Marine multi-source observation data fusion assimilation method and system

Abstract

The invention discloses an ocean multi-source observation data fusion assimilation method and system, and relates to the technical field of data processing, and the method comprises the steps: obtaining multi-source ocean observation data; uniformly converting the timestamps and the space grids of the multi-source ocean observation data into a data format of an FVCOM numerical model; constructing a data fusion assimilation model; the data fusion assimilation model comprises an FVCOM physical model and an AC-PKAN neural network; training the data fusion assimilation model, and adjusting the weight parameter of the data fusion assimilation model according to the dynamic observation data currently acquired by the edge device; and fusing and assimilating the multi-source marine observation data by using the data fusion assimilation model to obtain multi-source fusion marine observation data. According to the method, the timestamps and the space grids of the multi-source marine observation data are synchronized, and then the data fusion assimilation model is constructed for fusion assimilation, so that the fusion accuracy of the multi-source marine observation data can be improved.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to a method and system for fusing and assimilating multi-source ocean observation data. Background Art

[0002] Ocean observation data is characterized by being multi-source, heterogeneous, highly dynamic, and strongly spatio-temporally correlated, covering various observation means such as satellite remote sensing, buoy sensors, sonars, and underwater robots. However, traditional ocean observations face the following technical problems:

[0003] (1) Data island problem: The data formats, spatio-temporal resolutions, and benchmarks of different observation platforms vary significantly, resulting in difficulties in fusion. (2) Influence of complex environment: The ocean environment (such as ocean currents, temperature stratification, and underwater noise) easily causes distortion or noise interference in the observation data, affecting the fusion quality. (3) Computational efficiency bottleneck: Traditional fusion methods (such as Kalman filtering and principal component analysis) are inefficient in processing high-dimensional and non-linear data and are difficult to meet the real-time requirements. (4) Insufficient decision-level fusion: Existing methods mostly focus on data or feature-level fusion and lack the comprehensive discrimination ability of multi-source information, resulting in a relatively high target misjudgment rate. Summary of the Invention

[0004] The main purpose of the embodiments of this application is to propose a method and system for fusing and assimilating multi-source ocean observation data to improve the accuracy of fusing multi-source ocean observation data.

[0005] To achieve the above object, on the one hand, an embodiment of this application proposes a method for fusing and assimilating multi-source ocean observation data, and the method includes the following steps:

[0006] Obtain multi-source ocean observation data; wherein, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices;

[0007] Uniformly convert the timestamps and spatial grids of the multi-source ocean observation data to the data format of the FVCOM numerical model;

[0008] Construct a data fusion and assimilation model; wherein, the data fusion and assimilation model includes an FVCOM physical model and an AC-PKAN neural network;

[0009] Train the data fusion and assimilation model, and then adjust the weight parameters of the data fusion and assimilation model according to the dynamic observation data currently obtained by the edge device;

[0010] Use the data fusion and assimilation model to fuse and assimilate the multi-source ocean observation data to obtain multi-source fused ocean observation data.

[0011] In some embodiments, the acquisition of multi-source ocean observation data includes the following steps:

[0012] Obtain the ocean current and temperature field output by the FVCOM numerical model;

[0013] Obtain the sea surface height, sea surface temperature, and chlorophyll concentration obtained by satellite remote sensing inversion;

[0014] Obtain the salinity, sea surface current field, vertical velocity, and temperature profile collected by buoys and submersible buoys in the edge device;

[0015] Obtain the beach topography and local current field collected by drones and unmanned surface vehicles in the edge device.

[0016] In some embodiments, the unified conversion of the timestamps and spatial grids of the multi-source ocean observation data to the data format of the FVCOM numerical model includes the following steps:

[0017] Remove the noise and outliers in the multi-source ocean observation data, and supplement the missing values in the multi-source ocean observation data;

[0018] Perform dimension unification, spatial coordinate alignment, and time series alignment on the multi-source ocean observation data to obtain normalized data;

[0019] Using the velocity field output by the FVCOM numerical model as a benchmark, fuse the normalized data through weighted average, so that the timestamps and spatial grids of the normalized data are unified and converted to the triangular unstructured grid of the FVCOM numerical model.

[0020] In some embodiments, the construction of the data fusion and assimilation model includes the following steps:

[0021] Embed the Navier-Stokes equation and the mass transport equation in the FVCOM model, and set the configuration parameters and boundary conditions of the FVCOM model to obtain the FVCOM physical model;

[0022] Construct the AC-PKAN neural network based on the Chebyshev basis transformation layer and the dual attention mechanism; wherein, the Chebyshev basis transformation layer is used to map the multi-source ocean observation data to the Chebyshev polynomial space; the dual attention mechanism includes internal attention and external attention, the internal attention is used to dynamically weight the importance of different Chebyshev basis functions, and the external attention is used to associate the spatio-temporal correlation of the physical equation residuals with the observation data.

[0023] In some embodiments, the expression for constructing the Chebyshev basis transformation layer is as follows:

[0024] z = T kf(x) = cos(karccos(x));

[0025] where T k is the k-th order Chebyshev basis function, which is used to map the multi-source ocean observation data to the Chebyshev polynomial space;

[0026] Adjust the order k dynamically through the TensorFlow custom layer;

[0027] Construct the expression of the internal attention as follows:

[0028]

[0029] where e ij = LeakyReLU(W2 tanh(W1Z i ⊙Z j + b1) + b2); The residual connection is to concatenate the weighted feature and the original feature;

[0030] The expression of the external attention is:

[0031]

[0032] where f pq = V2 tanh(V1E p ⊙O q + b3) + b4; E p is the physical residual, and O q is the observation error;

[0033] Design the loss function of the AC-PKAN neural network as:

[0034] L1 = λ p L physics + λ d L data + μL attention ;

[0035] where L1 is the loss function of the AC-PKAN neural network, L physics is the physical constraint term, L data is the data assimilation term, L attention is the attention regularization term, and λ p 、λ d 、μ are the coefficients of the corresponding terms respectively.

[0036] In some embodiments, the step of using the data fusion and assimilation model to fuse and assimilate the multi-source ocean observation data to obtain the multi-source fused ocean observation data includes the following steps:

[0037] Construct the background error covariance and dynamically update the background error covariance through BiLSTM to dynamically update the weights of the data fusion and assimilation model;

[0038] The expression for updating the background error covariance is as follows:

[0039] B t = BiLSTM(B t-1 , ut, vt);

[0040] where B t is the background error covariance at time t, B t-1 is the background error covariance at time t - 1, and ut and vt are the ocean current velocity component and the tidal component, respectively;

[0041] Construct the observation error covariance and dynamically adjust the observation error covariance through the BiLSTM time series prediction residual to dynamically adjust the weights of the data fusion and assimilation model;

[0042] The expression for updating the observation error covariance is as follows:

[0043] R t = α·R t-1 + β·C;

[0044] where R t is the observation error covariance at time t, R t-1 is the observation error covariance at time t - 1, α and β are different forgetting factors, and C is the residual variance;

[0045] Minimize the prediction error and the positive definiteness constraint of the covariance matrix; the covariance matrix includes the background error covariance and the observation error covariance;

[0046] The loss function for the minimization is:

[0047] L2 = ∥y t - H(x t )∥ 2 2+ λ∥W∥ 2 F ;

[0048] where L2 is the loss function for the minimization, H is the observation operator, and λ is the regularization term coefficient;

[0049] Use the data fusion and assimilation model with adjusted weights to fuse and assimilate the multi-source ocean observation data to obtain the multi-source fused ocean observation data.

