Intelligent reconciliation error repairing method in financial management system

Abstract

The invention relates to the technical field of financial management analysis, in particular to an intelligent reconciliation error repairing method in a financial management system, which comprises the following steps: obtaining and processing financial data to generate a standardized reconciliation node set, and providing a data basis for reconciliation error analysis; by establishing a time sequence enhanced reconciliation relation graph and identifying a suspected error node cluster, a node set which may have errors is locked; by carrying out feature extraction and root cause classification on the suspected error node cluster, an error root source is judged, and classification confidence is given; an optimal repair scheme is screened out by matching an optimal repair action sequence and virtually simulating a rehearsal consequence, and the actual repair risk is reduced; the execution sequence can be flexibly adjusted through dynamic priority rearrangement of the optimal repair action sequence; the repair effect is verified based on double indexes, and the repair effect can be evaluated from account checking accuracy and business rule conformity; and through the iterative optimization of the reinforcement learning mechanism, the reconciliation error repair capability and accuracy are continuously improved.

Description

Technical Field

[0001] This invention relates to the field of financial management analysis technology, and in particular to an intelligent method for correcting reconciliation errors in a financial management system. Background Technology

[0002] In today's context of digital operations for enterprises, the stable and efficient operation of financial management systems is of paramount importance to the survival and development of enterprises. It is not only the core hub for the flow of corporate funds, but also a key support for enterprises to make strategic decisions, manage risks, and achieve sustainable growth.

[0003] The existing financial management system's reconciliation process has significant flaws that severely impact the accuracy and efficiency of financial management, creating numerous obstacles to the company's financial operations. These flaws include: Firstly, multi-source, heterogeneous financial data, due to inconsistent formats and standards, is difficult to integrate and effectively utilize, leading to a chaotic data foundation during reconciliation and an inability to accurately reflect the true financial situation. Secondly, during data acquisition and transmission, system malfunctions may cause missing or erroneous financial data, severely affecting the reliability of reconciliation results and hindering smooth reconciliation. Thirdly, traditional reconciliation methods often rely on manual operation and simple rule matching, which is inadequate when dealing with large amounts of complex data. The processing of financial data is slow, consuming significant manpower and time. Manual reconciliation is also prone to oversight and errors, further reducing efficiency. When discrepancies arise, the lack of effective analytical tools and methods makes it difficult to quickly and accurately pinpoint the error. Traditional methods may require item-by-item data verification, which is labor-intensive and prone to overlooking crucial information, leading to delays in reconciliation resolution. Furthermore, existing repair strategies are typically fixed and simplistic, unable to adapt flexibly to different error types and business scenarios. The lack of intelligent repair mechanisms prevents automatic analysis and selection of the optimal repair solution, resulting in unsatisfactory repair outcomes. Therefore, this invention proposes an intelligent repair method for reconciliation errors in financial management systems. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art by proposing an intelligent method for correcting reconciliation errors in a financial management system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Intelligent methods for correcting reconciliation errors in financial management systems include: S1. Acquire multi-source heterogeneous financial data, perform spatiotemporal alignment and structuring processing, and generate a standardized set of reconciliation nodes; S2. Based on the multi-dimensional attributes of reconciliation nodes, establish a time-enhanced reconciliation relationship graph and identify suspected error node clusters through dynamic weight calculation; S3. A hierarchical graph attention network is used to perform deep feature extraction and error root cause classification on clusters of suspected error nodes, and the classification confidence and error root cause type are output. S4. Based on the error root cause classification results, the optimal repair action sequence is matched from the adaptive repair strategy library, and a virtual simulation module is introduced to preview the repair consequences. S5. Utilize a multi-head attention mechanism to dynamically prioritize the optimal repair action sequence and generate the final repair execution plan; S6. Execute the final repair plan and verify the repair effect based on the dual indicators of reconciliation consistency index increment and business rule compliance. S7. Based on the repair results, iteratively optimize the adaptive repair strategy library and the hierarchical graph attention network model through reinforcement learning mechanism.

[0006] Furthermore, in step S1, the process of acquiring multi-source heterogeneous financial data, performing spatiotemporal alignment and structuring processing, and generating a standardized reconciliation node set includes: Obtain multi-source heterogeneous financial data from bank interfaces, core business systems, and third-party payment platform databases. This multi-source heterogeneous financial data includes transaction flow data, account balance data, and historical reconciliation log data. Cleaning and standardization of the acquired multi-source heterogeneous financial data; Based on at least one association key, the multi-source heterogeneous financial data after cleaning is fused and associated to identify data records belonging to the same business entity. Using a single transaction or a logically related set of transactions formed through fusion and association operations as the basic unit, reconciliation nodes are divided, and a globally unique identifier and attribute set are generated for each reconciliation node, thereby forming a standardized set of reconciliation nodes. The attribute information of the reconciliation node includes at least: node ID, transaction amount, currency, lending direction, transaction timestamp, local account number, counterparty account number, transaction summary text, and spatial location index in the reconciliation relationship graph.

