Accounting data intelligent processing method and system for enterprise financial audit

Abstract

The invention discloses an accounting data intelligent processing method and system for enterprise financial auditing, and relates to the technical field of accounting data intelligent processing, and the method comprises the steps: obtaining multi-mode enterprise financial data, and carrying out the preprocessing; carrying out multi-modal semantic understanding analysis on the unstructured text and image data; constructing an enterprise financial space-time knowledge graph containing time attributes; inputting into an anomaly analysis model, extracting spatial structure characteristics of the financial entity in the topological network, and extracting dynamic characteristics of the financial relationship evolved along with the time sequence; identifying an abnormal source, evaluating a systematic risk value and generating an abnormal propagation path; and integrating to generate a final audit report. According to the method, structured and bill images are fused, identifiers and time calibers are unified, abnormal source and propagation are positioned based on the space-time knowledge graph, closed-loop counter-knock and cross-period anomalies are identified, the auditing accuracy and coverage rate are remarkably improved, the workload of false report, missing report and manual recheck is reduced, and a traceable structured report is quickly generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent processing of accounting data, in particular to an accounting data intelligent processing method and system for enterprise financial audit. BACKGROUND

[0002] In recent years, enterprise financial audit has gradually evolved from manual sampling and rule checking to data-driven and intelligentization. With the popularity of enterprise resource planning systems, electronic invoices, image filing and document management systems, accounting data presents the multi-modal characteristics of structured account books and unstructured bills / contracts coexisting. The audit-oriented technical system has developed from early report checking, keyword search and threshold warning to intelligent methods represented by optical character recognition, layout understanding, knowledge graph, graph neural network and time series modeling, which can identify risks and verify compliance on a larger data volume and a finer time granularity. At the same time, the digitalization of supervision and the sharing of finance have gradually unified the data standards and interfaces, providing a computable data basis for intelligent audit.

[0003] However, the existing technology still mainly focuses on single-modal or structured account book-centered analysis, and lacks deep semantic understanding of unstructured text and images. There is a lack of unified identification, time scope and consistency checking between cross-modal and structured data, leading to misalignment of entity alignment and evidence correlation. Most anomaly detection methods rely on static thresholds and rule triggers, and cannot depict the diffusion path of anomalies and systemic risks in the graph structure and time dimension. The traceability and interpretability of risk evidence are limited, making it difficult to form a closed loop from "point anomaly" to "link evidence". The existing process is mostly manually patched and reported, making it difficult to automatically generate structured audit reports under a unified evidence chain and drive subsequent sampling and review resource allocation.

[0004] Therefore, the current technology cannot achieve the goal of simultaneously achieving results in multi-modal data semantic alignment, spatiotemporal propagation modeling and executable report solidification, limiting the landing of large-scale, low-false alarm and verifiable enterprise financial intelligent audit. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] Therefore, the technical problem solved by the present application is that the existing accounting data intelligent processing / intelligent audit method lacks semantic understanding and entity / relationship extraction capabilities for unstructured text and images, and lacks alignment and consistency checking between cross-modal and structured data, leading to incomplete evidence chain, lack of abnormal propagation modeling and quantitative risk assessment mechanism based on spatiotemporal knowledge graph, and how to automatically generate structured audit reports under the constraint of traceable evidence chain and support audit resource allocation.

[0007] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the embodiments of the present application provide an accounting data intelligent processing method for enterprise financial audit, comprising: acquiring multi-modal enterprise financial data, and performing standardized preprocessing on the acquired data to form a preprocessed data set; the multi-modal enterprise financial data includes structured database transaction data and unstructured text and image data; Performing multi-modal semantic understanding analysis on the unstructured text and image data in the preprocessed data set, and extracting financial entities, relationship attributes and key event information therefrom; Associating and fusing the extracted financial entities, relationship attributes and key event information with the structured database transaction data in the preprocessed data set to construct an enterprise financial spatio-temporal knowledge graph containing time attributes; Inputting the enterprise financial spatio-temporal knowledge graph into an anomaly analysis model, extracting spatial structure features of financial entities in a topological network using a graph convolution network model, and extracting dynamic features of financial relationships evolving over time sequences using a time convolution model; Based on the obtained spatial structure features and dynamic features, calculating the anomaly probability of the risk assessment layer calculation node and the propagation weight of the edge through the attention mechanism, identifying the abnormal source, evaluating the systematic risk value and generating the abnormal propagation path; Based on the abnormal source, the systematic risk value and the abnormal propagation path, integrating to generate a final audit report.

[0008] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, the multi-modal enterprise financial data includes structured database transaction data and unstructured text and image data. The structured database transaction data includes accounting voucher data, total ledger data, detailed ledger data, subject balance table data and financial statement data. The unstructured text and image data includes invoice image data, purchase and sale contract text data, bank document image data, business voucher scan data and audit report text data.

[0009] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, the standardized preprocessing of the acquired data includes data cleaning, format unification, key field extraction and time stamp standardization of the structured database transaction data. The unstructured text and image data is subjected to optical character recognition, image sharpness enhancement, text encoding unification and time information extraction and standardization of unstructured data.

[0010] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, wherein: the multi-modal semantic understanding analysis includes: analyzing the text data in the pre-processed data set, performing named entity recognition based on a pre-trained financial field natural language processing model, and extracting financial entities including company name, personnel, amount, date, and subject; The text data in the pre-processed data set is analyzed, the relationship extraction is performed based on a pre-trained financial field natural language processing model, and the association relationship between the financial entities is extracted; The image data in the pre-processed data set is analyzed, the key information is located and extracted based on a convolutional neural network and an optical character recognition technology, and entity attribute information including invoice code, invoice number, invoice date, amount, and purchase / sale party information is identified; The features of the financial entities, the association relationship and the entity attribute information are fused by using an attention mechanism to realize collaborative understanding and semantic alignment of multi-modal data in the same business scenario; Based on a rule engine and a deep learning model, key event information including large amount of fund flow, abnormal transaction mode, associated party transaction and important contract clause change event is detected and extracted from the fused features.

[0011] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, wherein: the construction of the enterprise financial space-time knowledge graph containing time attributes includes: performing entity alignment between the financial entities and corresponding entities in the structured database transaction data in the pre-processed data set based on a unified identifier to obtain an aligned entity set; The association relationship is associated and mapped with the transaction records in the structured database transaction data to obtain a mapped association relationship set; The entity attribute information is attribute fused with the corresponding entities in the aligned entity set to obtain an attribute-rich entity set; The key event information is associated with related entities and relationships in the attribute-rich entity set and the mapped association relationship set to obtain an event-entity-relationship association network; Based on a graph database construction technology, the attribute-rich entity set is taken as a graph node, the mapped association relationship set and the association relationships contained in the event-entity-relationship association network are taken as graph edges, a knowledge graph basic framework is constructed, and time attributes extracted from the structured database transaction data and the key event information are added to each node and edge; The knowledge graph basic framework is completed in potential correlation by a knowledge reasoning algorithm, and an enterprise financial space-time knowledge graph containing entity dimension, relationship dimension, event dimension and time dimension is generated.

