A multi-dimensional cross-checking-based intelligent auditing and early warning method for finance and tax vouchers

Abstract

This invention relates to the field of accounting technology, specifically to an intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation. The method includes the following steps: extracting multi-dimensional elements from vouchers and forming a total chain by interleaving them in a time sequence; constructing a sparse vector through frequency statistics using sliding window slicing; identifying high-risk vouchers by comparing them with compliance benchmarks; tracing related documents and subject status; asynchronously aligning multi-source sequences; handling timeout gaps and status discontinuities; dynamic programming scoring; and issuing early warnings when vouchers fall below a threshold. In this invention, core voucher elements are converted into feature metacharacters and interleaved in a time sequence to form a total voucher feature chain. Through variable sliding window slicing and frequency statistics, fragment combinations and distribution shifts are characterized, improving the accuracy of hidden anomaly identification. Asynchronous time-series alignment is performed on related documents and subject status of high-risk vouchers, and discontinuities and abrupt changes are evaluated using elastic alignment and affine space penalty scoring. This ensures that the early warning basis covers multi-source continuous relationships, enhancing the completeness of traceability and the interpretability of risk warnings.

Description

Technical Field

[0001] This invention relates to the field of accounting technology, and in particular to an intelligent audit and early warning method for financial and tax documents based on multi-dimensional cross-validation. Background Technology

[0002] Tax and financial document auditing typically encompasses the verification of the authenticity and compliance of tax and financial documents, the accuracy of accounting entries, and the identification and alerting of potential tax risks. Overall, it involves constructing standardized business rules and data verification logic to collect and classify all original documents generated in the daily operations of an enterprise, such as invoices, receipts, contracts, and bank statements. This allows for a systematic review of the logical relationships and compliance status of the document data. Traditional tax and financial document auditing methods, in financial and tax management activities, address technical issues such as low efficiency in document auditing and the risk of financial fraud by using manual verification or single-rule matching for prevention and correction. Traditional methods often involve financial personnel comparing each document with textual information such as the invoice header, tax number, amount in the VAT invoice management system, and payment terms in paper contracts, or using the reimbursement limit rules and fixed account correspondences built into the financial software to verify the legality and reasonableness of the documents.

[0003] In practice, existing technologies primarily rely on manual verification of each voucher and matching based on fixed rules. The focus is typically on the fields on each individual voucher, the reimbursement amount, the correspondence between accounts, and the consistency of textual information. There is a lack of a unified perspective on the temporal connections between different vouchers, the continuity of the transaction chain, and changes in the status of the entities involved. Because manual verification depends on the experience, business familiarity, and workload of financial personnel, it is prone to misinterpreting outwardly compliant documents as routine vouchers when there is a concentrated increase in the number of vouchers, frequent changes in business entities, or discrepancies between contractual agreements and cash flow cycles. It is difficult to detect abnormal risks hidden in the combination of multiple vouchers in a timely manner. Although fixed rule matching can intercept explicit problems such as excessive reimbursement, incorrect accounts, and inconsistent fields, its judgment boundary is relatively fixed. When faced with complex business scenarios such as split accounting, cross-period submission, continuous reimbursement for similar purposes, and short-term abnormal status of related entities, it can often only make judgments based on single conditions. It is difficult to form a comprehensive risk identification result for the entire transaction process, which leads to delayed risk identification, uneven allocation of review resources, difficulty in timely locking of abnormal transaction chains, and increased costs for subsequent tax adjustments, accounting corrections, and compliance rectification. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides an intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation, comprising the following steps: S1: Extract the multidimensional core elements of the voucher to be reviewed and convert them into corresponding feature metacharacters. Then, perform interleaved serialization and splicing according to the temporal polymorphism form to generate the voucher feature chain. S2: Scan the total chain of voucher features by using a variable sliding window slice to obtain a set of sub-character fragments, and construct a sparse feature vector based on the frequency of occurrence of each sub-character fragment; S3: Set the compliance voucher baseline vector and security threshold, calculate the distribution deviation between the sparse feature vector and the compliance voucher baseline vector, and if the distribution deviation exceeds the security threshold, the voucher to be reviewed is determined to be a high-risk suspected voucher, triggering a full-chain traceability instruction; S4: Responding to the full-link tracing command, capture the transaction documents and main status associated with high-risk suspected vouchers and convert them into external association sequences. Introduce an asynchronous buffer with a time delay water level mechanism. In the asynchronous buffer, align the voucher feature chain with each external association sequence according to the time axis to obtain a multi-source observation matrix. When the corresponding external association sequence is not reached after timeout, leave blanks at the corresponding time sequence positions in the multi-source observation matrix. Use dynamic programming algorithm to flexibly align the matrix. If a gap is caused by blanks or sudden state changes, insert empty positions and perform affine empty position penalty. Combine the comparison of feature metacharacters to calculate the overall alignment score. S5: Set a compliance baseline threshold. If the overall alignment score is lower than the compliance baseline threshold, an audit warning instruction will be output.

