Construction Method of Marine Ecosystem Multi-Agents and Its Interaction System

Abstract

This application relates to the field of marine ecological technologies, particularly to a construction method of marine ecological multi-agents and its interaction system. Among them, the construction method includes: a step of constructing a marine ecological large model, constructing a fine-tuning data set, and training a marine ecological large model based on the fine-tuning data set; a step of constructing a parallel application module, constructing a domain knowledge base, marine ecological multi-agents, and building a multi-modal large model, deploying the domain knowledge base, marine ecological multi-agents, multi-modal large model of the parallel application module and the marine ecological large model on a GPU server, where the marine ecological large model is used to parse user requests and distribute them to the parallel application module, triggering the parallel application module to generate response results and feedback them to the marine ecological large model for semantic integration and then output. Through this application, the intelligent level in the field of marine ecology is improved, the prediction ability of the model is enhanced, and the prediction results are made more accurate and reliable.

Description

Technical Field

[0001] This application relates to the field of marine ecological technologies, and particularly to a construction method for marine ecological multi - agents and their interaction system. Background Art

[0002] As a global issue, marine ecological protection has attracted increasing attention, especially in the research and application combined with AI technology. However, the existing "AI + marine ecology" technologies still have many deficiencies and fail to fully solve the key problems in the field.

[0003] In the "AI + marine ecology" field, traditional machine learning methods such as support vector machine (SVM), random forest (RF), and artificial neural network (ANN) have been applied to the research of species distribution prediction, helping scientists predict the potential distribution of species, evaluate species richness, and explore the relationship between species habitat selection and landscape pattern changes. At the same time, image recognition technology has also been applied in aspects such as species recognition and quantity statistics, and marine pollutant monitoring. These AI technologies provide decision - making support for marine ecological resource management, ecosystem prediction, and marine environmental protection, showing increasing application potential.

[0004] However, despite some progress made by AI technology in the field of marine ecological protection, its application and development still face a series of problems and challenges:

[0005] For example, the data resources in the marine ecological field are limited, and the phenomenon of data islands is serious. The effective training of AI models often depends on large - scale and high - quality data. However, since the acquisition of marine ecological data requires a large amount of field investigation work, which is not only time - consuming but also costly, the data scale is difficult to meet the training requirements of AI models, thus affecting the prediction accuracy and generalization ability of the models.

[0006] Another example is that the existing "AI + marine ecology" models mostly adopt methods such as traditional machine learning, convolutional neural network (CNN), and long short - term memory network (LSTM). The technology is relatively lagging and cannot cope with the complex and changeable marine ecological environment. Especially when dealing with interference factors such as typhoons, storm surges, and human activities, the accuracy and stability of the models perform poorly, and the business - oriented database and Internet data are not fully utilized.

[0007] In addition, currently, the AI models in sub - fields such as marine biological distribution prediction, species recognition, population analysis, and pollutant monitoring are often implemented by multiple independent small models respectively. There is a lack of data sharing and coupling between these models, and a unified analysis framework cannot be formed, resulting in the inability of the "AI + marine ecology" application to develop systematically, which greatly limits the progress of scientific research and industrial application. Summary of the Invention

[0008] An embodiment of the present application provides a method for constructing a marine ecological multi-agent and its interaction system, aiming to solve key problems such as data scarcity, poor model accuracy and stability, and application fragmentation, and provide more efficient and accurate technical support for marine ecological protection.

[0009] In a first aspect, an embodiment of the present application provides a method for constructing a marine ecological multi-agent, including:

[0010] Steps for constructing a marine ecological large model: constructing a fine-tuning data set, and training a selected large language model based on the fine-tuning data set to obtain a marine ecological large model;

[0011] Steps for constructing a parallel application module: constructing a domain knowledge base, a marine ecological multi-agent, and building a multi-modal large model, and deploying the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model of the parallel application module and the marine ecological large model in a diversified parallel deployment manner on a GPU server,

[0012] Among them, the knowledge data set of the domain knowledge base includes: marine environmental data, human activity pressure data, and the human activity pressure data includes: socio-economic data, human activity data. The marine ecological multi-agent includes: a pre-encapsulated basic function agent, an Internet function agent, and a marine ecological function agent. Among them, the basic function agent is used to realize information retrieval of a business database, mathematical calculation, server basic settings, operation and maintenance, etc. The Internet function agent is used to realize functions such as language translation, weather query, search engine, map service, paper retrieval, Wikipedia, etc. The marine ecological function agent is used to realize functions in the fields of fishery resources and marine ecological protection, and can support functions such as quantity statistics, species identification, habitat simulation, species distribution prediction, pollutant monitoring, extreme weather prediction, and ecological disaster warning; the multi-modal large model includes a multi-modal model, a text-to-image model, a text-to-speech model, and a text-to-video model, and is used to support the multi-modal large model to realize the mutual conversion and generation of visual question answering, text-to-picture, speech, and video;

[0013] The marine ecological large model is used to parse a user request and distribute it to the parallel application module, trigger any one or any combination of the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model in the parallel application module to generate a response result, and feedback it to the marine ecological large model for semantic integration and then output.

[0014] Based on the above steps, in the embodiments of the present application, in order to solve the problems of lagging existing model technology, poor model accuracy and stability, etc., a large marine ecological model is constructed and a large model parallel application architecture is adopted to improve the intelligent level in the field of marine ecology. The large marine ecological model is used to capture subtle changes and long-term trends that are difficult for traditional models to capture, thereby improving the prediction accuracy of the model. The parallel application architecture can support larger-scale data processing and more complex model training, further enhancing the prediction ability of the model, making the prediction results more accurate and reliable. The parallel application architecture can make full use of the computing resources of multiple computing nodes to achieve parallel training and inference of the model, thereby significantly shortening the time for model training and inference and improving the response speed of the model.