[0012] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, the method comprises the following steps: the spatial structure features of the financial entities in the topological network are extracted by using a graph convolution network model, and the dynamic features of the financial relationship evolution along the time sequence are extracted by using a time convolution model. The node feature matrix and the adjacency matrix are input into the graph convolution network model for spatial feature extraction, the spatial dependence relationship of the financial entities within the multi-hop neighbor range is captured through multi-layer graph convolution operation, and the spatial structure features of the financial entities in the topological network are extracted. The enterprise financial space-time knowledge graph is time series sliced, and the feature change sequence of each entity and relationship is extracted according to the time dimension, and the time series feature data is extracted. The time series feature data is input into the time convolution model for dynamic feature extraction, the evolution law of the financial relationship on different time scales is captured through dilated convolution operation, and the dynamic features of the financial relationship evolution along the time sequence are extracted.

[0013] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, the method comprises the following steps: the spatial structure features and the dynamic features are fused at the feature level to obtain space-time fusion features. The space-time fusion features are input into an attention mechanism, the attention weight of each node in the graph and the propagation weight of each edge are calculated, and a node attention weight matrix and an edge propagation weight matrix are obtained. The node attention weight matrix and the edge propagation weight matrix are input into a risk assessment layer, the abnormal probability of each node and the risk propagation intensity of each edge are calculated through a multi-layer neural network, and a node abnormal probability distribution and an edge risk propagation intensity distribution are obtained. Based on the node abnormal probability distribution, the nodes with an abnormal probability exceeding a preset threshold are identified as abnormal sources. Based on the edge risk propagation intensity distribution and the node abnormal probability distribution, the system risk value is calculated through a risk propagation model, and the overall risk level is evaluated. According to the edge risk propagation intensity distribution and the node abnormal probability distribution, a path search algorithm is used to generate an abnormal propagation path starting from the abnormal source. The risk analysis result containing the abnormal source identification, the system risk value and the abnormal propagation path is output.

[0014] As a preferred scheme of the accounting data intelligent processing method for enterprise financial audit, wherein: the integration to generate the final audit report comprises, based on the abnormal source identification, extracting corresponding abnormal entity information and associated transaction records, and generating detailed audit abnormal clues; Based on the systematic risk value, the risk level is determined through a risk level division model, and a risk level evaluation result is generated; Based on the abnormal propagation path, an abnormal propagation path graph is constructed through a visualization engine, and a visualized propagation path is generated; The audit abnormal clues, the risk level evaluation result and the visualized propagation path are integrated to generate a structured audit report; A timestamp, an audit range identifier and a model version information are added to the structured audit report to form a final audit report.

[0015] In a second aspect, the embodiments of the present application provide an accounting data intelligent processing system for enterprise financial audit, comprising: A multi-modal acquisition and cleaning module: acquires multi-modal enterprise financial data, and performs standardized preprocessing on the acquired data to form a preprocessed data set; the multi-modal enterprise financial data includes structured database transaction data and unstructured text and image data; A semantic analysis and extraction module: performs multi-modal semantic understanding analysis on the unstructured text and image data in the preprocessed data set, and extracts financial entities, relationship attributes and key event information therefrom; A space-time graph construction module: associates and fuses the extracted financial entities, relationship attributes and key event information with the structured database transaction data in the preprocessed data set, and constructs an enterprise financial space-time knowledge graph containing time attributes; A space-time feature extraction module: inputs the enterprise financial space-time knowledge graph into an abnormal analysis model, uses a graph convolution network model to extract spatial structure features of financial entities in a topological network, and uses a time convolution model to extract dynamic features of financial relationships evolving with time sequences; A risk propagation analysis module: based on the obtained spatial structure features and dynamic features, calculates abnormal probabilities of nodes and propagation weights of edges through an attention mechanism and a risk evaluation layer, identifies abnormal sources, evaluates systematic risk values and generates abnormal propagation paths; An audit report generation module: based on the abnormal sources, the systematic risk values and the abnormal propagation paths, integrates to generate a final audit report.

[0016] The application can recognize closed-loop pairs, rapid pairs and cross-period abnormalities that are difficult to find by traditional threshold and single-mode methods from multi-hop association and multi-scale fluctuations, realize abnormal source positioning, diffusion path description and systematic risk quantitative evaluation, and improve the explainability and robustness of abnormal judgment based on attention mechanism superimposed business consistency constraints (subject name, amount caliber, time proximity), significantly reduce false positives and false negatives caused by inconsistent caliber, OCR misidentification or master data differences, and ensure that the conclusion is reviewable and traceable by attaching source location, confidence and rule hit information to each entity, relationship and event. On this basis, the risk analysis results are automatically solidified into structured audit reports and associated evidence lists, which can directly drive the resource arrangement of sampling and review, shorten the closed-loop cycle from risk identification to report generation, improve audit coverage and accuracy, reduce manual review workload and implementation cost, and have good cross-industry and cross-account set migration and engineering landing. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings. Figure 1 A whole flow chart of an accounting data intelligent processing method for enterprise financial audit is provided for the first embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings in the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0019] Embodiment 1, refer to Figure 1 For an embodiment of the present application, an accounting data intelligent processing method for enterprise financial audit is provided, which comprises: S1: acquiring multi-modal enterprise financial data, and performing standardization preprocessing on the acquired data to form a preprocessed data set; the multi-modal enterprise financial data comprises structured database transaction data and unstructured text and image data.

[0020] The multi-modal enterprise financial data includes structured database transaction data and unstructured text and image data; The structured database transaction data includes accounting voucher data, general ledger data, detailed ledger data, subject balance sheet data, and financial statement data. The unstructured text and image data includes invoice image data, purchase and sale contract text data, bank document image data, business voucher scan data, and audit report text data.

[0021] The structured database transaction data is subjected to data cleaning, format unification, key field extraction, and timestamp standardization. The unstructured text and image data is subjected to optical character recognition, image sharpness enhancement, text encoding unification, and unstructured data time information extraction and standardization.

[0022] It should be noted that in the multi-modal enterprise financial data, the structured database transaction data is derived from an enterprise resource planning (ERP) system or a financial software backend database, and is extracted through a structured query language (SQL) interface or an application programming interface (API). The accounting voucher data includes fields such as debit and credit subject codes, amounts, transaction dates, etc. The general ledger data includes this period's debit total, credit total, and end-of-period balance for each subject. The detailed ledger data provides detailed records of each transaction under a subject. The subject balance sheet data reflects the balance information of each subject at a certain point in time. The financial statement data includes balance sheets, profit and loss statements, and cash flow statements, and the data is stored in two-dimensional table form, with field names and data types complying with the enterprise accounting standards.

[0023] The preprocessing of structured data includes: in the data cleaning stage, missing values, outliers, and duplicate records are identified and processed, for example, using the PauTa Criterion to remove abnormal amount records that deviate significantly from the mean; in the format unification stage, the date format in different data sources is unified to ISO 8601 standard (YYYY-MM-DD), and the amount unit is unified to yuan; in the key field extraction stage, fields related to financial risk analysis are selected, such as "customer number", "supplier number", "transaction amount", etc., to construct a feature vector; in the timestamp standardization stage, all transaction times are converted to the same time zone timestamp, and time features such as year, month, day, and week are extracted to form regular time series data.

