Environmental data processing method and system based on ocean engineering

Abstract

The invention discloses an environmental data processing method and system based on ocean engineering, and relates to the technical field of data processing, and the method comprises the steps: receiving an original observation data flow through a multi-source data preprocessing module, and carrying out the dynamic noise filtering and abnormal value adaptive detection; fusing the multi-source heterogeneous data through a multi-scale data fusion module, and embedding the fused multi-source heterogeneous data into a marine kinetic equation as a soft constraint; non-linear evolution features are extracted from the fusion data through a feature extraction and state representation module, and a high-dimensional environment state vector is constructed; real-time prediction of model parameters is executed through online learning and an inference engine; and performing uncertainty propagation calculation on the processing flow through a confidence evaluation module and generating a final environment state report. According to the method, the adaptive capacity of data preprocessing can be remarkably improved, the physical consistency of multi-source data fusion is improved, the nonlinear evolution law of ocean phenomena is accurately captured, and continuous online optimization and edge side low-delay response of model parameters are achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an environmental data processing method and system based on marine engineering. Background Technology

[0002] Marine engineering, as a crucial component of emerging industries, encompasses several key areas, including offshore oil and gas development, marine renewable energy utilization, deep-sea resource exploration, and marine environmental monitoring. Its development heavily relies on the accurate perception, efficient processing, and intelligent analysis of complex marine environmental data. The marine environment is characterized by high dynamism, strong nonlinearity, and multi-scale coupling, involving multi-dimensional heterogeneous data sources such as hydrology, meteorology, geology, and ecology. Data acquisition is typically accomplished collaboratively through multiple platforms, including buoys, underwater moorings, satellite remote sensing, underwater robots, and shore-based radar. However, traditional marine environmental data processing methods generally suffer from fragmented processing workflows, low information fusion efficiency, delayed anomaly identification, and insufficient model generalization ability when dealing with large-scale, high-concurrency, and highly spatiotemporally heterogeneous observational data. These shortcomings make it difficult to meet the stringent requirements of modern marine engineering for real-time performance, robustness, and decision support accuracy.

[0003] Among these, environmental data processing based on marine engineering focuses on extracting environmental state characteristics with physical significance and engineering value from multi-source heterogeneous observation systems, and supporting subsequent core tasks such as structural safety assessment, operational window prediction, disaster early warning, and resource scheduling. The core objective of this technical direction is to construct an intelligent processing framework that can adapt to dynamic changes in the marine environment, effectively integrate spatiotemporally unevenly distributed data streams, and possess online learning and error correction capabilities, so as to achieve a closed-loop transformation from raw observation to higher-order cognition.

[0004] Existing technologies for marine environmental data processing suffer from multiple structural defects: First, data preprocessing typically employs static thresholds or fixed filtering windows, failing to adapt to variations in noise characteristics across different sea areas, seasons, and extreme weather conditions, leading to over-smoothing of effective signals or missed outliers. Second, multi-source data fusion often relies on simple weighted averaging or linear interpolation, ignoring the fundamental differences in spatial resolution, temporal synchronization, and physical dimensions among different sensors, resulting in distorted fusion results. Third, feature extraction often employs general statistical indicators or shallow machine learning models, lacking the embedding of inherent physical constraints on marine dynamic processes, making it difficult to capture the nonlinear evolution of key marine phenomena such as eddies, internal waves, and fronts. Furthermore, existing system architectures are mostly offline batch processing modes, lacking edge computing and cloud-edge collaboration mechanisms, failing to support the low-latency response requirements of offshore platforms. Finally, the lack of a mechanism for quantifying the uncertainty of processing results leaves downstream applications without reliable confidence bases for risk assessment and decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide an environmental data processing method and system based on marine engineering, so as to solve the problems of insufficient adaptive capability of data preprocessing, poor physical consistency of multi-source data fusion, and inaccurate nonlinear feature extraction in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] An environmental data processing system based on marine engineering, comprising the following components:

[0008] The multi-source data preprocessing module is used to receive raw observation data streams from buoys, underwater moorings, satellite remote sensing, underwater robots and shore-based radar, and perform dynamic noise filtering and adaptive outlier detection based on physical constraints.

[0009] The multi-scale data fusion module is connected to the multi-source data preprocessing module and is used to fuse the preprocessed multi-source heterogeneous data under a unified spatiotemporal grid framework. The fusion process embeds ocean dynamics equations as soft constraints.

