Business and financial data processing method and system based on large model, and storage medium

Abstract

The invention relates to the technical field of data processing, and discloses a business financial data processing method and system based on a large model and a storage medium, and a converged communication method comprises the following steps: S1, data acquisition and real-time preprocessing; s2, unified semantic space mapping; s3, performing causal analysis and knowledge graph construction; s4, automatically creating and dynamically updating a business-finance mapping rule; s5, risk identification, decision quantification and ROI optimization strategy generation: carrying out risk identification and decision quantification on the industry and financial multi-source data according to the business-financial mapping rule through a decision maker of the industry large model in combination with an expert voting mechanism, identifying a potential loss chain and generating an executable scheme; through cooperative work of the data acquisition module, the cloud computing module and the business and financial decision module, full-process intelligence of business and financial data from acquisition and processing to decision support is realized, and the efficiency and effect of business and financial cooperative management are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a business data processing method, system, and storage medium based on a large model. Background Technology

[0002] In enterprise operations, business data and financial data are crucial indicators of a company's operating status, and their effective integration and in-depth analysis are essential for informed decision-making. However, current business and financial data processing faces numerous challenges: Firstly, business and financial data come from a wide range of sources and are diverse in type, including structured data (such as sensor time-series data and financial vouchers) and unstructured data (such as equipment maintenance reports and supply chain contract texts). The semantic spaces of these different data types differ significantly, making effective integration difficult. Secondly, the mapping relationship between business and finance is complex and dynamically changing. Traditional methods often rely on manual analysis, which is inefficient and prone to errors. Furthermore, in risk identification and decision-making, existing technologies often struggle to quickly uncover potential risks from massive amounts of business and financial data. The decision-making process lacks quantitative support, resulting in insufficient targeting and effectiveness of countermeasures. Simultaneously, existing technologies are ill-suited for real-time data analysis, adapting to agile business changes, and directly generating optimization strategies with ROI analysis to drive business actions.

[0003] To address the above issues, we propose a business and financial data processing method, system, and storage medium based on a large model. Summary of the Invention

[0004] This invention provides a business data processing method, system, and storage medium based on a large model, which helps to solve the problems mentioned in the background art.

[0005] on the one hand

[0006] This invention provides the following technical solution: a business and financial data processing method based on a large model, the integrated communication method comprising the following steps:

[0007] S1. Data Acquisition and Real-time Preprocessing: Collect multi-source business and financial data, which includes structured data and unstructured data, and preprocess the collected data. The structured data includes sensor time-series data and financial vouchers, while the unstructured data includes equipment maintenance reports and supply chain contract texts.

[0008] S2. Unified Semantic Space Mapping: The encoder, which uses an industry-wide large model, maps the collected multi-source business and financial data to a unified semantic space. The encoder uses a deep neural network model based on Transformer.

[0009] S3. Causal Analysis and Knowledge Graph Construction: The parser of the industry's large model is used to perform causal analysis on multi-source data in a unified semantic space, to mine the causal relationships between the data, and to construct a causal knowledge graph based on the causal relationships.

[0010] S4. Automatic creation and dynamic updating of business-financial mapping rules: Based on causal knowledge graphs, the generator of the industry big model automatically creates business-financial mapping rules and adopts an incremental learning mechanism to adapt to agile business changes;

[0011] S5. Risk Identification, Decision Quantification, and ROI Optimization Strategy Generation: By combining the decision-maker of the industry-wide model with an expert voting mechanism and based on the business-financial mapping rules, risk identification and decision quantification are performed on multi-source business and financial data to identify potential loss chains and generate executable solutions.

[0012] By adopting the above technical solutions, we can overcome the shortcomings of existing technologies, such as difficulty in integrating business and financial data, reliance on manual maintenance of mapping relationships, difficulty in adapting to agile business changes, lack of intelligent and quantitative support for risk identification and decision-making, and inability to output optimization strategies with ROI analysis.

