An artificial intelligence-based unit financial data intelligent analysis system

Abstract

This invention relates to the field of data processing and information auditing technology, and discloses an intelligent analysis system for unit financial data based on artificial intelligence. The system extracts logs from the source system through a log collection and time series calibration module and reconstructs the time series based on foreign key dependencies of documents; a cardinality mapping module solves matching combinations within a defined matching interval based on capital conservation constraints, outputting a sequence of composite state macro nodes; a causal graph construction module generates a baseline directed acyclic graph structure and a state transition probability matrix based on historical sequences and time prior constraints; a weighted entropy calculation module combines a normalized financial amount scalar with a probability matrix to output a value-weighted causal information entropy; and a judgment and diagnosis output module generates an interrupt control command when the information entropy meets the triggering condition and performs differential operations to output diagnostic records. This invention corrects cross-system time series deviations, solves the problem of complex financial data mapping, reduces the false alarm rate of anomaly interception, and achieves accurate location of anomaly nodes.

Description

Technical Field

[0001] This invention relates to the field of data processing and information auditing technology, specifically to an intelligent analysis system for corporate financial data based on artificial intelligence. Background Technology

[0002] As informatization deepens, internal business processes and financial accounting within organizations are often completed collaboratively across multiple independent subsystems. When integrating and analyzing this cross-system data, discrepancies in the physical clocks of each system, coupled with network transmission delays, often result in inverted timestamps in the aggregated source log data. Directly relying on these absolute physical times to construct data sequences would distort the true chronological order of business processes, leading to subsequent logical analysis being built on an incorrect data foundation.

[0003] After data extraction, the system needs to match discrete business events with financial events. Existing data mapping mechanisms mostly rely on rigid checks for absolute equality based on static attributes. When dealing with complex many-to-many reconciliation scenarios involving total amount tolerances, they are prone to matching failures due to slight amount discrepancies. Furthermore, performing combined searches on the entire dataset without time or space limitations consumes excessive system memory and computing resources.

[0004] In the abnormal business detection and interception phase, conventional systems typically trigger blocking commands based on pre-coded fixed amount thresholds or single-dimensional frequency rules. This approach severs the connection between fund changes and the underlying business probability distribution, making it difficult to adapt to dynamically changing real-world business scenarios and resulting in a high false alarm rate. Furthermore, when an anomaly is triggered, the system usually only outputs isolated alarm signals, lacking a means of comparison with the normal business baseline topology map. This prevents business personnel from directly identifying specific missing or redundant upstream logic nodes in the flow path, increasing the difficulty of subsequent manual investigation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent analysis system for unit financial data based on artificial intelligence. This system solves the problems of existing financial analysis systems, such as disordered time sequence reconstruction and difficulty in complex tolerance matching when processing cross-system data, as well as the high false alarm rate of the anomaly interception mechanism and the difficulty in accurately locating missing nodes in the flow path.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides an intelligent analysis system for corporate financial data based on artificial intelligence, comprising:

[0008] The log collection and time series calibration module is configured to extract source log data from the source system to generate discrete event tuples and reconstruct the time series based on the foreign key dependencies of the documents.

[0009] The cardinality mapping module is configured to solve for matching combinations of the discrete event tuples within the matching interval defined in the spatiotemporal dimension, based on the capital conservation constraint, and fuse and output a composite state macro node sequence.

[0010] The causal graph construction module is configured to generate a baseline directed acyclic graph structure and a state transition probability matrix based on the historical compliant composite state macro node sequence and time prior constraints.

[0011] The weighted entropy calculation module is configured to perform nonlinear normalization on the financial amount scalar of the real-time generated composite state macro node sequence, and output the value-weighted causal information entropy by combining the state transition probability matrix.

[0012] The judgment and diagnosis output module is configured to generate an interrupt control instruction when the value-weighted causal information entropy satisfies the interception trigger condition, and to perform differential operation in conjunction with the baseline directed acyclic graph structure to output a diagnostic record.

[0013] By reconstructing the time sequence through foreign key dependencies in documents, interference caused by the asynchronous physical clocks across systems is corrected; dynamic programming matching is performed using the capital conservation constraint to aggregate scattered single-point events into a whole node containing business and financial relationships; a directed acyclic graph and state transition probability matrix are established by combining causal discovery algorithms and time prior constraints to construct a probabilistic model of business flow paths; value-weighted causal information entropy is introduced as an evaluation indicator, and flow probability information is fused with financial scalars to form a numerical benchmark for measuring the degree of anomalies in business links, and differential comparison is used to accurately output diagnostic differences.

[0014] Furthermore, the log collection and time-series calibration module extracts absolute physical timestamps and foreign key dependencies from the source log data. Based on the foreign key dependencies, it compares the absolute physical timestamps of the upstream and downstream nodes corresponding to the discrete event tuples. If the absolute physical timestamp of the downstream node is detected to be earlier than that of the upstream node, a time-series inversion conflict is determined to be triggered. A local directed dependency graph is constructed for the data that triggered the time-series inversion conflict, the reverse association edges are disconnected, and a topological sorting operation is performed to generate incremental logical time-series variables for each event tuple, thereby reconstructing the time series. This mechanism eliminates the node order inversion problem caused by network data transmission delays by resetting the variable time sequence.

[0015] Furthermore, the cardinality mapping module extracts the business entity identifiers of the discrete event tuples, calculates corresponding hash keys using a hash function, and stores the discrete event tuples in independent entity space buckets based on the hash keys. A sliding time window is configured within each entity space bucket. Using the target financial event as a benchmark, a subset of business events and a subset of financial events are extracted within the sliding time window to generate a candidate set of business events and a local subset of financial events, constituting the matching interval. This operation adds spatial and temporal boundary constraints at the data retrieval level, thereby narrowing the search interval for subsequent matching and reducing system computational resource overhead.

[0016] Furthermore, the cardinality mapping module sets an amount tolerance threshold, using the amount of financial events within the local financial event subset as the knapsack capacity and the amount of business events in the candidate business event set as the item weight, constructing a one-dimensional state transition array; using a dynamic programming algorithm model, it searches for candidate combinations of total amount differences within the amount tolerance threshold range in the matching interval; if multiple candidate combinations are detected, the one with the smallest event timestamp span variance is selected as the optimal matching solution, and the composite state macro node sequence is fused and output. This processing flow enables the system to converge to a unique matching result under the condition of time clustering degree, even in the presence of amount recording deviations.

[0017] Furthermore, the causal graph construction module introduces a PC causal discovery algorithm to construct an initial fully connected undirected graph for the composite state macro node sequence. By comparing the order of two macro nodes with a connection in the time series, if the first macro node is determined to be later than the second macro node in time, a directed edge constraint is applied at the algorithm's underlying layer, forcibly blocking causal edges from the first macro node to the second macro node, while retaining directed connections from the second macro node to the first macro node, thus generating the baseline directed acyclic graph structure. The system thereby utilizes the irreversibility of time to constrain the direction of the network structure, avoiding the generation of reverse flow logic edges due to coincidental data distribution.