[0010] The feature extraction and state characterization module, connected to the multi-scale data fusion module, is used to extract nonlinear evolution features characterizing key marine phenomena from the fused data and construct a high-dimensional environmental state vector.

[0011] An online learning and inference engine, connected to the feature extraction and state representation module, is deployed on edge computing nodes and cloud servers to perform online updates of model parameters and real-time prediction of environmental states.

[0012] The confidence assessment module, connected to the online learning and inference engine, is used to perform uncertainty propagation calculations on the output results of each stage in the processing flow, and generate a final environmental status report with confidence intervals.

[0013] To achieve the above objectives, the present invention provides the following technical solution: On the other hand, an environmental data processing method based on marine engineering, the specific steps of which are as follows:

[0014] Step S110: The multi-source data preprocessing module receives and processes raw marine environmental data from multiple observation platforms, and performs physical constraint-based dynamic noise filtering and adaptive outlier detection.

[0015] In step S120, the multi-source heterogeneous data processed in step S110 is fused under a unified spatiotemporal grid framework through the multi-scale data fusion module. The fusion process embeds ocean dynamics equations as soft constraints.

[0016] Step S130: Through the feature extraction and state characterization module, nonlinear evolution features characterizing key ocean phenomena such as vortices, internal waves, and fronts are extracted from the fused data obtained in step S120, and a high-dimensional environmental state vector is constructed.

[0017] Step S140: Using an online learning and inference engine, based on the environment state vector constructed in step S130, perform online updates of model parameters and real-time prediction of environment state.

[0018] In step S150, the confidence assessment module performs uncertainty propagation calculations on the processing flow from steps S110 to S140 to generate a final environmental status report with confidence intervals.

[0019] Preferably, the multi-source data preprocessing module includes a dynamic noise model construction unit and a physical constraint anomaly detection unit. The dynamic noise model construction unit establishes a time-varying noise power spectral density model based on historical data statistical characteristics for different sensor types and observation sea areas. For hydrological sensors, the noise model parameters are correlated with historical statistical values ​​of ocean current velocity and temperature gradient; for meteorological sensors, the noise model parameters are correlated with sliding window statistical values ​​of wind speed and air pressure change rate. This unit dynamically adjusts the cutoff frequency and window length of the filtering algorithm according to the real-time observation data stream. The physical constraint anomaly detection unit does not rely on a fixed threshold but instead substitutes the observation data into simplified ocean physical conservation equations for residual calculation. For temperature, salinity, and depth data, residuals are detected based on the mass and thermohaline conservation equations; for ocean current data, residuals are detected based on a simplified form of the momentum equation. When the residual sequence exceeds a multiple of the standard deviation of its historical statistical distribution... When the data point at that moment is determined to be a physically inconsistent outlier, it is marked and interpolated for repair.

[0020] Furthermore, the multi-scale data fusion module employs a variational assimilation framework, with its objective function defined as the difference between the observed data and the background field, and the degree to which the data satisfies physical constraints. This module constructs a unified three-dimensional spatiotemporal grid, projecting observed data of different resolutions and time lengths onto this grid through the observation operator H. Objective function Represented as: in, Let these be the state variables to be analyzed. As background scene, The background error covariance matrix, For the observation vector, The observation error covariance matrix, These are the physical constraint weighting coefficients. For physical constraints, Represents state variables The discrete form of the ocean dynamics governing equations to be satisfied. This is achieved by minimizing the objective function. The goal is to achieve the optimal balance between data fitting and adherence to physical laws, thus realizing mechanism-guided data fusion.

[0021] Furthermore, the feature extraction and state representation module employs a hybrid architecture combining deep neural networks and empirical mode decomposition (EMD). This module first performs EMD on the fused multidimensional data field, separating intrinsic mode functions (EMFs) representing different time scales. Subsequently, the EMFs at each scale, along with the original data, are input into a long short-term memory (LSTM) network with an attention mechanism. This attention mechanism is designed to focus on regions with large spatial gradients and rapid temporal changes, typically corresponding to vortex edges or frontal positions. The network ultimately outputs a high-dimensional feature vector that simultaneously encodes the multi-scale statistical properties of the data, nonlinear dynamic evolution patterns, and spatial distribution information of key physical structures.