[0013] Furthermore, in step S1, when preprocessing the collected data, the Apache Flink stream processing algorithm is used to perform windowing processing on the sensor time-series data, as shown in the following formula:

[0014] ,in, for The set of sliding window data at each moment. For time series data points within the window, For data collection timestamps, The data is cleaned and its features are extracted in real time using the Apache Flink stream processing algorithm, with outliers removed and the data format standardized using the Z-score method, given a window length.

[0015] By adopting the above technical solutions, noise and format differences in the original data can be effectively eliminated, enabling the collected multi-source data to reach a unified standard in terms of data quality and feature representation. This lays a solid and reliable data foundation for achieving accurate data semantic alignment using the industry large model encoder in the subsequent unified semantic space mapping step.

[0016] Furthermore, the specific strategy of the Transformer-based deep neural network model in S2 is as follows:

[0017] For unstructured data, the formula for semantic embedding is: ,in, It is unstructured text. a semantic vector of the text;

[0018] For structured data, a feature embedding network is used to convert it into a vector: wherein, is a structured data feature matrix, is a multi-layer perceptron, is a semantic vector of the structured data;

[0019] The distribution of the semantic vector of the text and the semantic vector of the structured data is optimized by a metric learning loss function: wherein, The data pair has a semantic association label, and 1 represents correlation and 0 represents irrelevance, and cos is a cosine similarity function.

[0020] By using the above technical solution, the business and financial data sources are different, and the term definitions differ (for example, “cost” may refer to production cost on the business side, and includes allocated expenses on the financial side), and unified semantic mapping can standardize multi-source data terms and avoid data processing errors caused by semantic confusion.

[0021] Further, the parser in the S3 uses a causal graph neural network (CGNN) algorithm for causal analysis, specifically:

[0022] In the causal strength calculation, the CGNN quantifies the causal influence degree between variables, and the formula is: wherein, represents the causal strength of variable a on variable b, and the larger the value, the more significant the causal effect of variable a on variable b; is an intervention operation in causal inference, simulating the impact of artificially controlling the value of variable a on the system; is a probability distribution function, used to describe the likelihood of an event occurring;

[0023] In the causal knowledge graph construction phase, the system first sets a causal strength threshold to filter out statistically significant causal relationships. Then, the entities in the business data are taken as nodes, and the causal relationships that meet the conditions are taken as directed edges to construct a structured causal knowledge graph G=(V,E), wherein V is a node set containing business entities and financial indicators, and E is a relationship edge set labeled with causal strength and direction.

[0024] By using the above technical solution, by comparing the probability difference of variable b under the two states of intervention variable a and non-intervention variable a, the confounding factors are eliminated, and the true causal relationship is accurately identified; the graph can intuitively present the causal transmission path between business and financial data, and provide explainability basis for subsequent risk prediction and decision support.

[0025] Further, the generator in S4 adopts a reinforcement learning driven T5 generation model, taking the rule generation accuracy as the reward function R, with the formula being: wherein, is the rule matching accuracy, is the proportion of business scenarios covered by the rule, , is the weight coefficient;

[0026] The generator parameters are optimized by a deep reinforcement learning algorithm, i.e., DQN, to realize automatic updating of the mapping rules, which are represented as: wherein, is the th mapping rule, is the business indicator, is the financial indicator, is the business indicator trigger condition, is the financial indicator change amount, is the mapping function.

[0027] By using the above technical solution, the mapping relationship is automatically created to avoid manual configuration, the industry and financial data conversion time is shortened, the data logic is verified in real time relying on the algorithm, the financial data deviation caused by human error is reduced, the mapping rules are updated in real time according to business changes (such as adding a product line or adjusting the settlement method), and it is ensured that the financial accounting is always consistent with the business scenario.

[0028] Further, the risk identification in S5 adopts an anomaly detection algorithm based on an autoencoder to calculate the data reconstruction error as an anomaly score: wherein, is the input data vector, is the autoencoder reconstruction vector, is the anomaly score, when is determined as a risk point, and is the anomaly threshold;

[0029] The decision quantification introduces an analysis model, with the formula being: wherein, is the expected profit increment after executing the optimization strategy, is the strategy execution cost;

[0030] The expert voting mechanism adopts a weighted voting algorithm to aggregate expert opinions, with the formula being: wherein, is the final decision quantification value, is the weight of the th expert, is the expert score, is the number of experts;

[0031] Generate an optimization strategy with ROI analysis through the decision maker to drive business action.