[0007] Furthermore, in step S2, the process of establishing a time-enhanced reconciliation graph and identifying suspected error node clusters through dynamic weight calculation includes: An initial graph structure is constructed using each reconciliation node in the standardized reconciliation node set as a vertex; the feature vector of each vertex consists of its attribute information, including at least numerical features and textual features. Based on the attribute information of reconciliation nodes, various types of edges are established between vertices to create a time-enhanced reconciliation relationship graph; among them, the edge types include at least: fund flow edges, time-related edges, and semantically similar edges; For each edge in the graph, a comprehensive dynamic weight value is calculated. For multiple edges of different types connecting the same pair of vertices, their corresponding similarities are weighted and fused to form the final weight of the comprehensive edge between the vertex pair. For edges of only a single type, the similarity of that edge is the final weight. Based on dynamic weight values, the weighted SCAN graph clustering algorithm is used to cluster the vertices in the reconciliation graph; the set partitioning of vertices is completed based on the structural similarity of vertices and edge weights; and the set of vertices that meet the preset conditions is identified as a cluster of suspected erroneous nodes.

[0008] Furthermore, in step S3, the process of using a hierarchical graph attention network to perform deep feature extraction and error root cause classification on clusters of suspected erroneous nodes, and outputting classification confidence and error root cause type, includes: Hierarchical graph attention networks consist of node-level attention layers and graph-level attention layers; From the time-enhanced reconciliation graph, extract all vertices and edges within each cluster of suspected erroneous nodes and its K-hop neighborhood to form multiple subgraphs to be analyzed. For each subgraph, a deep feature representation of each node is learned through a node-level attention layer; a global representation of the entire subgraph is aggregated through a graph-level attention layer. The global representation vector of each subgraph is input into a fully connected neural network classifier, which outputs the raw scores corresponding to each error root cause category. The raw scores are converted into a probability distribution by the Softmax function, thereby obtaining the classification confidence of the suspected error node cluster to each category and the final error root cause type. At the same time, the confidence of each classification result is output.

[0009] Furthermore, in step S4, the process of matching the optimal repair action sequence from the adaptive repair strategy library and introducing the virtual simulation module to pre-simulate the repair consequences includes: Based on the output error root cause classification results and the context attributes of suspected error node clusters, retrieve all matching <error root cause-context condition-repair action> triples from the adaptive repair strategy library; The retrieved repair actions are initially sorted based on their historical execution success rates to form a candidate set of repair actions; The current time-series enhanced reconciliation graph state is replicated as the basis of the virtual simulation environment. Within this environment, an independent simulation branch is created for each candidate repair action sequence in the repair action candidate set. The corresponding repair action sequence is executed in each simulation branch, and evaluation metrics are calculated. These metrics include the estimated repair time (RT), the impact range index (NI) on upstream and downstream nodes, and the business rule conflict probability (BP). Based on historical log data, the average time required to execute all actions in the sequence is calculated to obtain the estimated repair time (RT). After executing the repair actions, the impact range index (NI) on upstream and downstream nodes is obtained by calculating the ratio of the number of nodes whose state has changed to the total number of nodes in the virtual simulation environment. A node refers to a cluster of suspected erroneous nodes in the reconciliation graph and all nodes within its first-order neighborhood. The repaired reconciliation graph is verified using a rule engine. The ratio of the number of conflicting rules to the total number of verified rules is the business rule conflict probability (BP). The rule engine is used to verify whether the repaired virtual reconciliation graph violates predefined business rules. The comprehensive evaluation score for each candidate sequence is calculated using CAS. The repair sequence with the highest comprehensive evaluation score (CAS) is selected as the optimal repair sequence.

[0010] Furthermore, in step S5, the process of dynamically prioritizing the optimal repair action sequence using a multi-head attention mechanism to generate the final repair execution plan includes: Each action in the optimal repair action sequence is represented as an action feature vector; Arrange the action feature vectors of the entire optimal repair action sequence in the initial order to form a sequence matrix; and generate a position encoding vector for each position in the sequence. Add the position encoding vector to the corresponding action feature vector to inject the order information of the action in the sequence. The sequence matrix after injecting position information is input into a multi-head attention network for calculation; the operation analysis is performed on each attention head in the multi-head attention network; Each action's context-enhanced action feature vector, after being enhanced by an attention mechanism, is mapped to a scalar action priority score through a fully connected layer. The scalar action priority score is then weighted and combined with the system's real-time load factor to generate the final execution priority of each action. The optimal repair action sequence is then rearranged in descending order of final execution priority to generate the final repair execution plan.

