Intelligent accounting auditing system based on AI and block chain

Abstract

The invention discloses an intelligent accounting auditing system based on AI and a block chain, and particularly relates to the technical field of accounting auditing and computer application. The system comprises a data acquisition module, a data processing module, an AI auditing module and a block chain storage module, wherein the data acquisition module acquires original accounting data from an enterprise internal information system through a RESTful API interface and an ETL technology; the data processing module cleans the original accounting data by using a large language model and an unsupervised learning algorithm, and generates audit data in combination with a machine learning classification algorithm and an accounting data standardization rule base; the AI auditing module constructs an AI auditing model based on a graph neural network in combination with reinforcement learning, and processes an auditing graph constructed through the auditing data through the AI auditing model to generate an auditing result; and the block chain storage module adopts AES-256 to encrypt and store the original accounting data, the audit data and the audit result. According to the invention, intelligent auditing and safe storage of the accounting data are realized, and auditing efficiency and data credibility are improved.

Description

Technical Field

[0001] The present invention specifically relates to the fields of accounting, auditing and computer application technology, and more specifically to an intelligent accounting and auditing system based on AI and blockchain. Background Art

[0002] As enterprises accelerate their digitalization, the shortcomings of traditional accounting and auditing systems in data processing and secure storage are becoming increasingly apparent. Data collection relies on manual labor and lacks standardized interfaces, resulting in low accounting data processing efficiency and high error rates. Audit models are often based on simple algorithms, struggling to handle complex business relationships and lacking risk identification capabilities. Centralized storage methods carry the risk of data tampering and leakage, failing to meet regulatory traceability requirements. Furthermore, the lack of integration between general standards and enterprise-specific needs makes cross-industry audits difficult.

[0003] While the application of artificial intelligence and blockchain technology in the financial sector has been increasingly widespread in recent years, the integration of these technologies into accounting and auditing scenarios is insufficient, and there is a lack of intelligent solutions for the entire process, making it difficult to effectively improve audit efficiency and data credibility. To address this issue, this paper proposes an intelligent accounting and auditing system based on AI and blockchain to address these issues. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent accounting and auditing system based on AI and blockchain to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] An intelligent accounting and auditing system based on AI and blockchain, including a data acquisition module, a data processing module, an AI auditing module and a blockchain storage module;

[0007] Data collection module, used to obtain original accounting data from the enterprise's internal information system;

[0008] The data processing module is used to clean the original accounting data and standardize the cleaned original accounting data to obtain audit data;

[0009] The AI ​​audit module is used to construct an audit graph based on audit data and process the audit graph through the AI ​​audit model to obtain audit results;

[0010] The blockchain storage module is used to store original accounting data, audit data and audit results in the blockchain.

[0011] Preferably, the data collection module comprises a data collection interface adopting a standardized data interface protocol RESTful API, which realizes cross-platform communication connection with the enterprise internal information system through the HTTP protocol, supports standard interface methods including GET, POST, PUT and DELETE, and can adapt to heterogeneous data interaction requirements of ERP systems, financial systems and supply chain management systems.

[0012] The data collection interface extracts original accounting data from the database of the enterprise internal information system according to a preset time period and rules through ETL technology, including structured fields and unstructured fields, wherein the preset time period is a fixed time point per day or transaction real-time triggering, and the preset rule is a data update mark or an incremental synchronization threshold.

[0013] The ETL technology is a data integration technology.

[0014] Preferably, the method for data cleaning processing of the original accounting data comprises the following steps:

[0015] The original text data is automatically divided into text type accounting data and numerical type accounting data by combining regular expression matching and field attribute annotation. For mixed data, the numerical part is extracted by named entity recognition technology and added to the numerical type accounting data, and the remaining text part is classified into text type data.

[0016] For ambiguous data in the text type data, the LLM generates multiple version correction suggestions through context semantic association, which is automatically replaced after being confirmed by the auditor. For fuzzy data, the LLM trains a numerical mapping model based on similar business data of the enterprise history, and converts the fuzzy data into accurate numerical values through the numerical mapping model. The large language model LLM is GPT-4o or BERT-wwm.

