An ai-based marine solution toolkit system and method

Abstract

The application relates to an AI-based marine solution toolkit system and method. The system comprises a data fusion layer, an application module layer and a decision support layer. The data fusion layer comprises a data fusion model for fusion calculation of marine data to generate a marine data set; the application module layer comprises multiple functional modules for generating intermediate data with decision reference value based on the marine data set; and the decision support layer comprises a large language model for receiving user input decision requirements, performing intent recognition on the decision requirements, and performing logical reasoning on the recognized intent based on the intermediate data and the marine data set to generate a decision scheme. Through the collaborative work of the data fusion model and the large language model, the application realizes deep fusion of multi-source marine data and intelligent decision support, significantly improves the scientificity, efficiency and ease of use of marine problem decision-making, and is suitable for multiple fields such as marine navigation safety, ecological protection and blue economy.

Description

Technical Field

[0001] This invention belongs to the field of marine problem technology, and in particular relates to an AI-based marine solution toolkit system and method. Background Technology

[0002] Against the backdrop of climate change, the lives and production of people in coastal cities are directly affected by various marine disasters, prompting the exploration of solutions to various marine problems long ago. In traditional methods, the generation of solutions mainly relies on human experience for decision-making, which places high demands on the decision-maker's past experience, knowledge reserves, and value orientation, and the decision results are often highly subjective.

[0003] In recent years, with the rapid development of artificial intelligence technology, smart ocean technologies based on the Internet of Things and big data have been developed. Decision-makers can improve the efficiency and scientific rigor of decision-making by building various auxiliary decision-making systems, such as storm surge warning systems, seaweed disaster management systems, deep-sea aquaculture monitoring systems, and maritime search and rescue platforms. These systems use computers to analyze data collected by various sensors and then make decisions based on the analysis results.

[0004] However, existing decision support systems typically have the following drawbacks: (1) Existing decision support systems usually only target a single problem and rely on a single type of data source that exists independently, and cannot integrate multiple types of data such as field observation, satellite remote sensing, and numerical simulation.

[0005] (2) The existing decision support systems are isolated, with different systems belonging to different management departments. Data is not interconnected and cannot be shared, forming data silos and system silos.

[0006] (3) The existing auxiliary decision-making system has a low degree of integration. Different marine issues require the use of multiple different independent systems to deal with them. Users need to operate them one by one, which is inefficient and makes it difficult to obtain comprehensive decision support.

[0007] (4) The existing human-computer interaction methods of auxiliary decision-making systems are still mainly menu operation and data query, which cannot understand the user's natural language needs and have limited intelligence. Summary of the Invention

[0008] The purpose of this invention is to solve one of the above-mentioned technical problems by providing an AI-based marine solution toolkit system and method. By combining data fusion models and large language models, the scientific nature, efficiency and ease of use of marine problem decision-making are significantly improved.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An AI-based marine solutions toolkit system includes: The data fusion layer contains a pre-trained data fusion model, which is used to receive ocean data from different data sources and call the data fusion model to perform fusion calculations on the ocean data to generate an ocean dataset covering the target sea area. The application module layer contains multiple functional modules for different marine application scenarios. These functional modules provide marine data to the data fusion layer and generate intermediate data with decision-making reference value based on the marine dataset. The decision support layer contains a pre-trained large language model, which is configured to: receive decision requirements input by the user in natural language form; identify the intent of the decision requirements and, based on the identification results, call one or more functional modules in the application module layer to obtain intermediate data; and, based on the intermediate data and the ocean dataset, perform logical reasoning on the identified intent to generate a decision scheme.

[0010] In some embodiments of the present invention, the data fusion module includes: Preprocessing unit: used to regrind marine data to unify the spatiotemporal reference and resolution of the original marine data; Univariate fusion unit: Used to extract the optimal part of each observation data of the same ocean variable collected by different sensors using a multi-objective decision-making method, and generate the fusion variable corresponding to the ocean variable; The consistency check unit is used to perform physical consistency verification and correction on the fusion results of different ocean variables; The deep learning network unit uses ocean data processed by the preprocessing unit, univariate fusion unit, and consistency check unit as input features, and uses pre-stored high-precision observation data as training targets to correct and fill in the input features; The dataset construction unit constructs an ocean dataset based on the output of the deep learning network unit.