[0032] By adopting the above technical scheme, through the systematic risk identification mechanism, the potential business risks and financial risks are made explicit, and the subjective decision is converted into data-driven objective judgment by combining the quantitative analysis tool, the decision error rate is effectively reduced, and the ROI-based optimization strategy can accurately locate the high-value business scene, dynamically adjust the resource allocation, avoid the inefficient investment of funds and manpower, and improve the overall operation efficiency of the enterprise.

[0033] On the other hand

[0034] An industry and finance data processing system based on a large model, comprising:

[0035] A data acquisition module for data acquisition and real-time preprocessing, and the data acquisition module is built-in Apache Flink stream processing engine, and the real-time reception and preprocessing of data are realized through distributed nodes.

[0036] A cloud computing module, the input end of the cloud computing module is electrically connected with the output end of the data acquisition module, and the cloud computing module is used for unified semantic space mapping, causal analysis and construction of knowledge graph, creation of business-finance mapping rules and dynamic updating, the cloud computing module comprises an industry large model, and the cloud computing module comprises an encoder, an analyzer and a generator.

[0037] An industry and finance decision module, the output end of the cloud computing module is electrically connected with the input end of the industry and finance decision module, and the industry and finance decision module is used for risk identification, decision quantification and ROI optimization strategy generation.

[0038] By adopting the above technical scheme, the distributed processing capability of Apache Flink can effectively cope with the high concurrency characteristics of industry and finance data, and guarantee the real-time performance of data acquisition.

[0039] Further, the encoder adopts BERT-Base pre-training model, adapts to the industry and finance field semantics through fine-tuning, the model structure contains 12 layers of Transformer encoder, the hidden layer dimension is 768, and the multi-head attention mechanism is adopted to capture the semantic association of text and structured data, and the attention weight calculation formula is: Wherein, , , Query, key, value matrix respectively, is the attention head dimension;

[0040] The analyzer adopts causal graph neural network (CGNN), which contains graph convolution layer and causal reasoning layer, and the calculation formula of the graph convolution layer is: ,in, Represents a node In the The hidden states of a layer are used to store feature information of nodes during neural network computation. Representative node The set of neighbors, that is, the set of nodes The set of other directly connected nodes. It is the first The layer's weight matrix is ​​used to perform linear transformations on the node features, adjusting the weights of different features. For activation functions;

[0041] The generator adopts the T5-3B generation model, and performs domain adaptation through reinforcement learning, namely DQN. The action space is candidate tokens for rule generation, and the state is the current knowledge graph subgraph. The rule generation strategy is optimized through the Q-learning algorithm.

[0042] By adopting the above technical solutions, the BERT-Base pre-trained model has powerful natural language understanding capabilities. Through domain fine-tuning, it can accurately capture the semantic features of business and financial texts. The multi-head attention mechanism can extract data features from multiple dimensions, effectively integrate structured data and unstructured text information, provide more accurate semantic representations for subsequent data parsing and generation, and improve the intelligence level of business and financial data processing.

[0043] Furthermore, the business and financial decision-making module includes a decision-maker and an expert voting mechanism unit. The decision-maker adopts an ROI optimization strategy, and the expert voting mechanism unit deploys a WebSocket real-time interactive interface to support online expert voting and aggregate opinions through a weighted algorithm. Finally, it outputs an optimization strategy with ROI analysis and connects with the enterprise business system through an API interface to drive business actions.

[0044] on the other hand

[0045] A storage medium storing a computer program that, when read and executed by a processor, performs steps of a business data processing method based on a large model.

[0046] The present invention has the following beneficial effects:

[0047] Real-time data acquisition and preprocessing are achieved through the Apache Flink stream processing algorithm. By combining the BERT encoder and feature embedding network, multi-source data is mapped to a unified semantic space, solving the problem of real-time fusion of structured and unstructured data and improving the timeliness and consistency of data processing.

