An industrial risk identification method based on hotspot clustering algorithm and industrial large model intelligent agent

Abstract

The present invention belongs to the field of industrial risk identification technology, and proposes an industrial risk identification method based on a hotspot clustering algorithm and an industrial large model intelligent agent. The clustering algorithm and the large model are combined to discover useful information under massive data and accurately identify industrial risks. First, through the hotspot clustering algorithm, the position weighted average pooling and the bias tensor contrast fine-tuning method are used to realize the vectorization of text semantics, and hot topics are found from massive industrial text data through connected graph clustering and split optimization. Secondly, based on the semantic understanding ability of the large model, an industrial large model intelligent agent is constructed, and the industrial large model intelligent agent is used to judge the hot topics, including sentiment analysis and industry judgment, so as to screen out hot topics involving industrial risks. Thirdly, the inductive and summarizing ability of the large model is applied to summarize industrial development suggestions based on the summarized industrial risk hot topics.

Description

Technical Field

[0001] The present invention relates to the field of industrial risk identification, and in particular to an industrial risk identification method based on a hotspot clustering algorithm and an industrial large model intelligent body. Background Art

[0002] Large language models (LLMs) are neural network models based on deep learning, typically with billions or even hundreds of billions of parameters. These models are pre-trained using self-supervised or semi-supervised learning on large amounts of text data. By learning the grammatical, semantic, and contextual information of text from massive amounts of data, they are able to perform well in natural language processing (NLP) tasks and are widely used in fields such as text generation, machine translation, question-answering systems, and sentiment analysis.

[0003] The vast amount of text data available online harbors a wealth of hidden information. How to accurately and timely discover the information needed from this vast amount of data is a challenge researchers are dedicated to solving. Topic identification and clustering algorithms can be used to aggregate diverse and chaotic text data (such as news and forum data) into hot topics, thereby uncovering the useful information hidden within the data. Summary of the Invention

[0004] In order to solve the technical problem that the existing technology cannot accurately find and filter out hot topics involving industrial risks from massive industrial text data and summarize industrial development suggestions, the present invention proposes an industrial risk identification method based on a hot spot clustering algorithm and an industrial large model intelligent body, which realizes the accurate identification of industrial risks and provides industrial development suggestions.

[0005] The specific plan is as follows:

[0006] An industrial risk identification method based on hotspot clustering algorithm and industrial large model intelligent agent,

[0007] S1, constructing an industry dataset: constructing an industry dataset by using industry databases and crawling online data; the industry dataset reflects the characteristics of a specified industry and contains data with negative opinions;

[0008] S2, hotspot clustering; semantic vectorization processing is performed on the data in the industry dataset described in S1, and similarity calculation is performed to evaluate data relevance; data clustering analysis is performed through a connectivity graph, and clustering results are optimized by splitting subgraphs; clusters in the clustering results are sorted in reverse order based on the number of members in each class to obtain industry hotspot text;

[0009] S3, Industry Risk Hotspot Screening: Based on the industry knowledge base and the big model, an industry big model agent is constructed and tool components are configured. The industry hotspot text is input into the industry big model agent, and the industry big model agent is run to perform sentiment analysis and industry-related judgments to obtain hot topics related to industry risks.

[0010] S4, industry recommendation generation: Generate industry recommendations based on large models for industry risk hotspots.

[0011] Preferably, the method for semantic vectorization of data in step S2 is:

[0012] S21, build model: build a neural network model with a GPT architecture, wherein the neural network model includes at least one Transformer decoder layer and is configured with a position weighted average pooling module;

[0013] S22, parameter optimization: optimizing the parameters of the neural network model using a bias tensor comparison fine-tuning method, and introducing a learnable bias tensor into the loss function to perform similarity comparison calculation;

[0014] S23, generating semantic vectors: traversing all text elements Ti in the text set T to be analyzed, performing semantic encoding on each Ti through the neural network model, and generating corresponding semantic vectors Vi;

[0015] S24, construct a semantic vector set: aggregate the semantic vectors Vi corresponding to all text elements Ti, and construct a structured semantic vector set V = {Vi|Ti∈T}.

[0016] Preferably, the similarity calculation method in step S2 is as follows: traverse any two vector combinations in the semantic vector set V and calculate the cosine value between any two vectors using the vector cosine calculation formula. Based on the input similarity threshold and the cosine value between the vectors, determine whether there is a correlation between any two vectors. If the cosine value exceeds the threshold, it is determined that there is a correlation.