[0005] As a further aspect of the present invention, the multidimensional core elements are divided into business, financial, and behavioral dimensions, and then converted into feature metacharacters of the corresponding dimensions.

[0006] As a further aspect of the present invention, the temporal multicell is formed by binding multiple feature metacharacters of different dimensions under the same business detail item and arranged according to the business occurrence sequence; When there is a lack of clear business occurrence sequence, the order of details in the voucher to be reviewed is used as a substitute sequence marker. Then, the feature metacharacters in each time sequence polyp are interleaved and spliced ​​according to different dimensions to generate the voucher feature chain.

[0007] As a further embodiment of the present invention, the variable sliding window includes at least two sliding windows of different preset lengths. Each sliding window performs a slicing scan of the total chain of voucher features with a feature metacharacter as the sliding step size, and extracts sub-character fragments at different scales respectively.

[0008] As a further aspect of the present invention, in step S2, when constructing a sparse feature vector based on the set of sub-character segments, a high-dimensional feature space basis mapping relationship is pre-set so that different sub-character segments correspond to different vector dimensions, and the occurrence frequency of each sub-character segment is written into the corresponding vector dimension. When a sub-character fragment contains an unknown feature metacharacter, the sub-character fragment is mapped to a reserved generalized smoothing dimension.

[0009] As a further aspect of the present invention, the compliance certificate baseline vector is generated from historical compliance certificate samples using the same processing method as S1 and S2, and maintains the same dimension as the sparse feature vector. The security threshold is preset based on the distribution deviation range of historical compliance certificate samples.

[0010] As a further aspect of the present invention, in step S3, when calculating the distribution deviation, the sparse feature vector and the compliant voucher reference vector are first normalized, and then the arithmetic mean vector of the two is calculated. The first relative entropy of the normalized sparse feature vector relative to the arithmetic mean vector and the second relative entropy of the normalized compliant voucher reference vector relative to the arithmetic mean vector are calculated respectively. The first relative entropy and the second relative entropy are summed with equal weights to obtain the distribution deviation.

[0011] As a further aspect of the present invention, the asynchronous buffer with a time delay level mechanism determines the time level line based on the event timestamps of the external associated sequence and sets the maximum tolerable time delay. When the time waterline reaches the preset cutoff condition corresponding to the current high-risk suspected evidence or the waiting time exceeds the maximum tolerable delay, the waiting for unreached external related sequences will stop, and blank spaces will be maintained at the corresponding time sequence points in the multi-source observation matrix.

[0012] As a further aspect of the present invention, the dynamic programming algorithm transforms the faults caused by blanking or abrupt state changes into empty spaces when flexibly aligning the multi-source observation matrix. If the vacancy is the initial vacancy in a continuous fault, then the vacancy opening penalty is applied; If the empty space is a continuation of an existing empty space, then an empty space extension penalty is applied, and the alignment comprehensive score is calculated by summing the matching score, mismatch penalty, empty space opening penalty, and empty space extension penalty.

