Financial audit compliance system based on multi-dimensional rule engine

Abstract

The invention discloses a financial auditing compliance system based on a multi-dimensional rule engine, and the system comprises a collection and construction module which is used for collecting multi-source business data, and generating a transaction account structure causal graph; the rule compiling module is used for generating a rule vector, a hit sequence and an exemption mask; the compliance reasoning module is used for outputting state representation by using a graph frequent differential compliance model; the risk assessment module is used for constructing compliance feature vectors, generating dynamic confidence coefficients and grading the dynamic confidence coefficients; the compliance optimization module is used for generating a minimum modification rule and instruction to form a modified draft; the rechecking and evidence storage module is used for rechecking the correction record, recompiling reasoning and storing the zero-knowledge proof; and the updating and displaying module is used for outputting and displaying the risk grade, the correction rule and the certification identifier. The invention also discloses a method. According to the method, the multi-dimensional rule engine is combined with the Scheng frequent differential model, so that intelligent compliance auditing and automatic correction of financial transactions are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial technology and intelligent compliance audit, and particularly relates to a financial audit compliance system based on a multi-dimensional rule engine. BACKGROUND

[0002] The existing financial business compliance audit technology mainly relies on rule engines and batch processing analysis methods, and matches and compares transaction information, account information, merchant information, etc. through pre-configured regulatory rules and compliance rules to discover potential violations. However, this kind of method usually adopts an offline computing mode, and the rule execution process is not differentiable, which leads to the fact that rule optimization and model training cannot form a unified optimization closed loop. The existing technology usually adopts manual coding or simple logic analysis in the rule analysis process, and has limited structured processing capability for regulatory provisions and internal rules, and is difficult to meet the regulatory requirements of cross-domain, multi-dimensional and dynamic changes. Although the traditional graph analysis method can construct a network structure for detection by using transaction relationships, it has deficiencies in time continuity modeling, and cannot effectively capture the complex patterns of transaction behavior evolution over time, resulting in insufficient accuracy and real-time performance of abnormality identification.

[0003] In terms of rule execution and modeling, the existing technology generally lacks a differentiable rule compilation mechanism, so that the rule satisfaction degree cannot directly participate in the training and reasoning process of the compliance model, and it is difficult to realize the deep integration of rule constraints and data-driven modeling. The abnormal score mechanism usually adopts a static threshold, and cannot combine uncertainty evaluation and dynamic confidence strategy for hierarchical judgment, which is easy to cause misjudgment and omission. For the correction and compliance approximation of high-risk transactions, the existing technology usually relies on manual intervention or fixed rule adjustment, lacks optimal counterfactual solution and automatic rule generation mechanism in the business feasible domain, and is difficult to ensure compliance while considering business feasibility.

[0004] Therefore, how to provide a financial audit compliance system based on a multi-dimensional rule engine is a problem that those skilled in the art need to solve. SUMMARY

[0005] One purpose of the present application is to propose a financial audit compliance system based on a multi-dimensional rule engine, which integrates multiple technical links such as multi-source business data preprocessing, differentiable regulatory rule compilation, transaction account structure causal graph construction, dynamic confidence evaluation, counterfactual compliance modification, etc. The whole process from the structured analysis of regulatory provisions and internal compliance rules, to the generation of rule satisfaction and dissatisfaction vectors, to the joint inference and abnormal severity scoring in the graph neural differential equation model is described in detail, realizing the deep integration of regulatory rules and data-driven models. The present application has the advantages of high rule analysis accuracy, high abnormality identification accuracy, high compliance modification automation degree and strong real-time response capability, and can ensure regulatory compliance while considering business feasibility and audit efficiency.

[0006] According to the financial audit compliance system based on the multi-dimensional rule engine of the embodiment of the present application, the following modules are included: The acquisition and construction module is used for acquiring and preprocessing multi-source business data, generating and updating the transaction account structure causal graph; The rule compilation module is used for analyzing regulatory provisions and internal compliance rules, generating rule satisfaction vectors, rule dissatisfaction vectors and rule hit sequences, and outputting exemption condition masks; The compliance reasoning module is used for inputting the transaction account structure causal graph, the rule satisfaction vector and the rule dissatisfaction vector into the graph neural differential equation model, and outputting a continuous time state representation; The risk assessment module is used for constructing a joint compliance feature vector, performing uncertainty evaluation to generate a dynamic confidence coefficient, calculating an abnormal severity score and completing risk classification; The compliance optimization module is used for generating a minimum modified modified rule and a modified instruction within a business feasible domain and a constraint set, and forming a modified transaction record draft; The review and storage module is used for reviewing the modified transaction record, re-executing the differentiable regulatory rule compilation and reasoning, and generating and storing a zero-knowledge proof after confirming compliance; The update and display module is used for outputting the risk classification result, the modified rule and the zero-knowledge proof identifier, and displaying the related results and evidence chain information through the interface.