[0011] In some embodiments of the present invention, the remeshing process is implemented using an interpolation method, which includes at least one of inverse distance weighted interpolation, kriging interpolation, or nearest neighbor interpolation.

[0012] In some embodiments of the present invention, the deep learning network unit adopts a neural network structure comprising an input layer, at least two hidden layers, and an output layer; The hidden layer uses the ReLU activation function, and a Dropout layer is set after the hidden layer to prevent overfitting.

[0013] In some embodiments of the present invention, the data fusion module includes: The feature selection unit performs coordinate unification and missing value removal on the marine data processed by the preprocessing unit, univariate fusion unit, and consistency check unit, and inputs the processed marine data into the deep learning network unit as input features.

[0014] In some embodiments of the present invention, the large language model is configured as follows: Parse the natural language input from the user to extract key information and decision-making scenarios; Based on key information, the corresponding target functional module is matched and called in the application module layer to generate the corresponding intermediate data; Retrieve marine data relevant to decision-making scenarios from marine datasets; Based on intermediate data and retrieved ocean data, at least one decision-making scheme is generated through logical reasoning and presented in the form of natural language and / or visual charts.

[0015] In some embodiments of the present invention, the large language model adopts a BERT-based text classification model, and its knowledge retrieval mechanism is a hybrid retrieval strategy that combines vector retrieval and keyword retrieval.

[0016] In some embodiments of the present invention, the application module layer includes: The first type of application module is used to provide marine natural environment data for the data fusion layer; The second type of application module is used to provide marine human environment data for the data fusion layer; The third type of application module is used to generate intermediate data with decision-making reference value based on ocean datasets.

[0017] In some embodiments of the present invention, the third type of application module includes: The marine navigation safety module is used to provide navigation risk forecasting and route planning information based on marine big data. The nearshore ecological health module is used to provide information on ecological early warning, pollution source tracing, and aquaculture environment assessment based on marine big data. The Blue Economy Support Module is used to provide resource potential assessment and environmental impact information based on marine big data datasets. The Ocean Literacy module is used to provide ocean knowledge information based on ocean big data datasets.

[0018] In some embodiments of the present invention, marine data includes, but is not limited to, on-site drifting buoy data, shipborne and airborne data, and satellite remote sensing data.

[0019] Some embodiments of the present invention further provide an AI-based ocean solution generation method, characterized in that, for the aforementioned AI-based ocean solution toolkit system, the method includes the following steps: S1: Acquire ocean data from different sources; S2: Input ocean data into a pre-trained data fusion model to generate an ocean dataset; S3: Receive the user's decision requirements in natural language and use a pre-trained large language model to perform intent recognition; S4: Based on the results of intent recognition, determine the target functional module from multiple preset functional modules for different marine application scenarios and generate intermediate data; S5: Based on ocean datasets and intermediate data, perform logical reasoning to generate decision-making solutions.

[0020] The beneficial effects of this invention are as follows: 1. The data fusion model constructed in this invention achieves deep fusion of marine data from various sources and in various formats through multi-stage processing such as regrinding, univariate fusion, physical consistency verification, and deep learning correction. The resulting marine dataset has both broad coverage and close accuracy to actual measurements in key parameter variables, providing comprehensive and reliable data support for problem decision-making. 2. The decision-making system provided by this invention calls multiple functional modules for different marine scenarios, providing marine data in multiple fields such as marine navigation safety, nearshore ecological health, blue economy support, and marine literacy for the decision-making of large language models, which is conducive to the generation of decision-making schemes that take into account multiple dimensions by large language models. 3. The large language model provided by this invention supports users to directly input decision-making needs in natural language. The model can automatically complete intent recognition, module calling and data retrieval, which greatly reduces the threshold for using the decision-making system and adapts to the usage needs of different users in multiple fields.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A schematic diagram of the structure of the AI-based marine solution toolkit system provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the data fusion model provided in the embodiments of this application; Figure 3A schematic diagram of the decision-making process of the Big Prophecy model provided in the embodiments of this application; Figure 4 A schematic diagram of the decision-making process of a big oracle model provided in a specific embodiment of this application; Figure 5 A schematic diagram of the functional modules provided in the embodiments of this application; Figure 6 A flowchart illustrating the AI-based marine solution generation method provided in this application embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] The technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0028] As attached Figure 1 -Appendix Figure 5 As shown in the illustration, in one illustrative embodiment of an AI-based marine solution toolkit system of the present invention, the decision support system includes a data fusion layer, an application module layer, and a decision support layer.