[0017] An unsupervised learning model fused with isolated forest and local anomaly factor is used to construct a numerical distribution model. The numerical type accounting data is input into the numerical type data distribution model. The isolated forest divides the numerical space through random binary tree, calculates the path length of continuous data in the numerical type accounting data, and marks the data with a path length less than 2 times the average standard deviation as abnormal data. The local anomaly factor calculates the local reachable density of multi-dimensional data in the numerical type accounting data. When the density of a numerical point is less than 30% of the neighbor points, it is determined as abnormal data. Accordingly, the abnormal data is corrected through the large language model LLM to obtain the original accounting data after data cleaning processing.

[0018] Preferably, the method for standardizing the original accounting data after data cleaning processing comprises the following steps:

[0019] The preset accounting data standardization rule library is stored in a distributed storage architecture, integrates international general accounting standards, national audit standards, and customized rules based on the characteristics of finance, manufacturing, and service industries and enterprise scale, and builds a rule association graph through a semantic network to realize dynamic retrieval.

[0020] The cleaned original accounting data is classified by a machine learning classification algorithm supporting a support vector machine or a random forest algorithm. For the support vector machine algorithm, a radial basis kernel function RBF is selected, and a penalty parameter C and a kernel function parameter γ are optimized by grid search. For the random forest, the number of decision trees is set to 50 to 100, and each tree randomly selects one-third of the feature dimension for splitting. The cleaned original accounting data is divided into eleven basic categories including accounting subject categories, transaction record categories, business process categories, enterprise department categories, personnel categories, fund flow categories, business logic categories, subject association categories, transaction association categories, department business association categories, and personnel business association categories by the machine learning classification algorithm. After classification, the original accounting data is converted into uniform audit data in an XBRL standard format by an ETL tool through a rule engine matching mechanism, and the standardized mapping from the original accounting data to the audit data is realized.

[0021] Preferably, the audit data includes accounting subject data, transaction record data, business process data, enterprise department data, personnel data, fund flow data, business logic data, subject association data, transaction association data, department business association data, and personnel business association data. The accounting subject data includes subject number, subject name, subject category, initial and final balance, and current period occurrence. The transaction record data includes transaction number, transaction date, transaction type, transaction amount, associated subject, transaction party information, and transaction voucher number. The business process data includes business process number, business process name, business process stage, responsible person name, and start and end time. The enterprise department data includes enterprise department number, department name, and enterprise department function. The personnel data includes employee number, employee name, employee position, and employee authority. The fund flow data includes fund flow between accounting subjects, fund flow between transaction records, amount, and time. The business logic data includes business relationship between business processes and transaction records and business relationship between business processes and transaction records. The transaction association data includes transaction relationship between transaction records. The department business association data includes functional relationship between enterprise departments and business processes and functional relationship between enterprise departments and transaction records. The personnel business association data includes participation relationship between personnel and business processes and participation relationship between personnel and transaction records.

[0022] Preferably, the process of constructing the audit graph includes:

[0023] extracting each accounting subject in the accounting subject data, taking each accounting subject as an accounting subject node, and setting corresponding accounting subject attributes, wherein the accounting subject attributes include subject number, subject name, subject category, initial and final balance, and this period's occurrence;

[0024] extracting each transaction record in the transaction record data, taking each transaction as a transaction record node, and setting corresponding transaction record attributes, wherein the transaction record attributes include transaction number, transaction date, transaction type, transaction amount, associated subject, transaction party information, and transaction voucher number;

[0025] extracting each business process in the business process data, taking each business process as a business process node, and setting corresponding business process attributes, wherein the business process attributes include business process number, business process name, business process stage, responsible person's name, and start and end time;

[0026] extracting each enterprise department in the enterprise department data, taking each enterprise department as an enterprise department node, and setting corresponding enterprise department attributes, wherein the enterprise department attributes include enterprise department number, department name, and enterprise department function;

[0027] extracting each person in the personnel data, taking each person as a personnel node, and setting corresponding personnel attributes, wherein the personnel attributes include employee number, employee name, employee position, and employee authority;

[0028] According to the fund flow relationship between each accounting subject and each transaction record in the fund flow data, the business relationship between each business process and each transaction record and each accounting subject in the business logic data, the transaction relationship between each transaction record in the transaction association data, the functional relationship of each enterprise department to each business process and each transaction record in the department business association data, and the participation relationship of each person to each business process and each transaction record in the personnel business association data, the accounting subject node, the transaction record node, the business process node, the enterprise department node, and the personnel node are connected through directed edges to obtain an audit graph.