[0017] Preferably, the method for implementing cluster analysis of data by using correlation through a connectivity graph in step S2 is:

[0018] S2a, constructing a topological graph: using the text data in the industry dataset as vertices to construct a vertex set, using the correlation between the text data as the basis for whether there is an edge between the vertices to construct an adjacency matrix, using the cosine value between the text data as the weight of the edge, and constructing a weighted undirected graph G based on the text similarity relationship of the dataset T;

[0019] S2b, connected graph clustering: The interconnected vertices in the weighted undirected graph G have semantically related relationships; based on the connectivity between the vertices of the weighted undirected graph G, the graph G is divided into multiple connected graphs, each of which can be regarded as a cluster, and the vertices Ti contained in the connected graph are the class member objects contained in the cluster;

[0020] S2c, split optimization: If the size of a cluster exceeds the set range, remove the relationships with lower weights and re-split the subgraph according to the connectivity of the graph to obtain new clusters.

[0021] Preferably, in step 2, the data of each cluster in the clustering results have semantic correlation, and the number of class members, that is, the amount of data contained in the cluster, reflects the popularity of the corresponding topic; the clusters are sorted in reverse order according to the number of class members, and the clusters before the threshold are set as hot topics contained in the industry data set, and output as industry hot text.

[0022] Preferably, in step 3, the method for constructing the industrial large model intelligent agent is:

[0023] S31, constructing an industry knowledge base: the industry knowledge base comprises data including but not limited to: industry-related policies and regulations, industry knowledge graphs, and industry historical news data; performing vectorized feature extraction and storage on the data in the industry knowledge base;

[0024] S32, Large Language Model Base Selection and Loading: Select a large language model as the large model base for the industrial large model agent, and perform supervised fine-tuning on the large model base by analyzing examples with labeled data;

[0025] S33, configure tool components to build a processing flow: configure tool components of the agent, including but not limited to: input nodes, large model calls, vector library retrieval, external links, logical processing, special result output, and data processing;

[0026] S34, release functional interface: The industrial large model intelligent body provides services to the outside world as a functional interface.

[0027] Preferably, in step 3, the method for running the industrial large model agent is:

[0028] S3a, first input the pre-processed industry hot text as a parameter into the industry large model intelligent agent;

[0029] S3b, the industry model agent performs industry knowledge base retrieval and network retrieval based on the input industry hot text, and uses the retrieved data as background knowledge;

[0030] S3c, Sentiment Analysis: Based on the large-scale model's ability to analyze positive and negative sentiment, we construct positive and negative analysis prompts through the prompting project. We then incorporate background knowledge and industry hotspot text into the large-scale model prompts to determine whether the industry hotspot text contains negative sentiment.

[0031] S3d, Industry Judgment: Based on the logical judgment ability of the large model, it constructs industry-related judgment prompt words, fills background knowledge and industry hot text into the large model prompt words, and obtains and marks whether the industry hot text is relevant to the industry;

[0032] S3e: Based on the labeling results of steps S3c and S3d, irrelevant hot topic data are filtered, and filtered data that contains both derogatory sentiment and is related to the industry is selected as hot topics related to industry risks and output.

[0033] Preferably, in step 4, the method for generating industry recommendations based on the large model is:

[0034] S41, summarize hot topics of industry risks: select representative data from each topic obtained in step 3 and splice them;

[0035] S42, constructing prompt words: generating prompt words for industry suggestions through prompt engineering, and filling risk hotspot data into the prompt words;

[0036] S43, call the big model to generate industry recommendations for industry risk hotspots and output them.

[0037] An industrial risk identification framework based on a hotspot clustering algorithm and an industrial large-scale model agent, including:

[0038] Application layer: supports interface calls and embedded calls; deployed as a service through middleware, supports external applications to call through the REST protocol; provides support for encapsulation as a method library, supports external applications to make embedded calls by introducing the method library and calling the interface;

[0039] Industrial large model intelligent agent: including calling external industrial knowledge base, large language model services, prompt words for engineering construction and tool components for improving large model capabilities;

[0040] The industrial risk hotspot identification service unit includes: a hotspot clustering module for mining hot topics from industrial data, a semantic vectorization module for converting text data into semantic vectors, and a data classification module for preprocessing industrial data and filtering out interference data; a sentiment analysis module that calls the large language model service to perform positive and negative analysis on hot topic data, an industrial relevance identification module that calls the large language model service to judge the industrial relevance of hot topic data, and an industrial suggestion generation module that calls the large language model service to generate industrial development suggestions for negative industrial hot spot data.

