Artificial intelligence-based digital finance and tax auditing method

Abstract

The present application relates to a kind of digital finance and tax auditing methods based on artificial intelligence, comprising the following steps: S1: obtaining electric power finance and tax multi-source data, and preprocessing, obtain the electric power finance and tax multi-source data after preprocessing;S2: according to the electric power finance and tax multi-source data after preprocessing, construct finance and tax knowledge graph;S3: construct multi-modal risk detection model, according to finance and tax knowledge graph and real-time electric power finance and tax multi-source data, obtain risk event list and whole-course data traceability path;S4: according to risk event and traceability, based on rule model, generate structured decision instruction set, contain all key information and traceability link;S5: structured decision instruction set and finance and tax knowledge graph are converted into standard relationship diagram input, render through D3.js, dynamically show inter-enterprise fund flow, risk event, abnormal aggregation and traceability path.The present application significantly improves the electric power digital finance and tax auditing efficiency and reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, and in particular to a digital financial and tax auditing method based on artificial intelligence. BACKGROUND

[0002] With the continuous development of digital economy, the power industry, as an important basic industry, is undergoing an important stage of digital transformation and business innovation. In recent years, the business scope of power enterprises is increasingly diversified, covering power generation, power transmission, power distribution, and power sales. It also extensively involves new industry forms such as energy services and financial investment. The internal fund flow, invoice flow, and energy flow of the industry are highly complex, with long business chains, a large number of subjects, and a large number of cross-regional, cross-department, and cross-platform fund and bill transactions.

[0003] Under this background, the financial and tax compliance management and risk prevention and control of the power industry face more severe challenges. On the one hand, traditional financial and tax auditing methods highly rely on manual review and rule setting based on a single system, making it difficult to achieve automatic comparison and full-link penetration analysis of multi-link and multi-source data. With the emergence of new compliance risks such as fake invoices, energy and fund flow inconsistency, and fund circulation, simply relying on historical rules and case experience cannot meet the needs of auditing and risk identification in complex scenarios.

[0004] On the other hand, the business data generated in the daily operation of power enterprises are of various types, including financial accounts, bill texts, contract images, power time series, and device operation logs. These data are scattered in different business systems and management platforms, and the data islands phenomenon is serious. The data standards are not unified. In the face of multi-modal, multi-source heterogeneous industry data, how to break down various information barriers and achieve unified collection, fusion modeling, and full-range audit analysis is the technical key and difficulty of digital financial and tax auditing in the power industry.

[0005] In addition, the tax risk in the power industry is not only closely related to invoices and accounting data, but also has a high coupling characteristic with actual energy flow and physical flow. For example, fake power invoices, serious inconsistency between invoices and power data, and abnormal aggregation of fund flow often hide in massive transaction and energy data. Relying on a single data source or simple rules cannot reveal the essence, making it difficult to discover and trace irregularities in a timely manner. SUMMARY

[0006] To solve the above problems, the purpose of the present application is to provide a digital financial and tax auditing method based on artificial intelligence, which significantly improves the efficiency of digital financial and tax auditing in the power industry.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] An artificial intelligence-based digital finance and tax auditing method, comprising the following steps:

[0009] S1: Obtain power finance and tax multi-source data and preprocess to obtain preprocessed power finance and tax multi-source data;

[0010] S2: Construct a finance and tax knowledge graph according to the preprocessed power finance and tax multi-source data;

[0011] S3: Construct a multi-modal risk detection model, and obtain a risk event list and a full-process data traceability path according to the finance and tax knowledge graph and real-time power finance and tax multi-source data;

[0012] S4: According to the risk event and the traceability, generate a structured decision instruction set based on a rule model, which contains all key information and traceability links;

[0013] S5: Convert the structured decision instruction set and the finance and tax knowledge graph into a standard relationship graph input, render it through D3.js, and dynamically display the inter-enterprise capital flow, risk events, abnormal aggregation, and traceability paths.

[0014] Further, the power finance and tax multi-source data includes structured data, unstructured data, and stream data, which are as follows:

[0015] The structured data adopts a multi-channel data extraction strategy, and is obtained through database direct connection, API interface, and file transmission mode from ERP systems and the Jintax system; for the ERP system, total account details, accounts payable, purchase orders, and payment document data are obtained; for the Jintax system, invoice issuance lists, input tax deduction lists, and tax return data are obtained;

[0016] The unstructured data is obtained by establishing a document management system, collecting bill images and contract documents through batch scanning, mobile phone shooting and uploading, and email attachment extraction; a file monitoring mechanism is set up to automatically identify new files and trigger the processing flow; a file naming specification and storage directory structure are established to ensure file traceability;

[0017] The stream data is obtained by establishing a real-time data transmission channel with the bank system, obtaining bank stream data through bank-enterprise direct interface, and setting up a message queue buffer mechanism to handle data transmission delays and network fluctuations.

[0018] Further, the structured data preprocessing is as follows:

[0019] For the structured data set, calculate the field integrity index:

[0020] ;

[0021] Wherein, CI jLet I(⋅) be the integrity index of the j-th field, and d be the indicator function. i,j This represents the value of the j-th field in the i-th record, where n′ represents the total number of records;

[0022] Outliers are identified using the quartile method for numeric fields DX:

[0023] Q1=P 25 (DX),Q3=P 75 (DX);

[0024] IQR = Q3 - Q1;

[0025] Outlier boundary = [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]; data points exceeding the boundary are marked as outliers and subject to manual review or rule correction.