[0011] Furthermore, in step S6, the final repair execution plan is executed, and the process of verifying the repair effect based on the dual indicators of reconciliation consistency index increment and business rule compliance includes: In the virtual simulation environment, the final repair execution plan is executed to obtain the repaired reconciliation graph state; based on this state, the node matching logic used when constructing the time-series enhanced reconciliation graph is re-executed to calculate the successfully matched node pairs among all reconciliation nodes; After executing the final repair plan, calculate the post-repair reconciliation consistency index. Its value is the ratio of the number of successfully matched node pairs to the total number of reconciliation nodes; calculate the reconciliation consistency index increment ΔC; The repaired reconciliation relationship diagram status is re-entered into the rule engine for verification; the business rule compliance CR is defined as the ratio of the number of rules that pass verification to the total number of rules in the rule engine; A comprehensive judgment is made based on the reconciliation consistency index increment ΔC and the business rule compliance CR. Based on the judgment results, the repair case, along with its input features, the sequence of repair actions performed, and the final (ΔC,CR) result, is stored in the training set as a sample with a clear result label.

[0012] Furthermore, in step S7, the process of iteratively optimizing the adaptive repair strategy library and the hierarchical graph attention network model through reinforcement learning based on the repair effect includes: Each repair task is abstracted into a reinforcement learning environment state. This state is characterized by the subgraph features of the suspected faulty node cluster and the contextual features of the repair action sequence. The adaptive repair policy library can be viewed as a policy function π of a reinforcement learning agent, where the policy function depends on the environment state. Output repair action The combined result of the reconciliation consistency index increment ΔC and the business rule compliance CR is quantified into a reward signal. ; The parameters of the policy function π are updated using a proximal policy optimization algorithm with the goal of maximizing cumulative reward. The model update cycle is preset; samples stored in the training set are periodically used as labeled supervised learning data to fine-tune the hierarchical graph attention network. For a verified and marked as a fully successful repair case, its corresponding <root cause-context condition-repair action> triple is treated as a new knowledge unit. If the unit is not in the adaptive repair strategy library, it is added to the library; if it already exists, its selection probability is updated based on the near-end strategy optimization results.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: By processing multi-source heterogeneous data to generate a standardized reconciliation node set and constructing a time-enhanced reconciliation relationship graph, dynamic weights are used to identify clusters of suspected erroneous nodes, enabling precise location of reconciliation errors and providing accurate targets for subsequent repairs. By employing a hierarchical graph attention network to perform deep feature extraction and error root cause classification on clusters of suspected erroneous nodes, outputting classification confidence and error root cause types, it is easier to deeply understand the causes of errors and provide a basis for formulating targeted repair strategies. By matching the optimal repair action sequence from the adaptive repair strategy library and introducing a virtual simulation module to pre-enact the repair consequences, a multi-head attention mechanism is used to dynamically prioritize and generate the final repair execution plan, improving the efficiency and rationality of the repair strategy. The repair effect is verified based on the incremental reconciliation consistency index and business rule compliance as dual indicators, allowing for a comprehensive evaluation of repair effectiveness from both reconciliation accuracy and business rule compliance aspects, ensuring that the repair results are both accurate and compliant. Through iterative optimization of the adaptive repair strategy library and the hierarchical graph attention network model using a reinforcement learning mechanism, the system can continuously adapt to new reconciliation scenarios and error types, improving overall repair capabilities and accuracy. Attached Figure Description

[0014] Figure 1 This is a flowchart of the intelligent repair method for reconciliation errors in the financial management system proposed in this invention. Detailed Implementation

[0015] The technical solutions of 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.

[0016] Reference Figure 1 Intelligent methods for correcting reconciliation errors in financial management systems include: S1. Acquire multi-source heterogeneous financial data, perform spatiotemporal alignment and structuring processing, and generate a standardized set of reconciliation nodes; S2. Based on the multi-dimensional attributes of reconciliation nodes, establish a time-enhanced reconciliation relationship graph and identify suspected error node clusters through dynamic weight calculation; S3. A hierarchical graph attention network is used to perform deep feature extraction and error root cause classification on clusters of suspected error nodes, and the classification confidence and error root cause type are output. S4. Based on the error root cause classification results, the optimal repair action sequence is matched from the adaptive repair strategy library, and a virtual simulation module is introduced to preview the repair consequences. S5. Utilize a multi-head attention mechanism to dynamically prioritize the optimal repair action sequence and generate a final repair execution plan with controllable risk. S6. Execute the final repair plan and verify the repair effect based on the dual indicators of reconciliation consistency index increment and business rule compliance. S7. Based on the repair results, iteratively optimize the adaptive repair strategy library and the hierarchical graph attention network model through reinforcement learning mechanism.