[0041] Beneficial effects:

[0042] This paper proposes an industrial risk identification method based on a hotspot clustering algorithm and an industrial big-model agent. This combination of the clustering algorithm and the big model uncovers useful information from massive amounts of data and accurately identifies industrial risks. First, the hotspot clustering algorithm vectorizes text semantics through position-weighted average pooling and bias tensor contrast fine-tuning. Hot topics are then identified from massive amounts of industrial text data using connected graph clustering and split optimization. Second, leveraging the semantic understanding capabilities of the big model, an industrial big-model agent is constructed. This agent conducts judgment on hot topics, including sentiment analysis and industry judgment, thereby identifying hot topics associated with industrial risks. Compared to using the big model and clustering algorithm alone, this synergistic effect enables more accurate identification of hot topics related to industrial risks. Furthermore, the clustered hot topics are fed into the industrial big model. Compared to directly feeding the text into the big model, this process is optimized to save processing time and improve efficiency. Third, the big model's summarization capabilities are leveraged to formulate industrial development recommendations based on the identified hot topics of industrial risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of an industrial risk identification method based on a hotspot clustering algorithm and an industrial large model intelligent agent in an embodiment.

[0044] Figure 2 4 is a hotspot clustering flow chart in the embodiment.

[0045] Figure 3 It is an example graph of the undirected graph G in the embodiment.

[0046] Figure 4 This is a flowchart of industrial risk judgment in the embodiment.

[0047] Figure 5 Generate a flow chart for industry recommendations in the examples.

[0048] Figure 6 The following is a framework diagram of an industrial risk identification program based on a hotspot clustering algorithm and an industrial large model intelligent body in an embodiment. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown in the figure, an industrial risk identification method based on hotspot clustering algorithm and industrial large model intelligent agent is proposed.

[0051] S1, constructing an industry dataset: constructing an industry dataset by using industry databases and crawling online data; the industry dataset reflects the characteristics of a specified industry and contains data with negative opinions;

[0052] S2, hotspot clustering; semantic vectorization processing is performed on the data in the industry dataset described in S1, and similarity calculation is performed to evaluate data relevance; data clustering analysis is performed through a connectivity graph, and clustering results are optimized by splitting subgraphs; clusters in the clustering results are sorted in reverse order based on the number of members in each class to obtain industry hotspot text;

[0053] S3, Industry Risk Hotspot Screening: Based on the industry knowledge base and the big model, an industry big model agent is constructed and tool components are configured. The industry hotspot text is input into the industry big model agent, and the industry big model agent is run to perform sentiment analysis and industry-related judgments to obtain hot topics related to industry risks.

[0054] S4, industry recommendation generation: Generate industry recommendations based on large models for industry risk hotspots.

[0055] 1. Build an industry dataset

[0056] To identify hot topics in a specific industry, you first need to obtain a certain amount of industry-related text data. This data includes, but is not limited to, industry news, industry public opinion data, and industry-related regulations. Industry data generally comes from existing industry databases or data scraped from the internet.

[0057] By gathering all the acquired industry data together, we can construct the industry dataset to be analyzed. The industry dataset generally needs to include the following characteristics:

[0058] 1) The data contained in the industry dataset must be relevant to the designated industry.

[0059] 2) The industry dataset needs to be of a certain size to support the subsequent clustering steps to form industry hotspots.

[0060] 3) The data in the industrial data set needs to have good data quality. The length of each data should not be too short and should try to avoid containing interference information.

[0061] 4) The industry data set needs to contain a certain amount of data with negative views on the specified industry, otherwise it will be impossible to analyze industry risk topics.

[0062] 2. Clustering of industry data hotspots

[0063] The main process of industrial data hotspot clustering is as follows: Figure 2 As shown:

[0064] 1) Data Semantic Vectorization

[0065] The main purpose of semantic vectorization is to extract the semantic features contained in the data text and encode them into vector representations in a low-dimensional space through a model. This facilitates subsequent similarity calculations and correlation judgments.

