A digital financial management method and system based on deep learning

Abstract

The application discloses a kind of digital financial management method and system based on deep learning, it is related to deep learning technical field, including: collecting digital financial business data, pre-processing forms unified financial event data;Pulse sequence chain folding axis and flow mark notch section profile are executed, and form flow mark index table;Improved Markov Random Field model is constructed, and structure field atlas is obtained;The node risk evidence set is calculated, and period alignment and difference calculation are executed;Based on generalized approximate message passing inference, output posterior risk result;Matrix cross cross-checking and responsibility chain encapsulation are carried out, and action set is generated;Each mirror image is compared, and feedback sample and verification result are obtained.The application realizes the explainable audit of multi-source heterogeneous financial data by the iterative posterior inference of improved Markov Random Field model and generalized approximate message passing inference, closed-loop compliance decision.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a digital financial management method and system based on deep learning. Background Technology

[0002] Existing digital financial management typically covers transaction processing, account management, fund allocation, reconciliation and compliance review, and aggregates and analyzes data through data warehouses, streaming computing, or reporting platforms. However, financial business data is characterized by diverse sources, heterogeneous formats, significant differences in semantic definitions, fluctuating data quality, and unstable arrival times. During data collection, cleaning, deduplication, missing data completion, standardization coding, and cross-source association and fusion, issues such as inconsistent field definitions, inconsistent primary keys, inconsistent time definitions, and difficulty in resolving duplicate and conflicting records can arise, causing the same business indicator to differ across different stages. This makes it difficult to reuse offline feature definitions in real-time online processing, and it hinders the closure of feature generation rules, version management, and data lineage tracing. Consequently, online inference latency increases, result reproduction becomes difficult, and anomaly localization relies on manual investigation, making it difficult to meet the requirements of low-latency online management and end-to-end traceability, and also failing to support continuously evolving digital financial management scenarios.

[0003] Current risk identification and prediction methods primarily rely on single-point features or localized statistics, often focusing on the identification of individual transactions or accounts. This makes it difficult to simultaneously depict the complex relationships and risk propagation paths between accounts, transactions, merchants, devices, and time slices. Even with graph-based modeling or probabilistic inference, iterative inference is prone to oscillations, slow convergence, or instability in results across different batches when the graph is large, has many loops, and strong coupling, affecting the reliability of real-time decision-making. The model output often lacks a verifiable consistency verification process with compliance rules and handling procedures, making it difficult to form auditable chains of evidence, responsibility, and replayable snapshots. Management actions lack structured basis, and the review conclusions and handling results after action execution are difficult to standardize and rewrite back to ensure consistency with historical archives. This makes it difficult to construct an explainable, auditable, closed-loop, adaptively optimized, and executable decision-making chain, thereby affecting the overall effectiveness and regulatory compliance of risk control, reconciliation management, and compliance management.

[0004] Therefore, how to provide a digital financial management method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a digital financial management method and system based on deep learning. This invention comprehensively utilizes real-time access and consistent processing of multi-source heterogeneous financial business data, entity graph structured modeling, improved Markov random field model structure innovation, and generalized approximate message passing inference technology. It details the entire process from collecting and preprocessing financial business data to form unified financial event data, constructing an entity graph and generating a flow trace index table, obtaining a structural field map through a variable topological self-consistent domain, an entropy path bifurcation latent flow module, and a self-evolving roll-up graph, calculating node risk evidence and performing iterative posterior estimation, then generating an action set through matrix cross-verification and responsibility chain encapsulation and archiving, and forming a feedback loop through mirror verification of reconciliation trajectories. This invention improves the stability and real-time performance of iterative inference through rotating projection replanes and random cyclotron aftershock channels, and enhances traceability and auditability through flow trace indexes, responsibility chain snapshots, and mirror reconciliation. Compared with existing technologies, it has advantages such as low-latency online inference, guaranteed feature consistency, clear evidence chains, audit replayability, and strong closed-loop optimization capabilities.

[0006] A digital financial management method based on deep learning according to an embodiment of the present invention includes: Collect digitized financial business data, preprocess the financial business data, and form unified financial event data; Based on unified financial event data, an entity graph structure is generated, a set of node attributes and a set of edge relationships are established, and pulse sequence chain folding and flow mark engraving profile processing is performed on the set of edge relationships to form a flow mark index table. An improved Markov random field model is constructed. Based on the variable topological self-consistent domain, the entity graph structure is partitioned into domains. The set of edge relationships within the domain is reconnected and deleted. The entropy path bifurcation latent flow module is used to perform risk information bifurcation and convergence. A self-evolving rolled-up graph is introduced to perform the unfolding and rolling-up of the graph, resulting in a structure field map. The node risk evidence set is calculated based on the node event sequence and the structural field map, and the structural field map is written back. Period alignment and differential calculation are performed on the node event sequence, and cross-curtain accumulation and merging are performed on the low-frequency weak signal segments in the node event sequence. Based on generalized approximate message passing inference, iterative posterior estimation processing is performed on the structural field spectrum. Orthogonal rotation, projection and inverse transformation are performed on the iterative message vector by rotating the projection replane. Random interval injection and attenuation write-back are performed using the random cyclotron aftershock channel to output the posterior risk result. Based on the posterior risk results, flow trace index table, and node risk evidence set, matrix cross-verification and responsibility chain encapsulation and archiving are performed to generate an action set. By performing mirror verification of the reconciliation trajectory, the action set is compared one by one to generate mirror reconciliation results, and feedback samples and verification results are obtained.

[0007] Optionally, the formation of unified financial event data includes: Collect financial business data, which includes transaction flow data, account balance data, fund transfer data, counterparty data, merchant data, equipment data, geographical location data, contract and invoice data, reconciliation discrepancy data, and compliance list data. The financial business data is then aggregated according to data source identifier, subject identifier, event identifier, and timestamp to generate a set of original event records. Preprocessing is performed on the original event record set, including deduplication, missing value handling, outlier handling, time alignment, and correlation fusion. A unified event generation process is performed on the preprocessed set of original event records. Field merging and conflict resolution are performed on the original event records with the same event association key to form unified financial event data. Feature version identifiers are generated for the unified financial event data.

[0008] Optionally, forming the flow mark index table includes: Construct an entity graph structure based on unified financial event data, and generate an edge relationship set according to the entity relationships in the unified financial event data; The edge relationship set is processed by pulse sequence chain folding. The unified financial event data is sliced ​​according to the period to obtain the periodic sequence. Based on the periodic sequence, the edge records with the same entity pair identifier are concatenated in time order to obtain the pulse sequence chain. The entity pair identifier is the identifier obtained by concatenating the start node identifier and the end node identifier of an edge in the edge relationship set in order. The pulse sequence chain is connected by folding axis according to the periodic boundary. The folding axis connection is to connect the first edge record and the last edge record in adjacent periods to generate a cross-period folding axis edge and write it into the edge relationship set. Flow mark profile processing is performed on the edge relationship set. For each edge, the event sequence associated with the edge is extracted to generate an event trajectory record. The event trajectory records are arranged in chronological order to form a trajectory vector. The trajectory vector is sequentially encoded to generate a flow mark identifier. The flow mark identifier is associated with the edge identifier and written into an index structure to generate a flow mark index table.