[0007] Optionally, the modules are realized through the following methods: S1, multi-source business data is collected, and the multi-source business data is preprocessed to construct a transaction account structure causal graph; S2, regulatory provisions and internal compliance rules are analyzed, differentiable regulatory rule compilation is performed, rule satisfaction vectors and rule dissatisfaction vectors are generated, and rule hit sequences are recorded; S3, input the structural features of the transaction account structure causal graph, the rule satisfaction degree vector and the rule non-satisfaction degree vector into a graph neural differential equation model for joint training and reasoning to obtain a continuous time state representation; S4, based on the continuous time state representation, the rule satisfaction degree vector and the rule non-satisfaction degree vector, a joint compliance feature vector is constructed, an uncertainty evaluation is performed on the joint compliance feature vector to obtain a dynamic confidence coefficient, an abnormal severity score is calculated and graded; S5, when the abnormal severity score exceeds a corresponding threshold value, a counterfactual compliance simulator is triggered, a minimum modification scheme is solved based on the transaction account structure causal graph, the rule satisfaction degree vector and the rule non-satisfaction degree vector within a business feasible region, and a revised rule and a revision instruction are generated; S6, the revision instruction is applied to the corresponding transaction record of the transaction account structure causal graph to generate a revised transaction record, and the joint reasoning of the differentiable regulatory rule compilation and the graph neural differential equation model is re-executed to confirm that all rule satisfaction degree vectors and grading thresholds are met, and zero-knowledge proofs are generated and stored for interval and threshold rules; S7, output the risk grading result, the revised rule and the zero-knowledge proof identifier, and update the differentiable regulatory rule compilation parameters, the graph neural differential equation model parameters, the dynamic confidence coefficient strategy and the abnormal severity score grading threshold according to the audit feedback.

[0008] Optionally, the multi-source business data specifically includes transactions, accounts, devices, merchants, channels, geographic locations and policy events.

[0009] Optionally, the preprocessing of the multi-source business data specifically includes time alignment, currency unification, missing value filling and outlier processing.

[0010] Optionally, the construction of the transaction account structure causal graph specifically refers to setting transactions, accounts, devices, merchants, channels, geographic locations and policy events as nodes after preprocessing the multi-source business data, establishing directed causal edges according to transaction fund flow and exogenous events, configuring static attributes and time-varying attributes for nodes and edges, and generating a traceable causal graph structure by time window increment.

[0011] Optionally, S2 specifically includes: S21, structurally analyze the regulatory provisions and internal compliance rules, extract subject elements, behavior elements and constraint elements, label threshold conditions, logical connection relationships, time windows and exemption conditions, generate a differentiable compliance intermediate representation and simultaneously generate an exemption condition mask; S22, based on the differentiable compliance intermediate representation, mapping the threshold class, the ratio class and the count class constraints therein into differentiable satisfaction functions, using monotonic differentiable smoothing mapping to convert the feature and threshold gap, the numerator and denominator ratio deviation, and the cumulative count and threshold gap within the time window into satisfaction values between zero and one, generating an initial rule satisfaction vector; S23, performing differentiable aggregation on the entries in the initial rule satisfaction vector that involve time windows and timing constraints, performing weighted average on the satisfaction values at each time in the window according to weights decreasing with time intervals, obtaining a time constraint aggregated satisfaction vector; S24, differentiably combining the time constraint aggregated satisfaction vector and the satisfaction values not involving time constraints according to the logical relationships in the compliance intermediate representation, using continuous logical operators to aggregate the logical relationships including and, or, implication and exemption, generating a logical combination satisfaction vector; S25, calculating a weighted coefficient for each rule in the logical combination satisfaction vector according to data integrity, freshness and source credibility, and limiting it to the range of zero to one through a monotonic differentiable mapping, multiplying the rule satisfaction value and the weighted coefficient to generate a rule satisfaction vector, and subtracting the corresponding weighted value from the maximum value of the rule satisfaction to generate a rule non-satisfaction vector; S26, generating a continuous hit indication score for the rule satisfaction vector according to the preset hierarchical threshold, sorting from high to low to form a rule hit sequence, and storing the rule satisfaction vector, the rule non-satisfaction vector and the rule hit sequence.

[0012] Optionally, the S3 specifically comprises: S31, extracting time-ordered node features, edge features and global features from the transaction account structure causal graph, organizing them into a structure feature sequence according to time order, and merging the rule satisfaction vector and the rule non-satisfaction vector into a rule feature sequence in order; S32, taking the structure feature sequence and the rule feature sequence as input, establishing the initial state of the graph neural differential aggregation model, and fixing the initial time point and the solution interval; S33, performing double-channel constraint injection in the graph neural differential aggregation model, injecting the rule satisfaction vector into the node state update channel and the rule non-satisfaction vector into the edge state update channel, and disabling the exempted constraints in both channels according to the exemption condition mask; S34, adaptively scheduling the solution step length according to the risk indication in the rule non-satisfaction vector and the structure feature sequence, setting a smaller step length for time slices with high non-satisfaction degree or drastic structure changes, and setting a larger step length for time slices with low non-satisfaction degree and stable structure, generating a continuous time state trajectory; S35, set up structural causal intervention hooks in the continuous-time state trajectory solving process, perform intervention on specified time slices based on exogenous events and policy events in the transaction account structural causal graph, and record the continuous-time state representation after intervention; S36, perform constraint consistency projection on the continuous-time state representation of each time slice, preferentially satisfy mandatory constraints, sequentially process mutually conflicting constraints according to priority tables, and minimize deviation from constraints that have been satisfied in the previous time slice, to obtain a consistent continuous-time state representation sequence; S37, jointly train and infer the consistent continuous-time state representation sequence and the rule feature sequence, and output the continuous-time state representation of the target time slice.

