Financial voucher automatic generation and verification method and system based on intelligent accounting

Abstract

The invention provides a multi-source financial data intelligent classification and automatic accounting optimization processing method and system, and relates to the technical field of enterprise financial management, and the method comprises the steps: carrying out the textualized preprocessing of multi-source enterprise financial data, constructing a transaction scene map, carrying out the scene clustering, forming a transaction classification rule set, and carrying out the classification of transactions, according to the method, labels and accounting subject codes are matched for conventional transaction data, a transaction traceability graph is constructed to analyze the integrity of a transaction link, and an accounting voucher is generated to realize automatic accounting, so that the accuracy and efficiency of financial data processing are improved, the manual intervention cost is reduced, and the traceability of financial data is enhanced.

Description

Technical Field

[0001] The present invention relates to intelligent financial management technology, and in particular to a method and system for automatically generating and verifying financial vouchers based on intelligent accounting. Background Art

[0002] With the rapid development of information technology, financial management is gradually undergoing a digital transformation. Traditionally, financial voucher processing relies primarily on manual operations by financial personnel, including identifying business document information, preparing accounting entries according to accounting standards, generating financial vouchers, and conducting audits and verifications. In recent years, intelligent financial systems have begun to apply technologies such as artificial intelligence and natural language processing to automatically process financial data. Some companies have also begun exploring the application of blockchain technology in the financial field to improve the security and credibility of financial data.

[0003] However, the current technology for automatic generation and verification of financial vouchers still has obvious shortcomings. First, the accuracy of information extraction from business documents by existing technologies is not high, especially the processing ability of non-standardized documents is limited, making it difficult to accurately identify and convert complex and diverse transaction information, resulting in the accuracy of subsequent accounting entry generation being affected. Secondly, the accounting subject matching process lacks semantic understanding capabilities and mainly relies on keyword matching or simple rule judgment. It cannot effectively handle the differences in subject allocation for the same transaction in different business scenarios, and it is also difficult to identify potential abnormal transaction patterns. Third, the validity verification and secure storage mechanism of financial vouchers are imperfect. Traditional database storage methods are easily tampered with, lack a reliable verification mechanism and complete audit tracking capabilities, and cannot fully guarantee the authenticity and integrity of financial data. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for automatically generating and verifying financial vouchers based on intelligent accounting, which can solve the problems in the prior art.

[0005] Scan and identify business documents, obtain transaction information from the business documents, and convert the transaction information into standardized transaction data including transaction amount, transaction time, counterparty information, transaction type, product information, and service information; Using natural language processing technology to extract semantic features from the standardized transaction data, similarity matching the extracted semantic features with a preset accounting subject feature library is performed to generate initial accounting entries; a transaction relationship network is constructed, transaction anomalies are identified through graph structure analysis, the initial accounting entries are optimized, and final accounting entries are generated; Generate a financial voucher for the final accounting entry, analyze the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyze the transaction path involved in the financial voucher through graph computing, calculate the rule matching degree of the financial voucher based on a preset accounting standard rule library, and determine the validity of the financial voucher; The financial voucher that has passed the validity judgment is encrypted in blocks, and the encrypted data is stored in a Merkle tree structure. A smart contract that supports zero-knowledge proof is deployed in a distributed network to verify the authenticity of the financial voucher. The evidence information of the financial voucher is written into the blockchain network based on a consensus mechanism; The financial vouchers that have completed the notarization are written into the enterprise financial accounting system, and the verification records, notarization records, and accounting records of the financial vouchers are written into the blockchain network.

[0006] The steps of extracting semantic features from the standardized transaction data using natural language processing technology, performing similarity matching between the extracted semantic features and a preset accounting subject feature library, and generating initial accounting entries include: Performing word segmentation and part-of-speech tagging on the standardized transaction data, extracting transaction entity information, and generating a semantic feature vector; Construct a multi-layer subject decision tree according to accounting standards, set the first-level accounting subject as the root node of the multi-layer subject decision tree, set the detailed accounting subject as the leaf node of the multi-layer subject decision tree, and establish an accounting subject attribution relationship between the root node and the leaf node; Starting from the root node, the semantic feature vector is matched with each layer of accounting subjects, the matching probability of each layer of accounting subjects is calculated, and the matching weight of the next layer of accounting subjects is dynamically adjusted based on the matching results of each layer to generate an accounting subject matching sequence; Calculating the semantic similarity between the semantic feature vector and the preset accounting subject feature library, calculating the business similarity between the transaction rule feature and the preset accounting subject feature library, analyzing the causal relationship strength between the transaction feature and the accounting subject selection, and weightedly fusing the semantic similarity, business similarity, and causal relationship strength to obtain a comprehensive similarity; Determine the final matching accounting subject based on the accounting subject matching sequence and the comprehensive similarity, determine the debit and credit direction of the accounting subject based on the causal relationship strength, and generate initial accounting entries; The matching confidence of the initial accounting entry is calculated, and when the matching confidence is lower than a preset confidence threshold, the initial accounting entry is marked as pending manual review.

[0007] The steps of constructing a transaction relationship network, identifying transaction anomalies through graph structure analysis, optimizing the initial accounting entries, and generating final accounting entries include: Build a transaction relationship network, construct transaction-related information into nodes, and establish edges based on transaction associations; Calculating a weighted degree distribution feature for each node in the transaction relationship network, calculating a path importance feature based on the cumulative edge weights of the transaction path, calculating the network density of the transaction-related subgraph to obtain a global topological feature, and fusing the weighted degree distribution feature, path importance feature, and global topological feature to obtain a graph structure feature; The graph structure features are used to calculate the link anomaly score of the transaction path, the fluctuation anomaly score of the transaction amount, and the time series anomaly score of the transaction frequency. The link anomaly score, the fluctuation anomaly score, and the time series anomaly score are weighted and combined, and adaptively normalized based on the transaction scale to obtain the anomaly score. For transactions with anomaly scores exceeding a first preset threshold, correction rules are generated using historical normal transaction patterns, and the initial accounting entries are rematched and amounts redistributed based on the correction rules. For transactions with anomaly scores exceeding a second preset threshold, an expert rule library is introduced for in-depth optimization to generate optimized accounting entries. The basic compliance score of the optimized accounting entry is calculated based on the accounting standards rules, and the transaction credibility score is calculated in combination with the anomaly score. The basic compliance score and the transaction credibility score are integrated to generate an entry quality assessment value. When the entry quality assessment value is greater than the preset quality threshold, the optimized accounting entry is determined as the final accounting entry.

[0008] The steps of generating a financial voucher for the final accounting entry, analyzing the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyzing the transaction path involved in the financial voucher through graph computing, and calculating the rule matching degree of the financial voucher based on a preset accounting standard rule library to determine the validity of the financial voucher include: Generating a financial voucher for the final accounting entry, and converting the subject information and amount information of the financial voucher into a subject feature vector and an amount feature vector respectively; Calculating attention weights for subject nodes in the subject feature vector, and calculating the association strength between subject nodes based on the attention weights to obtain a subject attention vector; using a gated recurrent unit to fuse the context information of historical credentials with the current credential state to obtain a context vector; and concatenating the subject attention vector with the context vector to obtain a subject logical relationship score; Performing an amount matching and integrity check on the financial voucher to generate an amount matching score and a voucher integrity score; By constructing an initial transaction graph and performing temporal expansion, a dynamic transaction relationship graph with temporal connections is obtained. The node features in the graph are iteratively updated based on an attention propagation mechanism based on transaction size, transaction timing, and transaction frequency. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths and perform sub-path decomposition. Sub-path scores are calculated by combining structural importance and temporal correlation, and transaction path anomaly scores are obtained through deep fusion. Calculating the rule matching degree of the financial voucher based on a preset accounting standard rule base; Adaptively weighted fusion of the subject logic relationship score, amount matching relationship score, voucher integrity score, transaction path anomaly score, and rule matching degree is performed to generate a comprehensive score for the financial voucher; When the comprehensive score is greater than a preset score threshold, the financial voucher is determined to be valid.

[0009] An initial transaction graph is constructed and time-series expanded to obtain a dynamic transaction relationship graph with time-series connections. The node features in the graph are iteratively updated based on an attention propagation mechanism based on transaction scale, transaction sequence, and transaction frequency. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths and perform sub-path decomposition. Sub-path scores are calculated by combining structural importance and time-series correlation. The steps of obtaining transaction path anomaly scores through deep fusion include: Constructing an initial transaction graph based on the subject feature vector and the amount feature vector, setting the transaction entities in the voucher as graph nodes, and setting the fund transaction relationships between the transaction entities as graph edges, wherein the initial weights of the graph edges are determined according to the transaction amount and transaction frequency; Performing temporal expansion on the initial transaction graph, establishing temporal connections between nodes of the same transaction entity in different time windows, and calculating dynamic weights of the temporal connections based on historical transaction data; Node features are iteratively updated using attention propagation. For each node, an attention coefficient is calculated based on the transaction scale, transaction sequence, and transaction frequency of its neighboring nodes. The neighboring node features are weighted and aggregated according to the attention coefficient. The aggregated features are then combined with the node's own features to generate updated node features. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths, decompose the transaction paths into transaction sub-paths, and calculate the structural importance score and temporal correlation score for each sub-path. The structural importance score, temporal correlation score and feature sequence of the sub-path are deeply fused and calculated with the feature sequence of the path nodes to generate a transaction path anomaly score.