[0022] Preferably, the online learning and inference engine employs a strategy combining federated learning and model distillation. A lightweight student model is deployed on edge computing nodes, responsible for real-time inference and preliminary anomaly detection on local high-frequency data. A complex teacher model is deployed on a cloud server, responsible for integrating data and model updates from multiple edge nodes and training and optimizing the global model. The student model is updated periodically by receiving distilled parameters from the teacher model. The teacher model updates its global parameters by aggregating model gradients uploaded from each edge node (processed with differential privacy protection) and using a federated averaging algorithm. This mechanism ensures that the model can continuously adapt to the dynamic changes in different marine environments while meeting the requirements of low latency and data privacy protection at the edge.

[0023] Furthermore, the confidence evaluation module is based on Bayesian deep learning and the Monte Carlo Dropout method. This module enables Dropout layers in both the nonlinear feature extraction and state representation network and the prediction network of the online learning and inference engine, and maintains these layers during the inference phase. For the same input, it performs... Each forward propagation involves randomly discarding different neurons, thus obtaining... A different set of predicted outputs Final predicted value Take the mean of this set to predict uncertainty. This is obtained by calculating the variance of the set:

[0024] This module also integrates the observation error covariance matrix R and the background error covariance matrix from the multi-scale data fusion module. The uncertainty is propagated step by step to the final status report through the error propagation chain law, generating confidence intervals for each environmental state parameter (such as flow rate and temperature).

[0025] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0026] This invention overcomes the limitations of static thresholding methods by constructing a dynamic noise model and anomaly detection mechanism based on physical constraints. It significantly improves the adaptability of the data preprocessing stage to different sea areas, seasons and extreme weather conditions, and effectively avoids loss of effective signals and missed detection of outliers.

[0027] This invention employs a variational assimilation framework and embeds ocean dynamics equations as soft constraints for multi-source data fusion, fundamentally solving the problem that traditional weighted averaging or linear interpolation methods ignore the essential differences and physical laws of sensors, and significantly improving the physical consistency and spatial continuity of the fusion results.

[0028] This invention extracts nonlinear features through a hybrid architecture of deep neural networks and empirical mode decomposition, and uses an attention mechanism to focus on key physical structures, which can more accurately capture the nonlinear evolution laws of ocean phenomena such as eddies and internal waves, thereby improving the depth of state representation and engineering value.

[0029] The cloud-edge collaborative online learning and inference engine designed in this invention combines federated learning and model distillation to achieve continuous online optimization of model parameters and low-latency response on the edge side, meeting the stringent requirements of offshore platforms for real-time intelligent processing.

[0030] This invention introduces a full-process uncertainty quantification mechanism based on Bayesian deep learning, providing clear confidence levels for each processing result and the final decision, greatly enhancing the reliability and scientific rigor of marine engineering risk assessment and emergency decision-making. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall technical solution architecture of the environmental data processing method based on marine engineering proposed in this invention;

[0032] Figure 2 This is a schematic diagram of the core principle framework of the multi-scale data fusion module in this invention. Detailed Implementation

[0033] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0034] Example 1

[0035] In offshore oil and gas exploration and development projects, particularly in the production area of ​​a deep-water oil and gas field in the South China Sea, real-time and accurate perception and prediction of the marine environment are crucial for ensuring the safety of drilling platforms, optimizing production operations, and preventing environmental risks such as oil spills. This sea area has complex sea conditions, frequently experiencing mesoscale eddies, strong internal waves, and rapidly changing ocean fronts. These phenomena directly affect the stability of subsea production systems and the stress state of subsea pipelines. Traditional environmental monitoring methods relying on single data sources or simple data splicing are ill-equipped to handle the significant differences in spatiotemporal scale, accuracy, and physical consistency between multi-source heterogeneous data, and cannot quantify the uncertainty of prediction results, leading to blind spots and risks in engineering decision-making. This embodiment will detail the specific implementation process of the system and method of the present invention in this complex engineering scenario.

[0036] See Figure 1 This system includes a multi-source data preprocessing module, a multi-scale data fusion module, a feature extraction and state representation module, an online learning and inference engine, and a confidence assessment module. These modules are connected sequentially to form a complete processing chain from raw data input to the generation of a confidence-based environmental state report.