[0013] Optionally, the S4 specifically comprises: S41, align the continuous-time state representation with the rule satisfaction degree vector and the rule non-satisfaction degree vector one by one according to the indexes of nodes and edges, form a joint compliance feature vector, and retain the time stamp and causal path identifier; S42, perform multi-dimensional uncertainty evaluation on the joint compliance feature vector, generate uncertainty measures based on graph neural differential equation model perturbation results, data integrity and freshness indicators, and rule stability indicators within a time window, and aggregate them into a dynamic confidence coefficient set; S43, run the counterfactual generative simulator in evaluation mode, apply the minimum feasible shadow modification to the joint compliance feature vector and observe the output change, obtain the counterfactual sensitivity factor, calibrate and fuse it with the dynamic confidence coefficient set to generate a dynamic confidence coefficient vector; S44, use the dynamic confidence coefficient vector as a weight to perform path aggregation on the rule non-satisfaction degree on the causal path of the transaction account structural causal graph, generate node-level, edge-level and global abnormal severity scores, and set the minimum level requirement for risk classification for entries marked as mandatory constraints; S45, adaptively adjust the classification threshold according to the historical false positive rate, audit processing capacity, regulatory priority and business time period load, map the abnormal severity score to the risk classification label, and store the dynamic confidence coefficient vector, the abnormal severity score and the risk classification label.

[0014] Optionally, the S5 specifically comprises: S51, extract the joint compliance feature vector, the transaction account structural causal graph local subgraph, the rule satisfaction degree vector and the rule non-satisfaction degree vector of the abnormal severity score threshold object; S52, construct the business feasible region and the invariable constraint, clear the adjustable item as the amount, the frequency, the time interval, the channel selection and the identity check field, set the upper and lower limit, the step size and the value set for each adjustable item, and convert the exemption condition mask into the disabled constraint on the object, form the constraint set; S53, locate the causal path in the time window in the local subgraph of the transaction account structure causal graph, group the affected rules according to the rule dissatisfaction vector and determine the modification priority order, generate the modification order table, and construct the candidate change with the smallest amplitude on the causal path in sequence; S54, perform multi-stage minimum modification solving, carry out the feasibility stage to obtain a candidate scheme that satisfies all mandatory rules, carry out the business cost convergence stage to select a candidate scheme with fewer modification steps and lower business cost, and call the differentiable regulatory rule compilation and graph neural differential equation model in the virtual environment for rapid checking, eliminate the candidates that reappear rule dissatisfaction in the checking, and output the optimal modified rule and modification instruction; S55, generate the field-level change list, execution order and rollback plan from the optimal modified rule and modification instruction, and form the minimum modification proof record.

[0015] The beneficial effects of the present application are: The present application introduces the differentiable regulatory rule compilation technology, converts complex regulatory provisions and internal compliance rules into computable and optimized rule satisfaction and dissatisfaction vectors, and deeply integrates with the structural characteristics of the transaction account structure causal graph, uses the graph neural differential equation model to realize continuous time compliance state modeling and dynamic reasoning of transaction behavior, and significantly improves the intelligent and fine level of compliance audit.

[0016] Through dynamic confidence coefficient evaluation and abnormal severity grading, real-time quantification and graded response to potential violation risks are realized, effectively reducing the missed detection and misjudgment rate.

[0017] The present application introduces the counterfactual compliance simulator, which can automatically generate the minimum modification scheme in the business feasible region and complete the modification verification, ensuring that the compliance rectification not only meets the regulatory requirements, but also maintains business continuity and reasonableness. Compared with the prior art, the present application not only improves the accuracy and efficiency of regulatory rule execution, but also realizes the closed-loop processing of compliance checking, risk warning and automatic correction, with the significant advantages of fast response, strong adaptability and good scalability. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0019] Figure 1This is a schematic diagram of the structure of a financial audit compliance system based on a multi-dimensional rule engine proposed in this invention; Figure 2 This is a flowchart illustrating a financial audit compliance method based on a multi-dimensional rule engine proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 A financial audit compliance system based on a multi-dimensional rule engine, comprising the following modules: The data collection and construction module is used to collect multi-source business data and preprocess it to generate and update the causal graph of the transaction account structure. The rule compilation module is used to parse regulatory provisions and internal compliance rules, generate rule satisfaction vectors, rule non-satisfaction vectors and rule hit sequences, and output exemption condition masks. The compliance reasoning module is used to input the causal graph of the transaction account structure, the rule satisfaction vector, and the rule non-satisfaction vector into the graph-divine constant differential compliance model, and output a continuous-time state representation. The risk assessment module is used to construct a joint compliance feature vector, perform uncertainty assessment to generate dynamic confidence coefficients, calculate anomaly severity scores, and complete risk classification. The compliance optimization module is used to generate minimum modification rules and instructions within the business feasible domain and constraint set, and to form a draft of the modified transaction record; The verification and evidence storage module is used to verify and correct transaction records, re-execute the compilation and reasoning of differentiable regulatory rules, and generate and store zero-knowledge proofs after confirming compliance. The updated display module outputs risk classification results, corrected rules, and zero-knowledge proof identifiers, while also displaying relevant results and evidence chain information through an interface.

[0022] refer to Figure 2 A financial audit compliance method based on a multi-dimensional rule engine includes the following steps: S1. Collect multi-source business data, preprocess the multi-source business data, and construct a cause-effect graph of the transaction account structure; S2. Parse regulatory provisions and internal compliance details, execute the compilation of differentiable regulatory rules, generate rule satisfaction vectors and rule non-satisfaction vectors, and record the rule hit sequence; S3. Input the structural features, rule satisfaction vector, and rule non-satisfaction vector of the causal graph of the transaction account structure into the graph god frequent differential compliance model for joint training and inference to obtain a continuous-time state representation. S4, constructing a joint compliance feature vector based on the continuous-time state representation, the rule satisfaction vector and the rule dissatisfaction vector, performing uncertainty evaluation on the joint compliance feature vector to obtain a dynamic confidence coefficient, calculating an abnormal severity score and grading; S5, when the abnormal severity score exceeds a corresponding threshold value, triggering a counterfactual rule synthesizer, solving a minimum modification scheme based on the transaction account structure causal graph, the rule satisfaction vector and the rule dissatisfaction vector within a business feasible region, generating a revised rule and a revision instruction; S6, applying the revision instruction to the corresponding transaction record of the transaction account structure causal graph to generate a revised transaction record, re-executing the differentiable regulatory rule compilation and the graph neural differential rule model joint inference, confirming that all rule satisfaction vectors and grading thresholds are satisfied, and generating and storing zero-knowledge proofs for interval and threshold rules; S7, outputting the risk grading result, the revised rule and the zero-knowledge proof identifier, and updating the differentiable regulatory rule compilation parameters, the graph neural differential rule model parameters, the dynamic confidence coefficient strategy and the abnormal severity score grading threshold according to the audit feedback.