[0026] Establish a cross-system data consistency verification mechanism, perform cross-validation on key fields, and calculate consistency metrics:

[0027]

[0028] Where m is the number of matched records, ERP i and Tax i These are the corresponding records for the ERP system and the Golden Tax system, respectively.

[0029] Further unstructured data preprocessing is performed as follows:

[0030] Preprocessing of the ticket image involves optimizing image quality through histogram equalization and denoising algorithms to improve OCR recognition accuracy.

[0031] Calculate the confidence level of the OCR recognition results. Assessment:

[0032] ;

[0033] in, For the first The recognition confidence score of each character, where N is the total number of characters; when the overall confidence score is lower than the threshold, a manual review process is triggered.

[0034] Based on template matching and regular expressions, key information fields from invoices and contracts are extracted. k :

[0035] Field k =argmax F Similarity(Template k ,F)

[0036] Among them, Templatek is the template of the k-th field, F is the candidate field region; Similarity is the template matching function;

[0037] A field format verification mechanism is established to perform format checking on the extraction result:

[0038] ;

[0039] wherein f is the extraction result; for the field that does not meet the format requirement, the edit distance algorithm is used for error correction; is the effectiveness indication function for judging whether f passes the verification or not;

[0040] The subject model is used for subject classification of the contract document, and the TF-IDF algorithm is used for identifying the key clauses of the contract.

[0041] Further, the stream data preprocessing is as follows:

[0042] Based on the business rules, the repeated transaction records are identified :

[0043]

[0044] wherein Hash(⋅) is the record hash function, t a , t b are the times of transactions r a , r b , and θ is the time threshold;

[0045] The sliding window statistical method is used to detect abnormal transactions, and the Z-score Z score is calculated:

[0046] ;

[0047] wherein 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;

[0048] When |Z score | is greater than the preset value, it is marked as a suspected abnormal transaction;

[0049] The Fourier transform is used to identify the periodic pattern of the transaction amount, and the moving average method is used to extract the trend;

[0050] A multi-dimensional index Index of enterprise Company, time Time, amount Amount, and transaction type Type is established to improve the query efficiency:

[0051] Index=f'(Company,Time,Amount,Type);

[0052] Where f' is the function for constructing the index key when performing multidimensional retrieval.

[0053] Furthermore, based on the preprocessed multi-source electricity finance and tax data, a finance and tax knowledge graph is constructed, as follows:

[0054] For the input text T=[w1,w2,...,w n ], where w n For the nth word segment; BERT uses multiple layers of Transformers to encode context-dependent semantics for each word segment:

[0055] H = BERT(T) = [h1, h2, ..., h n ];

[0056] Each h n Represents w n Deep semantic representation;

[0057] Using BERT output as input to BiLSTM improves the ability to mine local context order:

[0058] Forward LSTM:

[0059] ;

[0060] in, The output of the hidden state of the forward LSTM at the i′′ word; h i′′ Let i be the input feature vector at the i′′ position.

[0061] Inverse LSTM:

[0062] ;

[0063] in, This is the hidden state output of the inverse LSTM at position i′′;

[0064] splicing the two together:

[0065] ;

[0066] in, This is a spliced ​​output of a bidirectional LSTM;

[0067] For n word segments, each with L entity tags, use CRF layers to jointly model the optimal tag sequence y=[y1,...,y...] for the entire sentence. n ], y n The optimal tag for the nth word segment:

[0068] Scoring function :

[0069] ;

[0070] where, denotes the score of the i''th word being tagged y i′′ , obtained by a BiLSTM output followed by a linear layer; is the transition score from label y i′′ to label y i′′+1 , i.e. the label transition matrix learned by the CRF layer; n is the sequence length;

[0071] Sequence probability:

[0072] ;

[0073] where, is the probability of the optimal label sequence y given the input text T; Y T is the set of all possible label sequences given the input text T; y' is any label sequence drawn from Y T ;

[0074] The optimal label path inference is achieved by maximizing the probability of the real annotated sequence, and the entity label is output;

[0075] The structured relationship alignment based on rules and business primary keys, including enterprise ID-invoice number, bank card number, payment amount, realizes the cross-table mapping of the same business entity:

[0076] Rstruct(x,y)=I(Key x =Key y );

[0077] where, I(⋅) is an indicator function, Key x and Key y are the business primary keys of business x and business y respectively, and the directed association is obtained; Rstruct(x,y) is a structure consistency indicator function;

[0078] For unstructured text, a relationship extraction model classifier is used to extract relationship triples and map them to the graph relationship edge;

[0079] After entity disambiguation, relationship fusion rules are established;

[0080] The combined relationship is described by the path method:

[0081] Path(A,D)=A→transaction B→ticket C→money flow D;

[0082] Set up relationship credibility hierarchies, combined with data quality:

[0083] Edge_Score(e1,r,e2)=α1×Key_Match+α2×Text_Match+α3×Operation Log

[0084] Where α1, α2, and α3 are weighting parameters; Key_Match is the key consistency matching score; Text_Match is the text content similarity score; 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;

[0085] Finally, the financial and tax knowledge graph is defined as G=(V,E), where V=the set of entity nodes and E=the set of relation edges with type and attributes;

[0086] A graph database is used to implement relational storage for knowledge graphs, with self-describing attributes for nodes and edges, supporting fast path queries and multi-dimensional data retrieval.

