Big model-based enterprise data intelligent analysis method and system

Abstract

This invention relates to the field of enterprise data processing technology, and discloses an intelligent enterprise data analysis method and system based on a large model. It integrates RPA automated data acquisition components to construct multi-channel data access channels adapted to enterprise financial accounting systems, electronic tax bureaus, invoicing systems, and social security systems, enabling automated acquisition of multi-source heterogeneous data. The acquired multi-source heterogeneous data is cleaned, standardized, and correlated to obtain preprocessed data. The structured and unstructured data in the preprocessed data are fused across modalities, and feature extraction using an attention mechanism is employed to uncover potential correlations between data, resulting in an enterprise feature vector. This enterprise feature vector is input into a risk assessment model to identify enterprise operational risk points and generate a risk assessment report. The DQN algorithm is used to generate tiered response plans, outputting both the risk assessment report and the tiered response plans. This invention improves the efficiency and automation level of enterprise data analysis.

Description

Technical Field

[0001] This invention relates to the field of enterprise data processing technology, specifically to an intelligent analysis method and system for enterprise data based on a large model. Background Technology

[0002] Currently, enterprise operational data is scattered across multiple independent business systems such as financial accounting, tax invoicing, and social security management, resulting in a large amount of multi-source heterogeneous data. This data generally suffers from problems such as data redundancy, inconsistent formats, and poor correlation. Traditional data analysis methods rely heavily on manual processing and rule screening, and can only perform simple statistics on structured data. They cannot achieve cross-modal fusion processing of structured data and unstructured text data, making it difficult to deeply explore the potential correlation characteristics of multi-source data. Furthermore, their identification of enterprise operational risks is limited in scope and has low relevance, failing to meet the needs of enterprises for refined and intelligent operational analysis and risk management. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an intelligent analysis method and system for enterprise data based on a large model.

[0004] The first aspect of this invention provides an intelligent analysis method for enterprise data based on a large model, the method comprising the following steps: Integrate RPA automated data collection components to build a multi-channel data access channel that is compatible with enterprise financial accounting systems, electronic tax bureaus, invoicing systems, and social security systems, and to automatically collect multi-source heterogeneous data; The collected multi-source heterogeneous data is cleaned, standardized, and correlated to obtain preprocessed data. The preprocessed data, including both structured and unstructured data, is fused across modalities. An attention mechanism is then used for feature extraction to uncover potential relationships between the data and obtain enterprise feature vectors. By inputting enterprise feature vectors into a risk assessment model that integrates knowledge graphs and XGBoost algorithms, the risk points of enterprise operations are identified, and a risk assessment report is generated. Based on the risk assessment report, the DQN algorithm is used to generate a tiered response plan, and the risk assessment report and the tiered response plan are output through a natural language interaction interface.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the integration of the RPA automated data collection component to construct a multi-channel data access channel adapted to the enterprise financial accounting system, electronic tax bureau, invoicing system, and social security system, and to perform automated collection of multi-source heterogeneous data, includes: Pre-build interface feature libraries and page element feature libraries for each business system, and match the access paths, data field formats and permission verifications of the financial accounting system, electronic tax bureau, invoicing system and social security system one by one; Based on the preset collection dimensions, initialize three types of task instructions: timed collection, incremental collection, and triggered collection, determine the range of the dataset to be collected, and define the collection boundaries between structured and unstructured data. Based on RPA automated data collection components, form data, report data, log data, and text attachment data are traversed in layers to capture structured ledger data and unstructured documents, announcements, and notes from various systems, and integrated to obtain multi-source heterogeneous data.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of cleaning, standardizing, and associating the collected multi-source heterogeneous data to obtain preprocessed data includes: For multi-source heterogeneous data, hash matching is performed using a unique identifier field to remove duplicate data, and missing values ​​are identified and filled by traversing all fields. The 3σ principle is used to identify and filter multi-source heterogeneous data. The mean and dispersion range of data distribution for each field are calculated. Data that exceeds the preset fluctuation range is identified as abnormal data and is removed. The cleaned data is processed using range normalization to obtain normalized data. Based on the enterprise entity, business documents, and time nodes, the normalized data is mapped to obtain preprocessed data.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the cross-modal fusion of structured and unstructured data in the preprocessed data includes: The preprocessed data is categorized according to its data storage attributes to obtain structured and unstructured data. Structured and unstructured data are encoded and converted separately. Based on the business relevance dimension, the encoded structured and unstructured data are matched at the position and semantic level to establish corresponding relationships between the data. By combining and superimposing paired structured and unstructured representation information layer by layer, a cross-modal fusion data set is obtained.