[0010] The steps of encrypting the financial voucher that has passed the validity judgment in blocks, storing the encrypted data in a Merkle tree structure, deploying a smart contract that supports zero-knowledge proof in a distributed network to verify the authenticity of the financial voucher, and writing the evidence information of the financial voucher into the blockchain network based on a consensus mechanism include: Structurally dividing the financial voucher according to the subject dimension, amount dimension, and time dimension to obtain multiple data blocks, generating block indexes for the data blocks, and establishing a mapping relationship between the data blocks and the block indexes; generating a data block encryption key based on a master key, the block index, and a random salt value, performing authenticated encryption on the data block and additional authentication data using the data block encryption key to obtain a ciphertext block, and combining the ciphertext block with an initialization vector and an authentication tag to generate an encrypted data block; Perform hash calculation on the encrypted data block and metadata information to obtain a leaf node hash value, perform pairing calculation on adjacent leaf node hash values ​​to obtain a parent node hash value, perform recursive calculation until a root hash value is generated, and generate a Merkle proof for all nodes on the path from the leaf node hash value to the root hash value; Constructing an arithmetic circuit based on the encryption verification process, substituting the encrypted data block and the leaf node hash value as public inputs, and the proof key and the proof secret information as private inputs into the arithmetic circuit to generate a zero-knowledge proof, and deploying the zero-knowledge proof and the root hash value to a verification smart contract; The distributed nodes verify the zero-knowledge proof and the root hash value based on the verification key to obtain a verification result, and perform weighted calculation on the verification result according to the weight of each node to obtain a comprehensive verification score; The root hash value, timestamp and signature information are constructed into a proof transaction, and the proof transaction is packaged into a block based on a practical Byzantine fault-tolerant consensus mechanism and broadcast synchronously in the blockchain network.

[0011] The steps of substituting the encrypted data block and the leaf node hash value as public inputs, and the proof key and the proof secret information as private inputs into the arithmetic circuit to generate a zero-knowledge proof, and deploying the zero-knowledge proof and the root hash value to a verification smart contract include: Building an arithmetic circuit based on the encryption verification process, the arithmetic circuit including a data verification layer, an encryption verification layer, and a hash verification layer; Inputting the encrypted data block into the data verification layer and the encryption verification layer to generate data integrity constraints and cryptographic operation constraints respectively, and inputting the leaf node hash value into the hash verification layer to generate a Merkle path constraint; Generating sub-proofs for the data integrity constraint, cryptographic operation constraint, and Merkle path constraint based on the proof key and the proof secret information, and combining the sub-proofs to generate the original proof; Calculating a proof complexity based on the circuit size and witness data dimensions of the original proof, determining a proof segment size based on the proof complexity and a preset memory limit, and dividing the original proof into multiple proof segments according to the proof segment size; For each proof segment, the number of proof items and the relevance of the proof items are calculated, the proof segment is restructured based on the number of proof items and the relevance of the proof items to obtain an optimized proof segment, and the optimized proof segments are recursively merged according to a tree structure to obtain a merged proof; Calculating the proof feature of the combined proof, and performing directionally compressing the combined proof based on the proof feature to obtain a compressed proof; Obtaining static assessment results of circuit scale and witness data dimensions, obtaining dynamic assessment results of threat type and historical security events, and weightedly combining the static and dynamic assessment results to obtain a security level; Dynamically adjust security parameters according to the security level, perform security enhancement on the compressed proof based on the security parameters to obtain a final proof, and deploy the final proof to the verification smart contract.

[0012] A second aspect of an embodiment of the present invention provides a system for automatically generating and verifying financial vouchers based on intelligent accounting, comprising: The first unit is configured to scan and identify a business document, obtain transaction information from the business document, and convert the transaction information into standardized transaction data including transaction amount, transaction time, counterparty information, transaction type, product information, and service information; The second unit is configured to extract semantic features from the standardized transaction data using natural language processing technology, perform similarity matching between the extracted semantic features and a preset accounting subject feature library, and generate initial accounting entries; construct a transaction relationship network, identify transaction anomalies through graph structure analysis, optimize the initial accounting entries, and generate final accounting entries; The third unit is configured to generate a financial voucher for the final accounting entry, analyze the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyze the transaction path involved in the financial voucher through graph computing, calculate the rule matching degree of the financial voucher based on a preset accounting standard rule library, and determine the validity of the financial voucher; The fourth unit is configured to encrypt the financial voucher that has passed the validity judgment in blocks, store the encrypted data using a Merkle tree structure, deploy a smart contract that supports zero-knowledge proof in a distributed network to verify the authenticity of the financial voucher, and write the stored information of the financial voucher into the blockchain network based on a consensus mechanism; The fifth unit is used to write the financial voucher that has completed the evidence storage into the enterprise financial accounting system, and write the verification record, evidence storage record, and accounting record of the financial voucher into the blockchain network.

[0013] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0015] The method for automatic generation and verification of financial vouchers based on intelligent accounting provided by the present invention intelligently identifies business documents and converts them into standardized transaction data, and automatically generates accurate accounting entries by combining natural language processing and graph structure analysis technology, thereby greatly improving the degree of automation and accuracy of financial accounting and effectively reducing the workload and error rate of manual operations.

[0016] This invention uses deep neural networks and graph computing technology to perform multi-dimensional verification of financial vouchers. Based on the preset accounting standards rule library, the matching degree of calculation rules is calculated to build a comprehensive and rigorous voucher validity judgment mechanism, which can timely detect and prevent financial errors and fraudulent behaviors, and significantly improve the reliability and compliance of financial data.

[0017] This invention combines blockchain technology with zero-knowledge proof to perform distributed storage and verification of financial credentials, achieving the immutability and traceability of credential data while protecting the security of sensitive financial information. It provides efficient, secure, and transparent technical support for corporate financial management, and effectively solves the data security and trust issues in traditional financial systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the process of automatically generating and verifying financial vouchers based on intelligent accounting according to an embodiment of the present invention; Figure 2 Flowchart for blockchain storage and zero-knowledge proof verification of financial vouchers. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG is a flow chart of a method for automatically generating and verifying financial vouchers based on intelligent accounting according to an embodiment of the present invention. Figure 1 As shown, the method includes: Scan and identify business documents, obtain transaction information from the business documents, and convert the transaction information into standardized transaction data including transaction amount, transaction time, counterparty information, transaction type, product information, and service information; Using natural language processing technology to extract semantic features from the standardized transaction data, similarity matching the extracted semantic features with a preset accounting subject feature library is performed to generate initial accounting entries; a transaction relationship network is constructed, transaction anomalies are identified through graph structure analysis, the initial accounting entries are optimized, and final accounting entries are generated; Generate a financial voucher for the final accounting entry, analyze the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyze the transaction path involved in the financial voucher through graph computing, calculate the rule matching degree of the financial voucher based on a preset accounting standard rule library, and determine the validity of the financial voucher; The financial voucher that has passed the validity judgment is encrypted in blocks, and the encrypted data is stored in a Merkle tree structure. A smart contract that supports zero-knowledge proof is deployed in a distributed network to verify the authenticity of the financial voucher. The evidence information of the financial voucher is written into the blockchain network based on a consensus mechanism; The financial vouchers that have completed the notarization are written into the enterprise's financial accounting, and the verification records, notarization records, and accounting records of the financial vouchers are written into the blockchain network.

[0022] In an optional embodiment, the steps of extracting semantic features from the standardized transaction data using natural language processing technology, performing similarity matching on the extracted semantic features with a preset accounting subject feature library, and generating initial accounting entries include: Performing word segmentation and part-of-speech tagging on the standardized transaction data, extracting transaction entity information, and generating a semantic feature vector; Construct a multi-layer subject decision tree according to accounting standards, set the first-level accounting subject as the root node of the multi-layer subject decision tree, set the detailed accounting subject as the leaf node of the multi-layer subject decision tree, and establish an accounting subject attribution relationship between the root node and the leaf node; Starting from the root node, the semantic feature vector is matched with each layer of accounting subjects, the matching probability of each layer of accounting subjects is calculated, and the matching weight of the next layer of accounting subjects is dynamically adjusted based on the matching results of each layer to generate an accounting subject matching sequence; Calculating the semantic similarity between the semantic feature vector and the preset accounting subject feature library, calculating the business similarity between the transaction rule feature and the preset accounting subject feature library, analyzing the causal relationship strength between the transaction feature and the accounting subject selection, and weightedly fusing the semantic similarity, business similarity, and causal relationship strength to obtain a comprehensive similarity; Determine the final matching accounting subject based on the accounting subject matching sequence and the comprehensive similarity, determine the debit and credit direction of the accounting subject based on the causal relationship strength, and generate initial accounting entries; The matching confidence of the initial accounting entry is calculated, and when the matching confidence is lower than a preset confidence threshold, the initial accounting entry is marked as pending manual review.