[0015] In some of the embodiments, the construction process of the fine-tuning data set includes:

[0016] A data acquisition step of acquiring literature in the field of marine ecology and parsing and extracting the text content in the literature in the field of marine ecology;

[0017] A literature data processing step of splitting the text content into segment contents and correspondingly extracting the summary of each segment content;

[0018] A fine-tuning data set construction step of storing the segment content and its summary in Json format to obtain a fine-tuning data set.

[0019] In some of the embodiments, the construction process of the domain knowledge base includes:

[0020] A knowledge data set acquisition step of acquiring marine environment data and human activity pressure data through a document loader;

[0021] A data vectorization step of using a tokenizer to convert the marine environment data and human activity pressure data into text chunks and performing vectorization processing based on an Embedding model to obtain text vectors;

[0022] Wherein, when the domain knowledge base receives a distributed user request, the user request is vectorized and similarity matching is performed on the text vectors in the domain knowledge base. After using a Rerank model to sort the matching results to obtain text vectors with high similarity, the text is content-integrated using prompt engineering and then sent to the large marine ecological model.

[0023] In some of the embodiments, in the large marine ecological model construction step, the LORA fine-tuning method is adopted for training based on the Llama_factory framework.

[0024] In some of these embodiments, the workflow of the parallel application module is enabled or disabled through a start node and an end node. The main models of the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model are set as the marine ecological large model through an LLM node, and a switch node is configured for each of the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model to control whether the workflows of the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model are enabled respectively.

[0025] In some of these embodiments, the marine ecological multi-agent is further configured with a trigger, which is configured for keyword triggering. When the switch node of the marine ecological multi-agent is enabled, the corresponding agents in the marine ecological multi-agent are enabled through the trigger.

[0026] In some of these embodiments, the basic function agents include: a business-oriented database processing agent, a mathematical calculation agent, and a server basic setting agent, where:

[0027] The business-oriented database processing agent is obtained by encapsulating the API interface of the marine ecological business-oriented database. The marine ecological business-oriented database includes: a marine ranch monitoring database, a pollutant monitoring database, and a global biodiversity database;

[0028] The mathematical calculation skill question is obtained by encapsulating a calculator as an agent;

[0029] The server basic setting agent is obtained by encapsulating the server system query command as an agent.

[0030] In some of these embodiments, the Internet function agents include a language translation agent, a weather query agent, a search engine agent, a map service agent, a paper retrieval agent, and a knowledge encyclopedia agent, where,

[0031] The language translation agent is obtained by encapsulating the API interface of a translation platform as an agent;

[0032] The weather query agent is obtained by encapsulating the API interface of a weather prediction platform as an agent;

[0033] The search engine agent is obtained by encapsulating the API of a search engine as an agent;

[0034] The map service agent is obtained by encapsulating the POI service API of a map platform as an agent;

[0035] The paper retrieval agent is obtained by encapsulating the API of a library database as an agent;

[0036] The knowledge encyclopedia agent is obtained by encapsulating the Wikipedia database API interface as an agent.

[0037] In some of the embodiments, the marine ecological function agent includes: a quantity statistics agent, a species identification agent, a habitat simulation agent, and a species distribution prediction agent, which are used to achieve the purpose of fishery resource protection; among them,

[0038] The quantity statistics agent is obtained by encapsulating a fish fine-grained recognition model as an agent;

[0039] The species identification agent is obtained by encapsulating a marine species identification model as an agent;

[0040] The habitat simulation agent is obtained by encapsulating a marine habitat prediction and simulation analysis model as an agent;

[0041] The species distribution prediction agent is obtained by encapsulating a species distribution prediction model as an agent.

[0042] In some of the embodiments, the marine ecological function agent further includes: a pollutant monitoring agent, an extreme weather prediction agent, and an ecological disaster warning agent, which are used to achieve the purpose of pollutant monitoring; among them,

[0043] The pollutant monitoring agent is obtained by acquiring the seawater quality monitoring and seawater bath water quality information of the marine environmental monitoring center and encapsulating them as an agent;

[0044] The extreme weather prediction agent is obtained by encapsulating an extreme weather prediction model as an agent;

[0045] The ecological disaster warning agent is obtained by encapsulating the ecological disaster monitoring and warning system interface as an agent.

[0046] In a second aspect, an embodiment of the present application provides a marine ecological multi-agent interaction system. Based on the construction method of the marine ecological multi-agent described in the first aspect above, it includes:

[0047] A data layer, which is used to generate fine-tuning data, a domain knowledge base, and a business database based on data sources. The data sources include the fine-tuning data set, the knowledge data set, and the data sources of the marine ecological multi-agent;

[0048] The model layer is used to build the sub - agents required for the marine ecological large model, the domain knowledge base, the multi - modal large model, and the encapsulation of the basic function agent, the Internet function agent, and the marine ecological function agent. The sub - agents include: the business database processing agent, the mathematical calculation agent, the server basic setting agent, the language translation agent, the weather query agent, the search engine agent, the map service agent, the paper retrieval agent, the knowledge encyclopedia agent, the quantity statistics agent, the species identification agent, the habitat simulation agent, and the species distribution prediction agent;

[0049] The agent layer is used to build the basic function agent, the Internet function agent, and the marine ecological function agent by integrating the sub - agents on the basis of the model layer;

[0050] The application layer includes at least the fishery resource protection application module and the ecological protection application module. The fishery resource protection application module includes: the marine organism quantity statistics unit, the species identification unit, the habitat simulation unit, and the species distribution prediction unit. The ecological protection application module includes: the pollutant monitoring unit, the extreme weather prediction unit, and the ecological disaster warning unit;

[0051] The interaction layer provides the management of model configuration, multi - modal dialogue, domain knowledge dialogue, and knowledge base through a front - end interaction interface.