[0029] The directed edges include fund flow edges, business logic edges, subject association edges, transaction association edges, department business edges, and personnel business edges, which are obtained based on the fund flow data, the business logic data, the subject association data, the transaction association data, the department business association data, and the personnel business association data.

[0030] Preferably, the AI audit model is constructed by a graph neural network combined with a reinforcement learning mechanism, wherein the graph neural network is a graph convolution network or a graph attention network;

[0031] The graph convolution network adopts a two-layer GCN layer architecture, and the message passing mechanism of each layer is:

[0032] ;

[0033] in, Nodes of the audit graph In the The feature vector of the layer, Nodes of the audit graph The first-order neighbor set of and is the trainable weight matrix, is the ReLU activation function, is a node in the audit graph In the The feature vector of the layer;

[0034] The graph attention network adopts an 8-head attention mechanism, and the single-head attention calculation method is:

[0035] ;

[0036] in, Nodes of the audit graph For neighbor nodes The attention weight, is the attention parameter vector, is the feature transformation matrix, For vector splicing, Nodes of the audit graph The first-order neighbor set of and is the node feature vector;

[0037] The audit graph is input into the AI ​​audit model, and the graph convolutional network in the AI ​​audit model performs topological analysis on the audit graph to obtain a set of node feature vectors. and the global eigenvector of the graph Audit status , based on the audit status Potential risk points are obtained through local anomaly detection. Local anomaly detection calculates the Euclidean distance between the node feature vector and its neighboring nodes. When the distance exceeds 2.5 standard deviations of the average distance, it is marked as a local anomaly node, i.e., a potential risk point.

[0038] The reinforcement learning mechanism is based on the audit status. - the greedy strategy selects a suitable audit strategy from a pre-set audit strategy set, and evaluates the potential risk points through the audit strategy to obtain an audit result including problem description, problem nature, involved amount, risk level and audit suggestion, wherein the pre-set audit strategy set includes a risk verification strategy, a scope expansion strategy and a depth analysis strategy;

[0039] The The parameterization of the greedy strategy is:

[0040] ;

[0041] wherein, and , is an action value function output by the deep Q network, is an audit strategy selected in the tth iteration.

[0042] Preferably, the original accounting data, the audit data and the audit result are encrypted by an AES-256 symmetric encryption algorithm and stored in a blockchain.

[0043] Thanks to the above technical solutions, the present application has the following technical progress compared with the prior art:

[0044] 1. The present application adopts a RESTful API standardized interface and ETL technology to realize automatic data extraction from heterogeneous systems such as enterprise ERP and finance, improve collection efficiency, and combine a large language model with an unsupervised learning algorithm cleaning mechanism to accurately identify text ambiguity data and numerical anomaly data, and through a standardized rule library integrating international accounting standards and industry customized rules, complete data classification and XBRL format conversion based on a support vector machine or a random forest algorithm, solving the problems of non-uniform data format and low cleaning efficiency of traditional systems.

[0045] 2. The present application constructs an audit graph through the processed data, performs topological analysis on the audit graph using GCN or GAT, and combines a reinforcement learning - the greedy strategy decision mechanism can dynamically identify potential risk points such as fund flow abnormality and business logic conflict, and greatly improves the accuracy of the traditional rule engine model audit. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0047] Figure 1A schematic diagram of a system functional module of the present application. DETAILED DESCRIPTION

[0048] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0049] Embodiments, such as Figure 1 The AI and blockchain-based intelligent accounting auditing system, as described above, includes a data acquisition module, a data processing module, an AI auditing module and a blockchain storage module, which cooperatively complete the accounting auditing task.

[0050] The data acquisition module is configured to acquire original accounting data from an enterprise internal information system.

[0051] The data processing module is configured to perform data cleaning processing on the original accounting data, and perform standardization processing on the original accounting data after the data cleaning processing, to obtain auditing data.

[0052] The AI auditing module is configured to construct an audit graph based on the auditing data, and process the audit graph through an AI auditing model to obtain an audit result.

[0053] The blockchain storage module is configured to store the original accounting data, the auditing data and the audit result in a blockchain.

[0054] Further, the working principle of the present application will be illustrated below through embodiments:

[0055] Taking a large manufacturing enterprise as an example, the system architecture thereof includes a data acquisition module, a data processing module, an AI auditing module and a blockchain storage module, each module is deployed in an enterprise private cloud platform through Docker containerization, and the intelligent processing of the whole process of accounting auditing is realized.