[0023] For example, when performing basic text processing on standardized transaction data, conventional lexical analysis methods are used to segment the transaction description text into word units. For example, the transaction description "Pay office rent 5,000 yuan" is segmented into basic units such as "pay," "office," "rent," and "5,000 yuan." Rule matching is then used to identify the parts of speech of these words, such as "pay" as a verb, "office" and "rent" as nouns, and "5,000 yuan" as a numerical expression. Based on this basic lexical information, key transaction information such as the type of action, the transaction object, and the amount are extracted according to predefined templates. These elements are mapped to numerical representations, such as using simple one-hot encoding or word frequency statistics to represent the characteristics of each element. For example, "rent" is represented as a position index in a preset vocabulary. Finally, these basic features are organized into feature vectors according to a predetermined format for subsequent subject matching.

[0024] When constructing a multi-level account decision tree according to accounting standards, first-level accounting accounts (such as assets, liabilities, owner's equity, costs, expenses, and revenue) are set as the tree's root node. Second-level accounting accounts (such as current assets and non-current assets under assets) are set as second-level nodes. Third-level and lower-level detailed accounts (such as cash, bank deposits, and accounts receivable under current assets) are set as deeper nodes, until the most specific detailed account becomes a leaf node. Account attribution relationships are established between nodes to ensure that the account structure complies with the hierarchical relationships specified in accounting standards. For example, the "Administrative Expenses - Office Expenses" node is a child node of the "Administrative Expenses" node, which in turn is a child node of the "Expenses" node. The decision tree is stored in the database and contains fields such as account code, account name, and parent account ID to facilitate account matching and traversal.

[0025] When matching accounts starting from the root node, the matching probability between the semantic feature vector and the first-level accounting account is first calculated. For example, for the transaction "Pay Office Rent," the matching probability between the transaction and the first-level accounts "Assets" and "Expenses" is calculated. The matching probability for "Expenses" is 0.85, and the matching probability for "Assets" is 0.15. Based on the matching results for the first-level accounts, the matching weights for the second-level accounts are adjusted. Because the matching probability for "Expenses" is high, the weights for expense-related second-level accounts such as "Administrative Expenses" and "Selling Expenses" are increased. In the second-level account matching, the matching probability for "Administrative Expenses" is calculated to be 0.78, and the matching probability for "Selling Expenses" is 0.22. Continuing down the hierarchy, the matching probability for the leaf node "Office Expenses" is finally found under "Administrative Expenses," with a matching probability of 0.92. This hierarchical matching method generates an account matching sequence: "Expenses → Administrative Expenses → Office Expenses," which records the optimal matching path from the root node to the leaf node.

[0026] The cosine similarity between the semantic feature vector and the features of each account in the preset accounting account feature library is calculated. The accounting account feature library contains features such as each account's name, definition, scope of application, and common business descriptions. For example, the cosine similarity between the semantic feature vector of "paying office rent" and the features of the "office expenses" account is 0.88. Next, the business rule features of the transaction, such as transaction type, counterparty, and amount range, are analyzed and matched with the business rules of the accounting account to calculate business similarity. For example, the business similarity between transactions such as rent payments and the "office expenses" account is 0.85. Furthermore, the causal relationship strength between transaction features and accounting account selection is analyzed based on historical data. For example, the causal relationship strength between "rent" and "office expenses" is 0.90, indicating that rent transactions are highly likely to be included in the office expenses account. The semantic similarity (0.88), business similarity (0.85), and causal relationship strength (0.90) are weighted and fused according to preset weights (e.g., 0.3, 0.3, and 0.4), resulting in a comprehensive similarity of 0.88.

[0027] When determining the final accounting account, the account matching sequence and overall similarity are comprehensively considered, and the leaf node account with the highest overall similarity is selected as the final matching result. In this example, "Administrative Expenses - Office Expenses" is selected as the final matching account. Based on causal relationship analysis, the debit and credit directions of the account are determined. For example, if analysis shows that rent payments should be debited to "Administrative Expenses - Office Expenses" and credited to "Bank Deposits," the initial accounting entry is generated: Debit: Administrative Expenses - Office Expenses 5,000 yuan, Credit: Bank Deposits 5,000 yuan.

[0028] The matching confidence level of the initial accounting entry is calculated, taking into account factors such as the subject matching probability, overall similarity, and historical matching accuracy. For example, the matching confidence level for this accounting entry is calculated to be 0.86. When the matching confidence level falls below a preset threshold (such as 0.75), the entry is marked for manual review. In this example, because the matching confidence level of 0.86 is higher than the threshold of 0.75, the accounting entry is automatically approved and does not require manual review. For cases with lower confidence levels, such as a debit entry for "Administrative Expenses - Office Expenses" generated for a transaction "Purchase of Office Equipment for 10,000 Yuan," the confidence level is only 0.62, and the entry is marked for manual review, prompting the user to record it as "Fixed Assets" rather than "Office Expenses."

[0029] This invention achieves precise matching of accounting subjects through a multi-layer subject decision tree and comprehensive similarity calculation, significantly improving the accuracy of accounting entry generation. The weighted fusion of semantic features, business similarity, and causal strength addresses the inability of traditional methods to handle complex business contexts. Furthermore, the confidence assessment mechanism intelligently identifies situations requiring human intervention, ensuring the reliability of accounting processing while maintaining automated processing efficiency, and providing more accurate subject matching capabilities for corporate financial management.

[0030] In an optional embodiment, the steps of constructing a transaction relationship network, identifying transaction anomalies through graph structure analysis, optimizing the initial accounting entries, and generating final accounting entries include: Build a transaction relationship network, construct transaction-related information into nodes, and establish edges based on transaction associations; Calculating a weighted degree distribution feature for each node in the transaction relationship network, calculating a path importance feature based on the cumulative edge weights of the transaction path, calculating the network density of the transaction-related subgraph to obtain a global topological feature, and fusing the weighted degree distribution feature, path importance feature, and global topological feature to obtain a graph structure feature; The graph structure features are used to calculate the link anomaly score of the transaction path, the fluctuation anomaly score of the transaction amount, and the time series anomaly score of the transaction frequency. The link anomaly score, the fluctuation anomaly score, and the time series anomaly score are weighted and combined, and adaptively normalized based on the transaction scale to obtain the anomaly score. For transactions with anomaly scores exceeding a first preset threshold, correction rules are generated using historical normal transaction patterns, and the initial accounting entries are rematched and amounts redistributed based on the correction rules. For transactions with anomaly scores exceeding a second preset threshold, an expert rule library is introduced for in-depth optimization to generate optimized accounting entries. The basic compliance score of the optimized accounting entry is calculated based on the accounting standards rules, and the transaction credibility score is calculated in combination with the anomaly score. The basic compliance score and the transaction credibility score are integrated to generate an entry quality assessment value. When the entry quality assessment value is greater than the preset quality threshold, the optimized accounting entry is determined as the final accounting entry.

[0031] For example, a transaction relationship network is constructed. For each transaction, information such as the transaction subject (such as supplier, customer), transaction subject, and transaction time is constructed as a node. Node attributes include subject type, industry attribute, registered capital, etc. Directed edges are established based on the transaction relationship, and edge attributes include transaction amount, transaction frequency, and transaction time. For example, when enterprise A pays enterprise B 1 million yuan for goods, an edge from A to B is established in the network with an edge weight of 1 million yuan, an attribute of "goods payment", and a time stamp of the specific transaction date.

[0032] The in-degree weighted degree represents the total amount of transactions received by a node, while the out-degree weighted degree represents the total amount of transactions initiated by the node. For example, a supplier node's in-degree weighted degree of 5 million yuan represents the total amount of payments received by the supplier, while its out-degree weighted degree of 3 million yuan represents the total amount of payments made by the supplier. By calculating the weighted degree distribution ratio of each node, nodes with abnormal transaction amounts can be identified. For example, if 95% of a company's transactions are concentrated with two suppliers, and these two suppliers have significant financial transactions with another related entity, this concentration abnormality should be flagged.

[0033] For each transaction path in the network, the weights of each edge along the path are accumulated and normalized based on the path length. For example, if the funds flow path from enterprise A to enterprise C through intermediate enterprise B is 800,000 yuan and the transaction amount from A to B is 780,000 yuan, and the time interval is short (e.g., within one day), then the importance of this path is high, indicating the presence of funds transfer or related-party transactions. The path importance score is calculated as (80 + 78) / (2 × average enterprise transaction amount) × time decay factor. The time decay factor is determined by the time interval between transactions, with shorter intervals increasing the factor.