[0029] The data fusion layer includes a pre-trained AI-based multi-source heterogeneous ocean data fusion model. This model has been trained and optimized using massive amounts of ocean data. The data fusion layer is configured to receive multi-source heterogeneous ocean data from different data sources and call a pre-trained data fusion model to perform fusion calculations on the ocean data, generating an ocean dataset for the target sea area with wide spatial coverage, high temporal resolution, and high data accuracy.

[0030] The application module layer contains multiple functional modules for different marine application scenarios. These modules provide raw marine data to the data fusion layer and generate intermediate data with decision-making reference value based on the marine dataset, such as navigation risk indices, pollution source tracing paths, and resource potential distribution.

[0031] The decision support layer contains a pre-trained AI large language model.

[0032] The large language model is configured to receive decision-making requests from users in natural language. Examples include, "Is the Arctic shipping route safe at a certain time?" or "Where is it most advantageous to build offshore wind power?"

[0033] Identify the intent behind the decision-making needs and extract key information, such as time, location, and problem type.

[0034] Based on the recognition results, one or more corresponding functional modules in the application module layer are called to obtain intermediate data.

[0035] Based on intermediate data and ocean datasets, key information is extracted and comprehensively analyzed to generate decision-making solutions.

[0036] In this embodiment, a three-layer architecture design achieves a complete closed loop from data fusion and domain application to intelligent decision-making. The data fusion layer solves the problem of integrating multi-source heterogeneous marine data, the application module layer provides professional domain knowledge support, and the decision support layer enables natural language interaction, significantly reducing the user's learning curve and improving the scientific rigor and efficiency of decision-making.

[0037] In some embodiments of the present invention, multi-source heterogeneous ocean data includes, but is not limited to, field observation data, satellite remote sensing data, numerical simulation data, and AI computation data.

[0038] The data from on-site observations of drifting buoys is in .txt format; shipborne and airborne data is in .dat format, and the data is in point data format. Satellite remote sensing data is in .tif format, and the data is in orbital data format, or in .nc format, and the data is in preprocessed two-dimensional planar data format. Data obtained from numerical simulations and AI calculations is in .nc format, and the data is in two-dimensional planar data format.

[0039] Taking sea ice thickness SIT data as an example, the correspondence between various types of ocean data and data formats, travel, intervals, and sources is shown in Table 1.

[0040] Table 1. Correspondence Table of Ocean Data

[0041] Specifically, the on-site observation data of drifting buoys refers to the SIMBA buoy data from the MOSAiC program. This data covers 17 buoys from 2019 to 2020, recording their daily positions and corresponding ice and snow thicknesses. Similar to OIB, SIMBA offers high observational accuracy but has limited coverage, only representing the conditions along the buoy's drift path. In this embodiment, SIMBA is used to evaluate the model's generalization ability in independent periods.

[0042] The shipborne and airborne data specifically refers to Operation IceBridge (OIB) data. The measured data for Operation IceBridge comes from the NASA Operation IceBridge mission's IceBridge Sea IceFreeboard, Snow Depth, and Thickness Quick Look dataset. This dataset acquires information on sea ice surface and snow cover in the Arctic region using airborne laser altimeters (ATMs), snow radar, and digital imaging systems (DMS), from which sea ice thickness and snow thickness are calculated. The data covers the period from 2012 to spring 2019, with WGS 84 as the spatial reference frame. The spatial resolution is distributed along the flight path, with a measurement accuracy of approximately ±0.1 meters. Compared to satellite or model data, OIB offers higher accuracy at specific points, but its spatial and temporal coverage is limited, distributed only along the flight path. This study uses data from 2011 to 2015 as a high-precision reference for the validation and fusion of sea ice satellite and model products.

[0043] The satellite remote sensing data (CS2SMOS) specifically utilizes the CS2SMOS fused sea ice thickness product released by the European Space Agency. The high temporal resolution daily SMOS SIT in the sea ice thickness SIT comes from the SMOS satellite, with a short data update cycle but limited coverage. The better spatial coverage Weekly CS2SMOS SIT is a fused product of the CS2 and SMOS satellites, offering a wider spatial coverage. These two products are used to generate daily SIT data for the study area. This product combines observations from the CryoSat-2 radar altimeter and the SMOS passive microwave radiometer, providing sea ice thickness fields for the Northern Hemisphere at a spatial resolution of 25 km. The temporal resolution is weekly composite, covering the winter half-year (approximately October to April of the following year), and it has been continuously updated since November 2010. Compared to single-satellite products, CS2SMOS offers more complete spatial coverage and generally exhibits relatively better performance, but its thickness estimation has some bias: it may be lower than measured values ​​in areas with thick ice.