[0024] Unstructured text and image data are acquired through enterprise document management systems or scanning devices. Invoice image data and bank statement image data are converted into machine-readable text through optical character recognition (OCR) technology, specifically using a recognition model based on convolutional neural networks (CNN) and long short-term memory networks (LSTM), such as the CRNN (Convolutional Recurrent Neural Network) architecture. Image clarity enhancement uses histogram equalization and non-local mean denoising algorithms to improve the signal-to-noise ratio (SNR) of low-quality scanned images. Text encoding unification converts recognized text into UTF-8 encoding to avoid issues with garbled characters. Time information extraction for unstructured data uses named entity recognition (NER) technology to identify date entities from contract texts or audit reports and standardizes them into a unified format.

[0025] The preprocessed dataset is finally stored in the form of a feature matrix, where each row represents a sample (such as a transaction or a document), and each column represents a feature variable (such as the amount, date, transaction counterpart, etc.). Missing values are filled in with interpolated or default values to ensure the integrity and consistency of the dataset, providing high-quality input for subsequent financial risk analysis models.

[0026] S2: Perform multi-modal semantic understanding analysis on the unstructured text and image data in the preprocessed dataset to extract financial entities, relationship attributes, and key event information.

[0027] Parse the text data in the preprocessed dataset and perform named entity recognition based on a pre-trained financial domain natural language processing model to extract financial entities including company names, personnel, amounts, dates, and subjects. Parse the text data in the preprocessed dataset and perform relationship extraction based on a pre-trained financial domain natural language processing model to extract the association relationships between the financial entities. Parse the image data in the preprocessed dataset and perform key information positioning and extraction based on convolutional neural networks and optical character recognition technology to identify entity attribute information including invoice code, invoice number, invoice date, amount, and buyer-seller information. Use attention mechanisms to fuse the features of financial entities, association relationships, and entity attribute information to achieve collaborative understanding and semantic alignment of multi-modal data in the same business scenario. Based on rule engines and deep learning models, detect and extract key event information from the fused features, including large-scale fund transactions, abnormal transaction patterns, related party transactions, and important contract clause change events.

[0028] In the embodiment of the present application, step S2 is used to convert the preprocessed unstructured text and image data into a structured result of "entity-relation attribute-key event" that can be directly called by the subsequent construction of the spatio-temporal knowledge graph. The text side uses a pre-trained natural language processing model in the financial field: the model is first pre-trained on general Chinese corpus, then performs domain adaptive training on desensitized financial contract text, audit draft text, invoice recognition text, account letter and financial report notes, and performs supervised fine-tuning on named entity recognition, relation extraction and event extraction respectively. The categories of named entity recognition cover company name, personnel, amount, date, subject, invoice code, invoice number, document number, tax number, and currency; the relationship types are limited to invoice correspondence, transaction counterpart, subject entry, payment, contract constraints, and bill references; the event types are limited to large amount of funds, abnormal transaction mode, related party transaction, and contract clause change, and the argument filling is performed on the trigger time, involved subject, involved amount, voucher number, and associated document number. To ensure the consistency of character interpretation and field range, the text side also performs anaphora resolution and term standardization, which unifies the buyer, purchaser, and party A to the subject role, and standardizes the amount and date according to the units and formats agreed in step S1. The model inference output is given in a structured record form, including category label, original text character start and end position, source paragraph position and confidence, which facilitates evidence tracing and manual extraction.

[0029] The image side first performs quality enhancement and geometric correction, including contrast enhancement, denoising, tilt correction and resolution resampling, to improve the stability of subsequent recognition; then a convolutional neural network is used for bill layout analysis and region positioning, and combined with optical character recognition to extract entity attribute information such as invoice code, invoice number, invoice date, amount, buyer name, seller name, tax number, two-dimensional code containing verification string, and bank return number from invoice images, bank document images and business voucher scans. Each field synchronously retains the page number and boundary box coordinates, and performs validity check according to the check digit rule, amount interval rule and date legality rule; when the entity attributes extracted by the text side (such as amount and date) conflict with the same attribute recognized by the image side, the system performs the following conflict resolution rules: first compare the confidence of the two, and prefer the data source with higher confidence; if the difference in confidence is less than 0.1, the corresponding record in the structured database is preferred; if it still cannot be decided, the target is marked as "high conflict risk", and is automatically included in the pending review list, while retaining the evidence sources and original values of both sides.

[0030] The cross-modal fusion adopts an attention mechanism to align the financial entities and relationships identified on the text side with the entity attributes on the image side under the same business fact. Business consistency constraints, including at least subject name consistency, amount consistency, and time proximity, are introduced in the fusion process to improve the reliability of the matching decision. When there is an inconsistency in the amount or date, the system does not discard it directly, but generates a consistency score and retains the conflict entry and its source location as an important input for subsequent audit clues. The cross-modal matching threshold and consistency threshold are calibrated through historical period playback or validation set, and if necessary, they are re-estimated according to the same group benchmark for each account set, thereby ensuring reproducibility and robustness under different enterprises and different data quality conditions.

[0031] The output after fusion is stored in an object-oriented structure oriented towards graph construction. Entity records contain unique identifiers, categories, standardized names, source channels (text or image), source locations, and confidence levels. Relationship records contain subjects, objects, relationship types, occurrence times, source channels, and confidence levels. Event records contain event types, argument sets, trigger times, involved amounts, evidence references, and consistency scores.

[0032] It should also be noted that the pre-trained financial domain natural language processing model, whose core architecture can be the FinBERT model based on Transformer, can use 12-layer hidden layers, 768-dimensional hidden states, and 12 attention heads. The learning rate for domain adaptation training can be set to 2e-5, the batch size is 32, and the maximum sequence length is 512. The named entity recognition and relationship extraction tasks are connected to linear classifiers in the output layer during fine-tuning, and cross-entropy loss functions are used for optimization.

[0033] The convolutional neural network used for image layout analysis can use the ResNet-50 architecture, with its weights pre-initialized on the ImageNet dataset. The optical character recognition task uses the CRNN model, where the CNN part uses a VGG16-like structure, the RNN part is a bidirectional LSTM (with 256 hidden units), and the Connectionist Temporal Classification (CTC) loss function is used for end-to-end training.

[0034] S3: Associate and fuse the extracted financial entities, relationship attributes, and key event information with the structured database transaction data in the preprocessed data set to construct an enterprise financial spatio-temporal knowledge graph containing time attributes.

[0035] Align the financial entities with the corresponding entities in the structured database transaction data in the preprocessed data set based on a unified identifier to obtain an aligned entity set; The association relationship is associated and mapped with transaction records in the structured database transaction data to obtain a mapped association relationship set; The entity attribute information is attribute fused with corresponding entities in the aligned entity set to obtain an attribute-rich entity set. The key event information is associated with related entities and relationships in the attribute-rich entity set and the mapped association relationship set to obtain an event-entity-relationship association network. Based on a graph database construction technology, the attribute-rich entity set is taken as a graph node, the mapped association relationship set and the association relationship contained in the event-entity-relationship association network are taken as graph edges, a knowledge graph basic framework is constructed, and a time attribute extracted from the structured database transaction data and the key event information is added to each node and edge. The knowledge graph basic framework is associated and fused with the key event information and the preprocessed structured database transaction data to form an enterprise financial space-time knowledge graph with a time attribute.