[0009] Optionally, obtaining the structure field map includes: An improved Markov random field model is constructed. Based on the entity graph structure, a variable topological self-consistent domain, an entropy path bifurcation latent flow module, and a self-evolving roll-up graph are set up. The variable topological self-consistent domain is set up to form a set of domains on the set of nodes in the undirected graph. The entropy path bifurcation latent flow module is set up to add a set of latent flow nodes and a set of latent flow edges in each domain. The self-evolving roll-up graph is set up to establish a roll-up graph sequence on the time order of unified financial event data. Perform variable topology self-consistent domain processing, generate an initial domain set based on the edge relationship set and assign a domain identifier to each node, perform iterative merging and iterative splitting on the initial domain set according to the intra-domain connectivity determination to obtain a self-consistent domain set, and perform topology update on the intra-domain edge set within the self-consistent domain set; The entropy path bifurcation latent flow module is executed to generate a set of latent flow nodes in each self-consistent domain set and establish a set of latent flow edges with the nodes in the domain. The node information entropy is calculated based on the marginal probability distribution of the node state variables. Based on the rate of change of the node information entropy, bifurcation and convergence operations are performed on the set of latent flow edges. Perform self-evolving roll-up chart processing, write unified financial event data into the current chart in chronological order and generate the current chart identifier, mark the charts that arrive earlier than the current chart identifier as historical charts and generate historical chart identifiers, perform stability determination on historical charts, roll up the historical charts that are determined to be stable, and unfold the current chart. The improved Markov random field model is trained by extracting a training sample set from unified financial event data based on sample identifiers and generating a supervision label set. The potential function set parameters are updated using a pseudo-likelihood training method. In each training round, the corresponding training sample subset is selected based on the current curtain diagram identifier and the conditional probability of the node state variable is calculated. The pseudo-likelihood value is calculated based on the conditional probability. The potential function set parameters are updated based on the pseudo-likelihood value, and a structure field map is generated after training is completed.

[0010] Optionally, the calculation of the node risk evidence set based on the node event sequence and structural field map includes: Based on unified financial event data and rollup snapshots, determine the event subset corresponding to the current rollup chart, and merge the event subsets according to node identifiers to generate a node event sequence; A set of node risk evidence is generated based on the node attribute set of node event sequence and structural field map, and the structural field map is written back; Periodic alignment and differential calculation are performed on the node event sequence to generate an anomaly recording table. Cross-curtain accumulation and merging are performed on low-frequency weak signal segments to generate a fine-grained trace library. Low-frequency weak signal segments are event subsequences whose event count does not exceed a preset count within the time span. Cross-curtain accumulation and merging involves splicing low-frequency weak signal segments with the same node identification and event type identification in different curtain diagrams in chronological order and generating a trace summary, which is then written into the fine-grained trace library.

[0011] Optionally, the output posterior risk result includes: Iterative posterior estimation of the structural field graph is performed based on generalized approximate message passing inference. The node set and edge relationship set of the current curtain graph are extracted from the structural field graph. The observation vector is extracted and the state vector to be estimated is generated based on the node risk evidence set. A linear combination relationship is established such that the linear combination result vector is equal to the product of the linear combination matrix and the state vector to be estimated. An observation relationship is established such that the conditional probability of each element in the observation vector is determined by the corresponding linear combination result vector element. Perform rotation, projection, and replane processing. In each iteration, perform orthogonal rotation and projection on the variable-side message vector and the observation-side message vector respectively, and perform inverse transformation and write back. Orthogonal rotation is to multiply the message vector with the orthogonal matrix to obtain the rotated message vector. Projection is to multiply the rotated message vector with the projection matrix to obtain the projected message vector. Inverse transformation and write back is to multiply the projected message vector with the transpose of the orthogonal matrix to obtain the write-back message vector and replace the original message vector. Perform random cyclotron aftershock channel processing, record the residual vector of each iteration and calculate the residual energy, generate a trigger sequence based on the residual energy and determine the set of trigger iteration numbers, inject historical residual information into the update term of the observation end in the iteration corresponding to the set of trigger iteration numbers and perform attenuation write-back, stop the iteration when the L2 norm of the difference vector between two adjacent iterations of the state vector to be estimated is zero, and output the posterior risk result.

[0012] Optionally, the set of generated actions includes: A corroborating dataset is generated based on the posterior risk results, the flow mark index table, and the node risk evidence set. The posterior risk results are expanded according to the node identifier and the edge identifier to obtain the risk item set. The trajectory item set is retrieved from the flow mark index table according to the flow mark pattern identifier. The evidence item set is retrieved from the node risk evidence set according to the node identifier. The risk item set, trajectory item set, and evidence item set are associated and spliced ​​to generate the corroborating dataset. Perform matrix cross-verification on the verification dataset to generate a consistency verification matrix, and calculate the consistency value for each matrix element in the consistency verification matrix; A set of actions is generated based on the consistency verification matrix and the set of compliance rules. A rule judgment value is generated for each compliance rule. An action instruction is generated based on the rule judgment value. An action identifier, an action object identifier, a trigger basis identifier, and a timestamp are generated for each action instruction. Perform a chain of responsibility encapsulation and archiving on the action set to generate a chain of responsibility snapshot and write it to the archive library, and calculate a check value for the chain of responsibility snapshot.

[0013] Optionally, obtaining the feedback samples and verification results includes: Receive the execution results and review conclusions of the action set, and generate a feedback sample set according to the action identifier and action object identifier; Based on the feedback sample identifier set, the node identifiers and edge identifiers of the structure field map are located. Update entries are generated and written into the node attribute set and edge relationship set according to the execution status identifier and the review conclusion identifier. Update snapshots are generated and written into the rollup snapshot based on the update entries. Perform mirror verification of the reconciliation trajectory on the updated snapshot and the chain of responsibility snapshot in the archive to generate mirror reconciliation results. The mirror reconciliation results consist of a line-by-line comparison result, which is a Boolean value obtained by verifying the consistency of fields between the updated entry corresponding to the same action identifier and the action instruction in the chain of responsibility snapshot.

[0014] A deep learning-based digital financial management system according to an embodiment of the present invention includes the following modules: The data preprocessing module is used to collect financial business data and complete preprocessing to generate unified financial event data; The entity graph construction module is used to generate an entity graph structure from unified financial event data, forming a flow trace index table. The structure field generation module is used to construct an improved Markov random field model and generate structure field maps. The risk evidence generation module is used to calculate and write back the node risk evidence set based on the node event sequence and structural field map, and generate anomaly and trace results. The posterior inference module is used to complete iterative posterior estimation based on generalized approximate message passing and output the posterior risk result. The compliance decision archiving module is used to perform matrix cross-verification and responsibility chain encapsulation archiving based on the results of subsequent risk assessments, and generate a set of actions. The mirror reconciliation feedback module is used to perform mirror comparison on the action set to generate mirror reconciliation results and output feedback samples and verification results.