[0066] This method uses a GPT-structured vectorization model and a Transformer-structured decoder layer to achieve text semantic vectorization through position-weighted average pooling and bias tensor contrast fine-tuning.

[0067] We denote the text set to be analyzed as T, and Ti is a member of T, that is, Ti∈T. This method traverses all data in T and extracts the semantic vector Vi of the text based on Ti through a vectorization method, thereby constructing a semantic vector set V.

[0068] 2) Data similarity calculation and correlation judgment

[0069] Since texts with similar semantics also have similar vectors, the semantic similarity between two texts can be determined by calculating the distance between their semantic vectors, and whether the two texts are related can be determined based on the similarity threshold.

[0070] This method traverses any two vector combinations in the semantic vector set V and calculates the cosine value between any two vectors using the vector cosine calculation formula. Based on the input similarity threshold and the cosine value between the vectors, it determines whether there is a correlation between any two vectors. If the cosine value exceeds the threshold, it is determined to be correlated.

[0071] 3) Build a topology map

[0072] like Figure 3 As shown in the figure, this method uses text data as vertices to construct a vertex set, uses the correlation between data as the basis for determining whether there is an edge between vertices to construct an adjacency matrix, uses the cosine value between data as the weight of the edge, and constructs a weighted undirected graph G based on text similarity relationships of the dataset T.

[0073] 4) Connected Graph Clustering

[0074] In graph theory, if there is a path from vertex i to vertex j in an undirected graph, then i and j are said to be connected. If any two points in a graph are connected, then the graph is called a connected graph.

[0075] This method assumes that interconnected vertices in an undirected graph G have semantically related relationships. Based on the connectivity between the vertices in the undirected graph G, the graph G is divided into multiple connected graphs. Each connected graph can be regarded as a cluster, and the vertices Ti contained in the connected graph are the class member objects contained in the cluster.

[0076] 5) Split Optimization

[0077] This method analyzes the clustering results of the connected graph. If the size of a cluster exceeds a certain range, the cluster may need to be split and optimized to prevent individual outliers from clustering semantically unrelated data. For clusters that need optimization, low-weight relationships are removed and subgraphs are split based on the graph's connectivity to obtain more compact clusters.

[0078] 6) Results collation

[0079] Through the above steps, we cluster the industry data into multiple clusters. The data within each cluster is semantically related. This method assumes that these clusters represent the topics within the industry dataset. The number of cluster members, or the amount of data within a cluster, reflects the popularity of the corresponding topic. We sort the clusters in descending order by the number of cluster members. The clusters ranked first represent the hot topics within the industry dataset.

[0080] 3. Identification of hot topics of industrial risks based on large language models

[0081] Because the large language model learns and stores a vast amount of knowledge during training, including common sense, professional knowledge, and cultural knowledge, it possesses a certain level of logical reasoning ability, enabling it to reason and make judgments based on given information. Through the above steps, we have identified hot topics in industry data. This method, leveraging the semantic understanding and reasoning capabilities of the large model, allows us to filter out topics related to industry risks from these hot topics.

[0082] 1) Data preprocessing

[0083] For the hot topics identified through clustering, we need to perform data preprocessing to convert the clustering results into a data format that is convenient for large-scale model analysis. Because each cluster contains multiple industry data texts as class members, and the prompts that the large-scale model can process are limited in length, we need to sample several data texts from each hot topic and splice them together for easy subsequent use.

[0084] 2) Build an intelligent entity for large industrial models

[0085] In order to fully activate the semantic understanding and logical reasoning capabilities of the large language model, this method needs to first build an industrial large model intelligent agent based on the large language model. The main workflow for building the industrial large model intelligent agent is as follows:

[0086] a) Building an industry knowledge base

[0087] Optionally, to improve the accuracy of the large language model, we need to build an external industry knowledge base to expand the industry knowledge available to the large model. The industry knowledge base mainly includes, but is not limited to, industry-related policies and regulations, industry knowledge graphs, and historical industry news data.

[0088] After acquiring industry-related data, this method performs vectorized feature extraction and storage on these data, making it easier to improve the accuracy, relevance, and timeliness of the analysis results through the subsequent use of RAG (Retrieval-Augmented Generation) technology.