[0036] In the embodiment of the application, step S3 is to associate and fuse the financial entity, relationship attribute and key event information obtained in step S2 with the preprocessed structured database transaction data to form an enterprise financial space-time knowledge graph with a time attribute. First, entity alignment is performed: a unified identifier system is established for each type of entity, the identifier is composed of enterprise master data coding, source system identifier, table name and original primary key, and the fields such as company name or unified social credit code, supplier or customer number, subject code, personnel number are combined for deduplication and merging; when different sources have name differences or coding inconsistencies for the same entity, mapping is completed through a master data mapping table and a string standardization rule, and conflict marks and source evidence are reserved for entries that are difficult to automatically merge. The result of entity alignment forms an aligned entity set, and the source channel, source time, alignment confidence and data bloodline are recorded for each entity to support subsequent tracing and checking.

[0037] In the relationship mapping stage, the relationship attributes output in step S2 are correspondingly associated with the transaction records in the structured database transaction data, and a set of "entity-entity" financial association relationships is established; the relationship types include at least invoice correspondence, transaction counterpart, subject entry, money receipt and payment, contract constraint, and bill reference. When mapping, multi-key matching is performed according to the voucher number, document number, contract number, invoice number, bank return slip number, and opposite main body code, and secondary verification is performed in combination with the amount consistency and the period consistency, the items that do not satisfy the consistency constraint are retained as low-trust relationships and the conflict reasons are added. Then, the entity attributes extracted from the image and text sides are fused with the aligned entities, and the attribute source priority can be set according to the business rules, and the typical method is to give priority to structured transaction data and supplement missing fields with document extraction results; when the same attribute comes from multiple channels, the confidence weighted or "main data priority, document verification" strategy is used to determine the final attribute value, and the replaced and retained values and their evidence positions are recorded to realize explainable attribute synthesis.

[0038] The key event information is associated with the related entities and the mapped relationships to form an association network of event-entity-relationship. The event record contains event type, trigger time, involved subject, involved amount, corresponding voucher and associated document, and the association between the event and the entity and the relationship is explicitly assigned through the event participation role, so that the event can be used as evidence of the relationship and can also become a driving force for introducing new relationships. Based on the graph database construction technology, the attribute-rich entity set is used as the graph node, and the mapped financial association relationships and event association relationships are used as the graph edge, to establish a knowledge graph basic framework. The modeling of the time attribute adopts a multi-granularity coexistence strategy, and the node and the edge record the occurrence time, the accounting time and the effective period respectively; the snapshot sequence is maintained according to the accounting period on the time axis, and incremental update of types such as addition, modification and invalidation is supported, to realize history playback and point consistency verification. In order to ensure the query and training efficiency, the unified identifier, the voucher number, the document number and the time field are indexed at the graph database level, and the uniqueness constraint and the reference integrity constraint of the node and the edge are set.

[0039] Knowledge reasoning is used to complete potential correlation and improve the connectivity of the graph. The reasoning method can adopt one of rule-based reasoning and statistical learning-based reasoning or a combination of both: the former generates verifiable derived relationships according to accounting and business process constraints, such as the consistency of the same invoice number and voucher number within the period, the dependence of receipt and payment on the issuance of invoices, and the constraints of contract terms on revenue recognition; the latter proposes candidate edges to be verified according to the structure and attributes of the constructed graph to learn the possibility of the existence of relationships. All candidate relationships obtained by reasoning must pass four types of checks, including amount caliber consistency, period coincidence, subject matching, and evidence existence, before being added to the graph. Candidate relationships that do not pass the checks are retained as a set of to-be-verified and are not added to the production graph. The final output is a spatio-temporal knowledge graph containing entity dimension, relationship dimension, event dimension, and time dimension, with each node and edge attached with source channel, time label, confidence, and evidence reference. The above processing fuses heterogeneous information across systems and modalities under a unified identifier and unified time axis, eliminating the alignment difficulties caused by coding differences and time mismatches, and providing a stable structure, consistent caliber, and traceable input for subsequent spatio-temporal feature learning and anomaly propagation analysis.

[0040] S4: inputting the enterprise financial spatio-temporal knowledge graph into an anomaly analysis model, using a graph convolution network model to extract spatial structure features of financial entities in a topological network, and using a time convolution model to extract dynamic features of financial relationships evolving over time sequences.

[0041] topological structure analysis is performed on the enterprise financial spatio-temporal knowledge graph to obtain a node feature matrix and an adjacency matrix of the graph; The node feature matrix and the adjacency matrix are input into the graph convolution network model for spatial feature extraction, and the spatial dependence relationship of the financial entities within the multi-hop neighbor range is captured through multi-layer graph convolution operation to extract the spatial structure features of the financial entities in the topological network; The enterprise financial spatio-temporal knowledge graph is time series sliced to extract the feature change sequence of each entity and relationship according to the time dimension, and time series feature data is extracted; The time series feature data is input into the time convolution model for dynamic feature extraction, and the evolution law of the financial relationship over different time scales is captured through dilated convolution operation to extract the dynamic features of the financial relationship evolving over time sequences.

[0042] In the embodiment of the present application, step S4 takes the enterprise financial space-time knowledge graph obtained in step S3 as input, first constructs a graph data that can be directly consumed by the model. The graph is topologically analyzed to generate a node feature matrix and an adjacency matrix. The node feature vector is composed of two types of information: one is the standardized coding of entity attributes, including entity category coding, subject or principal subject data coding, currency identification, and amount range identification; the other is the quantitative features calculated in a unified time window according to the graph records, including the number of transactions during the period, the amount of debit, the amount of credit, the net amount during the period, the amount fluctuation range (determined by the dispersion index of the rolling window), the type and confidence of the latest key event, the average score of cross-modal consistency, the proportion of low-confidence fields, and the shortest path length to the labeled high-risk entity. The above quantitative features are all derived from the standardized and verified fields in steps S1 and S2, as well as the node, edge, and event records in step S3. The calculation range and time granularity are consistent with step S1 to ensure reviewability. The adjacency matrix is generated from the association in the graph, and the edge weight adopts a composite weight that integrates transaction frequency, amount scale, evidence confidence, and event intensity. After normalization, it is written into the matrix, and self-loops are retained to enhance the stability of aggregation; a type coding vector is also generated for the relationship type to distinguish the influence of different relationships during convolution. To avoid representation bias caused by high-connectivity nodes, their neighbor sets are sampled in batches and degree-normalized.