[0015] The beneficial effects of this invention are: This invention unifies the financial event processing of financial business data, completing deduplication, missing and anomaly handling, time alignment and correlation fusion, and forming traceable data lineage and feature version identifiers in the same link; it constructs an entity graph structure and establishes a set of node attributes and edge relationships, generates cross-period folded axis edges through pulse sequence chain folding processing to characterize periodic recurring relationships, and generates flow mark pattern identifiers for edge relationships through flow mark pattern profile processing to form a flow mark index table, thereby improving the correlation expression capability and index retrieval efficiency in the scenario of real-time access to multi-source heterogeneous data; further, it combines the domain partitioning of variable topology self-consistent domain and the reconnection and deletion within the domain, and the unfolding and rolling up of self-evolving rollup graph processing, so that the structure field graph maintains controllable graph size and computational load when transaction flows arrive continuously, reduces latency fluctuations caused by repeated calculations and structural drift, and improves the low latency characteristics of online inference and the replayability of historical snapshots.

[0016] This invention calculates and writes back a set of node risk evidence on a structural field map, generates an anomaly recording table by combining periodic alignment and differential calculation, and performs cross-curtain accumulation and merging on low-frequency weak signal segments to generate a fine-grained trace library, enabling weak anomalies and hidden risks to be continuously accumulated and form searchable evidence. In the posterior inference stage, iterative posterior estimation is performed based on generalized approximate message passing. Orthogonal rotation, projection, and inverse transformation of the iterative message vector are performed on the rotating projection replane to reduce oscillations caused by correlation. Random interval injection and attenuated write-back of residual information are performed through a random cyclotron aftershock channel to weaken the risk of spurious convergence. To improve the stability of inference and the consistency of results, in the decision-making process, the execution matrix is ​​cross-verified based on the posterior risk results, the flow trace index table and the node risk evidence set to form a consistency verification matrix. The responsibility chain is further encapsulated and archived to generate an action set. The consistency verification between the action execution results and the archived responsibility chain is realized through the mirror verification of the reconciliation trajectory, forming feedback samples and verification results. This enables an executable decision-making link that is interpretable, auditable and optimizable in closed loop between risk results and compliance rules and management processes, overcoming the technical bottlenecks of scattered evidence, difficulty in audit reproduction and unstable closed loop updates in the existing technology. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based digital financial management method proposed in this invention; Figure 2 This is a block diagram of an improved Markov random field model for a deep learning-based digital financial management method proposed in this invention. Figure 3 This is a functional diagram of a deep learning-based digital financial management system proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1 and Figure 2 A deep learning-based digital financial management method includes: Collect digitized financial business data, preprocess the financial business data, and form unified financial event data; Based on unified financial event data, an entity graph structure is generated, a set of node attributes and a set of edge relationships are established, and pulse sequence chain folding and flow mark engraving profile processing is performed on the set of edge relationships to form a flow mark index table. An improved Markov random field model is constructed. Based on the variable topological self-consistent domain, the entity graph structure is partitioned into domains. The set of edge relationships within the domain is reconnected and deleted. The entropy path bifurcation latent flow module is used to perform risk information bifurcation and convergence. A self-evolving rolled-up graph is introduced to perform the unfolding and rolling-up of the graph, resulting in a structure field map. The node risk evidence set is calculated based on the node event sequence and the structural field map, and the structural field map is written back. Period alignment and differential calculation are performed on the node event sequence, and cross-curtain accumulation and merging are performed on the low-frequency weak signal segments in the node event sequence. Based on generalized approximate message passing inference, iterative posterior estimation processing is performed on the structural field spectrum. Orthogonal rotation, projection and inverse transformation are performed on the iterative message vector by rotating the projection replane. Random interval injection and attenuation write-back are performed using the random cyclotron aftershock channel to output the posterior risk result. Based on the posterior risk results, flow trace index table, and node risk evidence set, matrix cross-verification and responsibility chain encapsulation and archiving are performed to generate an action set. By performing mirror verification of the reconciliation trajectory, the action set is compared one by one to generate mirror reconciliation results, and feedback samples and verification results are obtained.

[0020] In this embodiment, the formation of unified financial event data includes: Collect financial business data, which includes transaction flow data, account balance data, fund transfer data, counterparty data, merchant data, equipment data, geographical location data, contract and invoice data, reconciliation discrepancy data, and compliance list data. The financial business data is then aggregated according to data source identifier, subject identifier, event identifier, and timestamp to generate a set of original event records. Preprocessing is performed on the original event record set. This preprocessing includes deduplication, missing value handling, outlier handling, time alignment, and correlation fusion. Specifically, the preprocessing of the original event record set includes: Grouping records by event identifier and data source identifier, deduplication is completed by retaining only the record with the earliest timestamp and deleting the record for records with completely identical field values ​​within the same group. Then, missing records are checked field by field. Records with missing key fields such as subject identifier and timestamp are directly removed. For records with missing numeric fields, the median of the field within the time window of the same subject is used to fill the missing records. If there is no valid value within the window, the median of the entire data is used to fill the missing records. Outlier handling is performed on numeric fields. The first quartile and the third quartile are calculated within the time window of the same subject. The interquartile range is the third quartile minus the first quartile. Values ​​less than the first quartile minus 1.5 times the interquartile range or greater than the third quartile plus 1.5 times the interquartile range are replaced with the corresponding boundary values. Timestamps are uniformly converted to the same time zone and aligned by granularity. The aligned timestamp is the result of dividing the timestamp by the granularity in seconds, rounding it down, and then multiplying it by the granularity in seconds. Event association keys are generated based on the subject identifier and the aligned timestamp. Records with the same event association key are merged according to field priority. Conflicting fields are retained in the following order: priority of compliance list hit, priority of reconciliation difference, and priority of latest timestamp. A unified event generation process is performed on the preprocessed set of original event records. Field merging and conflict resolution are performed on the original event records with the same event association key to form unified financial event data. Feature version identifiers are generated for the unified financial event data.