[0008] Optionally, in the fourth implementation of the first aspect of the present invention, the feature extraction using an attention mechanism to mine potential correlations between data and obtain enterprise feature vectors includes: Weights are calculated for all feature dimensions in the cross-modal fusion dataset using a multi-head attention mechanism; Based on the calculated attention weights, potential business relationships between different data are mined to obtain target features, and these target features are then aggregated and encoded. The aggregated target features are encoded and transformed to generate enterprise feature vectors.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the step of inputting the enterprise feature vector into a risk assessment model that integrates a knowledge graph and the XGBoost algorithm to identify enterprise operational risk points and generate a risk assessment report includes: The enterprise feature vector is input into the risk assessment model, and the enterprise feature vector is mapped and matched with the entity nodes of the pre-built knowledge graph to associate the business scenario, risk dimension and risk entity corresponding to the feature vector; The enterprise feature vectors that have completed the graph association are input into the XGBoost model. The input features are split and iteratively analyzed to distinguish between normal operating features and abnormal risk features, and to identify multi-dimensional risk points and risk types in the enterprise's finance, taxation, compliance, and social security. By combining the risk weights of the knowledge graph with the output of the XGBoost model, the severity of the identified risks is quantified and graded, distinguishing between minor risks, general risks, severe risks, and major risks. The risk level, risk type, and risk location are then integrated to generate a risk assessment report.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the step of generating a tiered response plan based on the risk assessment report using the DQN algorithm, and outputting the risk assessment report and the tiered response plan through a natural language interaction interface, includes: The risk level, risk type, and risk location in the risk assessment report are used as environmental status characteristics and input into the DQN algorithm. The DQN algorithm uses the current state characteristics to call the main Q network and the target Q network, iteratively calculates the value and benefits of risk management actions, and selects the optimal combination of management actions. Based on the optimal combination of actions and the severity of the risk, corresponding response strategies are matched for different levels of risk, and converted into text format. Risk assessment reports and tiered response plans are then output through a natural language interaction interface.

[0011] A second aspect of the present invention provides an intelligent enterprise data analysis system based on a large model, the system comprising: The data acquisition module is used to integrate RPA automated acquisition components to build a multi-channel data access channel that is compatible with enterprise financial accounting system, electronic tax bureau, invoicing system and social security system, and to automatically collect multi-source heterogeneous data. The data processing module is used to clean, standardize, and correlate the collected multi-source heterogeneous data to obtain preprocessed data. The feature extraction module is used to perform cross-modal fusion of structured and unstructured data in the preprocessed data, and to extract features using an attention mechanism to mine potential correlations between data and obtain enterprise feature vectors. The model recognition module is used to input enterprise feature vectors into a risk assessment model that integrates knowledge graphs and XGBoost algorithms to identify enterprise operational risk points and generate risk assessment reports. The results output module is used to generate tiered response plans based on the risk assessment report using the DQN algorithm, and outputs the risk assessment report and tiered response plans through a natural language interaction interface.

[0012] A third aspect of the present invention provides an intelligent enterprise data analysis device based on a large model, the intelligent enterprise data analysis device based on a large model includes a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the intelligent enterprise data analysis device based on a large model to perform the various steps of the intelligent enterprise data analysis method based on a large model as described in any of the preceding claims.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the enterprise data intelligent analysis method based on a large model as described in any of the preceding claims.

