Digital finance and tax auditing method based on artificial intelligence

Abstract

The invention relates to a digital finance and taxation auditing method based on artificial intelligence, and the method comprises the following steps: S1, obtaining electric finance and taxation multi-source data, and carrying out the preprocessing of the data, and obtaining the preprocessed electric finance and taxation multi-source data; s2, constructing a finance and taxation knowledge graph according to the preprocessed power finance and taxation multi-source data; s3, constructing a multi-modal risk detection model, and obtaining a risk event list and a whole-course data traceability path according to the finance and taxation knowledge graph and the real-time electric finance and taxation multi-source data; s4, according to the risk event and traceability, based on a rule model, generating a structured decision instruction set including all key information and traceability links; and S5, converting the structured decision instruction set and the finance and tax knowledge graph into a standard relation graph for input, and dynamically displaying fund flow, risk events, abnormal aggregation and traceability paths among enterprises through D3. Js rendering. According to the invention, the efficiency and reliability of electric power digital finance and taxation auditing are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a digital finance and tax auditing method based on artificial intelligence. Background Art

[0002] With the continuous development of the digital economy, the power industry, as a key foundational industry, is undergoing a critical phase of digital transformation and business innovation. In recent years, the business scope of power companies has become increasingly diverse, encompassing all aspects of power generation, transmission, distribution, and sales, and also expanding into new business models such as energy services and financial investment. The industry's internal capital flows, invoice flows, and energy flows are highly complex, with long business chains and a large number of entities, resulting in extensive cross-regional, cross-departmental, and cross-platform capital and bill transactions.

[0003] Against this backdrop, the power industry faces even more severe challenges in fiscal and tax compliance management and risk prevention. Traditional fiscal and tax auditing methods rely heavily on manual review and single-system rule-based auditing, making it difficult to automatically compare and analyze data from multiple links and sources, including across multiple processes. With the emergence of new compliance risks such as false invoices, discrepancies between energy and cash flows, and capital circulation, relying solely on historical rules and case experience is no longer sufficient to meet the audit and risk identification needs in complex scenarios.

[0004] On the other hand, power companies generate a rich variety of business data in their daily operations, including financial accounts, bill documents, contract images, power time series, equipment operation logs, and other structured and unstructured information. This data is scattered across various business systems and management platforms, resulting in severe data silos and inconsistent data standards. Faced with multimodal, multi-source, and heterogeneous industry data, breaking down these various information barriers to achieve unified collection, integrated modeling, and comprehensive audit analysis is a key technical challenge and a major challenge in digital power finance and taxation auditing.

[0005] Furthermore, tax risks in the power sector are not only closely linked to invoices and accounting data but are also highly coupled to actual energy and physical flows. For example, issues such as false electricity invoices, significant discrepancies between bills and electricity consumption data, and abnormal capital flow accumulation are often hidden within massive volumes of transaction and energy data. Relying solely on a single data source or simple rules makes it difficult to gain insight into the underlying issues, making it difficult to promptly identify and trace violations. Summary of the Invention

[0006] In order to solve the above problems, the purpose of the present invention is to provide a digital fiscal and tax audit method based on artificial intelligence, which significantly improves the efficiency of digital fiscal and tax audit of electricity.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A digital finance and tax audit method based on artificial intelligence, comprising the following steps: S1: Acquire and preprocess the multi-source data on electricity finance and taxation to obtain the pre-processed multi-source data on electricity finance and taxation; S2: Construct a finance and taxation knowledge graph based on the pre-processed electricity finance and taxation multi-source data; S3: Build a multimodal risk detection model to obtain a list of risk events and the entire data traceability path based on the finance and taxation knowledge graph and real-time electricity finance and taxation multi-source data; S4: Based on risk events and traceability, a structured decision instruction set is generated based on the rule model, which includes all key information and traceability links; S5 converts structured decision-making instruction sets and financial and tax knowledge graphs into standard relationship diagram inputs, and dynamically displays inter-enterprise capital flows, risk events, abnormal aggregations, and traceability paths through D3.js rendering.

[0008] Furthermore, we can obtain multi-source data on electricity finance and taxation, including structured data, unstructured data, and streaming data, as follows: Structured data uses a multi-channel data extraction strategy to obtain structured data from the ERP system and the Golden Tax system through direct database connection, API interface, and file transfer. For the ERP system, general ledger details, accounts payable, purchase orders, and payment document data are obtained; for the Golden Tax system, invoice issuance lists, input tax deduction lists, and tax return data are obtained. Unstructured data acquisition: By establishing a document management system, we can collect images of bills and contract documents through batch scanning, mobile photo upload, and email attachment extraction. We also set up a file monitoring mechanism to automatically identify new files and trigger processing. We also establish file naming standards and storage directory structures to ensure file traceability. Stream data acquisition: establish a real-time data transmission channel with the bank system, obtain bank transaction data through the bank-enterprise direct connection interface, and set up a message queue buffer mechanism to handle data transmission delays and network fluctuations.