[0013] As a further aspect of the present invention, the audit warning instruction is also used to trigger distributed circuit breaker control. When the alignment comprehensive score is lower than the compliance baseline threshold, an abnormal interruption control signal is generated based on the audit warning instruction, and the abnormal interruption control signal is sent to the business flow control link to suspend the approval flow interface, payment flow interface or database write interface related to the currently judged high-risk suspected voucher.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the core elements of the voucher are converted into feature metacharacters and spliced ​​together in a time-sequential manner to form a voucher feature chain. Through variable sliding window slicing and frequency statistics, the combination and distribution offset of fragments are characterized, improving the accuracy of hidden anomaly identification. Asynchronous time-series alignment is performed on high-risk voucher-related documents and main status, and elastic alignment and affine gap penalty are used to evaluate faults and abrupt changes, so that the early warning basis covers multi-source continuous relationships, enhancing the traceability integrity and risk warning interpretability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] Please see Figure 1 This invention provides an intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation, comprising the following steps: S1: Extract the multidimensional core elements of the voucher to be reviewed and convert them into corresponding feature metacharacters. Then, perform interleaved serialization and splicing according to the temporal polymorphism form to generate the voucher feature chain. When physical paper-based tax vouchers are scanned at high speed and concurrently by the workstation's linear array image sensor, high-resolution image bitmap data is captured in real time. Common image preprocessing algorithms are executed, including adaptive binarization, noise reduction filtering, and image position tilt correction, to ensure that the text lines on the invoice are basically parallel to the horizontal baseline and meet the requirements for subsequent character recognition. Subsequently, the character recognition engine is called to perform a full-area scan of the voucher image, extracting fields such as invoice code, invoice number, invoice date, and invoice amount. At the same time, the multi-dimensional core elements of the voucher to be reviewed are extracted. The elements are divided into three dimensions: business, financial, and behavioral. The business elements include industry standard classification codes, commodity tax classification codes, and metacharacters formed by standardizing commodity names. The financial elements include the applicable tax rate, total amount excluding tax, value-added tax, and the numerical series of reimbursement amount after logarithmic series transformation. The behavioral elements include the internal job level label of the person seeking reimbursement, the geographical location grid code of the consumption location, and the absolute value of the time difference between invoicing and reimbursement application submission. By using discretized encoding mapping rules and a fixed-length feature metacharacter mapping table, the extracted business, financial, and behavioral elements are transformed into corresponding fixed-length feature metacharacters, such as double-byte fixed-length feature metacharacters. Specifically, the industry code in the business element is transformed into business feature metacharacters after being retrieved from the mapping table, and the commodity tax classification code is transformed into another set of business feature metacharacters after being retrieved from the mapping table. In the processing of financial elements, tax rates are directly classified, and different tax rates are directly mapped to corresponding financial feature metacharacters. The amount value is transformed into a logarithmic series, taking the logarithm to the base 10 and rounding down. This logarithmic series is used to represent the scale and order of magnitude of funds and mapped to financial feature metacharacters. In the processing of behavioral elements, the job level of the reimbursement applicant is retrieved and mapped to the corresponding behavioral feature metacharacters, and the consumption geographical location is transformed into an encoding through a gridding algorithm and mapped to the corresponding behavioral feature metacharacters. This makes the voucher elements standardized into a set of fixed-length feature metacharacters. After the transformation of feature metacharacters is completed, the interleaved serialization concatenation logic is initiated. It uses a temporal multi-cell pair format for encapsulation, binding business character groups, financial character groups, and behavioral character groups under the same business detail item to form a tightly coupled feature multi-cell set. Subsequently, these are prioritized according to the absolute temporal order of occurrence in physical transactions. When a single voucher lacks a clear business occurrence sequence, the top-to-bottom spatial arrangement of the voucher detail lines is used as a substitute sequence marker. These metacharacters are then connected end-to-end according to an interleaved topology structure of business, finance, behavior, business, finance, and behavior. The resulting long string is the voucher feature chain. Due to the use of interleaved serialization concatenation in the form of temporal multi-cell pairs, elements in different dimensions within the voucher feature chain are physically adjacent within local areas of the string. This ensures that subsequent sliding windows can simultaneously capture cross-dimensional association patterns at any slice position, thus enabling the identification of abnormal business combinations through local string arrangement features in the first stage, generating the voucher feature chain.