[0023] In the embodiment, the multi-source business data specifically includes transactions, accounts, devices, merchants, channels, geographic locations and policy events.

[0024] In the embodiment, the preprocessing of the multi-source business data specifically includes time alignment, currency unification, missing value filling and outlier processing.

[0025] In the embodiment, the construction of the transaction account structure causal graph specifically refers to setting transactions, accounts, devices, merchants, channels, geographic locations and policy events as nodes after preprocessing the multi-source business data, establishing directed causal edges according to transaction fund flow and exogenous events, configuring static attributes and time-varying attributes for nodes and edges, and generating a traceable causal graph structure by time window increment.

[0026] In the embodiment, S2 specifically includes: S21, structurally analyzing the regulatory provisions and internal compliance rules, extracting subject elements, behavior elements and constraint elements, labeling threshold conditions, logical connection relationships, time windows and exemption conditions, generating a differentiable compliance intermediate representation and synchronously generating an exemption condition mask; S22, based on the differentiable compliance intermediate representation, mapping the threshold class, the ratio class and the count class constraints therein into differentiable compliance functions, using monotonic differentiable smoothing mapping to convert the gap between the feature and the threshold, the ratio deviation between the numerator and the denominator, and the cumulative count within the time window to the compliance value between zero and one, generating an initial rule compliance vector, wherein the differentiable compliance function refers to a function constructed for the compliance degree between the regulatory rule provisions and the transaction data, the output value of which continuously changes between 0 and 1, wherein 0 represents complete non-compliance and 1 represents complete compliance, and the differentiable compliance function has differentiability; S23, performing differentiable aggregation on the items in the initial rule compliance vector involving time windows and time sequence constraints, and performing weighted average on the compliance values at each time within the window according to the weights decreasing with time intervals to obtain a time constraint aggregated compliance vector; S24, performing differentiable combination on the time constraint aggregated compliance vector and the compliance values not involving time constraints according to the logical relationships in the compliance intermediate representation, and performing aggregation on the logical relationships including and, or, implication and exemption using continuous logical operators to generate a logical combination compliance vector; S25, calculating a weighted coefficient for each rule in the logical combination compliance vector according to data integrity, freshness and source credibility, and limiting the weighted coefficient to the range of zero to one through monotonic differentiable mapping, multiplying the rule compliance value by the weighted coefficient to generate a rule compliance vector, and subtracting the corresponding weighted value from the maximum value of the rule compliance to generate a rule non-compliance vector, wherein the data integrity refers to the completeness of the information content of the multi-source business data in the process of collection and transmission, including whether the field is missing and whether the value is truncated or tampered; the freshness refers to the closeness of the multi-source business data to the collection time, which reflects the timeliness of the data; and the source credibility refers to the reliability of the multi-source business data collection source in terms of historical records, identity verification and secure transmission; S26, generating a continuous hit indication score for the rule compliance vector according to a preset hierarchical threshold, sorting from high to low to form a rule hit sequence, and storing the rule compliance vector, the rule non-compliance vector and the rule hit sequence, wherein the generation of the continuous hit indication score for the rule compliance vector according to the preset hierarchical threshold refers to comparing each rule compliance value in the rule compliance vector with the corresponding hierarchical threshold, and mapping the difference value relationship into a continuous value between zero and one according to the difference value relationship, which reflects the relative strength of the rule hit degree, so as to form a sortable hit indication result between different rules.

[0027] In the embodiment, the S3 specifically includes: S31, extract time-ordered node features, edge features and global features from the transaction account structure causal graph, organize them into a structure feature sequence in chronological order, and merge the rule satisfaction degree vector and the rule non-satisfaction degree vector into a rule feature sequence in order; S32, input the structure feature sequence and the rule feature sequence to establish the initial state of the graph neural differential equation model, fix the initial time point and solve the interval; S33, perform double-channel constraint injection in the graph neural differential equation model, inject the rule satisfaction degree vector into the node state update channel, inject the rule non-satisfaction degree vector into the edge state update channel, and disable the exempted constraints in the two channels according to the exemption condition mask; S34, adaptively schedule the solving step length according to the risk indication in the rule non-satisfaction degree vector and the structure feature sequence, set a smaller step length for time slices with high non-satisfaction degree or dramatic structure changes, and set a larger step length for time slices with low non-satisfaction degree and stable structure, and generate a continuous time state trajectory; S35, set a structure causal intervention hook in the continuous time state trajectory solving process, execute intervention on specified time slices based on exogenous events and policy events in the transaction account structure causal graph, and record the continuous time state representation after intervention, wherein the structure causal intervention hook is a pre-set insertable position marker in the continuous time state trajectory solving process, used to plant the defined exogenous events and policy events in the transaction account structure causal graph into the calculation process at a specific time slice, so that the graph neural differential equation model synchronously contains the intervention information when generating the state representation of the time slice, and the state result after intervention is retained in the state trajectory; S36, perform constraint consistency projection on the continuous time state representation of each time slice, preferentially satisfy mandatory constraints, sequentially process mutually conflicting constraints according to the priority table, and minimize the deviation from the constraints already satisfied in the previous time slice, to obtain a consistent continuous time state representation sequence, wherein the constraint consistency projection is a process of mapping the state representation into a feasible region that satisfies the given constraint conditions, so that it meets both mandatory constraints and optional constraints sorted by priority, and when there is a conflict, it is processed one by one according to the preset priority table, ensuring that the deviation from the constraints already satisfied in the previous time slice is minimized, thereby forming a consistent state representation result that satisfies all constraint requirements; S37, jointly train and infer the consistent continuous time state representation sequence and the rule feature sequence to output the continuous time state representation of the target time slice, specifically: Pair the consistent continuous time state representation sequence and the rule feature sequence of the corresponding time slice according to the time alignment method to form a time sequence input sample, and keep the structure correlation and sequence consistency of each time slice; The time sequence input sample is input into a graph neural differential compliance model, and features are extracted across time slices and rules by modeling dynamic state evolution in the time dimension and causal dependence in the structure dimension. Based on the time sequence feature result of the output, a feature vector corresponding to a target time slice is selected to generate a continuous time state representation of the time slice.