[0009] Further, structured data preprocessing is as follows: For structured datasets, calculate the field completeness index: ; Among them, CI j is the integrity index of the jth field, I(•) is the indicator function, d i,j It represents the jth field value of the i-th record, and n′ represents the total number of records; Use the quartile method to identify outliers. For the numeric field DX: Q1=P 25 (DX),Q3=P 75 (DX); IQR=Q3-Q1; Outlier boundary = [Q1-1.5×IQR, Q3+1.5×IQR]; data points outside the boundary are marked as outliers and manually reviewed or rule modified; Establish a cross-system data consistency verification mechanism, cross-validate key fields, and calculate consistency indicators:

[0010] Among them, m is the number of matching records, ERP i and Tax i These are the corresponding records of the ERP system and the Golden Tax system respectively.

[0011] Further, unstructured data preprocessing is as follows: Pre-process the bill image, optimize the image quality through histogram equalization and denoising algorithms, and improve OCR recognition accuracy: Confidence level of OCR recognition results Assessment: ; in, For the The recognition confidence of characters, N is the total number of characters; when the overall confidence is lower than the threshold, the manual review process is triggered; Extract key information fields of bills and contracts based on template matching and regular expressions k : Field k =argmax F Similarity(Template k ,F) Among them, Template k is the template of the kth field, F is the candidate field region; Similarity is the template matching function; Establish a field format verification mechanism to perform format verification on the extraction results: ; Where f is the extraction result; for fields that do not meet the format requirements, the edit distance algorithm is used for error correction; A validity indicator function to determine whether f has passed the verification; The topic model is used to classify the contract documents, and the key terms of the contract are identified based on the TF-IDF algorithm.

[0012] Further, stream data preprocessing is as follows: Identify duplicate transaction records based on business rules :

[0013] Among them, Hash(•) is the record hash function, t a , t b Transaction r a 、r b time, θ is the time threshold; Use sliding window statistical method to detect abnormal transactions and calculate Z score Z score : ; Among them, Amount t is the transaction amount at time t; μ window and σ window are the mean and standard deviation of the transaction amount in the current sliding window respectively; When | Z score ∣When it is greater than the preset value, it is marked as a suspected abnormal transaction; Fourier transform is used to identify periodic patterns in transaction amounts, and moving average method is used to extract trends; Create a multi-dimensional index based on Company, Time, Amount, and Transaction Type to improve query efficiency: Index=f'(Company,Time,Amount,Type); Among them, f' is the index key function used for multi-dimensional retrieval.

[0014] Furthermore, based on the pre-processed electricity finance and taxation multi-source data, a finance and taxation knowledge graph is constructed, as follows: For the input text T=[w1,w2,...,w n ], where w n is the nth word segment; BERT uses a multi-layer Transformer to encode context-related semantics for each word segment: H=BERT(T)=[h1,h2,...,h n ]; Each h n Represents w n Deep semantic representation of ; BERT output is used as BiLSTM input to improve the ability to mine local context order: Forward LSTM: ; in, is the hidden state output of the forward LSTM at the i′′th word; h i′′is the input feature vector at the i′′th position; Backward LSTM: ; in, is the hidden state output of the reverse LSTM at the i′′th position; Splicing the two: ; in, is the concatenated output of the bidirectional LSTM; For n word segments, each with L entity labels, the CRF layer is used to jointly model the optimal label sequence y=[y1,...,y n ], y n The optimal label for the nth word: Scoring function : ; in, Indicates that the i′′th word is marked y i′′ The score is obtained by the linear layer after BiLSTM output; From label y i′′ Move to label y i′′+1 The transfer score is the label transfer matrix learned by the CRF layer; n is the sequence length; Sequence probability: ; in, is the probability of the optimal label sequence y given the input text T; Y T is the set of all possible label sequences for a given input text T; y′ is Y T Any label sequence extracted from ; By maximizing the probability of the true annotation sequence, the optimal label path is inferred and the entity label is output; Align structured relationships based on rules and business primary keys, including enterprise ID-invoice number, bank card number, and payment amount, to achieve cross-table mapping of the same business entity: Rstruct(x,y)=I(Key x =Key y ); Among them, I(•) is the indicator function, Key x and Key y are the business primary keys of business x and business y respectively, and a directed association is obtained; Rstruct(x,y) is the structural consistency indicator function; For unstructured text, a relation extraction model classifier is used to extract relation triplets and map them to graph relation edges. After entity disambiguation, relationship fusion rules are established; The combination relationship is described using the path method: Path(A,D)=A→Transaction B→Bill C→Fund Flow D; Set up relationship credibility tiers, incorporating data quality: Edge_Score(e1,r,e2)=α1×Key_Match+α2×Text_Match+α3×Operation Log Among them, α1, α2, and α3 are weighted parameters; Key_Match is the primary key consistency matching score; Text_Match is the similarity score of the text content; Edge_Score(e1,r,e2) represents the matching score of edge (e1,r,e2); e1 and e2 are entities, and r is the association relationship; Finally, the financial and tax knowledge graph is defined as G=(V,E), where V=entity node set, E=relationship edge set with type and attribute; A graph database is used to implement relational storage of knowledge graphs, with self-description of node and side attributes, supporting fast path query and multi-dimensional data retrieval.