[0087] Furthermore, a multimodal risk detection model is constructed. Based on the fiscal and tax knowledge graph and real-time multi-source power fiscal and tax data, a list of risk events and a complete data traceability path are obtained, as detailed below:

[0088] Real-time electricity and tax data from multiple sources are mapped to corresponding entities and relationships, aligned with knowledge graph nodes, and multi-features are integrated for risk label modeling and anomaly detection.

[0089] Based on the multimodal risk detection model, typical risk labels are identified, and a list of risk events is obtained;

[0090] By combining financial and tax knowledge graphs with log metadata, a chain record is formed from event triggering to the source of the data.

[0091] 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 document anomaly detection sub-model, a suspected fraudulent invoice identification sub-model, a three-document discrepancy detection sub-model, and a business fraud detection sub-model, as detailed below:

[0092] The bill anomaly detection sub-model combines CR recognition, field validation, and dual indexing analysis of bills. It marks bills with anomalies in format, number, or amount and obtains a set of abnormal bills.

[0093] The suspected fraudulent invoice identification sub-model uses a graph neural network to detect suspicious fund reflux loops. ;

[0094] Among them, Loop grepresents a loop consisting of g nodes; v1, v2,..., vg g represents a node in the graph neural network; v g represents the gth node; EdgeType(e a,a+1 ) is the type of edge e a,a+1 ;

[0095] RiskLabel(Loop) = 1, RiskLabel is used to determine whether the closed loop (Loop) belongs to the risk type label function;

[0096] Three single inconsistency detection sub-model, three single amount, subject, time, commodity item intersection consistency detection of procurement contract, invoice and bank return:

[0097] where, InconsistentTriplet is the inconsistent triplet set; c, b are two entities (such as contract, invoice, and document); r is the association relationship; A(c), 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.

[0098] Company(c), Company(b) are the company identifiers corresponding to c and b; ∨ represents logical or, and any inconsistent condition is satisfied to be inconsistent;

[0099] Business inauthenticity detection sub-model, combined with power time series and bill data, check the high mismatch degree FakeBusinessScore of procurement / production / electricity consumption and income / invoicing: ;

[0100] where, RealPower(t0, t1) is the actual electricity consumption in the time interval [t0, t1], and ReportedProd(t0, t1) is the theoretical electricity consumption converted from the declared production or production value in the interval [t0, t1];

[0101] If the FakeBusinessScore score is higher than the threshold, it is marked as suspected business inauthenticity;

[0102] According to the detection results of the multi-modal risk detection model, combine multiple labels to automatically generate the list:

[0103] ;

[0104] Wherein, the RiskEventList is a risk event list, indicating a collection of all events identified as risks; l represents an event type, L is a collection of all risk event types; Event t,l represents a specific event at time t , event type l ; RiskLabel l ( t ) determines whether the event Event t,l is a risk event at time t . If the value is 1, it indicates that the event is determined to be a risk event.

[0105] Further, combined with the financial and tax knowledge graph and log metadata, a link record from event triggering to the most source data is formed, as follows:

[0106] Using the financial and tax knowledge graph, all causal links related to the risk event are found, with e s as the starting point and e t as the end point, and the traceability path being all valid paths: ;

[0107] Wherein, determines whether the edge e i,i+1 is a relevant edge; p is a path;

[0108] The highest weight path p *: is obtained by using weight multiplication;

[0109] Wherein, is the risk weight of edge e i,i+1 ;

[0110] For each risk label determination event, the original data used, the triggering module, the reference rules and the output fields are recorded to form a complete traceability chain.

[0111] Further, according to the risk event and the traceability, a structured decision instruction set is generated based on the rule model, containing all key information and traceability links, as follows:

[0112] According to the risk event list and the traceability path data, for each risk event, the system automatically extracts structured features, which are then used as inputs to the rule model;

[0113] Load the rule model, including static rules and dynamic rule models, and match each risk event with the rules in the rule model;

[0114] The matched rules are conditionally evaluated, corresponding actions are triggered, and the triggered actions and related information are structured into decision instructions.

[0115] The present application has the following advantages:

[0116] 1. The present application uses a financial and tax knowledge graph to model core objects such as enterprises, accounts, invoices, and transactions, and supports dynamic adaptation of complex entity relationships, providing a solid and comprehensive data and knowledge base for auditing analysis.

[0117] 2. The present application uses multi-modal artificial intelligence technology to intelligently detect risks such as invoice abnormalities, suspected fake invoices, discrepancies between three documents, and unrealistic business. Based on a hybrid engine combining rules and models, it automatically outputs structured and traceable decision instructions, improving the accuracy, automation, and decision transparency of auditing.