[0021] In this embodiment, forming the flow mark index table includes: An entity graph structure is constructed based on unified financial event data, and an edge relationship set is generated according to the entity relationships in the unified financial event data. Specifically, the entity graph structure is constructed based on the unified financial event data as follows: Parse the account identifier, transaction identifier, merchant identifier, device identifier, and time slice identifier from each event record, generate corresponding nodes and write them into the node set. Generate entity association relationships according to event fields and form an edge relationship set. For each event, generate occurrence edges between account nodes and transaction nodes, acceptance edges between transaction nodes and merchant nodes, initiation edges between transaction nodes and device nodes, and attribution edges between account nodes and time slice nodes. Write amount, direction, event type, and timestamp into the edge attributes. Merge the edge relationship set according to the starting node identifier, ending node identifier, and event type. Calculate the occurrence count and cumulative amount of the merged edges. The occurrence count is the number of edge records in the merged group, and the cumulative amount is the sum of the amount fields in the merged group. When the occurrence count is not less than 3 times or the cumulative amount is not less than 10,000 yuan, the current edge is retained and the occurrence count and cumulative amount are written into the edge attributes. Otherwise, the edge is deleted. Solidify the retained node set and edge relationship set into an entity graph structure. The edge relationship set is processed by pulse-order chain folding. Time slices are applied to the unified financial event data according to the period to obtain a periodic sequence. Based on the periodic sequence, edge records with the same entity pair identifier are concatenated in chronological order to obtain a pulse-order chain. The entity pair identifier is obtained by sequentially concatenating the start and end node identifiers of an edge in the edge relationship set. The pulse-order chain is then folded along the periodic boundaries. Folding the chain involves connecting the first and last edge records in adjacent periods to generate a cross-period folded edge, which is then written into the edge relationship set. Specifically, the pulse-order chain folding process for the edge relationship set is as follows: The data is sliced ​​into time segments at fixed periods using timestamps of unified financial event data, with each period being one day. Each edge record is assigned to the corresponding daily period according to its timestamp to obtain a periodic sequence. The edge relationship set is grouped according to entity pair identifiers. The entity pair identifier is obtained by concatenating the start node identifier and the end node identifier in sequence. Within each entity pair identifier group, the edge records are sorted in ascending order by timestamp. The difference between the timestamps of two adjacent edge records is used as the interval value to form a pulse sequence chain that alternates between edge records and interval values. The period boundary is used as the fold axis point. A fold axis connection is performed on the pulse sequence chains of the same entity pair identifier group in two adjacent periods. The fold axis connection is generated by taking the last edge record of the group in the previous period and the first edge record of the group in the next period to generate a cross-period fold axis edge. The start node identifier and the end node identifier of the cross-period fold axis edge remain unchanged. The timestamp of the cross-period fold axis edge is taken from the timestamp of the first edge record in the next period. The interval value of the cross-period fold axis edge is taken from the timestamp of the first edge record in the next period minus the timestamp of the last edge record in the previous period. The cross-period fold axis edge is written into the edge relationship set. Flow mark profiling is performed on the edge relation set. For each edge, the event sequence associated with the edge is extracted to generate an event trajectory record. These event trajectory records are arranged chronologically to form a trajectory vector. Sequential encoding is performed on the trajectory vector to generate flow mark identifiers. The flow mark identifiers are then associated with the edge identifiers and written into an index structure to generate a flow mark index table. For each edge, an event sequence associated with the edge is extracted to generate an event trajectory record. Specifically, the starting node identifier, ending node identifier, and edge type identifier of the edge are located using the edge identifier. Event records that simultaneously meet the requirements of consistent starting node identifier, consistent ending node identifier, and consistent event type and edge type are selected from the unified financial event data to form a candidate set. The candidate set is sorted in ascending order by timestamp to obtain the event sequence. For each event sequence, the timestamp, amount, direction identifier, event type identifier, and counterparty identifier are extracted to form a trajectory entry. The timestamp difference between adjacent trajectory entries is calculated as an interval field, which is the number of seconds obtained by subtracting the previous timestamp from the next timestamp. The trajectory entries are written into the event trajectory record in sequence, and the trajectory start time, trajectory end time, number of trajectory entries, and cumulative amount value are generated. The cumulative amount value is the sum of the amount fields in the trajectory entries. The flow mark identifier is generated by sequential encoding of the trajectory vector. Specifically, the trajectory entries in the event trajectory record are converted into a fixed-length sequence with a length of 32. For entries longer than 32, only the most recent 32 are retained, and those shorter than 32 are padded with zeros at the end. Each trajectory entry is concatenated according to the field order to form an entry code. The entry code is a string obtained by concatenating the time interval field, amount field, direction identifier, event type identifier, and opponent identifier in sequence. The entry codes are concatenated in sequence to obtain the trajectory code string. The check value of the trajectory code string is calculated as the flow mark identifier. The check value is a fixed-length string obtained by performing a hash operation on the byte sequence of the trajectory code string. The flow mark identifier, the edge identifier, the trajectory start and end time, and the number of trajectory entries are written into the flow mark index table.