[0015] Furthermore, a multimodal risk detection model is constructed. Based on the finance and taxation knowledge graph and real-time electricity finance and taxation multi-source data, a list of risk events and the entire data traceability path are obtained. The details are as follows: Map real-time electricity finance and taxation multi-source data into corresponding entities and relationships, align them with knowledge graph nodes, and integrate multiple features for risk label modeling and anomaly detection; Based on the multimodal risk detection model, typical risk labels are identified and a list of risk events is obtained; Combining the financial and tax knowledge graph with log metadata, a link record is formed from event triggering to the source of data.

[0016] Furthermore, based on the multimodal risk detection model, typical risk labels are identified and a list of risk events is obtained, as follows: The multimodal risk detection model includes a bill anomaly detection sub-model, a suspected false invoice identification sub-model, a three-invoice discrepancy detection sub-model, and a false business detection sub-model, as follows: The bill anomaly detection sub-model combines CR recognition, field verification, and bill dual index analysis to flag any abnormalities in format, number, or amount, and obtain a collection of abnormal bills. The suspected fraudulent invoice identification sub-model uses a graph neural network to detect suspicious capital reflux loops: ; Among them, Loopg represents a loop consisting of g nodes; v1,v2,...,v g is a node in the graph neural network; v g Indicates the g-th node; EdgeType(e a,a+1 ) is edge e a,a+1 Type; If the risk loop accounts for a large proportion of the amount, set RiskLabel(Loop)=1. RiskLabel is a label function used to determine whether the closed loop (Loop) belongs to the risk type; The three-document discrepancy detection sub-model checks the consistency of the intersection of the purchase contract, invoice, and bank receipt in terms of amount, subject, time, and item:

[0017] InconsistentTriplet is a set of inconsistent triples; c and b are two entities (such as contracts, invoices, documents, etc.); r is an association relationship; A(c) and A(b) are the amount (or volume) fields of c and b; is the amount (or volume) difference threshold; T(c), is the time field of c and b; θ t is the time difference threshold.

[0018] Company(c) and Company(b) are the company identifiers of c and b respectively; ∨ represents logical OR, and if any inconsistency condition is met, it is considered inconsistent; The fake business detection sub-model combines power time series and bill data to check the mismatch between procurement / production / power consumption and revenue / invoicing. ; Among them, RealPower(t0,t1) is the actual power consumption in the time interval [t0,t1], and ReportedProd(t0,t1) is the theoretical power consumption converted from the output or output value reported by the enterprise in the interval [t0,t1]; If the FakeBusinessScore score is higher than the threshold, it is marked as suspected fake business; Based on the detection results of the multimodal risk detection model, multiple labels are combined to automatically generate a checklist: ; Among them, RiskEventList is the risk event list, which refers to the collection of all events identified as risks; l Indicates the event type, L is the set of all risk event types; Eventt,l Indicates time t , event type l A specific event below; RiskLabel l ( t ) Judge the event Event t,l In time t A label indicating whether it is a risk event. If its value is 1, it means it is a risk event.

[0019] Furthermore, by combining the financial and tax knowledge graph with log metadata, a link record from event triggering to the source of data is formed, as follows: Use the financial and tax knowledge graph to find all causal links related to risk events. s As the starting point, e t is the end point, and the traceability path is all valid paths: ; in, Judge edge e i,i+1 Is it a related edge? p is a path; Use weight multiplication to obtain the highest weight path : ; in, For edge e i,i+1 risk weights; For each risk label determination event, the original data, trigger module, reference rules and output fields used are recorded to form a complete traceability chain.

[0020] Furthermore, based on risk events and traceability, a structured decision instruction set is generated based on the rule model, which contains all key information and traceability links, as follows: Based on the risk event list and traceability path data, the system automatically captures the structured features of each risk event and then uses them as input for the rule model; The loaded rule model includes static rules and dynamic rule models, and each risk event is matched with the rules in the rule model; Perform conditional evaluation on the matching rules, trigger corresponding actions, and structure the triggered actions and related information into decision instructions.