[0118] 3. The present application uses a dynamic graph to visually and dynamically display the flow of funds between enterprises, risk events, abnormal aggregations, and traceable links, greatly enhancing the intuitiveness and interactivity of risk discovery. All analysis and decision results support full traceability and responsibility attribution, significantly improving auditing efficiency, compliance, and user experience. BRIEF DESCRIPTION OF DRAWINGS

[0119] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0120] The present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0121] REFERENCE Figure 1 In this embodiment, a digital financial and tax auditing method based on artificial intelligence is provided, including the following steps:

[0122] S1: Obtain power financial and tax multi-source data and preprocess to obtain preprocessed power financial and tax multi-source data.

[0123] S2: Construct a financial and tax knowledge graph based on the preprocessed power financial and tax multi-source data.

[0124] S3: Construct a multi-modal risk detection model, obtain a risk event list (including labels such as invoice abnormalities, suspected fake invoices, discrepancies between three documents, and unrealistic business) and a full data traceability path based on the financial and tax knowledge graph and real-time power financial and tax multi-source data.

[0125] S4: Based on the risk events and traceability, generate a structured decision instruction set based on rules and models, including all key information and traceability links.

[0126] S5 converts the structured decision-making instruction set and the tax knowledge graph into a standard relationship graph input, renders it through D3.js, and dynamically displays the capital flow between enterprises, risk events, abnormal aggregation, and traceability paths.

[0127] In this embodiment, multi-source data of power finance and taxation is acquired, including structured data, unstructured data, and stream data, as follows:

[0128] Structured data is acquired through multi-channel data extraction strategies, such as database direct connection, API interface, and file transfer, to obtain structured data from ERP systems and the Jintax system. For ERP systems, total account details, accounts payable, purchase orders, and payment documents are acquired. For the Jintax system, invoice issuance lists, input tax deduction lists, and tax return data are acquired.

[0129] Unstructured data is acquired by establishing a document management system, collecting bill images and contract documents through batch scanning, mobile phone shooting and uploading, and email attachment extraction, setting up a file monitoring mechanism to automatically identify new files and trigger processing procedures, establishing file naming standards and storage directory structures to ensure file traceability.

[0130] Stream data is acquired by establishing a real-time data transmission channel with the bank system, obtaining bank stream data through bank-enterprise direct interfaces, and setting up a message queue buffer mechanism to handle data transmission delays and network fluctuations.

[0131] In this embodiment, structured data preprocessing is performed as follows:

[0132] For structured data sets, field completeness indicators are calculated:

[0133] ;

[0134] where CI j is the completeness indicator of the jth field, I(⋅) is the indicator function, d i,j represents the jth field value of the ith record, and n' represents the total number of records.

[0135] Quartile method is used to identify outliers. For numerical fields DX:

[0136] Q1=P 25 (DX), Q3=P 75 (DX);

[0137] IQR=Q3-Q1;

[0138] where Q1 is the first quartile, Q3 is the third quartile, and IQR is the interquartile range.

[0139] Outlier boundary = [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]; data points exceeding the boundary are marked as outliers and subject to manual review or rule correction.

[0140] Establish a cross-system data consistency verification mechanism, perform cross-validation on key fields, and calculate consistency metrics:

[0141]

[0142] Where m is the number of matched records, ERP i and Tax i These are the corresponding records for the ERP system and the Golden Tax system, respectively.

[0143] In this embodiment, the unstructured data preprocessing is as follows:

[0144] Preprocessing of the ticket image involves optimizing image quality through histogram equalization and denoising algorithms to improve OCR recognition accuracy.

[0145] Calculate the confidence level of the OCR recognition results. Assessment:

[0146] ;

[0147] in, For the first The recognition confidence score of each character, where N is the total number of characters; when the overall confidence score is lower than the threshold, a manual review process is triggered.

[0148] Based on template matching and regular expressions, key information fields from invoices and contracts are extracted. k :

[0149] Field k =argmax F Similarity(Template k ,F)

[0150] Among them, Template k Let F be the template for the k-th field, F be the candidate field region, and Similarity be the template matching function.

[0151] Establish a field format validation mechanism to verify the format of the extracted results:

[0152] ;

[0153] Where f is the extraction result; This is a validity indicator function to determine whether f has passed the validation.

[0154] For fields that do not meet the format requirements, the edit distance algorithm is used for error correction;

[0155] The contract document is classified by topic model, and the key clauses of the contract are identified based on TF-IDF algorithm;

[0156] In this embodiment, the flow data is preprocessed, specifically as follows:

[0157] Based on business rules, repeated transaction records are identified :

[0158]

[0159] Wherein, Hash(⋅) is the record hash function, t a , t b are the time of transaction r a , r b , and θ is the time threshold;

[0160] The sliding window statistical method is used to detect abnormal transactions, and the Z-score Z score :

[0161] ;

[0162] Wherein, 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;

[0163] When |Z score | is greater than the preset value, it is marked as a suspected abnormal transaction;

[0164] Fourier transform is used to identify the periodic pattern of transaction amount, and moving average method is used to extract the trend;

[0165] A multi-dimensional index Index of enterprise Company, time Time, amount Amount, and transaction type Type is established to improve the query efficiency:

[0166] Index=f′(Company,Time,Amount,Type);

[0167] Wherein, f′ is the key function used for multi-dimensional retrieval when constructing the index.