[0050] In some embodiments, the method further includes the following steps:

[0051] Output at least one of a three-dimensional flow field animation, a salinity contour map, or a heatwave diffusion simulation path based on the multi-source fusion ocean observation data, and then determine the results of ocean environment prediction and decision-making.

[0052] To achieve the above object, on the other hand, an embodiment of the present application provides a system for assimilating multi-source ocean observation data, the system includes:

[0053] A data acquisition unit, configured to acquire multi-source ocean observation data; wherein, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices;

[0054] A format conversion unit, configured to uniformly convert the timestamps and spatial grids of the multi-source ocean observation data into the data format of the FVCOM numerical model;

[0055] A model construction unit, configured to construct a data assimilation model; wherein, the data assimilation model includes an FVCOM physical model and an AC-PKAN neural network;

[0056] A model training and adjustment unit, configured to train the data assimilation model, and then adjust the weight parameters of the data assimilation model according to the dynamic observation data currently acquired by the edge device;

[0057] A data fusion unit, configured to fuse and assimilate the multi-source ocean observation data by using the data assimilation model to obtain multi-source fusion ocean observation data.

[0058] To achieve the above object, on the other hand, an embodiment of the present application provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the above method is implemented.

[0059] To achieve the above object, on the other hand, an embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above method is implemented.

[0060] The embodiments of the present application at least include the following beneficial effects:

[0061] This application can obtain multi-source ocean observation data; among them, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices; the timestamps and spatial grids of the multi-source ocean observation data are uniformly converted into the data format of the FVCOM numerical model; a data fusion and assimilation model is constructed; among them, the data fusion and assimilation model includes the FVCOM physical model and the AC-PKAN neural network; the data fusion and assimilation model is trained, and then the weight parameters of the data fusion and assimilation model are adjusted according to the dynamic observation data currently obtained by the edge device; the multi-source ocean observation data is fused and assimilated by using the data fusion and assimilation model to obtain multi-source fused ocean observation data. By synchronizing the timestamps and spatial grids of the multi-source ocean observation data and then constructing a data fusion and assimilation model for fusion and assimilation, this application can improve the fusion accuracy of the multi-source ocean observation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0063] Figure 1 It is a schematic flowchart of a method for fusing and assimilating multi-source ocean observation data provided by an embodiment of the present application;

[0064] Figure 2 It is an overall framework diagram of a multi-source ocean observation data fusion and assimilation solution provided by an embodiment of the present application;

[0065] Figure 3 It is a schematic structural diagram of a multi-source ocean observation data fusion and assimilation system provided by an embodiment of the present application;

[0066] Figure 4 It is a schematic hardware structure diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will further elaborate on the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application. When the following description involves the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the embodiments of the present application. They are only examples of systems and methods that are consistent with some aspects of the embodiments of the present application detailed in the appended claims.

[0068] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, words such as "if" and "when" as used herein may be interpreted as "when...", "while...", or "in response to a determination".

[0069] The terms "at least one", "a plurality", "each", "any one", etc. used in this application, "at least one" includes one, two, or more than two, "a plurality" includes two or more than two, "each" refers to each one of the corresponding plurality, and "any one" refers to any one of the plurality.

[0070] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0071] Before elaborating on the embodiments of this application in detail, some related technologies involved in the embodiments of this application are described first as follows:

[0072] To solve the above problems, the technology of marine multi-source observation data fusion and assimilation has emerged as the times require. This technology integrates multi-source data and numerical models, combines intelligent algorithms to optimize data quality and improve model accuracy, and finally realizes high-precision marine environment prediction and decision support. The general situation of existing multi-source observation data fusion and assimilation technology solutions is as follows:

[0073] (I) Related technologies:

[0074] (1) Multi-source information fusion hierarchical data-level fusion: directly processes raw observation data (such as satellite remote sensing images, sonar signals), and realizes preliminary integration through methods such as weighted average and Kalman filtering, but is vulnerable to noise interference and has a large amount of calculation. Feature-level fusion: extracts the common features of multi-source data (such as target morphology, spectral features), and uses machine learning (such as SVM, deep learning) for feature matching and classification, but may lose some detailed information. Decision-level fusion: based on the comprehensive judgment of multi-source features (such as Bayesian inference, D-S evidence theory), generates the final decision result, but has a strong dependence on prior knowledge.

[0075] (2) Representative data assimilation methods Variational methods (such as four-dimensional variational assimilation, 4D-Var): Optimize the objective function to fuse observational data and model predictions, applicable to the global scale but with high computational complexity. Filtering methods (such as Ensemble Kalman Filter, EnKF; Ensemble Optimal Interpolation, EnOI, etc.): Successively update state estimates based on the Bayesian framework, applicable to non-linear and non-Gaussian systems, but sensitive to initial errors.

[0076] (II) Existing representative technical solutions:

[0077] (1) DI&M-Engine platform:

[0078] The technical characteristics of this platform are integrating multi-source data from "space-air-ground-sea" (satellites, drones, buoys, etc.), improving the analysis efficiency by more than 70% through data cleaning and trend prediction algorithms, and supporting three-dimensional visualization decision-making. The limitation is that it relies on manual weight setting and does not fully combine prior knowledge to optimize the fusion quality.

[0079] (2) Regional high-resolution multi-sphere coupled assimilation system:

[0080] The technical characteristics of this system are supporting the assimilation of new observational data such as underwater gliders and Argo buoys, with a horizontal resolution reaching the 1 km level, and improving the forecast accuracy by more than 5% by combining atmosphere-ocean-land surface coupling algorithms. The limitation is the insufficient adaptability to complex environments such as reefs and limited acoustic signal assimilation ability.

[0081] (3) Traditional Ensemble Kalman Filter (EnKF):

[0082] The technical characteristics of this method are combining ensemble forecasting and Kalman filtering to process non-linear systems, and it is widely used in ocean data assimilation. The limitation is that the accuracy decreases when the ensemble size is limited, and it is difficult to process high-dimensional data.