[0089] b) Large language model base selection and loading

[0090] The core of an industrial large-scale intelligent agent is a large language model. We need to select an appropriate open-source or private large language model as the foundation for this large-scale intelligent agent. If conditions permit, we can use labeled data analysis examples to perform supervised fine-tuning (SFT) on the large-scale model foundation, adapting the model to the analysis task and improving its accuracy and stability.

[0091] c) Build processing flow through tool components

[0092] The agent adds additional tool components to the large language model, expanding its capabilities to accommodate more complex processing flows. Common tool components typically include: input nodes, large model calls, vector library retrieval, external links, logic processing, special result output, and data processing. Based on these components and the capabilities of the large model, this method plans the agent's processing flow and constructs an agent for identifying industry risk hotspots.

[0093] d) Publish a functional interface

[0094] After the intelligent body is built, it needs to be released to the outside world as a callable service. In the subsequent actual implementation process, the industrial large model intelligent body provides services to the outside world as a functional interface.

[0095] 3) Industry risk identification and industry large-scale model intelligent body operation process

[0096] Figure 4 The specific operation process of the industrial large model intelligent body is demonstrated:

[0097] a) First, the pre-processed industry hotspot text is input into the agent as a parameter.

[0098] b) The intelligent agent performs knowledge base retrieval and network retrieval based on the input industry hot text, and uses the retrieved data as background knowledge.

[0099] c) Based on the large-scale model's ability to analyze positive and negative sentiment, prompts for positive and negative sentiment analysis are constructed through prompt engineering. Background knowledge and industry hotspot text are incorporated into the large-scale model's prompts to detect and label whether the industry hotspot text contains negative sentiment.

[0100] d) Based on the logical judgment capabilities of the large model, construct industry-related judgment prompt words. Fill the large model prompt words with background knowledge and industry hot text, and obtain and mark whether the industry hot text is relevant to the industry.

[0101] e) Based on the marking results of c and d, filter out irrelevant hot topic data and select hot topics related to industrial risks.

[0102] 4. Generate industry recommendations based on large models

[0103] Through the above steps, we have analyzed the hot topics related to industry risks from the industry data set. Based on the summarization ability of the large language model, this method summarizes all the discovered hot topics of industry risks and generates suggestions for the subsequent development of the industry. Figure 5 As shown, the main process is as follows:

[0104] 1) Summarize hot topics of industry risks, select several data from each topic as representative topics, and splice all the data as the data basis for generating recommendations for the large model.

[0105] 2) Through prompt engineering, a prompt word (Prompt) is constructed to generate industry suggestions, and risk hotspot data is filled into the prompt word.

[0106] 3) Call the big model to generate and output industry recommendations for industry risk hotspots.

[0107] like Figure 6 As shown in the figure, an industrial risk identification framework based on hotspot clustering algorithm and industrial large model intelligent agent includes:

[0108] Application layer: supports interface calls and embedded calls; deployed as a service through middleware, supports external applications to call through the REST protocol; provides support for encapsulation as a method library, supports external applications to make embedded calls by introducing the method library and calling the interface;

[0109] Industrial large model intelligent agent: including calling external industrial knowledge base, large language model services, prompt words for engineering construction and tool components for improving large model capabilities;

[0110] The industrial risk hotspot identification service unit includes: a hotspot clustering module for mining hot topics from industrial data, a semantic vectorization module for converting text data into semantic vectors, and a data classification module for preprocessing industrial data and filtering out interference data; a sentiment analysis module that calls the large language model service to perform positive and negative analysis on hot topic data, an industrial relevance identification module that calls the large language model service to judge the industrial relevance of hot topic data, and an industrial suggestion generation module that calls the large language model service to generate industrial development suggestions for negative industrial hot spot data.

[0111] 1. This method relies on the capabilities of a large language model, with the underlying industry large model agent based on the large model. This includes an external industry knowledge base (industry knowledge vector library), a large language model service, prompt words constructed through prompt engineering, and tool components for improving the capabilities of the large model.

[0112] 2. In the actual implementation process, this method provides services to the application side as an independent service, mainly including the following functional modules:

[0113] 1) Hotspot clustering module is responsible for mining hot topics from industry data

[0114] 2) The semantic vectorization module is responsible for converting text data into semantic vectors.

[0115] 3) The data classification module is mainly responsible for industrial data preprocessing and filtering interference data.