[0021] S2: Scan the total chain of voucher features by using a variable sliding window slice to obtain a set of sub-character fragments, and construct a sparse feature vector based on the frequency of occurrence of each sub-character fragment; Based on the voucher feature chain, a variable sliding window with a multi-scale preset length and a sliding step of one feature metacharacter is configured. In the stream processing, three sets of parallel windows with feature metacharacter lengths of three, five, and seven feature metacharacters are run simultaneously. The short window with three feature metacharacters is responsible for extracting micro-combination features between internal or adjacent local fields of temporal polymorphic pairs, such as whether the combination of a certain item category and its applicable tax rate conforms to normal financial logic. The long windows with five and seven feature metacharacters cross the physical boundaries of temporal polymorphic pairs and are used to capture medium- and long-range composite anomalies across item details and time periods, such as the combination features of continuous geographical location mutations and abnormal fluctuations in the amount series. The step size of the sliding window is strictly limited to one feature metacharacter, ensuring that the window moves to the right by one feature metacharacter encoding unit each time, thereby achieving high overlap and no dead-angle slicing scan of the voucher feature chain. During the sliding process of the variable sliding window, continuous feature metacharacters within its coverage area are extracted at the current pointer position. The substring is defined as a subcharacter fragment. During the sliding process, subcharacter fragments are obtained at scales of three feature metacharacters, five feature metacharacters, and seven feature metacharacters, and then merged into a set to obtain the subcharacter fragment set of the voucher. Based on the set of sub-character fragments, a high-dimensional feature space basis mapping table is set up. Each dimension in the space uniquely corresponds to a specific scale of sub-character fragment combination pattern. The set of sub-character fragments is traversed, and the absolute frequency values ​​of different sub-character fragments appearing in the set are counted. Then, a high-dimensional numerical vector with the same size as the total dimension of the feature space is dynamically constructed. The index number of the high-dimensional numerical vector corresponds to each dimension of the feature space. The value of the high-dimensional numerical vector at a specific index position is determined by the frequency of occurrence of the sub-character fragment. The sparse feature vector of the voucher to be reviewed is constructed in this way. Since the character length of the total feature chain of a single voucher is limited, the number of unique sub-character fragments scanned is much smaller than the total dimension of the feature space. Therefore, the values ​​at a very small number of positions in the sparse feature vector are greater than zero, and the rest are all zero. In computer memory, a sparse storage structure is used for high-density and lightweight storage and representation. In actual production environments, due to the emergence of new businesses or occasional recognition errors in the character recognition engine, the total feature chain of vouchers may contain unknown feature metacharacters that have not appeared in the historical compliance sample library. Directly ignoring unknown characters can lead to vector dimension conflicts or abnormal deviations in calculation accuracy. To address this, a generalized smoothing dimension is reserved at the end of the high-dimensional sparse feature vector space. When the sliding window scans a sub-character fragment containing an unknown metacharacter, it automatically stops performing regular spatial dimension indexing and mapping, and uniformly redirects it to this generalized smoothing dimension. At the same time, a smoothing factor is introduced to correct the frequency only for the frequency statistics of this generalized smoothing dimension. A small baseline weight determined by the smoothing factor is assigned to the scanned sub-character fragment containing the unknown feature metacharacter. Through the generalized smoothing dimension processing mechanism, spatial vector dimension truncation caused by unknown inputs is avoided, ensuring computational robustness in a high-concurrency continuous transaction processing environment.