[0083] The deficiencies of the existing technologies are mainly reflected in: (1) Limited fusion quality: The quality differences and systematic biases of multi-source data are not fully considered, resulting in local distortion of the fusion results. (2) Poor adaptability to complex environments: The ability to capture small-scale physical processes (such as vortices and turbulence) in areas such as reefs and deep seas is insufficient. (3) Real-time performance and scalability: Traditional methods are difficult to adapt to edge computing devices (such as buoys and ships), and it is difficult to meet the real-time analysis requirements.

[0084] Therefore, the embodiments of the present application provide a method and system for fusing and assimilating multi-source ocean observation data, which relates to the technical field of data processing. The method and system for fusing and assimilating multi-source ocean observation data provided by the embodiments of the present application can be applied to a terminal, or to a server, or can also be software running on a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, a vehicle-mounted terminal, etc., but is not limited thereto; the server side can be configured as an independent physical server, or can be configured as a server cluster or a distributed system composed of multiple physical servers, or can also be configured as a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements a method and system for fusing and assimilating multi-source ocean observation data, etc., but is not limited to the above forms.

[0085] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and so on. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0086] Refer to Figure 1 , the embodiments of the present application provide a method for fusing and assimilating multi-source ocean observation data, which may include but is not limited to S100 to S140, as follows:

[0087] S100: Obtain multi-source ocean observation data; wherein, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices;

[0088] S110: Uniformly convert the timestamps and spatial grids of the multi-source ocean observation data into the data format of the FVCOM numerical model;

[0089] S120: Construct a data fusion and assimilation model; wherein, the data fusion and assimilation model includes an FVCOM physical model and an AC-PKAN neural network;

[0090] S130: Train the data fusion and assimilation model, and then adjust the weight parameters of the data fusion and assimilation model according to the dynamic observation data currently obtained by the edge device;

[0091] S140: Use the data fusion and assimilation model to fuse and assimilate the multi-source ocean observation data to obtain multi-source fused ocean observation data.

[0092] Optionally, the obtaining of the multi-source ocean observation data includes the following steps:

[0093] Obtain the sea current and temperature field output by the FVCOM numerical model;

[0094] Obtain the sea surface height, sea surface temperature, and chlorophyll concentration obtained by satellite remote sensing inversion;

[0095] Obtain the salinity, sea surface current field, vertical velocity, and temperature profile collected by buoys and submersible buoys in the edge device;

[0096] Obtain the beach topography and local current field collected by drones and unmanned boats in the edge device.

[0097] Optionally, the unifying the timestamps and spatial grids of the multi-source ocean observation data to the data format of the FVCOM numerical model includes the following steps:

[0098] Remove the noise and outliers in the multi-source ocean observation data, and supplement the missing values in the multi-source ocean observation data;

[0099] Perform dimension unification, spatial coordinate alignment, and time series alignment on the multi-source ocean observation data to obtain normalized data;

[0100] Use the velocity field output by the FVCOM numerical model as a benchmark, fuse the normalized data through weighted average, and then unify the timestamps and spatial grids of the normalized data to the triangular unstructured grid of the FVCOM numerical model.

[0101] Optionally, the constructing of the data fusion and assimilation model includes the following steps:

[0102] Embed the Navier-Stokes equation and the mass transport equation into the FVCOM model, and set the configuration parameters and boundary conditions of the FVCOM model to obtain the FVCOM physical model;

[0103] Construct the AC-PKAN neural network based on the Chebyshev basis transformation layer and the dual attention mechanism; wherein, the Chebyshev basis transformation layer is used to map the multi-source ocean observation data to the Chebyshev polynomial space; the dual attention mechanism includes internal attention and external attention, the internal attention is used to dynamically weight the importance of different Chebyshev basis functions, and the external attention is used to associate the spatio-temporal correlation between the physical equation residuals and the observation data.

[0104] Optionally, the expression for constructing the Chebyshev basis transformation layer is as follows:

[0105] z = T k (x) = cos(karccos(x));

[0106] where, T k is the k-th order Chebyshev basis function, which is used to map the multi-source ocean observation data to the Chebyshev polynomial space;

[0107] Through the TensorFlo w custom layer, dynamically adjust the order k;

[0108] The expression for constructing the internal attention is as follows:

[0109]

[0110] where, e ij = LeakyReLU(W2 tanh(W1Z i ⊙Z j + b1) + b2); The residual connection is to splice the weighted feature and the original feature;

[0111] The expression for the external attention is:

[0112]

[0113] where, f pq = V2 tanh(V1E p ⊙O q + b3) + b4; E p is the physical residual, and O q is the observation error;

[0114] Design the loss function of the AC-PKAN neural network as:

[0115] L1 = λ p L physics + λ d L data + μL attention ;

[0116] Among them, L1 is the loss function of the AC-PKAN neural network, L physics is the physical constraint term, L data is the data assimilation term, L attention is the attention regularization term, λ p 、λ d 、μ are the coefficients of the corresponding terms respectively.

[0117] Optionally, the step of fusing and assimilating the multi-source ocean observation data by using the data fusion and assimilation model to obtain the multi-source fused ocean observation data includes the following steps:

[0118] Construct the background error covariance, and dynamically update the background error covariance through BiLSTM to dynamically update the weights of the data fusion and assimilation model;

[0119] The expression for updating the background error covariance is as follows:

[0120] B t = BiLSTM(B t-1 , ut, vt);

[0121] Among them, B t is the background error covariance at time t, B t-1 is the background error covariance at time t-1, ut and vt are the sea current velocity component and the tidal component respectively;

[0122] Construct the observation error covariance, and dynamically adjust the observation error covariance through the BiLSTM time series prediction residual to dynamically adjust the weights of the data fusion and assimilation model;

[0123] The expression for updating the observation error covariance is as follows:

[0124] R t = α·R t-1 + β·C;

[0125] Among them, R t is the observation error covariance at time t, R t-1 is the observation error covariance at time t-1, α and β are different forgetting factors, and C is the residual variance;

[0126] Minimize the prediction error and the positive definiteness constraint of the covariance matrix; the covariance matrix includes the background error covariance and the observation error covariance;

[0127] The loss function to be minimized is:

[0128] L2 = ∥y t - H(x t )∥ 22 + λ∥W∥ 2 F ;

[0129] where L2 is the minimized loss function, H is the observation operator, and λ is the regularization term coefficient;

[0130] Using the data fusion and assimilation model with adjusted weights to fuse and assimilate the multi-source ocean observation data to obtain the multi-source fused ocean observation data.

[0131] Optionally, the method further includes the following steps:

[0132] Outputting at least one of a three-dimensional flow field animation, a salinity isohyet map, or a heatwave diffusion simulation path according to the multi-source fused ocean observation data, and further determining the result of ocean environment prediction and decision-making.

[0133] Next, the solution of the embodiment of the present application will be introduced and described in detail in combination with specific application examples.