[0037] First, the multi-source data preprocessing module begins operation. Deployed on the edge server of the offshore platform, this module directly receives raw data streams from various observation platforms deployed around the oil and gas field. This data includes: ocean current velocity, direction, temperature, salinity, and pressure profiles collected every second by moored buoys and bottom-dwelling moorings; remote sensing images of sea surface height anomalies, sea surface temperature, and chlorophyll concentration provided several times daily by over-the-top satellites; temperature, salinity, and depth profiles and subtle seabed topographic changes collected every six hours by autonomous underwater vehicles cruising along predetermined routes; and sea surface current vector maps retrieved every ten minutes by shore-based high-frequency ground-wave radar. These data exhibit significant differences in transmission format, sampling frequency, spatial coverage, and observation error characteristics.

[0038] The multi-source data preprocessing module includes a dynamic noise model construction unit and a physical constraint anomaly detection unit. The dynamic noise model construction unit establishes a time-varying noise power spectral density model offline, based on one year of historical observation data, for each type of sensor and its specific sea area deployment. For current meters, the noise model parameters are correlated with the square root of the historically statistically analyzed average current velocity and vertical shear strength at that location; for temperature, salinity, and depth sensors, the noise model parameters are correlated with the standard deviations of historical temperature and salinity gradients. In real-time processing, this unit uses a five-minute sliding window to calculate the statistical characteristics of the real-time data stream within that window and dynamically adjusts the parameters of subsequent filtering algorithms. For example, when an increased fluctuation in the current velocity is detected, the cutoff frequency of the low-pass filter is automatically relaxed to avoid filtering out the real high-frequency signals caused by vortices; conversely, during periods of stable flow, a more stringent filter is used to suppress instrument noise.

[0039] The physical constraint anomaly detection unit synchronously performs physical consistency checks on the input data stream. Instead of using fixed numerical thresholds, this unit substitutes real-time observations into simplified, discretized ocean physics conservation equations to calculate residuals. For temperature, salinity, and depth (TDM) data sequences, it detects residuals based on the mass and thermohaline conservation equations of the one-dimensional advection diffusion equation; for ocean current vector data, it detects residuals based on the simplified momentum equation (considering Coriolis force, pressure gradient force, and linear friction). This unit maintains a 30-day historical residual sequence sliding window and calculates the standard deviation of the residuals within this window in real time. When the absolute value of the residual calculated for a newly arrived data point exceeds three times the current standard deviation, the data point is determined to be a physically inconsistent anomaly. For example, if a buoy reports an anomaly of a sudden 10-degree Celsius temperature drop, but surrounding moorings and satellite data show no corresponding cold signals, and this change causes the thermohaline conservation equation residuals to far exceed the historical fluctuation range, then this point is marked as an anomaly. For the marked outliers, this unit employs a spatiotemporal kriging interpolation method to repair them using surrounding normal observation points and background field information, generating physically more reasonable replacement values. After processing by this module, the original observation data stream is transformed into a standardized data product with spatiotemporal alignment, noise suppression, and preliminary assurance of physical consistency, laying the foundation for subsequent fusion.

[0040] Next, the multi-scale data fusion module begins operation. See also... Figure 2 This module employs a variational assimilation framework. Its core task is to optimally fuse multi-source heterogeneous data from the aforementioned preprocessing module on a unified three-dimensional spatiotemporal analysis grid, ensuring that the fusion result conforms as closely as possible to known ocean dynamics. For the oil and gas field area in this embodiment, this module constructs a three-dimensional spatiotemporal grid with a horizontal resolution of one kilometer, a vertical division into twenty layers, and a time step of half an hour, covering a sea area with a radius of fifty kilometers centered on the platform.

[0041] This module defines an objective function to measure the deviation between the analysis field and the background field, the fit between the analysis field and various observation data, and the degree to which the analysis field satisfies physical constraints. Specifically, the objective function is expressed as the weighted sum of squared deviations between the state variables to be analyzed and the background field, the weighted sum of squared deviations between the observation vector and the analysis field after projection through the observation operator, and a physical constraint penalty term. The background field is obtained from the previous day's assimilation results through short-term forecasts using ocean numerical models, and its error characteristics are described by the background error covariance matrix. The observation error covariance matrix integrates the uncertainty information from various data sources provided by the multi-source data preprocessing module after evaluation by a dynamic noise model. The physical constraint penalty term is a key innovation; its weight coefficients are dynamically adjusted based on sea area stability and data density. In this embodiment, the physical constraint term is based on a simplified version of the original equations, requiring the fused three-dimensional temperature, salinity, and current field to satisfy certain soft constraints in terms of continuity and momentum and mass conservation.