[0034] Global topological features are calculated. For each trading entity, a subgraph is constructed based on its n-hop neighborhood. The ratio of the actual number of edges in the subgraph to the theoretical maximum number of edges is calculated. For example, if a company and its 10 trading partners have a large number of trading relationships among these partners (e.g., 45 edges, while the theoretical maximum number is 10 × 9 / 2 = 45), then the network density is 1, indicating a high concentration of trading relationships and the presence of a related-party trading ring.

[0035] A weighted average approach is used to calculate the weighted degree distribution, path importance, and global topology features, assigning weights to different features (e.g., 0.4 for the weighted degree distribution, 0.35 for the path importance, and 0.25 for the global topology) to generate comprehensive graph structural features. For nodes or edges with abnormal feature values, anomaly scores are further calculated.

[0036] Graph structure features are used to calculate transaction anomaly scores. Link anomaly scores are calculated based on the degree to which transaction path patterns deviate from historical norms. For example, if a company has never transacted with company X, but suddenly experiences a large transaction, and company X immediately transfers the funds to company Y, the link anomaly score will be high. The volatility anomaly score measures the degree of fluctuation in transaction amounts relative to the historical average. For example, if the transaction amount in a particular month suddenly increases by 500%, the volatility anomaly score will be high. The time series anomaly score is based on changes in transaction frequency. For example, if large payments that were once quarterly suddenly become weekly, the time series anomaly score will increase.

[0037] The link anomaly score, fluctuation anomaly score, and timing anomaly score are weighted and combined with weights of 0.3, 0.4, and 0.3, respectively. The final anomaly score is then adaptively normalized based on transaction size. For large transactions (e.g., exceeding five times the company's average monthly transaction volume), the normalization factor is adjusted to 0.8, making it easier to detect large, abnormal transactions.

[0038] For transactions with anomaly scores exceeding the first preset threshold (e.g., 0.75), correction rules are generated based on historical normal transaction patterns. For example, if a "management expense" expense is 10 times the historical average and the payment is to a newly added supplier, the transaction description and amount characteristics will be used to determine that the transaction is more consistent with "fixed asset acquisition" and the account will be reclassified from "management expense" to "fixed asset." Amount redistribution involves splitting or combining unreasonable amounts according to accounting standards. For example, an abnormal advertising expense of 5 million yuan could be split into reasonable advertising expenses of 3 million yuan and service fees of 2 million yuan. For transactions with anomaly scores exceeding the second preset threshold (e.g., 0.9), an expert rule base is introduced for in-depth optimization. This pre-built, multi-layered expert rule base contains specialized processing rules for different industries and business scenarios. For example, for cross-border related-party transactions, the rule base includes transfer pricing assessment standards, industry profit margin reference values, and tax compliance requirements. When a multinational company is detected to have paid a service fee of 20 million yuan to an overseas related party, the system will match applicable rules such as "related party service fees should not exceed 20% of the market price for similar entities" and determine the reasonable range to be 15-17 million yuan based on industry benchmark data. The excess amount will then be adjusted to the "investment" or "profit distribution" account, and corresponding adjustment basis and handling suggestions will be generated for reference by financial personnel.

[0039] For optimized accounting entries, we check whether account usage complies with accounting account setup specifications, whether debit / credit balances are met, and whether attached documents are complete. A compliance score between 0 and 1 is assigned. The transaction credibility score is calculated based on the anomaly score: credibility score = (1 - anomaly score) × transaction size adjustment factor, where the transaction size adjustment factor decreases with increasing transaction amount. Combining the basic compliance score and transaction credibility score, the calculation formula is: Entry Quality Assessment Value = 0.6 × Basic Compliance Score + 0.4 × Transaction Credibility Score. When the entry quality assessment value exceeds a preset quality threshold (e.g., 0.85), the optimized accounting entry is finalized.

[0040] Existing financial anomaly detection methods primarily rely on statistical thresholds or simple rule matching. These methods lack in-depth analysis of transaction dependencies and are difficult to detect complex anomaly patterns, particularly hidden risks such as related transactions involving multiple parties or capital circulation. This paper addresses the limitations of existing technologies by proposing a graph structure analysis method centered on transaction relationship networks. This method transforms dispersed transaction data into structured network relationships and reveals potential anomaly patterns through node features, path analysis, and topological structure analysis. Its innovation lies in the multidimensional fusion of weighted degree distribution features, path importance features, and global topological features to achieve comprehensive perception of transaction anomalies. Furthermore, a graded threshold strategy distinguishes between different levels of anomalies, employing targeted optimization measures to avoid misjudgments caused by simplistic approaches. This significantly improves the accuracy and comprehensiveness of transaction anomaly identification. Compared with traditional methods, it can discover complex capital flow patterns and related transaction networks, thus preventing the risk of financial fraud. The multidimensional feature fusion mechanism improves the robustness of anomaly detection and reduces the false alarm rate. A graded optimization strategy based on the degree of anomaly enables precise adjustment of accounting entries, improving the quality of financial data. The quality assessment mechanism ensures that the optimization results comply with accounting standards, enhances the reliability of financial reports, and provides strong support for corporate financial risk management.

[0041] In an optional embodiment, generating a financial voucher for the final accounting entry, analyzing the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyzing the transaction path involved in the financial voucher through graph computing, and calculating the rule matching degree of the financial voucher based on a preset accounting standard rule library, the step of determining the validity of the financial voucher includes: Generating a financial voucher for the final accounting entry, and converting the subject information and amount information of the financial voucher into a subject feature vector and an amount feature vector respectively; Calculating attention weights for subject nodes in the subject feature vector, and calculating the association strength between subject nodes based on the attention weights to obtain a subject attention vector; using a gated recurrent unit to fuse the context information of historical credentials with the current credential state to obtain a context vector; and concatenating the subject attention vector with the context vector to obtain a subject logical relationship score; Performing an amount matching and integrity check on the financial voucher to generate an amount matching score and a voucher integrity score; By constructing an initial transaction graph and performing temporal expansion, a dynamic transaction relationship graph with temporal connections is obtained. The node features in the graph are iteratively updated based on an attention propagation mechanism based on transaction size, transaction timing, and transaction frequency. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths and perform sub-path decomposition. Sub-path scores are calculated by combining structural importance and temporal correlation, and transaction path anomaly scores are obtained through deep fusion. Calculating the rule matching degree of the financial voucher based on a preset accounting standard rule base; Adaptively weighted fusion of the subject logic relationship score, amount matching relationship score, voucher integrity score, transaction path anomaly score, and rule matching degree is performed to generate a comprehensive score for the financial voucher; When the comprehensive score is greater than a preset score threshold, the financial voucher is determined to be valid.

[0042] For example, when generating financial vouchers for the final accounting entries, the account information in the vouchers is converted into account feature vectors, and the amount information is converted into amount feature vectors. Account information can include debit accounts, credit accounts, and account levels. This discrete information is mapped into a continuous vector space through the embedding layer. For example, the "accounts receivable" account can be represented as a vector of [0.32, -0.45, 0.78, 0.21]. Amount information is vectorized based on features such as the amount size and its positive or negative sign. For example, 10,000 yuan can be represented as [0.10, 0.35, -0.42].

[0043] A multi-head self-attention mechanism is used to calculate attention weights for account feature vectors. Each account node is treated as a node in the graph, and the strength of the association between nodes is calculated. The account feature vector is converted into a query vector, a key vector, and a value vector using the query matrix, key matrix, and value matrix. For the two accounts "Sales Revenue" and "Accounts Receivable," the attention weight between them is calculated to be 0.85, indicating a high association; while the attention weight between "Sales Revenue" and "Fixed Assets" is only 0.12, indicating a low association. Based on the calculated attention weights, the account attention vector is obtained.

[0044] A gated recurrent unit is used to process historical voucher information. For a sales voucher, contextual information from similar historical transactions is incorporated. For example, the company's sales voucher records from the past three months are encoded into a state vector and fused with the current voucher state to generate a context vector that captures historical transaction patterns. The subject attention vector and the context vector are concatenated and passed through a fully connected layer to obtain a subject logical relationship score. For example, a normal sales voucher scores 0.92, while a voucher that does not conform to the standard subject combination scores 0.35.

[0045] Financial vouchers are checked for amount matching. Normal vouchers should have equal debit and credit amounts. The total difference between the debit and credit amounts is calculated and normalized to generate an amount matching score. For example, a voucher with balanced debits and credits receives a score of 1.0, while one with minor discrepancies receives a score of 0.95. Vouchers are also checked for completeness, including the completion of required fields and attachments. This generates a voucher completeness score, such as 0.88, indicating that the voucher is generally complete but with room for improvement.

[0046] By constructing a dynamic graph of transaction relationships, we analyze the transaction paths involved in the vouchers. Taking inter-enterprise fund transactions as an example, we consider each enterprise as a node in the graph and transactions as edges to construct an initial transaction graph. Through time expansion, we link transactions at different time points to form a dynamic graph. For transactions with similar payments to the same supplier for three consecutive months, we identify temporal connections. We design an attention propagation mechanism based on transaction size, transaction timing, and transaction frequency, assigning higher attention weights to large, frequent, or tightly timed transactions. We iteratively update node features through multiple rounds of message passing.

