current step while the CPU calculates the next step boundary conditions.

[0250] FPGA edge acceleration:

[0251] Customized IP core: Design a dedicated pre-processing pipeline for sensor data (such as Kalman filtering and outlier detection), reducing latency to 5 μs.

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

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

[0254] Feature decoupling and migration:

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

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

[0257] Active learning strategy: Dynamically select key areas for supplementary observations based on uncertainty quantification (such as Monte Carlo Dropout) to reduce data acquisition costs.

[0258] Technical principles and core ideas:

[0259] Cross-modal data distillation uses knowledge transfer and feature decoupling to compress the knowledge of a high-precision model (teacher) into a lightweight model (student), while simultaneously fusing heterogeneous data from multiple sources (satellite, buoy, and simulation). 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] Multimodal fusion backbone: DenseNet-201 is used as the basic architecture, and the input layer processes different modal data in different paths. Satellite remote sensing: 3D convolution processes height-time series (such as 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: Contrastive loss is used to constrain the consistency of different modal features in the latent space.

[0267] Student Model Design:

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

[0269] Multi-scale attention: Introducing a lightweight CBAM module (channel + spatial attention) to focus on key areas (such as the 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 l-th layer features of teachers and students respectively.

[0275] Relation Distillation: The relationship matrix (Gram matrix) of the teacher model feature map is used to guide the student model to learn feature correlations.

[0276] Cross-modal data alignment:

[0277] Space-time interpolation alignment: Cubic spline interpolation is used to unify satellite data (1 km resolution) and buoy point data to the FVCOM grid.

[0278] Adversarial Domain Adaptation: A domain classifier is added to the teacher model to eliminate the inter-modality distribution difference through the Gradient Reversal Layer (GRL).

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

[0280] Uncertainty sampling:

[0281] Monte Carlo Dropout: Calculate the prediction variance through multiple forward propagations and select the area with the largest variance to make additional observations.

[0282] Boundary-sensitive sampling: Prioritizes the collection of data points with large flow field gradients (such as fronts and vortex edges).

[0283] Incremental training strategy:

[0284] Elastic weight solidification: Preserve important parameters of existing knowledge and only fine-tune the network parts relevant to new data.

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

[0286] Verification and performance optimization.

[0287] (a) Performance indicators:

[0288] Model compression rate and accuracy preservation: PSNR decreases by <0.5 dB and vortex center positioning error increases by <0.1 km.

[0289] Inference speed: Edge devices (such as Jetson Xavier) can process a single frame in less than 20 ms.

[0290] (b) Optimization techniques.

[0291] Dynamic channel pruning: Dynamically close redundant channels based on neuron activation strength, accelerating inference.

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

[0293] Full process integration and challenge response:

[0294] Collaborative Workflow:

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

[0296] Heterogeneous computing scheduling: Kubernetes allocates resources based on task labels, with the priority being: real-time tasks > training tasks > simulation tasks.

[0297] Distillation and Inference: Offline training of the teacher model → Knowledge transfer to the student model → Deployment on edge devices.

[0298] Core challenges and solutions:

[0299] Heterogeneous hardware synchronization solution: The NCCL communication library is used to achieve high-speed communication between GPUs, and RDMA is used between CPUs and GPUs to reduce latency.

[0300] Inter-modal semantic gap solution: Introduce cross-modal contrastive learning (such as CLIP style) to build a unified semantic space.

[0301] Edge device resource limitation solution: Use TensorRT to optimize the inference engine, combined with INT8 quantization and layer fusion technology.

[0302] 4. System implementation process and verification.

[0303] (1) Full process steps:

[0304] a. Data input: Integrate multi-source data such as satellite remote sensing, buoys, and numerical simulations, and perform normalization and spatiotemporal alignment. Figure 2 Schematic diagram of data input and preprocessing.

[0305] b. Coarse grid FVCOM precalculation: Generate background flow field and identify high gradient areas to trigger grid refinement.

[0306] c. PINN-GAN joint training: alternately optimize physical loss and adversarial loss, iterating until convergence.

[0307] d. Heterogeneous computing task scheduling: Dynamically allocate CPU / GPU resources through Kubernetes to prioritize the processing of real-time data streams.

[0308] e. Dynamic verification and feedback: Compare the reconstruction results with independent observation datasets (such as Argo floats) and calculate indicators such as RMSE and eddy capture rate. Figure 3 Schematic diagram for dynamic verification and feedback optimization.