[0044] The numerical model simulation data used was the OCEANUS product, a forecasting system independently developed by my country's Ministry of Natural Resources. It can simulate global ocean processes over the past few decades and forecast ocean conditions for the next few days to months, with a spatial resolution of 0.1°. This system couples the evolution of sea ice and ocean based on the physical processes of the ocean and sea ice, providing long-term Arctic sea ice thickness data covering the entire Arctic region north of 60°N. The data is in nc file format and updated daily. While the model product offers advantages in temporal continuity and physical consistency, its simulation results depend on the model assimilation and parameterization schemes, and may exhibit systematic biases at the regional scale.

[0045] In this embodiment, the ocean parameter variables used for data fusion are shown in Table 2.

[0046] Table 2. Oceanographic Parameter Variables

[0047] In some embodiments of the present invention, the data fusion module includes a preprocessing unit, a univariate fusion unit, a consistency check unit, and a deep learning network unit.

[0048] The preprocessing unit is used to regrind ocean data from different sources, in different formats, and at different resolutions, unifying the spatiotemporal reference and resolution of the original ocean data. Specifically, interpolation is used to map various types of data to the same grid system, eliminating scale differences between data.

[0049] For example, sea ice concentration (SIC) or sea ice thickness (SIT) products from different sources differ in spatial and temporal coverage and resolution, and there is uncertainty in the temporal sampling of weekly CS2SMOS SIT data. Therefore, it is necessary to upgrade or downscale these products through regrinding to remap them to the required coverage and resolution.

[0050] The univariate fusion unit is used to extract the optimal part of each observation data of the same ocean variable collected by different sensors using a multi-objective decision-making method, and generate the fusion variable corresponding to that ocean variable.

[0051] For example, the NIS ice map performs better in accurately depicting low sea ice concentration (SIC) regions, while the AMSR2 SIC is more continuous and accurate in depicting SIC concentrations far from low SIC regions. Similarly, the weekly CS2SMOSSIT has lower uncertainty in thick sea ice, while the SMOS SIT has lower uncertainty in thin sea ice. By merging the SMOS SIT and the weekly CS2SMOS SIT, a refined daily SIT observation of sea ice thickness can be provided that is applicable to both thin and thick sea ice.

[0052] The consistency check unit is used to perform physical consistency verification and correction on the fusion results of different ocean variables. For example, such as Figure 2 As shown, since weather filters typically underestimate or remove sea ice near the sea ice edge, the SIC can be improved by using sea ice maps, and the uncertainty of SIT can be estimated by utilizing the empirical relationship between the SIC and SIV of thin sea ice. SIT is retained when in-grid SIC exists; SIT is removed when in-grid SIC does not exist.

[0053] The deep learning network unit uses ocean data processed by the preprocessing unit, univariate fusion unit, and consistency check unit as input features (such as longitude, latitude, CS2SMOS thickness, PIOMAS thickness, etc.) and pre-stored high-precision observation data (such as IceBridge airborne data, MOSAiC buoy data, etc.) as training targets to correct and fill in the input features. The dataset construction unit builds a comprehensive and accurate ocean dataset based on the output of deep learning network units.

[0054] In this embodiment, the complementary advantages of multi-source data are fully utilized through the collaborative work of multiple processing units. Specifically, the preprocessing unit eliminates data heterogeneity, the univariate fusion unit extracts optimal information, the consistency check unit ensures the physical consistency of the data, and the deep learning unit improves accuracy. The resulting ocean dataset not only has broad coverage of satellite / model data but also approximates the measured accuracy on key variables.

[0055] In some embodiments of the present invention, the remeshing process is implemented using spatial interpolation.

[0056] Spatial interpolation can be divided into two algorithms: spatial interpolation and spatial extrapolation. Spatial interpolation algorithms extrapolate data for unknown points in the same region from data for known points, while spatial extrapolation algorithms extrapolate data for other regions from data for known regions. These algorithms include inverse distance weighted interpolation, kriging interpolation, and nearest neighbor interpolation.