[0042] The fusion process involves iteratively solving a numerical optimization algorithm to minimize the aforementioned objective function, thereby obtaining the optimal environmental state analysis field on the spatiotemporal grid. The observation operator maps the state variables on the grid to the actual locations and types of various observations. For example, it maps temperature, salinity, and ocean current velocity on the grid to the specific observation depths and locations of buoys, moorings, and underwater robots, and maps sea surface height to the footprint of a satellite altimeter. Through this process, high spatiotemporal resolution satellite remote sensing data, high-precision fixed-point buoy and mooring data, and underwater robot data with flexible spatial coverage are organically fused together to generate a physically consistent, spatiotemporally continuous, and more accurate comprehensive three-dimensional map of the marine environment than any single data source.

[0043] Subsequently, the feature extraction and state characterization module performs in-depth analysis on the fused 3D spatiotemporal data field. This module employs a hybrid architecture combining deep neural networks and empirical mode decomposition (EMD). First, it performs EMD on the temperature, salinity, and flow time series of key areas (such as platform locations and pipeline routes), adaptively decomposing the signal into several eigenmode functions ranging from high to low frequencies and a trend term. This process decomposes complex non-stationary signals into fluctuation components at different time scales; for example, it separates periodic fluctuations of tens of minutes caused by internal waves, periodic fluctuations of several hours to tens of hours caused by tides, and low-frequency variations of several days to several weeks caused by mesoscale eddies.

[0044] Then, these multi-scale modal function sequences, along with the spatial distribution maps of the original fused data field on key vertical layers and horizontal slices, are input into a specially designed long short-term memory (LSTM) network. This network incorporates a spatial attention mechanism, trained to automatically focus on grid points or regions with large spatial gradients and dramatic temporal variations. In engineering practice, these regions often correspond to the location of ocean fronts, the edge shear zones of vortices, or the leading edges of internal wave propagation, and are the areas most significantly affected by stress on platforms and pipelines. By learning the correlation between these features in historical data and the evolution of known ocean phenomena, the LTM network extracts highly abstract nonlinear evolutionary features from the input multi-modal, multi-scale data. Ultimately, the network outputs a 512-dimensional environmental state feature vector. This vector not only encodes statistical information on basic physical quantities such as temperature, salinity, and flow velocity at the current moment, but also more deeply encodes the dynamic representation of key engineering concerns such as the rotational intensity and direction of movement of vortices, the amplitude and propagation path of internal waves, and the intensity and gradient direction of fronts, achieving a transformation from raw data to a high-level interpretable engineering state.

[0045] The online learning and inference engine is responsible for real-time prediction and adaptive model updates using the aforementioned state representations. This engine employs an architecture combining federated learning and model distillation. A lightweight student model is deployed on the edge computing nodes of the offshore platform. Essentially, this student model consists of the long short-term memory portion of the aforementioned nonlinear feature extraction and state representation network, along with a simplified prediction head. Using the environmental state feature vector from the previous time step and recent historical sequences as input, this student model can predict the current profile, significant wave height, and identify any abnormal current patterns such as vortices or internal waves at the platform's location within the next hour, with a latency of less than one second, thus supporting immediate operational decisions for the platform.

[0046] In a cloud data center on land, a complete and complex teacher model is deployed, encompassing a comprehensive feature extraction and prediction network. This teacher model does not directly access the raw edge data; instead, it periodically receives model parameter updates uploaded from various edge nodes. These updates consist of gradient information generated by the edge student models after training on new local data. Before uploading, these gradients undergo differential privacy processing, adding specific noise to ensure the original data is not leaked. The cloud server uses a federated averaging algorithm to aggregate encrypted gradients from multiple oil and gas field edge nodes, updating the parameters of the global teacher model. Subsequently, the teacher model uses knowledge distillation to compress the complex knowledge it has learned, generating a new, lightweight set of parameters, which is then distributed to all edge student models for updates. This mechanism allows edge models deployed in different sea areas and facing varying environmental characteristics to quickly adapt to the uniqueness of their respective regions through local learning, while also sharing global knowledge through cloud aggregation, collectively improving the model's generalization ability and prediction accuracy. The entire online learning process continues in six-hour cycles.