[0052] In some embodiments, the front - end interaction interface at least includes:

[0053] The model configuration unit is used to select the marine ecological large model;

[0054] The multi - modal dialogue configuration unit is used to configure the multi - modal large model;

[0055] The domain knowledge base dialogue configuration unit is used to configure at least one knowledge base and provide question - answering services based on the configured knowledge base. The customization of new knowledge bases and the editing of existing knowledge bases can be realized through a domain knowledge base management unit;

[0056] The multi - agent configuration unit obtains the multi - agent enabling request through a selection box and provides a multi - agent list for selecting at least one agent;

[0057] The dialogue interaction unit uploads files through a text input box or a file transfer button and conducts a dialogue with the marine ecological large model, obtains the user request, sends it to the marine ecological large model, and obtains the output result after the marine ecological large model integrates the semantics.

[0058] The details of one or more embodiments of the present application are set forth in the following drawings and description to make the other features, objects, and advantages of the present application more comprehensible. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation to the present application. In the drawings:

[0060] Figure 1 is a flowchart of a method for constructing a marine ecological multi-agent according to an embodiment of the present application;

[0061] Figure 2 is a schematic diagram of the construction process of a fine-tuning data set for a method for constructing a marine ecological multi-agent according to an embodiment of the present application;

[0062] Figure 3 is a schematic diagram of the construction process of a domain knowledge base for a method for constructing a marine ecological multi-agent according to an embodiment of the present application;

[0063] Figure 4 is a schematic diagram of the parallel application principle of a marine ecological large model according to an embodiment of the present application;

[0064] Figure 5 is a schematic diagram of the construction principle of a domain knowledge base according to an embodiment of the present application;

[0065] Figure 6 is a schematic diagram of the working principle of a domain knowledge base according to an embodiment of the present application;

[0066] Figure 7 is a schematic diagram of the workflow control principle of a parallel application module according to an embodiment of the present application;

[0067] Figure 8 is a schematic diagram of the principle of a marine ecological multi-agent interaction system according to an embodiment of the present application;

[0068] Figure 9 is a schematic diagram of the interaction interface of a marine ecological multi-agent interaction system according to an embodiment of the present application.

[0069] In the figure:

[0070] 1. Data layer; 2. Model layer; 3. Agent layer; 4. Application layer; 5. Interaction layer;

[0071] 501. Model configuration unit; 502. Multimodal dialogue configuration unit; 503. Domain knowledge base dialogue configuration unit; 504. Multi-agent configuration unit; 505. Dialogue interaction unit. Detailed implementation manners

[0072] To make the objectives, technical solutions and advantages of this application more clear and understandable, the following describes and explains this application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.

[0073] Obviously, the accompanying drawings in the following description are only some examples or embodiments of this application. For those of ordinary skill in the art, without creative efforts, this application can also be applied to other similar scenarios based on these drawings. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in this application, some design, manufacturing or production changes based on the technical content disclosed in this application are only conventional technical means and should not be understood as the content disclosed in this application being insufficient.

[0074] Referring to "embodiment" in this application means that the specific features, structures or characteristics described in combination with the embodiment can be included in at least one embodiment of this application. The appearance of this phrase in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those of ordinary skill in the art explicitly and implicitly understand that the embodiments described in this application can be combined with other embodiments without conflict.

[0075] Unless otherwise defined, the technical terms or scientific terms involved in this application should have the ordinary meaning understood by those with ordinary skills in the technical field to which this application belongs. The words such as "a", "an", "one kind", "the" and the like involved in this application do not indicate a quantity limitation and can represent a singular or plural number. The terms "including", "comprising", "having" and any variations thereof involved in this application are intended to cover non-exclusive inclusion.

[0076] This embodiment provides a method for constructing a marine ecological multi-agent. Figure 1 is a flowchart of the method for constructing a marine ecological multi-agent according to the embodiment of this application, as Figure 1 shown, this process includes the following steps:

[0077] Marine ecological large model construction step S1, construct a fine-tuning data set, and train the selected large language model based on the fine-tuning data set to obtain a marine ecological large model. The large language model selected in this embodiment is Qwen2-7B to meet the requirements of factors such as performance, deployment framework compatibility and agent compatibility, specifically Qwen-vl-chat, glm-4v, etc.;

[0078] Parallel application module construction step S2: Construct a domain knowledge base, a marine ecological multi-agent, and build a multimodal large model. Deploy the domain knowledge base, marine ecological multi-agent, and multimodal large model of the parallel application module and the marine ecological large model on the GPU server in a diversified parallel deployment manner. Optionally, the GPU server includes at least 8 H800 GPUs with 64G video memory, and each H800 GPU card can run about 3 - 5 large models of different magnitudes.

[0079] Among them, the knowledge data set of the domain knowledge base includes: marine environmental data, human activity pressure data, and human activity pressure data includes: socioeconomic data, human activity data. The marine ecological multi-agent includes: pre-encapsulated basic function agents, Internet function agents, and marine ecological function agents. Among them, the basic function agents are used to implement business database information retrieval, mathematical calculations, server basic settings, operation and maintenance, etc. The Internet function agents are used to implement language translation, weather query, search engines, map services, paper retrieval, Wikipedia, etc. The marine ecological function agents are used to implement functions in the fields of fishery resources and marine ecological protection, and can support functions such as quantity statistics, species identification, habitat simulation, species distribution prediction, pollutant monitoring, extreme weather prediction, and ecological disaster warning; The multimodal large model includes a multimodal model, a text-to-image model, a text-to-speech model, and a text-to-video model, which are used to support the multimodal large model to realize the mutual conversion and generation of visual question answering, text-to-picture, speech, and video; Optionally, the supported text-to-image models are sd-turbo, sdxl-turbo, sd-medium, the supported text-to-speech models are ChatTTS, whisper-base, the supported Reranker models are Bge-reranker-base, jina-reranker-v2, the supported Embedding models are Bge-m3, bge-base-zh, gte-base, and the supported text-to-video models are Keling, Vidu. In addition, small models and other interfaces are also supported.