[0134] Referring to Figure 2 , the overall framework of a scheme for fusing and assimilating multi-source ocean observation data provided in this embodiment is divided into four layers, which are, from bottom to top, the data acquisition layer, the preprocessing layer, the model layer, and the application layer. The detailed construction steps are as follows:

[0135] (1) The components of the data acquisition layer include a satellite remote sensing system (such as Sentinel-2), a buoy network (Argo / CTD), sensors carried by unmanned aerial vehicles, weather stations, and the FVCOM numerical model. The function is to collect multi-source data such as ocean current velocity, temperature, salinity, chlorophyll concentration, and wave height in real time, and the spatial resolution covers the global to local areas (such as a 1 km grid).

[0136] The data acquisition layer has 5 parts:

[0137] ① Data acquisition: satellite → buoy → unmanned aerial vehicle → FVCOM model → edge device. Satellite remote sensing is applicable to obtaining macroscopic parameters such as sea surface height and temperature using Sentinel-1 (SAR) and Aqua MODIS (optical). Buoys / submersibles / high-frequency ground wave radars, Argo buoys / CTD submersibles, etc. collect salinity, sea surface current field, vertical current velocity, and temperature profiles at fixed points or in local sea areas. Unmanned aerial vehicles / unmanned boats, etc., carry SAR and ADCP equipment to achieve shallow water terrain mapping and local current field measurement. ② Transmission layer: Iridium / 5G → satellite communication → LoRaWAN. Use API interfaces (such as the FVCOM model output interface) to obtain numerical simulation data. Deploy ETL scripts (such as the pandas library in Python) to implement multi-source data format conversion. Realize second-level data transmission through Iridium communication (buoy) and 5G network (unmanned aerial vehicle).

[0138] ③ Storage layer: The distributed database (Ceph) stores the raw data and assimilation results. The Ceph storage system is used to save the raw data (NetCDF format) and assimilation results, and a data dictionary is established to record key information such as sensor models and acquisition timestamps.

[0139] ④ Processing layer: ETL tools (such as Talend) clean the data → Spark cluster transformation → Hive storage. The local binary pattern (LBP) encoding is used to enhance the texture features of satellite SAR images. Trilinear interpolation is used to unify different data sources into the FVCOM grid system. The Kalman filtering algorithm is used to identify sudden fluctuations in buoy salinity data.

[0140] ⑤ Application layer: Output 3D flow field animations and contour maps of hydrometeorological elements such as salinity to support decision-making. The observation data (such as satellite SST) is combined with the FVCOM model error covariance matrix to optimize the initial conditions. The MobileAttention-Lite model is deployed on the drone side to achieve real-time feature extraction.

[0141] (2) The data preprocessing layer unifies the timestamps and spatial grids of different data sources into the FVCOM system through spatio-temporal interpolation (such as trilinear interpolation). Integrate the numerical simulation data of ocean currents, temperature fields, etc. output by the FVCOM model, the ocean surface parameters retrieved by satellite remote sensing (such as sea surface temperature, chlorophyll concentration), and the real-time observation data of buoys / ships (such as salinity, wave height). Through interpolation and resampling techniques, the timestamps and spatial grids of different data sources are unified into the FVCOM grid system. The satellite images are enhanced by flipping / rotating, and the time series data is denoised by wavelet transform. The implementation process of the data preprocessing layer includes: data acquisition layer → data cleaning (noise / missing values / outliers) → data standardization (dimension / dimensionality reduction) → feature extraction (LBP / wavelet transform) → spatio-temporal alignment (interpolation / projection transformation) → multi-source data fusion → real-time preprocessing (edge computing) → metadata management → distributed storage. The corresponding implementation steps are introduced as follows. The steps of data cleaning and outlier processing are as follows:

[0142] Step 1.1: Noise filtering Median filtering (window size 3×3) is used to remove random noise in SAR images from satellite remote sensing data; moving average filtering (window size 5 hours) is used to smooth high-frequency fluctuating noise in buoy data; wavelet threshold denoising (Daubechies-4 wavelet) is used to eliminate instrument errors in drone ADCP data. Step 1.2: Missing value filling The spatio-temporal interpolation method uses trilinear interpolation to fill the blank areas in the FVCOM grid; the time series model predicts the missing buoy salinity data based on LSTM (such as the gap during the drift of Argo buoys).

[0143] Step 1.3: Outlier Detection and Correction:

[0144] Based on statistical methods, the 3σ principle is used to eliminate outliers (such as sea surface temperature exceeding the historical mean ±3σ); the Isolation Forest in machine learning model training is used to detect abnormal data points (such as sudden turbidity peaks).

[0145] The steps of data standardization and normalization are as follows:

[0146] Step 2.1: Dimensional Unification Physical quantity conversion is to convert the sea current velocity from [m / s] to [knots] (1 knot = 0.51444 m / s); unit normalization is to standardize salinity (PSU), flow velocity, etc. to the interval [-1, 1]. Step 2.2: Spatial Coordinate Alignment Projection conversion is to convert the WGS84 coordinate system of satellite images to the local rectangular coordinate system used by the FVCOM model; grid matching maps the UAV ADCP data to the 1km FVCOM grid through nearest neighbor interpolation. Step 2.3: Time Series Alignment Resampling is to downsample the satellite daily data to the hourly level using the resample method of pandas; timestamp correction is to synchronize the timestamps of the buoy and the UAV based on UTC time and the NTP protocol.

[0147] The steps of data augmentation and feature extraction are as follows:

[0148] Step 3.1: Multimodal Feature Fusion:

[0149] Image feature extraction uses the LBP algorithm of OpenCV to extract the local binary pattern (LBP) features of SAR images to characterize the texture of the shoal area; spectral feature extraction calculates NDVI (Normalized Difference Vegetation Index) and CHL-a (chlorophyll-a concentration) from MODIS satellite images.

[0150] Step 3.2: Temporal Feature Mining:

[0151] Wavelet transform uses the PyWavelets library to perform wavelet decomposition on the buoy salinity time series data to extract high-frequency (<24 hours) and low-frequency (>7 days) components; autocorrelation analysis calculates the autocorrelation coefficient of the sea current velocity to capture periodic fluctuations (such as tidal cycles).

[0152] Step 3.3: Data Augmentation Techniques:

[0153] Image enhancement performs random rotation (±15°) and horizontal flipping on optical images to expand the training dataset; temporal data augmentation generates synthetic samples through time translation (±1 hour) to improve the robustness of the model to emergencies.

[0154] The steps of spatio-temporal alignment and data integration are as follows:

[0155] Step 4.1: Multi-source data fusion. Physical constraint fusion takes the flow velocity field output by the FVCOM model as the benchmark and fuses satellite SST data through weighted average; data-driven fusion uses a graph convolutional network (GCN) to associate the spatial correlation between buoy salinity and UAV ADCP data. Step 4.2: Unified grid construction. The implementation of FVCOM grid adaptation is to project satellite images and buoy data onto the triangular unstructured grid of FVCOM; resolution balancing is to downsample UAV high-resolution data (such as 10m grid) to 1km to avoid computational redundancy.