[0047] A bidirectional search strategy is used to identify transaction paths within the dynamic transaction relationship graph. For example, if Company A transfers funds to Company B, followed by Company B immediately transferring funds to Company C, a forward search is performed starting from A and a reverse search is performed starting from C, intersecting at B to identify the complete path A→B→C. This path is then decomposed into two sub-paths, A→B and B→C, and sub-path scores are calculated based on structural importance and temporal relevance. For example, if the transaction A→B is 1 million yuan and is the first transaction, the score is 0.75; if the transaction B→C is 980,000 yuan and occurs within a very short period of time, the score is 0.35, indicating high anomaly. By deeply integrating the sub-path scores, an overall transaction path anomaly score is generated, such as 0.40, indicating a high anomaly risk.

[0048] The system evaluates the rule-matching degree of financial documents based on a pre-set accounting standards rule base. This rule base contains applicable accounting standards and internal corporate policies, such as the rule that "depreciation of fixed assets should be charged to the corresponding cost and expense account." Documents are checked for compliance with these rules and the rule-matching degree is calculated. For example, a document that fully complies with the rules has a matching degree of 0.95, while a document that violates a rule has a matching degree of 0.60.

[0049] Adaptively weight the account logic relationship score, amount matching relationship score, document integrity score, transaction path anomaly score, and rule matching degree to achieve a fusion score. Dynamically adjust the weight of each indicator based on different business scenarios. For example, for routine business documents, the account logic relationship weight is 0.3, the amount matching relationship weight is 0.25, the document integrity weight is 0.15, the transaction path anomaly weight is 0.1, and the rule matching degree weight is 0.2. For cross-border transaction documents, the transaction path anomaly weight is increased to 0.25. The weighted average is calculated to obtain a comprehensive score, such as 0.85.

[0050] Based on historical data and risk tolerance, a scoring threshold, such as 0.75, is set. Credentials above this threshold are considered valid and can proceed to the next processing step. Credentials below this threshold are flagged as potential anomalies and require manual review or further analysis.

[0051] This paper establishes a comprehensive mechanism for verifying the validity of financial vouchers. By analyzing logical relationships between accounts, amount matching, and voucher integrity through deep neural networks and combining graph computing technology to analyze transaction paths, this system comprehensively assesses the compliance of financial vouchers. An adaptive weighted fusion strategy for multi-dimensional scoring dynamically adjusts the importance of various indicators based on different business scenarios, improving the accuracy and adaptability of verification results, effectively preventing the risk of financial fraud, and ensuring the authenticity and integrity of corporate financial data.

[0052] In an optional embodiment, an initial transaction graph is constructed and time-series expanded to obtain a dynamic transaction relationship graph with time-series connections. The node features in the graph are iteratively updated based on an attention propagation mechanism based on transaction scale, transaction sequence, and transaction frequency. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths and perform sub-path decomposition. Sub-path scores are calculated by combining structural importance and time-series correlation. The steps of obtaining a transaction path anomaly score through deep fusion include: Constructing an initial transaction graph based on the subject feature vector and the amount feature vector, setting the transaction entities in the voucher as graph nodes, and setting the fund transaction relationships between the transaction entities as graph edges, wherein the initial weights of the graph edges are determined according to the transaction amount and transaction frequency; Performing temporal expansion on the initial transaction graph, establishing temporal connections between nodes of the same transaction entity in different time windows, and calculating dynamic weights of the temporal connections based on historical transaction data; Node features are iteratively updated using attention propagation. For each node, an attention coefficient is calculated based on the transaction scale, transaction sequence, and transaction frequency of its neighboring nodes. The neighboring node features are weighted and aggregated according to the attention coefficient. The aggregated features are then combined with the node's own features to generate updated node features. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths, decompose the transaction paths into transaction sub-paths, and calculate the structural importance score and temporal correlation score for each sub-path. The structural importance score, temporal correlation score and feature sequence of the sub-path are deeply fused and calculated with the feature sequence of the path nodes to generate a transaction path anomaly score.

[0053] For example, the subject feature vector and amount feature vector are extracted from the voucher data provided by the enterprise, and then the initial transaction graph is constructed, time series expansion is performed to form a dynamic graph, node features are updated using attention propagation, and finally abnormal transaction paths are identified through bidirectional search.

[0054] When constructing the initial transaction graph, the accounts in each transaction voucher are first parsed, and transaction entities such as companies, individuals, and bank accounts are assigned as graph nodes. Each node is assigned initial attributes, including node type (internal corporate account, external corporate account, personal account, etc.) and initial balance. The fund flows recorded in the voucher are assigned as graph edges, with the direction of the edge representing the flow of funds. The initial weight of the edge is determined based on the transaction amount and transaction frequency. This is calculated by dividing the single transaction amount by the average transaction amount during the accounting period to obtain a normalized amount, and then multiplying the amount by the number of transactions for the transaction pair during the observation period. For example, if Company A transfers 1 million yuan to Company B, and the average transaction amount during the period is 500,000 yuan, and there are three transactions between Company A and Company B during the observation period, the initial weight of the edge is 2 × 3 = 6.

[0055] When performing temporal expansion on the initial transaction graph, a one-month time window is set, and temporal connections are established between nodes of the same transaction entity within adjacent windows. The dynamic weight of the temporal connection is calculated based on historical transaction data, specifically the ratio of the deviation between the transaction amount in the current time window and the historical average transaction amount. For example, if the total transaction amount of Company A to Company B in the current month is 3 million yuan, and the monthly average transaction amount over the past six months is 2 million yuan, the dynamic weight of the temporal connection is (300-200) / 200=0.5, indicating that the current transaction intensity is 50% higher than the historical average.

[0056] When iteratively updating node features using the attention propagation mechanism, three factors are considered for each node regarding its neighboring nodes: transaction size, transaction timing, and transaction frequency. Transaction size is represented by the ratio of transaction amount to the enterprise's overall transaction volume. Transaction timing considers the time interval pattern of transactions, converting changes in transaction frequency over the past seven days, one month, and three months into time series features. Transaction frequency is the number of transactions per unit time. An attention coefficient is calculated based on these three factors. For example, for node A's neighbor B, if the transaction amount between A and B accounts for 30% of A's total transaction amount and the transaction frequency has recently increased significantly, B is given a higher attention weight. Specifically, the normalized values ​​of these three factors are summed and subjected to a nonlinear transformation to obtain the attention coefficient. These coefficients are then used to perform a weighted aggregation of the neighboring node features. For example, if node A has three neighbors, B, C, and D, with attention coefficients of 0.5, 0.3, and 0.2, respectively, then A's updated feature is the weighted sum of these three neighboring features according to these ratios, which is then concatenated with A's own features to generate the final feature. After three rounds of iterative updates, the node features fully incorporate network structure information.

[0057] A bidirectional search strategy is used to identify abnormal transaction paths within the dynamic transaction relationship graph. The search begins simultaneously at the suspicious source and destination of funds, with a depth limit of four hops. When the search paths in both directions intersect, a complete transaction path is formed. The identified transaction path is decomposed into subpaths of length two. For each subpath, a structural importance score and a temporal relevance score are calculated. Structural importance is assessed by the centrality and connection strength of the subpath within the entire graph. Subpaths with nodes in key positions and significant transaction amounts are considered to have high structural importance. Temporal relevance measures the temporal correlation of transactions within a subpath. If two transactions are very close in time (for example, less than 24 hours apart), the temporal relevance is high.

[0058] For example, suppose a transaction path is detected: Company A → Company B → Individual C → Company D. This path is decomposed into three sub-paths: A → B, B → C, and C → D. For sub-path B → C, if B is a core enterprise with transactions with multiple entities, and C is a newly emerged individual account, this sub-path has a high structural importance score of 0.85. If the transfer from B to C occurs only two hours before the transfer from C to D, the temporal correlation score is 0.92. Combining these scores with the feature sequences of the path nodes and calculating them through a deep fusion network, the transaction path is ultimately assigned an anomaly score of 0.88, exceeding the set threshold of 0.75 and being identified as a suspicious related-party transaction.

[0059] Existing financial transaction analysis technologies often rely on fixed rules or simple statistical indicators when processing complex transaction paths, making it difficult to detect carefully designed fund transfer paths and abnormal transaction chains that are closely linked in time. This invention integrates the time dimension into graph structure analysis, capturing the temporal evolution of transaction patterns by establishing temporal connections between identical entity nodes within different time windows. An attention propagation mechanism based on transaction size, transaction timing, and transaction frequency is introduced to enable dynamic updating of node features and integration of network structure information. Furthermore, a bidirectional search and subpath decomposition strategy, combined with structural importance and temporal correlation assessment, enhances the accuracy of abnormal transaction path identification. This significantly improves the precision and efficiency of financial transaction anomaly detection.