[0168] In this embodiment, according to the preprocessed power finance and tax multi-source data, a finance and tax knowledge graph is constructed, specifically as follows:

[0169] For input text T=[w1,w2,...,w n ], wherein, wn For the nth word; BERT uses multiple layers of Transformers to encode the context-dependent semantics of each word:

[0170] H = BERT(T) = [h1, h2, ..., h n ];

[0171] Each h n Represents w n Deep semantic representation;

[0172] Using BERT output as input to BiLSTM improves the ability to mine local context order:

[0173] Forward LSTM:

[0174] ;

[0175] in, The output of the hidden state of the forward LSTM at the i′′ word; h i′′ Let i be the input feature vector at the i′′ position.

[0176] Inverse LSTM:

[0177] ;

[0178] in, This is the hidden state output of the inverse LSTM at position i′′;

[0179] splicing the two together:

[0180] ;

[0181] in, This is a spliced ​​output of a bidirectional LSTM;

[0182] For n word segments, each with L entity tags, use CRF layers to jointly model the optimal tag sequence y=[y1,...,y...] for the entire sentence. n ], y n The optimal tag for the nth word segment:

[0183] Scoring function :

[0184] ;

[0185] in, This indicates that the i′′-th word is marked y. i′′ The score is obtained from the wiring layer after the BiLSTM output; For from label y i′′ Transfer to label yi′′+1 transition score of the CRF layer, i.e., the label transition matrix learned by the CRF layer; n is the sequence length;

[0186] Sequence probability:

[0187]

[0188] wherein, is the probability of the optimal label sequence y under the given input text T; Y T is the set of all possible label sequences under the given input text T; y' is any label sequence extracted from Y T ;

[0189] By maximizing the probability of the real annotation sequence, the optimal label path inference is realized, and the entity labels (enterprise, supplier, commodity, tax rate, amount, etc.) are outputted;

[0190] Typical label paths include: transaction relationship (enterprise-supplier, enterprise-commodity), bill ownership (enterprise-bill, supplier-bill), fund flow direction (paying account-receiving account), cross-document / cross-table association (contract-invoice-serial number);

[0191] Structured relationship alignment based on rules and business primary keys, including enterprise ID-invoice number, bank card number, payment amount, realizing cross-table mapping of the same business entity:

[0192] Rstruct(x, y) = I(Key x = Key y );

[0193] wherein, I(·) is an indicator function, Key x and Key y are the business primary keys of the business x and the business y, respectively, to obtain a directed association; Rstruct(x, y) is a structure consistency indicator function;

[0194] For unstructured text (such as invoice remarks, contract terms, etc.), a relationship extraction model (BERT+Span) classifier is used to extract relationship triples and map them to graph relationship edges;

[0195] Relation = (e1, r, e2); wherein Relation represents a relationship, e1 and e2 are entities, and r is an associated relationship; for example: “A pays C amount to B” → (A, payment, B);

[0196] After entity disambiguation, relationship fusion rules are established, such as: a bill is associated with multiple businesses / invoices / accounts; fund inflow and outflow loop tracking (multi-step path analysis);

[0197] ​The combination relationship is described by the path method:

[0198] Path(A, D) = A→Trade B→Invoice C→Fund Flow D;

[0199] The relationship credibility is set to be layered, and the data quality is combined:

[0200] Edge_Score(e1, r, e2) = a1 x Key_Match + a2 x Text_Match + a3 x Operation Log

[0201] Wherein, a1, a2, a3 are weighted parameters, which are configured according to actual business scenarios; Key_Match is the consistency matching score of the primary key (such as ID, number, account, etc.); Text_Match is the similarity score of the text content (such as name, description, abstract); Edge_Score(e1, r, e2) represents the matching score of the edge (e1, r, e2);

[0202] Node (entity): enterprise, supplier, commodity, invoice, contract, bank account, …;

[0203] Edge (relationship): supply, purchase, use, hold, invoice, payment, fund flow, etc.

[0204] Finally, the tax knowledge graph is defined as G=(V,E), wherein V=entity node set, E=relationship edge set with type and attribute;

[0205] The graph database (Neo4j, GraphDB, etc.) is used to realize the relational storage of the knowledge graph, the nodes and edges are self-described with attributes, and the fast path query and multi-dimensional data retrieval are supported.

[0206] In the embodiment, a multi-modal risk detection model is constructed, according to the tax knowledge graph and real-time power tax multi-source data, a risk event list (including invoice abnormality, suspected false opening, three-unmatched, business unreal, etc. label) and full-process data traceability path are obtained, and the specific process is as follows:

[0207] The real-time power tax multi-source data is mapped to the corresponding entity and relationship, aligned with the knowledge graph node, and fused with multiple features to model the risk label and detect the abnormality;

[0208] According to the multi-modal risk detection model, the typical risk label is identified, and the risk event list is obtained;

[0209] The tax knowledge graph and the log metadata are combined to form a link record from event triggering to the most source data.