[0022] S3: Set the compliance voucher baseline vector and security threshold, calculate the distribution deviation between the sparse feature vector and the compliance voucher baseline vector, and if the distribution deviation exceeds the security threshold, the voucher to be reviewed is determined to be a high-risk suspected voucher, triggering a full-chain traceability instruction; Based on a large number of real voucher samples that have passed all historical audits, extracted from the financial shared service center, and through offline statistics, clustering or machine learning, combined with element extraction, staggered serialization and multi-scale sliding window vectorization processes, the mathematical expectation distribution center of the high-dimensional sparse feature vectors of a large number of real voucher samples is obtained, and the baseline vector of compliant vouchers is obtained, whose dimension is consistent with the sparse feature vector of the vouchers to be audited. After constructing the sparse feature vector of the voucher to be reviewed, the existence of distribution distortion in the high-dimensional feature space of the voucher is evaluated. A distribution deviation calculation process is introduced. In order to eliminate the influence of invoice details on the absolute frequency value, normalization scaling is performed on the current sparse feature vector and the compliant voucher baseline vector, so that the sum of the component values ​​inside the two vectors is strictly equal to one, thereby transforming them into a standard probability distribution vector. The feature comparison task is transformed into a probability distribution similarity measurement task. In order to quantify the distribution deviation, this embodiment adopts a divergence calculation architecture with symmetric properties to define the distribution deviation. The specific calculation process is as follows: calculate the arithmetic mean vector of the probability distribution vector of the voucher to be reviewed and the compliant voucher baseline vector. Then, using the sparse acceleration algorithm, traverse the non-zero dimension union of the probability distribution vector of the voucher to be reviewed and the compliant voucher baseline vector, respectively calculate the first relative entropy of the probability distribution of the voucher to be reviewed relative to the average vector, and the second relative entropy of the compliant voucher baseline vector relative to the average vector. Finally, the calculated first relative entropy and second relative entropy are summed with equal weight to obtain the final value representing the difference in distribution between the two in the high-dimensional geometric space. This value is defined as the distribution deviation. The calculated distribution deviation value is compared with the pre-configured safety threshold in real time. The safety threshold is a deterministic floating-point number set based on the statistical confidence interval of historical risk control boundaries. Deterministic diversion control logic is executed according to the comparison result: if the distribution deviation value of the current voucher does not exceed the safety threshold, the combination pattern of the voucher's business, financial and behavioral elements is highly consistent with the historical compliance benchmark, a spatial distribution compliance mark is generated, and a release settlement instruction for the regular financial accounting system is generated. If the distribution deviation value of the current voucher exceeds the safety threshold, the voucher shows obvious geometric distribution anomalies in the high-dimensional feature space, which means that it has potential risks of false reimbursement, cross-category packaging or fictitious business. Its one-click release branch is immediately stopped, and the voucher to be reviewed is judged as a high-risk suspected voucher. Once a document is identified as high-risk, a full-chain tracing instruction is immediately triggered. The full-chain tracing instruction is a structured message payload containing a high-priority execution marker, and it is mandatory to include the unique system serial number of the current high-risk document, the invoice date and timestamp, the unique employee ID of the reimbursement applicant, and the unified social credit code of the invoicing supplier.