[0060] In an optional embodiment, the financial voucher that has passed the validity judgment is encrypted in blocks, the encrypted data is stored in a Merkle tree structure, a smart contract that supports zero-knowledge proof is deployed in a distributed network to verify the authenticity of the financial voucher, and the steps of writing the stored information of the financial voucher into the blockchain network based on a consensus mechanism include: Structurally dividing the financial voucher according to the subject dimension, amount dimension, and time dimension to obtain multiple data blocks, generating block indexes for the data blocks, and establishing a mapping relationship between the data blocks and the block indexes; generating a data block encryption key based on a master key, the block index, and a random salt value, performing authenticated encryption on the data block and additional authentication data using the data block encryption key to obtain a ciphertext block, and combining the ciphertext block with an initialization vector and an authentication tag to generate an encrypted data block; Perform hash calculation on the encrypted data block and metadata information to obtain a leaf node hash value, perform pairing calculation on adjacent leaf node hash values ​​to obtain a parent node hash value, perform recursive calculation until a root hash value is generated, and generate a Merkle proof for all nodes on the path from the leaf node hash value to the root hash value; Constructing an arithmetic circuit based on the encryption verification process, substituting the encrypted data block and the leaf node hash value as public inputs, and the proof key and the proof secret information as private inputs into the arithmetic circuit to generate a zero-knowledge proof, and deploying the zero-knowledge proof and the root hash value to a verification smart contract; The distributed nodes verify the zero-knowledge proof and the root hash value based on the verification key to obtain a verification result, and perform weighted calculation on the verification result according to the weight of each node to obtain a comprehensive verification score; The root hash value, timestamp and signature information are constructed into a proof transaction, and the proof transaction is packaged into a block based on a practical Byzantine fault-tolerant consensus mechanism and broadcast synchronously in the blockchain network.

[0061] Figure 2 This is a flowchart for blockchain storage and zero-knowledge proof verification of financial documents. For example, financial documents are structured and divided into multiple data blocks based on account, amount, and time dimensions. For example, a document containing multiple accounting items can be divided into asset account data blocks, liability account data blocks, income account data blocks, and expense account data blocks. Based on the amount dimension, it can be divided into large-value transaction data blocks (e.g., those exceeding 100,000 yuan) and small-value transaction data blocks. Based on the time dimension, it can be divided into monthly data blocks, quarterly data blocks, and so on. A unique block index is generated for each data block, such as "Assets-2023Q1-Large" for large-value asset transaction data for the first quarter of 2023. A mapping table between data blocks and block indexes is established in the database to facilitate subsequent retrieval and verification.

[0062] A 256-bit master key (e.g., "a8b7c6d5e4f3g2h1i0j9k8l7m6n5o4p3") is generated, and a 32-bit random salt value (e.g., "r7d9e2f4") is assigned to each block index. A key derivation function, using the master key, block index, and random salt as input, generates a 128-bit encryption key specific to each data block. This encryption key is used to encrypt the data block and additional authentication data (e.g., document number, issuer information, etc.) in AEAD (authenticated encryption and decryption) mode, producing a ciphertext block. A 16-byte initialization vector (e.g., "9f8e7d6c5b4a3210") and a 16-byte authentication tag are generated. The ciphertext block is combined with the initialization vector and authentication tag to produce the final encrypted data block.

[0063] The encrypted data block and metadata are hashed to obtain the leaf node hash value. For example, for the encrypted "Asset-2023Q1-Large Amount" data block, the SHA-256 hash algorithm is used to calculate its hash value, which serves as the leaf node value (e.g., "7f83b1657ff1fc53b92dc18148a1d65dfc2d4b1fa3d677284addd200126d9069"). All leaf nodes are arranged in a predefined order, and the hash values ​​of adjacent leaf nodes are paired and concatenated. The hash values ​​are then calculated again to obtain the parent node hash value. For example, the hash values ​​of leaf nodes A and B are paired to obtain the hash value of parent node P. This process is recursively executed until a single root hash value is generated (e.g., "d82c4eb5261cb9c8aa9855edd67d1bd10482f41529858d925094d173fa662aa7"). A Merkle proof is generated containing all nodes on the path from the leaf node to the root hash value, which can be used to subsequently verify the integrity and attribution of a specific data block.

[0064] The encrypted data block and leaf node hash value are used as public inputs, while the proving key (e.g., "e5f6g7h8i9j0k1l2") and proving secret information (e.g., master key, random salt) are used as private inputs in an arithmetic circuit to generate a zero-knowledge proof. This proof demonstrates knowledge of all the information required to correctly encrypt the plaintext financial document into the ciphertext block without revealing this sensitive information. The generated zero-knowledge proof and root hash value are then deployed to a verification smart contract on the blockchain network, which contains verification logic in its code.

[0065] Distributed nodes cryptographically verify the zero-knowledge proof using a verification key (a 256-bit public key, such as "b9c8a76f5e4d3c2b1a0f9e8d7c6b5a4"). Each node executes the same verification algorithm: first, verifying the cryptographic integrity of the zero-knowledge proof, then checking that the root hash value is consistent with the value promised in the zero-knowledge proof. Node weights are determined based on a preset trust level: authoritative nodes (such as regulatory agencies and accounting firms) receive a weight of 0.5, core enterprise nodes receive a weight of 0.3, and ordinary verification nodes receive a weight of 0.1. Each node's verification result (1 for pass, 0 for fail) is multiplied by its corresponding weight and summed to create a comprehensive verification score. When the score exceeds a preset threshold of 1.5, the entire verification is considered passed.

[0066] A 32-byte root hash value, an 8-byte UTC timestamp, and a 64-byte ECDSA digital signature are packaged into a standard format for a proof transaction. In the PBFT consensus process, the master node first broadcasts a pre-prepare message containing the proof transaction. Verifying nodes check the transaction format and signature validity before broadcasting a prepare message. When the number of prepare messages exceeds 2f+1 (f is the number of malicious nodes allowed), the node broadcasts a confirmation message. Once a node receives 2f+1 confirmation messages, it considers the transaction to have reached consensus and is added to the block candidate set. When the number of candidate transactions reaches the preset block size or the time interval reaches 10 seconds, these transactions are packaged into blocks, the hash value of the previous block is added to form a chain structure, and then broadcast to all nodes via the P2P network. After verification, each node stores the blocks persistently in its local ledger.

[0067] This invention establishes a highly secure blockchain-based financial voucher storage mechanism. Through multi-dimensional data segmentation and authenticated encryption, it ensures the integrity and privacy of voucher data. The application of Merkle tree structures and Merkle proofs improves data verification efficiency. The introduction of zero-knowledge proof technology enables verification of voucher authenticity without revealing specific content, addressing the privacy protection issues inherent in traditional voucher storage methods. Distributed verification based on a consensus mechanism ensures the credibility of the voucher storage process, providing an unalterable trust foundation for financial data.

[0068] In an optional embodiment, the steps of substituting the encrypted data block and the leaf node hash value as public inputs, and the proof key and the proof secret information as private inputs into the arithmetic circuit to generate a zero-knowledge proof, and deploying the zero-knowledge proof and the root hash value to the verification smart contract include: Building an arithmetic circuit based on the encryption verification process, the arithmetic circuit including a data verification layer, an encryption verification layer, and a hash verification layer; Inputting the encrypted data block into the data verification layer and the encryption verification layer to generate data integrity constraints and cryptographic operation constraints respectively, and inputting the leaf node hash value into the hash verification layer to generate a Merkle path constraint; Generating sub-proofs for the data integrity constraint, cryptographic operation constraint, and Merkle path constraint based on the proof key and the proof secret information, and combining the sub-proofs to generate the original proof; Calculating a proof complexity based on the circuit size and witness data dimensions of the original proof, determining a proof segment size based on the proof complexity and a preset memory limit, and dividing the original proof into multiple proof segments according to the proof segment size; For each proof segment, the number of proof items and the relevance of the proof items are calculated, the proof segment is restructured based on the number of proof items and the relevance of the proof items to obtain an optimized proof segment, and the optimized proof segments are recursively merged according to a tree structure to obtain a merged proof; Calculating the proof feature of the combined proof, and performing directionally compressing the combined proof based on the proof feature to obtain a compressed proof; Obtaining static assessment results of circuit scale and witness data dimensions, obtaining dynamic assessment results of threat type and historical security events, and weightedly combining the static and dynamic assessment results to obtain a security level; Dynamically adjust security parameters according to the security level, perform security enhancement on the compressed proof based on the security parameters to obtain a final proof, and deploy the final proof to the verification smart contract.

[0069] For example, an arithmetic circuit is constructed based on the cryptographic verification process, comprising a data verification layer, an encryption verification layer, and a hash verification layer. This circuit design takes into account the requirements for data integrity verification, cryptographic operation verification, and Merkle tree path verification. For example, a blockchain containing 100 user transaction records, each of which includes fields such as user ID, transaction amount, and timestamp, has a total data volume of approximately 20KB.