[0014] The technical solution provided by this invention integrates RPA automated data acquisition components to construct a multi-channel data access channel adapted to enterprise financial accounting systems, electronic tax bureaus, invoicing systems, and social security systems, enabling automated acquisition of multi-source heterogeneous data. The acquired multi-source heterogeneous data is cleaned, standardized, and correlated to obtain preprocessed data. The structured and unstructured data in the preprocessed data are fused across modalities, and feature extraction using an attention mechanism is employed to uncover potential correlations between data, resulting in enterprise feature vectors. These enterprise feature vectors are input into a risk assessment model that integrates knowledge graphs and the XGBoost algorithm to identify enterprise operational risk points and generate a risk assessment report. Based on the risk assessment report, a DQN algorithm is used to generate tiered response plans, and the risk assessment report and tiered response plans are output through a natural language interaction interface. This invention achieves a fully intelligent closed-loop process for enterprise data acquisition, feature analysis, risk identification, and decision output, significantly improving the efficiency of enterprise data analysis, automation level, and the professionalism and standardization of risk management. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart of an intelligent enterprise data analysis method based on a large model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an enterprise data intelligent analysis system based on a large model, provided in an embodiment of the present invention. Figure 3 A schematic diagram of the structure of an intelligent enterprise data analysis device based on a large model, provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the intelligent enterprise data analysis method based on a large model provided in this embodiment of the invention includes the following steps: Step 101: Integrate RPA automated data collection components to build a multi-channel data access channel that is compatible with the enterprise's financial accounting system, electronic tax bureau, invoicing system, and social security system, and to carry out automated collection of multi-source heterogeneous data; In this embodiment, multi-business system adaptation modeling is completed in advance, and a unified system interface feature library and page element feature library are constructed to adapt to the enterprise's business scenarios. The interface protocols, page control elements, data return formats, and interaction logic parameters within the feature library are batch-entered and solidified. Comprehensive feature matching is carried out for four mainstream enterprise business systems: financial accounting system, electronic tax bureau, invoicing system, and social security system. The network access path, core data field definition, field storage format, system login mechanism, and permission verification logic of each business system are verified and confirmed one by one, thus completing the pre-adaptation and feature binding of multi-source business systems. After completing the multi-system feature adaptation, based on the enterprise's preset data analysis dimensions and business collection requirements, the data collection tasks are initialized, decomposed, and configured. Three independent automated task instructions are built and started: timed periodic collection, incremental data update collection, and business event triggered collection. The execution timing, collection frequency, and data update rules for each type of collection task are clearly defined. The specific data set range required by each business system is accurately identified. The collection boundaries and data classification standards of numerical and ledger-type structured data and text and attachment-type unstructured data are strictly distinguished, and the fine-grained configuration of the collection tasks is completed. Based on the configured data collection task instructions, the RPA automated data collection component is invoked to perform a full-domain data crawling operation. According to the data storage hierarchy of the business system, the system's internal form filling data, statistical report data, system operation log data, and various text attachment data are traversed layer by layer from top to bottom. The structured ledger data and business data retained by each business system are crawled and collected simultaneously. At the same time, unstructured business documents, public notice texts, business remarks and other data retained by the system are collected in batches. All the crawled raw data are uniformly collected and temporarily stored, and finally integrated to obtain a complete enterprise multi-source heterogeneous raw dataset.