[0023] S4: Responding to the full-link tracing command, capture the transaction documents and main status associated with high-risk suspected vouchers and convert them into external association sequences. Introduce an asynchronous buffer with a time delay water level mechanism. In the asynchronous buffer, align the voucher feature chain with each external association sequence according to the time axis to obtain a multi-source observation matrix. When the corresponding external association sequence is not reached after timeout, leave blanks at the corresponding time sequence positions in the multi-source observation matrix. Use dynamic programming algorithm to flexibly align the matrix. If a gap is caused by blanks or sudden state changes, insert empty positions and perform affine empty position penalty. Combine the comparison of feature metacharacters to calculate the overall alignment score. Based on the end-to-end traceability instruction, and using the supplier's unified social credit code in the instruction as an index, the system retrieves contract text data signed with the supplier from the enterprise's legal contract database. Simultaneously, it retrieves bank statement data corresponding to the current reimbursement amount level, the reimbursing person, or the supplier from the bank-enterprise direct connection system's cash flow database. Furthermore, it retrieves corresponding reimbursement application records, approval records, or historical reimbursement behavior records using the applicant's unique employee ID. In parallel, it retrieves the supplier's business registration change logs for the specified time period before and after the invoice date from the credit information disclosure data interface or a pre-connected entity status database. Because these external data sources are distributed across different network nodes, and each system... The data updates and transmissions are time-varying, resulting in highly asynchronous and time-delayed data streams. The captured contract text data, bank statement data, and business registration change logs are transformed into corresponding external association sequences using time-series multi-cell mapping and interleaved serialization methods. Specifically, the contract subject category, contract amount level, signing date, and supplier entity identifier in the contract text data are transformed into contract feature sequences; the payment amount level, payment time, payee account, and payment status in the bank statement data are transformed into transaction feature sequences; and the status of existence, abnormal operation, cancellation, revocation, change of business scope, and effective time in the business registration change logs are transformed into log feature sequences. Based on external associated sequences, an asynchronous buffer is opened, and a flow control protocol with a time delay level mechanism is introduced into the asynchronous buffer. The time delay level mechanism measures the progress of global time by monitoring the event timestamps of the external associated sequences and sets a maximum tolerable time delay parameter. As various external associated sequences continuously flow into the asynchronous buffer, the maintained event time water level line increases monotonically. When the event time water level line crosses the target deadline timestamp corresponding to the current high-risk suspected credential, or the physical timer of the buffer exceeds the waiting limit of the maximum tolerable time delay parameter, the time delay level mechanism sends a forced cutoff output control signal to the asynchronous buffer. Upon receiving the forced cutoff output control signal from the time-delay water level mechanism, within the asynchronous buffer, the voucher feature chain of the currently pending voucher is used as the baseline axis. All external related sequences, such as the contract feature sequence, transaction feature sequence, and log feature sequence, which are concurrently captured and transformed, are arranged horizontally in a grid on the time axis according to their respective absolute timestamp information, constructing a multi-row, multi-column two-dimensional grid data structure. This structure is defined as a multi-source observation matrix. At the moment of truncation in the construction of the multi-source observation matrix, a timeout blanking strategy is executed. Whether it is the external related sequence corresponding to the transaction document or the external related sequence corresponding to the main body status, once the time-delay water level mechanism cuts off or the timer times out, if the corresponding external related sequence fails to fully reach the buffer, the waiting is immediately stopped. Predictive filling of missing data is prohibited during the construction phase of the multi-source observation matrix. Physical blanking is directly forced at the corresponding time sequence position of the multi-source observation matrix, and blanking control characters are injected into the corresponding matrix cells, transforming the physical data time delay into a blanking discontinuity state that can be identified at the algorithm level. After the multi-source observation matrix is ​​constructed, the dynamic programming algorithm is used to perform elastic alignment calculations on multiple sequences of the multi-source observation matrix. The dynamic programming algorithm stretches or compresses the time axis to eliminate the overall time axis misalignment caused by reasonable extension of payment period, cross-month settlement, and late arrival of settlement documents in normal business behavior, so as to achieve elastic alignment between sequences and distinguish between intentional fictitious trade chain break risks and legitimate business delays. Then, a recursive score matrix is ​​constructed. For multiple externally related sequences, a two-dimensional recursive score matrix is ​​constructed between the total chain of voucher features and each externally related sequence. The scores of each two-dimensional recursive score matrix are summarized according to the preset weights, or multiple externally related sequences are projected into a unified multi-source joint sequence before constructing the recursive score matrix. In the recursive alignment process of the dynamic programming algorithm, if character inconsistencies or gaps occur, a nonlinear penalty based on the affine gap penalty model is executed. In this embodiment, the affine gap penalty model sets the gap penalty score as an affine function of the length of consecutive gaps. The gap penalty value is composed of two parts: a gap opening penalty and a gap extension penalty. The gap opening penalty is set as the first penalty weight, and the gap extension penalty is set as the second penalty weight. The value of the first penalty weight is significantly greater than the value of the second penalty weight. In the recursive state transition of dynamic programming, cumulative recursion is performed through set boundary conditions. To distinguish between gaps and main state abrupt changes at the logical computation layer, and to fit the two types of gap mechanisms (gap or state abrupt change), the algorithm performs differentiated judgment: when encountering physical gaps caused by transaction documents not being reached due to timeout, the algorithm inserts a gap identifier in the alignment stream. Here, gaps are the original missing states in the multi-source observation matrix, and gaps are alignment identifiers inserted during the dynamic programming elastic alignment process for the gaps or state abrupt changes. If the gap identifier is not immediately adjacent to the existing... Once a gap is found, it is determined that a new gap has been opened and a gap opening penalty is triggered, deducting a heavy score determined by the first penalty weight. When the associated sequence corresponding to the main state has been reached on time, but the main state value changes drastically in the middle of the timeline, such as the supplier's main state changing from existing to cancelled after invoicing, or from normal state to abnormal operation, revocation, or other preset abnormal state, it is determined that a state change has occurred. Similarly, a gap identifier is inserted into the alignment stream and the corresponding affine penalty is triggered. If the gap identifier is a continuous fault immediately following an existing gap, the algorithm determines that it is a continuous extension of the same fault and executes a gap extension penalty, deducting only a slight baseline score determined by the second penalty weight each time to avoid repeatedly applying the gap opening penalty to continuous missing values. At the same time, if the characteristic characters of the two sequences are inconsistent at the same time point, it is determined that there is an information mismatch, and a preset fixed substitution score is deducted. Through the above state transition process, the full cell accumulation and recursion are performed between the various dimensional tracks of the multi-source observation matrix, and finally the alignment comprehensive score of the multi-source sequence elastic alignment is calculated.