[0080] Reference Figure 4 As shown, the marine ecological large model is used to parse the user request and distribute it to the parallel application module, triggering any one or any combination of the domain knowledge base, marine ecological multi-agent, and multimodal large model in the parallel application module to generate a response result and feedback it to the marine ecological large model for semantic integration and then output.

[0081] Based on the above steps, in order to solve the problems of lagging existing model technology, poor model accuracy and stability, the embodiment of the present application constructs a large marine ecological model and adopts a large model parallel application architecture to improve the level of intelligence in the marine ecological field. The large marine ecological model is used to capture subtle changes and long-term trends that are difficult to capture with traditional models, thereby improving the prediction accuracy of the model. The parallel application architecture can support larger-scale data processing and more complex model training, further enhance the prediction ability of the model, and make the prediction results more accurate and reliable. The parallel application architecture can make full use of the computing resources of multiple computing nodes to realize parallel training and reasoning of the model, thereby significantly shortening the time of model training and reasoning, and improving the response speed of the model.

[0082] refer to Figure 2 As shown in Figure 2, the construction process of the fine-tuning dataset includes:

[0083] Data acquisition step S201, write a python crawler to obtain marine ecology literature in .pdf format, including Chinese papers related to marine ecology in HowNet, and parse and extract text content in marine ecology literature; optionally, use the Nougat model to extract text content in the pdf file, remove other content, and save the text content in makedown format.

[0084] The document data processing step S202 divides the text content into segments and extracts the summary of each segment accordingly. Optionally, a summary generation model under the PaddleNLP framework is used to perform batch summary extraction.

[0085] In the fine-tuning dataset construction step S203, the segment content and its summary are stored in Json format to obtain a fine-tuning dataset. Optionally, the fine-tuning dataset content is saved as three fields in the Json file, where the instruction field is the summary content, the input field is empty, and the output field is the detailed content corresponding to the summary.

[0086] refer to Figure 3 , Figure 5 As shown in the figure, the construction process of the domain knowledge base includes:

[0087] Step S301 for obtaining the knowledge dataset: Obtain marine environmental data and human activity pressure data in formats such as CSF and PDF through a document loader. Among them, marine environmental data is divided into four types: seabed topography and geomorphology, marine physics, marine chemistry, and nutrients. Seabed topography and geomorphology data includes information such as coastline distance and elevation; marine physics data includes information such as mean sea surface temperature, sea surface temperature range, sea surface temperature in the warmest month, sea surface temperature in the coldest month, mean sea surface salinity, sea surface salinity range, water flow velocity, mean bottom illumination, bottom illumination range, and wave height; marine chemistry data includes information such as chlorophyll a, primary productivity, pH, and total suspended solids; nutrient data includes information such as nitrate, phosphate, dissolved oxygen, and silicate. Specific data sources of marine environmental data are shown in Table 1 as an example, and the data format is markdown and PDF. Data in other formats can be saved as markdown format after format conversion to optimize the data structure.

[0088] Table 1 Marine Environmental Data

[0089]

[0090] Specific data sources of human activity pressure data are shown in Table 2 as an example. Human activity pressure data includes socio-economic data and human activity data. Socio-economic data includes information such as GDP, coastal population density, and fishery economy. Human activity data includes information such as aquaculture, shipping density, distance from the port, coastal land use, aquaculture pond expansion, and marine protected areas. The data format is markdown and PDF.

[0091] Table 2 Human Activity Pressure Data

[0092]

[0093] Data vectorization step S302: Use a tokenizer to convert marine environment data and human activity pressure data into text chunks, and perform vectorization processing based on the Embedding model to obtain text vectors, which are stored in a vector database, such as the Chroma vector database or the faiss vector database. The combination of the vector database with the Embedding model and the tokenizer will directly affect the accuracy of question answering in the domain knowledge base. Through practice, the Chroma vector database with the highest matching degree and the best actual effect with the Embedding model and the tokenizer is adopted; to balance the effectiveness of Chinese and English vectorization, the Bge-m3 model is used as the Embedding model; in addition, for the unique tree structure of makedown format document data, the UnstructuredMarkdownLoader loader is used to reduce the logical errors of the tree structure during document loading, and the markdown_header_metadata_splitter tokenizer is used to ensure the integrity and context logic of the text chunks after segmentation;

[0094] Among them, as shown in Figure 6 When the domain knowledge base receives the distributed user request, it vectorizes the user request and performs similarity matching on the text vectors in the domain knowledge base. After using the Rerank model to sort the matching results to obtain highly similar text vectors, the text is integrated using prompt engineering and then sent to the marine ecological large model. Optionally, the question answering parameters of the domain knowledge base are set to 5, and the top 5 highly similar text vectors before sorting are selected as candidates for the final answer, and the context is set to 1024, considering up to 1024 characters or words before and after the user's query as the context.

[0095] In another embodiment, data such as marine ecological news, laws and regulations, papers, and books can also be used in the construction of the domain knowledge base.

[0096] Based on the above steps, the embodiment of the present application constructs a fine-tuning dataset for the marine ecological large model covering various marine environment elements and human activity elements, builds and optimizes the domain knowledge base in the marine ecological field, solves the problem of lack of marine ecological data, enables the model to perform operations based on more comprehensive and accurate data, and thus improves the accuracy and reliability of the model.

