A method for aggregating and processing digital and electrical tickets based on direct connection of industry, finance and tax

Abstract

The application provides a method for aggregating and analyzing electronic invoices based on direct connection of industry, finance and tax, comprising: S101, configuring a secure authentication gateway, obtaining business data, financial data and tax data of a target enterprise from a business management system, a financial management system and a tax platform through the secure authentication gateway, and collecting the data into original industry, finance and tax data; S102, storing the original industry, finance and tax data into a data lake, constructing a cross-domain unified data model, and associating and mapping the electronic invoices of the target enterprise with the original industry, finance and tax data through the cross-domain unified data model; S103, monitoring key operations of the electronic invoices initiated by the target enterprise through an intelligent agent, analyzing potential risks of the electronic invoices based on the cross-domain unified data model, combining a rule engine and AI analysis, generating a risk analysis report and outputting. The method combines the rule engine and the artificial intelligence technology, aggregates and analyzes the potential risks of the electronic invoices based on the direct connection of the industry, finance and tax data, and can identify and block the risks in time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital electronic invoice management, and in particular to a digital electronic invoice aggregation analysis processing method based on industry-finance-tax direct connection. BACKGROUND

[0002] With the deepening of the digital upgrade of tax collection and management and the reform of electronic invoice, comprehensive digital electronic invoices have become the new normal for enterprise operation and tax management. The life cycle of digital electronic invoices usually includes issuing, receiving, storing, accounting, checking, and archiving. Under normal circumstances, after completing a transaction in its own business system, an enterprise needs to manually log in to the electronic platform of the tax bureau or use an interface to issue a digital electronic invoice. The recipient of the digital electronic invoice can obtain the information on the invoice or connect through email, SMS, etc., manually download the PDF, OFD, or XML format file of the digital electronic invoice, and manually or through OCR (Optical Character Recognition) technology enter the information on the digital electronic invoice into the financial accounting system for checking, accounting, and archiving.

[0003] The current digital electronic invoice processing process has some significant problems: first, as the business scale of enterprises continues to expand, the number of digital electronic invoices also increases, and the operation of each link of digital electronic invoices highly depends on manual work, occupying a large number of human resources, and the work content is tedious and repetitive, which is prone to errors due to fatigue and other personal factors; second, the enterprise business, financial, and tax systems are independent of each other, the data standards are different, the connection cost is high, the digital electronic invoice and its associated business data, financial data are isolated from each other, forming a data island; third, the verification of digital electronic invoices mainly focuses on authenticity and consistency of basic content elements, lacks association verification with business, financial, and other data, and it is difficult to identify deep risks such as repeated reimbursement and false transactions in real time, and relies mainly on post-auditing, with lagging risk control.

[0004] Therefore, it is necessary to study a technical solution that can associate and aggregate business, financial, and tax data to realize integrated automatic analysis of digital electronic invoices, in order to solve the many defects existing in the prior art. SUMMARY

[0005] The present application aims to provide a digital electronic invoice aggregation analysis processing method based on industry-finance-tax direct connection, which can construct a cross-domain unified data model based on multi-source heterogeneous industry-finance-tax data, and further combine rule engines and artificial intelligence technologies to automatically aggregate and analyze potential deep risks of digital electronic invoices after associating digital electronic invoices, to identify and block risks in a timely manner.

[0006] To achieve the above-mentioned application purposes, the technical solutions provided by the present application are as follows:

[0007] A number and electricity invoice aggregation analysis processing method based on industry finance and tax direct connection, the method comprises:

[0008] S101, configure a security authentication gateway, obtain the business data, financial data and tax data of the target enterprise from the business management system, financial management system and tax platform through the security authentication gateway, and collect the original industry finance and tax data;

[0009] S102, store the original industry finance and tax data into the data lake, construct a cross-domain unified data model, and associate and map the number and electricity invoices of the target enterprise with the original industry finance and tax data through the cross-domain unified data model;

[0010] S103, monitor the key operation of the number and electricity invoices initiated by the target enterprise through the intelligent agent, analyze the potential risks of the number and electricity invoices based on the cross-domain unified data model, combine the rule engine and AI analysis, generate a risk analysis report and output.

[0011] Further, the verification dispatching module and the verification decision module are extracted in the container respectively;

[0012] S202, the verification dispatching module is configured to the local of the target enterprise, and the verification decision module is configured to the local of the business management system, the financial management system and the tax platform respectively;

[0013] S203, the verification dispatching module obtains the private attribute information of the target enterprise, and writes the encrypted private attribute information into the blockchain;

[0014] S204, when the verification decision module receives the data acquisition request of the security authentication gateway, determines the verification strategy, and sends the verification strategy identification information to the security authentication gateway;

[0015] S205, the security authentication gateway generates a verification content acquisition request based on the verification strategy identification information, and sends it to the verification dispatching module;

[0016] S206, the verification dispatching module generates verification content based on the target enterprise private attribute information according to the verification content acquisition request, and dispatches the verification content to the security authentication gateway;

[0017] S207, the security authentication gateway provides the verification content to the verification decision module, the verification decision module verifies the verification content based on the target enterprise private information stored in the blockchain and the verification strategy, and responds to the data acquisition request after the verification is passed.