[0017] It should be further explained that, in the specific implementation process, step S1, which involves acquiring multi-source heterogeneous financial data, performing spatiotemporal alignment and structuring processing, and generating a standardized set of reconciliation nodes, includes: Obtain multi-source heterogeneous financial data from bank interfaces, core business systems, and third-party payment platform databases. This multi-source heterogeneous financial data includes transaction flow data, account balance data, and historical reconciliation log data. The acquired multi-source heterogeneous financial data is cleaned and standardized, including: removing invalid records, filling in missing fields, and standardizing the format of amounts and dates; Based on at least one association key, the multi-source heterogeneous financial data after cleaning is fused and associated to identify data records belonging to the same business entity. Using a single transaction or a logically related set of transactions formed through fusion and association operations as the basic unit, reconciliation nodes are divided, and a globally unique identifier and attribute set are generated for each reconciliation node, thereby forming a standardized set of reconciliation nodes. The attribute information of the reconciliation node includes at least: node ID, transaction amount, currency, lending direction, transaction timestamp, local account number, counterparty account number, transaction summary text, and spatial location index in the reconciliation relationship graph.

[0018] It should be further explained that, in the specific implementation process, step S2, which involves establishing a time-enhanced reconciliation graph and identifying suspected erroneous node clusters through dynamic weight calculation, includes: An initial graph structure is constructed using each reconciliation node in the standardized reconciliation node set as a vertex; the feature vector of each vertex consists of its attribute information, including at least numerical features and textual features. Based on the attribute information of reconciliation nodes, various types of edges are established between vertices to create a time-enhanced reconciliation graph. The edge types include at least: fund flow edges, time-related edges, and semantically similar edges, and multiple types of edges are allowed to exist simultaneously between any two vertices. Specifically, fund flow edges are established by creating directed edges based on the correspondence between the user's account and the counterparty's account, according to the direction of fund flow. Time-related edges are established by creating directed edges between transaction nodes that occur within a preset time window, based on the order of transaction timestamps. Semantically similar edges are established by creating undirected edges between transaction nodes whose similarity exceeds a threshold, based on the semantic similarity of the transaction summary text. For each edge in the graph, a comprehensive dynamic weight value is calculated. For multiple edges of different types connecting the same pair of vertices, their corresponding similarities are weighted and fused to form the final weight of the comprehensive edge between the vertex pair. For edges of only a single type, the similarity of that edge is the final weight. Based on dynamic weight values, a weighted SCAN graph clustering algorithm is used to cluster vertices in the reconciliation graph. This algorithm, building upon the classic SCAN algorithm's reliance on structural similarity, introduces edge weights as a clustering criterion. Vertex set partitioning is completed based on vertex structural similarity (i.e., the similarity of shared neighbors) and edge weights. Vertex sets meeting preset conditions are identified as suspected erroneous node clusters: Preset condition 1: The average dynamic weight value of all edges within the cluster is less than 50% of the global average edge weight; Preset condition 2: There are no connections between this cluster and any other cluster in the graph whose size exceeds 5% of the total number of nodes, where the edge weight is higher than the weight threshold.

[0019] It should be further explained that, in the specific implementation process, step S3, which uses a hierarchical graph attention network to perform deep feature extraction and error root cause classification on clusters of suspected erroneous nodes, and outputs classification confidence and error root cause type, includes: Hierarchical graph attention networks consist of node-level attention layers and graph-level attention layers; From the time-enhanced reconciliation graph, extract all vertices and edges within each cluster of suspected erroneous nodes and its K-hop neighborhood to form multiple subgraphs to be analyzed. For each subgraph, a deep feature representation of each node is learned through a node-level attention layer, specifically: A1. For each target vertex in the subgraph Calculate its relationship with each of its neighboring vertices. The original attention coefficients between The calculation formula is as follows: In the formula, Represents the target vertex and neighboring vertices The initial feature vector; Index the target vertex; For the target vertex A neighboring vertex index; For the shared weight matrix; || denotes the vector concatenation operation; To characterize the target vertex With neighboring vertices The edge feature vectors of all edge types and their weights; It is a learnable attention vector; This is the matrix transpose. For activation functions; A2. Use the Softmax function to target vertex. Normalize the original attention coefficients of all neighbors to obtain normalized node-level attention weights. The calculation formula is as follows: In the formula, For the target vertex The set of neighbors in the subgraph; It is an exponential function; For summation index, to distinguish it from external specific indexes , indicating a set of neighbors Sum the results of all vertices within the range; Represents the target vertex Its neighboring vertices The unnormalized attention coefficients between them; understandably, the Softmax function is used to normalize the attention coefficients into node-level attention weights in order to perform weighted aggregation of neighboring node features; A3. The features of neighboring nodes are weighted and summed according to the node-level attention weights, and then a non-linear activation function is used to obtain the target vertex. New feature representation The calculation formula is as follows: In the formula, For activation functions; The global representation of the entire subgraph is aggregated through a graph-level attention layer, specifically: B1. Update all vertex features after the node-level attention layer. As input (where, This indicates the total number of vertices in the currently processed subgraph. This indicates that in this subgraph, from the 1st to the th... (a set of updated eigenvectors for all vertices). B2. Calculate the graph-level attention weight of each vertex for the current subgraph. The global representation vector of the subgraph is obtained by weighted summation. ; The global representation vector of each subgraph Input a fully connected neural network classifier, which outputs raw scores corresponding to each error root cause category. The raw scores are converted into a probability distribution using the Softmax function, thus obtaining the classification confidence of the suspected error node cluster to each category and the final error root cause type. The error root cause types include: amount entry error, account mapping deviation, time sequence logic conflict, duplicate accounting, and compound error. The confidence of each classification result is also output.