[0097] In some of these embodiments, in the steps of constructing the marine ecological large model, the LORA fine-tuning method is adopted for training based on the Llama_factory framework. The fine-tuning parameter settings for the training process are shown in Table 3 below. In this embodiment, considering that a training round of 3.0 will lead to overfitting and affect the basic performance of Qwen2-7B, in order to avoid overfitting, it is necessary to screen the LORA weights for each training round. Qwen2-7B loads each LORA weight in turn and conducts a question-and-answer test to eliminate the overfitted and logically incorrect weights. Select the LORA weight with the smallest number of training rounds on the premise of having the fine-tuning function.

[0098] Table 3 Fine-tuning parameter settings

[0099]

[0100] Furthermore, the weights of Qwen2-7B are merged with the LORA weights. When merging, the maximum block size is set to 2GB, and the export quantization level is set to none. To ensure the model performance, no quantization processing is performed. The merged weights are subjected to an inference test. Using the manual test method, the accuracy of the large model's question and answer is actually verified. After the accuracy meets the standard, the construction of the marine ecological large model is completed. The inference parameter settings are shown in Table 4 below.

[0101] Table 4 Inference parameter settings

[0102]

[0103] In some of these embodiments, as shown in Figure 7 , the workflow of the parallel application module is turned on or off by configuring the start node and the end node. Through an LLM node, the domain knowledge base, the marine ecological multi-agent, and the main model of the multi-modal large model are set as the marine ecological large model, and a switch node is configured for the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model respectively, which is used to control whether the workflows of the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model are turned on. When the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model are all turned off, the marine ecological large model will directly output the result according to the user request and will no longer trigger the domain knowledge base, the marine ecological multi-agent, and the multi-modal large model.

[0104] In some of these embodiments, the marine ecological multi-agent is also configured with a trigger, and the trigger is configured for keyword triggering. When the switch node of the marine ecological multi-agent is turned on, the corresponding agent in the marine ecological multi-agent is turned on through the trigger. In another embodiment, the trigger can also be set for timed triggering or event triggering.

[0105] In some of these embodiments, the basic function agents include: a business database processing agent, a mathematical calculation agent, and a server basic setting agent, where:

[0106] The business database processing agent is obtained by encapsulating the API interface of the marine ecological business database into an agent. The marine ecological business database includes: the marine ranch monitoring database and the pollutant monitoring database based on MySQL, and the global biodiversity database based on Neo4J. The agent is started, and the corresponding agent can be triggered when the user request involves keywords such as "marine ranch", "pollutants", "monitoring" and "biodiversity";

[0107] Mathematical calculation skill questions are obtained by encapsulating a calculator into an intelligent agent. When the intelligent agent is started and an operator is input, the intelligent agent is triggered and run, giving the marine ecological large model computing capabilities, serving the marine ecological refined data analysis scenario, and making up for the shortcomings of the large model's mathematical calculation performance;

[0108] The server basic setting agent is obtained by encapsulating the server system query command into an agent, and performs operation performance detection, CPU usage query, video memory and memory usage analysis, disk usage analysis, and network performance analysis through the server system query command. When the agent is started and keywords such as "CPU usage", "video memory", "memory", and "disk usage" are entered, this agent is triggered and run.

[0109] In some embodiments, the Internet functional agent includes a language translation agent, a weather query agent, a search engine agent, a map service agent, a paper retrieval agent, and a knowledge encyclopedia agent, wherein:

[0110] The language translation agent is obtained by encapsulating the translation platform API interface into an agent. When the agent is started and keywords such as "translation" are input, the agent is triggered and run to obtain the translation result of the specified translation content, giving the marine ecological large model the ability to translate multiple languages, especially the translation effect of minority languages.

[0111] The weather query agent is obtained by encapsulating the weather forecast platform API interface into an agent. For example, if you know the weather, start the agent and enter keywords such as "city" and "weather", the agent will be triggered and run to obtain the weather conditions of the specified city, giving the marine ecological model the ability to forecast weather.

[0112] The search engine agent is obtained by encapsulating the search engine API into an agent, such as Bing search and Baidu. When the agent is started and keywords such as "Internet search" are input, the agent is triggered and run to obtain Internet search results, thus giving the marine ecological model the ability to search for Internet information.

[0113] The map service agent is obtained by encapsulating the POI service API of the map platform as an agent. For example, for AutoNavi Map, when the agent is started and keywords for location-based services such as "location", "hotel", "restaurant", etc. are input, this agent is triggered and run to obtain the corresponding location service results, endowing the marine ecological large model with the ability of location service;

[0114] The paper retrieval agent is obtained by encapsulating the library database API as an agent. For example, for the Arxiv database, when the agent is started and information such as "paper name", "keyword", etc. are input, this agent is triggered and run to obtain information such as the corresponding paper title, abstract, etc.;

[0115] The knowledge encyclopedia agent is obtained by encapsulating the Wikipedia database API interface as an agent. Here, the Wikipedia database can also be replaced by other knowledge query databases, such as the DuXiu knowledge base. When the agent is started and keywords such as "noun to be queried", "Wikipedia" are input, it is triggered and run to return the encyclopedic explanation content corresponding to the noun to be queried.