[0156] The steps of real-time data processing and edge computing are as follows:

[0157] Step 5.1: Edge device preprocessing. The lightweight model deployment is to deploy the MobileAttention-Lite model on the UAV side to achieve real-time LBP feature extraction; the streaming data processing uses Apache Kafka to build a message queue to buffer real-time data streams (such as buoy data). Step 5.2: Online feature update. Dynamically load the pre-trained model through TensorFlow Serving and perform fine-tuning on new observation data; parameter synchronization realizes the synchronization of attention weight parameters between the edge device and the cloud model based on the MQTT protocol.

[0158] The steps of metadata management and data storage are as follows:

[0159] Step 6.1: Metadata annotation. Use the ISO 19115 standard to record sensor information (such as model, calibration date), data acquisition parameters (such as sampling frequency). Step 6.2: Distributed storage architecture. Store the original data in the MinIO object storage system (supporting PB-level data); store the processed feature data in ClickHouse, supporting efficient spatio-temporal queries.

[0160] (3) The model construction layer has two core modules: the FVCOM physical model and the AC-PKAN neural network. FVCOM physical model: Embed the Navier-Stokes equation and the mass transport equation as the underlying dynamic framework. The AC-PKAN neural network contains a Chebyshev basis transformation layer and a double attention mechanism. Chebyshev basis transformation layer: Map multi-modal data to the Chebyshev polynomial space, and the formula is:

[0161] z = T k (x) = cos(karccos(x));

[0162] where T k is the k-th order Chebyshev basis function, and x is the normalized input feature.

[0163] The dual attention mechanism includes: internal attention: dynamically weighting the importance of different Chebyshev basis functions; external attention: correlating the spatio-temporal correlation between the physical equation residuals and the observed data.

[0164] The AC-PKAN framework aims to overcome the problems of gradient vanishing, gradient explosion, and lack of interpretability in PINNs by introducing Chebyshev polynomials and attention mechanisms and combining the approximation ability of Kolmogorov-Arnold Networks (KANs). The Chebyshev polynomials accelerate convergence: mapping the solution space of the PDE to the Chebyshev polynomial basis and utilizing its orthogonality and compact support properties to significantly reduce the computational complexity.

[0165] The mathematical expression is:

[0166]

[0167] where T k is the Chebyshev polynomial and ξ(x) is the normalized coordinate.

[0168] The dual attention collaboration mechanism includes internal attention and external attention. Internal attention is only dynamically weighting the importance of different Chebyshev basis functions; external attention is correlating the spatio-temporal correlation between the physical equation residuals and the observed data.

[0169] The architecture of the three-layer model (input layer + hidden layer + output layer) is:

[0170] Input layer: concatenation of multi-modal data (such as velocity field + temperature field + terrain features);

[0171] Hidden layer: (1) Chebyshev basis transformation layer: mapping the input to the polynomial space; (2) attention module: internal attention: calculating the weights of the basis functions; external attention: correlating the physical residuals and the observed errors; (3) residual connection layer: enhancing the non-linear expression ability of the deep network.

[0172] Output layer: predicting the coefficient vector C of the PDE solution k .

[0173] The loss function is:

[0174] L = L physics + λL data + μL attention ;

[0175] where L physics is the PDE residual, L data is the observed error, and L attention is the attention weight regularization term.

[0176] Optimization Strategy: The adaptive AdamW algorithm is adopted, combined with DropBlock to prevent overfitting. The implementation process of the model construction layer is as follows: FVCOM model compilation → grid generation and optimization → AC-PKAN network construction → dual attention mechanism embedding → loss function design → distributed training → online update and edge deployment. The implementation steps for the corresponding parts are as follows:

[0177] Part 1: Construction of the FVCOM Physical Model:

[0178] (1) Environment Deployment and Compilation, Operating System Configuration: Use the Ubuntu 20.04 LTS system, deploy the Linux environment through WSL2 or virtual machines. Install dependent components: Intel OneAPI toolkit (including compiler, MPI library), NETCDF, HDF5, etc. Model Compilation: Download the FVCOM source code, configure the compilation options through the configure script (such as enabling hydrodynamic, temperature-salinity, and wave modules). Use the make command to generate the executable file, and the optimization options include -O3 and -march=native1.

[0179] (2) Grid Generation and Optimization, Unstructured Triangular Grid: Use SMS (Scientific Modeling System) to generate the grid, with key encryption in the nearshore shoal area (resolution 10 - 50m). Adopt the wet-dry discrimination technology to handle the moving boundary of the tidal flat and ensure mass conservation. Vertical Stratification: Use the σ coordinate system in the deep water area (such as σ = 0 at the seabed), and switch to the z coordinate system in the shoal area.

[0180] (3) Parameter Setting and Boundary Conditions, Initial Field Configuration: Sea Surface Temperature (SST) and Salinity (SSS) adopt satellite remote sensing inversion data (such as MODIS). Tidal boundary conditions are obtained through the TPXO database, and wind field data uses ECMWF reanalysis products. Substance Transport Module: In the oil spill diffusion simulation, set the particle release file (including location, time, pollutant type)

[0181] Part 2: Construction of the AC-PKAN Neural Network:

[0182] (1) Chebyshev Basis Transformation Layer:

[0183] Mathematical Expression: z = T k (x) = cos(karccos(x));

[0184] Where Tk is the k-th order Chebyshev basis function, used to map multimodal data to the low-dimensional feature space.

[0185] Through the TensorFlow custom layer, support dynamic adjustment of the order k (such as using higher-order terms in the shoal area to enhance texture features).

[0186] (2) Dual attention mechanism:

[0187] Internal attention (correlation between features): Dynamic weight calculation:

[0188]

[0189] where, e ij = LeakyReLU(W2 tanh(W1Z i ⊙Z j + b1) + b2);

[0190] Residual connection: Concatenate the weighted features with the original features to enhance the network's non - linear expression ability.

[0191] External attention (spatiotemporal correlation):

[0192] Spatiotemporal weight calculation:

[0193]

[0194] where, f pq = V2 tanh(V1E p ⊙O q + b3) + b4;

[0195] E p is the physical residual, and O q is the observation error.

[0196] (3) Loss function design:

[0197] Multi - objective optimization: L1 = λ p L physics + λ d L data + μL attention ;

[0198] Physical constraint term: Weighted FVCOM momentum equation residual (λ p = 0.7).

[0199] Data assimilation term: Weighted MSE between the observed data and the model prediction (λ d = 0.2).

[0200] Attention regularization term: Suppress redundant features (such as salinity fluctuations far from the coast, μ = 0.1).