[0021] The present invention has the following beneficial effects: 1. This invention uses the financial and tax knowledge graph to comprehensively model core objects such as enterprises, accounts, invoices, and transactions, and their business relationships. It supports dynamic adaptation of complex entity relationships and provides a solid and comprehensive data and knowledge foundation for audit analysis. 2. This invention uses multimodal artificial intelligence technology to intelligently detect risks such as bill anomalies, suspected fraudulent invoices, discrepancies between three documents, and fraudulent transactions. A hybrid engine based on rules and models automatically outputs structured, traceable decision instructions, improving audit accuracy, automation, and decision transparency. 3. This invention uses dynamic graphs to intuitively and dynamically display inter-enterprise capital flows, risk events, abnormal aggregations, and traceability links, greatly enhancing the intuitiveness and interactivity of risk discovery. All analysis and decision-making results support full traceability and responsibility attribution, significantly improving audit efficiency, compliance, and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a digital tax audit method based on artificial intelligence is provided, comprising the following steps: S1: Acquire and preprocess the multi-source data on electricity finance and taxation to obtain the pre-processed multi-source data on electricity finance and taxation; S2: Construct a finance and taxation knowledge graph based on the pre-processed electricity finance and taxation multi-source data; S3: Build a multimodal risk detection model. Based on the finance and taxation knowledge graph and real-time electricity finance and taxation multi-source data, obtain a list of risk events (including labels such as bill anomalies, suspected false invoices, discrepancies between three invoices, and false business) and the entire data traceability path; S4: Based on risk events and traceability, a structured decision instruction set is generated based on the rule model, which includes all key information and traceability links; S5 converts structured decision-making instruction sets and financial and tax knowledge graphs into standard relationship diagram inputs, and dynamically displays inter-enterprise capital flows, risk events, abnormal aggregations, and traceability paths through D3.js rendering.

[0024] In this embodiment, multi-source data on electricity finance and taxation is obtained, including structured data, unstructured data, and stream data, as follows: Structured data uses a multi-channel data extraction strategy to obtain structured data from the ERP system and the Golden Tax system through direct database connection, API interface, and file transfer. For the ERP system, general ledger details, accounts payable, purchase orders, and payment document data are obtained; for the Golden Tax system, invoice issuance lists, input tax deduction lists, and tax return data are obtained. Unstructured data acquisition: By establishing a document management system, we can collect images of bills and contract documents through batch scanning, mobile photo upload, and email attachment extraction. We also set up a file monitoring mechanism to automatically identify new files and trigger processing. We also establish file naming standards and storage directory structures to ensure file traceability. Stream data acquisition: establish a real-time data transmission channel with the bank system, obtain bank transaction data through the bank-enterprise direct connection interface, and set up a message queue buffer mechanism to handle data transmission delays and network fluctuations.

[0025] In this embodiment, structured data preprocessing is specifically as follows: For structured datasets, calculate the field completeness index: ; Among them, CI j is the integrity index of the jth field, I(•) is the indicator function, d i,j It represents the jth field value of the i-th record, and n′ represents the total number of records; Use the quartile method to identify outliers. For the numeric field DX: Q1=P 25 (DX),Q3=P 75 (DX); IQR=Q3-Q1; Among them, Q1 is the first quartile; Q3 is the third quartile; IQR is the interquartile range; Outlier boundary = [Q1-1.5×IQR, Q3+1.5×IQR]; data points outside the boundary are marked as outliers and manually reviewed or rule modified; Establish a cross-system data consistency verification mechanism, cross-validate key fields, and calculate consistency indicators:

[0026] Among them, m is the number of matching records, ERP i and Tax i These are the corresponding records of the ERP system and the Golden Tax system respectively.

[0027] In this embodiment, unstructured data preprocessing is as follows: Pre-process the bill image, optimize the image quality through histogram equalization and denoising algorithms, and improve OCR recognition accuracy: Confidence level of OCR recognition results Assessment: ; in, For the The recognition confidence of characters, N is the total number of characters; when the overall confidence is lower than the threshold, the manual review process is triggered; Extract key information fields of bills and contracts based on template matching and regular expressions k : Field k =argmax F Similarity(Template k ,F) Among them, Template k is the template of the kth field, F is the candidate field region; Similarity is the template matching function; Establish a field format verification mechanism to perform format verification on the extraction results: ; Among them, f is the extraction result; A validity indicator function to determine whether f has passed the verification; For fields that do not meet format requirements, the edit distance algorithm is used for error correction; Use topic models to classify contract documents and identify key contract terms based on the TF-IDF algorithm; In this embodiment, the stream data is pre-processed as follows: Identify duplicate transaction records based on business rules :

[0028] Among them, Hash(•) is the record hash function, t a , t b Transaction r a 、r b time, θ is the time threshold; Use sliding window statistical method to detect abnormal transactions and calculate Z score Z score : ; Among them, Amount t is the transaction amount at time t; μ window and σ window are the mean and standard deviation of the transaction amount in the current sliding window respectively; When | Z score ∣When it is greater than the preset value, it is marked as a suspected abnormal transaction; Fourier transform is used to identify periodic patterns in transaction amounts, and moving average method is used to extract trends; Create a multi-dimensional index based on Company, Time, Amount, and Transaction Type to improve query efficiency: Index=f′(Company,Time,Amount,Type); Among them, f′ is the index key function used for multi-dimensional retrieval.