[0043] The node feature matrix and the adjacency matrix are input into a graph convolution network model for spatial feature extraction. The model obtains the structural representation within the multi-hop neighborhood through multiple layers of aggregation, and adds normalization and residual connection between layers to alleviate the over-smoothing phenomenon; the edge weight participates in the weighted aggregation, so that high-confidence, high-amount, and high-frequency relationships have a greater contribution to the representation; for known normal samples, structure reconstruction or link discrimination can be used as training signals, and for historically labeled abnormal samples, supervised signals can be used for contrast training. When there is a lack of labels, subgraph comparison or masked reconstruction can be used as self-supervised targets to ensure the implementability under different data availability conditions. This step can explicitly capture multi-hop dependencies and closed-loop structures across departments and accounts, identify short-term new relationships, rapid buy-sell, loop backflow, and other graph patterns, and provide spatial evidence for abnormal source localization.

[0044] The time dimension processing adopts time series slicing. Ordered snapshots are generated for the graph at an accounting period or daily granularity, and the feature change sequence is extracted for each entity and relationship along the time axis. The sequence at least includes the transaction count, the amount aggregation, the number of key event triggers, the cross-modal consistency score, and the rule hit mark; explicit missing indicators are set for missing segments, and the length is completed by using the adjacent period value carrying method to avoid introducing false values. The obtained time series feature data is one-to-one corresponding to the entity and relationship index, and the timestamp is consistent with the standard in step S1, which is convenient for alignment with the spatial representation.

[0045] The time series feature data is input into a time convolution model for dynamic feature extraction. The model uses multi-scale convolution channels to cover three time spans of week, month and season, and only uses past information for representation through causal convolution or boundary constraint to avoid information leakage; in the presence of seasonality and account switching, the sequence is first segmented and normalized and the cross-period connection is processed to improve the sensitivity to structural jumps. The training target can be one of the next time point prediction, sequence reconstruction or deviation minimization from the historical steady-state baseline, and the output reflects the dynamic representation of short-term fluctuations, long-term trends and periodic characteristics.

[0046] The spatial structure features and dynamic features are aligned in entity and relationship indexes and time indexes, and are output with source time range and feature generation time, retaining missing indication and numerical range standardization information to ensure that the threshold of the subsequent risk assessment layer is not affected by scale bias. Compared with the processing method that relies on single-period statistical threshold, this step simultaneously uses multi-hop spatial dependence and multi-scale time patterns under the constraint of unified identification and unified time axis, and can provide interpretable abnormal basis through structural reconstruction deviation and time deviation signal under the condition of label scarcity, providing stable, consistent, traceable model input for abnormal probability calculation, propagation weight measurement and abnormal source identification in step S5.

[0047] It should also be noted that the graph convolution network model can use a two-layer GraphSAGE structure, with an output dimension of 128 for each layer, using a ReLU activation function and mean aggregation for aggregation. During training, negative sampling and cross-entropy loss are used for self-supervised learning of the link prediction task.

[0048] The time convolution model can use a 4-layer dilated causal convolution structure with a hidden layer dimension of 64, a dilation coefficient sequence of [1, 2, 4, 8], and a convolution kernel size of 3. This structure ensures that the model's receptive field grows exponentially with the number of layers, while avoiding future information leakage. The training target is the prediction of the next time step feature, and the mean square error (MSE) is used as the loss function.

[0049] S5: Based on the obtained spatial structure features and dynamic features, the abnormal probability of the node and the propagation weight of the edge are calculated through the attention mechanism and the risk assessment layer, the abnormal source is identified, the systemic risk value is evaluated, and the abnormal propagation path is generated.

[0050] The spatial structure features and dynamic features are fused at the feature level to obtain spatio-temporal fusion features; The spatio-temporal fusion features are input into the attention mechanism to calculate the attention weight of each node and the propagation weight of each edge in the graph, obtaining a node attention weight matrix and an edge propagation weight matrix; input the node attention weight matrix and the edge propagation weight matrix into a risk assessment layer, calculate an abnormal probability of each node and a risk propagation intensity of each edge through a multi-layer neural network, and obtain a node abnormal probability distribution and an edge risk propagation intensity distribution; based on the node abnormal probability distribution, identify a node with an abnormal probability exceeding a preset threshold as an abnormal source; based on the edge risk propagation intensity distribution and the node abnormal probability distribution, calculate a systemic risk value through a risk propagation model, and evaluate an overall risk level; based on the edge risk propagation intensity distribution and the node abnormal probability distribution, generate an abnormal propagation path starting from the abnormal source through a path search algorithm; output a risk analysis result containing an abnormal source identifier, a systemic risk value and an abnormal propagation path.

[0051] In the embodiment of the application, step S5 completes risk calculation and propagation analysis on the basis of the spatial structure features and dynamic features output in step S4. First, feature-level fusion is performed. The spatial structure features and dynamic features of the same node and its associated edges in the same time window are aligned according to a unified identifier, and are spliced or weighted and merged in a fixed order, while the source identifier, time range and missing indication are retained, to form spatio-temporal fusion features. The fusion strategy is determined through the training set and the evaluation set, aiming to maintain the time sensitivity while not weakening the distinguishing power of multi-hop structure information. Then, the spatio-temporal fusion features are input into an attention mechanism to calculate the contribution of the neighborhood of the node and the contribution of the edge to risk diffusion. The attention score is determined not only by the feature similarity, but also by introducing business consistency constraints as prior signals, including at least three items of subject name consistency, amount consistency and time proximity. During training, the weight of a low-consistency sample is reduced to suppress the influence of mis-matching caused by source noise. The attention output is given in the form of two types of weights: node attention weight for explaining the influence intensity of each neighbor on the representation of the target node, and edge propagation weight for explaining the accessibility of each relationship in risk diffusion. Both of them retain the source index and time range for audit tracing.

[0052] In the risk assessment layer, the node side and the edge side are scored by multi-layer neural networks respectively, outputting node anomaly probability distribution and edge risk propagation intensity distribution. Node anomaly probability is used to describe the possibility of a single entity being abnormal in the current period, and the training signal can come from historical verified anomalies, semi-supervised labels or pseudo-labels based on structural and sequence reconstruction bias; edge risk propagation intensity is used to describe the carrying capacity of the relationship when the risk spreads along the graph, and the edge weight, event intensity and cross-modal consistency score are considered simultaneously during training. To ensure the usability of the output, the risk assessment layer is calibrated afterwards to make the probability consistent with the actual hit frequency. The calibration method is selected through historical period playback and cross-validation, and is re-calibrated with the same group benchmark in different account sets, so as to maintain a stable threshold under the influence of industry differences and seasonality. The identification of abnormal sources is based on the node anomaly probability distribution and the business set threshold, and the threshold is set by maximizing the comprehensive index on the validation data, and is re-evaluated in a rolling window after going online; the nodes judged as abnormal sources output unique identification, time range, main evidence field and attention explanation vector, which is convenient for audit review.

[0053] The evaluation of systemic risk value is based on the joint completion of edge risk propagation intensity distribution and node anomaly probability distribution. The risk propagation model regards the graph as a network with direction and weight, simulates the spread of risk on the relationship in chronological order, aggregates the risk volume in the whole graph within a given time window and normalizes it to obtain a comparable overall risk level. This value reflects the potential impact range and cumulative intensity of anomalies in the organizational structure, which can be compared with the historical stable baseline to trigger key audits at the group level or business unit level. The generation of abnormal propagation path is based on the joint scoring of edge propagation weight and node anomaly probability, and through path search algorithm, the highest risk accumulation value is selected from the candidate path set, and the path length and time span are limited to avoid misjudgment of low-value long links as key propagation chains. The output path is presented in the form of node sequence and edge sequence, accompanied by timestamp, amount range, evidence reference and attention explanation component of each link to ensure checkability.