[0018] Furthermore, the causal graph construction module sets a tracking window with a step size parameter greater than 1, and statistically analyzes the transition frequency between cross-step nodes based on the historical compliant composite state macro-node sequence. Based on the transition frequency, a Laplace smoothing mechanism is introduced to calculate the conditional probability of the previous node transitioning to the current node, generating the state transition probability matrix with directional orientation. The smoothing mechanism is mainly used to handle sparse paths not observed in the samples and to prevent zero-value truncation during the conditional probability calculation.

[0019] Furthermore, the weighted entropy calculation module extracts the financial amount scalar attached to the composite state macro node sequence and takes its absolute value; it then uses a logarithmic nonlinear function containing a preset numerical smoothing constant to perform a mapping calculation on the financial amount scalar after taking the absolute value, converting it into a financial value coefficient in the range of 0 to 1. This mapping process can converge financial data of different magnitudes to a unified calculation range.

[0020] Furthermore, the weighted entropy calculation module retrieves the sequence of preceding nodes of the current macro node from a pre-configured graph database, extracts the corresponding conditional probability values ​​from the state transition probability matrix, calculates the self-information of the step jump, performs a multiplicative fusion operation between the financial value coefficient and the self-information, and accumulates the weighted information of the effective preceding paths to output the value-weighted causal information entropy. By fusing the probability of event occurrence and the monetary weight, the system transforms complex business flow characteristics into quantifiable scalar indicators for easy comparison.

[0021] Furthermore, the judgment and diagnosis output module performs statistical analysis on the value-weighted causal information entropy generated during historical compliance periods based on business type attributes, and dynamically calculates a baseline entropy threshold by combining the mean and standard deviation. When the received real-time value-weighted causal information entropy is greater than the baseline entropy threshold, it determines that the interception trigger condition is met, generates the interruption control command, and sends it to the enterprise service bus to freeze fund allocation. This operation automatically sets a dynamic interception baseline based on historical distribution status, reducing misjudgments caused by manually hard-coded fixed thresholds.

[0022] Furthermore, the judgment and diagnosis output module extracts the intercepted current flow path, converts each node instance within it into macro node state types to construct a real-time directed edge set; it extracts the baseline expected directed edge set corresponding to the current flow start and end paths in the baseline directed acyclic graph structure; it calculates the actual redundancy difference set and the baseline missing difference set for the real-time directed edge set and the baseline expected directed edge set, respectively, locates unmatched edges based on the difference set, and generates the diagnostic record containing the attribute information of the missing preceding node. Using set difference operations, the specific missing node information that triggered the interception event can be directly output.

[0023] Furthermore, before extracting the absolute physical timestamp and foreign key dependencies, the system extracts the payload content of the source log data and calculates a feature digest; a Bloom filter is constructed in memory, and the feature digest is used to perform an existence check on the arriving source log data; if it is determined that the feature digest already exists in the Bloom filter, the source log data is marked as a retransmission redundancy and discarded. This verification step is added to intercept redundant data generated by the network retry mechanism of the source system, maintaining the uniqueness of time-series data.

[0024] Furthermore, the arrival rate of discrete event tuples within the entity space bucket is periodically calculated to determine the data distribution density. When the data distribution density exceeds a preset congestion threshold, an exponential decay function is used to shorten the span of the sliding time window. When the data distribution density falls below a preset sparsity threshold, a linear amplification algorithm is used to extend the span of the sliding time window. By adjusting the time window span, the computational load of the matching interval can adapt to changes in business concurrency at different times.

[0025] Furthermore, the timestamps corresponding to the circulation frequencies are obtained, and their physical time differences from the current system time are calculated. These physical time differences are then substituted into a preset half-life forgetting function to add a weight decay penalty coefficient less than 1 to the historical circulation frequencies. The state transition probability matrix is ​​updated using the equivalent frequencies after adding the decay penalty coefficient. This forgetting design aims to reduce the influence weight of distant historical data on the current probability matrix, making the baseline model biased towards recently occurring business logic.

[0026] Furthermore, an open feedback interface is provided to receive audit confirmation tags for fund allocation and freezing operations. When the received audit confirmation tag indicates that the interruption control command was falsely intercepted, the deviation gradient between the current value-weighted causal information entropy and the baseline entropy threshold is calculated. Based on the deviation gradient, the calculation weight parameters of the baseline entropy threshold are fine-tuned using a gradient descent algorithm with a reverse penalty. This constructs a threshold correction feedback loop, using external input data to reduce the probability of false alarms in similar business scenarios in the future.

[0027] Furthermore, a baseline network topology map generated based on the baseline directed acyclic graph structure is loaded into the graphical user interface; a color mapping engine is used to perform feature rendering on the diagnostic records, rendering the directed edges corresponding to the actual redundant difference sets as warning colors, and rendering the directed edges corresponding to the baseline missing difference sets as dashed flashing lines; the rendered topology map is merged with the diagnostic records and exported as a financial analysis report. This process converts the underlying difference data into visual charts to intuitively identify abnormal topological structures in the flow path.

[0028] This invention provides an intelligent analysis system for corporate financial data based on artificial intelligence. It has the following beneficial effects:

[0029] 1. This invention extracts foreign key dependencies from documents and compares upstream and downstream physical timestamps to construct a local directed dependency graph for topological sorting in case of inversion conflicts. This mechanism breaks away from the strong dependence on the absolute physical clocks of each business subsystem, corrects the data inversion phenomenon caused by network latency or system clock asynchrony from the logical level, ensures the accurate reconstruction of the execution order of cross-system events, and provides a reliable data foundation for subsequent causal logic determination.

[0030] 2. This invention introduces a dimensionality reduction strategy of entity space binning and sliding time window in the discrete event matching stage, and constructs a state transition array based on capital conservation constraints for dynamic programming solution. This approach can effectively reduce the retrieval space in massive data environments to reduce memory overhead. Furthermore, given reasonable monetary deviations, it utilizes time clustering features to select the optimal matching combination, thus solving the mapping and fusion problem of many-to-many financial accounting in complex business processes.