[0029] In this embodiment, a finance and taxation knowledge graph is constructed based on the pre-processed electricity finance and taxation multi-source data, as follows: For the input text T=[w1,w2,...,w n ], where w n is the nth word; BERT uses a multi-layer Transformer to encode the context-dependent semantics of each word: H=BERT(T)=[h1,h2,...,h n ]; Each h n Represents w n Deep semantic representation of ; BERT output is used as BiLSTM input to improve the ability to mine local context order: Forward LSTM: ; in, is the hidden state output of the forward LSTM at the i′′th word; h i′′ is the input feature vector at the i′′th position; Backward LSTM: ; in, is the hidden state output of the reverse LSTM at the i′′th position; Splicing the two: ; in, is the concatenated output of the bidirectional LSTM; For n word segments, each with L entity labels, the CRF layer is used to jointly model the optimal label sequence y=[y1,...,y n ], y n The optimal label for the nth word: Scoring function : ; in, Indicates that the i′′th word is marked y i′′The score is obtained by the linear layer after BiLSTM output; From label y i′′ Move to label y i′′+1 The transfer score is the label transfer matrix learned by the CRF layer; n is the sequence length; Sequence probability: ; in, is the probability of the optimal label sequence y given the input text T; Y T is the set of all possible label sequences for a given input text T; y′ is Y T Any label sequence extracted from ; By maximizing the probability of true annotation sequences, the optimal label path is inferred and entity labels (enterprise, supplier, product, tax rate, amount, etc.) are output. Typical tag paths include: transaction relationships (enterprise-supplier, enterprise-product), bill ownership (enterprise-bill, supplier-bill), fund flow (paying account-receiving account), and cross-document / cross-table associations (contract-invoice-transaction). Align structured relationships based on rules and business primary keys, including enterprise ID-invoice number, bank card number, and payment amount, to achieve cross-table mapping of the same business entity: Rstruct(x,y)=I(Key x =Key y ); Among them, I(•) is the indicator function, Key x and Key y are the business primary keys of business x and business y respectively, and a directed association is obtained; Rstruct(x,y) is the structural consistency indicator function; For unstructured text (such as invoice notes, contract terms, etc.), a relation extraction model (BERT+Span) classifier is used to extract relation triplets and map them to graph relation edges. Relation = (e1, r, e2); where Relation represents a relationship, e1 and e2 are entities, and r is the association relationship; for example: "A pays B an amount of C" → (A, payment, B); After entity disambiguation, relationship fusion rules are established, such as: linking a single bill to multiple businesses / invoices / accounts; tracking capital inflows and outflows (multi-step path analysis); The combination relationship is described using the path method: Path(A,D)=A→Transaction B→Bill C→Fund Flow D; Set up relationship credibility tiers, incorporating data quality: Edge_Score(e1,r,e2)=α1×Key_Match+α2×Text_Match+α3×Operation Log Among them, α1, α2, and α3 are weighted parameters, which are configured according to the actual business scenario; Key_Match is the consistency matching score of the primary key (such as ID, number, account number, etc.); Text_Match is the similarity score of the text content (such as name, description, summary); Edge_Score(e1,r,e2) represents the matching score of the edge (e1,r,e2); Node (entity): enterprise, supplier, product, invoice, contract, bank account...; Side (relationship): supply, procurement, receipt, holding, invoicing, payment, capital flow, etc.; Finally, the financial and tax knowledge graph is defined as G=(V,E), where V=entity node set, E=relationship edge set with type and attribute; Graph databases (Neo4j, GraphDB, etc.) are used to implement relational storage of knowledge graphs, with self-describing node and side attributes, supporting fast path queries and multi-dimensional data retrieval.

[0030] In this embodiment, a multimodal risk detection model is constructed. Based on the finance and taxation knowledge graph and real-time electricity finance and taxation multi-source data, a risk event list (including labels such as bill anomalies, suspected false invoices, discrepancies between three documents, and false business) and the entire data traceability path are obtained. The details are as follows: Map real-time electricity finance and taxation multi-source data into corresponding entities and relationships, align them with knowledge graph nodes, and integrate multiple features for risk label modeling and anomaly detection; Based on the multimodal risk detection model, typical risk labels are identified and a list of risk events is obtained; Combining the financial and tax knowledge graph with log metadata, a link record is formed from event triggering to the source of data.

[0031] In this embodiment, typical risk labels are identified and a list of risk events is obtained based on a multimodal risk detection model. Specifically, the multimodal risk detection model includes a bill anomaly detection sub-model, a suspected fraudulent invoice identification sub-model, a three-invoice discrepancy detection sub-model, and a fraudulent business detection sub-model. Specifically, The bill anomaly detection sub-model combines CR recognition, field verification, and bill dual index analysis to flag any abnormalities in format, number, or amount, and obtain a collection of abnormal bills. The suspected fraudulent invoice identification sub-model uses a graph neural network to detect suspicious capital reflux loops: ; Among them, Loop g represents a loop consisting of g nodes; v1,v2,...,vg is a node in the graph neural network; v g Indicates the g-th node; EdgeType(e a,a+1 ) is edge e a,a+1 Type; If the risk loop accounts for a large proportion of the amount, set RiskLabel(Loop)=1. RiskLabel is a label function used to determine whether the closed loop (Loop) belongs to the risk type; The three-document discrepancy detection sub-model checks the consistency of the intersection of the purchase contract, invoice, and bank receipt in terms of amount, subject, time, and item: ; InconsistentTriplet is a set of inconsistent triples; c and b are two entities (such as contracts, invoices, documents, etc.); r is an association relationship; A(c) and A(b) are the amount (or volume) fields of c and b; is the amount (or volume) difference threshold; T(c), is the time field of c and b; θ t is the time difference threshold.