[0047] Finally, the confidence assessment module performs a comprehensive uncertainty assessment of the entire chain from data preprocessing to final prediction. Based on Bayesian deep learning, this module enables the Monte Carlo Dropout method in both the nonlinear feature extraction and state representation network and the prediction network of the marginal student model. Specifically, during the model inference phase, the Dropout layers in these networks remain active. For the same set of input data, the module controls the network to perform one hundred forward propagation calculations, randomly discarding a certain proportion of neurons in the network during each forward propagation. This results in one hundred slightly different prediction outputs, forming a prediction set. The final prediction value is the average of these one hundred predictions, and the uncertainty of this prediction is quantified by calculating the variance of these one hundred predictions. The larger the variance, the more uncertain the model's prediction for that input data.

[0048] Furthermore, this module constructs a complete error propagation chain. It progressively propagates and synthesizes the errors from various data sources estimated by the dynamic noise model in the multi-source data preprocessing module, the analytical errors implied in the observation error covariance matrix and background error covariance matrix in the multi-scale data fusion module, and the uncertainties in the model parameters in the online learning and inference engine, using a chain rule. Ultimately, for each key parameter in the environmental status report, such as the eastward current velocity at a depth of 100 meters at the platform, the module not only provides a predicted value but also a 95% confidence interval. For example, the report shows that "the predicted eastward current velocity at a depth of 100 meters is 0.5 meters per second for the next three hours, with a 95% confidence interval of 0.3 to 0.7 meters per second." This quantified uncertainty information is crucial for engineering decisions; for example, when deciding whether to proceed with underwater equipment installation, decision-makers can clearly understand the boundaries of predicted risk.

[0049] Through the collaborative work of the above five modules, this embodiment system realizes closed-loop processing of complex marine engineering marine environments, from multi-source data access, physical constraint fusion, intelligent feature extraction, online learning and prediction to full-process uncertainty quantification, providing highly reliable and timely intelligent decision support for safe production, emergency response and operation optimization in oil and gas fields.

[0050] Example 2

[0051] In the application of intelligent route planning and energy efficiency management for transoceanic container fleets, accurate forecasting of ocean currents, wind fields, and waves over vast ocean areas is crucial for optimizing speed, saving fuel, ensuring schedules, and avoiding severe sea conditions. Fleet operations involve multiple global routes, traversing vastly different marine environments. Available real-time observational data is sparse and uneven, and traditional atmospheric and oceanic numerical models relied upon for meteorological navigation services suffer from initial field errors and model biases. This embodiment illustrates another implementation of the system and method of the present invention in a collaborative scenario between shipborne edge devices and shore-based cloud centers.

[0052] The shipborne edge computing unit incorporates a multi-source data preprocessing module and a student model within a lightweight cloud-edge collaborative online learning and inference engine. It receives real-time data on wind speed, direction, air pressure, and temperature from the shipborne weather station; ocean current profile data from the ship's acoustic Doppler current profiler; and intermittently received buoy reports from surrounding waters and satellite remote sensing data on sea surface wind field and sea surface temperature via satellite communication. The multi-source data preprocessing module operates at the shipborne edge, and for shipborne sensors, its dynamic noise model is correlated with the ship's navigation status, adjusting the filtering intensity of wind speed and current data, for example, during high-speed navigation or in high winds and waves. The physical constraint anomaly detection unit utilizes limited ocean background field information near the ship's trajectory to perform rapid physical consistency checks on the received sparse buoy and satellite data, marking suspicious data.

[0053] Preprocessed data is asynchronously uploaded to the multi-scale data fusion module at the shore-based cloud center via satellite communication, in compressed and encrypted form. This module operates in the cloud, with a unified spatiotemporal grid covering the predicted sea area along the target shipping route for the next seven days. The background field uses the output of a global operational ocean-atmosphere forecasting model. Since direct observations along shipping routes are very sparse, the role of physical constraint embedding is even more critical. The strongly constrained analysis field in the fusion process must satisfy physical laws such as large-scale geostrophic equilibrium and temperature-salinity relationships. This allows for the effective correction and refinement of errors in the large-scale model forecast background field using limited, fragmented shipborne observations and satellite remote sensing data, particularly in the prediction of the location and intensity of mesoscale eddies and oceanic fronts.