[0031] 3. This invention calculates the value-weighted causal information entropy by fusing a constructed state transition probability matrix with a normalized financial amount scalar, and dynamically generates a baseline threshold based on historical distribution. This evaluation method, which combines circulation probability and capital weight, replaces the traditional fixed interception rules, reducing the false alarm rate of anomaly detection. Simultaneously, by utilizing the set difference operation of the real-time path and baseline directed acyclic graph, the system can directly output the specific missing edges and preceding node information that triggered the interception, providing a clear basis for locating anomaly business investigations. Attached Figure Description

[0032] Figure 1 This is a system framework diagram of the present invention;

[0033] Figure 2 This is a flowchart of the method of the present invention;

[0034] Figure 3 This is a schematic diagram of the data processing principle of the log acquisition and timing calibration module of the present invention;

[0035] Figure 4 This is a schematic diagram illustrating the data dimensionality reduction and matching principle of the cardinality mapping module of the present invention;

[0036] Figure 5 This is a schematic diagram illustrating the working principle of the cause-effect graph construction module of the present invention;

[0037] Figure 6 This is a schematic diagram of the data calculation principle of the weighted entropy calculation module of the present invention;

[0038] Figure 7 This is a schematic diagram illustrating the operating principle of the determination and diagnosis output module of this invention.

[0039] Figure 8 This is a timing diagram of the anomaly diagnosis of the present invention;

[0040] Figure 9 This is a comparison curve of the ROC performance of the present invention. Detailed Implementation

[0041] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] See attached document Figure 1 , Figure 1 This is a schematic diagram of the architecture of an intelligent analysis system for corporate financial data based on artificial intelligence, according to an embodiment of the present invention. The present invention provides an intelligent analysis system for corporate financial data based on artificial intelligence, comprising:

[0043] The log collection and time series calibration module 10 is configured to extract source log data from the front-end business system and the back-end financial system to generate discrete event tuples, and reconstruct the time series of cross-system data based on document lineage foreign key dependencies.

[0044] The cardinality mapping module 20 is configured to receive discrete event tuples, perform spatiotemporal data partitioning operations to narrow the search matching interval, solve the matching combination according to the capital conservation constraint within the defined matching interval, and merge and package multiple business events with business mapping relationships with financial events to output a composite state macro node sequence.

[0045] The causal graph construction module 30 is configured to obtain the sequence of composite state macro nodes during the historical compliance period, combine the time prior constraints to perform causal discovery algorithm to generate a directed acyclic graph structure, and count the flow frequency between adjacent nodes to generate a state transition probability matrix.

[0046] The weighted entropy calculation module 40 is configured to obtain the composite state macro nodes generated in real time during the current business flow process, extract the attached financial amount scalar and perform nonlinear normalization calculation, combine it with the conditional probability values ​​in the state transition probability matrix for joint operation, and output the value weighted causal information entropy of the current flow path.

[0047] The judgment and diagnosis output module 50 is configured to receive value-weighted causal information entropy, execute feature value comparison logic, generate an interrupt control command when the interception trigger condition is met, and perform differential operation on the real-time flow path and the baseline directed acyclic graph structure to output diagnostic records.

[0048] This AI-based intelligent analysis system for corporate financial data runs on a server cluster environment containing multi-core processors. Its underlying architecture relies on a distributed memory computing engine to handle high-concurrency data streams and deploys a graph database service process to support the storage and retrieval of topology network data within the system.

[0049] See attached documentFigure 2 , Figure 2 This is a flowchart of an intelligent analysis method for unit financial data based on artificial intelligence according to an embodiment of the present invention.

[0050] This invention provides an intelligent analysis method for corporate financial data based on artificial intelligence. This method, applied to the aforementioned intelligent analysis system for corporate financial data based on artificial intelligence, includes the following steps:

[0051] S10: The log collection and time series calibration module 10 extracts source log data from the front-end business system and the back-end financial system to generate discrete event tuples, and reconstructs the time series of cross-system data based on document lineage foreign key dependencies.

[0052] S20 receives discrete event tuples through the cardinality mapping module 20, performs data partitioning operations in the spatiotemporal dimension to narrow the search matching interval, solves the matching combination according to the capital conservation constraint within the defined matching interval, and merges and packages multiple business events with business mapping relationships with financial events to output a composite state macro node sequence.

[0053] S30: The composite state macro node sequence during the historical compliance period is obtained through the causal graph construction module 30. The causal discovery algorithm is executed in combination with the time prior constraints to generate a directed acyclic graph structure, and the flow frequency between adjacent nodes is counted to generate a state transition probability matrix.

[0054] S40, through the weighted entropy calculation module 40, obtains the composite state macro node generated in real time in the current business flow process, extracts the attached financial amount scalar and performs nonlinear normalization calculation, combines it with the conditional probability value in the state transition probability matrix for joint operation, and outputs the value weighted causal information entropy of the current flow path.

[0055] S50 receives value-weighted causal information entropy through the judgment and diagnosis output module 50, executes feature value comparison logic, generates an interrupt control command when the interception trigger condition is met, and performs differential operation on the real-time flow path and baseline directed acyclic graph structure to output diagnostic records.

[0056] To make the technical solution of the present invention clearer, the specific implementation details of each module are discussed in detail below in conjunction with the above-described workflow.

[0057] See attached document Figure 3 , Figure 3 This is a data processing principle diagram of a log acquisition and timing calibration module according to an embodiment of the present invention.

[0058] In this embodiment, the log collection and timing calibration module 10 performs the access and timing cleaning of underlying system data. This process specifically includes the following steps:

[0059] S101, to establish a cross-system data acquisition path, the log acquisition and timing calibration module 10 configures a bypass listening component on the network side of the front-end business system and the back-end financial system, or connects to the enterprise service bus to perform message subscription, thereby acquiring unstructured operation logs generated by different systems. For the specific network protocol configuration for using the bypass listening component or message subscription mechanism to capture and parse logs from heterogeneous systems, those skilled in the art can set it according to the interface specifications of the source system; this is well-known technology in the field and will not be elaborated upon here.

[0060] S102, after acquiring the underlying logs, the system needs to structure them uniformly for subsequent calculations. Specifically, the log acquisition and time-series calibration module 10 performs data extraction and field mapping processing on the acquired unstructured operation logs, converting them into a standardized discrete event tuple format. The log acquisition and time-series calibration module 10 stores the continuously generated discrete event tuples as a global discrete event set. Any discrete event tuple in the set. The data structure is defined as follows:

[0061] ;

[0062] In the formula, It represents a global business serial number that spans different application systems and is used to associate upstream and downstream operations of the same business. Indicates the source system identifier that generated and recorded this log; It indicates the business type attribute of the document (such as daily expense reimbursement, bulk purchase, asset depreciation, etc.), and is used to provide the underlying data support for classification alignment and independent benchmark threshold calculation for subsequent similar business data; This represents the absolute physical timestamp generated by the underlying hardware clock of the source system. As a preferred method, this timestamp uses a Unix timestamp numerical format with millisecond precision. The character representing the specific operation action corresponding to this log; This field represents a scalar value representing the financial amount extracted from the log payload. Its data type is usually set to double-precision floating-point. When the operation event does not contain a financial variable attribute, this field is assigned the constant 0. It represents a foreign key representing the document lineage, which is essentially an identifier of a preceding document that must be completed before the current event is executed, extracted from the log payload.