[0032] Company(c) and Company(b) are the company identifiers of c and b respectively; ∨ represents logical OR, and if any inconsistency condition is met, it is considered inconsistent; The fake business detection sub-model combines power time series and bill data to check the mismatch between procurement / production / power consumption and revenue / invoicing. ; Among them, RealPower(t0,t1) is the actual power consumption in the time interval [t0,t1], and ReportedProd(t0,t1) is the theoretical power consumption converted from the output or output value reported by the enterprise in the interval [t0,t1]; If the FakeBusinessScore score is higher than the threshold, it is marked as suspected fake business; Based on the detection results of the multimodal risk detection model, multiple labels are combined to automatically generate a checklist: ; Among them, RiskEventList is the risk event list, which refers to the collection of all events identified as risks; l Indicates the event type, L is the set of all risk event types; Event t,l Indicates time t , event type lA specific event below; RiskLabel l ( t ) Judge the event Event t,l The label of whether it is a risk event at time t. If its value is 1, it means it is judged as a risk event.

[0033] In this embodiment, the financial and tax knowledge graph and log metadata are combined to form a link record from event triggering to the source of data, as follows: Use the financial and tax knowledge graph to find all causal links related to risk events. s As the starting point, e t is the end point, and the traceability path is all valid paths : ; in, Judge edge e i,i+1 Is it a related edge? p is the path. Use weight multiplication to obtain the highest weight path. : ; in, For edge e i,i+1 risk weights; For each risk label determination event, the original data, trigger module, reference rules and output fields used are recorded to form a complete traceability chain.

[0034] In this embodiment, based on risk events and traceability, a structured decision instruction set is generated based on a rule model, which contains all key information and traceability links, as follows: Based on the risk event list and traceability path data, the system automatically captures the structural features of each risk event (such as subject type, data field, exception type, amount range, time, frequency, link level, etc.) and then uses them as input to the rule model; The loaded rule model includes static rules and dynamic rule models, and each risk event is matched with the rules in the rule model; The rule model includes static rules and dynamic rule models: Static rules: These are hard-coded rules based on current regulations, internal control requirements, and business logic processes. For example, if multiple consecutive transfers of the same amount to the same company occur within a single day, the risk is marked as high. The Drools engine is used to execute over 5,000 tax rules. Dynamic rules: Adaptive rules derived from historical data, machine learning model reasoning, or real-time monitoring. For example, "If the risk score is greater than 0.8 and a capital reflux path appears in the recent network relationship graph, it is considered an abnormal capital flow."

[0035] Perform conditional evaluation on the matching rules, trigger corresponding actions, and structure the triggered actions and related information into decision instructions.

[0036] Each instruction includes the following key information: Instruction ID: uniquely identifies this decision instruction; Risk types: such as "false invoices", "discrepancies between three documents", and "circuitous capital flow"; Triggering subject / entity: such as the company involved, bank account, specific bill number; Recommended actions: such as "report", "freeze account", "manual review", "generate audit report", etc. Risk priority / confidence level: helps with subsequent automatic diversion and sequential processing; Traceability link ID and detailed path: Clearly record the entire chain of data paths, entities involved, and original data sources that lead to risk events, for subsequent visualization, accountability, or review.

[0037] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0038] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0039] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0040] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. The digital finance and tax audit method based on artificial intelligence is characterized by: The following steps are involved: S1: Acquire and preprocess the multi-source data on electricity finance and taxation to obtain the pre-processed multi-source data on electricity finance and taxation; S2: Construct a finance and taxation knowledge graph based on the pre-processed electricity finance and taxation multi-source data; S3: Build a multimodal risk detection model to obtain a list of risk events and the entire data traceability path based on the finance and taxation knowledge graph and real-time electricity finance and taxation multi-source data; S4: Based on risk events and traceability, a structured decision instruction set is generated based on the rule model, which includes all key information and traceability links; S5 converts structured decision-making instruction sets and financial and tax knowledge graphs into standard relationship diagram inputs, and dynamically displays inter-enterprise capital flows, risk events, abnormal aggregations, and traceability paths through D3.js rendering.

2. The digital finance and tax audit method based on artificial intelligence according to claim 1 is characterized in that: The acquisition of multi-source data on electricity finance and taxation, including structured data, unstructured data and streaming data, is as follows: Structured data uses a multi-channel data extraction strategy to obtain structured data from the ERP system and the Golden Tax system through direct database connection, API interface, and file transfer. For the ERP system, general ledger details, accounts payable, purchase orders, and payment document data are obtained; for the Golden Tax system, invoice issuance lists, input tax deduction lists, and tax return data are obtained. Unstructured data acquisition: By establishing a document management system, we can collect images of bills and contract documents through batch scanning, mobile photo upload, and email attachment extraction. We also set up a file monitoring mechanism to automatically identify new files and trigger processing. We also establish file naming standards and storage directory structures to ensure file traceability. Stream data acquisition: establish a real-time data transmission channel with the bank system, obtain bank transaction data through the bank-enterprise direct connection interface, and set up a message queue buffer mechanism to handle data transmission delays and network fluctuations.