[0022] In this embodiment, obtaining the structure field map includes: An improved Markov random field model is constructed. Based on the entity graph structure, a variable topological self-consistent domain, an entropy path bifurcation latent flow module, and a self-evolving roll-up graph are set up. The variable topological self-consistent domain is set up to form a set of domains on the set of nodes in the undirected graph. The entropy path bifurcation latent flow module is set up to add a set of latent flow nodes and a set of latent flow edges in each domain. The self-evolving roll-up graph is set up to establish a roll-up graph sequence on the time order of unified financial event data. Perform variable topology self-consistent domain processing: generate an initial domain set based on the edge relation set and assign a domain identifier to each node; perform iterative merging and iterative splitting of the initial domain set according to the intra-domain connectivity determination to obtain a self-consistent domain set; and perform topology update on the intra-domain edge set within the self-consistent domain set, where: Based on the intra-domain connectivity determination, an iterative merging and splitting process is performed on the initial domain set to obtain a self-consistent domain set. Specifically, an intra-domain adjacency list is constructed with nodes as vertices and intra-domain edges as undirected edges. Each connected component is treated as an initial domain, and all nodes are assigned the same domain identifier. The intra-domain connectivity is calculated for any domain, which is the result of dividing the number of intra-domain edges by the number of intra-domain nodes minus one. When the intra-domain connectivity is less than 0.3, the current domain is split, re-dividing it into multiple subdomains according to connected components and updating the domain identifiers. When any node is connected to an external node... If a node has an edge and the edge appears at least 3 times in the last 7 days, mark the current domain and the domain containing the node outside the domain as a pair of domains to be merged. Calculate the inter-domain coupling degree for the pair of domains to be merged. The inter-domain coupling degree is the result of dividing the number of cross-domain edges by the sum of the number of nodes in the two domains. When the inter-domain coupling degree is greater than 0.2, perform the merge, merge the nodes of the two domains into the same domain and unify the domain identifier. Repeat this process in the order of splitting and merging until all domains satisfy the condition that the intra-domain connectivity degree is not less than 0.3 and there are no pairs of domains to be merged with an inter-domain coupling degree greater than 0.2, thus obtaining a set of self-consistent domains. Within the self-consistent domain set, a topology update is performed on the edge set within the domain. Specifically, for each self-consistent domain, edges are extracted and merged according to entity pair identifier and edge type identifier. If multiple edges exist within the same merge group, the edge with the largest cumulative amount is retained and the remaining edges are deleted. The activity level of each edge within the domain is calculated, which is the number of edge-related events in the last 7 days. Edges with an activity level of 0 are deleted. The consistency of each edge within the domain is calculated by taking the counterparty set of the edge's starting node and the counterparty set of the edge's ending node in the last 7 days, and calculating the intersection-union ratio of the two sets. The intersection-union ratio is the result of dividing the number of intersection elements by the number of union elements. When the intersection-union ratio is less than 0.1 and the edge activity level is less than 3, the edge is deleted. For edges with the same entity pair identifier that appear consecutively in adjacent periods, the timestamp of the edge is updated to the latest timestamp and the number of occurrences and cumulative amount are added to complete the topology update. The entropy path bifurcation latent flow module is executed to process the latent flow node set within each self-consistent domain set and establish a latent flow edge set with the nodes within the domain. The node information entropy is calculated based on the marginal probability distribution of the node state variables. Based on the rate of change of the node information entropy, bifurcation and convergence operations are performed on the latent flow edge set, where: Within each self-consistent domain set, a set of latent flow nodes is generated, and a set of latent flow edges is established with the nodes within the domain. Specifically, the risk uncertainty is calculated for each node within the self-consistent domain. The risk uncertainty is the difference between the maximum and second-largest probability values ​​of the marginal probability distribution of the node's state variables. Nodes are sorted from smallest to largest risk uncertainty, and the top 10% are selected as candidate latent flow nodes. The degree of each candidate latent flow node is calculated based on its edge relationship set. The degree is the number of edges within the domain connected to the node. The two nodes with the largest degrees are selected to generate the set of latent flow nodes. For each node within the self-consistent domain, the shortest path length to each latent flow node in the set of latent flow nodes is calculated. The shortest path length is the minimum number of edges. The latent flow node with the smallest shortest path length is selected, and a latent flow edge is established, pointing the node to the selected latent flow node. The path length and establishment time are written into the latent flow edge. The node information entropy is calculated based on the marginal probability distribution of the node state variables. Specifically, for each node, the set of values ​​of the state variables and the corresponding probability set are read. The sum of all probabilities in the probability set is one. For each value probability, the probability is multiplied by the natural logarithm of that probability. The sum of all values ​​is taken and the negative value is used to obtain the information entropy. When the value probability is zero, the probability corresponding to the value is multiplied by the natural logarithm and set to zero. The obtained information entropy is used as the node information entropy and written into the node attribute set. At the same time, the calculation time is recorded to form an information entropy time series. Based on the rate of change of node information entropy, bifurcation and convergence operations are performed on the set of latent edges. Specifically, the rate of change is obtained by dividing the difference in information entropy between two consecutive calculation times of the same node by the time difference between the two calculation times, with the time difference in seconds. When the rate of change of a latent edge is greater than 0.01, all latent edges pointing to the latent edge are grouped according to the risk category value of the source node, the set of latent edges is split into two latent paths and bifurcation node identifiers are generated. The bifurcation node identifiers are associated with the set of source nodes of each latent path. When the rate of change of a latent edge is less than 0.002, multiple latent paths corresponding to the latent edge are merged, the latent edges of each latent path are uniformly rewritten to point to the same latent edge and convergence node identifiers are generated. The bifurcation and convergence results are written into the risk bifurcation node set and the risk convergence node set, respectively. Perform self-evolving curtain graph processing: write unified financial event data into the current curtain graph in chronological order and generate the current curtain identifier; mark curtain graphs arriving earlier than the current curtain identifier as historical curtain graphs and generate historical curtain identifiers; perform stability determination on historical curtain graphs; perform roll-up processing on historical curtain graphs determined to be stable; and perform unfolding processing on the current curtain graph. Specifically: The unified financial event data is written into the current curtain graph in chronological order and a current curtain identifier is generated. Specifically, the unified financial event data is sorted by event timestamp, the curtain duration is set to 30 minutes, and the most recent complete curtain duration is taken as the current curtain time interval. Events whose timestamps fall into the current curtain time interval are written into the event set of the current curtain graph. The node and edge identifiers involved are obtained based on the event parsing, and the node and edge sets of the current curtain graph are updated respectively. The current curtain identifier is generated by concatenating the start and end timestamps of the current curtain time interval in sequence. Events whose timestamps are earlier than the start timestamp of the current curtain time interval are assigned to the corresponding historical curtain time interval and a historical curtain identifier is generated. The historical curtain identifier is generated by concatenating the start and end timestamps of the historical curtain time interval in sequence. A stability check is performed on historical charts. For historical charts deemed stable, they are rolled up. The current chart is unfolded. Specifically: For each historical curtain identifier, the historical curtain graph corresponding to the node state variable edge probability distribution and edge state variable edge probability distribution obtained from the last two rounds of inference are stored. The distribution difference between the two rounds is calculated. The distribution difference is a scalar obtained by summing the absolute differences of the probability of each state value of the same node or the same side. When the distribution difference between all nodes and all edges in the historical curtain graph is less than 0.001 and satisfies this condition for 3 consecutive rounds, the historical curtain graph is considered stable. For historical curtain graphs that are determined to be stable, roll-up processing is performed. Roll-up processing generates a summary snapshot of the posterior risk results of the summary nodes, the set of key associated paths, the set of undercurrent bifurcation nodes, and the set of undercurrent convergence nodes of the historical curtain graph. The event set and detailed edge set of the historical curtain graph are deleted, and only the summary snapshot and historical curtain identifier are retained. For the current curtain graph, unfold processing is performed. Unfolding processing parses the event set of the current curtain graph into a set of nodes and a set of edges and writes them into the structural field graph. An improved Markov random field model is trained by extracting a training sample set from unified financial event data based on sample identifiers and generating a supervision label set. A pseudo-likelihood training method is used to update the potential function set parameters. In each training round, a corresponding subset of training samples is selected based on the current curtain diagram identifier, and the conditional probabilities of the node state variables are calculated. The pseudo-likelihood value is calculated based on the conditional probability, and gradient updates are performed on the potential function set parameters based on the pseudo-likelihood value. After training, a structure field graph is generated, where: The potential function set parameters are updated using a pseudo-likelihood training method. Specifically, the potential function set parameters are defined as node potential function parameters, intra-domain edge potential function parameters, inter-domain edge potential function parameters, and latent edge potential function parameters. The node potential function takes the node feature vector as input and outputs the energy value of each state. The edge potential function takes the state pair of two adjacent nodes as input and outputs the joint energy value. The training sample set is extracted from the unified financial event data according to the sample identifier. The supervision label set and the training sample set are aligned according to the node identifier. The training sample subset whose timestamp falls into the curtain interval is selected according to the current curtain diagram identifier. The entity graph and domain partition corresponding to the training sample subset are used as the training graph. The current state value or label value of the remaining nodes except the target node is fixed on the training graph. The potential function parameters are updated node by node. The process involves calculating the conditional probabilities of the node state variables, calculating the pseudo-likelihood values ​​based on these probabilities, performing gradient updates on the potential function set parameters based on the pseudo-likelihood values, and generating a structure field map after training. Specifically: For any target node, calculate the local energy for all candidate states. The local energy is obtained by adding two parts: the first part is the energy value output by the node potential function in the candidate state, and the second part is the sum of the energy values ​​output by the edge potential functions of the node and all its neighbors in the combination of the candidate state and the current state of the neighbors. Take the negative of the local energy corresponding to each candidate state and perform an exponential operation to obtain the unnormalized probability. Add the unnormalized probabilities of each candidate state to obtain the normalized denominator. Divide the unnormalized probability by the normalized denominator to obtain the conditional probability. For all nodes in the training sample subset, take the conditional probability corresponding to the true label of each node, take the natural logarithm, and sum them to obtain the pseudo-likelihood value. Using pseudo-likelihood as the objective function, gradients are calculated for the node potential function parameters, intra-domain edge potential function parameters, inter-domain edge potential function parameters, and latent current edge potential function parameters. Parameter updates are performed with a learning rate of 0.001. After updating all training sample subsets corresponding to the current curtain diagram identifiers, the updated potential function set parameters are used to infer and summarize the structural field generation process, generating a structural field map.