[0201] Part Three: Model training and optimization:

[0202] (1) Data input and augmentation:

[0203] Satellite SAR images are used to extract texture features through LBP encoding, and wavelet transform is used to denoise buoy time series data. The tf.data API is used to build a data pipeline, which supports dynamic augmentation (such as random rotation and time translation).

[0204] (2) Training strategy:

[0205] Multi-GPU parallelism is implemented based on the Horovod framework, and the Ring AllReduce algorithm is used for gradient aggregation. The internal attention weights use AdamW (learning rate η in = 10 -4 ), and the external attention weights use a larger learning rate (η ex = 10 -3 ).

[0206] (3) Regularization and acceleration:

[0207] DropBlock is introduced in the attention module, and the block size is dynamically adjusted according to the grid resolution (e.g., 3x3 blocks for a 1km grid). Quantum computing is used to optimize the calculation of attention weights, reducing the training time of a model with hundreds of billions of parameters to the minute level.

[0208] Part 4: Online update and edge deployment:

[0209] 1. Dynamic model fine-tuning, triggering mechanism: After receiving new observation data every hour, the model fine-tuning is triggered through edge devices (such as NVIDIA Jetson AGX Xavier), and only the attention weight parameters are updated. Parameter synchronization: The attention weights trained in the cloud are synchronized to the edge device using the MQTT protocol, with a latency of less than 50ms.

[0210] 2. Lightweight deployment, model compression: The model size is compressed through TensorFlow Lite, and the power consumption is reduced to less than 5W. Real-time inference: The MobileAttention-Lite model is deployed on the drone side to achieve sub-second flow field prediction and oil spill diffusion warning.

[0211] (4) Assimilation application layer:

[0212] Online update mechanism: After receiving new observation data every hour, the model fine-tuning is triggered through edge devices (such as drones), and only the attention weight parameters are updated (the Chebyshev basis transformation layer is frozen). Decision support: Output 3D flow field animations and salinity contour maps to support scenarios such as port scheduling and fishery resource management.

[0213] Through the collaboration of Ensemble Optimal Interpolation (EnOI) and the dynamic error covariance update of BiLSTM, high-precision and strong real-time ocean environment prediction are achieved, focusing on solving the following problems: Lack of error covariance flow dependence: Traditional static error covariance cannot capture the spatio-temporal evolution of ocean dynamic processes (such as the change of the Kuroshio path); Computational efficiency bottleneck: The optimization requirements of large-scale matrix operations (such as SVD decomposition) on heterogeneous computing platforms; Extreme event response delay: Insufficient assimilation timeliness of processes such as subsurface heat transport.

[0214] The technical process of the assimilation application layer is as follows: Data collection → Preprocessing (tide separation / LBP enhancement) → EnOI assimilation (B / R covariance dynamic update) → Model update → Result verification and visualization. The specific implementation steps of the core three parts are as follows:

[0215] The first part: Implementation of Ensemble Optimal Interpolation (EnOI) assimilation:

[0216] Step 1.1: Construction of the error covariance matrix:

[0217] Background error covariance (B): Calculate the spatial correlation function based on FVCOM historical forecast data (samples in the past 10 years), and use heterogeneous computing to accelerate the SVD decomposition (cooperation of CPU + GPU);

[0218] Introduce the BiLSTM dynamic update flow dependence weight, and the formula is as follows:

[0219] B t = BiLSTM(B t-1 , ut, vt);

[0220] Where ut and vt are the sea current velocity components and tidal components.

[0221] Observation error covariance (R):

[0222] Construct the initial covariance matrix based on the satellite data quality assessment report (such as RMSE = 5cm of Jason-3); Dynamically adjust the weight through the BiLSTM time series prediction residual, and the formula is as follows:

[0223] R t = α·R t-1 + β·C;

[0224] Where α = 0.8 and β = 0.2 are forgetting factors.

[0225] Step 1.2: Optimization of the assimilation algorithm for incremental update strategy: After receiving new data every hour, only update the attention weight parameters (freeze the physical model), and the computational efficiency is increased by 40%; Adopt the dynamic observation scheduling algorithm: Based on the extreme weather probability predicted by LSTM, automatically trigger the encrypted monitoring of the drone cluster.

[0226] Part 2: BiLSTM Dynamic Error Covariance Update:

[0227] Step 2.1: Data Input and Feature Engineering Input sequence: ocean current velocity (10-day history), tidal phase (3-day prediction), temperature gradient (vertical stratification); Feature extraction: Use a BiLSTM bidirectional network to capture long-term dependencies and output flow dependence weights.

[0228] Step 2.2: Model Training and Optimization Taking the South China Sea assimilation data from 2015 to 2025 (1000+ samples) as an example; Minimize the prediction error and the positive definiteness constraint of the covariance matrix. The loss function formula:

[0229] L2 = ∥y t -H(x t )∥ 2 2 + λ∥W∥ 2 F ;

[0230] where H is the observation operator and λ = 0.01 is the regularization term coefficient.

[0231] Part 3: Result Verification and Visualization:

[0232] Step 3.1: Verification metrics, such as Kuroshio flow error, prediction accuracy of marine heatwaves (such as Blob events), etc.

[0233] Step 3.2: Interactive display:

[0234] The platform functions support 3D flow field animation, salinity isohypse maps, and heatwave diffusion simulation paths; Provide a data download interface (NetCDF format).

[0235] In summary, the technical solutions of this embodiment include:

[0236] (1) Based on satellite orbit error correction and buoy drift prediction, dynamically adjust the data source access priority (such as preferentially calling Argo buoy data in case of sudden salinity anomalies) to improve the timeliness of data fusion. Break through the traditional static data source selection mechanism and introduce a real-time multi-source data quality evaluation model.

[0237] (2) UAV SAR image enhancement technology, using local binary pattern (LBP) coding and wavelet transform for denoising, and combining the deployment of the MobileAttention-Lite model on edge devices to achieve real-time extraction of shoal terrain features. Combine image processing algorithms with edge computing to solve the problem of real-time parsing of high-resolution data in shoal areas.

[0238] (3) Space-time alignment interpolation optimization method. Based on the FVCOM grid system, it uses trilinear interpolation and adaptive resampling techniques to downsample satellite daily data to the hourly level, and dynamically updates the flow dependence weights through BiLSTM. By fusing statistical interpolation and machine learning time series prediction, it solves the problem of spatio-temporal resolution differences in multi-source data.

[0239] (4) Outlier intelligent correction mechanism. After removing outliers based on the 3σ principle, it uses the LSTM time series model to fill the salinity gap caused by the drift of Argo floats, and combines the isolation forest to detect sudden turbidity peaks. Construct "statistical filtering + time series prediction".