[0063] S103. Because the front-end business system and the back-end financial system are deployed on different physical device nodes, and there are often network delays or situations where system clocks are not strictly synchronized, the system clock deviation between devices has a high probability of causing a discrepancy between the absolute physical timestamp and the actual order of business transactions. To troubleshoot such conflicts, the log collection and time-series calibration module 10 uses a global business serial number. As an aggregate primary key, perform grouping and comparison operations on discrete event tuples.

[0064] The log collection and timing calibration module 10 retrieves dependencies between tuples within the same group. When an event is detected... Included document lineage foreign keys The target data record is an event. At this time, this dependency indicates the event at the business logic level. It occurred earlier than the event. The log collection and time-series calibration module 10 extracts the absolute physical timestamps of both and performs numerical comparison. If the comparison result shows... If the size relationship is determined, then the data in that group is determined to have triggered a timing reversal conflict.

[0065] S104, the log collection and timing calibration module 10 initiates a logical calibration mechanism for grouped data that triggers timing inversion conflicts. The log collection and timing calibration module 10 uses event tuples within the group as vertices of a graph structure, with document lineage foreign keys as the vertices. The dependency relationship of the pointer is constructed by directed edges pointing from preceding events to subsequent events, and the document lineage foreign key is linked. The relationships between the elements are used as directed edges to construct a local directed dependency graph in system memory. Considering that actual business data may contain mutual references between documents due to abnormal system interactions, the log collection and time-series calibration module 10 performs directed cycle detection after constructing the dependency graph. If a circular dependency cycle is detected in the graph, the system compares the absolute physical timestamps of related events within the cycle. This forces the disconnection of reverse connections between events with later timestamps and events with earlier timestamps, thereby degrading the network into a compliant acyclic network.

[0066] Subsequently, the log collection and timing calibration module 10 performs a topological sorting operation on the local directed dependency graph. The output of the topological sorting results in all vertices with directed edge constraints forming a linear sequence, and the starting vertex of each directed edge is always positioned before the ending vertex in the sequence. Based on the position of the sorted linear sequence, the log collection and timing calibration module 10 generates incrementally increasing logical timing variables for each event tuple. Logical timing variables In the data flow structure, it is configured to replace the original absolute physical timestamp. This generates a standardized time series that eliminates timestamp drift for downstream modules to use.

[0067] See attached document Figure 4 , Figure 4 This is a schematic diagram of the data dimensionality reduction and matching principle of a cardinality mapping module according to an embodiment of the present invention.

[0068] In this embodiment, the cardinality mapping module 20 receives discrete event tuples containing logical time-series variables and performs cross-domain cardinality mapping to resolve the data misalignment problem caused by the many-to-many correspondence between business and financial documents. This process specifically includes the following steps:

[0069] S201, the cardinality mapping module 20 divides the discrete event tuples into business event subsets based on their system source attributes. subset of financial events When faced with massive amounts of system logs, using full cross-comparison can easily lead to a significant consumption of computing resources or even system blockage.

[0070] To this end, the cardinality mapping module 20 performs spatial dimension isolation and dimensionality reduction operations. Considering that in real-world financial transaction rules, only documents for the same business entity will be merged or split for settlement, the cardinality mapping module 20 extracts the business entity identifier from the discrete event tuple. As a preferred approach, this business entity identifier uses the unified social credit code or the supplier association number within the system.

[0071] The radix mapping module 20 uses a hash function to calculate the corresponding hash key for the extracted business entity identifier. Based on the hash key, the radix mapping module 20 distributes the global discrete event tuples into independent entity space buckets, so that the event data of different business entities are isolated from each other in physical memory.

[0072] S202, based on spatial isolation, the cardinality mapping module 20 further performs time dimension reduction operations within each entity's spatial bucket. The cardinality mapping module 20 configures a sliding time window for the data within each bucket. The specific value of this parameter is usually set according to the company's regular financial settlement cycle, for example, it is set to 30 days to 90 days.

[0073] For any financial event that needs to be addressed at present The cardinality mapping module 20 uses the logical time-series variables of the financial event as a benchmark and extracts a subset of business events within a set sliding time window. and subset of financial events Data to generate a candidate business event set And a subset of local financial events. The logic for generating the candidate business event set satisfies the following conditions:

[0074] ;

[0075] In the formula, This represents a single event within a subset of business events. This represents the logical time-series variable of the target financial event. By introducing a sliding time window constraint, the system reduces the matching retrieval interval from the entire database to a fixed period, thereby reducing the time complexity of the matching algorithm to constant level.

[0076] S203, within the candidate matching interval constructed by spatiotemporal dual dimensionality reduction, the cardinality mapping module 20 solves for many-to-many matching combinations based on the capital conservation constraint. Conventional financial settlement operations require that the amounts after merging or splitting front-end business documents must be consistent with the back-end financial accounting amounts. The cardinality mapping module 20 sets the candidate business event set... The dynamic programming algorithm model is used to solve the problem by inputting a subset of local financial events.

[0077] In its implementation, the system uses the target financial amount as the knapsack capacity and the amounts of candidate business events as the item weights, constructing a one-dimensional state transition array for combined summation. Considering that minor numerical deviations often occur in actual business transactions due to currency conversion errors or tax rounding, the system does not employ absolute equality judgments but introduces a monetary tolerance mechanism. The decision constraint inequality of this dynamic programming algorithm is defined as follows:

[0078] ;

[0079] In the formula, Represents a single business event tuple that constitutes a subset of business events; This represents a single financial event tuple that constitutes a subset of financial events. and These represent the subsets of business events and financial events extracted within the dimensionality reduction interval, respectively. and These represent the corresponding event tuples. and Extracted financial amount scalar; As a preferred approach, the monetary tolerance threshold set for the system can be configured to be between 0.01 yuan and 1.00 yuan.

[0080] S204, During the dynamic programming solution process, if an event combination satisfying the above constraint equation is detected, the cardinality mapping module 20 triggers a fusion and packaging instruction. Considering that there may be multiple candidate combination solutions satisfying the tolerance equation within a local interval, the cardinality mapping module 20 extracts the logical time-series variables of the event tuples in each candidate combination and calculates the span variance of the timestamps of each event within the same combination. The cardinality mapping module 20 selects the combination with the smallest variance value as the optimal matching solution, in order to conform to the actual business operation pattern that similar businesses are often centrally settled.

[0081] Furthermore, if a combination satisfying the constraints cannot be found within the current sliding time window, the system will not directly discard the financial event. Instead of processing new business events, the events are moved to a waiting queue. As new business events are added and the time window slides, the events in the waiting queue will be re-entered into the next round of matching operations, thereby ensuring the completeness and closed loop of the data processing logic.