[0020] It should be further explained that, in the specific implementation process, step S4, which involves matching the optimal repair action sequence from the adaptive repair strategy library and introducing the virtual simulation module to pre-simulate the repair consequences, includes: Based on the output error root cause classification results and the context attributes of suspected error node clusters, retrieve all matching <error root cause-context condition-repair action> triples from the adaptive repair strategy library; where the context attributes include at least: the account type involved in the error, the amount range, and the time sensitivity; The retrieved repair actions are initially sorted based on their historical execution success rates to form a candidate set of repair actions; where the historical execution success rate represents the proportion of successful executions of a particular repair action in the historical records out of the total number of executions. The current time-series enhanced reconciliation graph state is replicated as the basis of the virtual simulation environment. Within this environment, an independent simulation branch is created for each candidate repair action sequence in the repair action candidate set. The corresponding repair action sequence is executed in each simulation branch, and evaluation metrics are calculated. These metrics include the estimated repair time (RT), the impact range index (NI) on upstream and downstream nodes, and the business rule conflict probability (BP). Based on historical log data, the average time required to execute all actions in the sequence is calculated to obtain the estimated repair time (RT). After executing the repair actions, the impact range index (NI) on upstream and downstream nodes is obtained by calculating the ratio of the number of nodes whose state has changed to the total number of nodes in the virtual simulation environment. A node refers to a cluster of suspected erroneous nodes in the reconciliation graph and all nodes within its first-order neighborhood. The repaired reconciliation graph is verified using a rule engine. The ratio of the number of conflicting rules to the total number of verified rules is the business rule conflict probability (BP). The rule engine is used to verify whether the repaired virtual reconciliation graph violates predefined business rules, which at least include: debit and credit balance rules, non-negative account balance rules, and transaction time sequence continuity rules. The comprehensive evaluation score (CAS) for each candidate sequence is calculated using the following formula: In the formula, This is the normalized repair time; This is the normalized influence range index; The preset decision weight coefficients, and ; The repair sequence with the highest comprehensive evaluation score (CAS) is selected as the optimal repair sequence.

[0021] It should be further explained that, in the specific implementation process, step S5, which uses a multi-head attention mechanism to dynamically prioritize the optimal repair action sequence and generate the final repair execution plan, includes: Each action in the optimal repair action sequence is represented as an action feature vector, wherein the dimensions of the action feature vector include at least: action type encoding, normalized estimated execution time, required resource level, classification confidence of the error root cause type it targets, and the influence range index corresponding to the action in the virtual simulation pre-play. Arrange the action feature vectors of the entire optimal repair action sequence in the initial order to form a sequence matrix; and generate a position encoding vector for each position in the sequence. Add the position encoding vector to the corresponding action feature vector to inject the order information of the action in the sequence. The sequence matrix after injecting positional information is input into a multi-head attention network for calculation. Operational analysis is performed on each attention head in the multi-head attention network: D1. The action feature vector of each action in the sequence is linearly mapped to a sequence query vector, a sequence key vector, and a sequence value vector; D2. For each target action in the sequence, its sequence query vector is multiplied by the sequence key vectors of all actions in the sequence (including itself) to obtain the unnormalized sequence attention score; D3. Based on the preset dependency constraint matrix between the repair actions, the sequence attention scores corresponding to actions that do not conform to the dependency relationship are set to negative infinity to ensure the rearranged sequence... Without violating the causal logic between actions; D4, normalize the sequence attention scores after masking to obtain the sequence attention weights; D5, sum all sequence value vectors according to the sequence attention weights to obtain the context-enhanced action feature vector of the target action under this attention head; concatenate the context-enhanced action feature vectors of multiple attention heads and perform another linear transformation to obtain the final feature representation of the repair action sequence enhanced with context information; where, the context-enhanced action feature vector refers to the feature of a repair action, which, after being calculated by the attention mechanism, incorporates new feature representations that are more context-aware and integrate information from all other actions in the sequence; Each action, enhanced by an attention mechanism, is transformed into a context-enhanced action feature vector and mapped to a scalar action priority score through a fully connected layer. The scalar action priority score is obtained by inputting the context-enhanced action feature vector, representing all key information about the action, into a single-neuron fully connected layer; the output is the scalar action priority score. This score integrates the importance of the action itself and its contextual relationship with other actions, and is compressed into a single comparable value through a learnable linear transformation. The scalar action priority score is then weighted and combined with the system's real-time load factor to generate the final execution priority for each action. The system's real-time load factor is a quantitative indicator reflecting the current system's computational resource strain, ranging from 0 to 1, with higher values ​​indicating a busier system. Finally, the optimal repair action sequence is rearranged according to its final execution priority from highest to lowest to generate the final repair execution plan.