[0309] (2) Performance advantages:

[0310] Improved accuracy: In the South China Sea vortex case, compared with traditional interpolation methods, this system reduces the RMSE of the velocity field and the vortex center positioning error.

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

[0312] 5. Key technological innovations.

[0313] Physics-data dual-driven fusion: By embedding the FVCOM control equations through PINN, the mechanism deviation problem of pure data-driven models in data-sparse areas is solved.

[0314] Multi-discriminator GAN architecture: Spatial-temporal dual-path supervision significantly improves the ability to capture transient processes such as mesoscale eddies and internal waves.

[0315] Grid-computing collaborative optimization: The combination of dynamically nested grids and heterogeneous computing achieves Pareto optimality in resolution and efficiency.

[0316] This solution includes the following key technical solutions:

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

[0318] A unified cross-scale spatiotemporal benchmarking method is proposed to address the challenge of adaptive alignment between satellite (low-frequency / large-scale), buoy (high-frequency / point), and numerical simulation field (mesoscale / grid) data. Unlike traditional linear interpolation, this method employs a combined cubic spline temporal interpolation and inverse distance weighted spatial fusion (IDW) algorithm to achieve continuous mapping of satellite data (1-day period) and buoy data (5000Hz sampling rate) at a unified time step (1 second), while also implementing dynamic weight assignment. Z-score normalization and isolation forest anomaly detection (threshold 3σ) are also introduced to reduce data fusion errors.

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

[0320] A grid refinement mechanism triggered by flow field gradients (velocity gradient, vorticity) is proposed, breaking through the limitations of traditional fixed multi-scale grids. By defining gradient threshold trigger conditions and combining them with a vertical stratification strategy, this method enables adaptive switching of grid resolution from kilometers (background field) to meters (vortex region), improving computational efficiency. Differentiating from existing methods, a maximum side-to-side ratio constraint for adjacent grids is introduced to avoid topological distortion between the refined region and the unstructured grid.

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

[0322] The hard constraints of the Navier-Stokes equations are embedded in the GAN generator to construct a PINN-GAN dual-driven model. A physical loss term and a dynamic weight adjustment strategy are designed to address the physical inconsistencies of traditional data-driven models. The generator adopts a U-Net structure, and the discriminator introduces a spatiotemporal 3D convolution kernel (5×5×5) to reduce spatiotemporal discrimination errors.

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

[0324] A real-time scheduling algorithm based on task priorities and hardware characteristics was proposed. Task allocation rules were defined: the CPU handles FVCOM coarse-grid simulation (priority 1), the GPU performs PINN-GAN inference (priority 2), and the FPGA handles ADCP data parsing. A pipeline timing synchronization protocol was developed to enable end-to-end flow field generation with multi-device collaboration, achieving speed improvements compared to traditional CPU clusters.

[0325] (5) Active sampling and model incremental update mechanism driven by dynamic feedback.

[0326] Build a buoy path planning algorithm guided by prediction uncertainty (based on Monte Carlo Dropout). Define a formula for the active sampling radius and set trigger conditions for model updates. Propose a lightweight teacher-student model distillation method to reduce the time required for incremental updates.

[0327] (6) Cross-domain feature alignment technology for multimodal contrastive learning.

[0328] A cross-modal (current velocity, temperature, salinity) comparative loss function was designed to address the feature space differences between satellite, buoy, and simulated data. Triplet Loss (anchor = satellite data, positive sample = buoy data, negative sample = simulated background field) was used to constrain the cosine similarity of feature vectors, enabling efficient mapping of multi-source data into a unified latent space and improving feature alignment accuracy.

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

[0330] A collaborative framework for FPGA edge preprocessing and GPU cloud inference is proposed. A dedicated logic unit for ADCP signal analysis is defined to support real-time flow rate analysis. Multi-GPU parallel inference engines are deployed in the cloud, and data synchronization is achieved through the gRPC protocol.

[0331] (8) Physical constraint enhancement method for high-precision vortex and front capture.

[0332] For mesoscale eddies and temperature fronts, a vorticity conservation constraint and front gradient enhancement loss are proposed. Combined with the PINN-GAN generator, the vortex center positioning error is reduced, resulting in improved temperature front resolution.

[0333] (9) Bidirectional conversion algorithm from unstructured grid to structured tensor.

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

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

[0336] This method is extended to industrial scenarios, and a fusion scheme is proposed for high-frequency PIV (10 kHz) and RANS simulation. A dynamic turbulent viscosity correction model is designed and combined with GAN super-resolution reconstruction to achieve high-precision capture of flow fields in special situations.