[0082] S205, for the successfully determined optimal matching solution, the cardinality mapping module 20 calls the mapping function. The discrete event set in the determined optimal matching solution is encapsulated to generate composite state macro nodes. The mapping relationship is expressed as follows:

[0083] ;

[0084] Composite state macro nodes Internally, it encapsulates multiple underlying cross-domain events with deterministic mapping relationships. The system continuously outputs multiple such composite state macro nodes based on the data processing progress, forming a standardized sequence of composite state macro nodes. This macro node sequence directly eliminates the technical obstacle of the asymmetry between the underlying business and financial event bases, and provides a convergent, structured node foundation for subsequent time-series graph topology algorithms.

[0085] See attached document Figure 5 , Figure 5 This is a schematic diagram illustrating the working principle of a cause-effect graph construction module according to an embodiment of the present invention.

[0086] In this embodiment, the causal graph construction module 30 establishes a baseline topology model based on the aligned macro node sequence to quantify the temporal causal relationships under normal system operation. This process specifically includes the following steps:

[0087] S301, In order to construct a benchmark model that can reflect the regular business flow pattern, the causal graph construction module 30 extracts the composite state macro node sequence from the system database during the historical compliance period, and uses it as the sequence sample set for constructing the benchmark topology network. .

[0088] Here, the historical compliance period typically refers to a past period in which the system has not experienced any major anomalies or alarms and the business has operated smoothly, such as the most recent full fiscal year. The causal graph construction module 30 extracts various macro nodes from the sequence sample set into a vertex set for the graph structure and introduces the PC causal discovery algorithm to construct the initial fully connected undirected graph. The specific calculation process for performing conditional independence tests and graph structure initialization using the PC causal discovery algorithm can be set by those skilled in the art based on standard causal inference theory; this is well-known technology in the field and will not be elaborated upon here.

[0089] S302. Conventional causal discovery algorithms, when processing financial flow data with strong business logic, are often susceptible to interference from purely statistical correlations, generating reverse causal edges that violate the actual order of business transactions. To address this issue, the causal graph construction module 30 injects time-prior constraints into the algorithm model.

[0090] In practical operation, the cause-effect graph construction module 30 extracts the logical time-series variables encapsulated inside the macro nodes. For any two macro nodes in the initial graph structure that have a connection relationship. and The system compares the logical timing variables of the two. If the determination is... This indicates macro nodes Later in time series The causal graph construction module 30 applies hard directed edge constraints at the algorithm's underlying level, forcibly blocking and prohibiting the generation of edges from... point to The causal edge, while establishing and preserving from point to Directed connections are used to maintain topological connectivity.

[0091] S303, after establishing the topological network structure, the causal graph construction module 30 traverses the network paths of this directed acyclic graph structure, based on the historical sequence sample set. The actual frequency of data transfer between statistical nodes is recorded. The cause-effect graph construction module 30 sets a tracking window during the statistical process. As a preferred approach, the step size parameter of this tracking window is configured to be 2 to 5 nodes. The reason for introducing a tracking window instead of relying solely on adjacent nodes for statistics is that actual corporate financial approval processes often involve non-standard paths such as special approval skips, expedited processing, or bypassing of levels for return. If only the first-order Markov assumption is followed and adjacent nodes are considered, these long-term dependent business characteristics will be lost.

[0092] By expanding the step size statistics range, the system not only counts the flow frequency between adjacent nodes, but also counts the flow frequency between cross-step nodes based on the tracking window, thereby capturing business causal relationships with long-range dependency characteristics.

[0093] S304, the cause-effect graph construction module 30 calculates the conditional probability of a previous node transitioning to the current node based on the statistically derived flow frequency, and generates a state transition probability matrix with directional orientation. To avoid deadlock errors where the denominator is zero due to certain rare flow paths not appearing in historical samples during actual calculations, the causal graph construction module 30 introduces a Laplace smoothing mechanism to correct the probability calculation logic. (Matrix) The formula for calculating the conditional probability of an interior element is defined as follows:

[0094] ;

[0095] In the formula, This indicates that a preceding macro node has already occurred. Under these conditions, subsequent business processes will be routed to the current macro node. The conditional probability; This represents the graph structure stride size between the previous macro node and the current macro node, and its value is affected by the tracking window. Parameter limitations; This indicates the preceding macro node in the historical sequence sample set. Flow to the current macro node The total number of times; Indicates the preceding macro node The total frequency of occurrence in the historical sequence sample set; The smoothing constant, which is used to prevent division by zero errors, is typically set to 1. This represents the total number of types of macro-node states in the directed acyclic graph structure of the system.

[0096] S305, after the aforementioned topology construction and parameter statistical calculations, the causal graph construction module 30 synchronously writes the directed acyclic graph structure containing complete node relationships and the state transition probability matrix quantifying the flow probabilities into the system's graph database for persistent storage. These data structures serve as baseline reference models representing the normal flow logic of the business, providing a comparison benchmark for subsequent anomaly diagnosis of the real-time computing flow.

[0097] See attached document Figure 6 , Figure 6 This is a schematic diagram of the data calculation principle of a weighted entropy calculation module according to an embodiment of the present invention.

[0098] In this embodiment, the weighted entropy calculation module 40 performs a fusion operation on multi-dimensional data features, jointly quantifying the graph topology information of the pure probability dimension with the scale of funds generated by actual business operations. This process specifically includes the following steps:

[0099] S401, the weighted entropy calculation module 40 continuously monitors the business flow process and acquires the real-time generated composite state macro node stream. In actual corporate financial scenarios, the amount range of different types of documents is usually large. For example, daily expense reimbursements may only be tens of yuan, while large-scale procurement settlements often reach tens of millions of yuan. If the original amount values ​​are directly introduced into the information entropy model, large-value documents will cause excessive value skew, thereby reducing the model's sensitivity to anomalies in normal business. Based on this, the weighted entropy calculation module 40 extracts the currently acquired composite state macro nodes in real time. The accompanying financial amount is scalarized and its absolute value is taken before calculation to represent the actual scale of fund transfers regardless of the direction of lending.

[0100] S402, to unify the vastly different financial data to a similar operational scale, the weighted entropy calculation module 40 performs nonlinear normalization calculations on the processed financial amount scalars, converting them into financial value coefficients mapped to a specified interval. In this embodiment, the system uses a logarithmic nonlinear function containing a smoothing constant to achieve the dimensional transformation. The specific mathematical model for this nonlinear normalization calculation is defined as follows:

[0101] ;

[0102] In the formula, This represents the macronodes generated after nonlinear normalization. The financial value coefficient ranges from 0 to 1. Indicates extraction from the current macro node The absolute value of the financial amount; This represents a preset numerical smoothing constant. Configuring this constant prevents the logarithmic function from reporting an error when the amount is 0. As a preferred method, this smoothing constant... The value is a constant 1; This represents the benchmark for the maximum single transaction amount derived from a historical compliance period sequence sample set. This benchmark is dynamically adjusted as historical compliance data is periodically updated.