[0022] It should be further explained that, in the specific implementation process, step S6, which involves executing the final repair plan and verifying the repair effect based on both the reconciliation consistency index increment and the business rule compliance indicator, includes: In the virtual simulation environment, the final repair execution plan is executed to obtain the repaired reconciliation graph state; based on this state, the node matching logic used when constructing the time-series enhanced reconciliation graph is re-executed to calculate the successfully matched node pairs among all reconciliation nodes; After executing the final repair plan, calculate the post-repair reconciliation consistency index. Its value is the ratio of the number of successfully matched node pairs to the total number of nodes in the full reconciliation; the increment of the reconciliation consistency index is calculated. ,in In the formula, The initial reconciliation consistency index was calculated based on the original data before the repair. The repaired reconciliation relationship diagram status is re-entered into the rule engine for verification; the business rule compliance CR is defined as the ratio of the number of rules that pass verification to the total number of rules in the rule engine; Based on the increment of the reconciliation consistency index A comprehensive judgment is made based on the business rule compliance (CR) index: if the reconciliation consistency index increases... If the consistency index increment is greater than or equal to the preset effective threshold δ (e.g., within the range of [0.05, 0.15]) and the business rule compliance CR = 1, then the repair is considered completely successful; if the consistency index increment is greater than or equal to the preset effective threshold δ, then the repair is considered completely successful. If the business rule compliance rate (CR) is less than 1 but the preset effective threshold (δ) is greater than or equal to 1, then the rule conflict resolution is considered a failure, triggering a manual review process to focus on checking the conflicting items in the rule engine report; if the reconciliation consistency index increases... If the reconciliation consistency index is less than the effective threshold δ but the business rule compliance CR is 1, it is determined that the reconciliation consistency repair is insufficient, triggering a manual review process, with a focus on checking unsuccessfully matched node pairs; if the reconciliation consistency index increases... If the effective threshold δ and the business rule compliance CR is less than 1, the comprehensive repair is deemed to have failed, triggering the manual review process. Based on the judgment results, this repair case, along with its input features, the sequence of repair actions performed, and the final result, will be included. As a result, a sample with a clear result label is stored in the training set for optimization of the model and policy library in S7.

[0023] It should be further explained that, in the specific implementation process, step S7, the process of iteratively optimizing the adaptive repair strategy library and the hierarchical graph attention network model through reinforcement learning mechanism based on the repair effect includes: Each repair task is abstracted into a reinforcement learning environment state. (in The time step represents the first step in the repair task sequence. (One repair task cycle), this state is jointly characterized by the subgraph features of the suspected faulty node cluster and the context features of the repair action sequence; The adaptive repair policy library can be viewed as a policy function π of a reinforcement learning agent, where the policy function depends on the environment state. Output repair action (i.e., the action sequence retrieved from the strategy library); increment the reconciliation consistency index. The comprehensive assessment result of compliance with business rules (CR) is quantified into a reward signal. The calculation formula is as follows: In the formula, As a reward weighting coefficient, The CR value represents the consistency index increment, and the CR value represents the business rule compliance. The proximal policy optimization algorithm is used to update the parameters of the policy function π with the goal of maximizing cumulative reward. Specifically, the selection probability of each <root cause-context condition-repair action> triple in the policy library is regarded as the output probability of the policy function π. The PPO algorithm is then used to further optimize the policy function π. Data sequences, iteratively adjusting these probability distributions, that is, improving performance under environmental conditions. You can get high rewards. Repair actions Increase the probability of selection and reduce the probability of selection for low-reward actions; The model update cycle is preset (e.g., every M accumulated repair cases or every L hours); the samples stored in the training set are periodically used as labeled supervised learning data to fine-tune the hierarchical graph attention network; specifically, the subgraph features in the sample are used as input, the error root cause classification of the sample is finally confirmed as the target label, the cross-entropy loss function is used as the optimization objective, and the weights of the hierarchical graph attention network are backpropagated to improve the accuracy of its misclassification. For a verified and marked as a fully successful repair case, its corresponding <root cause-context condition-repair action> triple is treated as a new knowledge unit. If the unit is not in the adaptive repair strategy library, it is added to the library; if it already exists, its selection probability is updated based on the near-end strategy optimization results.

[0024] Furthermore, the formulas mentioned above are all dimensionless calculations, derived from software simulation using a large amount of collected data to approximate the real situation. The weighting coefficients and preset thresholds in the formulas are set by those skilled in the art based on actual conditions or obtained through large-scale data simulation. The magnitude of the weighting coefficient is a specific value obtained by quantifying each parameter to facilitate subsequent comparison. The magnitude of the weighting coefficient depends on the amount of sample data and the processing coefficients initially set by those skilled in the art for each set of sample data. As long as it does not affect the proportional relationship between the parameter and the quantified value, it is acceptable.