[0054] To improve robustness, step S5 performs weight reduction processing on low-confidence inputs, retains explicit masks for missing features and includes mask vectors in the evaluation layer to avoid false signals from filling; enables a penalty factor for high-connectivity nodes to prevent disproportionate amplification in attention aggregation; sets version numbers and effective periods for threshold values and calibration parameters to facilitate backtracking and comparison when regulations or data distributions change. The final output risk analysis result includes at least abnormal source identification, systemic risk value, abnormal propagation path, and evidence list, and records model version, training batch, and time range. Compared with schemes based only on single-period statistics or only on local rules, this step uses spatiotemporal fusion features and interpretable attention weights to cover both multi-hop structures across departments and long-term and short-term evolution patterns, and converts "point anomalies" into "link evidence" through risk propagation and path search, providing executable decision-making basis for subsequent audit sampling and resource allocation.

[0055] It should also be noted that the risk assessment layer can be a three-layer fully connected neural network with an input dimension of spatiotemporal fusion feature dimension (e.g., 256 dimensions), hidden layer dimensions of 128 and 64, and an output layer using a Sigmoid activation function to output node anomaly probability (between 0 and 1). The calculation of edge risk propagation intensity shares the first two layers of the same network, and finally outputs through an independent linear layer.

[0056] The preset threshold of node anomaly probability is determined by the maximum F1 score on the validation set, usually set in the range of 0.7 to 0.9 (e.g., 0.85). The grade division threshold of the system risk value is dynamically calibrated according to the distribution of historical data in the same industry, for example: low risk [0, 0.3), medium risk [0.3, 0.6), high risk [0.6, 1.0].

[0057] S6: Based on the abnormal source, the systemic risk value, and the abnormal propagation path, integrate to generate the final audit report.

[0058] Based on the abnormal source identification, extract the corresponding abnormal entity information and associated transaction records to generate detailed audit abnormal clues; Based on the systemic risk value, determine the risk level through the risk level division model to generate the risk level evaluation result; Based on the abnormal propagation path, construct an abnormal propagation path graph through a visualization engine to generate a visualized propagation path; Integrate the audit abnormal clues, the risk level evaluation result, and the visualized propagation path to generate a structured audit report; Add a timestamp, an audit range identifier, and a model version information to the structured audit report to form the final audit report.

[0059] In the embodiment of the present application, step S6 takes the abnormal source, the systemic risk value and the abnormal propagation path output in step S5 as input, and completes the structured integration and solidification of the audit results. First, based on the abnormal source identification, the information and associated transaction records of the corresponding entity are extracted from the enterprise financial space-time knowledge graph and the structured database transaction data. The identifier system and timestamp standard established in step S1 are uniformly adopted, and the entity master data, account or subject information, amount and balance within the period, corresponding vouchers and document numbers, cross-modal consistency scores, low-trust field markers and triggered key event records are aggregated. In order to avoid repetition and inconsistent standards, de-duplication and standard checking are performed according to the voucher number, document number, opposite subject and period. For the conflicting items, the conflict reason and source position are retained, and they are listed in the evidence list of the audit abnormal clues.

[0060] The systemic risk value is used to generate a risk level assessment result. The risk level division model determines the threshold value based on historical period playback and same group benchmark calibration, maps the systemic risk value to the level interval, and outputs the basis for the formation of the level, including the benchmark period used, the sample size, the industry and size grouping standard, the threshold version and the effective period. In order to enhance the interpretability, the risk value and baseline value, the change rate of the last three periods, and the path and node set with the largest contribution are given in the report, and the corresponding attention explanation elements and edge propagation weight sources are marked, so as to facilitate the audit personnel to understand the trigger reason and the influence range.

[0061] The presentation of the abnormal propagation path relies on the visualization engine to generate a path graph and corresponding data objects. The path graph adopts a hierarchical layout of nodes and edges, preferentially displays links with higher propagation weights, and imposes constraints on path length and time span to ensure readability. Each node and edge is attached with a unique identifier, time range, amount standard, evidence reference and consistency score. The path graph and its data objects can be restored to each other, supporting both graphical viewing and subsequent machine consumption of automated audit procedures. In order to serve subsequent sampling and review, the system synchronously generates executable audit action suggestions, including transaction sampling proportion of high-risk nodes, types and fields of documents that need to be supplemented and verified, conflict pair list for priority verification, and suggested serial and parallel operation sequence.

[0062] The structured audit report is organized in a unified data structure and at least includes a set of audit anomaly clues, risk level assessment results, anomaly propagation path data objects, evidence lists, specification and threshold value descriptions, data sources and time coverage. To ensure traceability and tamper resistance, the report is embedded with model version, training batch identification, feature engineering version, threshold and benchmark version, rule base version, uniform identifier mapping table version, generation timestamp and audit range identification; a digital fingerprint hash value is generated for the report subject and evidence list and the storage location and access permission policy are recorded. The output is provided in machine-readable format and human-readable version, the former is used to interface with the audit management system and the work order system, and the latter is convenient for review and archiving.

[0063] The technical effect of the above integration process is that: through the consistent identification and time specification of steps S1 to S5, the point identification result is solidified into a linkable and verifiable audit product; through the calibration information, path explanation elements and evidence coordinates of the level division, the explainability and verifiability of the audit conclusion are improved; through the action suggestion and structured output, the result can directly drive the subsequent audit sampling and review process, shorten the closed-loop cycle from identification to execution, and reduce the landing friction caused by heterogeneous data sources and specification differences.

[0064] Embodiment 2 is a second embodiment of the present application, which is different from the previous embodiment: If the function is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the prior art that make essential contributions or parts of the current technical solutions can be embodied in the form of a software product, and the current computer software product is stored in a storage medium, including a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and various program code storage media.

[0065] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that can be later executed by a computer. In this context, a "computer-readable medium" can be any means that can store the program for use by or in connection with the instruction execution system, apparatus, or device.

[0066] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that can be later executed by a computer.

[0067] Embodiment 3, as an embodiment of the present application, provides an accounting data intelligent processing system for enterprise financial audit, comprising a multi-modal acquisition and cleaning module, a semantic analysis and extraction module, a space-time graph construction module, a space-time feature extraction module, a risk propagation analysis module and an audit report generation module.