[0018] Further, the original industry finance and tax data is stored into the data lake, which specifically includes the following operations:

[0019] S301, the security authentication gateway acquires the original industry finance and tax data in real time, generates a lake entry job based on the acquired original industry finance and tax data, and sends the lake entry job to the job manager;

[0020] S302, the job manager converts the lake entry job into multiple lake entry tasks, and assigns the lake entry tasks to different task nodes;

[0021] S303, each task node writes the corresponding original industry-finance-tax data into the data lake according to the assigned lake entry task for storage.

[0022] Further, the job manager converts the lake entry job into multiple lake entry tasks, and assigns the lake entry tasks to different task nodes, which specifically includes the following operations:

[0023] S401, the job manager collects the computing resource usage data of each task node in real time;

[0024] S402, input the computing resource usage data of the task node into the first double-gated recurrent unit for processing, and iteratively update the hidden state of the computing resource usage data;

[0025] S403, convert the output of the last time step of the first double-gated recurrent unit into a prediction result through linear transformation, and the prediction result is used to represent the computing resource usage of the task node in the future period of time;

[0026] S404, input the lake entry task content into the resource prediction model for processing to obtain a computing resource occupation estimation result;

[0027] S405, according to the computing resource occupation estimation result of each lake entry task and the prediction result, assign the lake entry task to different task nodes.

[0028] Further, after step S405, the following operations are performed:

[0029] S501, difference calculation is performed on the actual computing resource usage and the prediction result of the same task node in the same time period to obtain a calculation result, and the number of times that the calculation result is greater than a preset difference threshold is periodically counted, and if the number of times counted in the period is greater than a preset number threshold, the next step is performed;

[0030] S502, create an asynchronous thread, and build a second double-gated recurrent unit in the asynchronous thread, which has the same structure and parameters as the first double-gated recurrent unit;

[0031] S503, use all historical lake entry tasks and historical computing resource usage of the task nodes generated when the asynchronous thread is created as a training set to train the second double-gated recurrent unit, and optimize and update the parameters;

[0032] S504, the asynchronous thread transmits the optimized parameters of the second double-gated recurrent unit to the first double-gated recurrent unit for application.

[0033] Further, a cross-domain unified data model is constructed, and the target enterprise's number and electricity invoice and the original industry and financial tax data are associated and mapped through the cross-domain unified data model, specifically including the following operations:

[0034] S601, input the number and electricity invoice and the original industry and financial tax data into the pre-trained language model respectively to obtain the vector representation of the content;

[0035] S602, the vector representation of the content is subjected to data enhancement through feature fusion to obtain enhanced feature information;

[0036] S603, the enhanced feature information is input into the encoding layer based on the recurrent neural network for encoding to output the encoding feature vector;

[0037] S604, the relationship between different entity labels is modeled through the multi-head self-attention layer;

[0038] S605, the encoding feature vector is decoded under the relationship model constraint of the entity label through the decoding layer to output the entity recognition result;

[0039] S606, the relationship between different entities is extracted, and the association mapping between the number and electricity invoice and the original industry and financial tax data is established according to the relationship between different entities and the entity source.

[0040] Further, step S103 specifically includes the following operations:

[0041] S701, when the agent monitors the key operation of the number and electricity invoice, the key operation type is identified;

[0042] S702, the cross-domain unified data model associated with the number and electricity invoice is called from the data lake;

[0043] S703, the cross-domain unified data model called in the previous step is input into the rule engine, and the rule engine calls the corresponding rule set for processing according to the key operation type, if the rule judgment result is successfully generated, the rule judgment result is output, otherwise the next step is executed;

[0044] S704, the cross-domain unified data model is input into the AI analysis module, and the AI analysis module performs corresponding analysis on the cross-domain unified data model according to the key operation type to output the analysis result;

[0045] S705, the rule engine processing result and the analysis result are input into the decision module, and the decision module analyzes the potential risk of the number and electricity invoice based on the preset decision rule, generates a risk analysis report and outputs.