[0103] Compared to conventional linear normalization, by introducing a logarithmic function, the system can effectively smooth out the long-tail distribution characteristics of financial funds, so that the monetary features of different orders of magnitude are mapped relatively evenly to the same computational space.

[0104] S403, based on the obtained financial value coefficient, the weighted entropy calculation module 40 establishes a data reading connection with the graph database. Specifically, the weighted entropy calculation module 40 calculates data based on the current composite state macro nodes. The system retrieves the preceding node sequence from the graph database based on the business identifier attribute, and simultaneously retrieves the baseline state transition probability matrix containing historical state transition information. The weighted entropy calculation module 40 extracts the representation of the transition from the preceding node to the current node from the state transition probability matrix. The conditional probability values ​​for each item.

[0105] S404, furthermore, the weighted entropy calculation module 40 constructs a value-weighted causal information entropy joint calculation model. Conventional information entropy theory only focuses on the uncertainty of an event's occurrence, neglecting the scale of the event's physical impact. To accurately characterize the degree of risk in the financial process, the weighted entropy calculation module 40 performs a multiplicative fusion operation between the information entropy calculation formula and the aforementioned financial value coefficients. For the current macro node... Value-weighted causal information entropy The joint computation logic is defined as follows:

[0106] ;

[0107] In the formula, This represents the value-weighted causal information entropy of the current node, which combines the characteristics of the capital scale with the probability of state transition. This represents the financial value coefficient output from the pre-calculation. This represents the threshold value for the tracking window step size configured in the causal graph construction phase of the system; Indicates the step index number of the preceding node within the tracking window; Indicates the location at the current node No. ahead Preceding macro nodes of the step; This represents the value extracted from the state transition probability matrix at the preceding node. Under the condition of flow to the current node The conditional probability; This represents the base-2 logarithmic function, used to quantify the self-information of the probability dimension.

[0108] It is particularly important to note that, considering the objective reality in business processes that the initial node (such as the initial creation of a business application form) may lack preceding nodes, directly substituting the above formula would cause the algorithm to fail to obtain the conditional probability, resulting in a computational dead zone. Therefore, the weighted entropy calculation module 40 dynamically determines the preceding path length of the current node before performing the summation. If the current node has preceding nodes but their number is less than the tracking window... The system only sums the actual preceding node items; if the current node is the first node with no preceding nodes, the system skips the accumulation process and directly assigns the node a preset initial baseline entropy constant, thereby ensuring the computational continuity and completeness of the algorithm under various boundary scenarios.

[0109] S405, through the above model calculation, the weighted entropy calculation module 40 fuses multi-dimensional features into a single scalar indicator. Within this calculation framework, if a financial transaction is a rare event in the historical probability distribution (conditional probability approaches 0, leading to increased self-information), and is simultaneously accompanied by a large expenditure (financial value coefficient approaches 1), the fusion of the two will output a higher value-weighted causal information entropy. The weighted entropy calculation module 40 continuously calculates the entropy value of each node along the real-time business flow path, outputting a value-weighted causal information entropy array containing temporal characteristics, and pushes it downstream to the anomaly diagnosis network, providing a basic quantitative basis for judging the compliance of the current business flow chain.

[0110] See attached document Figure 7 , Figure 7 This is a schematic diagram illustrating the operating principle of a determination and diagnosis output module according to an embodiment of the present invention.

[0111] In this embodiment, the judgment and diagnosis output module 50 receives the feature quantification data calculated by the front end and performs the final risk interception and anomaly backtracking analysis. This process specifically includes the following steps:

[0112] S501, the judgment and diagnosis output module 50 sets and dynamically updates the benchmark entropy threshold based on historical security benchmarks. Considering the inherent differences in process complexity and fund distribution among different types of financial transactions, using a single global threshold could easily lead to the averaged masking of business characteristics. Therefore, in this embodiment, the system classifies the historical compliance period sequence sample set according to the business type attributes of the documents. For each of the divided independent business sets, the system extracts the value-weighted causal information entropy of its composite state macro nodes and performs statistical operations to obtain data distribution characteristics. The calculation logic of this benchmark entropy threshold is defined as follows:

[0113] ;

[0114] In the formula, This represents the baseline entropy threshold used by the system to determine anomalies for a specific business type. This represents the statistical mean of the value-weighted causal information entropy generated during historical compliance periods under the corresponding business type; The standard deviation of the historical entropy distribution; This represents the sensitivity adjustment coefficient configured in the system. As a preferred approach, this coefficient is typically set between 2 and 3 to cover a confidence interval of approximately 95% or more for normal business operations. As the compliant sample data is periodically expanded, the judgment and diagnosis output module 50 will periodically re-execute the aforementioned statistical calculations to achieve dynamic adaptive updates to the baseline entropy threshold.

[0115] S502, the judgment and diagnosis output module 50 continuously receives the value-weighted causal information entropy of the current flow path pushed in real time by the weighted entropy calculation module 40, and calls the internal Boolean logic comparator to perform feature value comparison. When the value-weighted causal information entropy calculated in real time is less than or equal to the baseline entropy threshold... When the system determines that the current business action conforms to the normal flow pattern, it allows the data flow to proceed.

[0116] If the real-time value-weighted causal information entropy exceeds the threshold, the system determines that the current process has a high-risk anomaly and immediately generates an interruption control command. This interruption control command is directed to the enterprise service bus and related business execution system interfaces. The system suspends the current financial flow process at the underlying level and freezes the flow and scheduling of related funds, thereby achieving system-level blocking of potentially risky business operations.

[0117] S503, after triggering the interception action, the judgment and diagnosis output module 50 initiates a topology network differential operation to trace back and locate the root cause of the abnormal process. The judgment and diagnosis output module 50 extracts the intercepted service flow path and extracts each node instance within it, converting it into the corresponding macro node state type. It then parses this into a real-time directed edge set consisting of the current node and all its preceding service nodes within a defined window. Simultaneously, the judgment and diagnosis output module 50 interacts with the graph database, retrieves the baseline directed acyclic graph structure that has been fixed and saved during the causal graph construction phase, and extracts the set of expected directed edges corresponding to the start and end paths of the current business flow. By mapping from node instances to state types in a dimensionality reduction manner, the system ensures the isomorphism of the real-time path and the baseline model in terms of data dimensions, thereby eliminating the type mismatch gap caused by direct comparison of concrete instances and abstract rules.