[0054] The feature extraction and state characterization module also runs in the cloud. It analyzes the fused marine environmental field, focusing on extracting key features that affect ship resistance and navigation safety. For example, it extracts the intensity and extent of downstream or upstream regions, the effective wave height and period of wind-generated waves, and the resonant spectral features that may cause ship rolling. These features are encoded into state vectors.

[0055] The teacher model at the shore-based cloud center integrates observational data uploaded from multiple ships in the fleet with local learning results, continuously optimizing its global forecast model through federated learning. The optimized model knowledge is distilled to generate lightweight prediction model parameter packages optimized for different typical sea area characteristics, which are periodically distributed to each ship via satellite. The student model at the ship's edge receives and updates its parameters, enabling it to utilize the ship's real-time observations and received sparse data to make higher-precision, lower-latency localized predictions of ocean currents and waves along the ship's forward route for the next 24 to 48 hours, while also providing prediction uncertainties. Based on this, the captain and shore-based operations center can make optimal trade-offs between fuel costs, sailing time, and risks, dynamically adjusting routes and speeds.

[0056] The confidence assessment module operates both in the cloud and on the ship. In the cloud, it provides spatially distributed confidence maps for large-scale fused forecast results. On the ship, it provides specific confidence intervals for local predictions made by edge models. For example, predicting a two-meter downstream current in a certain sea area with a confidence interval of 1.5 to 2.5 meters provides a risk boundary for assessing fuel-saving potential; predicting a significant wave height exceeding four meters in a certain area with a 90% confidence probability provides a clear basis for detour decisions.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An environmental data processing system based on marine engineering, characterized in that, The system includes the following components: The multi-source data preprocessing module is used to receive raw observation data streams from buoys, underwater moorings, satellite remote sensing, underwater robots and shore-based radar, and perform dynamic noise filtering and adaptive outlier detection based on physical constraints. The multi-scale data fusion module is connected to the multi-source data preprocessing module and is used to fuse the preprocessed multi-source heterogeneous data under a unified spatiotemporal grid framework. The fusion process embeds ocean dynamics equations as soft constraints. The feature extraction and state characterization module, connected to the multi-scale data fusion module, is used to extract nonlinear evolution features characterizing key marine phenomena from the fused data and construct a high-dimensional environmental state vector. An online learning and inference engine, connected to the feature extraction and state representation module, is deployed on edge computing nodes and cloud servers to perform online updates of model parameters and real-time prediction of environmental states. The confidence assessment module, connected to the online learning and inference engine, is used to perform uncertainty propagation calculations on the output results of each stage in the processing flow, and generate a final environmental status report with confidence intervals.

2. The environmental data processing system based on marine engineering according to claim 1, characterized in that, The multi-source data preprocessing module includes a dynamic noise model construction unit and a physical constraint anomaly detection unit. The dynamic noise model construction unit establishes a time-varying noise power spectral density model based on historical data statistical characteristics for different sensor types and observation areas. For hydrological sensors, the noise model parameters are correlated with historical statistical values ​​of ocean current velocity and temperature gradient; for meteorological sensors, the noise model parameters are correlated with sliding window statistical values ​​of wind speed and air pressure change rate. This unit dynamically adjusts the cutoff frequency and window length of the filtering algorithm according to the real-time observation data stream. The physical constraint anomaly detection unit substitutes the observation data into simplified ocean physical conservation equations for residual calculation. For temperature, salinity, and depth data, it detects residuals based on mass and thermohaline conservation equations; for ocean current data, it detects residuals based on a simplified form of the momentum equation. When the residual sequence exceeds 3.0 times the standard deviation of its historical statistical distribution, the data point at that moment is determined to be a physically inconsistent anomaly, and is marked and interpolated for repair.

3. The environmental data processing system based on marine engineering according to claim 1, characterized in that, The multi-scale data fusion module employs a variational assimilation framework, with its objective function defined as the difference between the observed data and the background field, and the degree to which the data satisfies physical constraints. This module constructs a unified three-dimensional spatiotemporal grid, integrating observed data of different resolutions and time lengths through observation operators. Projected onto the grid, the objective function Represented as: in, Let these be the state variables to be analyzed. As background scene, The background error covariance matrix, For the observation vector, The observation error covariance matrix, These are the physical constraint weighting coefficients. For physical constraints, Represents state variables The discrete form of the ocean dynamics governing equations that must be satisfied is determined by minimizing the objective function. The goal is to achieve the optimal balance between data fitting and adherence to physical laws, thus realizing mechanism-guided data fusion.