[0046] Further, before inputting the cross-domain unified data model into the rule engine, the cross-domain unified data model is cut, specifically including the following operations:

[0047] S801, establish a graph model based on each preset rule set of the rule engine and the cross-domain unified data model, the graph model being composed of nodes and edges, and an adjacency matrix being established based on the relationship between the nodes and the edges;

[0048] S802, construct a feature propagation model based on the graph model and the adjacency matrix;

[0049] S803, configure a graph convolutional neural network, the input of which is the graph model and the output of which is the data necessary for rule judgment of the rule set;

[0050] S804, train the graph convolutional neural network through pre-labeled sample data, and match the feature propagation model with the judgment rules of different rule sets to generate training labels;

[0051] S805, construct a loss function to calculate the error between the training result of the graph convolutional neural network and the training labels, and when the error is lower than a preset error threshold, cut the cross-domain unified data model according to the training result.

[0052] Further, after step S103, the following operations are further included:

[0053] S901, record the key operations of the electronic invoice, perform a hash operation on the recorded content to obtain a first hash operation result;

[0054] S902, mark the content of the cross-domain unified data model involved in the key operations of the electronic invoice to obtain marked content, perform a hash operation on the marked content to obtain a second hash operation result;

[0055] S903, store the first hash operation result and the second hash operation result into a blockchain.

[0056] Compared with the prior art, the present application has the following beneficial effects:

[0057] The application provides a method for aggregating and analyzing electronic invoices based on industry, finance and tax, which centrally connects business, financial and tax systems through a unified security authentication gateway to improve information interaction reliability, eliminate data barriers, and store original industry, finance and tax data obtained by the security authentication gateway in a data lake to realize real-time writing and real-time retrieval and analysis of data, and maintain data consistency by building a cross-domain unified data model. The application also monitors key operations of electronic invoices of target enterprises by intelligent agents to realize real-time response. After monitoring key operations of electronic invoices, the application analyzes potential risks of electronic invoices based on a cross-domain unified data model combined with a rule engine and AI analysis, generates a risk analysis report and outputs it, so that the application can realize automatic aggregation and analysis of electronic invoices, liberate financial personnel from repetitive and tedious invoice processing work, improve the utilization of human resources, and on the other hand, realize cross-validation combined with multi-source heterogeneous industry, finance and tax data, shift from post-audit to in-process early warning, and strengthen risk control. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only preferred embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0059] Figure 1 It is a whole process schematic diagram of the method for aggregating and analyzing electronic invoices based on direct connection of industry, finance and tax provided by the embodiment of the application. DETAILED DESCRIPTION

[0060] The principles and characteristics of the application will be described below in combination with the drawings, and the listed embodiments are only used to explain the application, and are not used to limit the scope of the application.

[0061] Referring to Figure 1 The embodiment provides a method for aggregating and analyzing electronic invoices based on direct connection of industry, finance and tax, which comprises the following operations:

[0062] S101, a security authentication gateway is configured, and business data, financial data and tax data of a target enterprise are obtained from a business management system, a financial management system and a tax platform through the security authentication gateway, and are collected as original industry, finance and tax data.

[0063] In the embodiment, the business management system and the financial management system of the enterprise, and the tax platform are connected through the single interface of the security authentication gateway, which is convenient for unified management of collected data, realizes data interaction and reliable verification, and reduces the cost of system construction.

[0064] S102, store the original industry, financial and tax data into the data lake, build a cross-domain unified data model, and associate and map the target enterprise's electronic invoice with the original industry, financial and tax data through the cross-domain unified data model.

[0065] As a flexible and scalable data storage and management system, the data lake can ingest and store raw data from different sources in raw format, and provide maintenance, query processing and data analysis in a dynamic manner with the help of rich metadata. After storing the original industry, financial and tax data into the data lake, the embodiment further builds a cross-domain unified data model based on the original industry, financial and tax data, thereby unifying the original industry, financial and tax data of multiple sources and heterogeneous into the model. Further, the electronic invoice can be associated and mapped with business order information, business contract information, payment information, etc. based on key fields such as order number.

[0066] S103, monitor the key operation of the electronic invoice initiated by the target enterprise through the intelligent agent, analyze the potential risk of the electronic invoice based on the cross-domain unified data model, combine the rule engine and AI analysis, generate a risk analysis report and output.

[0067] For example, the key operation of the electronic invoice includes, but is not limited to, invoicing, aggregation, value-added tax deduction, value-added tax refund, invoice entry, invoice inspection, etc.

[0068] The method provided by the embodiment can automatically collect the original industry, financial and tax data through the security authentication gateway, store the original industry, financial and tax data through the data lake to realize real-time writing and real-time reading of data flow, update data and call analysis. Further, a cross-domain unified data model is built to associate and map the electronic invoice with the original industry, financial and tax data, so as to facilitate subsequent retrieval and analysis of associated data. At the same time, the intelligent agent monitors whether the target enterprise initiates the key operation of the electronic invoice. When the key operation of the electronic invoice is monitored, the potential risk of the electronic invoice is analyzed based on the cross-domain unified data model, combined with the rule engine and AI analysis, thereby realizing automatic triggering of the analysis process, reducing the investment of human resources and cost in repetitive and tedious work, improving the processing efficiency of the electronic invoice, and avoiding errors caused by personal reasons.