[0048] The causal knowledge graph is constructed by using a causal graph neural network, and the internal correlation between industry and financial data is accurately mined through a causal strength calculation formula; the generator automatically creates and updates mapping rules based on a T5 model and a reinforcement learning algorithm, a reward function ensures the accuracy and scene coverage rate of the rules, does not need manual maintenance, can quickly adapt to agile changes of the business, and the rule updating efficiency is obviously improved.

[0049] The decision maker combines the self-encoder anomaly detection and ROI analysis model, accurately identifies the risk points through an anomaly score formula, and quantifies the decision benefits by means of an ROI calculation formula; the expert voting mechanism aggregates opinions by using a weighted algorithm, so that the decision is more objective. The generated optimization strategy directly includes the ROI analysis result, can be driven by an API interface to execute the business system, and shortens the conversion cycle from decision to action;

[0050] Through the cooperative work of the data acquisition module, the cloud computing module and the industry and financial decision module, the whole process of intelligentization of industry and financial data from acquisition, processing to decision support is realized, and the efficiency and effect of the industry and financial collaborative management are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 It is a method flowchart of the present application.

[0052] Figure 2 It is a system flowchart of the present application.

[0053] In the figure: 1, data acquisition module; 2, cloud computing module; 3, industry and financial decision module. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0055] Embodiment 1

[0056] Reference Figure 1 A large model-based industry and financial data processing method, a fusion communication method comprises the following steps:

[0057] S1. Data acquisition and real-time preprocessing: acquiring industry and financial multi-source data, the industry and financial multi-source data includes structured data and unstructured data, and the acquired data is preprocessed, wherein the structured data includes sensor time series data and financial vouchers, and the unstructured data includes equipment maintenance reports and supply chain contract texts;

[0058] In S1, the collected data is preprocessed, Apache Flink stream processing algorithm is used for window processing of sensor time series data, and the formula is as follows:

[0059] wherein, is the sliding window data set at time t, is the time series data point in the window, is the data collection timestamp, is the window length, and the Apache Flink stream processing algorithm is used to realize real-time cleaning and feature extraction of data, remove outliers, and use Z-score standardization method to unify data format.

[0060] By using the above technical scheme, the noise and format difference in the original data can be effectively eliminated, and the multi-source data collected can be unified in data quality and feature expression to lay a solid and reliable data foundation for subsequent unified semantic space mapping step, and use industry large model encoder to realize accurate data semantic alignment.

[0061] S2. Unified semantic space mapping: using the encoder of the industry large model to map the collected industry and financial multi-source data to a unified semantic space, the encoder uses a deep neural network model based on Transformer;

[0062] The specific strategy of the deep neural network model based on Transformer in S2 is as follows:

[0063] For unstructured data, the formula for semantic embedding is as follows: wherein, is the unstructured text, is the semantic vector of the text;

[0064] For structured data, a feature embedding network is used to convert it into a vector: wherein, is the feature matrix of structured data, is a multilayer perceptron, is the semantic vector of structured data;

[0065] The distribution of the semantic vector of the text and the semantic vector of the structured data is optimized by a metric learning loss function: wherein, is the semantic association label of the data pair , and 1 represents correlation and 0 represents irrelevant, and cos is the cosine similarity function;

[0066] ​S3. Causal Analysis and Knowledge Graph Construction: The parser of the industry's large model is used to perform causal analysis on multi-source data in a unified semantic space, to mine the causal relationships between the data, and to construct a causal knowledge graph based on the causal relationships.

[0067] In S3, the parser uses a causal graphical neural network, specifically the CGNN algorithm, for causal parsing.

[0068] To mitigate the causal strength calculation, the formula for quantifying the degree of causal influence between variables using CGNN is: ,in, This indicates the causal strength of variable a on variable b; the larger the value, the more significant the causal effect of variable a on variable b. It is an intervention operation in causal inference, simulating the impact on the system when the value of the human-controlled variable 'a' is taken. This is a probability distribution function used to describe the likelihood of an event occurring;

[0069] In the causal knowledge graph construction phase, the system first sets a causal strength threshold to filter out causal relationships with statistical significance. Subsequently, entities in the business data are used as nodes, and causal relationships that meet the conditions are used as directed edges to construct a structured causal knowledge graph G=(V,E), where V is the set of nodes containing business entities and financial indicators, and E is the set of relation edges labeled with causal strength and direction.