[0024] S5: Set a compliance baseline threshold. If the overall alignment score is lower than the compliance baseline threshold, an audit warning instruction will be issued. Retrieve the pre-set and fixed compliance baseline threshold, and compare the alignment comprehensive score with the compliance baseline threshold. If the alignment comprehensive score is not lower than the compliance baseline threshold, it means that the current multi-source link has not met the triggering conditions for the fault warning, and the current full-link fault warning can be lifted, or the voucher can be transferred to the regular review process. If the alignment comprehensive score is lower than the compliance baseline threshold, it means that the score has been significantly reduced due to the high penalty for opening a gap, and it is determined that there is a substantial risk of fraud due to the flow of time and space faults in the current business, and an audit warning instruction is generated. If the score is determined to be below the compliance threshold, the reverse backtracking path optimization mechanism is immediately activated. The mechanism searches backward along the optimal path of the dynamic programming score matrix to locate the matrix space coordinates and absolute time axis coordinates of the empty position with the most severe deduction. The specific characteristic character connotation corresponding to the coordinates is then parsed in reverse to locate the specific fraud risk break point. The data parsed in reverse is automatically packaged and an audit warning instruction with the break point mark is output. In one optional implementation, the audit warning instruction is also used to execute multi-level cascaded distributed circuit breaker control. Based on the audit warning instruction, a control status word with the highest interrupt priority is generated and cascaded to the business flow kernel. By injecting abnormal interrupt control signaling into the kernel routing queue, software-level circuit breaker control is implemented at the business protocol layer and interface layer. This suspends and locks the application programming interface access permissions and database write permissions of the current suspected credential during the flow process, thereby blocking the approval process.

[0025] In a further optional implementation, the audit warning instruction is also used to link with the security access control module to restrict the access permissions of the payment interface, fund payment channel, and sensitive data transmission channel related to the current voucher. This extends the interface locking at the software layer to the path isolation at the network access layer, thereby restricting the outflow of potentially abnormal funds.