[0028] In this embodiment, the S4 specifically includes: S41, align the continuous time state representation with the rule satisfaction degree vector and the rule non-satisfaction degree vector according to the index of nodes and edges to form a joint compliance feature vector, and keep the timestamp and causal path identifier; S42, perform multi-dimensional uncertainty evaluation on the joint compliance feature vector, generate uncertainty metrics based on the graph neural differential compliance model perturbation result, data integrity and freshness indicators, and rule stability indicators within a time window, and aggregate them into a dynamic confidence coefficient set, specifically: (1) under the condition of keeping the structure of the joint compliance feature vector unchanged, slightly perturb the graph neural differential compliance model parameters and input, record the variance and distribution difference of multiple inference results, and generate model perturbation uncertainty metrics; (2) calculate the data integrity ratio and data freshness score of the data relied on by the joint compliance feature vector within the current analysis window, and combine them to form a data quality uncertainty metric; (3) statistics the hit frequency, satisfaction degree fluctuation range and trend stability indicators of the rule within the set time window, generate the rule stability uncertainty metric, and normalize and aggregate to form a dynamic confidence coefficient set; S43, run the counterfactual compliance simulator in evaluation mode, apply the minimum feasible shadow modification to the joint compliance feature vector and observe the output change to obtain the counterfactual sensitivity factor, calibrate and fuse with the dynamic confidence coefficient set to generate a dynamic confidence coefficient vector. The counterfactual compliance simulator refers to performing virtual parameter adjustment of a specific variable in the joint compliance feature vector under the condition of keeping the overall topology of the transaction account structure causal graph and the business feasibility domain constraint unchanged, constructing a "shadow transaction scenario" that does not occur in actual business but is logically possible, and inputting the virtual scenario into the compliance inference process. Compare the difference between the output results before and after adjustment to depict the sensitivity of each feature to the compliance result and generate the counterfactual sensitivity factor; S44, use the dynamic confidence coefficient vector as a weight to aggregate the rule non-satisfaction degree on the causal path of the transaction account structure causal graph, generate node-level, edge-level and global abnormal severity scores, and set the minimum level requirement of risk classification for the entries marked as mandatory constraints; S45, mapping the anomaly severity score to a risk classification label according to the historical false positive rate, audit processing capacity, regulatory priority and business time period load self-adaptive adjustment classification threshold, and storing the dynamic confidence coefficient vector, anomaly severity score and risk classification label.

[0029] In this embodiment, S5 specifically includes: S51, extracting the joint compliance feature vector of the anomaly severity score exceeding threshold object, the transaction account structure causal graph local subgraph, the rule satisfaction degree vector and the rule non-satisfaction degree vector; S52, constructing a business feasible region and an invariable constraint, specifying the adjustable items as amount, frequency, time interval, channel selection and identity verification field, setting upper and lower bounds, step size and value set for each adjustable item, and converting the exemption condition mask into a disable constraint on the object, forming a constraint set, wherein the business feasible region and the invariable constraint refer to the parameter value range and fixed condition set determined by business logic, regulatory provisions and system running conditions in a specific financial business scenario, wherein the business feasible region covers all legal parameter combination spaces that meet business logic and compliance requirements, and the invariable constraint refers to the core rules and conditions that must remain unchanged in any case; S53, locating the causal path in the local subgraph of the transaction account structure causal graph according to the time window, grouping the affected rules according to the rule non-satisfaction degree vector and determining the correction priority order, generating a correction order table, and constructing the candidate change with the smallest amplitude on the causal path in order; S54, perform multi-stage minimum modification solving, obtain candidate schemes satisfying all mandatory rules in the feasibility stage, select candidate schemes with fewer modification steps and lower business cost in the business cost convergence stage, and call the differentiable regulatory rule compilation and graph neural differential model in the virtual environment for rapid checking, eliminate candidates that do not meet the rules again in the checking, and output the optimal correction rule and correction instruction, wherein obtaining candidate schemes satisfying all mandatory rules in the feasibility stage specifically refers to inputting all candidate modification schemes generated based on the business feasible domain and the constraint set into the differentiable regulatory rule compilation process one by one in the initial stage of the minimum modification solving process, detecting the satisfaction degree of each rule, and only when the candidate scheme meets the preset satisfaction degree threshold on all mandatory rules, it is marked as a qualified scheme and retained. The candidate scheme that does not meet the requirement is directly eliminated to form a candidate scheme set that satisfies all mandatory rules, wherein selecting candidate schemes with fewer modification steps and lower business cost in the business cost convergence stage specifically refers to selecting a scheme with fewer modification steps as a candidate based on the modification step and business cost indicators corresponding to each candidate scheme in the candidate scheme set that has passed the feasibility stage, and then comparing the business cost values in the candidate schemes to select a scheme with lower business cost as a priority reservation scheme, until the final candidate scheme set that meets the business cost minimization target is converged; S55, generate a field-level change list, an execution order, and a rollback plan based on the optimal correction rule and the correction instruction, and form a minimum modification proof record.