3. The digital finance and tax audit method based on artificial intelligence according to claim 2 is characterized in that: Structured data preprocessing, as follows: For structured datasets, calculate the field completeness index: ； Among them, CI j is the integrity index of the jth field, I(•) is the indicator function, d i,j It represents the jth field value of the i-th record, and n′ represents the total number of records; Use the quartile method to identify outliers. For the numeric field DX: Q1=P 25 (DX),Q3=P 75 (DX); IQR=Q3-Q1; Among them, Q1 is the first quartile; Q3 is the third quartile; IQR is the interquartile range; Outlier boundary = [Q1-1.5×IQR, Q3+1.5×IQR]; data points outside the boundary are marked as outliers and manually reviewed or rule modified; Establish a cross-system data consistency verification mechanism, cross-validate key fields, and calculate consistency indicators: ； Among them, m is the number of matching records, ERP i and Tax i These are the corresponding records of the ERP system and the Golden Tax system respectively.

4. The digital finance and tax audit method based on artificial intelligence according to claim 2 is characterized in that: Unstructured data preprocessing, as follows: Pre-process the bill image, optimize the image quality through histogram equalization and denoising algorithms, and improve OCR recognition accuracy: Confidence level of OCR recognition results Assessment: ； in, For the The recognition confidence of characters, N is the total number of characters; when the overall confidence is lower than the threshold, the manual review process is triggered; Extract key information fields of bills and contracts based on template matching and regular expressions k : Field k =argmax F Similarity(Template k ,F); Among them, Template k is the template of the kth field, F is the candidate field region; Similarity is the template matching function; Establish a field format verification mechanism to perform format verification on the extraction results: ； Where f is the extraction result; for fields that do not meet the format requirements, the edit distance algorithm is used for error correction; A validity indicator function to determine whether f has passed the verification; The topic model is used to classify the contract documents, and the key terms of the contract are identified based on the TF-IDF algorithm.

5. The digital finance and tax audit method based on artificial intelligence according to claim 2 is characterized in that: Stream data preprocessing, as follows: Identify duplicate transaction records based on business rules : ； Among them, Hash(•) is the record hash function, t a , t b Transaction r a 、r b time, θ is the time threshold; Use sliding window statistical method to detect abnormal transactions and calculate Z score Z score : ； Among them, Amount t is the transaction amount at time t; μ window and σ window are the mean and standard deviation of the transaction amount in the current sliding window respectively; When | Z score ∣When it is greater than the preset value, it is marked as a suspected abnormal transaction; Fourier transform is used to identify periodic patterns in transaction amounts, and moving average method is used to extract trends; Create a multi-dimensional index based on Company, Time, Amount, and Transaction Type to improve query efficiency: Index=f'(Company,Time,Amount,Type); Among them, f' is the index key function used for multi-dimensional retrieval.

6. The digital finance and tax audit method based on artificial intelligence according to claim 1 is characterized in that: The financial and taxation knowledge graph is constructed based on the pre-processed power finance and taxation multi-source data, as follows: For the input text T=[w1,w2,...,w n ], where w n is the nth word segment; BERT uses a multi-layer Transformer to encode context-related semantics for each word segment: H=BERT(T)=[h1,h2,...,h n ]; Each h n Represents w n Deep semantic representation of ; BERT output is used as BiLSTM input to improve the ability to mine local context order: Forward LSTM: ； in, is the hidden state output of the forward LSTM at the i′′th word; h i′′ is the input feature vector at the i′′th position; Backward LSTM: ； in, is the hidden state output of the reverse LSTM at the i′′th position; Splicing the two: ； in, is the concatenated output of the bidirectional LSTM; For n word segments, each with L entity labels, the CRF layer is used to jointly model the optimal label sequence y=[y1,...,y n ], y n The optimal label for the nth word: Scoring function : ； in, Indicates that the i′′th word is marked y i′′ The score is obtained by the linear layer after BiLSTM output; From label y i′′ Move to label y i′′+1 The transfer score is the label transfer matrix learned by the CRF layer; n is the sequence length; Sequence probability: ； in, is the probability of the optimal label sequence y given the input text T; Y T is the set of all possible label sequences for a given input text T; y′ is Y T Any label sequence extracted from ; By maximizing the probability of the true annotation sequence, the optimal label path is inferred and the entity label is output; Align structured relationships based on rules and business primary keys, including enterprise ID-invoice number, bank card number, and payment amount, to achieve cross-table mapping of the same business entity: Rstruct(x,y)=I(Key x =Key y ); Among them, I(•) is the indicator function, Key x and Key y are the business primary keys of business x and business y respectively, and a directed association is obtained; Rstruct(x,y) is the structural consistency indicator function; For unstructured text, a relation extraction model classifier is used to extract relation triplets and map them to graph relation edges. After entity disambiguation, relationship fusion rules are established; The combination relationship is described using the path method: Path(A,D)=A→Transaction B→Bill C→Fund Flow D; Set up relationship credibility tiers, incorporating data quality: Edge_Score(e1,r,e2)=α1×Key_Match+α2×Text_Match+α3×Operation Log Among them, α1, α2, and α3 are weighted parameters; Key_Match is the primary key consistency matching score; Text_Match is the similarity score of the text content; Edge_Score(e1,r,e2) represents the matching score of edge (e1,r,e2); e1 and e2 are entities, and r is the association relationship; Finally, the financial and tax knowledge graph is defined as G=(V,E), where V=entity node set, E=relationship edge set with type and attribute; A graph database is used to implement relational storage of knowledge graphs, with self-description of node and side attributes, supporting fast path query and multi-dimensional data retrieval.