[0026] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent auditing and early warning of financial and tax vouchers based on multi-dimensional cross-validation, characterized in that, Includes the following steps: S1: Extract the multidimensional core elements of the voucher to be reviewed and convert them into corresponding feature metacharacters. Then, perform interleaved serialization and splicing according to the temporal polymorphism form to generate the voucher feature chain. S2: Scan the total chain of voucher features by using a variable sliding window slice to obtain a set of sub-character fragments, and construct a sparse feature vector based on the frequency of occurrence of each sub-character fragment; S3: Set the compliance voucher baseline vector and security threshold, calculate the distribution deviation between the sparse feature vector and the compliance voucher baseline vector, and if the distribution deviation exceeds the security threshold, the voucher to be reviewed is determined to be a high-risk suspected voucher, triggering a full-chain traceability instruction; S4: Responding to the full-link tracing command, capture the transaction documents and main status associated with high-risk suspected vouchers and convert them into external association sequences. Introduce an asynchronous buffer with a time delay water level mechanism. In the asynchronous buffer, align the voucher feature chain with each external association sequence according to the time axis to obtain a multi-source observation matrix. When the corresponding external association sequence is not reached after timeout, leave blanks at the corresponding time sequence positions in the multi-source observation matrix. Use dynamic programming algorithm to flexibly align the matrix. If a gap is caused by blanks or sudden state changes, insert empty positions and perform affine empty position penalty. Combine the comparison of feature metacharacters to calculate the overall alignment score. S5: Set a compliance baseline threshold. If the overall alignment score is lower than the compliance baseline threshold, an audit warning instruction will be output.

2. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, The multidimensional core elements are divided into business, financial, and behavioral dimensions, and then transformed into feature metacharacters for the corresponding dimensions.

3. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, The temporal multi-cell pair is formed by binding multiple feature metacharacters of different dimensions under the same business detail item and arranged according to the time sequence of business occurrence; When there is a lack of clear business occurrence sequence, the order of details in the voucher to be reviewed is used as a substitute sequence marker. Then, the feature metacharacters in each time sequence polyp are interleaved and spliced ​​according to different dimensions to generate the voucher feature chain.

4. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, The variable sliding window includes at least two sliding windows of different preset lengths. Each sliding window uses a feature metacharacter as the sliding step size to slice and scan the total chain of voucher features, and extracts sub-character fragments at different scales.

5. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, In S2, when constructing sparse feature vectors based on the set of sub-character fragments, a high-dimensional feature space basis mapping relationship is pre-set so that different sub-character fragments correspond to different vector dimensions, and the occurrence frequency of each sub-character fragment is written into the corresponding vector dimension. When a sub-character fragment contains an unknown feature metacharacter, the sub-character fragment is mapped to a reserved generalized smoothing dimension.

6. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, The compliance certificate baseline vector is generated from historical compliance certificate samples using the same processing method as S1 and S2, and maintains the same dimension as the sparse feature vector. The security threshold is preset based on the distribution deviation range of historical compliance certificate samples.

7. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, In step S3, when calculating the distribution deviation, the sparse feature vector and the compliance certificate reference vector are first normalized, and then the arithmetic mean vector of the two is calculated. The first relative entropy of the normalized sparse feature vector relative to the arithmetic mean vector and the second relative entropy of the normalized compliance certificate reference vector relative to the arithmetic mean vector are calculated respectively. The first relative entropy and the second relative entropy are summed with equal weights to obtain the distribution deviation.

8. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, The asynchronous buffer with a time delay level mechanism determines the time level based on the event timestamps of the external associated sequence and sets the maximum tolerable time delay. When the time waterline reaches the preset cutoff condition corresponding to the current high-risk suspected evidence or the waiting time exceeds the maximum tolerable delay, the waiting for unreached external related sequences will stop, and blank spaces will be maintained at the corresponding time sequence points in the multi-source observation matrix.

9. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation as described in claim 1, characterized in that, When performing elastic alignment of the multi-source observation matrix, the dynamic programming algorithm transforms the faults caused by blanking or abrupt state changes into empty slots. If the vacancy is the initial vacancy in a continuous fault, then the vacancy opening penalty is applied; If the empty space is a continuation of an existing empty space, then an empty space extension penalty is applied, and the alignment comprehensive score is calculated by summing the matching score, mismatch penalty, empty space opening penalty, and empty space extension penalty.

10. The intelligent audit and early warning method for financial and tax vouchers based on multi-dimensional cross-validation according to claim 1, characterized in that, The audit warning instruction is also used to trigger distributed circuit breaker control. When the overall alignment score is lower than the compliance baseline threshold, an abnormal interruption control signal is generated based on the audit warning instruction, and the abnormal interruption control signal is sent to the business flow control link to suspend the approval flow interface, payment flow interface or database write interface related to the currently judged high-risk suspected voucher.

Images