[0240] (5) AC-PKAN neural network architecture. It embeds the Chebyshev basis transformation layer and the double attention mechanism, and realizes multi-modal data fusion through physical equation residual weighting (λp = 0.7) and observation error weighting (λd = 0.2). The convergence speed is significantly improved compared with traditional PINN. It breaks through the weak constraint of traditional neural networks on physical laws and establishes a convergence guarantee supported by mathematical proofs.

[0241] (6) Dynamic error covariance update algorithm. Based on the BiLSTM network to predict the long-term dependence relationship between sea current velocity and tidal components, it dynamically adjusts the flow dependence weights of the error covariance matrix (formula: Bt = BiLSTM(Bt-1, ut, vt)). By introducing the time series prediction model into the dynamic update of the error covariance, it solves the assimilation bottleneck in complex scenarios such as the change of the Kuroshio path.

[0242] (7) Real-time decision support system. While outputting 3D flow field animations and salinity isograms, it constructs an oil spill diffusion path simulation module to support minute-level responses in scenarios such as port scheduling and fishery resource management. It directly embeds the assimilation results into the ocean management business system to achieve a closed-loop of "data - model - decision".

[0243] (8) Edge computing-driven online update mechanism. It receives new observation data every hour through the MQTT protocol, and only fine-tunes the attention weight parameters (freeze the physical model). The parameter synchronization delay is less than 50ms. It proposes a "freeze - fine-tune" incremental update strategy to balance computational efficiency and model accuracy.

[0244] (9) Heterogeneous data fusion storage architecture. It uses Ceph distributed storage for raw data (NetCDF format), ClickHouse for storing feature data, and the MinIO object storage system to support PB-level data management, and realizes automatic annotation and version control of data dictionaries. It constructs a three-level storage system of "raw data - feature data - metadata" to solve the problem of multi-source heterogeneous data management.

[0245] (10) Heterogeneous computing accelerates SVD decomposition, uses CPU + GPU to collaboratively batch-process the SVD decomposition of a parameter model with hundreds of billions of parameters, and improves computing efficiency. A method for optimizing a heterogeneous computing platform for ocean assimilation, which supports dynamic task allocation and load balancing. A matrix operation acceleration algorithm based on a GPU-like card reduces the operation latency of matrices with hundreds of billions of elements.

[0246] (11) Quantum acceleration assimilation algorithm, uses quantum computing to optimize the assimilation algorithm, shortens the training time of a parameter model with hundreds of billions of parameters to the minute level, and improves the response speed.

[0247] (12) Adaptive regularization technology, introduces DropBlock regularization in the BiLSTM layer, and the block size is dynamically adjusted according to the grid resolution (e.g., a 3x3 block is taken for a 1km grid) to suppress overfitting.

[0248] (13) Multilevel fusion architecture, adopts a hierarchical fusion strategy at the data level (such as SAR image registration), feature level (such as LBP texture extraction), and decision level (such as typhoon path prediction) to improve the overall perception accuracy.

[0249] (14) Lightweight deployment driven by edge computing, deploys the MobileAttention-Lite model on the drone side, realizes second-level flow field prediction and oil spill diffusion warning, and the power consumption is lower than 5W.

[0250] (15) Ensemble Optimal Interpolation (EnOI) assimilation framework, combines BiLSTM dynamic error covariance update to achieve real-time optimization of mesoscale ocean processes (such as Kuroshio intrusion). Claims: A method for dynamically updating the flow-dependent background error covariance, adjusts the weights of BiLSTM time series prediction. An incremental observation scheduling algorithm triggers encrypted monitoring of drones based on the extreme weather probability predicted by LSTM.

[0251] Refer to Figure 3 , the embodiment of the present application also provides a system for fusing and assimilating multi-source ocean observation data, which can implement the above-mentioned method for fusing and assimilating multi-source ocean observation data. The system includes:

[0252] A data acquisition unit for acquiring multi-source ocean observation data; wherein, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices;

[0253] A format conversion unit for uniformly converting the timestamps and spatial grids of the multi-source ocean observation data into the data format of the FVCOM numerical model;

[0254] A model construction unit for constructing a data fusion and assimilation model, where the data fusion and assimilation model includes an FVCOM physical model and an AC-PKAN neural network;

[0255] A model training and adjustment unit for training the data fusion and assimilation model and then adjusting the weight parameters of the data fusion and assimilation model according to the dynamic observation data currently obtained by the edge device;

[0256] A data fusion unit for fusing and assimilating the multi-source ocean observation data by using the data fusion and assimilation model to obtain multi-source fused ocean observation data.

[0257] It can be understood that the content in the above method embodiments is applicable to the system embodiments of the present application. The functions specifically implemented in the system embodiments of the present application are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those in the above method embodiments.

[0258] An embodiment of the present application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the method of the embodiment of the present application is implemented. The electronic device can be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.

[0259] It can be understood that the content in the above method embodiments is applicable to the device embodiments of the present application. The functions specifically implemented in the device embodiments of the present application are the same as those of the method of the present application, and the beneficial effects achieved are also the same as those of the method of the present application.

[0260] Please refer to Figure 4 , Figure 4 which shows the hardware structure of an electronic device in another embodiment. The electronic device includes:

[0261] A processor 401, which can be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solution provided by the embodiment of the present application;

[0262] The memory 402 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), etc. The memory 402 can store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 402 and are called by the processor 401 to execute the methods of the embodiments of this application;

[0263] The input / output interface 403 is used to implement information input and output;

[0264] The communication interface 404 is used to implement communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or through wireless means (such as mobile network, WIFI, Bluetooth, etc.);

[0265] The bus 405 transmits information between the various components of the device (such as the processor 401, the memory 402, the input / output interface 403, and the communication interface 404);

[0266] Among them, the processor 401, the memory 402, the input / output interface 403, and the communication interface 404 are communicatively connected to each other inside the device through the bus 405.

[0267] The embodiments of this application also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the method of this application.

[0268] It can be understood that the content in the above method embodiments is applicable to the embodiments of this storage medium. The functions specifically implemented by the embodiments of this storage medium are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above method embodiments.

[0269] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory can include high-speed random access memory, and can also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory optionally includes a memory remotely disposed relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0270] The embodiments described in the embodiments of the present application are to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0271] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than those shown in the figures, or combine certain steps, or different steps.

[0272] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0273] Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices can be implemented as software, firmware, hardware, and their appropriate combinations.

[0274] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily need to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0275] It should be understood that in this application, "at least one (item)" means one or more, and "a plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships can exist. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist at the same time. Among them, A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one (item) of the following" or its similar expression refers to any combination of these items, including any combination of single item (item) or plural items (items). For example, at least one (item) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0276] In several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the above division of units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of systems or units can be in electrical, mechanical or other forms.