[0070] By adopting the above technical solution, and by comparing the probability difference of variable b under the two states of intervention variable a and non-intervention variable a, the interference of confounding factors is eliminated, and the true causal relationship is accurately identified. This graph can intuitively present the causal transmission path between business and financial data, and provide interpretable basis for subsequent risk prediction and decision support.

[0071] S4. Automatic creation and dynamic updating of business-financial mapping rules: Based on causal knowledge graphs, the generator of the industry big model automatically creates business-financial mapping rules and adopts an incremental learning mechanism to adapt to agile business changes;

[0072] The generator in S4 uses a reinforcement learning-driven T5 generation model, with the rule-based generation accuracy as the reward function R, as shown in the formula: ,in, For rule matching accuracy, The proportion of business scenarios covered by the rules. , These are the weighting coefficients;

[0073] The generator parameters are optimized using a deep reinforcement learning algorithm, DQN, to achieve automatic updating of the mapping rules, which are represented as follows: ,in, For the first a business indicator mapping rule, for a business indicator, for a financial indicator, for a business indicator trigger condition, for a financial indicator change amount, for a mapping function;

[0074] S5. Risk identification, decision quantification, and ROI optimization strategy generation: Through the decision maker of the industry large model, combined with the expert voting mechanism, according to the business-finance mapping rule, the risk identification and decision quantification are carried out on the multi-source data of industry and finance, the potential loss chain is identified, and the executable scheme is generated;

[0075] In S5, the risk identification adopts an anomaly detection algorithm based on a self-encoder to calculate the data reconstruction error as an anomaly score: wherein, is the input data vector, is the self-encoder reconstruction vector, is the anomaly score, when is determined as a risk point, and is an anomaly threshold;

[0076] The decision quantification introduces analysis model, and the calculation formula is: wherein, is the expected profit increment after executing the optimization strategy, is the strategy execution cost;

[0077] The expert voting mechanism adopts a weighted voting algorithm to aggregate expert opinions, and the formula is: wherein, is the final decision quantification value, is the weight of the expert, is the expert score, is the number of experts;

[0078] The decision maker generates an optimization strategy with ROI analysis to drive business action.

[0079] Embodiment 2

[0080] Referring to Figure 2 , a big model-based industry and finance data processing system comprises:

[0081] A data acquisition module 1 is used for data acquisition and real-time preprocessing, and the Apache Flink stream processing engine is built-in the data acquisition module 1, and the real-time reception and preprocessing of data are realized through distributed nodes;

[0082] A cloud computing module 2, an input end of the cloud computing module 2 is electrically connected with an output end of the data acquisition module 1, and the cloud computing module 2 is used for unified semantic space mapping, causal analysis and construction of a knowledge graph, creation of business-finance mapping rules and dynamic updating, the cloud computing module 2 includes an industry large model, and the cloud computing module 2 includes an encoder, a parser and a generator;

[0083] The encoder adopts a BERT-Base pre-training model, adapts the industry-finance field semantics through fine-tuning, the model structure includes 12 layers of Transformer encoders, the hidden layer dimension is 768, the multi-head attention mechanism is adopted to capture the semantic association of text and structured data, and the attention weight calculation formula is as follows: , , , query, key and value matrices respectively, is an attention head dimension;

[0084] The parser adopts a causal graph neural network (CGNN), includes a graph convolution layer and a causal reasoning layer, and the graph convolution layer calculation formula is as follows: , represents the hidden state of a node in the first layer, used for storing the feature information of the node in the neural network calculation process, represents a neighbor set of a node , that is, a set of other nodes directly connected with the node , is a weight matrix of the first layer, used for linear transformation of the node features and adjustment of the weights of different features, is an activation function;

[0085] The generator adopts a T5-3B generation model, performs field adaptation through reinforcement learning (DQN), the action space is a candidate token for rule generation, the state is a current knowledge graph subgraph, and the rule generation strategy is optimized through a Q-learning algorithm.