[0070] When an encrypted data block is input into the data validation layer, the integrity of the data structure is checked to ensure that each field is correctly formatted and its value is within the valid range. For example, user IDs are 8-bit strings, transaction amounts are positive integers not exceeding 1,000,000 units, and timestamps conform to a standard time format. This process generates data integrity constraints, which include approximately 200 logical constraints.

[0071] The encrypted data block is also fed into the encryption verification layer, which uses the AES-256 encryption algorithm for cryptographic verification and generates cryptographic operation constraints. In this example, 16 rounds of iteration are used, each of which includes byte substitution, row shift, column mixing, and round key addition operations, generating a total of approximately 3,000 constraints related to the encryption operation.

[0072] The leaf node hash values ​​are fed into the hash verification layer, which uses the SHA-256 hash algorithm to calculate the Merkle tree path and generate Merkle path constraints. For a Merkle tree with 100 records and a height of 7, six hash operations from the leaf node to the root node need to be verified. Each hash operation generates approximately 1,000 constraints, for a total of approximately 6,000 constraints.

[0073] The proving key (containing the circuit's public parameters) and the proving secret (e.g., the user's private key) are used to generate sub-proofs. For the data integrity constraint, a 2KB sub-proof is generated; for the cryptographic operation constraint, a 30KB sub-proof is generated; and for the Merkle path constraint, a 60KB sub-proof is generated. These three sub-proofs are combined into a single 92KB original proof.

[0074] The proof complexity was calculated based on the circuit size of the original proof (approximately 9,200 constraints) and the witness data dimensions (approximately 3,000 variables), resulting in a complexity index of 27.6MB. Given the default memory limit of 64MB, each proof segment should be no larger than 16KB, so the 92KB original proof was divided into six segments.

[0075] For each proof segment, the number of proof items and the degree of correlation between them are calculated. For example, the first segment contains 32 proof items with an average correlation of 3.5; the second segment contains 28 proof items with an average correlation of 4.2. Based on these metrics, the proof segments are restructured, combining highly correlated proof items to obtain six optimized proof segments. These optimized segments are recursively merged according to a tree structure. First, segments 1 and 2 are merged into intermediate proof A, segments 3 and 4 are merged into intermediate proof B, segments 5 and 6 are merged into intermediate proof C, intermediate proofs A and B are then merged into intermediate proof D, and finally intermediate proofs D and C are merged to obtain a merged proof of 45KB.

[0076] The combined proof's characteristics were calculated, including its element distribution entropy (approximately 5.6), structural complexity (approximately 12.3), and redundancy (approximately 0.25). Based on these characteristics, a targeted compression algorithm was used to compress the combined proof, removing redundant information and optimizing its structure, resulting in a compressed proof of 28KB.

[0077] Parameters for the arithmetic circuit were obtained, including the number of constraints (9,200), the number of variables (3,000), the circuit depth (32 layers), and the witness vector length (12,800 bytes). A baseline risk value of 5.0 was set, with the risk value increasing by 0.1 for every 1,000 constraints added to the circuit size, 0.2 for every 5 layers of circuit depth, and 0.15 for every 2,000 bytes of witness data. The calculated static risk index was 6.8. A dynamic risk assessment was also conducted. Recent threat intelligence was obtained from the security monitoring center database, identifying major threat types such as replay attacks (12 occurrences in the past seven days), man-in-the-middle attacks (8 occurrences in the past seven days), and hash collision attacks (3 occurrences in the past seven days). A search of the historical security incident database revealed 23 incidents of exploitation of security vulnerabilities similar to zero-knowledge proofs in the past 30 days, five of which were classified as high-risk. The dynamic risk index was calculated to be 7.5 based on the frequency of threat types (weighted 0.4) and the severity of historical security incidents (weighted 0.6). Based on the current network environment and the importance of the verification task, the weighting ratio of static and dynamic assessments was determined to be 4:6. The weighted average was calculated as: 6.8 × 0.4 + 7.5 × 0.6 = 7.22. After rounding, the security level was 7.2. According to the security level standard (0-3 low risk, 3-6 medium risk, 6-8 high risk, 8-10 extremely high risk), this verification process was classified as "high risk."

[0078] Based on this security level, security parameters are dynamically adjusted, including increasing the number of random challenges from the default 16 to 24 and the number of verification rounds from 8 to 12. Based on these enhanced security parameters, the compressed proof is security-enhanced by adding additional verification information and tamper-evident markers, ultimately producing a final proof of 32KB. This final proof is deployed to the verification smart contract at the contract address 0x7B4A...E2F3 for on-chain verification. This approach achieves efficient and secure verification of encrypted data while maintaining data privacy.

[0079] A second aspect of an embodiment of the present invention provides a system for automatically generating and verifying financial vouchers based on intelligent accounting, comprising: The first unit is configured to scan and identify a business document, obtain transaction information from the business document, and convert the transaction information into standardized transaction data including transaction amount, transaction time, counterparty information, transaction type, product information, and service information; The second unit is configured to extract semantic features from the standardized transaction data using natural language processing technology, perform similarity matching between the extracted semantic features and a preset accounting subject feature library, and generate initial accounting entries; construct a transaction relationship network, identify transaction anomalies through graph structure analysis, optimize the initial accounting entries, and generate final accounting entries; The third unit is configured to generate a financial voucher for the final accounting entry, analyze the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyze the transaction path involved in the financial voucher through graph computing, calculate the rule matching degree of the financial voucher based on a preset accounting standard rule library, and determine the validity of the financial voucher; The fourth unit is configured to encrypt the financial voucher that has passed the validity judgment in blocks, store the encrypted data using a Merkle tree structure, deploy a smart contract that supports zero-knowledge proof in a distributed network to verify the authenticity of the financial voucher, and write the stored information of the financial voucher into the blockchain network based on a consensus mechanism; The fifth unit is used to write the financial voucher that has completed the evidence storage into the enterprise financial accounting system, and write the verification record, evidence storage record, and accounting record of the financial voucher into the blockchain network.

[0080] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0081] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0082] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically generating and verifying financial vouchers based on intelligent accounting, characterized in that: include: Scan and identify business documents, obtain transaction information from the business documents, and convert the transaction information into standardized transaction data including transaction amount, transaction time, counterparty information, transaction type, product information, and service information; Using natural language processing technology to extract semantic features from the standardized transaction data, similarity matching the extracted semantic features with a preset accounting subject feature library is performed to generate initial accounting entries; a transaction relationship network is constructed, transaction anomalies are identified through graph structure analysis, the initial accounting entries are optimized, and final accounting entries are generated; Generate a financial voucher for the final accounting entry, analyze the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyze the transaction path involved in the financial voucher through graph computing, calculate the rule matching degree of the financial voucher based on a preset accounting standard rule library, and determine the validity of the financial voucher; The financial voucher that has passed the validity judgment is encrypted in blocks, and the encrypted data is stored in a Merkle tree structure. A smart contract that supports zero-knowledge proof is deployed in a distributed network to verify the authenticity of the financial voucher. The evidence information of the financial voucher is written into the blockchain network based on a consensus mechanism; The financial vouchers that have completed the notarization are written into the enterprise financial accounting system, and the verification records, notarization records, and accounting records of the financial vouchers are written into the blockchain network.

2. The method according to claim 1, characterized in that The steps of extracting semantic features from the standardized transaction data using natural language processing technology, performing similarity matching between the extracted semantic features and a preset accounting subject feature library, and generating initial accounting entries include: Performing word segmentation and part-of-speech tagging on the standardized transaction data, extracting transaction entity information, and generating a semantic feature vector; Construct a multi-layer subject decision tree according to accounting standards, set the first-level accounting subject as the root node of the multi-layer subject decision tree, set the detailed accounting subject as the leaf node of the multi-layer subject decision tree, and establish an accounting subject attribution relationship between the root node and the leaf node; Starting from the root node, the semantic feature vector is matched with each layer of accounting subjects, the matching probability of each layer of accounting subjects is calculated, and the matching weight of the next layer of accounting subjects is dynamically adjusted based on the matching results of each layer to generate an accounting subject matching sequence; Calculating the semantic similarity between the semantic feature vector and the preset accounting subject feature library, calculating the business similarity between the transaction rule feature and the preset accounting subject feature library, analyzing the causal relationship strength between the transaction feature and the accounting subject selection, and weightedly fusing the semantic similarity, business similarity, and causal relationship strength to obtain a comprehensive similarity; Determine the final matching accounting subject based on the accounting subject matching sequence and the comprehensive similarity, determine the debit and credit direction of the accounting subject based on the causal relationship strength, and generate initial accounting entries; The matching confidence of the initial accounting entry is calculated, and when the matching confidence is lower than a preset confidence threshold, the initial accounting entry is marked as pending manual review.