[0057] The main steps of inverse distance weighted interpolation include: constructing sample data, creating a grid for interpolation, and plotting the interpolation results.

[0058] In some embodiments of the present invention, the deep learning network unit adopts a neural network structure comprising an input layer, at least two hidden layers, and an output layer.

[0059] The input layer contains four features, two of which are longitude and latitude, and the other two can be selected from satellite remote sensing CS2SMOS thickness, numerical simulation PIOMAS thickness, and AI simulation calculation SIT.

[0060] The first hidden layer has 256 neurons, and the second hidden layer has 128 neurons. Each layer uses the ReLU activation function to enhance the network's ability to learn non-linear relationships. A Dropout layer with a dropout rate of approximately 0.092 is added after the hidden layers to prevent overfitting.

[0061] The output layer is a single neuron, which can correspond to one of the following: MoSAiC buoy observation, ocean observation, or aerial observation of IceBridge sea ice thickness, and can be configured differently according to different scenarios.

[0062] like Figure 2 As shown, in a specific embodiment, the input layer includes four features: longitude, latitude, satellite remote sensing CS2SMOS thickness, and numerical simulation PIOMAS thickness, and the output layer outputs the IceBridge sea ice thickness observed from air.

[0063] The aforementioned deep learning network has a simple structure, effectively improving the processing speed of multi-parameter data and accelerating the system's response efficiency. This network structure design saves computational costs while being sufficient to express the nonlinear relationship between the input and the target.

[0064] Taking SIT sea ice thickness data as an example, compared to PIOMAS data, from October 2019 to January 2020, the merged sea ice thickness data was generally higher in the Arctic region, with high-value differences concentrated in the waters adjacent to Greenland and the Canadian Arctic Archipelago. From February to April 2020, the merged data was lower in the Beaufort Sea and East Siberian Sea, and higher in the Kara Sea, Barents Sea, Greenland Sea, and Baffin Bay. Compared to CS2SMOS data, from October 2019 to April 2020, the merged data was higher in the Arctic marginal seas, particularly showing significant differences in the Kara Sea, Barents Sea, and Baffin Bay.

[0065] The mean squared error (MSE) is used as the loss function during the training of the data fusion model, and the Adam optimizer is employed. The dataset is randomly divided into training and validation sets, with 20% of the samples used as the validation set to monitor model performance during training. An early stopping strategy is introduced during training, terminating training prematurely when the validation set error fails to improve for several consecutive rounds. After each training round, the RMSE, MAE, and R² of the validation set are calculated to monitor model convergence, and the model weights with the minimum validation error are saved.

[0066] The mean absolute error (MAE) on the final validation set was 0.4518 m, and the root mean square error (RMSE) was 0.6053 m, indicating that the model can fit the observed data well.

[0067] In some embodiments of the present invention, the data fusion module includes a feature selection unit.

[0068] The feature selection unit is used to unify coordinates and remove missing values ​​from the ocean data processed by the preprocessing unit, univariate fusion unit and consistency check unit, and then inputs the processed ocean data into the deep learning network unit as input features.

[0069] Taking sea ice thickness SIT fusion as an example, the feature selection unit starts from the 4686 cleaned IceBridge samples, matches the thickness of CS2SMOS satellite products and PIOMAS model products for the corresponding date and location, unifies the longitude to the range [0, 360], removes missing values, and standardizes the input features. Simultaneously, the target variable IceBridge thickness is also standardized to ensure the stability of network training. It should be understood that all standardized parameters from all training stages are saved to ensure input consistency when predicting independent data subsequently.

[0070] The embodiments of this application ensure the stability and convergence speed of model training through standardized feature processing, and save the standardized parameters to ensure the consistency between training data and prediction data, so that the model has good generalization ability.

[0071] In some embodiments of the present invention, such as Figure 6 As shown, the application module layer includes a first type of application module, a second type of application module, and a third type of application module.

[0072] The first type of application module is used to provide marine natural environment data for the data fusion layer. In this embodiment, it is specifically the Ocean God Early Warning Module (Module 1), which is used to output marine physical parameters such as temperature, salinity, current velocity, wind speed, sea ice concentration, and sea ice thickness.