[0023] In this embodiment, the calculation of the node risk evidence set based on the node event sequence and the structural field map includes: Based on unified financial event data and rollup snapshots, determine the event subset corresponding to the current rollup chart, and merge the event subsets according to node identifiers to generate a node event sequence; The process of generating a node risk evidence set based on the node attribute set of the node event sequence and the structural field graph, and then writing it back to the structural field graph, specifically involves: For each node, within the statistical window of the node event sequence, calculate the event count, sum of amounts, mean amount, variance of amounts, mean of adjacent event intervals, and deduplication count of the counterparty. The event count is the number of events, the mean amount is the sum of amounts divided by the event count, and the variance of amounts is the sum of the squares of the differences between each amount and the mean amount divided by the event count. All statistical features and the static attributes of the nodes in the structural field map are concatenated in field order to form a node evidence vector. The vector is standardized field by field, and the standardized value is the field value minus the training mean divided by the training standard deviation. The standardized node evidence vector is multiplied element by element by the probability weight vector, summed, and then exponentially normalized to obtain the risk probability value. For each risk category, the node evidence vector and the category weight vector are multiplied element by element and summed to obtain the category score. The category corresponding to the largest category score is taken as the risk category value. The contribution is calculated by the absolute value of the product of the field standardized value and the corresponding weight of the probability weight. The field with the largest contribution is selected to form the evidence feature set. The risk probability value, risk category value and evidence feature set are written into the node risk evidence set. The node identifier is used as the key to write back to the node attribute set of the structural field map. The process involves performing period alignment and differential calculation on node event sequences to generate an anomaly recording table, and performing cross-curtain accumulation and merging on low-frequency weak signal segments to generate a fine-grained trace library. Low-frequency weak signal segments are event subsequences whose event count does not exceed a preset count within a time span. Cross-curtain accumulation and merging involves splicing low-frequency weak signal segments with the same node identifier and event type identifier in different curtain diagrams in chronological order and generating a trace summary, which is then written into the fine-grained trace library. Specifically, the process of performing period alignment and differential calculation on node event sequences to generate the anomaly recording table is as follows: The node event sequence is divided into continuous periodic segments according to the periodic granularity. A periodic feature vector is calculated for each periodic segment. The periodic feature vector consists of event count, sum of amount, mean of amount, variance of amount, and deduplication count of counterparty. The periodic feature vectors of two adjacent periodic segments are subtracted according to the same field position to obtain the difference vector. The difference value of each field of the difference vector is the field value of the next period minus the field value of the previous period. When the absolute value of the difference value of any field is greater than 0.2 of the absolute value of the field value of the previous period and the field difference value is not zero, the node identifier, field identifier, previous period identifier, next period identifier and field difference value are written into the anomaly recording table until the difference calculation is completed for all adjacent periodic segments.

[0024] In this embodiment, the output of the posterior risk result includes: Iterative posterior estimation of the structural field graph is performed based on generalized approximate message passing inference. The node set and edge relationship set of the current curtain graph are extracted from the structural field graph. Observation vectors are extracted from the node risk evidence set and a state vector to be estimated is generated. A linear combination relationship is established such that the linear combination result vector equals the product of the linear combination matrix and the state vector to be estimated. An observation relationship is established such that the conditional probability of each element in the observation vector is determined by the corresponding element of the linear combination result vector. Specifically, the extraction of observation vectors and generation of state vectors to be estimated based on the node risk evidence set is as follows: For each node in the current curtain diagram, read the risk probability value and risk category value. Convert the risk category value into a one-hot encoded vector in category order. Concatenate the risk probability value and the one-hot encoded vector in order to generate the observation entry for the node. Arrange all the observation entries for the nodes in ascending order according to the node identifier and connect them sequentially to form the observation vector. At the same time, perform standardization on each element of the observation vector. Establish a state entry for each node. The state entry consists of the node's potential risk intensity variable and potential risk consistency variable. The potential risk intensity variable takes the value range of zero to one, and the potential risk consistency variable takes the value range of negative one to one. Connect all the state entries for the nodes sequentially according to the same node sorting rule as the observation vector to form the state vector to be estimated. Set the initial value of the state vector to be estimated as the vector obtained by concatenating the risk probability value and the zero value. The rotation-projection replane processing is performed. In each iteration, orthogonal rotation and projection are performed on the variable-side message vector and the observation-side message vector, respectively, followed by inverse transformation and write-back. Orthogonal rotation is obtained by multiplying the message vector with an orthogonal matrix to obtain the rotated message vector. Projection is obtained by multiplying the rotated message vector with the projection matrix to obtain the projected message vector. Inverse transformation and write-back is obtained by multiplying the projected message vector with the transpose of the orthogonal matrix to obtain the write-back message vector, which replaces the original message vector. The orthogonal matrix is ​​specifically: An orthogonal matrix is ​​a square matrix used to rotate and transform message vectors. The column vectors are pairwise orthogonal and each column has a Euclidean norm of 1. This ensures that the transpose of the matrix multiplied by itself yields an identity matrix, guaranteeing that the Euclidean norm of the message vector remains unchanged after matrix multiplication. Random cyclotron aftershock channel processing is performed, the residual vector of each iteration is recorded and the residual energy is calculated. A trigger sequence is generated based on the residual energy, and a set of trigger iteration numbers is determined. In the iterations corresponding to the trigger iteration number set, historical residual information is injected into the observation-end update term and attenuation write-back is performed. Iteration stops when the L2 norm of the difference vector between two adjacent iterations of the state vector to be estimated is zero, and the posterior risk result is output, where: Record the residual vector for each iteration and calculate the residual energy. Specifically, after each iteration, the residual vector is calculated as the observation vector minus the current linear combination result vector. Each element of the residual vector is the difference between the corresponding observation element and the corresponding linear combination element. Write the current residual vector into the residual sequence cache and write the corresponding iteration number. Calculate the residual energy as a scalar obtained by summing the squares of each element of the residual vector and write the residual energy into the residual energy sequence. A trigger sequence is generated based on the residual energy, and a set of trigger iteration numbers is determined. In the iterations corresponding to the trigger iteration number set, historical residual information is injected into the observation-end update item, and decay write-back is performed. Specifically, the absolute value of the difference between two adjacent residual energies is calculated for the residual energy sequence to form a change sequence. Each item in the change sequence is the absolute value of the current residual energy minus the previous residual energy. At the beginning of each iteration, a uniformly distributed random number is generated, with a value range of zero to one. When the random number is less than 0.1 and the current change sequence item is less than the average of the last 5 change sequences, the iteration number is added to the trigger iteration number set to form a trigger sequence. In the iterations corresponding to the trigger iteration number set, the historical residual information is obtained by adding the two most recent residual vectors from the residual sequence cache. The historical residual information is multiplied by the decay coefficient 0.3 to obtain the injection vector. The injection vector is added element by element to the observation-end update item to complete the injection and decay write-back.