[0210] In this embodiment, according to the multi-modal risk detection model, typical risk label recognition is performed, and a risk event list is obtained, as follows: the multi-modal risk detection model includes a bill abnormality detection sub-model, a suspected false invoice identification sub-model, a three-single inconsistency detection sub-model, and a business inaccuracy detection sub-model, as follows:

[0211] The bill abnormality detection sub-model combines CR identification, field verification, and bill double-index analysis, and marks when encountering format, number, and amount abnormalities, to obtain an abnormal bill set;

[0212] The suspected false invoice identification sub-model uses a graph neural network to detect suspicious fund reflux loops: ;

[0213] wherein, Loop g represents a loop composed of g nodes; v1, v2,..., v g are nodes in the graph neural network; v g represents the gth node; EdgeType(e a,a+1 ) is the type of edge e a,a+1 ;

[0214] For a risk loop with a large amount, set RiskLabel(Loop)=1, and RiskLabel is a label function for determining whether the closed loop (Loop) belongs to a risk type;

[0215] The three-single inconsistency detection sub-model detects the amount, subject, time, and commodity item intersection consistency of the procurement contract, invoice, and bank return single: ;

[0216] wherein, InconsistentTriplet is a set of inconsistent triplets; c and b are two entities (such as contracts, invoices, and documents); r is a related relationship; A(c), 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.

[0217] Company(c), Company(b) are the company identifiers corresponding to c and b; ∨ represents logical or, and any inconsistent condition is satisfied to be determined as inconsistent;

[0218] The business inaccuracy detection sub-model jointly checks the power time sequence and bill data to check the high mismatch degree FakeBusinessScore of procurement / production / electricity consumption and income / invoicing: ;

[0219] Wherein, RealPower(t0, t1) is the actual power consumption in the time interval [t0, t1], ReportedProd(t0, t1) is the theoretical power consumption converted from the production or value declared by the enterprise in the interval [t0, t1];

[0220] If the FakeBusinessScore score is higher than the threshold, it is marked as suspected business fraud;

[0221] According to the detection result of the multi-modal risk detection model, a plurality of labels are combined to automatically generate a list:

[0222] ;

[0223] Wherein, RiskEventList is a risk event list, indicating a collection of all events identified as risks; l represents the event type, L is a collection of all risk event types; Event t,l represents a specific event under the time t , event type l ; RiskLabel l ( t ) determines whether the event Event t,l is a risk event at time t, if its value is 1, it means that it is determined to be a risk event.

[0224] In this embodiment, the financial and tax knowledge graph and the log metadata are combined to form a link record from event triggering to the most source data, as follows:

[0225] Using the financial and tax knowledge graph, all causal links related to the risk event are found, taking e s as the starting point, e t as the end point, and the source path as all valid paths : ;

[0226] Wherein, determines whether the edge e i,i+1 is a relevant edge; p is a path;

[0227] The highest weight path is obtained by multiplying the weights p *: ;

[0228] Wherein, is the risk weight of edge e i,i+1 ;

[0229] For each risk labeling decision event, the original data used, trigger module, reference rules and output fields are recorded to form a complete traceability chain.

[0230] In this embodiment, according to the risk event and traceability, a structured decision instruction set is generated based on the rule model, containing all key information and traceability links, as follows:

[0231] According to the risk event list and traceability path data, for each risk event, the system automatically extracts its structured features (such as subject type, data field, abnormal type, amount interval, time, frequency, link level, etc.), and then uses them as inputs to the rule model;

[0232] Load the rule model, including static rules and dynamic rule models, and match each risk event with the rules in the rule model;

[0233] The rule model includes static rules and dynamic rule models:

[0234] Static rules: hard decision rules based on existing regulations, enterprise internal control requirements and business logic processes. For example, "if the same enterprise has multiple transactions with the same amount and the same company as the counterparty within a day, it is marked as high risk." The Drools engine is used to execute 5000+ tax rules;

[0235] Dynamic rules: adaptive rules based on historical data, machine learning model inference or real-time monitoring. For example, "if the risk score is greater than 0.8 and there is a recent network relationship graph showing a money flow path, it is determined to be abnormal money flow."

[0236] Conditionally evaluate the matched rules to trigger corresponding actions, and structure the triggered actions and related information into decision instructions.

[0237] Each instruction includes the following key information:

[0238] Instruction ID: a unique identifier for this decision instruction;

[0239] Risk type: such as "fake invoicing", "three single not match", "money flow loop";

[0240] Trigger subject / entity: such as the involved enterprise, bank account, specific invoice number;

[0241] Suggested operation: such as "report", "freeze account", "manual review", "generate audit report", etc.;

[0242] Risk priority / confidence: to help subsequent automatic shunting and processing;

[0243] Traceability link ID and detailed path: clearly record the full-chain data path of the risk event, the entities involved, and the original data source, so as to be visualized, accountable, or reviewed later.

[0244] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage media, etc.) having computer usable program code embodied therein.

[0245] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application. It should be understood that each flow and / or block in the flowchart and / or block diagrams, as well as a combination of flows and / or blocks in the flowchart 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 apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus create means for implementing the functions specified in the flowchart and / or block diagrams block or blocks. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the function specified in the flowchart block or blocks.

[0246] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including instruction means, which implements the functions specified in the flowchart and / or block diagrams block or blocks. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the function specified in the flowchart block or blocks.

[0247] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide means for implementing the functions specified in the flowchart and / or block diagrams block or blocks. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the function specified in the flowchart block or blocks.