[0073] The second type of application module is used to provide marine human environment data for the data fusion layer. In this embodiment, it is specifically the coastal zone integrated management module (module 4), which is used to output marine social science parameters such as coastal population density, location of nature reserves, marine functional zoning, and management policies.

[0074] The third type of application module is used to generate intermediate data with decision-making reference value based on ocean datasets.

[0075] The third category of application modules specifically includes the marine navigation safety module (module 2), the nearshore ecological health module (module 3), the blue economy support module (module 5), and the marine literacy module (module 6).

[0076] Among them, the marine navigation safety module (module 2) is used to provide information such as the Arctic ice zone navigation risk index, optimal navigation route planning, and navigation risk heat map based on the marine big data set.

[0077] The nearshore ecological health module (Module 3) provides functions such as coral reef bleaching early warning, pollution source tracing and drift prediction, marine aquaculture environment assessment, and marine protected area selection based on marine big data.

[0078] The Blue Economy Support Module (Module 5) is used to provide products such as marine renewable energy potential distribution, blue carbon storage estimation, marine environmental impact assessment, and marine tourism index based on marine big data datasets.

[0079] The marine literacy module (Module 6) is used to provide functions such as the marine literacy alliance for primary and secondary schools, the marine museum alliance, marine knowledge dissemination, and best practices for harmony between humans and the sea based on the marine big data.

[0080] In this embodiment, a modular classification design achieves the separation and collaboration of data contribution and application. The first and second types of application modules provide basic parameters as data sources. After being fused by a data fusion model to form a high-quality dataset, this dataset is then used by the third type of application module to generate specialized intermediate data. This design avoids redundant construction, ensures the consistency of data sources across modules, and facilitates system expansion and maintenance.

[0081] In some embodiments of the present invention, the large language model is configured to perform the following steps.

[0082] Parse the natural language input from the user to extract key information and decision-making scenarios.

[0083] Based on key information, the corresponding target functional module is matched and invoked in the application module layer to generate the corresponding intermediate data.

[0084] Retrieve marine data relevant to decision-making scenarios from marine datasets.

[0085] Based on intermediate data and retrieved ocean data, at least one decision-making scheme is generated through logical reasoning and presented in the form of natural language and / or visual charts.

[0086] For example, such as Figure 3 As shown, the core process of user decision-making driven by large language models for ocean problem solutions includes the following steps: User input decision-making needs steps: Users ask questions or describe decision-making scenarios using natural language, such as "Help me predict the trajectory of a person who has fallen into the water for search and rescue."

[0087] This may include explicit conditions (e.g., yesterday, Qingdao sea area) or vague requirements (e.g., the most likely route).

[0088] Understanding and intent recognition steps include semantic parsing and contextual supplementation. Semantic parsing involves extracting keywords and classifying question types. Contextual supplementation specifically involves clarifying questions based on user history data or follow-up inquiries, such as "What was the specific time? What was the specific location?".

[0089] Information retrieval and knowledge retrieval steps include: calling pre-trained internal databases, such as ocean currents and wind fields, from an internal knowledge base; and acquiring external real-time data via plugins, such as whether a suspected drowning person has been found in a certain sea area.

[0090] Analysis and solution generation steps: Conduct multi-dimensional evaluation, generate the optimal solution, and generate alternative solutions.

[0091] The reasoning process is logical. If the ocean current is considered to play a major role, the optimal solution is adopted; if there are special circumstances and the wind is considered to play a greater role, the alternative solution is favored.

[0092] Use natural language to summarize, for example, "The search and rescue area is recommended to be located at [location]", or visualize the results, for example, draw a prediction map of the dynamic drift trajectory to present the decision-making options.

[0093] User feedback and iteration steps: Users revise / refine their requirements, and the model dynamically adjusts the output decision scheme.

[0094] Taking the problem of searching for and rescuing a person who has fallen into the water as an example, the system's workflow is as follows: Figure 4 As shown.

[0095] As attached Figure 6 As shown, some embodiments of the present invention further provide an AI-based marine solution generation method for controlling the aforementioned AI-based marine solution toolkit system, the method comprising the following steps.

[0096] S1: Acquire ocean data from different sources and input the ocean data into a pre-trained data fusion model to generate an ocean dataset.

[0097] S2: Receives the user's decision requirements in natural language and uses a pre-trained large language model to perform intent recognition.

[0098] S3: Based on the results of intent recognition, the target functional module is determined from multiple preset functional modules for different marine application scenarios, and intermediate data is generated.