[0116] In some of these embodiments, the marine ecological function agents include: a quantity statistics agent, a species identification agent, a habitat simulation agent, and a species distribution prediction agent, which are used to achieve the purpose of fishery resource protection; among them,

[0117] The quantity statistics agent is obtained by encapsulating the fish fine-grained recognition model as an agent. When the agent is started and keywords such as "fish quantity statistics" are input, it is triggered and run to return fish quantity statistics information such as "fish species", "quantity", "spatial coordinates", etc., endowing the marine ecological large model with the ability of fish fine-grained recognition and quantity statistics, so as to serve scenarios such as marine ranches and the protection of economic fish populations;

[0118] The species identification agent is obtained by encapsulating the marine species identification model as an agent. When the agent is started and keywords such as "species name", "marine ranch name" are input, it is triggered and run to return information such as the species name, spatial coordinates, etc., endowing the marine ecological large model with the ability to identify common species in marine ranches, and serving scenarios such as marine ranches;

[0119] The habitat simulation agent is obtained by encapsulating the marine habitat prediction and simulation analysis model as an agent. When the agent is started and keywords such as "marine ranch", "seawater bathing beach" or the name of a station are input, it is triggered and run to return information such as the temperature, salinity, sea current, depth and water quality indicators of the specified marine area, endowing the marine ecological large model with the prediction and analysis ability of temperature, salinity, sea current, depth and water quality indicators (including dissolved oxygen, total phosphorus, total nitrogen, ammonia nitrogen, chemical oxygen demand, chlorophyll a) for selected marine areas such as marine ranches, seawater bathing beaches, stations, etc.;

[0120] The species distribution prediction agent is obtained by encapsulating the species distribution prediction model as an agent, such as the Maxent model. Start the agent and input keywords such as "Maxent", "species distribution prediction", "species name", "latitude and longitude", and "environmental information", trigger and run it, and return information such as the suitable habitat area of the species, endowing the marine ecological big model with the ability to predict the distribution and suitable habitat area of specific species.

[0121] In some of these embodiments, the marine ecological function agent further includes: a pollutant monitoring agent, an extreme weather prediction agent, and an ecological disaster warning agent, which are used to achieve the purpose of pollutant monitoring; among them,

[0122] The pollutant monitoring agent is obtained by encapsulating the seawater quality monitoring and seawater bath water quality information of the marine environmental monitoring center as an agent. Start the agent and input keywords such as "seawater bath name" and "water quality", trigger and run it, and return information such as the seawater bath water quality, endowing the marine ecological big model with the ability to monitor seawater quality;

[0123] The extreme weather prediction agent is obtained by encapsulating the extreme weather prediction model as an agent. Start the agent and input keywords such as "sea area", "typhoon", and "storm surge", trigger and run it, and return information such as typhoons and storm surges in the specified sea area, endowing the marine ecological big model with the ability to predict typhoons and storm surges;

[0124] The ecological disaster warning agent is obtained by encapsulating the ecological disaster monitoring and warning system interface as an agent. Start the agent and input keywords such as "sea area", "red tide", "green tide", and "Enteromorpha prolifera", trigger and run it, and return information such as red tide, green tide, or Enteromorpha prolifera in the specified sea area, endowing the marine ecological big model with the ability to predict red tide, green tide, and Enteromorpha prolifera;

[0125] The fish fine-grained recognition model, marine species recognition model, marine habitat prediction and simulation analysis model, species distribution prediction model, extreme weather prediction model, and ecological disaster monitoring and warning system in the above embodiments can be pre-trained or introduced based on existing platforms. As the functions of each model are enriched, the marine ecological big model of this application is also further expanded accordingly;

[0126] Based on the above steps, the embodiments of this application build a marine ecological multi-agent, integrating the basic function agent, the Internet function agent, and the marine ecological function agent, which is beneficial to serving multiple scenarios in the fields of fishery resource protection and ecological protection, and providing a solution for the systematic application of the marine ecology.

[0127] The domain knowledge base, marine ecological multi-agent, multi-modal big model and the marine ecological big model in the above embodiments are deployed on the GPU server in a diversified parallel deployment manner, specifically including:

[0128] Configuration of the production environment, where the production environment is divided into five modules: large model server, small model server, database, API, and application;

[0129] The large model server is used to deploy large language models, multimodal large models, etc., and uses the Xinference framework for parallel deployment. Specific models include: marine ecological large model, Qwen-vl-chat, sd-turbo, ChatTTS, Bge-reranker-base, Bge-m3, etc.

[0130] The small model server includes various marine ecological artificial intelligence small models, mostly using Paddle and PyTorch frameworks, including: marine habitat prediction simulation analysis model, marine species identification model, marine habitat prediction simulation analysis model, species distribution prediction model, extreme weather prediction model, etc., and configures its production environment.

[0131] The database side is used to deploy various marine ecology-related relational databases and graph databases, including: marine ranch monitoring database and pollutant monitoring database, global biodiversity database, ecological disaster monitoring and early warning system, and configure its production environment.

[0132] The API end is used to uniformly call various API interfaces to ensure normal interface access, including: calculator, server system query commands, Keling API, seawater quality monitoring API, beach water quality weekly report API, Baidu translation API, Xinzhi Weather API, Bing search API, Amap API, Wikipedia database API, Arxiv API, etc., and configure its production environment.

[0133] The application side is mainly responsible for the construction of the front-end and supporting functions of the marine ecological multi-intelligence. The production environment to be configured includes: Chroma vector database, WebUI front-end, Langchain framework, Paddle framework, PyTorch framework, etc.

[0134] In addition, network and external network mapping port configuration is also required, including configuring the external network access interface of the marine ecological multi-intelligence body and configuring the API call interface of the large model server, small model server, database, API and other modules.

[0135] Based on the above configuration, the model is kept running normally, the production environment is connected to the Internet, and the model interface and API are available.

[0136] It should be noted that the steps shown in the above process or the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0137] Based on the above embodiments, this embodiment also provides a marine ecological multi-agent interaction system, and those already described will not be repeated. As used hereinafter, terms such as "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that can achieve a predetermined function.