[0025] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. The focus of each embodiment is on its differences from other embodiments. In particular, the apparatus embodiments are described simply because they are fundamentally based on the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0026] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0027] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0028] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0031] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligently correcting reconciliation errors in a financial management system, characterized by: S1. Acquire multi-source heterogeneous financial data, perform spatiotemporal alignment and structuring processing, and generate a standardized set of reconciliation nodes; S2. Based on the multi-dimensional attributes of reconciliation nodes, establish a time-enhanced reconciliation relationship graph and identify suspected error node clusters through dynamic weight calculation; S3. A hierarchical graph attention network is used to perform deep feature extraction and error root cause classification on clusters of suspected error nodes, and the classification confidence and error root cause type are output. S4. Based on the error root cause classification results, the optimal repair action sequence is matched from the adaptive repair strategy library, and a virtual simulation module is introduced to preview the repair consequences. S5. Utilize a multi-head attention mechanism to dynamically prioritize the optimal repair action sequence and generate the final repair execution plan; S6. Execute the final repair plan and verify the repair effect based on the dual indicators of reconciliation consistency index increment and business rule compliance. S7. Based on the repair results, iteratively optimize the adaptive repair strategy library and the hierarchical graph attention network model through reinforcement learning mechanism.

2. The intelligent repair method for reconciliation errors in a financial management system according to claim 1, characterized in that: In step S1, the process of acquiring multi-source heterogeneous financial data, performing spatiotemporal alignment and structuring processing, and generating a standardized reconciliation node set includes: Obtain multi-source heterogeneous financial data from bank interfaces, core business systems, and third-party payment platform databases. This multi-source heterogeneous financial data includes transaction flow data, account balance data, and historical reconciliation log data. Cleaning and standardization of the acquired multi-source heterogeneous financial data; Based on at least one association key, the multi-source heterogeneous financial data after cleaning is fused and associated to identify data records belonging to the same business entity. Using a single transaction or a logically related set of transactions formed through fusion and association operations as the basic unit, reconciliation nodes are divided, and a globally unique identifier and attribute set are generated for each reconciliation node, thereby forming a standardized set of reconciliation nodes. The attribute information of the reconciliation node includes at least: node ID, transaction amount, currency, lending direction, transaction timestamp, local account number, counterparty account number, transaction summary text, and spatial location index in the reconciliation relationship graph.

3. The intelligent repair method for reconciliation errors in a financial management system according to claim 2, characterized in that: In step S2, the process of establishing a time-enhanced reconciliation graph and identifying suspected error node clusters through dynamic weight calculation includes: An initial graph structure is constructed using each reconciliation node in the standardized reconciliation node set as a vertex; the feature vector of each vertex consists of its attribute information, including at least numerical features and textual features. Based on the attribute information of reconciliation nodes, various types of edges are established between vertices to create a time-enhanced reconciliation relationship graph; among them, the edge types include at least: fund flow edges, time-related edges, and semantically similar edges; For each edge in the graph, a comprehensive dynamic weight value is calculated. For multiple edges of different types connecting the same pair of vertices, their corresponding similarities are weighted and fused to form the final weight of the comprehensive edge between the vertex pair. For edges of only a single type, the similarity of that edge is the final weight. Based on dynamic weight values, the weighted SCAN graph clustering algorithm is used to cluster the vertices in the reconciliation graph; the set partitioning of vertices is completed based on the structural similarity of vertices and edge weights; and the set of vertices that meet the preset conditions is identified as a cluster of suspected erroneous nodes.

4. The intelligent repair method for reconciliation errors in a financial management system according to claim 3, characterized in that: In step S3, the process of using a hierarchical graph attention network to perform deep feature extraction and error root cause classification on clusters of suspected erroneous nodes, and outputting classification confidence and error root cause type, includes: Hierarchical graph attention networks consist of node-level attention layers and graph-level attention layers; From the time-enhanced reconciliation graph, extract all vertices and edges within each cluster of suspected erroneous nodes and its K-hop neighborhood to form multiple subgraphs to be analyzed. For each subgraph, a deep feature representation of each node is learned through a node-level attention layer; a global representation of the entire subgraph is aggregated through a graph-level attention layer. The global representation vector of each subgraph is input into a fully connected neural network classifier, which outputs the raw scores corresponding to each error root cause category. The raw scores are converted into a probability distribution by the Softmax function, thereby obtaining the classification confidence of the suspected error node cluster to each category and the final error root cause type. At the same time, the confidence of each classification result is output.