[0056] The data collection module establishes a communication connection with the enterprise ERP system, financial system and supply chain management system through the RESTful API interface, and uses the GET method of HTTP protocol to extract raw accounting data at 2 a.m. every day, for example, extracts raw material procurement records from the material management module of the ERP system, the interface request format is GET / api / v1 / erp / purchasing / transactions?date=2025-07-01, and the extracted raw data is cleaned through ETL technology, or the incremental synchronization threshold is set to 500 records, when the newly added transaction data exceeds the threshold, real-time extraction is triggered, and the raw accounting data obtained from the financial system includes structured fields such as subject number and transaction amount and unstructured fields such as invoice abstract text, for example, the raw data of a raw material procurement transaction is: {"transaction_id": "TRX20250701001", "amount":"128,500.00", "description":"purchase A type steel, used for production B product, supplier: XX steel limited company"}.

[0057] The numerical part in the transaction description is extracted from the text type data in the raw accounting data using regular expressions, and "purchase A type steel, quantity 10 tons, unit price 12850 yuan" is divided into text type data "purchase A type steel" and numerical type data "10 tons" "12850 yuan", the BERT-wwm model based on fine-tuning in the accounting field identifies ambiguous data, such as correcting "accounts payable-Company A (cancelled)" to "accounts payable-Company B", and the correction suggestion confidence reaches 96%; The isolated forest and local anomaly factor fusion model is used for numerical type data cleaning, and the raw material procurement amount is detected for anomaly, when the isolated forest path length of a transaction amount is less than 2.3 times the standard deviation of the average level, and the local reachable density is less than 35% of the neighbor point, it is judged as abnormal, and the correction suggestion "may be a batch purchase discount, suggest checking the contract" is generated by GPT-4o; The distributed storage accounting data standardization rule library is called, the manufacturing industry customization rules include "raw material procurement needs to be associated with production order number", the support vector machine classification algorithm is used, the radial basis kernel function is set, the penalty parameter C=100, and the kernel function parameter γ=0.1, the cleaned raw accounting data is divided into "purchase transaction class", and the XBRL standard "raw material procurement" template is matched through the rule engine to generate standardized audit data.

[0058] Nodes and edges are extracted from standardized audit data. The "Raw Materials Purchase" transaction record is used as the transaction node, with attributes including the transaction amount of 128,500 yuan and the associated account "Raw Materials - Steel A". The production department is used as the department node, with attributes including the department number "D001" and the function "Product Production". Based on the capital flow data, an expenditure directed edge is established between the bank deposit account node and the raw materials purchase transaction node. Based on the business logic data, an execution directed edge is established between the production department node and the raw materials purchase transaction node. The audit graph is obtained and processed using a two-layer graph convolutional network (GCN). The first-layer message passing formula is: , where the node feature vector It includes attributes such as transaction amount and account balance. After calculation through the ReLU activation function, potential risk points are obtained, such as the transaction amount of a supplier exceeds the annual procurement quota by 20%.

[0059] set up - Greedy strategy initial parameters is 0.1, press Attenuation, when an abnormal supplier transaction is detected, the reinforcement learning mechanism selects the "deep analysis strategy" with an 80% probability, calls historical transaction data for comparison, and generates an audit result including the problem description that the transaction amount of supplier XX Steel Co., Ltd. in July 2025 exceeded the annual quota by 22%, the nature of the problem is the abnormal procurement amount, the amount involved is 128,500, the risk level is medium risk, and the audit recommendation is to verify the procurement contract and approval process.

[0060] The original accounting data, standardized audit data, and audit results are encrypted with AES-256 and stored in the enterprise alliance chain, which is built on Hyperledger Fabric. For example, the above-mentioned procurement transaction data is encrypted and uploaded to the chain. The block header contains the hash value of the previous block, the timestamp (2025-07-01T03:15:22Z), and the Merkle root hash to ensure that the data cannot be tampered with. The blockchain browser can query the data traceability path, such as the full process record of "original data → cleaned data → audit results".

[0061] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An intelligent accounting and auditing system based on AI and blockchain, characterized by: include: Data collection module, used to obtain original accounting data from the enterprise's internal information system; The data processing module is used to clean the original accounting data and standardize the cleaned original accounting data to obtain audit data; The AI ​​audit module is used to construct an audit graph based on audit data and process the audit graph through the AI ​​audit model to obtain audit results; The blockchain storage module is used to store original accounting data, audit data and audit results in the blockchain.