[0019] Step 102: Clean, standardize, and perform correlation mapping on the collected multi-source heterogeneous data to obtain preprocessed data; In this embodiment, for the massive amount of multi-source heterogeneous raw data collected, a unified data duplication detection and missing data repair process is adopted. The globally unique identifier field of each data entry is extracted as the matching primary key. A full-domain hash comparison is used to traverse and match all data entries one by one, verifying the content, field information, and storage characteristics of each data entry. Data entries with complete duplication, partial duplication, or redundant replication are accurately identified and removed in batches. After deduplication, the system continues to traverse all data fields in the current dataset, detecting data defects such as empty field content, missing field information, and incomplete field content. Combining the original data logic of the same source business system with the corresponding business field attributes, the identified missing data is filled and repaired in compliance with regulations, achieving basic normalization processing of multi-source heterogeneous data. After deduplication and completion processing, a comprehensive anomaly detection and screening operation is performed on the multi-source heterogeneous data. The 3σ statistical discriminant principle is used as the standard for judging anomalies. For each independent data field in the dataset, the overall distribution of all sample data under that field is statistically analyzed. The mean parameter and overall discrete distribution range of the data samples for each field are calculated one by one to accurately define the normal fluctuation range of single field data. All sample data with values ​​exceeding the normal preset fluctuation range in a field are uniformly marked as deviation anomalies. The marked anomalies are reviewed and confirmed one by one. Under the premise of not damaging the normal business data structure, a batch removal operation is performed on all confirmed anomalies to completely remove deviation, error, and invalid data samples from the dataset. The clean datasets, after deduplication, missing data imputation, and anomaly filtering, are uniformly incorporated into a standardized processing flow. Throughout the process, a range normalization algorithm is employed to uniformly transform various business data with different dimensions, orders of magnitude, and value ranges. Range compression and unified mapping are performed on the values ​​of all data dimensions to ensure consistent value standards across all dimensions, resulting in a normalized dataset with consistent dimensions and dimensions. After data standardization, a cross-source data association and mapping process is initiated. Using the enterprise's business entity identity information, unique serial numbers of various business documents, and the timeline of business transactions as three core association links, and following the enterprise's actual business flow logic and data hierarchy, normalized data scattered across different business systems and data dimensions are comprehensively associated, bound, and structured, completing the full-domain integration of multi-source heterogeneous data. Finally, pre-processed data with complete structure, logical connections, and unified definitions is generated.

[0020] Step 103: Perform cross-modal fusion of structured and unstructured data in the preprocessed data, and use attention mechanism to extract features to mine potential correlations between data and obtain enterprise feature vectors; In this embodiment, the preprocessed enterprise data is comprehensively categorized and sorted. Based on the original data storage format, data organization form, and business representation attributes, the data is finely classified. The structured data, which is carried by numerical ledgers, form fields, and statistical indicators, is strictly distinguished from the unstructured data, which is carried by text documents, business notes, announcements, and contract content. The data categories are split by screening and classifying the data one by one across the entire domain, forming independent and complete subsets of structured data and subsets of unstructured data. A unified encoding and transformation process is performed on the two types of data with different attributes after splitting. This process adapts the numerical encoding rules for structured data and the semantic encoding rules for unstructured data, respectively, transforming the original data with different forms and representation logics into standardized representation vectors that can participate in model calculations. After encoding, based on the pre-defined association dimensions of enterprise business operations, bidirectional matching and verification are performed on the two types of coded representation data. Simultaneously, the numerical position alignment of structured data and the semantic content alignment of unstructured data are completed, achieving accurate cross-modal data matching. Business association mapping relationships are established between structured and unstructured data item by item, forming a bimodal paired dataset with corresponding logic. The structured and unstructured data representation information that has been accurately matched is processed in a hierarchical manner and fused. The feature superposition and information integration operations are carried out layer by layer from shallow to deep according to the data business weight level. The multi-dimensional information combination of the basic feature layer, the associated feature layer, and the deep semantic layer is completed in sequence. The core representation content of the two types of data is gradually integrated, eliminating the representation differences and dimensional barriers between different data modalities. Finally, a cross-modal fusion data set with complete information, unified features, and close correlation is generated.

[0021] In this embodiment, for the processed cross-modal fusion dataset, the trained multi-head attention mechanism model is invoked to perform full-domain weight calculation on all internal feature dimensions. The structured numerical feature dimensions and unstructured semantic feature dimensions included in the fusion data are fully covered. Through multi-head independent feature parsing, features of different types and levels are scored for correlation, and the initial attention weight parameters corresponding to each feature dimension are independently output, completing the fine-grained weight quantization and assignment operation for all feature dimensions. Based on the final output attention weight values ​​of each feature dimension, high-weight core features and low-weight redundant features are stratified and identified. The intrinsic relationships between features are deeply analyzed based on the weight distribution pattern to mine the surface data characteristics. The system identifies implicit business logic and potential business relationships that cannot be readily apparent. It then selects core target features that accurately reflect the company's operational and risk status, eliminates invalid and interfering features, and performs unified collection, orderly integration, and deep aggregation coding on all selected target features to form a well-structured, high-value feature set. This aggregated feature set undergoes standardized coding transformation, uniformly adjusting the dimensional structure and representation of the target features. It comprehensively organizes and maps multi-dimensional, scattered feature information, eliminating structural biases caused by differences in feature dimensions and representation forms. Finally, it completes the vector quantization transformation of all target features, generating a standardized enterprise operational feature vector that is dimensionally unified, information-condensed, and compatible with model input standards.