4. The environmental data processing system based on marine engineering according to claim 1, characterized in that, The feature extraction and state representation module adopts a hybrid architecture of deep neural network and empirical mode decomposition. The module first performs empirical mode decomposition on the fused multidimensional data field to separate the intrinsic mode functions representing different time scales. Then, the mode functions of each scale and the original data are input into a long short-term memory network with an attention mechanism. The attention mechanism of the network is designed to focus on regions with large spatial gradients and drastic temporal changes. The network finally outputs a high-dimensional feature vector, which simultaneously encodes the multi-scale statistical characteristics, nonlinear dynamic evolution mode and spatial distribution information of key physical structures of the data.

5. The environmental data processing system based on marine engineering according to claim 1, characterized in that, The online learning and inference engine employs a strategy combining federated learning and model distillation. Lightweight student models are deployed on edge computing nodes to perform real-time inference and preliminary anomaly detection on local high-frequency data. Complex teacher models are deployed on cloud servers to integrate data and model updates from multiple edge nodes and to train and optimize the global model. The student models are updated periodically by receiving distilled parameters from the teacher models, while the teacher models update their global parameters by aggregating model gradients uploaded from various edge nodes and using a federated averaging algorithm.

6. The environmental data processing system based on marine engineering according to claim 1, characterized in that, The confidence evaluation module is based on Bayesian deep learning and the Monte Carlo Dropout method. In both the nonlinear feature extraction and state representation network and the prediction network of the online learning and inference engine, the module enables Dropout layers and maintains their activation during the inference phase. For the same input, it performs 100 forward propagations, randomly discarding different neurons each time, thus obtaining 100 different sets of predicted outputs. Final predicted value Take the mean of this set to predict uncertainty. This is obtained by calculating the variance of the set: This module also includes the observation error covariance matrix from the multi-scale data fusion module. With background error covariance matrix The uncertainty is transmitted step by step to the final status report through the error propagation chain law.

7. A method for processing environmental data based on marine engineering, characterized in that, The method includes the following steps: Step S110: The multi-source data preprocessing module receives and processes raw marine environmental data from multiple observation platforms, and performs physical constraint-based dynamic noise filtering and adaptive outlier detection. In step S120, the multi-source heterogeneous data processed in step S110 is fused under a unified spatiotemporal grid framework through the multi-scale data fusion module. The fusion process embeds ocean dynamics equations as soft constraints. Step S130: Through the feature extraction and state characterization module, nonlinear evolution features characterizing key ocean phenomena such as vortices, internal waves, and fronts are extracted from the fused data obtained in step S120, and a high-dimensional environmental state vector is constructed. Step S140: Using an online learning and inference engine, based on the environment state vector constructed in step S130, perform online updates of model parameters and real-time prediction of environment state. In step S150, the confidence assessment module performs uncertainty propagation calculations on the processing flow from steps S110 to S140 to generate a final environmental status report with confidence intervals.

8. The environmental data processing method based on marine engineering according to claim 7, characterized in that, In step S110, the cutoff frequency and window length of the dynamic noise filtering algorithm are dynamically adjusted according to the statistical characteristics of the real-time observed data stream with a sliding window of 5 minutes; the outlier adaptive detection and maintenance of the historical residual sequence sliding window with a length of 30 days is used to repair the marked outliers using a spatiotemporal kriging interpolation method.

9. The environmental data processing method based on marine engineering according to claim 7, characterized in that, In step S120, the unified three-dimensional spatiotemporal grid has a horizontal resolution of 1 kilometer, is divided into 20 vertical layers, and has a time step of half an hour; the physical constraint weight coefficients The system is dynamically adjusted based on sea area stability and data density.

10. The environmental data processing method based on marine engineering according to claim 7, characterized in that, In step S150, the uncertainty propagation calculation constructs a complete error propagation chain, which transmits the data source error of the multi-source data preprocessing module, the observation error covariance matrix R and background error covariance matrix B of the multi-scale data fusion module, and the model parameter uncertainty of the online learning and inference engine step by step through the chain rule. Generate 95% confidence intervals for each key parameter in the environmental status report.

Images