3. The method according to claim 1, characterized in that The steps of constructing a transaction relationship network, identifying transaction anomalies through graph structure analysis, optimizing the initial accounting entries, and generating final accounting entries include: Build a transaction relationship network, construct transaction-related information into nodes, and establish edges based on transaction associations; Calculating a weighted degree distribution feature for each node in the transaction relationship network, calculating a path importance feature based on the cumulative edge weights of the transaction path, calculating the network density of the transaction-related subgraph to obtain a global topological feature, and fusing the weighted degree distribution feature, path importance feature, and global topological feature to obtain a graph structure feature; The graph structure features are used to calculate the link anomaly score of the transaction path, the fluctuation anomaly score of the transaction amount, and the time series anomaly score of the transaction frequency. The link anomaly score, the fluctuation anomaly score, and the time series anomaly score are weighted and combined, and adaptively normalized based on the transaction scale to obtain the anomaly score. For transactions with anomaly scores exceeding a first preset threshold, correction rules are generated using historical normal transaction patterns, and the initial accounting entries are rematched and amounts redistributed based on the correction rules. For transactions with anomaly scores exceeding a second preset threshold, an expert rule library is introduced for in-depth optimization to generate optimized accounting entries. The basic compliance score of the optimized accounting entry is calculated based on the accounting standards rules, and the transaction credibility score is calculated in combination with the anomaly score. The basic compliance score and the transaction credibility score are integrated to generate an entry quality assessment value. When the entry quality assessment value is greater than the preset quality threshold, the optimized accounting entry is determined as the final accounting entry.

4. The method according to claim 1, wherein The steps of generating a financial voucher for the final accounting entry, analyzing the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyzing the transaction path involved in the financial voucher through graph computing, and calculating the rule matching degree of the financial voucher based on a preset accounting standard rule library to determine the validity of the financial voucher include: Convert the account information and amount information of the financial voucher into the account feature vector and amount feature vector respectively; Calculating attention weights for subject nodes in the subject feature vector, and calculating the association strength between subject nodes based on the attention weights to obtain a subject attention vector; using a gated recurrent unit to fuse the context information of historical credentials with the current credential state to obtain a context vector; and concatenating the subject attention vector with the context vector to obtain a subject logical relationship score; Performing an amount matching and integrity check on the financial voucher to generate an amount matching score and a voucher integrity score; By constructing an initial transaction graph and performing temporal expansion, a dynamic transaction relationship graph with temporal connections is obtained. The node features in the graph are iteratively updated based on an attention propagation mechanism based on transaction size, transaction timing, and transaction frequency. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths and perform sub-path decomposition. Sub-path scores are calculated by combining structural importance and temporal correlation, and transaction path anomaly scores are obtained through deep fusion. Adaptively weighted fusion of the subject logic relationship score, amount matching relationship score, voucher integrity score, transaction path anomaly score, and rule matching degree is performed to generate a comprehensive score for the financial voucher; When the comprehensive score is greater than a preset score threshold, the financial voucher is determined to be valid.

5. The method according to claim 4, characterized in that By constructing an initial transaction graph and performing temporal expansion, a dynamic transaction relationship graph with temporal connections is obtained. The node features in the graph are iteratively updated based on the attention propagation mechanism based on transaction scale, transaction timing, and transaction frequency. The steps of using a bidirectional search strategy to identify transaction paths and decompose subpaths in the transaction relationship dynamic graph, calculating subpath scores by combining structural importance and temporal correlation, and obtaining transaction path anomaly scores through deep fusion include: Constructing an initial transaction graph based on the subject feature vector and the amount feature vector, setting the transaction entities in the voucher as graph nodes, and setting the fund transaction relationships between the transaction entities as graph edges, wherein the initial weights of the graph edges are determined according to the transaction amount and transaction frequency; Performing temporal expansion on the initial transaction graph, establishing temporal connections between nodes of the same transaction entity in different time windows, and calculating dynamic weights of the temporal connections based on historical transaction data; Node features are iteratively updated using attention propagation. For each node, an attention coefficient is calculated based on the transaction scale, transaction sequence, and transaction frequency of its neighboring nodes. The neighboring node features are weighted and aggregated according to the attention coefficient. The aggregated features are then combined with the node's own features to generate updated node features. A bidirectional search strategy is used in the transaction relationship dynamic graph to identify transaction paths, decompose the transaction paths into transaction sub-paths, and calculate the structural importance score and temporal correlation score for each sub-path. The structural importance score, temporal correlation score and feature sequence of the sub-path are deeply fused and calculated with the feature sequence of the path nodes to generate a transaction path anomaly score.

6. The method according to claim 1, characterized in that The steps of encrypting the financial voucher that has passed the validity judgment in blocks, storing the encrypted data in a Merkle tree structure, deploying a smart contract that supports zero-knowledge proof in a distributed network to verify the authenticity of the financial voucher, and writing the evidence information of the financial voucher into the blockchain network based on a consensus mechanism include: Structurally dividing the financial voucher according to the subject dimension, amount dimension, and time dimension to obtain multiple data blocks, generating block indexes for the data blocks, and establishing a mapping relationship between the data blocks and the block indexes; generating a data block encryption key based on a master key, the block index, and a random salt value, performing authenticated encryption on the data block and additional authentication data using the data block encryption key to obtain a ciphertext block, and combining the ciphertext block with an initialization vector and an authentication tag to generate an encrypted data block; Perform hash calculation on the encrypted data block and metadata information to obtain a leaf node hash value, perform pairing calculation on adjacent leaf node hash values ​​to obtain a parent node hash value, perform recursive calculation until a root hash value is generated, and generate a Merkle proof for all nodes on the path from the leaf node hash value to the root hash value; Constructing an arithmetic circuit based on the encryption verification process, substituting the encrypted data block and the leaf node hash value as public inputs, and the proof key and the proof secret information as private inputs into the arithmetic circuit to generate a zero-knowledge proof, and deploying the zero-knowledge proof and the root hash value to a verification smart contract; The distributed nodes verify the zero-knowledge proof and the root hash value based on the verification key to obtain a verification result, and perform weighted calculation on the verification result according to the weight of each node to obtain a comprehensive verification score; The root hash value, timestamp and signature information are constructed into a proof transaction, and the proof transaction is packaged into a block based on a practical Byzantine fault-tolerant consensus mechanism and broadcast synchronously in the blockchain network.

7. The method according to claim 6, characterized in that The steps of substituting the encrypted data block and the leaf node hash value as public inputs, and the proof key and the proof secret information as private inputs into the arithmetic circuit to generate a zero-knowledge proof, and deploying the zero-knowledge proof and the root hash value to a verification smart contract include: The arithmetic circuit includes a data verification layer, an encryption verification layer and a hash verification layer; Inputting the encrypted data block into the data verification layer and the encryption verification layer to generate data integrity constraints and cryptographic operation constraints respectively, and inputting the leaf node hash value into the hash verification layer to generate a Merkle path constraint; Generating sub-proofs for the data integrity constraint, cryptographic operation constraint, and Merkle path constraint based on the proof key and the proof secret information, and combining the sub-proofs to generate the original proof; Calculating a proof complexity based on the circuit size and witness data dimensions of the original proof, determining a proof segment size based on the proof complexity and a preset memory limit, and dividing the original proof into multiple proof segments according to the proof segment size; For each proof segment, the number of proof items and the degree of relevance of the proof items are calculated, the proof segments are restructured to obtain optimized proof segments, and the optimized proof segments are recursively merged according to a tree structure to obtain a merged proof; Calculating the proof feature of the combined proof, and performing directionally compressing the combined proof based on the proof feature to obtain a compressed proof; Obtain static assessment results of circuit scale and witness data dimensions, obtain dynamic assessment results of threat type and historical security events, and perform weighted combination to obtain security level; Dynamically adjust security parameters according to the security level, perform security enhancement on the compressed proof based on the security parameters to obtain a final proof, and deploy the final proof to the verification smart contract.

8. A financial voucher automatic generation and verification system based on intelligent accounting, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is configured to scan and identify a business document, obtain transaction information from the business document, and convert the transaction information into standardized transaction data including transaction amount, transaction time, counterparty information, transaction type, product information, and service information; The second unit is configured to extract semantic features from the standardized transaction data using natural language processing technology, perform similarity matching between the extracted semantic features and a preset accounting subject feature library, and generate initial accounting entries; construct a transaction relationship network, identify transaction anomalies through graph structure analysis, optimize the initial accounting entries, and generate final accounting entries; The third unit is configured to generate a financial voucher for the final accounting entry, analyze the subject logic relationship, amount matching relationship, and voucher integrity of the financial voucher through a deep neural network, analyze the transaction path involved in the financial voucher through graph computing, calculate the rule matching degree of the financial voucher based on a preset accounting standard rule library, and determine the validity of the financial voucher; The fourth unit is configured to encrypt the financial voucher that has passed the validity judgment in blocks, store the encrypted data using a Merkle tree structure, deploy a smart contract that supports zero-knowledge proof in a distributed network to verify the authenticity of the financial voucher, and write the stored information of the financial voucher into the blockchain network based on a consensus mechanism; The fifth unit is used to write the financial voucher that has completed the evidence storage into the enterprise financial accounting system, and write the verification record, evidence storage record, and accounting record of the financial voucher into the blockchain network.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images