[0086] A business-finance decision module 3, an output end of the cloud computing module 2 is electrically connected with an input end of the business-finance decision module 3, and the business-finance decision module 3 is used for risk identification, decision quantization and ROI optimization strategy generation.

[0087] The business-finance decision module 3 includes a decision maker and an expert voting mechanism unit; the decision maker adopts an ROI optimization strategy, the expert voting mechanism unit deploys a WebSocket real-time interaction interface, supports online voting of experts and aggregates opinions through a weighted algorithm, finally outputs an optimization strategy with ROI analysis, and is connected with an enterprise business system through an API interface to drive business actions.

[0088] In another aspect

[0089] A storage medium stores a computer program that, when read and run by a processor, performs the steps of the big model-based industry and finance data processing method.

[0090] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0091] The above description is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A large model-based industry and finance data processing method, characterized in that, The fusion communication method comprises the following steps: S1. Data acquisition and real-time preprocessing: collecting industry and finance multi-source data, the industry and finance multi-source data comprising structured data and unstructured data, and preprocessing the collected data, wherein the structured data comprises sensor time series data and financial vouchers, and the unstructured data comprises equipment maintenance reports and supply chain contract texts; S2. Mapping to a unified semantic space: mapping the collected industry and finance multi-source data to a unified semantic space by using an encoder of an industry large model, the encoder using a deep neural network model based on a Transformer; S3. Causal analysis and construction of a knowledge graph: performing causal analysis on the multi-source data in the unified semantic space by using an analyzer of the industry large model, mining the causal relationships between the data, and constructing a causal knowledge graph based on the causal relationships; S4. Automatic creation and dynamic updating of business-finance mapping rules: automatically creating business-finance mapping rules based on the causal knowledge graph by using a generator of the industry large model, and adapting to business agile changes by using an incremental learning mechanism; S5. Risk identification, decision quantification, and ROI optimization strategy generation: identifying risks and quantifying decisions according to the business-finance mapping rules by using a decision maker of the industry large model in combination with an expert voting mechanism, identifying potential loss chains, and generating executable solutions.

2. The big model-based industry and finance data processing method according to claim 1, characterized in that, When preprocessing the collected data in S1, the Apache Flink stream processing algorithm is used for windowing processing of the sensor time series data, and the formula is as follows: wherein, is a sliding window data set of time series data, is a time series data point within the window, is a data collection timestamp, is a window length, real-time data cleaning and feature extraction are achieved by Apache Flink stream processing algorithm, outliers are removed and Z-score standardization method is used to unify the data format.

3. The big model-based industry and finance data processing method according to claim 2, characterized in that, The specific strategy of the deep neural network model based on the Transformer in S2 is as follows: For unstructured data, the formula for semantic embedding is: where, is the unstructured text, is the semantic vector of the text; For structured data, a feature embedding network is employed to convert it into a vector: where, is the structured data feature matrix, is a multilayer perceptron, is the semantic vector for structured data; The distribution of the semantic vectors of the text and the semantic vectors of the structured data is optimized by a metric learning loss function: wherein, data pair has a semantic association label, and 1 represents relevance and 0 represents irrelevance, and cos is a cosine similarity function.

4. The big model-based industry and finance data processing method according to claim 3, characterized in that, The analyzer uses a causal graph neural network (CGNN) algorithm for causal analysis, and the specific strategy is as follows: In the causal strength calculation mitigation, the CGNN quantifies the degree of causal influence between variables formula is: Wherein, The causal strength of variable a to variable b is represented, the larger the value, the more significant the causal effect of variable a on variable b; Is the intervention operation in causal inference, simulating the influence on the system when the value of variable a is artificially controlled; It is a probability distribution function, which is used to describe the possibility of an event. In the causal knowledge graph construction stage, the system first sets a causal strength threshold to filter out statistically significant causal relationships, and then constructs a structured causal knowledge graph G=(V,E) by taking entities in business data as nodes and causal relationships meeting the conditions as directed edges, wherein V is a node set containing business entities and financial indicators, and E is a relationship edge set labeled with causal strength and direction.