[0248] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application to other forms, and any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution content of the present application, and according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. An artificial intelligence-based digital finance and tax auditing method, characterized in that, The method comprises the following steps: S1: Obtain power finance and tax multi-source data and pre-process to obtain pre-processed power finance and tax multi-source data; Specifically as follows: Structured data adopts a multi-channel data extraction strategy to obtain structured data of ERP systems and Jinsan systems through database direct connection, API interface and file transmission; For the ERP system, total account details, accounts payable, purchase orders and payment documents are obtained; For the Jinsan system, invoice issuing lists, input tax deduction lists and tax return table data are obtained; Unstructured data is obtained by establishing a document management system, collecting bill images and contract documents through batch scanning, mobile phone shooting and uploading, and email attachment extraction; A file monitoring mechanism is set up to automatically identify new files and trigger the processing flow; A file naming specification and storage directory structure are established to ensure file traceability; Flow data is obtained by establishing a real-time data transmission channel with the bank system, obtaining bank flow data through bank-enterprise direct interface, and setting up a message queue buffer mechanism to handle data transmission delays and network fluctuations; S2: Construct a finance and tax knowledge graph according to the pre-processed power finance and tax multi-source data; S3: Construct a multi-modal risk detection model, obtain a risk event list and a full-process data traceability path according to the finance and tax knowledge graph and real-time power finance and tax multi-source data; Specifically as follows: Map the real-time power finance and tax multi-source data to the corresponding entities and relationships, align them with the knowledge graph nodes, and fuse multiple features to model risk labels and perform anomaly detection; According to the multi-modal risk detection model, identify typical risk labels and obtain a risk event list; Combine the finance and tax knowledge graph with the log metadata to form a link record from event triggering to the most source data; S4: According to the risk event and traceability, generate a structured decision instruction set based on a rule model, which contains all key information and traceability links; Specifically as follows: According to the risk event list and traceability path data, for each risk event, the system automatically extracts structured features, which are then used as inputs to the rule model; Load the rule model, which includes static rules and dynamic rules, and match each risk event with the rules in the rule model; Conditionally evaluate the matched rules to trigger corresponding actions, and structure the triggered actions and related information into decision instructions; S5: Convert the structured decision instruction set and the finance and tax knowledge graph into a standard relationship graph input, render it through D3.js, and dynamically display the inter-enterprise capital flow, risk events, abnormal aggregation and traceability path.

2. The artificial intelligence-based digitalized fiscal audit method according to claim 1, characterized in that, Structured data preprocessing, specifically as follows: For structured data sets, calculate the field integrity index: ； wherein CI j is the integrity index of the jth field, I(·) is an indicator function, d i,j denotes the jth field value of the ith record, and n' denotes the total number of records. Use the quartile method to identify outliers for numerical fields DX: Q1 = P 25 (DX), Q3 = P 75 (DX); IQR = Q3 - Q1; Where Q1 is the first quartile, Q3 is the third quartile, and IQR is the interquartile range; The outlier boundary is [Q1-1.5*IQR, Q3+1.5*IQR]; data points outside the boundary are marked as outliers and manually reviewed or corrected; Establish a cross-system data consistency verification mechanism to cross-verify key fields and calculate the consistency index: ； wherein m is the number of matching records, ERP i and Tax i are the corresponding records of the ERP system and the Tax system, respectively. 3.The AI-based digitalized fiscal audit method according to claim 1, wherein, Unstructured data preprocessing, specifically as follows: Preprocessing of the ticket image involves optimizing image quality through histogram equalization and denoising algorithms to improve OCR recognition accuracy. Calculate the confidence level of the OCR recognition results. Assessment: ； wherein, is the recognition confidence of the th character, N is the total number of characters; when the overall confidence is lower than a threshold, an artificial review process is triggered; Extracting key information from invoices and contracts based on template matching and regular expressions k : Field k = argmax F Similarity(Template k , F) wherein Template k is a template for the kth field, F is a candidate field region; Similarity is a template matching function; Establish a field format validation mechanism to verify the format of the extracted results: ； Wherein, f is the extraction result; for the field that does not meet the format requirement, the edit distance algorithm is used for error correction; The validity indication function for judging whether f passes the verification; The contract documents are classified by topic model and key contract clauses are identified based on the TF-IDF algorithm.

4. The artificial intelligence-based digitalized fiscal audit method according to claim 1, characterized in that, Stream data preprocessing, specifically as follows: Identifying duplicate transaction records based on business rules : ； wherein Hash(·) is a record hash function, t a , t b are the times of transactions r a , r b , respectively, and θ is a time threshold; Anomalies are detected using a sliding window statistical method, calculating Z-scores Z score : ； where Amount t is the transaction amount at time t; μ window and σ window are the mean and standard deviation of the transaction amounts within the current sliding window, respectively. When |Z score | is greater than a preset value, it is marked as a suspected abnormal transaction; Fourier transform is used to identify periodic patterns in transaction amounts, and moving average is used to extract trends. Create multi-dimensional indexes based on Company, Time, Amount, and Transaction Type to improve query efficiency: Index=f'(Company,Time,Amount,Type); Where f' is the function for constructing the index key when performing multidimensional retrieval.