[0069] As a possible implementation, step S101 specifically includes the following operations:

[0070] S201, extract the verification dispatch module and the verification decision module from the micro-service container respectively.

[0071] S202, configure the verification dispatch module to the target enterprise locally, and configure the verification decision module to the business management system, the financial management system and the tax platform locally respectively.

[0072] In this embodiment, the verification dispatch module and the verification decision module installation package adapted to different operating system environments are pre-packaged in the micro-service container, and according to the operating system environment of the target enterprise, the business management system, the financial management system and the tax platform, the corresponding installation package is selected and downloaded to the local for installation.

[0073] In S203, the verification dispatch module obtains the private attribute information of the target enterprise, and writes the encrypted private attribute information into the block chain.

[0074] For example, the private attribute information is information capable of uniquely identifying the target enterprise, such as a taxpayer identification number or a social credit code, and the embodiment does not make specific limitations on the content of the private attribute information.

[0075] In S204, when the verification decision module receives the data acquisition request of the security authentication gateway, the verification strategy is determined, and the verification strategy identification information is sent to the security authentication gateway.

[0076] In this embodiment, the verification decision module is built-in with a plurality of verification strategies. Each time the data acquisition request of the security authentication gateway is received, the verification decision module randomly selects one verification strategy from all the verification strategies, and sends the identification information of the verification strategy to the security authentication gateway. It can be understood that the verification strategy identification information is used to uniquely identify the verification strategy.

[0077] For example, the verification strategy can be to verify the encrypted ciphertext obtained by homomorphic encryption of the private attribute information of the target enterprise, or to provide corresponding knowledge proof after blinding the private attribute information with the electronic signature of the target enterprise. Those skilled in the art can formulate different verification strategies according to actual needs, and the embodiment does not make specific limitations.

[0078] In S205, the security authentication gateway generates a verification content acquisition request based on the verification strategy identification information, and sends it to the verification dispatch module.

[0079] In S206, the verification dispatch module generates verification content based on the private attribute information of the target enterprise according to the verification content acquisition request, and dispatches the verification content to the security authentication gateway.

[0080] It can be understood that the security authentication gateway itself does not store the private attribute information of the target enterprise and does not store the verification policy in this embodiment. After receiving the verification policy identification information, the security authentication gateway sends the verification policy identification information to the verification dispatching module, the verification dispatching module identifies the verification policy corresponding to the verification policy identification information, and processes the private attribute information of the target enterprise according to the verification policy to generate verification content, and dispatches the verification content to the security authentication gateway.

[0081] S207, the security authentication gateway provides the verification content to the verification decision module, the verification decision module verifies the verification content based on the target enterprise private information stored in the block chain and the verification policy, and responds to the data acquisition request after the verification is passed.

[0082] In this embodiment, the verification dispatching module and the verification decision module are deployed in the form of micro-service architecture on the industry, finance and tax corresponding systems / platforms in the target enterprise. On the one hand, it can provide adaptive installation package for different operating system environments, realize lightweight deployment, and on the other hand, the function of micro-service module is easy to extend and update, which is convenient for subsequent maintenance and upgrade. In this embodiment, the industry, finance and tax corresponding systems / platforms need to verify the identity of the data requester before providing the original industry, finance and tax data. In the verification process, the security authentication gateway does not need to directly obtain the original text content of the private attribute information of the target enterprise throughout the process, and only plays a data transfer role between the verification dispatching module and the verification decision module, and transmits the verification content obtained based on the processing of the private attribute information, which can effectively protect the private attribute information, and at the same time, when the security authentication gateway interfaces with a large number of enterprises, it can also reduce the data storage pressure.

[0083] As another possible embodiment, the original industry, finance and tax data is stored in the data lake, which includes the following operations:

[0084] S301, the security authentication gateway acquires the original industry, finance and tax data in real time, generates a lake entry job based on the acquired original industry, finance and tax data, and sends the lake entry job to the job manager.

[0085] S302, the job manager converts the lake entry job into a plurality of lake entry tasks, and distributes the lake entry tasks to different task nodes.

[0086] S303, each task node writes the corresponding original industry, finance and tax data into the data lake for storage according to the assigned lake entry task.

[0087] When multiple enterprises are docked or the business scale of an enterprise is large, the amount of original industry-finance-tax data is also very large. In order to store these data in real time for subsequent retrieval and use, in this implementation, after obtaining the original industry-finance-tax data, a corresponding lake entry job is generated, and the lake entry job is converted into multiple lake entry tasks by the job manager. Different task nodes write the corresponding part of the original industry-finance-tax data into the data lake for storage according to the lake entry task assigned to them, thereby improving the data lake entry efficiency.