[0277] The units described above as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0278] In addition, each functional unit in various embodiments of this application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0279] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple 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 of the various embodiments of this application. The aforementioned storage medium includes: various media that can store programs such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0280] The preferred embodiments of the embodiments of this application have been described above with reference to the accompanying drawings, and thus do not limit the scope of rights of the embodiments of this application. Any modifications, equivalent replacements, and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of this application shall be within the scope of rights of the embodiments of this application.

Claims

1. A marine multi-source observation data fusion and assimilation method, characterized in that The method includes the following steps: Obtain multi-source ocean observation data; wherein, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices; Unify the time stamps and spatial grids of the multi-source ocean observation data to the data format of the FVCOM numerical model; Construct a data fusion and assimilation model; wherein, the data fusion and assimilation model includes an FVCOM physical model and an AC-PKAN neural network; Train the data fusion and assimilation model, and then adjust the weight parameters of the data fusion and assimilation model according to the dynamic observation data currently obtained by the edge device; Use the data fusion and assimilation model to fuse and assimilate the multi-source ocean observation data to obtain multi-source fused ocean observation data.

2. The marine multi-source observation data fusion and assimilation method according to claim 1, characterized in that The obtaining of the multi-source ocean observation data includes the following steps: Obtain the sea current and temperature field output by the FVCOM numerical model; Obtain the sea surface height, sea surface temperature, and chlorophyll concentration obtained by satellite remote sensing inversion; Obtain the salinity, sea surface current field, vertical flow velocity, and temperature profile collected by buoys and submersible buoys in the edge device; Obtain the beach topography and local flow field collected by unmanned aerial vehicles and unmanned boats in the edge device.

3. A method for fusing and assimilating marine multi-source observation data according to claim 1, characterized in that, The unifying the time stamps and spatial grids of the multi-source ocean observation data to the data format of the FVCOM numerical model includes the following steps: Remove noise and outliers from the multi-source ocean observation data, and supplement missing values in the multi-source ocean observation data; Perform dimension unification, spatial coordinate alignment, and time series alignment on the multi-source ocean observation data to obtain normalized data; Using the velocity field output by the FVCOM numerical model as a benchmark, fuse the normalized data through weighted average, and then unify the time stamps and spatial grids of the normalized data to the triangular unstructured grid of the FVCOM numerical model.

4. A method for fusing and assimilating marine multi-source observation data according to claim 1, characterized in that, The constructing of the data fusion and assimilation model includes the following steps: Embed the Navier-Stokes equation and the mass transport equation in the FVCOM model, and set the configuration parameters and boundary conditions of the FVCOM model to obtain the FVCOM physical model; Construct the AC-PKAN neural network based on the Chebyshev basis transformation layer and the dual attention mechanism; wherein, the Chebyshev basis transformation layer is used to map the multi-source ocean observation data to the Chebyshev polynomial space; the dual attention mechanism includes internal attention and external attention, the internal attention is used to dynamically weight the importance of different Chebyshev basis functions, and the external attention is used to associate the spatio-temporal correlation between the physical equation residuals and the observation data.

5. A method for fusing and assimilating marine multi-source observation data according to claim 4, characterized in that, The expression for constructing the Chebyshev basis transformation layer is as follows: z = T k (x) = cos(karccos(x)); where, T k is the k-th order Chebyshev basis function, which is used to map the multi-source ocean observation data into the Chebyshev polynomial space; Through the TensorFlow custom layer, dynamically adjust the order k; The expression for constructing the internal attention is as follows: where e ij = LeakyReLU(W2tanh(W1Z i ⊙Z j + b1)+ b2);; The residual connection is to concatenate the weighted feature and the original feature; The expression for the external attention is: where f pq = V2tanh(V1E p ⊙O q + b3)+ b4; E p is the physical residual, and O q is the observation error; Design the loss function of the AC-PKAN neural network as: L1 = λ p L physics + λ d L data + μL attention ; Among them, L1 is the loss function of the AC-PKAN neural network, L physics is the physical constraint term, L data is the data assimilation term, L attention is the attention regularization term, λ p 、λ d 、μ are the coefficients of the corresponding terms respectively.

6. The marine multi-source observation data fusion and assimilation method according to claim 1, wherein Fusing and assimilating the multi-source ocean observation data by using the data fusion and assimilation model, to obtain multi-source fused ocean observation data, includes the following steps: Construct a background error covariance, and dynamically update the background error covariance through BiLSTM, so as to dynamically update the weights of the data fusion and assimilation model; The expression for updating the background error covariance is as follows: B t = BiLSTM(B t-1 , ut, vt); Among them, B t is the background error covariance at time t, and B t-1 is the background error covariance at time t-1. ut and vt are the ocean current velocity component and the tidal component respectively; Construct an observation error covariance, and dynamically adjust the observation error covariance through the BiLSTM time series prediction residuals, so as to dynamically adjust the weights of the data fusion and assimilation model; The expression for updating the observation error covariance is as follows: R t = α·R t-1 + β·C; wherein, R t is the observation error covariance at time t, and R t-1 is the observation error covariance at time t-1, α and β are different forgetting factors respectively, and C is the residual variance; Minimize the prediction error and the positive definiteness constraint of the covariance matrix; the covariance matrix includes the background error covariance and the observation error covariance; The loss function for the minimization is: L2 = ∥y t -H(x t )∥ 2 2 + λ∥W∥ 2 F ; where L2 is the loss function for the minimization, H is the observation operator, and λ is the regularization term coefficient; Use the data fusion and assimilation model with adjusted weights to fuse and assimilate the multi-source ocean observation data, to obtain the multi-source fused ocean observation data.

7. A marine multi-source observation data fusion and assimilation method according to any one of claims 1 to 6, characterized in that The method further includes the following steps: Output at least one of a three-dimensional flow field animation, a salinity contour map, or a heat wave diffusion simulation path according to the multi-source fused ocean observation data, and further determine the results of ocean environment prediction and decision-making.

8. An ocean multi-source observation data fusion and assimilation system, characterized in that, The system includes: A data acquisition unit, configured to acquire multi-source ocean observation data; wherein, the multi-source ocean observation data includes numerical simulation data output by the FVCOM numerical model, ocean surface parameters obtained by satellite remote sensing inversion, and dynamic observation data of edge devices; A format conversion unit, configured to uniformly convert the timestamps and spatial grids of the multi-source ocean observation data into the data format of the FVCOM numerical model; A model construction unit, configured to construct a data fusion and assimilation model; wherein, the data fusion and assimilation model includes an FVCOM physical model and an AC-PKAN neural network; A model training and adjustment unit, configured to train the data fusion and assimilation model, and further adjust the weight parameters of the data fusion and assimilation model according to the dynamic observation data currently acquired by the edge device; A data fusion unit, configured to use the data fusion and assimilation model to fuse and assimilate the multi-source ocean observation data, to obtain multi-source fused ocean observation data.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

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

Images