[0022] Step 104: Input the enterprise feature vector into the risk assessment model that integrates knowledge graph and XGBoost algorithm to identify the enterprise's operational risk points and generate a risk assessment report; In this embodiment, the generated enterprise feature vectors are imported into a pre-deployed risk assessment model. The enterprise operational risk knowledge graph system is retrieved, and the various feature elements contained in the feature vectors are extracted sequentially. The content of each element is compared and matched with the various entity nodes stored within the knowledge graph. A two-way correspondence is established according to business affiliation logic. Based on the matching results, the actual business scenario to which the feature vector belongs is determined, the corresponding risk assessment dimension is defined, and the various risk entity objects are precisely bound to it. The enterprise feature vectors that have completed knowledge graph association and binding are then fed into the trained XGBoost algorithm model. The model performs layer-by-layer splitting and iterative analysis of the input feature data according to its built-in operation rules, comparing the numerical performance and change patterns of each feature to classify those that conform to normal business operations. By identifying both regular and unusual characteristics that deviate from standard patterns, a multi-dimensional screening process is conducted based on feature differences. This accurately pinpoints specific risk points within a company's financial revenue and expenditure, tax declaration, operational compliance, and social security payment sectors, while also determining the specific category of each risk. The system retrieves pre-set risk weighting criteria from a knowledge graph and combines them with feature analysis conclusions from iterative calculations using the XGBoost model. These two reference bases are then used to quantitatively assess the level of each identified risk. Risks are categorized into four levels—minor, general, severe, and major—based on their impact scope, probability of triggering, and degree of loss. All risk level definitions, specific risk categories, and actual risk location information are then compiled and arranged according to a standardized format to ultimately form a complete and standardized corporate risk assessment report.

[0023] Step 105: Based on the risk assessment report, use the DQN algorithm to generate a tiered response plan, and output the risk assessment report and the tiered response plan through a natural language interaction interface.

[0024] In this embodiment, the system performs structured parsing and feature extraction on the generated standardized risk assessment report, comprehensively deconstructs all the core risk elements contained in the report, and fully extracts the quantitative risk level, specific risk classification type, precise risk location, risk impact range, risk triggering factors, and other related information corresponding to each risk. All the parsed risk status information is integrated and encapsulated, and uniformly constructed as the environment state input features of the DQN reinforcement learning algorithm to complete the environment state initialization configuration before the algorithm runs. After receiving complete risk environment state characteristics, the DQN algorithm initiates a dual-network collaborative iterative computation mechanism, simultaneously calling the main Q-network and target Q-network built into the model. Based on the current real risk state environment of the enterprise, it traverses all preset risk disposal action sets, calculates the long-term value returns and scenario adaptability of different disposal actions, disposal order, and disposal combinations one by one, and continuously optimizes the value score of each disposal action through multiple rounds of parameter updates, error correction, and iterative convergence computation, eliminates inefficient and unsuitable disposal actions, and finally selects the risk disposal action combination that is suitable for the current enterprise's multi-dimensional risk scenarios and has the best comprehensive disposal effect. Based on the optimal risk management action combination output after the DQN algorithm iterative convergence, and combined with the risk severity level determined by the knowledge graph and XGBoost model in the early stage, differentiated, hierarchical, and targeted exclusive rectification and management strategies and closed-loop management processes are matched for minor risks, general risks, severe risks, and major risks respectively. The structured algorithm action instructions, management logic, rectification requirements, and management priorities are systematically sorted out. Through the large model semantic conversion mechanism, the machine-recognized structured data is uniformly converted into standardized, fluent, and directly executable natural language text content. Finally, through the natural language interaction interface configured by the system, a complete structured risk assessment report and refined graded response plans adapted to different risk levels are output to the outside world.