[0088] As a further possible implementation, in step S302, the job manager converts the lake entry job into multiple lake entry tasks, and assigns the lake entry tasks to different task nodes, specifically including the following operations:

[0089] S401, the job manager collects the computing resource usage data of each task node in real time.

[0090] Illustratively, the computing resource usage data includes, but is not limited to, CPU usage, memory usage, CPU core number, memory total capacity, etc.

[0091] S402, input the computing resource usage data of the task node into the first double gated recurrent unit for processing, and iteratively update the hidden state of the computing resource usage data.

[0092] In this implementation, the first double gated recurrent unit includes a front and a back two gated recurrent units, wherein the first gated recurrent unit is responsible for receiving the computing resource usage data of the current time step and the hidden state output of the first double gated recurrent unit at the last time step, and the second gated recurrent unit receives the hidden state output by the first gated recurrent unit and the hidden state output by itself at the last time step, so that the first double gated recurrent unit can continuously capture the local and global features of the computing resource usage data, and improve its modeling ability for the computing resource usage data with time sequence characteristics.

[0093] S403, convert the output of the last time step of the first double gated recurrent unit into a prediction result through linear transformation, and the prediction result is used to represent the computing resource usage of the task node in the future period of time.

[0094] S404, input the lake entry task content into the resource prediction model for processing to obtain a computing resource occupation estimation result.

[0095] In this implementation, the resource prediction model is a neural network model trained in advance using annotated sample lake entry task content and sample computing resource occupation data. Illustratively, the neural network model can be a recurrent neural network model or a Transformer model, and this implementation does not make specific limitations.

[0096] S405, according to the calculation resource occupation estimation result of each lake entry task and the prediction result, the lake entry task is distributed to different task nodes.

[0097] In this embodiment, the task node can be used to process other tasks in addition to processing the data lake entry task. Therefore, this embodiment first predicts the calculation resource usage of each task node in a future period of time through the first double-gated recurrent unit, then estimates the calculation resource required by the lake entry task according to the content of the lake entry task, and finally combines the prediction result and the estimation result to reasonably distribute the lake entry task to each task node, so as to ensure that the lake entry task can be completed quickly and the calculation resource of the task node can be used efficiently. For example, based on the prediction result and the estimation result, the distribution of the lake entry task can be realized through the whale algorithm, the genetic algorithm, or the optimized particle swarm algorithm, which is not limited in this embodiment.

[0098] As a further possible embodiment, after step S405, the following operations are performed:

[0099] S501, the actual calculation resource usage of the same task node in the same period of time is differentially calculated with the prediction result to obtain a calculation result, the number of times that the calculation result is greater than a preset difference threshold value is periodically counted, and if the number of times counted in a period is greater than a preset number threshold value, the next step is performed.

[0100] In this step, when calculating the difference between the actual calculation resource usage and the prediction result, the calculation resource occupation generated by the task node when executing the lake entry task needs to be considered.

[0101] S502, an asynchronous thread is created, and a second double-gated recurrent unit with the same structure and parameters as the first double-gated recurrent unit is built in the asynchronous thread.

[0102] S503, all historical lake entry tasks and historical calculation resource usage of the task node generated when the asynchronous thread is created are taken as a training set to train the second double-gated recurrent unit, and the parameters are optimized and updated.

[0103] S504, the asynchronous thread transmits the optimized parameters of the second double-gated recurrent unit to the first double-gated recurrent unit for application.

[0104] In this embodiment, when the difference between the actual computing resource usage of the task node and the prediction result continues to have a large deviation, the parameters of the first double-gated recurrent unit are updated and optimized in time to strengthen the adaptability of the model to the task demand. Meanwhile, through the second double-gated recurrent unit in the asynchronous thread, the parameters are iteratively optimized, and then synchronized to the first double-gated recurrent unit, which can avoid affecting the computing resource prediction process.

[0105] As another possible implementation, a cross-domain unified data model is constructed, and the electronic invoice of the target enterprise and the original industry and finance and tax data are associated and mapped through the cross-domain unified data model, specifically including the following operations:

[0106] S601, input the electronic invoice and the original industry and finance and tax data into a pre-trained language model respectively to obtain vector representations of the contents. The contents include the contents of the electronic invoice and the original industry and finance and tax data.

[0107] For example, the pre-trained language model can use a BERT model, or other language models, which are not limited in this embodiment.

[0108] S602, the vector representations of the contents are subjected to data enhancement through feature fusion to obtain enhanced feature information.

[0109] S603, the enhanced feature information is input into an encoding layer based on a recurrent neural network for encoding to output an encoded feature vector.

[0110] S604, the relationship between different entity labels is modeled through a multi-head self-attention layer.

[0111] S605, the encoded feature vector is decoded under the constraint of the relationship model of the entity labels through a decoding layer to output an entity recognition result.