5. The artificial intelligence-based digitalized fiscal audit method according to claim 1, wherein, The process of constructing a tax knowledge graph based on preprocessed multi-source electricity tax data is as follows: For the input text T=[w1,w2,...,w n ], where w n For the nth word segment; BERT uses multiple layers of Transformer to encode context-related semantics for each word segment: H = BERT(T) = [h1, h2,..., h n ]; each h n representative w n deep semantic representation; Using BERT output as input to BiLSTM improves the ability to mine local context order: Forward LSTM: ； wherein, is the hidden state output of the forward LSTM at the i" word; h i′′ is the input feature vector at the i" position; Inverse LSTM: ； wherein, is the hidden state output of the reverse LSTM at the i" position; splicing the two together: ； wherein, is the concatenated output of the bidirectional LSTM; For n word segments, each with L entity tags, use CRF layers to jointly model the optimal tag sequence y=[y1,...,y...] for the entire sentence. n ], y n The optimal tag for the nth word segment: Scoring function : ； in, This indicates that the i′′-th word is marked y. i′′ The score is obtained from the wiring layer after the BiLSTM output; For from label y i′′ Transfer to label y i′′+1 The transition score is the label transition matrix learned by the CRF layer; n is the sequence length. Sequence probability: ； wherein, P(y | T) is the probability of the optimal label sequence y given the input text T; and T Y is the set of all label sequences given the input text T; y' is any label sequence drawn from Y T Y is the set of all label sequences given the input text T; y' is any label sequence drawn from Y By maximizing the probability of the true labeled sequence, the optimal label path is inferred and the entity label is output. Based on structured relationships aligned according to rules and business primary keys, including enterprise ID-invoice number, bank card number, and payment amount, cross-table mapping of the same business entity is achieved: Rstruct(x, y) = I(Key x = Key y ); where I(·) is an indicator function, Key x and Key y are the business primary keys of business x and business y, respectively, resulting in a directed association; Rstruct(x, y) is a structure consistency indicator function; For unstructured text, a relation extraction model classifier is used to extract relation triples and map them to graph relation edges; After entity disambiguation, relationship fusion rules are established; Composition relationships are described using the path method: Path(A,D) = A→Transaction B→Bill C→Funds Flow D; Set up relationship credibility hierarchies, combined with data quality: Edge_Score(e1,r,e2)=α1×Key_Match+α2×Text_Match+α3×Operation Log Where α1, α2, and α3 are weighting parameters; Key_Match is the key consistency matching score; Text_Match is the text content similarity score; 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=the set of entity nodes and E=the set of relation edges with type and attributes; A graph database is used to implement relational storage for knowledge graphs, with self-describing node and side attribute attributes, supporting fast path query and multi-dimensional data retrieval.

6. The artificial intelligence-based digitalized fiscal audit method according to claim 1, wherein, The process involves identifying typical risk labels and obtaining a list of risk events based on a multimodal risk detection model, as detailed below: The multimodal risk detection model includes sub-models for detecting abnormal invoices, identifying suspected fraudulent invoices, detecting discrepancies between invoices, documents, and payment records, and detecting fraudulent business transactions, as detailed below: The bill anomaly detection sub-model combines CR recognition, field validation, and dual indexing analysis of bills. It marks bills with anomalies in format, number, or amount and obtains a set of abnormal bills. The suspected virtual invoice identification sub-model adopts a graph neural network to detect a suspected fund backflow loop: ; wherein Loop g represents a loop consisting of g nodes; v1, v2,..., vg g are nodes in a graph neural network; v g represents the gth node; EdgeType(e a,a+1 ) is the type of edge e a,a+1 . Set RiskLabel(Loop) = 1 for the risk loop with a large proportion of risk amount, and RiskLabel is a label function for determining whether the closed loop belongs to the risk type; Three single inconsistency detection sub-model, the procurement contract, invoice and bank return single three single amount, subject, time, commodity item intersection consistency detection: ; wherein InconsistentTriplet is a set of inconsistent triplets; c, b are two entities; r is a relation; A(c), A(b) are the amount or volume fields of c and b; is an amount or volume difference threshold; T(c), is a time field of c and b; θ t is a time difference threshold; Company(c), Company(b) are the company identifiers of c and b; ∨ represents logical or, and any inconsistent condition is satisfied to be inconsistent; The business misrepresentation detection sub-model, combining electricity time-series and invoice data, verifies the degree of mismatch between procurement / production / electricity consumption and revenue / invoicing (FakeBusinessScore). ; Where 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 reported production or output value in the interval [t0, t1]; If the FakeBusinessScore score is higher than the threshold, mark it as suspected business fraud; According to the detection results of the multi-modal risk detection model, combine multiple labels to automatically generate the list: ； wherein RiskEventList is a list of risk events, indicating a collection of all events identified as risks; l represents an event type, L is a collection of all risk event types; Event t,l represents a specific event at time t , event type l ; RiskLabel l ( t ) a label indicating whether an event Event t,l at time t is a risk event, and if its value is 1, it indicates that the event is determined to be a risk event.

7. The artificial intelligence-based digitalized fiscal audit method according to claim 1, characterized in that, The combination of financial and tax knowledge graph and log metadata forms a link record from event triggering to the most source data, as follows: Using the tax knowledge graph, find all causal links related to the risk event, take e s as the starting point, e t as the end point, and the traceable path is all valid paths: ; wherein, determining edge e i,i+1 whether it is a relevant edge; p is a path; Obtaining the highest weight path by weight multiplication p ;​ wherein is the risk weight for edge e i,i+1 of edge e For each risk label determination event, record the original data used, triggering module, reference rules, and output fields to form a complete traceability chain.

Images