[0025] Please see Figure 2 A schematic diagram of the structure of an enterprise data intelligent analysis system based on a large model provided in this embodiment of the invention. The system includes: The data acquisition module is used to integrate RPA automated acquisition components to build a multi-channel data access channel that is compatible with enterprise financial accounting system, electronic tax bureau, invoicing system and social security system, and to automatically collect multi-source heterogeneous data. The data processing module is used to clean, standardize, and correlate the collected multi-source heterogeneous data to obtain preprocessed data. The feature extraction module is used to perform cross-modal fusion of structured and unstructured data in the preprocessed data, and to extract features using an attention mechanism to mine potential correlations between data and obtain enterprise feature vectors. The model recognition module is used to input enterprise feature vectors into a risk assessment model that integrates knowledge graphs and XGBoost algorithms to identify enterprise operational risk points and generate risk assessment reports. The results output module is used to generate tiered response plans based on the risk assessment report using the DQN algorithm, and outputs the risk assessment report and tiered response plans through a natural language interaction interface.

[0026] Figure 3This is a schematic diagram of the structure of a large-scale enterprise data intelligent analysis device 300 provided in an embodiment of the present invention. The large-scale enterprise data intelligent analysis device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the large-scale enterprise data intelligent analysis device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the large-scale enterprise data intelligent analysis device 300 to implement the method provided in the above embodiment.

[0027] The enterprise data intelligent analysis device 300 based on a large model may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The structure of the enterprise data intelligent analysis device based on the large model shown does not constitute a limitation on the computer device provided by the present invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0028] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the various steps of the enterprise data intelligent analysis method based on the large model provided in the above embodiments.

[0029] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0030] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent analysis of enterprise data based on a large model, characterized in that: The method includes the following steps: Integrate RPA automated data collection components to build a multi-channel data access channel that is compatible with enterprise financial accounting systems, electronic tax bureaus, invoicing systems, and social security systems, and to automatically collect multi-source heterogeneous data; The collected multi-source heterogeneous data is cleaned, standardized, and correlated to obtain preprocessed data. The preprocessed data, including both structured and unstructured data, is fused across modalities. An attention mechanism is then used for feature extraction to uncover potential relationships between the data and obtain enterprise feature vectors. By inputting enterprise feature vectors into a risk assessment model that integrates knowledge graphs and XGBoost algorithms, the risk points of enterprise operations are identified, and a risk assessment report is generated. Based on the risk assessment report, the DQN algorithm is used to generate a tiered response plan, and the risk assessment report and the tiered response plan are output through a natural language interaction interface.

2. The intelligent enterprise data analysis method based on a large model as described in claim 1, characterized in that, The integrated RPA automated data acquisition component constructs a multi-channel data access channel adapted to enterprise financial accounting systems, electronic tax bureaus, invoicing systems, and social security systems, enabling automated acquisition of multi-source heterogeneous data, including: Pre-build interface feature libraries and page element feature libraries for each business system, and match the access paths, data field formats and permission verifications of the financial accounting system, electronic tax bureau, invoicing system and social security system one by one; Based on the preset collection dimensions, initialize three types of task instructions: timed collection, incremental collection, and triggered collection, determine the range of the dataset to be collected, and define the collection boundaries between structured and unstructured data. Based on RPA automated data collection components, form data, report data, log data, and text attachment data are traversed in layers to capture structured ledger data and unstructured documents, announcements, and notes from various systems, and integrated to obtain multi-source heterogeneous data.

3. The intelligent enterprise data analysis method based on a large model as described in claim 1, characterized in that, The process of cleaning, standardizing, and mapping the collected multi-source heterogeneous data to obtain preprocessed data includes: For multi-source heterogeneous data, hash matching is performed using a unique identifier field to remove duplicate data, and missing values ​​are identified and filled by traversing all fields. The 3σ principle is used to identify and filter multi-source heterogeneous data. The mean and dispersion range of data distribution for each field are calculated. Data that exceeds the preset fluctuation range is identified as abnormal data and is removed. The cleaned data is processed using range normalization to obtain normalized data. Based on the enterprise entity, business documents, and time nodes, the normalized data is mapped to obtain preprocessed data.