[0030] In the embodiment, the S6 specifically includes: S61, apply the correction instruction to the transaction account structure causal diagram corresponding to the transaction record, update the related node attributes and edge attributes, generate the corrected transaction record, and record the correction timestamp and the correction range; S62, re-execute the joint inference of the differentiable regulatory rule compilation and the graph neural differential model based on the corrected transaction record to obtain a new rule satisfaction degree vector, a rule dissatisfaction degree vector, a rule hit sequence, and a continuous time state representation, and check whether the new rule satisfaction degree vector satisfies all hierarchical threshold conditions; S63, generate a zero-knowledge proof for the interval class and threshold class rules involved in the corrected transaction record, record the zero-knowledge proof identifier, the zero-knowledge proof abstract, and the associated rule number, and store the zero-knowledge proof together with the corrected transaction record, the rule satisfaction degree vector, and the hierarchical threshold comparison result.

[0031] In the embodiment, the S7 specifically includes: S71, obtain the risk grading result, the correction rule, and the zero-knowledge proof identifier, output the risk grading result, the correction rule, and the zero-knowledge proof identifier to the visualized audit terminal and the regulatory platform, and perform data archiving. S72, receive audit feedback data, update the related parameters of the differentiable regulatory rule compilation according to the audit feedback content; S73, adjust the dynamic confidence coefficient strategy based on the audit feedback content, optimize the grading threshold of the abnormal severity score, and improve the accuracy and rationality of risk grading; S74, update the parameter configuration of the graph neural differential equation model in combination with the audit feedback, write the updated differentiable regulatory rule compilation parameters, graph neural differential equation model parameters, dynamic confidence coefficient strategy and abnormal severity score grading threshold into the parameter library, and push them to the business front end and the regulatory platform.

[0032] Embodiment 1: In order to verify the feasibility of the application in implementation, the application is applied to a certain commercial bank. The commercial bank has long used a compliance audit mode of "static rules + post-sampling" in cross-border payment and foreign currency settlement business. There are problems such as lag in updating regulatory rules, independent operation of rules and detection models, heavy pressure on manual review, and coexistence of false positives and false negatives. To solve the above problems, the bank has online the method and system of the application on the data center, and the complex regulatory provisions and internal risk control rules are compiled into differentiable regulatory rules to convert them into computable rule satisfaction and dissatisfaction vectors. After time synchronization, currency conversion and abnormality cleaning of multi-source business data, a transaction account structure causal graph composed of accounts, devices, merchants, transaction channels, geographic units and policy events is constructed. Based on this, the graph neural differential equation model is used to infer the risk state of transaction behavior in continuous time, and generate joint compliance features for uncertainty assessment, dynamically calculate the confidence coefficient, and then output the processing result according to the abnormal severity grading. For transactions exceeding the threshold, the counterfactual model simulator generates the minimum modified correction rules and execution instructions within the business feasible domain. For transactions that meet the interval type or threshold type rules, zero-knowledge compliance proof is automatically generated for internal audit and cross-institution review without exposing business details.

[0033] The system is deployed in the private cloud cluster of the bank's data center, and accesses the payment core system, channel convergence system, customer identity recognition and list checking system, cross-border clearing system and risk control label database. The business scope covers multiple domestic and international clearing channels, and loads nearly two years of historical transaction data and real-time incremental stream data. In the joint debugging stage, the system collects transaction streams through bypass mirroring for playback verification to ensure that the differentiable compilation results of the rule template are consistent with the historical manual audit conclusions; after verification is stable, switch to real-time online processing mode. The rules monitored by the system include: high-frequency cross-border payment in a certain time window, joint restriction of single and cumulative amount, and detection of split transaction mode around accounts and devices, and the corresponding zero-knowledge compliance evidence is generated synchronously.

[0034] In actual operation, when a customer transfers out foreign currency in multiple channels to the same region's associated account in several times within a short period of time, the traditional static rules may miss the report due to the single amount being lower than the threshold. The method of the present application models the multi-hop relationship and time interval closeness of the customer's transaction path in the transaction account structure causal graph, and the dissatisfaction accumulates rapidly in continuous time, the dynamic confidence coefficient decreases and triggers a high-risk alarm. The system then generates a minimum modification suggestion through the counterfactual compliance simulator: for example, merging part of the transactions, adjusting the time interval, supplementing the beneficiary identity field, or replacing the transaction channel. After online confirmation by business personnel, the correction instruction is reviewed in the sandbox and written back to the business system, so that the transaction meets all regulatory constraints without affecting the flow of funds. Similar compliance correction rules also apply to night high-frequency operations, cross-currency rapid switching, and multi-device concurrent operations. These scenarios, which previously required extensive manual log comparison to discover, can now be automatically identified and corrected by the system.