[0337] Reference Figure 4 Another aspect of the present application further provides a high spatiotemporal resolution refined flow field reconstruction device, the device comprising:

[0338] Preprocessing unit, used to normalize and align satellite remote sensing, buoy observation and numerical simulation data in time and space;

[0339] The flow field generation unit is used to perform pre-calculations on the unstructured nested grid through the FVCOM model based on the aligned data, and then dynamically refine the grid according to the flow field gradient and generate the background flow field;

[0340] The flow field reconstruction unit is used to input the background flow field into the PINN-GAN joint framework and output the refined flow field through physical constraint loss and dual discriminator adversarial training;

[0341] The optimization unit is used to schedule computing tasks using heterogeneous accelerators, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

[0342] It can be understood that the contents of the above method embodiments are all applicable to the present device embodiments, the functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0343] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flow chart of the present application are provided in an exemplary manner for 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. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0344] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present application. More specifically, 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 understood within the routine skills of an engineer. Therefore, a person skilled in the art can implement the present application as set forth in the claims using ordinary techniques without undue experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0345] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

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

[0347] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0348] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0349] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

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

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

Claims

1. A high temporal and spatial resolution refined flow field reconstruction method, characterized in that: The method comprises the following steps: Normalize and align satellite remote sensing, buoy observation and numerical simulation data in time and space; Based on the aligned data, the FVCOM model is used to perform pre-calculations on the unstructured nested grid, and then the grid is dynamically refined according to the flow field gradient to generate the background flow field. The background flow field is input into the PINN-GAN joint framework, and the refined flow field is output through physical constraint loss and dual discriminator adversarial training; Heterogeneous accelerators are used to schedule computing tasks, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

2. The high temporal and spatial resolution refined flow field reconstruction method according to claim 1, characterized in that: The method of performing precalculation on the unstructured nested grid by the FVCOM model based on the aligned data includes the following steps: Nested grid layering is performed using horizontal grids and vertical surface layers.

3. The high temporal and spatial resolution refined flow field reconstruction method according to claim 2, characterized in that: The method of dynamically encrypting the grid according to the flow field gradient and generating the background flow field comprises the following steps: The grid encryption operation is triggered by the flow field gradient threshold, and the FVCOM model is used on the encrypted unstructured nested grid to generate the background flow field.

4. The high spatiotemporal resolution refined flow field reconstruction method according to claim 1, characterized in that: The background flow field is input into the PINN-GAN joint framework, and a refined flow field is output through physical constraint loss and dual discriminator adversarial training, including the following steps: Implement physical law constraints by adding momentum equation and continuity equation residuals to the loss function; The 3D convolution kernel is used to process the spatial vortex features, and the Bi-LSTM network is used to process the tidal time series features.

5. The high temporal and spatial resolution refined flow field reconstruction method according to claim 1, characterized in that: The method of scheduling computing tasks using heterogeneous accelerators includes the following steps: The FVCOM coarse grid parallel computation is run on a CPU cluster, PINN-GAN is deployed on GPU nodes to achieve a single-frame reconstruction delay of less than 50 ms, and the real-time ADCP data stream is processed by an FPGA coprocessor.

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

7. The high spatiotemporal resolution refined flow field reconstruction method according to any one of claims 1 to 6, characterized in that: The real-time verification of the reconstructed refined flow field and feedback optimization includes the following steps: Quantitative evaluation indicators are calculated by comparing the reconstructed refined flow field with independent observation data sets, and the parameters of the FVCOM model are optimized based on the feedback of the indicator results.

8. High temporal and spatial resolution refined flow field reconstruction device, characterized by: The device comprises: Preprocessing unit, used to normalize and align satellite remote sensing, buoy observation and numerical simulation data in time and space; The flow field generation unit is used to perform pre-calculations on the unstructured nested grid through the FVCOM model based on the aligned data, and then dynamically refine the grid according to the flow field gradient and generate the background flow field; The flow field reconstruction unit is used to input the background flow field into the PINN-GAN joint framework and output the refined flow field through physical constraint loss and dual discriminator adversarial training; The optimization unit is used to schedule computing tasks using heterogeneous accelerators, verify the reconstructed refined flow field in real time, and provide feedback for optimization.

9. 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 according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The storage medium stores a program, and the program is executed by a processor to implement the method according to any one of claims 1 to 7.

Images