[0025] In this embodiment, the set of actions to be generated includes: A corroborating dataset is generated based on the posterior risk results, the flow mark index table, and the node risk evidence set. The posterior risk results are expanded according to the node identifier and the edge identifier to obtain the risk item set. The trajectory item set is retrieved from the flow mark index table according to the flow mark pattern identifier. The evidence item set is retrieved from the node risk evidence set according to the node identifier. The risk item set, trajectory item set, and evidence item set are associated and spliced ​​to generate the corroborating dataset. The matrix cross-verification process is performed on the corroboration dataset to generate a consistent corroboration matrix. A consistency value is calculated for each matrix element in the consistent corroboration matrix. Specifically, the matrix cross-verification process is performed on the corroboration dataset to generate the consistent corroboration matrix. The corroborating dataset is grouped along two axes: source and evidence. The source dimension is defined as the source of posterior risk, the source of flow trace, and the source of risk evidence. The evidence dimension is defined as the risk category, the key correlation path, and the evidence feature. A matrix framework is constructed with the source dimension as the row and the evidence dimension as the column. Each corroborating record is placed into the corresponding matrix cell to form an entry set according to the source dimension identifier and the evidence dimension identifier. A consistency value is calculated for each matrix cell. The consistency value is the ratio obtained by dividing the number of mutually aligned entries in the cell by the total number of entries in the cell. An action set is generated based on the consistency verification matrix and the set of compliance rules. A rule judgment value is generated for each compliance rule, and an action instruction is generated based on the rule judgment value. For each action instruction, an action identifier, an action object identifier, a trigger basis identifier, and a timestamp are generated, where: The action set is generated based on the consistency verification matrix and the set of compliance rules. Specifically, a set of rule elements is established for each compliance rule. The set of rule elements consists of source dimension identifier, evidence dimension identifier, target object type identifier, and action type identifier. For each rule, the consistency value of the matrix unit corresponding to the source dimension identifier and evidence dimension identifier is extracted from the consistency verification matrix. According to the target object type identifier in the rule element, the corresponding node identifier or edge identifier is retrieved in the verification dataset as an action object candidate. A rule calculation record is formed for each action object candidate. The rule calculation record includes the matrix unit consistency value, risk category value, key association path identifier, and flow mark identifier. For each compliance rule, a rule judgment value is generated, and an action instruction is generated based on the rule judgment value. Specifically, the rule judgment value is calculated for the rule calculation record. The rule judgment value is a Boolean value. When the matrix cell consistency value is not less than 0.7, the risk category value belongs to the set of risk categories allowed by the rule, and the key association path identifier contains a flow mark identifier, the rule judgment value is true; otherwise, it is false. For action object candidates with a true rule judgment value, an action instruction is generated. The action instruction field includes an action type identifier, an action object identifier, and a trigger basis identifier. The trigger basis identifier is generated by concatenating the rule identifier, matrix cell coordinates, matrix cell consistency value, and key association path identifier in sequence. An action identifier is generated for each action instruction and a generation timestamp is written. The action identifier is obtained by concatenating the action object identifier, action type identifier, and generation timestamp in sequence. The responsibility chain is encapsulated and archived on the action set to generate a responsibility chain snapshot, which is then written to the archive repository. A checksum is calculated on the responsibility chain snapshot. Specifically, the responsibility chain encapsulation and archiving of the action set to generate the responsibility chain snapshot involves: The action instructions are sorted according to the generation timestamp. For each action instruction, the corresponding consistency verification matrix unit coordinates, matrix unit consistency value, key association path identifier, flow mark identifier, and node risk evidence item are associated with the trigger basis identifier. The associated information, together with the data lineage identifier, feature version identifier, graph snapshot identifier, and roll-up snapshot identifier, are serialized in the field order to generate a responsibility chain snapshot byte sequence. The responsibility chain snapshot byte sequence is written to the archive library. The check value of the responsibility chain snapshot byte sequence is calculated. The check value is the string obtained by performing a hash operation on the byte sequence. The check value and the responsibility chain snapshot identifier are written to the archive library together.

[0026] In this embodiment, obtaining the feedback sample and verification result includes: Receive the execution results and review conclusions of the action set, and generate a feedback sample set according to the action identifier and action object identifier; Based on the feedback sample identifier set, the node identifiers and edge identifiers of the structure field map are located. Update entries are generated and written into the node attribute set and edge relationship set according to the execution status identifier and the review conclusion identifier. Update snapshots are generated and written into the rollup snapshot based on the update entries. The mirror reconciliation results are generated by performing mirror verification of the reconciliation trajectory on the updated snapshot and the chain of responsibility snapshot in the archive. These mirror reconciliation results consist of a line-by-line comparison, where each comparison is a Boolean value obtained by verifying the consistency of fields between the updated entry corresponding to the same action identifier and the action instruction in the chain of responsibility snapshot. Specifically, the mirror reconciliation results generated by performing mirror verification of the reconciliation trajectory on the updated snapshot and the chain of responsibility snapshot in the archive are as follows: Extract the list of updated entries from the update snapshot and index them by action identifier. Read the chain of responsibility snapshots within the same time period from the archive and parse them to obtain the list of action instructions. Index them by action identifier. Take the intersection of the two indexes to obtain the set of action identifiers to be compared. Extract the updated entries and action instructions for each action identifier and perform field consistency verification. The field consistency verification compares the action object identifier, action type identifier, trigger basis identifier and timestamp field by field. Output a true value when all fields are equal, otherwise output a false value. Write the verification result of each action identifier into the comparison result list. Count the true value count and the total number. Divide the true value count by the total number to obtain the consistency ratio. Write the comparison result list and the consistency ratio into the mirror reconciliation result.

[0027] refer to Figure 3 A deep learning-based digital financial management system includes the following modules: The data preprocessing module is used to collect financial business data and complete preprocessing to generate unified financial event data; The entity graph construction module is used to generate an entity graph structure from unified financial event data, forming a flow trace index table. The structure field generation module is used to construct an improved Markov random field model and generate structure field maps. The risk evidence generation module is used to calculate and write back the node risk evidence set based on the node event sequence and structural field map, and generate anomaly and trace results. The posterior inference module is used to complete iterative posterior estimation based on generalized approximate message passing and output the posterior risk result. The compliance decision archiving module is used to perform matrix cross-verification and responsibility chain encapsulation archiving based on the results of subsequent risk assessments, and generate a set of actions. The mirror reconciliation feedback module is used to perform mirror comparison on the action set to generate mirror reconciliation results and output feedback samples and verification results.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous enterprise-level fund compliance audit cycle. The system received multi-source financial business data streams, with 2480 original event records covering receipts and payments, transfers, reversals, reconciliation discrepancies, and list matching. The original data had 6.4% duplicate submissions, 4.9% missing key fields, inconsistent timestamp time zones and precision, and amounts with heavy tails and outliers. After processing according to this invention, unified financial event data was achieved, retaining 2120 records after deduplication. Missing logarithmic values ​​were filled using the median of the sliding window for the same subject, totaling 963 fields filled; 187 outliers were corrected using the interquartile range method; timestamps were unified to the same time zone and aligned to the second level, resulting in 11.8% of records that could be merged within the same subject and second after alignment. Event association keys were generated using the subject identifier and aligned time period, and conflicting fields were resolved, resulting in 2980 unified financial event data records, and a feature version identifier was generated.