[0035] After a complete settlement cycle of comparative evaluation, the results show that the false positive rate has dropped from the original two digits to the middle and low digits, the false negative rate has decreased significantly, the accuracy of high-risk identification has improved significantly, and the rule hit sequence, abnormal path and minimum modification proof output by the system can be directly used as audit and regulatory response materials. Zero-knowledge compliance evidence achieves verifiable sampling of cumulative amount and frequency rules, meeting audit needs while avoiding sensitive data leakage. For front-line business personnel, field-level correction rules and rollback plans generated by the system significantly reduce manual trial and error costs and communication rounds; for technical operation and maintenance, the linkage training of micro-scale rules and graph neural differential scale models reduces the burden of dual-track maintenance of rules and models, and version switching can be smoothly upgraded at the parameter and strategy level.

[0036] Table 1 Comparison of key indicators before and after the cross-border payment compliance audit goes online

[0037] As can be seen from the data in Table 1, the overall performance of the application after going online is obviously improved compared with that before going online. Before going online, the system processed an average of about 1,020,000 transactions per week, with an average processing delay of 215 milliseconds, 9,850 high-risk transactions, 8,920 of which were manually confirmed, a false positive rate of 9.4%, a false negative rate of 3.2%, and no zero-knowledge proof generation and counterfactual correction functions. In the first week after going online, the total number of transactions increased to 1,065,432, the average processing delay decreased to 178 milliseconds, the number of high-risk transactions was 10,112, of which 9,533 were manually confirmed, the false positive rate decreased to 5.7%, the false negative rate decreased to 2.5%, and 9,684 zero-knowledge proofs were newly generated, with a counterfactual correction success rate of 73%. During the second week of operation, the total number of transactions was 1,012,089, the average processing delay further decreased to 171 milliseconds, the number of high-risk transactions was 9,406, the number of manually confirmed high-risk transactions was 8,931, the false positive rate decreased to 5.1%, the false negative rate decreased to 2.3%, 9,217 zero-knowledge proofs were generated, and the counterfactual correction success rate increased to 75%. The third week of operation data performed best, with a total of 1,128,547 transactions, an average processing delay of 169 milliseconds, 10,885 high-risk transactions, 10,327 manually confirmed high-risk transactions, a false positive rate of 5.1%, a false negative rate of 2.1%, 10,104 zero-knowledge proofs generated, and a counterfactual correction success rate of 77%.

[0038] Overall, the application is significantly superior to the pre-online system in terms of processing efficiency, risk identification accuracy, and compliance proof generation, and the performance shows a stable and optimized trend.

[0039] The above describes only the preferred embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A multi-dimensional rule engine based financial audit compliance system characterized in that, The method comprises the following steps: S1, collecting multi-source business data, preprocessing the multi-source business data, and constructing a transaction account structure causal graph; S2, analyzing regulatory provisions and internal compliance rules, performing differentiable regulatory rule compilation, generating rule satisfaction vectors and rule non-satisfaction vectors, and recording rule hit sequences; S3, inputting the structural features of the transaction account structure causal graph, the rule satisfaction vectors and the rule non-satisfaction vectors into a graph neural differential compliance model for joint training and reasoning to obtain a continuous time state representation; S4, constructing a joint compliance feature vector based on the continuous time state representation, the rule satisfaction vectors and the rule non-satisfaction vectors, performing uncertainty evaluation on the joint compliance feature vector to obtain a dynamic confidence coefficient, calculating an abnormal severity score and grading; S5, when the abnormal severity score exceeds the corresponding threshold, triggering a counterfactual compliance model, solving a minimum modification scheme based on the transaction account structure causal graph, the rule satisfaction vectors and the rule non-satisfaction vectors within the business feasible region, generating modified rules and modified instructions; S6, applying the modified instructions to the corresponding transaction records of the transaction account structure causal graph to generate modified transaction records, re-executing the differentiable regulatory rule compilation and the joint reasoning of the graph neural differential compliance model, confirming that all rule satisfaction vectors and grading thresholds are satisfied, and generating and storing zero-knowledge proofs for interval and threshold rules; S7, outputting the risk grading results, the modified rules and the zero-knowledge proof identifiers, and updating the differentiable regulatory rule compilation parameters, the graph neural differential compliance model parameters, the dynamic confidence coefficient strategy and the abnormal severity score grading thresholds according to the audit feedback.

2. The financial audit compliance method based on the multi-dimensional rule engine, applied to the financial audit compliance system based on the multi-dimensional rule engine in claim 1, characterized in that, The multi-source business data specifically includes transactions, accounts, devices, merchants, channels, geographic locations and policy events. The preprocessing of the multi-source business data specifically includes time alignment, currency unification, missing value filling and abnormal value processing. ​ ​ ​ ​ ​ ​ 3. The financial audit compliance method based on multi-dimensional rule engine according to claim 2, wherein, ​ 4. The method of claim 2, wherein, ​ 5. The financial audit compliance method based on multi-dimensional rule engine according to claim 2, wherein, The constructing transaction account structure causal diagram specifically refers to, after pre-processing of multi-source business data is completed, setting transaction, account, device, merchant, channel, geographical location and policy event as nodes, establishing directed causal edges according to transaction fund flow direction and exogenous events, configuring static attributes and time-varying attributes for nodes and edges respectively, and generating a traceable causal graph structure by time window increment updating.