[0112] S606, the relationship between different entities is extracted, and the association mapping between the electronic invoice and the original industry and finance and tax data is established according to the relationship between different entities and the source of the entities.

[0113] This embodiment vectorizes the contents of the electronic invoice and the original industry and finance and tax data through a pre-trained language model, then realizes data enhancement through feature fusion of the vector representations of the contents, inputs the enhanced feature information into an encoding layer for encoding to obtain an encoded feature vector, models the relationship between different entity labels through a multi-head self-attention layer to enhance the attention to the association of the entity labels, and finally decodes the encoded feature vector under the constraint of the relationship model of the entity labels through a decoding layer. This embodiment can improve the recognition performance of the model for entities, so that when constructing a cross-domain unified data model, the associated entities in the electronic invoice and the original industry and finance and tax data can be accurately mapped.

[0114] As yet another possible implementation, step S103 specifically includes the following operations:

[0115] S701, the intelligent agent monitors the key operation of the electronic invoice, and identifies the type of the key operation.

[0116] S702, the cross-domain unified data model associated with the electronic invoice is retrieved from the data lake.

[0117] S703, the cross-domain unified data model retrieved in the previous step is input into the rule engine, and the rule engine retrieves the corresponding rule set for processing according to the type of the key operation. If a rule judgment result is successfully generated, the rule judgment result is output, otherwise the next step is executed.

[0118] In this embodiment, the rule engine has multiple pre-set rule sets built-in, and the rule sets and the types of the key operations can be in a one-to-one or one-to-many relationship. For example, one type of key operation can correspond to multiple rule sets, i.e., when the key operation is performed, it needs to be determined whether the cross-domain unified data model meets all the rule sets corresponding thereto. The same rule set can also be used for multiple key operations.

[0119] For example, the rule set includes, but is not limited to, a syntax rule set, a business logic rule set, and a tax rule set. The syntax rule set can be represented as: if the amount of the electronic invoice is empty, the verification fails. The business logic rule set can be represented as: if the amount of the electronic invoice is not equal to the contract amount or the order amount, the verification fails. The tax rule set can be represented as: if the commodity tax rate is not equal to any pre-set tax rate, the verification fails.

[0120] S704, the cross-domain unified data model is input into the AI analysis module, and the AI analysis module performs corresponding analysis on the cross-domain unified data model according to the type of the key operation, and outputs an analysis result.

[0121] In this embodiment, the AI analysis module is used to handle tasks that are difficult to determine by simple rules. For example, for non-directly connected paper invoice images or contract images, the rule engine cannot directly determine the content thereof, and the AI analysis module can convert them into text through OCR technology and realize information extraction. For diversified electronic invoice commodity names, the AI analysis module can identify them through a pre-trained algorithm model, and match them with tax classification codes to determine the correct tax rate. For potential risks of repeated reimbursement, the AI analysis module can perform similarity matching on different reimbursement invoices, and filter out invoices with a similarity greater than a pre-set threshold or with obvious modification traces.

[0122] S705, input the rule engine processing result and the analysis result into a decision module, the decision module analyzes potential risks of the electronic invoice based on preset decision rules, generates a risk analysis report, and outputs.

[0123] In this embodiment, first, the rule engine performs a quick analysis and judgment on the cross-domain unified data model according to simple business rules. If it is directly determined that the verification fails or there is a problem, the AI analysis module does not need to be further occupied. If the rule engine cannot accurately determine that there is a problem in the data in the cross-domain unified data model, the AI analysis module is called to analyze the problems that the rule group cannot handle, and the analysis result is output. Finally, the decision module combines the processing result and the analysis result of the rule engine to determine whether there is a potential risk in the electronic invoice. For example, when the processing result of the rule engine is that there may be a risk and it cannot be accurately determined, the analysis result of the AI analysis module is that there is no repeated reimbursement, and the commodity and tax classification code match, the decision module generates a risk analysis report based on the processing result and the analysis result, and outputs it to the staff for manual review. When the rule engine determines that the verification fails, the decision module directly generates a risk analysis report according to the reason for the verification failure.

[0124] As a further possible embodiment, before the cross-domain unified data model is input into the rule engine, the cross-domain unified data model is cut, specifically including the following operations:

[0125] S801, based on each preset rule group of the rule engine and the cross-domain unified data model, a graph model is established, the graph model is composed of nodes and edges, and an adjacency matrix is established based on the relationship between the nodes and the edges.

[0126] In this embodiment, the nodes of the graph model include master nodes and slave nodes, wherein the master nodes are rule groups, and the slave nodes are data in the cross-domain unified data model. The edges include edges between the master nodes and the slave nodes, and edges between the master nodes. The edges between the master nodes and the slave nodes are used to represent the data relied on by the rule group when performing rule judgment; the edges between the master nodes are used to represent the dependency relationship between the rule groups, for example, when the judgment of a rule group must rely on the judgment result of another rule group, the edges are connected between the corresponding master nodes of the two.