[0116] 4) The sentiment analysis module is responsible for performing positive and negative analysis on hot topic data based on the positive and negative analysis capabilities of the large model.

[0117] 5) The industry relevance identification module is responsible for judging the industry relevance of hot topic data based on the large model reasoning capability.

[0118] 6) The industry suggestion generation module is responsible for generating industry development suggestions based on the summarization and conclusion capabilities of the large model and targeting negative industry hot data.

[0119] 3. The method can be called by external applications in a variety of ways as a functional service:

[0120] 1) This method supports deployment as a service through middleware and supports external applications calling it through the REST protocol.

[0121] 2) This method supports encapsulation as a method library, and supports external applications to make embedded calls by introducing the method library and calling the interface.

[0122] Example 2: A program constructed based on the method of the present invention is applied to industry A for example:

[0123] 1. Obtain news data related to Industry A through online collection. Through keyword retrieval, time period filtering, and data preprocessing, a news dataset related to Industry A was constructed. The data volume is approximately 39,809 items, and the data fields include: text, title, time, unique value, etc.

[0124] 2. Using the hot topic clustering algorithm, we discovered hot topics in the dataset. By setting the parameter threshold, we found the top 200 hot topics as shown in Table 1. The topic name is the summarized hot topic information, and the topic popularity is the news data contained in the topic:

[0125]

[0126] 3. Utilizing the semantic understanding and logical judgment capabilities of the large model, we conducted sentiment analysis and industry relevance screening on hot topics. Several hot topics related to industry risks were selected from 200 hot topics, as shown in Table 2:

[0127]

[0128] 4. Leveraging the analytical capabilities of the big model, the industry risk hot topics obtained in the previous steps are input into the big model as public opinion information. The big model analyzes the industry risk summary obtained, and the analysis results output by the big model agent are as follows:

[0129] The risks currently faced by the {A industry mainly include the following:

[0130] 1. Risks of technical route and market acceptance

[0131] 2. Infrastructure and cost risks

[0132] 3. Policy risks

[0133] 4. Market competition risk

[0134] In summary, Industry A faces significant risks in terms of technology selection, infrastructure development, policy reliance, and market competition, requiring joint efforts from businesses, governments, and society to address them.

[0135] Based on the macro model's analysis of industry A's risks, the macro model can be further used to identify challenges and issues facing industry A with respect to the government, consumers, and businesses under the current risk situation, and to provide corresponding recommendations. The industry recommendations generated by the macro model in this example are as follows:

[0136] I. Challenges and Suggestions for the Government

[0137] 1. Challenges

[0138] Lagging infrastructure construction, insufficient policy stability, and supply chain security risks

[0139] 2. Recommendations

[0140] Strengthen infrastructure planning, improve policy support system, and ensure supply chain security.

[0141] 2. Challenges and Suggestions for Consumers

[0142] III. Challenges and Suggestions for Enterprises

[0143] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be included in the scope of protection of the patent for the present invention.

Claims

1. An industrial risk identification method based on hotspot clustering algorithm and industrial large model intelligent agent, characterized by: S1, constructing an industry dataset: constructing an industry dataset by using industry databases and crawling online data; the industry dataset includes data with specified industry characteristics and containing negative opinions; S2, hotspot clustering; semantic vectorization processing is performed on the industry dataset described in S1, and similarity calculation is performed to obtain semantic correlation relationships, and a connectivity graph is further constructed to obtain clustering results. The clustering results are optimized by splitting the subgraph; the clusters in the clustering results are sorted in reverse order based on the number of members of the class in the connectivity graph to obtain industry hotspot text; S3, screening of industry risk hotspots: Fine-tune the large model, configure the industry knowledge base and large model tool components to obtain an industry large model intelligent agent; input hot spot clustering into the industry large model intelligent agent to obtain the industry hot text, run the industry large model intelligent agent to perform sentiment analysis and industry-related judgments in turn, and screen out hot topics related to industry risks; S4, industry suggestion generation: summarize industry risk hot topics, construct prompt words, call the big model to generate industry suggestions for industry risk hotspots and output them.

2. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: The method for data semantic vectorization in step S2 is: S21, build model: build a neural network model with a GPT architecture, wherein the neural network model includes at least one Transformer decoder layer and is configured with a position weighted average pooling module; S22, parameter optimization: optimizing the parameters of the neural network model using a bias tensor comparison fine-tuning method, and introducing a learnable bias tensor into the loss function to perform similarity comparison calculation; S23, generating semantic vectors: traversing all text elements Ti in the text set T to be analyzed, performing semantic encoding on each Ti through the neural network model, and generating corresponding semantic vectors Vi; S24, construct a semantic vector set: aggregate the semantic vectors Vi corresponding to all text elements Ti, and construct a structured semantic vector set V = {Vi|Ti∈T}.

3. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: The similarity calculation method described in step S2 is: traverse any two vector combinations in the semantic vector set V, calculate the cosine value between any two vectors using the vector cosine calculation formula, and judge whether there is a correlation between any two vectors based on the input similarity threshold and the size of the cosine value between the vectors. If the cosine value exceeds the threshold, it is judged that there is a correlation.

4. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: The method for implementing data cluster analysis through connectivity graph using correlation in step S2 is: S2a, constructing a topological graph: using the text data in the industry dataset as vertices to construct a vertex set, using the correlation between the text data as the basis for whether there is an edge between the vertices to construct an adjacency matrix, using the cosine value between the text data as the weight of the edge, and constructing a weighted undirected graph G based on the text similarity relationship of the dataset T; S2b, connected graph clustering: The interconnected vertices in the weighted undirected graph G have semantically related relationships; based on the connectivity between the vertices of the weighted undirected graph G, the graph G is divided into multiple connected graphs, each of which can be regarded as a cluster, and the vertices Ti contained in the connected graph are the class member objects contained in the cluster; S2c, split optimization: If the size of a cluster exceeds the set range, remove the relationships with lower weights and re-split the subgraph according to the connectivity of the graph to obtain new clusters.

5. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: In step 2, the data of each cluster in the clustering results have semantic correlation, and the number of class members, that is, the amount of data contained in the cluster, reflects the popularity of the corresponding topic; the clusters are sorted in reverse order according to the number of class members, and the clusters before the threshold are set as hot topics contained in the industry data set, and output as industry hot text.

6. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: In step 3, the method for constructing the industrial large model intelligent agent is: S31, constructing an industry knowledge base: the industry knowledge base comprises data including but not limited to: industry-related policies and regulations, industry knowledge graphs, and industry historical news data; performing vectorized feature extraction and storage on the data in the industry knowledge base; S32, Large Language Model Base Selection and Loading: Select a large language model as the large model base for the industrial large model agent, and perform supervised fine-tuning on the large model base by analyzing examples with labeled data; S33, configure tool components to build a processing flow: configure tool components of the agent, including but not limited to: input nodes, large model calls, vector library retrieval, external links, logical processing, special result output, and data processing; S34, release functional interface: The industrial large model intelligent body provides services to the outside world as a functional interface.

7. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: In step 3, the method for running the industrial large model agent is: S3a, first input the pre-processed industry hot text as a parameter into the industry large model intelligent agent; S3b, the industry model agent performs industry knowledge base retrieval and network retrieval based on the input industry hot text, and uses the retrieved data as background knowledge; S3c, Sentiment Analysis: Based on the large-scale model's ability to analyze positive and negative sentiment, we construct positive and negative analysis prompts through the prompting project. We then incorporate background knowledge and industry hotspot text into the large-scale model prompts to determine whether the industry hotspot text contains negative sentiment. S3d, Industry Judgment: Based on the logical judgment ability of the large model, it constructs industry-related judgment prompt words, fills background knowledge and industry hot text into the large model prompt words, and obtains and marks whether the industry hot text is relevant to the industry; S3e: Based on the labeling results of steps S3c and S3d, irrelevant hot topic data are filtered, and filtered data that contains both derogatory sentiment and is related to the industry is selected as hot topics related to industry risks and output.

8. The industrial risk identification method based on the hotspot clustering algorithm and the industrial large model intelligent agent according to claim 1 is characterized in that: In step 4, the large model generates industry recommendations for hot topics by: S41, summarize hot topics of industry risks: select representative data from each topic obtained in step 3 and splice them; S42, constructing prompt words: generating prompt words for industry suggestions through prompt engineering, and filling risk hotspot data into the prompt words; S43, call the big model to generate industry recommendations for industry risk hotspots and output them.

Images