[0138] Figure 8 is a schematic diagram of the principle of the marine ecological multi-agent interaction system according to the embodiments of the present application, as Figure 8 shown, the system includes:

[0139] A data layer 1, configured to generate fine-tuning data, a domain knowledge base, and a business database based on data sources, where the data sources include a fine-tuning data set, a knowledge data set, and data sources of marine ecological multi-agents;

[0140] A model layer 2, configured to build the marine ecological large model, the domain knowledge base, the multi-modal large model, and the sub-agents required for encapsulating the basic function agent, the Internet function agent, and the marine ecological function agent. The sub-agents include: a business database processing agent, a mathematical calculation agent, a server basic setting agent, a language translation agent, a weather query agent, a search engine agent, a map service agent, a paper retrieval agent, a knowledge encyclopedia agent, a quantity statistics agent, a species identification agent, a habitat simulation agent, and a species distribution prediction agent;

[0141] An agent layer 3, configured to, based on the model layer, fuse the sub-agents to build the basic function agent, the Internet function agent, and the marine ecological function agent;

[0142] The application layer 4 includes at least a fishery resource protection application module and an ecological protection application module. The fishery resource protection application module includes: a marine organism quantity statistics unit, a species identification unit, a habitat simulation unit, and a species distribution prediction unit. The ecological protection application module includes: a pollutant monitoring unit, an extreme weather prediction unit, and an ecological disaster warning unit. The fishery resource protection application module is constructed based on the marine ecological function agent. The marine organism quantity statistics unit, the species identification unit, the habitat simulation unit, and the species distribution prediction unit are respectively connected to the quantity statistics agent, the species identification agent, the habitat simulation agent, and the species distribution prediction agent. The ecological protection application module is constructed based on the marine ecological function agent. The pollutant monitoring unit, the extreme weather prediction unit, and the ecological disaster warning unit are respectively connected to the pollutant monitoring agent, the extreme weather prediction agent, and the ecological disaster warning agent;

[0143] The interaction layer 5 provides management of model configuration, multi-modal dialogue, domain knowledge dialogue, and knowledge base through a front-end interaction interface.

[0144] In some embodiments, referring to Figure 9 as shown, the front-end interaction interface at least includes:

[0145] A model configuration unit 501 for selecting a marine ecological large model;

[0146] A multi-modal dialogue configuration unit 502 for configuring a multi-modal large model. Among them, for text-to-image, the sd-turbo model is selected; for text-to-speech conversion, the ChatTTS model is selected; for text-to-video, the Kelin large model is selected; for image semantic question and answer, the Qwen-vl-chat model is selected;

[0147] A domain knowledge base dialogue configuration unit 503 for configuring at least one knowledge base and providing question and answer services based on the configured knowledge base. The customization of a new knowledge base and the editing of an existing knowledge base can be achieved through a domain knowledge base management unit;

[0148] A multi-agent configuration unit 504 obtains a multi-agent enabling request through a selection box and provides a multi-agent list for selecting at least one agent;

[0149] A dialogue interaction unit 505 uploads a file through a text input box or a file transfer button and conducts a dialogue with the marine ecological large model, obtains a user request, and sends it to the marine ecological large model. The output result after the marine ecological large model integrates semantics is obtained and displayed in the dialogue content display area. The dialogue interaction unit also provides a clear button to clear the historical chat content.

[0150] It should be noted that the above-mentioned modules can be functional modules or program modules. The above-mentioned modules can be located in the same processor; or the above-mentioned modules can also be located in different processors respectively in any combination form.