[0068] The multi-modal acquisition and cleaning module: acquires multi-modal enterprise financial data, and performs standardized preprocessing on the acquired data to form a preprocessed data set; the multi-modal enterprise financial data includes structured database transaction data and unstructured text and image data; The semantic analysis and extraction module: performs multi-modal semantic understanding and analysis on the unstructured text and image data in the preprocessed data set, and extracts financial entities, relationship attributes and key event information therefrom; The space-time graph construction module: associates and fuses the extracted financial entities, relationship attributes and key event information with the structured database transaction data in the preprocessed data set, and constructs an enterprise financial space-time knowledge graph containing time attributes; The space-time feature extraction module: input the enterprise financial space-time knowledge graph into the anomaly analysis model, use the graph convolution network model to extract the spatial structure features of the financial entity in the topological network, and use the time convolution model to extract the dynamic features of the evolution of the financial relationship along the time sequence; The risk propagation analysis module: based on the obtained spatial structure features and dynamic features, through the attention mechanism and the abnormal probability of the risk evaluation layer calculation node and the propagation weight of the edge, the abnormal source, the system risk value and the abnormal propagation path are identified; The audit report generation module: based on the abnormal source, the system risk value and the abnormal propagation path, the final audit report is integrated and generated.

[0069] Embodiment 4 is an embodiment of the present application, which provides an accounting data intelligent processing method for enterprise financial audit. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation / contrast experiment are used for scientific demonstration.

[0070] This embodiment selects two sets of real business process account data similar to manufacturing and retailing, corresponding to the fourth quarter of 2024 and the first quarter of 2025 respectively. In the S1 stage, structured and unstructured data are synchronously extracted from enterprise resource planning systems and document management systems. For the structured side, data cleaning, uniformity of caliber, key field extraction and timestamp standardization are completed. For the unstructured side, image quality enhancement, optical character recognition and text encoding uniformity are completed, and time information is standardized from contracts and bills. In the S2 stage, the pre-trained financial field natural language processing model is used to analyze the named entities and relationships in the text side. Combining convolutional neural networks and optical character recognition, the invoice code, number, invoice date, amount, and purchase and sale subjects are located and extracted. Then, based on the attention mechanism, the text and image evidence under the same business fact are aligned for consistency, and the traceable results containing entities, relationship attributes and key events are output. In the S3 stage, the entities and structured transaction records are aligned according to the uniform identifier system, the relationships are mapped and the attributes are fused, the events and related entities and relationships are associated into an event-entity-relationship network, and a basic framework containing nodes, edges and time labels is constructed in a graph database. The potential relationships are completed by rule reasoning and statistical inference, and the enterprise financial spatio-temporal knowledge graph is formed. In the S4 stage, the graph is topologically analyzed to generate node feature matrix and adjacency matrix, and time series features are formed by slicing according to the accounting period. The spatial side learns multi-hop dependence through graph convolutional network, and the temporal side learns multi-scale evolution features through time convolution model, and the aligned spatial structure features and dynamic features are output. In the S5 stage, the two types of features are fused according to the uniform identifier, the attention mechanism is introduced to obtain node attention weight and edge propagation weight, and the node anomaly probability and edge risk propagation intensity are output through the risk evaluation layer. Based on the threshold, the abnormal source is identified, the systemic risk value is calculated by the propagation model, and the key abnormal propagation link is generated by path search. In the S6 stage, around the abnormal source, systemic risk value and propagation path, evidence is extracted from the graph and transaction library to form audit anomaly clues, risk level evaluation and path data objects, which are solidified into structured audit reports with time stamp, range identifier and model version.

[0071] To demonstrate the advantages of the method, two types of controls are set: control 1 is a traditional scheme relying only on structured threshold and rules, and control 2 is a single modal OCR+rule document-driven scheme; the three are compared and evaluated on the same account and the same caliber data.

[0072] The experimental reference data is shown in Table 1.

[0073] Table 1 Experimental data records

[0074] Compared with Control 1 and Control 2, the method of the present application has stable advantages on the two sets of accounts. First, the entity alignment accuracy reaches 96.2% and 95.1% respectively, which is 14.7-15.3 percentage points higher than Control 1 and 12.1-12.0 percentage points higher than Control 2. This difference is due to the linkage of S2 cross-evidence alignment and S3 unified identifier system, reducing the mismatch caused by name variation, ticket missing characters and master data differences. Second, the F1 of relationship extraction is 92.4% and 90.3% respectively in the two sets of accounts, which is 18.5-19.1 percentage points higher than Control 1 and 14.3-13.9 percentage points higher than Control 2, indicating that the multi-modal fusion of relationship evidence and S3 event-entity-relation connection significantly enhances the determinability of edge level. The key event detection rate under the present application reaches 88.6% and 85.2%, which is 26.8-26.5 percentage points higher than Control 1 and 19.2-19.1 percentage points higher than Control 2; this shows that the joint discrimination of S2 and the fusion of S5 spatio-temporal features can still maintain a high recall under the condition of sparse documents or poor scanning quality. The abnormality recognition AUC is 0.94 and 0.93 respectively, which is higher than 0.79 and 0.77 of Control 1 and 0.85 and 0.84 of Control 2, proving that the multi-hop spatial dependence of S4 graph convolution and the multi-scale dynamic representation of time convolution form a complement in risk discrimination, which is better than the threshold or single modal evidence mode. The systematic risk assessment bias (MAPE) is significantly reduced to 7.8% and 8.6%, which is about 10-12 percentage points lower than Control 1 and about 6-7 percentage points lower than Control 2, reflecting the more accurate estimation of S5 propagation weight and path search on the overall risk volume. In terms of precision, the precision of the audit abnormal clues of the present application reaches 89.1% and 87.3%, which is 15.5-15.4 percentage points higher than Control 1 and 7.9-8.3 percentage points higher than Control 2, directly reducing the interference of false positives to subsequent manual review. In terms of efficiency, the audit manual review workload is reduced to 42 and 38 hours, Control 1 is 128 and 114 hours, and Control 2 is 96 and 88 hours; the average report generation time is shortened to 16.4 and 18.1 minutes, which is significantly lower than 31.2 and 28.4 minutes of Control 1 and 26.5 and 24.7 minutes of Control 2, indicating that the structured integration and evidence solidification of S6 can directly drive the report output and reduce the manual assembly time.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for intelligent processing of accounting data for enterprise financial auditing, characterized in that, include: Acquire multimodal enterprise financial data and perform standardized preprocessing on the acquired data to form a preprocessed dataset; The multimodal enterprise financial data includes structured database transaction data and unstructured text and image data; Multimodal semantic understanding and parsing are performed on the unstructured text and image data in the preprocessed dataset to extract financial entities, relationship attributes and key event information; The extracted financial entity, relationship attributes, and key event information are associated and integrated with the structured database transaction data in the preprocessed dataset to construct a corporate financial spatiotemporal knowledge graph containing time attributes. The enterprise financial spatiotemporal knowledge graph is input into the anomaly analysis model. The spatial structure features of financial entities in the topological network are extracted using the graph convolutional network model, and the dynamic features of the evolution of financial relationships over time are extracted using the temporal convolutional model. Based on the obtained spatial structure and dynamic features, the abnormal probability of nodes and the propagation weight of edges are calculated through the attention mechanism and risk assessment layer to identify the source of anomalies, assess the systemic risk value, and generate anomaly propagation paths. Based on the source of the anomaly, the systemic risk value, and the anomaly propagation path, the final audit report is generated.