[0099] S4: Based on ocean datasets and intermediate data, perform logical reasoning to generate decision-making solutions.

[0100] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. An AI-based marine solutions toolkit system, characterized in that, include: The data fusion layer contains a pre-trained data fusion model, which is used to receive ocean data from different data sources and call the data fusion model to perform fusion calculations on the ocean data to generate an ocean dataset covering the target sea area. The application module layer contains multiple functional modules for different marine application scenarios; these functional modules are used to provide marine data to the data fusion layer and generate intermediate data with decision-making reference value based on the marine dataset. The decision support layer includes a pre-trained large language model configured to receive decision requirements input by the user in natural language. The decision-making requirements are identified by intent, and one or more functional modules in the application module layer are called based on the identification results to obtain intermediate data; based on the intermediate data and the ocean dataset, logical reasoning is performed on the identified intent to generate a decision scheme.

2. The AI-based marine solutions toolkit system according to claim 1, characterized in that, The data fusion module includes: Preprocessing unit: used to regrind the ocean data to unify the spatiotemporal reference and resolution of the ocean data; Univariate fusion unit: Used to extract the optimal part of each observation data of the same ocean variable collected by different sensors using a multi-objective decision-making method, and generate the fusion variable corresponding to the ocean variable; The consistency check unit is used to perform physical consistency verification and correction on the fusion results of different ocean variables; The deep learning network unit uses the ocean data processed by the preprocessing unit, the univariate fusion unit, and the consistency check unit as input features, and uses pre-stored high-precision observation data as training targets to correct and fill in the input features; The dataset construction unit constructs an ocean dataset based on the output of the deep learning network unit.

3. The AI-based marine solutions toolkit system according to claim 2, characterized in that, The remeshing process is implemented using an interpolation method, which includes at least one of inverse distance weighted interpolation, Kriging interpolation, or nearest neighbor interpolation.

4. The AI-based marine solutions toolkit system according to claim 2, characterized in that, The deep learning network unit adopts a neural network structure that includes an input layer, at least two hidden layers, and an output layer; The hidden layer uses the ReLU activation function, and a Dropout layer is set after the hidden layer to prevent overfitting.

5. The AI-based marine solutions toolkit system according to claim 2, characterized in that, The data fusion module includes: The feature selection unit performs coordinate unification and missing value removal on the ocean data processed by the preprocessing unit, the univariate fusion unit, and the consistency check unit, and inputs the processed ocean data into the deep learning network unit as input features.

6. The AI-based marine solutions toolkit system according to claim 1, characterized in that, The large language model is configured as follows: Parse the natural language input from the user to extract key information and decision-making scenarios; Based on the key information, the corresponding target function module is matched and called in the application module layer to generate the corresponding intermediate data. Retrieve marine data relevant to the decision-making scenario from the marine dataset; Based on the intermediate data and the retrieved ocean data, at least one decision-making scheme is generated through logical reasoning and presented in the form of natural language and / or visual charts.

7. The AI-based marine solutions toolkit system according to claim 1 or 6, characterized in that, The large language model adopts a BERT-based text classification model, and its knowledge retrieval mechanism is a hybrid retrieval strategy that combines vector retrieval and keyword retrieval.

8. The AI-based marine solutions toolkit system according to claim 1, characterized in that, The application module layer includes: The first type of application module is used to provide marine natural environment data for the data fusion layer; The second type of application module is used to provide marine human environment data for the data fusion layer; The third type of application module is used to generate intermediate data with decision-making reference value based on the ocean dataset.

9. The AI-based marine solutions toolkit system according to claim 1, characterized in that, The marine data includes, but is not limited to, on-site drifting buoy data, shipborne and airborne data, and satellite remote sensing data.

10. A method for generating ocean solutions based on AI, characterized in that, The system for controlling the AI-based marine solutions toolkit system according to any one of claims 1-9 includes the following steps: S1: Acquire ocean data from different sources; S2: Input the ocean data into a pre-trained data fusion model to generate an ocean dataset; S3: Receive the user's decision requirements in natural language and use a pre-trained large language model to perform intent recognition; S4: Based on the results of intent recognition, determine the target functional module from multiple preset functional modules for different marine application scenarios and generate intermediate data; S5: Based on the ocean dataset and the intermediate data, perform logical reasoning to generate a decision-making scheme.

Images