7. The digital finance and tax audit method based on artificial intelligence according to claim 1 is characterized in that: The multimodal risk detection model is constructed to obtain a list of risk events and a full data traceability path based on the finance and taxation knowledge graph and real-time electricity finance and taxation multi-source data, as follows: Map real-time electricity finance and taxation multi-source data into corresponding entities and relationships, align them with knowledge graph nodes, and integrate multiple features for risk label modeling and anomaly detection; Based on the multimodal risk detection model, typical risk labels are identified and a list of risk events is obtained; Combining the financial and tax knowledge graph with log metadata, a link record is formed from event triggering to the source of data.

8. The digital finance and tax audit method based on artificial intelligence according to claim 7 is characterized in that: The multimodal risk detection model is used to identify typical risk labels and obtain a list of risk events, specifically as follows: The multimodal risk detection model includes a bill anomaly detection sub-model, a suspected false invoice identification sub-model, a three-invoice discrepancy detection sub-model, and a false business detection sub-model, specifically as follows: The bill anomaly detection sub-model combines CR recognition, field verification, and bill dual index analysis to flag any abnormalities in format, number, or amount, and obtain a collection of abnormal bills. The suspected fraudulent invoice identification sub-model uses a graph neural network to detect suspicious capital reflux loops: ; Among them, Loop g represents a loop consisting of g nodes; v1,v2,...,v g is a node in the graph neural network; v g Indicates the g-th node; EdgeType(e a,a+1 ) is edge e a,a+1 Type; If the risk loop accounts for a large proportion of the amount, set RiskLabel(Loop)=1. RiskLabel is a label function used to determine whether the closed loop (Loop) belongs to the risk type; The three-document discrepancy detection sub-model checks the consistency of the intersection of the purchase contract, invoice, and bank receipt in terms of amount, subject, time, and item: ； InconsistentTriplet is a set of inconsistent triples; c and b are two entities; r is an association relationship; A(c) and A(b) are the amount or volume fields of c and b; is the amount or volume difference threshold; T(c), is the time field of c and b; θ t is the time difference threshold; Company(c) and Company(b) are the company identifiers of c and b respectively; ∨ represents logical OR, and if any inconsistency condition is met, it is considered inconsistent; The fake business detection sub-model combines power time series and bill data to check the mismatch between procurement / production / power consumption and revenue / invoicing. ； Among them, RealPower(t0,t1) is the actual power consumption in the time interval [t0,t1], and ReportedProd(t0,t1) is the theoretical power consumption converted from the output or output value reported by the enterprise in the interval [t0,t1]; If the FakeBusinessScore score is higher than the threshold, it is marked as suspected fake business; Based on the detection results of the multimodal risk detection model, multiple labels are combined to automatically generate a checklist: ； Among them, RiskEventList is the risk event list, which refers to the collection of all events identified as risks; l Indicates the event type, L is the set of all risk event types; Event t,l Indicates time t , event type l A specific event below; RiskLabel l ( t ) Judge the event Event t,l In time t A label indicating whether it is a risk event. If its value is 1, it means it is a risk event.

9. The digital finance and tax audit method based on artificial intelligence according to claim 7 is characterized in that: The combination of the financial and tax knowledge graph and log metadata forms a link record from event triggering to the source of data, as follows: Use the financial and tax knowledge graph to find all causal links related to risk events. s As the starting point, e t is the end point, and the traceability path is all valid paths: ； in, Judge edge e i,i+1 Is it a related edge? p is a path; Use weight multiplication to obtain the highest weight path : ; in, For edge e i,i+1 risk weights; For each risk label determination event, the original data, trigger module, reference rules and output fields used are recorded to form a complete traceability chain.

10. The digital finance and tax audit method based on artificial intelligence according to claim 1 is characterized in that: According to the risk events and traceability, based on the rule model, a structured decision instruction set is generated, which contains all key information and traceability links, as follows: Based on the risk event list and traceability path data, the system automatically captures the structured features of each risk event and then uses them as input for the rule model; The loaded rule model includes static rules and dynamic rule models, and each risk event is matched with the rules in the rule model; Perform conditional evaluation on the matching rules, trigger corresponding actions, and structure the triggered actions and related information into decision instructions.

Images