5. The intelligent repair method for reconciliation errors in a financial management system according to claim 4, characterized in that: In step S4, the process of matching the optimal repair action sequence from the adaptive repair strategy library and introducing the virtual simulation module to pre-simulate the repair consequences includes: Based on the output error root cause classification results and the context attributes of suspected error node clusters, retrieve all matching <error root cause-context condition-repair action> triples from the adaptive repair strategy library; The retrieved repair actions are initially sorted based on their historical execution success rates to form a candidate set of repair actions; The current time-series enhanced reconciliation graph state is replicated as the basis of the virtual simulation environment. Within this environment, an independent simulation branch is created for each candidate repair action sequence in the repair action candidate set. The corresponding repair action sequence is executed in each simulation branch, and evaluation metrics are calculated. These metrics include the estimated repair time (RT), the impact range index (NI) on upstream and downstream nodes, and the business rule conflict probability (BP). Based on historical log data, the average time required to execute all actions in the sequence is calculated to obtain the estimated repair time (RT). After executing the repair actions, the impact range index (NI) on upstream and downstream nodes is obtained by calculating the ratio of the number of nodes whose state has changed to the total number of nodes in the virtual simulation environment. A node refers to a cluster of suspected erroneous nodes in the reconciliation graph and all nodes within its first-order neighborhood. The repaired reconciliation graph is verified using a rule engine. The ratio of the number of conflicting rules to the total number of verified rules is the business rule conflict probability (BP). The rule engine is used to verify whether the repaired virtual reconciliation graph violates predefined business rules. The comprehensive evaluation score for each candidate sequence is calculated using CAS. The repair sequence with the highest comprehensive evaluation score (CAS) is selected as the optimal repair sequence.

6. The intelligent repair method for reconciliation errors in a financial management system according to claim 5, characterized in that: In step S5, the process of dynamically prioritizing the optimal repair action sequence using a multi-head attention mechanism to generate the final repair execution plan includes: Each action in the optimal repair action sequence is represented as an action feature vector; Arrange the action feature vectors of the entire optimal repair action sequence in the initial order to form a sequence matrix; and generate a position encoding vector for each position in the sequence. Add the position encoding vector to the corresponding action feature vector to inject the order information of the action in the sequence. The sequence matrix after injecting position information is input into a multi-head attention network for calculation; the operation analysis is performed on each attention head in the multi-head attention network; Each action's context-enhanced action feature vector, after being enhanced by an attention mechanism, is mapped to a scalar action priority score through a fully connected layer. The scalar action priority score is then weighted and combined with the system's real-time load factor to generate the final execution priority of each action. The optimal repair action sequence is then rearranged in descending order of final execution priority to generate the final repair execution plan.

7. The intelligent repair method for reconciliation errors in a financial management system according to claim 6, characterized in that: In step S6, the final repair execution plan is executed, and the repair effect is verified based on the dual indicators of reconciliation consistency index increment and business rule compliance. This process includes: In the virtual simulation environment, the final repair execution plan is executed to obtain the repaired reconciliation graph state; based on this state, the node matching logic used when constructing the time-series enhanced reconciliation graph is re-executed to calculate the successfully matched node pairs among all reconciliation nodes; After executing the final repair plan, calculate the post-repair reconciliation consistency index. Its value is the ratio of the number of successfully matched node pairs to the total number of reconciliation nodes; calculate the reconciliation consistency index increment ΔC; The repaired reconciliation relationship diagram status is re-entered into the rule engine for verification; the business rule compliance CR is defined as the ratio of the number of rules that pass verification to the total number of rules in the rule engine; A comprehensive judgment is made based on the reconciliation consistency index increment ΔC and the business rule compliance CR. Based on the judgment results, the repair case, along with its input features, the sequence of repair actions performed, and the final (ΔC,CR) result, is stored in the training set as a sample with a clear result label.

8. The intelligent repair method for reconciliation errors in a financial management system according to claim 7, characterized in that: In step S7, the process of iteratively optimizing the adaptive repair strategy library and the hierarchical graph attention network model through reinforcement learning based on the repair effect includes: Each repair task is abstracted into a reinforcement learning environment state. This state is characterized by the subgraph features of the suspected faulty node cluster and the contextual features of the repair action sequence. The adaptive repair policy library can be viewed as a policy function π of a reinforcement learning agent, where the policy function depends on the environment state. Output repair action The combined result of the reconciliation consistency index increment ΔC and the business rule compliance CR is quantified into a reward signal. ; The parameters of the policy function π are updated using a proximal policy optimization algorithm with the goal of maximizing cumulative reward. The model update cycle is preset; samples stored in the training set are periodically used as labeled supervised learning data to fine-tune the hierarchical graph attention network. For a verified and marked as a fully successful repair case, its corresponding <root cause-context condition-repair action> triple is treated as a new knowledge unit. If the unit is not in the adaptive repair strategy library, it is added to the library; if it already exists, its selection probability is updated based on the near-end strategy optimization results.

Images