2. The intelligent accounting data processing method for enterprise financial auditing as described in claim 1, characterized in that, The multimodal enterprise financial data includes structured database transaction data and unstructured text and image data; The structured database transaction data includes accounting voucher data, general ledger data, subsidiary ledger data, trial balance data, and financial statement data; The unstructured text and image data includes invoice image data, purchase and sales contract text data, bank document image data, business voucher scan data, and audit report text data.

3. The intelligent accounting data processing method for enterprise financial auditing as described in claim 2, characterized in that, The standardization preprocessing of the acquired data includes data cleaning, format unification, key field extraction, and timestamp standardization of the structured database transaction data. Optical character recognition, image sharpness enhancement, text encoding standardization, and time information extraction and standardization of the unstructured text and image data are performed on the unstructured text and image data.

4. The intelligent accounting data processing method for enterprise financial auditing as described in claim 3, characterized in that, The multimodal semantic understanding and parsing process includes parsing the text data in the preprocessed dataset, performing named entity recognition based on a pre-trained financial domain natural language processing model, and extracting financial entities including company name, personnel, amount, date, and account. The text data in the preprocessed dataset is parsed, and relation extraction is performed based on a pre-trained financial domain natural language processing model to extract the association relationships between the financial entities. The image data in the preprocessed dataset is parsed, and key information is located and extracted based on convolutional neural network and optical character recognition technology, identifying entity attribute information including invoice code, invoice number, invoice date, amount, and buyer and seller information; An attention mechanism is used to integrate features of financial entities, relationships, and entity attributes to achieve collaborative understanding and semantic alignment of multimodal data in the same business scenario; Based on a rule engine and a deep learning model, key event information is detected and extracted from the fused features. The key event information includes large-scale fund transactions, abnormal transaction patterns, related-party transactions, and changes to important contract terms.

5. The intelligent accounting data processing method for enterprise financial auditing as described in claim 4, characterized in that, The construction of the enterprise financial spatiotemporal knowledge graph containing time attributes includes aligning financial entities with corresponding entities in the structured database transaction data of the preprocessed dataset based on a unified identifier, to obtain an aligned entity set. The relationships are mapped to transaction records in the structured database transaction data to obtain a set of mapped relationships; The entity attribute information is fused with the corresponding entity in the aligned entity set to obtain an entity set with rich attributes. By associating key event information with relevant entities and relationships in a set of attribute-rich entities and a set of mapped relationships, an event-entity-relationship association network is obtained. Based on graph database construction technology, a knowledge graph basic framework is constructed using a set of entity nodes with rich attributes as graph nodes, a set of mapped relationships and the relationships contained in the event-entity-relationship network as graph edges, and time attributes extracted from the structured database transaction data and the key event information are added to each node and edge. By using knowledge reasoning algorithms to complete the basic framework of the knowledge graph with potential relationships, an enterprise financial spatiotemporal knowledge graph containing entity, relationship, event, and time dimensions is generated.

6. The intelligent accounting data processing method for enterprise financial auditing as described in claim 5, characterized in that, The step of extracting the spatial structure features of financial entities in the topological network using a graph convolutional network model and extracting the dynamic features of the evolution of financial relationships over time using a temporal convolutional model includes performing topological structure analysis on the enterprise financial spatiotemporal knowledge graph to obtain the node feature matrix and adjacency matrix of the graph. The node feature matrix and adjacency matrix are input into the graph convolutional network model for spatial feature extraction. The spatial dependencies of financial entities within the range of multi-hop neighbors are captured through multi-layer graph convolution operations, and the spatial structural features of financial entities in the topological network are extracted. Time series slices are performed on the enterprise financial spatiotemporal knowledge graph, and the feature change sequences of each entity and relationship are extracted according to the time dimension to obtain time series feature data. The time series feature data is input into the time convolution model for dynamic feature extraction. The evolution of financial relationships at different time scales is captured by dilated convolution operation, and the dynamic features of financial relationships evolving with time series are extracted.

7. The intelligent accounting data processing method for enterprise financial auditing as described in claim 6, characterized in that, The process of identifying the source of anomalies, assessing systemic risk values, and generating anomaly propagation paths includes fusing spatial structural features and dynamic features at the feature level to obtain spatiotemporal fusion features. The spatiotemporal fusion features are input into the attention mechanism to calculate the attention weights of each node and the propagation weights of each edge in the graph, thus obtaining the node attention weight matrix and the edge propagation weight matrix. The node attention weight matrix and edge propagation weight matrix are input into the risk assessment layer. The anomaly probability of each node and the risk propagation intensity of each edge are calculated through a multi-layer neural network to obtain the node anomaly probability distribution and the edge risk propagation intensity distribution. Based on the node anomaly probability distribution, nodes whose anomaly probability exceeds a preset threshold are identified as anomaly sources. Based on the edge risk propagation intensity distribution and the node anomaly probability distribution, the systemic risk value is calculated using the risk propagation model to assess the overall risk level. Based on the edge risk propagation intensity distribution and the node anomaly probability distribution, a path search algorithm is used to generate an anomaly propagation path starting from the anomaly source. The output includes risk analysis results such as anomaly source identification, systemic risk value, and anomaly propagation path.

8. The intelligent accounting data processing method for enterprise financial auditing as described in claim 7, characterized in that, The process of generating the final audit report includes extracting the corresponding abnormal entity information and related transaction records based on the anomaly source identifier, and generating detailed audit anomaly clues. Based on the systemic risk value, the risk level is determined through a risk level classification model, and a risk level assessment result is generated. Based on the anomaly propagation path, an anomaly propagation path graph is constructed using a visualization engine to generate a visualized propagation path. The audit anomaly clues, risk level assessment results, and visualized dissemination paths are integrated to generate a structured audit report; Add timestamps, audit scope identifiers, and model version information to the structured audit report to generate the final audit report.

9. An intelligent accounting data processing system for enterprise financial auditing, used to implement the intelligent accounting data processing method for enterprise financial auditing as described in any one of claims 1 to 8, characterized in that, include: Multimodal data acquisition and cleaning module: Acquires multimodal enterprise financial data and performs standardized preprocessing on the acquired data to form a preprocessed dataset; The multimodal enterprise financial data includes structured database transaction data and unstructured text and image data; Semantic parsing and extraction module: performs multimodal semantic understanding parsing on the unstructured text and image data in the preprocessed dataset to extract financial entities, relationship attributes and key event information; Spatiotemporal graph construction module: The extracted financial entity, relationship attributes and key event information are associated and integrated with the structured database transaction data in the preprocessed dataset to construct a corporate financial spatiotemporal knowledge graph containing time attributes; Spatiotemporal feature extraction module: The enterprise financial spatiotemporal knowledge graph is input into the anomaly analysis model. The spatial structure features of financial entities in the topological network are extracted using the graph convolutional network model, and the dynamic features of the evolution of financial relationships over time are extracted using the temporal convolutional model. Risk propagation analysis module: Based on the obtained spatial structure features and dynamic features, it calculates the anomaly probability of nodes and the propagation weight of edges through attention mechanism and risk assessment layer, identifies the source of anomalies, assesses the systemic risk value, and generates anomaly propagation paths; Audit report generation module: Based on the source of the anomaly, the systemic risk value, and the anomaly propagation path, it integrates and generates the final audit report.

Images