2. The intelligent accounting and auditing system based on AI and blockchain according to claim 1 is characterized in that: The data acquisition module includes a data acquisition interface using a standardized data interface protocol RESTful API for communicating with the enterprise's internal information systems including ERP systems, financial systems, and supply chain management systems; The data acquisition interface extracts original accounting data from the enterprise's internal information system according to preset rules and time periods through ETL technology.

3. The intelligent accounting and auditing system based on AI and blockchain according to claim 2 is characterized in that: The method for performing data cleaning processing on original accounting data: The original text data is divided into text-based accounting data and numerical accounting data. The text-based accounting data is semantically understood through the large language model (LLM), and ambiguous and fuzzy data are identified and corrected. At the same time, a numerical data distribution model is constructed through an unsupervised learning algorithm. The numerical accounting data is input into the numerical data distribution model, abnormal data is identified, and the abnormal data is corrected through the large language model (LLM), thereby obtaining the original accounting data after data cleaning.

4. The intelligent accounting and auditing system based on AI and blockchain according to claim 3 is characterized in that: The method for standardizing the original accounting data after data cleaning: Calling a preset accounting data standardization rule base that integrates generally accepted accounting principles, auditing standards, and industry and enterprise-scale customized rules; A machine learning classification algorithm is used to classify the original accounting data after data cleaning and match the corresponding standardization rules in the accounting data standardization rule library. Based on the corresponding standardization rules, the classified original accounting data after data cleaning is converted into a unified audit format to obtain audit data. The machine learning classification algorithm is a support vector machine or a random forest.

5. According to claim 4, an intelligent accounting and auditing system based on AI and blockchain is characterized in that: The audit data includes accounting subject data, transaction record data, business process data, enterprise department data, personnel data, capital flow data, business logic data, subject association data, transaction association data, department business association data and personnel business association data.

6. The intelligent accounting and auditing system based on AI and blockchain according to claim 5 is characterized in that: The process of constructing the audit graph: Extract each accounting subject from the accounting subject data, treat each accounting subject as an accounting subject node, and set the corresponding accounting subject attributes; Extract each transaction record from the transaction record data, treat each transaction as a transaction record node, and set the corresponding transaction record attributes; Extract each business process from the business process data, treat each business process as a business process node, and set the corresponding business process attributes; Extract each enterprise department from the enterprise department data, treat each enterprise department as an enterprise department node, and set the corresponding enterprise department attributes; Extract each person in the personnel data, treat each person as a person node, and set the corresponding person attributes; Based on this, according to the capital flow relationship between each accounting account and each transaction record in the capital flow data, the business relationship between each business process and each transaction record and each accounting account in the business logic data, the transaction relationship between each transaction record in the transaction association data, the functional relationship between each enterprise department to each business process and each transaction record in the department business association data, and the participation relationship between each personnel to each business process and each transaction record in the personnel business association data, the accounting account nodes, transaction record nodes, business process nodes, enterprise department nodes and personnel nodes are connected through directed edges to obtain an audit graph.

7. The intelligent accounting and auditing system based on AI and blockchain according to claim 6 is characterized in that: The AI ​​audit model is constructed by a graph neural network combined with a reinforcement learning mechanism, where the graph neural network is a graph convolutional network or a graph attention network.

8. The intelligent accounting and auditing system based on AI and blockchain according to claim 7 is characterized in that: The process of processing the audit graph through the AI ​​audit model: The audit graph is input into the AI ​​audit model. The graph convolutional network in the AI ​​audit model performs topological analysis on the audit graph to obtain potential risk points and the audit status including the node feature vector set and the global feature vector of the graph. The reinforcement learning mechanism is based on the audit status. - Greedy strategy selects appropriate audit strategies from a pre-set audit strategy set and uses the audit strategies to evaluate potential risk points to obtain audit results including problem description, problem nature, amount involved, risk level and audit recommendations; The audit status is an intermediate product of potential risk points obtained by performing topological analysis on the audit graph by the graph convolutional network.

9. The intelligent accounting and auditing system based on AI and blockchain according to claim 7 is characterized in that: The original accounting data, audit data and audit results are encrypted using the AES-256 symmetric encryption algorithm and stored in the blockchain.

Images