4. The intelligent enterprise data analysis method based on a large model as described in claim 1, characterized in that, The cross-modal fusion of structured and unstructured data in the preprocessed data includes: The preprocessed data is categorized according to its data storage attributes to obtain structured and unstructured data. Structured and unstructured data are encoded and converted separately. Based on the business relevance dimension, the encoded structured and unstructured data are matched at the position and semantic level to establish corresponding relationships between the data. By combining and superimposing paired structured and unstructured representation information layer by layer, a cross-modal fusion data set is obtained.

5. The intelligent enterprise data analysis method based on a large model as described in claim 4, characterized in that, The feature extraction using the attention mechanism mines potential correlations between data to obtain enterprise feature vectors, including: Weights are calculated for all feature dimensions in the cross-modal fusion dataset using a multi-head attention mechanism; Based on the calculated attention weights, potential business relationships between different data are mined to obtain target features, and these target features are then aggregated and encoded. The aggregated target features are encoded and transformed to generate enterprise feature vectors.

6. The intelligent enterprise data analysis method based on a large model as described in claim 1, characterized in that, The process of inputting enterprise feature vectors into a risk assessment model that integrates knowledge graphs and the XGBoost algorithm to identify enterprise operational risk points and generate a risk assessment report includes: The enterprise feature vector is input into the risk assessment model, and the enterprise feature vector is mapped and matched with the entity nodes of the pre-built knowledge graph to associate the business scenario, risk dimension and risk entity corresponding to the feature vector; The enterprise feature vectors that have completed the graph association are input into the XGBoost model. The input features are split and iteratively analyzed to distinguish between normal operating features and abnormal risk features, and to identify multi-dimensional risk points and risk types in the enterprise's finance, taxation, compliance, and social security. By combining the risk weights of the knowledge graph with the output of the XGBoost model, the severity of the identified risks is quantified and graded, distinguishing between minor risks, general risks, severe risks, and major risks. The risk level, risk type, and risk location are then integrated to generate a risk assessment report.

7. The intelligent enterprise data analysis method based on a large model as described in claim 1, characterized in that, The risk assessment report is used to generate a tiered response plan using the DQN algorithm. The risk assessment report and the tiered response plan are then output through a natural language interface, including: The risk level, risk type, and risk location in the risk assessment report are used as environmental status characteristics and input into the DQN algorithm. The DQN algorithm uses the current state characteristics to call the main Q network and the target Q network, iteratively calculates the value and benefits of risk management actions, and selects the optimal combination of management actions. Based on the optimal combination of actions and the severity of the risk, corresponding response strategies are matched for different levels of risk, converted into text format, and output as risk assessment reports and tiered response plans through a natural language interaction interface.

8. An intelligent enterprise data analysis system based on a large model, characterized in that: The system includes: The data acquisition module is used to integrate RPA automated acquisition components to build a multi-channel data access channel that is compatible with enterprise financial accounting system, electronic tax bureau, invoicing system and social security system, and to automatically collect multi-source heterogeneous data. The data processing module is used to clean, standardize, and correlate the collected multi-source heterogeneous data to obtain preprocessed data. The feature extraction module is used to perform cross-modal fusion of structured and unstructured data in the preprocessed data, and to extract features using an attention mechanism to mine potential correlations between data and obtain enterprise feature vectors. The model recognition module is used to input enterprise feature vectors into a risk assessment model that integrates knowledge graphs and XGBoost algorithms to identify enterprise operational risk points and generate risk assessment reports. The results output module is used to generate tiered response plans based on the risk assessment report using the DQN algorithm, and outputs the risk assessment report and tiered response plans through a natural language interaction interface.

9. An intelligent enterprise data analysis device based on a large model, characterized in that: The enterprise data intelligent analysis device based on a large model includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the enterprise data intelligent analysis device based on a large model to perform the various steps of the enterprise data intelligent analysis method based on a large model as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the enterprise data intelligent analysis method based on a large model as described in any one of claims 1-7.

Images