[0127] In this step, the value of the adjacency matrix of the nodes and the edges is taken according to whether there is a connection between the ith node and the jth edge, and the value range is (0, 1).

[0128] S802, based on the graph model and the adjacency matrix, a feature propagation model is constructed.

[0129] S803, configure a graph convolutional neural network, the input is the graph model, and the output is the data required for the rule group to perform rule judgment.

[0130] S804, train the graph convolutional neural network through the pre-labeled sample data, and match the feature propagation model with the judgment rules of different rule groups to generate training labels.

[0131] S805, construct a loss function to calculate the error between the training result of the graph convolutional neural network and the training label, and when the error is lower than a preset error threshold, cut the cross-domain unified data model according to the training result.

[0132] In this embodiment, the content of the cut cross-domain unified data model is the data content necessary for rule judgment, and the data amount is reduced compared with the original cross-domain unified data model.

[0133] In the foregoing embodiment, when the rule engine performs rule judgment, the cross-domain unified data model needs to be input into the rule engine to provide data support. When the scale of the cross-domain unified data model reaches a certain order of magnitude, the data input efficiency and the processing efficiency of the rule engine will decrease. In this embodiment, the graph convolutional neural network is constructed based on the cross-domain unified data model and the rule group of the rule engine, the data demand necessary for rule judgment of the rule group of the rule engine is predicted, the cross-domain unified data model is cut based on the spatial relationship between each rule group and the industry, finance, tax and invoice data, so as to effectively control the data amount input into the rule engine, retain the effective part, and further improve the data input efficiency and the processing efficiency of the rule engine.

[0134] As another possible implementation, after step S103, the following operations are further included:

[0135] S901, record the key operation of the numerical electric invoice, perform hash operation on the record content to obtain a first hash operation result.

[0136] S902, mark the content of the cross-domain unified data model involved in the key operation of the numerical electric invoice to obtain a marked content, perform hash operation on the marked content to obtain a second hash operation result.

[0137] S903, store the first hash operation result and the second hash operation result into the blockchain.

[0138] In this embodiment, the content of the key operation of the numerical electric invoice and the data model content involved are recorded and stored, so as to facilitate subsequent audit tracing. Meanwhile, in this embodiment, the content of the key operation of the numerical electric invoice and the data model content involved are first subjected to hash operation, and then the hash operation results are stored on the chain, so as to realize a lightweight on-chain strategy, which can balance the data transmission efficiency and data credibility.

[0139] The above merely describes preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for aggregating and analyzing electronic invoices based on direct connection between industry, finance, and taxation, characterized in that, The method includes: S101. Configure a security authentication gateway to obtain the target company's business data, financial data and tax data from the business management system, financial management system and tax platform respectively through the security authentication gateway, and aggregate them into raw business, financial and tax data. S102. Store the original business, financial and tax data in the data lake, construct a cross-domain unified data model, and use the cross-domain unified data model to associate and map the target enterprise's digital invoices with the original business, financial and tax data. S103. Monitor key operations of digital invoices initiated by target enterprises through intelligent agents, analyze potential risks of digital invoices based on cross-domain unified data models, and combine rule engines and AI analysis to generate and output risk analysis reports. Step S101 specifically includes the following operations: S201. Extract the verification dispatch module and verification decision module from the microservice container respectively; S202. Configure the verification and dispatch module to the target enterprise's local machine, and configure the verification decision module to the business management system, financial management system, and tax platform respectively. S203. The verification and distribution module obtains the target enterprise's private attribute information, encrypts the private attribute information, and writes it into the blockchain. S204. When the verification decision module receives a data acquisition request from the security authentication gateway, it determines the verification strategy and sends the verification strategy identification information to the security authentication gateway. S205. The security authentication gateway generates a verification content retrieval request based on the verification policy identification information and sends it to the verification dispatch module. S206. The verification and dispatch module generates verification content based on the target enterprise's private attribute information according to the verification content acquisition request, and dispatches the verification content to the security authentication gateway. S207. The security authentication gateway provides verification content to the verification decision module. The verification decision module verifies the verification content based on the target enterprise's private information and verification strategy stored in the blockchain. After the verification is passed, it responds to the data acquisition request. Construct a cross-domain unified data model to associate and map the target enterprise's electronic invoices with its original business, financial, and tax data. This includes the following operations: S601. Input the digital invoices and original business and tax data into the pre-trained language model to obtain the vector representation of the content; S602. Enhance the vector representation of the content through feature fusion to obtain enhanced feature information; S603. Input the enhanced feature information into the encoding layer based on the recurrent neural network for encoding, and output the encoded feature vector; S604. Model the relationship between different entity labels through a multi-head self-attention layer; S605. The decoder decodes the encoded feature vector under the constraint of the relational model of the entity label and outputs the entity recognition result. S606. Extract the relationships between different entities and establish a mapping between power transmission invoices and original business and tax data based on the relationships between different entities and the source of the entities. Step S103 specifically includes the following operations: S701: When the intelligent agent detects key operations on digital invoices, it identifies the type of key operation. S702. Retrieve the cross-domain unified data model associated with electronic invoices from the data lake; S703. Input the cross-domain unified data model retrieved in the previous step into the rule engine. The rule engine retrieves the corresponding rule group for processing according to the key operation type. If the rule judgment result is successfully generated, the rule judgment result is output; otherwise, proceed to the next step. S704. Input the cross-domain unified data model into the AI ​​analysis module. The AI ​​analysis module performs corresponding analysis on the cross-domain unified data model according to the key operation types and outputs the analysis results. S705. Input the processing and analysis results of the rule engine into the decision module. The decision module analyzes the potential risks of digital invoices based on preset decision rules, generates a risk analysis report, and outputs it.