6. The multi-dimensional rule engine based financial audit compliance method as claimed in claim 2, wherein, The S2 specifically includes: S21, structurally analyzing the regulatory provisions and internal compliance rules, extracting subject elements, behavior elements and constraint elements, labeling threshold conditions, logical connection relationships, time windows and exemption conditions, generating a differentiable compliance intermediate representation and synchronously generating an exemption condition mask; S22, based on the differentiable compliance intermediate representation, mapping the threshold, proportion and count constraints in the intermediate representation into differentiable satisfaction functions, using a monotonic differentiable smoothing mapping to convert the difference between the feature and the threshold, the proportion deviation between the numerator and the denominator, and the threshold difference of the cumulative count within the time window into a satisfaction value between zero and one, and generating an initial rule satisfaction vector; S23, differentiable aggregation of the entries in the initial rule satisfaction vector involving time windows and time sequence constraints, weighted average of the satisfaction values at each time within the window according to the time interval decreasing weight, to obtain a time constraint aggregated satisfaction vector; S24, differentiable combination of the time constraint aggregated satisfaction vector and the satisfaction value not involving time constraint according to the logical relationship in the compliance intermediate representation, using continuous logical operators to aggregate the logical relationships including and, or, implication and exemption, to generate a logical combination satisfaction vector; S25, calculating a weighted coefficient for each rule in the logical combination satisfaction vector according to data integrity, freshness and source credibility, and limiting it to the range of zero to one through a monotonic differentiable mapping, multiplying the rule satisfaction value and the weighted coefficient to generate a rule satisfaction vector, and subtracting the corresponding weighted value from the maximum rule satisfaction value to generate a rule non-satisfaction vector; S26, generating a continuous hit indication score for the rule satisfaction vector according to the preset hierarchical threshold, sorting from high to low to form a rule hit sequence, and storing the rule satisfaction vector, the rule non-satisfaction vector and the rule hit sequence.

7. The multi-dimensional rule engine based financial audit compliance method as claimed in claim 2, wherein, The S3 specifically includes: S31, extracting time-ordered node features, edge features and global features from the transaction account structure causal diagram, organizing them into a structure feature sequence according to time sequence, and merging the rule satisfaction vector and the rule non-satisfaction vector into a rule feature sequence in order; S32, taking the structure feature sequence and the rule feature sequence as input, establishing the initial state of the graph neural differential model, and fixing the initial time point and the solution interval; S33, performing double-channel constraint injection in the graph neural differential model, injecting the rule satisfaction vector into the node state update channel, injecting the rule non-satisfaction vector into the edge state update channel, and disabling the exempted constraints in both channels according to the exemption condition mask. S34, adaptively schedule the solving step length according to the rule non-compliance degree vector and the risk indication in the structural feature sequence, set a smaller step length for time slices with high non-compliance degree or severe structural changes, and set a larger step length for time slices with low non-compliance degree and stable structure, and generate a continuous time state trajectory; S35, set a structural causal intervention hook during the continuous time state trajectory solving process, execute intervention on specified time slices based on exogenous events and policy events in the transaction account structural causal graph, and record the continuous time state representation after intervention; S36, perform constraint consistency projection on the continuous time state representation of each time slice, preferentially satisfy mandatory constraints, sequentially process mutually conflicting constraints according to a priority table, and minimize the deviation from the constraints already satisfied in the previous time slice, to obtain a consistent continuous time state representation sequence; S37, jointly train and infer the consistent continuous time state representation sequence and the rule feature sequence, and output the continuous time state representation of the target time slice.

8. The multi-dimensional rule engine based financial audit compliance method as claimed in claim 2, wherein, The S4 specifically comprises: S41, align the continuous time state representation with the rule compliance degree vector and the rule non-compliance degree vector according to the index of nodes and edges one by one, form a joint compliance feature vector, and retain the timestamp and causal path identifier; S42, perform multi-dimensional uncertainty evaluation on the joint compliance feature vector, generate uncertainty measures based on graph neural differential equation model perturbation results, data integrity and freshness indicators, and rule stability indicators within a time window, and aggregate them into a dynamic confidence coefficient set; S43, run the counterfactual model simulator in evaluation mode, apply the smallest feasible shadow modification to the joint compliance feature vector and observe the output change, obtain the counterfactual sensitivity factor, calibrate and fuse it with the dynamic confidence coefficient set, and generate a dynamic confidence coefficient vector; S44, use the dynamic confidence coefficient vector as the weight to aggregate the rule non-compliance degree on the causal path of the transaction account structural causal graph, generate node-level, edge-level and global abnormal severity scores, and set the minimum level requirement for risk classification for entries marked as mandatory constraints; S45, adaptively adjust the classification threshold according to the historical false positive rate, audit processing capacity, regulatory priority and business time period load, map the abnormal severity score to the risk classification label, and store the dynamic confidence coefficient vector, the abnormal severity score and the risk classification label.

9. The multi-dimensional rule engine based financial audit compliance method as claimed in claim 2, wherein, The S5 specifically comprises: S51, extract the joint compliance feature vector, the transaction account structural causal graph local subgraph, the rule compliance degree vector and the rule non-compliance degree vector of the abnormal severity score threshold object; S52, construct the business feasible region and the invariable constraint, clearly define the adjustable items as amount, frequency, time interval, channel selection and identity verification field, set the upper and lower bounds, step size and value set for each adjustable item, and convert the exemption condition mask to a disable constraint on the object, to form a constraint set; S53, locate the causal paths in the time window within the local subgraph of the transaction account structure causal graph, group the affected rules according to the rule non-satisfaction degree vector and determine the priority order of the amendments, generate an amendment order table, and construct the candidate changes with the smallest amplitude on the causal paths in sequence; S54, perform a multi-stage minimum modification solution, obtain a candidate scheme that satisfies all mandatory rules in the feasibility stage, select a candidate scheme with fewer modification steps and lower business cost in the business cost convergence stage, and call the differentiable regulatory rule compilation and graph neural differential in the virtual environment to perform rapid checking, eliminate the candidates that reappear rule non-satisfaction in the checking, and output the optimal amendment rule and amendment instruction; S55, generate a field-level change list, an execution order, and a rollback plan for the optimal amendment rule and amendment instruction, and form a minimum modification proof record.

Images