[0151] The above-described embodiments merely represent several implementation manners of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for constructing a marine ecological multi-agent, characterized in that: include: The steps of constructing the marine ecological large model are to build a fine-tuning data set, train the selected large language model based on the fine-tuning data set to obtain the marine ecological large model, and the selected large language model is Qwen2-7B. The LORA fine-tuning method is used for training based on the Llama_factory framework, and the LORA weights of each training round are screened. Qwen2-7B loads each LORA weight in turn, and conducts a question-and-answer test to eliminate overfitting and logically incorrect weights. The LORA weight with the smallest number of training rounds under the premise of fine-tuning function is selected, and the weights of Qwen2-7B and LORA are merged. The maximum block size is set to 2GB when merging, and the export quantization level is set to none; The parallel application module construction step includes constructing a domain knowledge base, a marine ecological multi-agent, and a multimodal large model. The domain knowledge base, the marine ecological multi-agent, the multimodal large model of the parallel application module and the marine ecological large model are deployed on the GPU server in a diversified parallel deployment mode. The knowledge data set of the domain knowledge base includes: marine environment data, human activity pressure data, and the human activity pressure data includes: socio-economic data, human activity data, The marine ecological multi-agent comprises: The pre-packaged basic function intelligent body, Internet function intelligent body, and marine ecological function intelligent body, the marine ecological multi-intelligent body is also configured with a trigger, and the trigger is configured as a keyword trigger, The multimodal large model includes a multimodal model, a text-image model, a text-speech model, and a text-video model. The basic function agent includes: a business database processing agent, a mathematical calculation agent and a server basic setting agent. The business database processing agent is obtained by encapsulating the API interface of the marine ecological business database into an agent. The marine ecological business database includes: a marine ranch monitoring database, a pollutant monitoring database, and a global biodiversity database. When the user request involves the keywords "marine ranch", "pollutant", "monitoring" and "biodiversity", the corresponding agent can be triggered; the mathematical calculation agent is obtained by encapsulating a calculator into an agent; the server basic setting agent is obtained by encapsulating a server system query command into an agent; The Internet functional agents include language translation agents, weather query agents, search engine agents, map service agents, paper retrieval agents, and knowledge encyclopedia agents. The marine ecological functional intelligent agent includes: a quantitative statistics intelligent agent, a species identification intelligent agent, a habitat simulation intelligent agent, and a species distribution prediction intelligent agent. The marine ecological functional intelligent agent also includes: a pollutant monitoring intelligent agent, an extreme weather prediction intelligent agent, and an ecological disaster early warning intelligent agent; the quantitative statistics intelligent agent is obtained by encapsulating the fish fine-grained recognition model into an intelligent agent, starting the intelligent agent and inputting the keyword "fish quantity statistics", triggering and running, returning the fish quantity statistics information of "fish species", "quantity", and "spatial coordinates", and giving the marine ecological large model the ability of fine-grained fish recognition and quantitative statistics; The species identification agent is obtained by encapsulating the marine species identification model into an agent, starting the agent and inputting the keywords "species name" and "marine ranch name" to trigger and run, returning the species name and spatial coordinates, and giving the marine ecological large model marine ranch common species identification capabilities; The habitat simulation agent is obtained by encapsulating the marine habitat prediction simulation analysis model into an agent, starting the agent and inputting the name of "marine ranch", "sea bathing beach" or station, triggering and running, returning the temperature, salinity, current, depth and water quality indicators of the specified marine area, and giving the marine ecological big model the ability to predict and analyze the temperature, salinity, current, depth and water quality indicators of the selected marine ranch, sea bathing beach or station; The species distribution prediction agent is obtained by encapsulating the species distribution prediction model into an agent, starting the agent and inputting "Maxent", "species distribution prediction", "species name", "latitude and longitude", and "environmental information", triggering and running, and returning species suitable habitat information; The marine ecological big model is used to parse the user request and distribute it to the parallel application module, triggering any one of the domain knowledge base, marine ecological multi-agent, and multimodal big model in the parallel application module or any combination thereof to generate a response result and feed it back to the marine ecological big model for semantic integration and output. The workflow of the parallel application module is turned on or off through the start node and the end node configuration, and the main model of the domain knowledge base, marine ecological multi-agent, and multimodal large model is set as the marine ecological large model through an LLM node, and a switch node is configured for the domain knowledge base, marine ecological multi-agent, and multimodal large model respectively, which is used to control whether the workflow of the domain knowledge base, marine ecological multi-agent, and multimodal large model is turned on. When the domain knowledge base, marine ecological multi-agent, and multimodal large model are all turned off, the marine ecological large model will directly output the result according to the user request, and will no longer trigger the domain knowledge base, marine ecological multi-agent, and multimodal large model.

2. The method for constructing a marine ecological multi-intelligent entity according to claim 1, characterized in that: The process of constructing the fine-tuning dataset includes: A data acquisition step, acquiring documents in the field of marine ecology, and parsing and extracting text content in the documents in the field of marine ecology; The document data processing step is to divide the text content into segments and extract the abstract of each segment; The fine-tuning dataset construction step stores the segment content and its summary to obtain a fine-tuning dataset.

3. The method for constructing a marine ecological multi-agent according to claim 1, characterized in that: The construction process of the domain knowledge base includes: The knowledge dataset acquisition step is to obtain marine environment data and human activity pressure data through the document loader; A data vectorization step, using a word segmenter to convert the marine environment data and human activity pressure data into text blocks, and performing vectorization processing based on an Embedding model to obtain a text vector; When the domain knowledge base obtains the distributed user request, the user request is vectorized and the text vector in the domain knowledge base is matched by similarity. After sorting the matching results using the Rerank model to obtain text vectors with high similarity, the text is integrated using the prompt word project and sent to the marine ecological model.

4. The method for constructing a marine ecological multi-agent according to claim 1, characterized in that: The language translation agent is obtained by encapsulating the translation platform API interface into an agent; The weather query agent is obtained by encapsulating the weather forecast platform API interface into an agent; The search engine agent is obtained by encapsulating the search engine API into an agent; The map service agent is obtained by encapsulating the map platform POI service API into an agent; The paper retrieval agent is obtained by encapsulating the library database API into an agent; The knowledge encyclopedia agent is obtained by encapsulating the Wikipedia database API interface into an agent.

5. The method for constructing a marine ecological multi-intelligent entity according to claim 1, characterized in that: The pollutant monitoring intelligent agent is obtained by obtaining seawater quality monitoring and bathing beach water quality information and encapsulating it into an intelligent agent; The extreme weather prediction intelligent agent is obtained by encapsulating the extreme weather prediction model into an intelligent agent; The ecological disaster early warning intelligent agent is obtained by encapsulating the interface of the ecological disaster monitoring and early warning system into an intelligent agent.

6. A marine ecological multi-agent interactive system, based on the marine ecological multi-agent construction method according to any one of claims 1 to 5, characterized in that: include: The data layer is used to generate fine-tuning data, domain knowledge base and business database based on data sources, including the data sources of the fine-tuning data set, knowledge data set and marine ecological multi-agent; The model layer is used to build the marine ecological large model, domain knowledge base, multimodal large model and encapsulate the sub-agents required for the basic functional agent, Internet functional agent, and marine ecological functional agent; The agent layer is used to integrate the sub-agents to build the basic function agent, Internet function agent, and marine ecological function agent based on the model layer; The application layer includes at least a fishery resource protection application module and an ecological protection application module. The fishery resource protection application module includes: a marine biological quantity counting unit, a species identification unit, a habitat simulation unit, and a species distribution prediction unit. The ecological protection application module includes: a pollutant monitoring unit, an extreme weather prediction unit, and an ecological disaster early warning unit. The interaction layer provides model configuration, multimodal dialogue, domain knowledge dialogue and knowledge base management through a front-end interactive interface.

Images