2. The method for aggregation and analysis of electronic invoices based on direct connection between industry, finance, and taxation as described in claim 1, characterized in that, The original business, financial, and tax data is stored in the data lake, which specifically includes the following operations: S301, the security authentication gateway acquires raw business, financial and tax data in real time, generates lake entry operations based on the acquired raw business, financial and tax data, and sends the lake entry operations to the operation manager; S302, The job manager converts the lake entry job into multiple lake entry tasks and assigns the lake entry tasks to different task nodes; S303. Each task node writes the corresponding original business, financial and tax data into the data lake for storage according to the assigned data entry task.

3. The method for aggregation and analysis of electronic invoices based on direct connection between industry, finance, and taxation, as described in claim 2, is characterized in that... The job manager converts the lake entry job into multiple lake entry tasks, and assigns these tasks to different task nodes. Specifically, this includes the following operations: S401, the job manager collects computing resource usage data for each task node in real time; S402. Input the computing resource usage data of the task node into the first dual-gated loop unit for processing, and iteratively update the hidden state of the computing resource usage data. S403. The output of the last time step of the first dual-gated loop unit is converted into a prediction result through a linear transformation. The prediction result is used to characterize the computing resource usage of the task node in the future. S404. Input the content of the lake entry task into the resource prediction model for processing to obtain the estimated results of resource occupancy. S405. Based on the estimated computational resource consumption of each lake entry task and the prediction results, the lake entry tasks are assigned to different task nodes.

4. The method for aggregation and analysis of electronic invoices based on direct connection between industry, finance, and taxation, as described in claim 3, is characterized in that... After step S405, perform the following operations: S501. Calculate the difference between the actual computing resource usage of the same task node and the prediction result in the same time period, and obtain the calculation result. Periodically count the number of times the calculation result is greater than the preset difference threshold. If the number of times the calculation result is greater than the preset number threshold within the period, proceed to the next step. S502. Create an asynchronous thread, and build a second double-gated loop unit in the asynchronous thread with the same structure and parameters as the first double-gated loop unit. S503. Use all historical lake entry tasks and historical computing resource usage of task nodes generated when the asynchronous thread is created as the training set to train the second dual-gated loop unit, optimize and update the parameters. S504, the asynchronous thread transmits the optimized parameters of the second dual-gated loop unit to the first dual-gated loop unit for application.

5. The method for aggregation and analysis of electronic invoices based on direct connection between industry, finance, and taxation, as described in claim 1, is characterized in that... Before inputting the cross-domain unified data model into the rule engine, the cross-domain unified data model is trimmed, specifically including the following operations: S801. A graph model is established based on the various preset rule groups of the rule engine and the cross-domain unified data model. The graph model consists of nodes and edges, and an adjacency matrix is ​​established based on the relationship between nodes and edges. S802. Constructing a feature propagation model based on graph models and adjacency matrices; S803. Configure a graph convolutional neural network with a graph model as input and the data required for rule judgment of the rule group as output. S804. Train the graph convolutional neural network using pre-labeled sample data, and simultaneously match the feature propagation model with the judgment rules of different rule groups to generate training labels. S805. Construct a loss function to calculate the error between the training results and training labels of the graph convolutional upgrade network. When the error is lower than the preset error threshold, prune the cross-domain unified data model according to the training results.

6. The method for aggregation and analysis of electronic invoices based on direct connection between industry, finance, and taxation, as described in claim 1, is characterized in that... Following step S103, the following operations are also included: S901. Record key operations of digital electronic invoices, perform hash operations on the recorded content, and obtain the first hash operation result; S902. Mark the cross-domain unified data model content involved in the key operations of the digital electronic invoice, obtain the marked content, perform a hash operation on the marked content, and obtain the second hash operation result; S903. Store the results of the first hash operation and the second hash operation into the blockchain.

Images