5. The big model-based industry and finance data processing method according to claim 4, characterized in that, The generator in the S4 adopts a reinforcement learning driven T5 generation model, taking the rule generation accuracy rate as a reward function R, and the formula is: Wherein, is the rule matching accuracy rate, is the proportion of business scenarios covered by the rule, , is the weight coefficient; The generator parameters are optimized using a deep reinforcement learning algorithm, DQN, to achieve automatic updating of the mapping rules, which are represented as follows: ,in, For the first Mapping rules, For business metrics, For financial indicators, For business metric trigger conditions, For changes in financial indicators, This is a mapping function.

6. The big model-based industry and finance data processing method according to claim 5, characterized in that, The risk identification in S5 uses an autoencoder-based anomaly detection algorithm to calculate data reconstruction error as an anomaly score: wherein, is an input data vector, is an autoencoder reconstruction vector, is an anomaly score when is determined as a risk point, and is an anomaly threshold; Decision quantization introduction An analysis model is calculated by the formula: Wherein, is the expected profit increment after the execution of the optimization strategy, is the strategy execution cost; The expert voting mechanism uses a weighted voting algorithm to aggregate expert opinions. The formula is as follows: ,in, Quantified value for the final decision. For the first The weight of each expert Rate the experts For the number of experts; The decision maker generates an optimization strategy with ROI analysis to drive business actions.

7. A large model-based industry and finance data processing system, characterized by, It comprises: a data acquisition module (1) for data acquisition and real-time preprocessing, wherein the Apache Flink stream processing engine is built-in the data acquisition module (1), and the real-time reception and preprocessing of data are realized by distributed nodes; a cloud computing module (2), wherein the input end of the cloud computing module (2) is electrically connected with the output end of the data acquisition module (1), and the cloud computing module (2) is used for mapping to a unified semantic space, causal analysis and construction of a knowledge graph, creation and dynamic updating of business-finance mapping rules, and the cloud computing module (2) comprises an industry large model, and the cloud computing module (2) comprises an encoder, an analyzer, and a generator; A finance decision module (3) is electrically connected to the output end of the terminal computing module (2) and the input end of the finance decision module (3), and is used for risk identification, decision quantification and ROI optimization strategy generation.

8. The big model-based industry and finance data processing system according to claim 7, characterized in that, The encoder adopts a BERT-Base pre-training model, adapts the industry and financial field semantics through fine-tuning, the model structure comprises 12 layers of Transformer encoders, the hidden layer dimension is 768, the multi-head attention mechanism is adopted to capture the semantic association of text and structured data, and the attention weight calculation formula is as follows: Wherein, , , are query, key and value matrices respectively, is the attention head dimension. The parser adopts a causal graph neural network (CGNN) comprising a graph convolution layer and a causal reasoning layer, and the calculation formula of the graph convolution layer is: wherein, represents a node The hidden state of the first layer is used to store the feature information of the node in the neural network calculation process, represents a neighbor set of the node , i.e. a set of other nodes directly connected to the node , and is a weight matrix of the first layer, which is used to perform linear transformation on the node features and adjust the weights of different features, is an activation function. The generator adopts a T5-3B generation model, performs field adaptation through reinforcement learning, i.e., DQN, the action space is candidate tokens generated by rules, the state is a current knowledge graph subgraph, and the rule generation strategy is optimized through a Q-learning algorithm.

9. The big model-based industry and finance data processing system according to claim 8, characterized in that, The finance decision module (3) comprises a decision maker and an expert voting mechanism unit; the decision maker adopts an ROI optimization strategy, the expert voting mechanism unit deploys a WebSocket real-time interaction interface, supports online voting of experts, aggregates opinions through a weighted algorithm, finally outputs an optimized strategy with ROI analysis, and is connected to an enterprise business system through an API interface to drive business actions.

10. A storage medium, characterized by The storage medium stores a computer program, which is read and run by the processor and executes the steps of the industry and finance data processing method based on the large model in any one of claims 1-7.

Images