[0118] S504, for the two sets of acquired topological data, the judgment and diagnosis output module 50 performs a bidirectional difference operation on the set dimension. Specifically, the system calculates the actual redundant difference set and the baseline missing difference set respectively. The system will then use the real-time directed edge set... Remove those that exist in the set of compliant directed edges The system generates an actual redundant difference set after processing the elements in the data. The directed edges contained in this difference set indicate, at the business logic level, that the current data flow has executed unrecorded paths such as unauthorized skipping or special approval for violations. Simultaneously, the system will use the baseline expected directed edge set... Remove those that exist in the real-time directed edge set The baseline missing difference set is generated after the elements in the baseline. This difference set indicates that the current process has missed a necessary pre-approval or verification node specified in the baseline. The decision and diagnosis output module 50 marks the elements contained in the two difference sets as unmatched edges.

[0119] S505, based on the located unmatched edges, the judgment and diagnosis output module 50 extracts relevant node parameters and integrates them to output a structured diagnostic record. This diagnostic record includes the global business serial number that triggered the interruption command, the absolute physical timestamp when the interception occurred, the specific node combination of the violation jump, and the missing preceding node attribute information inferred through reverse comparison of the baseline graph structure. The judgment and diagnosis output module 50 persistently stores this diagnostic record and pushes it to the front-end system management panel, providing data location basis for business management personnel to subsequently perform compliance reviews and document repairs.

[0120] To further aid in understanding the technical solution of this invention, a specific application embodiment of this invention is given below, taking into account a specific enterprise-level financial procurement to payment business flow scenario.

[0121] Specific application examples:

[0122] See attached document Figure 8 , Figure 8 This is a timing sequence diagram for anomaly diagnosis in a purchase-to-payment business scenario according to an embodiment of the present invention.

[0123] In the specific application scenario of this embodiment, the supply chain management system (front-end business system) and the ERP financial core system (back-end financial system) of a large manufacturing enterprise are deployed separately.

[0124] In daily business operations, the log collection and time-series calibration module 10 continuously monitors the two heterogeneous systems mentioned above. When a procurement transaction occurs, the front-end system sequentially generates logs for Purchase Order Creation (PO) and Material Receipt (GR); the back-end financial system subsequently generates logs for Invoice Verification (INV) and Payment Settlement (PAY). Due to network latency in cross-system interactions, the physical timestamp of a certain transaction's GR may be later than the invoice verification timestamp. The system uses document lineage foreign keys to... Upon detecting that invoice verification depends on the inbound order, a timing logic calibration mechanism is triggered, forcibly resetting the logical timing variables of the inbound order. Placing it before invoice verification eliminates timing reversal errors caused by timestamp drift.

[0125] Since enterprises often adopt a settlement model of multiple receipts and centralized invoicing, the base mapping module 20 extracts the supplier's credit code as a hash key, and then maps the supplier's credit code to the credit code within the past 30 days (sliding time window). Multiple material receipts (business events) and a single total summary invoice (financial event) within a given timeframe are allocated to the same memory bucket. Using a dynamic programming algorithm, the system maintains a tolerance of 0.05 yuan. Within, the system successfully identified a combination of documents with an absolutely matching total amount and packaged them into a single composite status macro node. This reduces the original complex network of many-to-one documents to a standard linear sequence.

[0126] The causal graph construction module 30 uses historical compliance data and a PC algorithm constrained by temporal priors to construct the standard topology network for this business. The statistical matrix shows that the probability of an inbound order being transferred to invoice verification is extremely high, while the probability of a purchase order directly initiating fund clearing (skipping inbound and verification) is close to a minimum.

[0127] During a real-time monitoring session, the system detected a large, unusual purchase settlement. This transaction, consisting only of a purchase order, initiated a fund clearing operation that bypassed its intended level. The weighted entropy calculation module 40 extracted the financial amount of this transaction, performed logarithmic nonlinear normalization, and output a financial value coefficient close to its upper limit. Simultaneously, the algorithm searched the graph database and found that the conditional probability of this leap was extremely low, leading to an exponential amplification of self-information. After multiplicatively fusing the two, the system's output value-weighted causal information entropy exhibited an extreme value anomaly.

[0128] The judgment and diagnosis output module 50 detected that the real-time entropy value far exceeded the baseline entropy threshold set based on bulk procurement transactions. It immediately sent an interrupt control command to the enterprise service bus, physically intercepting the payment at the underlying gateway. Subsequently, the system performed a bidirectional differential operation, comparing the real-time cascading path with the baseline topology, accurately outputting a diagnostic record, clearly indicating the absence of necessary material receiving slips and invoice verification pre-nodes. Based on this structured diagnostic record, the reviewers quickly verified that the operation was an irregular payment, achieving precise risk prevention.

[0129] Experimental verification and effect comparison:

[0130] The experimental dataset uses anonymized real financial and business transaction logs from a group company, spanning 24 months and containing approximately 5.6 million discrete event records. The system extracts data from the first 23 months as samples for model training and baseline graph construction, and the data from the last month as the test set. 500 known anomalies (covering typical violation scenarios such as splitting amounts to circumvent approvals, cross-domain time-series reversal, and skipping key nodes) are manually injected into the test set.

[0131] The experiment introduced two other conventional methods:

[0132] Baseline Model A: Traditional static rule engine model (based on manually preset strong rule conditions for regular expression matching and interception).

[0133] Baseline Model B: Standard PC causal discovery + basic information entropy model (only performs graph topological probability calculation, without introducing time series calibration and financial value weighting mechanism).

[0134] Model C of this invention is a joint computational model that combines time-series calibration, cross-domain cardinality mapping, and value-weighted causal information entropy.

[0135] The experimental data are shown in Table 1: Table 1: Comparison of core performance indicators of the three diagnostic models: Model Name Recall False Positive Rate (FPR) Average Processing Delay Peak Throughput (TPS) Model A 68.4% 8.2% 180 milliseconds 6200 Model B 83.1% 21.5% 4500 milliseconds 450 Model C 97.2% 1.8% 120 milliseconds 5500 .

[0136] in conclusion: See attached document Figure 9 , Figure 9 This is a comparison curve of the ROC performance of three diagnostic models according to an embodiment of the present invention.

[0137] According to Table 1 and Figure 9 Comparative analysis of the data shows that the financial process anomaly diagnosis method based on value-weighted causal information entropy provided by this invention can achieve an effective balance between detection accuracy and computational efficiency when dealing with complex cross-system financial settlement scenarios. The specific conclusions are as follows:

[0138] Regarding the accuracy of anomaly diagnosis, this invention constructs a baseline topology network with time prior constraints and combines it with long-range tracking statistics to compensate for the weak generalization ability of static rule engines, effectively improving the ability to identify non-standard and hidden violation paths. At the same time, the solution introduces a non-linear normalized weighting mechanism based on the scale of financial funds, objectively removing statistical noise interference caused by low-value compliance anomalies, and solving the technical bottleneck of excessively high false alarm rate of pure probabilistic topology models in practical applications.

[0139] In terms of system computational performance, this invention addresses the computational overload problem that is easily caused by cross-domain many-to-many document matching by employing a dual dimensionality reduction strategy of spatial hash buckets and sliding time windows. This strategy transforms complex global graph search into local constant-level matching, significantly reducing the processing time of the underlying algorithm and ensuring that the solution still possesses millisecond-level real-time response and risk blocking capabilities when facing large-scale high-concurrency business throughput.

[0140] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent analysis system for corporate financial data based on artificial intelligence, characterized in that, include: The log collection and time series calibration module is configured to extract source log data from the source system to generate discrete event tuples and reconstruct the time series based on the foreign key dependencies of the documents. The cardinality mapping module is configured to solve for matching combinations of the discrete event tuples within the matching interval defined in the spatiotemporal dimension, based on the capital conservation constraint, and fuse and output a composite state macro node sequence. The causal graph construction module is configured to generate a baseline directed acyclic graph structure and a state transition probability matrix based on the historical compliant composite state macro node sequence and time prior constraints. The weighted entropy calculation module is configured to perform nonlinear normalization on the financial amount scalar of the real-time generated composite state macro node sequence, and output the value-weighted causal information entropy by combining the state transition probability matrix. The judgment and diagnosis output module is configured to generate an interrupt control instruction when the value-weighted causal information entropy satisfies the interception trigger condition, and to perform differential operation in conjunction with the baseline directed acyclic graph structure to output a diagnostic record.

2. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 1, characterized in that, The specific configuration for the log collection and time series calibration module to reconstruct the time series based on document foreign key dependencies is as follows: Extract the absolute physical timestamp and foreign key dependencies of the documents from the source log data; Based on the foreign key dependency relationship of the document, the absolute physical timestamps of the upstream and downstream nodes corresponding to the discrete event tuple are compared. If the absolute physical timestamp of the downstream node is detected to be earlier than the absolute physical timestamp of the upstream node, it is determined that a timing reversal conflict is triggered. A local directed dependency graph is constructed for the data that triggers the time-series inversion conflict. The reverse association edges are disconnected and a topological sorting operation is performed to generate an increasing logical time-series variable for each event tuple, thereby reconstructing the time series.

3. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 1, characterized in that, The specific configuration of the cardinality mapping module in defining the matching interval in the spatiotemporal dimension is as follows: Extract the business entity identifier of the discrete event tuple, calculate the corresponding hash key using a hash function, and then distribute and store the discrete event tuple into independent entity space buckets based on the hash key; A sliding time window is configured inside the physical space bucket. Based on the target financial event in the financial events, a subset of business events and a subset of financial events are extracted within the sliding time window to generate a candidate set of business events and a partial subset of financial events, which constitute the matching interval.

4. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 3, characterized in that, The specific configuration for the cardinality mapping module to solve for matching combinations based on the capital conservation constraint is as follows: Set a tolerance threshold for the amount, use the amount of financial events in the subset of local financial events as the knapsack capacity, and use the amount of business events in the set of candidate business events as the item weight to construct a one-dimensional state transition array. Using a dynamic programming algorithm model, candidate combination solutions are searched within the matching interval for the total amount difference to be within the amount tolerance threshold range; If multiple candidate combination solutions are detected, the one with the smallest event timestamp span variance is selected as the optimal matching solution, and the composite state macro node sequence is fused and output.

5. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 1, characterized in that, The specific configuration for the cause-effect graph construction module to generate a baseline directed acyclic graph structure based on time prior constraints is as follows: The PC causal discovery algorithm is introduced to construct an initial fully connected undirected graph for the composite state macro node sequence; By comparing the order of two macro nodes with a connection in the time series, if it is determined that the first macro node is later than the second macro node in time, a directed edge constraint is applied at the bottom layer of the algorithm to forcibly block the causal edge from the first macro node to the second macro node, while retaining the directed connection from the second macro node to the first macro node, thus generating the baseline directed acyclic graph structure.

6. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 5, characterized in that, The specific configuration for the cause-effect graph construction module to generate the state transition probability matrix is ​​as follows: Set a tracking window with a step size parameter greater than 1, and count the frequency of transitions between cross-step nodes based on the historical compliant composite state macro node sequence; Based on the aforementioned transition frequency, a Laplace smoothing mechanism is introduced to calculate the conditional probability of the previous node transitioning to the current node, thereby generating the state transition probability matrix with directional orientation.

7. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 1, characterized in that, The specific configuration for the weighted entropy calculation module to perform nonlinear normalization on the financial amount scalar is as follows: Extract the financial amount scalar attached to the composite state macro node sequence and take its absolute value; A logarithmic nonlinear function containing a preset numerical smoothing constant is used to perform mapping calculations on the financial amount scalar after taking the absolute value, converting it into a financial value coefficient in the range of 0 to 1.

8. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 7, characterized in that, The specific configuration for the weighted entropy calculation module to output the value-weighted causal information entropy by combining the state transition probability matrix operation is as follows: The sequence of predecessor nodes of the current macro node is obtained by retrieving the pre-configured graph database, and the corresponding conditional probability values ​​are extracted from the state transition probability matrix. Calculate the self-information of the step jump, perform a multiplicative fusion operation between the financial value coefficient and the self-information, and accumulate the weighted information of the effective preceding path to output the value-weighted causal information entropy.

9. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 1, characterized in that, The specific determination configuration of the determination and diagnosis output module regarding the satisfaction of the interception trigger condition is as follows: Based on the business type attributes, perform statistical analysis on the value-weighted causal information entropy generated during historical compliance periods, and dynamically calculate the benchmark entropy threshold by combining the mean and standard deviation; When the received real-time value-weighted causal information entropy is greater than the baseline entropy threshold, it is determined that the interception trigger condition is met, the interruption control instruction is generated and sent to the enterprise service bus to freeze fund scheduling.

10. The intelligent analysis system for unit financial data based on artificial intelligence according to claim 9, characterized in that, The specific configuration for the determination and diagnosis output module to perform differential operations and output diagnostic records in conjunction with the baseline directed acyclic graph structure is as follows: Extract the current flow path that has been intercepted, and convert each node instance within it into a macro node state type to construct a real-time directed edge set; Extract the set of expected directed edges of the baseline corresponding to the current flow start and end paths from the baseline directed acyclic graph structure; The actual redundancy difference set and the baseline missing difference set are calculated for the real-time directed edge set and the baseline expected directed edge set, respectively. The unmatched edges are located based on the difference set, and the diagnostic record containing the attribute information of the missing preceding node is generated.

Images