[0029] After unified financial event data is incorporated into the entity graph construction, it is parsed to generate 3184 nodes, including accounts, transactions, merchants, devices, and time slices, with 2905 initial edges. After edge merging, the frequency of occurrence and cumulative amount are calculated, retaining edges with at least 3 occurrences or at least 10000 cumulative amounts, resulting in 1142 retained edges. In the pulse sequence chain folding axis processing, consecutive edges of the same entity are connected across periods using a one-day cycle, generating 628 cross-period folding axis edges, with a return flow type accounting for approximately 41%. In the flow mark profile processing, associated event sequences are extracted from the retained edges to generate event trajectory records. The trajectory vector has a fixed length of 32. The entry encoding string is hashed to obtain flow mark identifiers, which are written into the flow mark index table. Training samples are stratified by subject to form 1600 training samples, 400 validation samples, and 500 test samples, with an outlier rate of 3.2%, ensuring that the same subject does not cross sets.

[0030] In the structural field generation phase, the system constructs an improved Markov random field model and performs variable topology self-consistent domain partitioning, resulting in 1260 self-consistent domains with an average of 252 nodes per domain and an edge density of 0.47. During domain topology updates, 874 duplicate entity edges are merged and deleted, 639 inactive edges are deleted, and 295 continuously related edges are reconnected and updated. The entropy path bifurcation latent flow module generates latent flow nodes and edges within the domain and calculates the information entropy change rate, recording 1980 bifurcation nodes and 1540 convergence nodes, forming a replayable bifurcation-convergence summary. The self-evolving rolled-up graph maintains the curtain sequence with a curtain duration of 30 minutes. Stability is determined by a distribution difference less than 0.001 and three consecutive rounds of this condition. In this cycle, 48 rolled-up snapshots and 36 rolled-up snapshots are generated. These snapshots preserve posterior summaries and critical path information to maintain long-term resource stability.

[0031] In the risk evidence calculation phase, statistical features are calculated based on the event sequence of nodes within the curtain and concatenated with the static attributes of the nodes to form a node evidence vector. After field-by-field standardization, risk probability values, risk category values, and evidence feature sets are obtained. Period alignment and differentiation generate an anomaly recording table with a daily cycle. Entries with an absolute difference value exceeding 20% ​​of the absolute value of the corresponding field in the previous cycle and with a difference not equal to 0 are written in, resulting in 246 anomaly recording entries. Low-frequency weak signals are judged based on an event count not exceeding 2 within a time span, generating segments and accumulating and merging them across curtains. The fine-grained trace library contains 780 entries, approximately 29% of which were confirmed as early abnormal clues during review.

[0032] In the posterior inference phase, iterative posterior estimation of the structural field spectrum is performed using generalized approximate message passing. The observation vector is obtained by concatenating and standardizing the risk probability value and risk category one-hot encoding. The state vector to be estimated consists of risk intensity and consistency variables. In the rotating projection replane, an orthogonal matrix satisfying the condition that its transpose multiplied by itself is the identity matrix is ​​used to complete the rotation and inverse transformation. In the cyclotron aftershock channel, the residual vector is the observation vector minus the linear combination result vector. The residual energy is the sum of the squares of the residual elements. When the random number is less than 0.1 and the change in residual energy is less than the average of the last 5 rounds, injection is triggered. The injection vector is the sum of the last two residuals multiplied by the attenuation coefficient of 0.3 and written back to the observation end update term.

[0033] During the decision-making phase, the system generates a corroborating dataset and constructs a consistency corroborating matrix based on the a posteriori risk results, flow trace index table, and node risk evidence set. The consistency value is calculated based on the number of alignable entries. Alignment conditions include consistent risk categories under the same node or edge, the inclusion of the same flow trace pattern in the critical path, or at least two matching evidence features. Compliance rules generate an action set based on a consistency value of not less than 0.7, the risk category falling into the rule set, and the inclusion of flow trace patterns in the critical path. This set is then encapsulated as a responsibility chain snapshot for archiving. After execution, the updated snapshot is mirrored against the responsibility chain snapshot line by line according to the action identifier. The action object, action type, triggering basis, and timestamp are verified, and the mirror reconciliation results and feedback samples are output.

[0034] In the comparative experiment, the traditional method employs conventional cleaning and fusion, static entity graphs, standard Markov random field inference, and conventional iterative message updates, directly triggering generation actions based on rules without performing matrix cross-verification or chain-of-responsibility mirror verification. Both methods use the same training, validation, and test samples. Anomaly identification accuracy is 84.6% for the traditional method and 92.8% for this invention; anomaly recall is 61.3% for the traditional method and 79.5% for this invention; false positive rate is 13.9% for the traditional method and 8.4% for this invention; reproducibility rate of key association paths is 52.1% for the traditional method and 95.6% for this invention; median single-event inference latency is 760 milliseconds for the traditional method and 190 milliseconds for this invention; inconsistency rate in feature consistency sampling is 3.1% for the traditional method and 0.2% for this invention; average audit review time is 41 minutes for the traditional method and 14 minutes for this invention; mirror reconciliation consistency ratio is 0.78 for the traditional method and 0.99 for this invention.

[0035] 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 digital financial management method based on deep learning, characterized in that, include: Collect digitized financial business data, preprocess the financial business data, and form unified financial event data; Based on unified financial event data, an entity graph structure is generated, a set of node attributes and a set of edge relationships are established, and pulse sequence chain folding and flow mark engraving profile processing is performed on the set of edge relationships to form a flow mark index table. An improved Markov random field model is constructed. Based on the variable topological self-consistent domain, the entity graph structure is partitioned into domains. The set of edge relationships within the domain is reconnected and deleted. The entropy path bifurcation latent flow module is used to perform risk information bifurcation and convergence. A self-evolving rolled-up graph is introduced to perform the unfolding and rolling-up of the graph, resulting in a structure field map. The node risk evidence set is calculated based on the node event sequence and the structural field map, and the structural field map is written back. Period alignment and differential calculation are performed on the node event sequence, and cross-curtain accumulation and merging are performed on the low-frequency weak signal segments in the node event sequence. Based on generalized approximate message passing inference, iterative posterior estimation processing is performed on the structural field spectrum. Orthogonal rotation, projection and inverse transformation are performed on the iterative message vector by rotating the projection replane. Random interval injection and attenuation write-back are performed using the random cyclotron aftershock channel to output the posterior risk result. Based on the posterior risk results, flow trace index table, and node risk evidence set, matrix cross-verification and responsibility chain encapsulation and archiving are performed to generate an action set. By performing mirror verification of the reconciliation trajectory, the action set is compared one by one to generate mirror reconciliation results, and feedback samples and verification results are obtained.

2. The digital financial management method based on deep learning according to claim 1, characterized in that, The formation of unified financial event data includes: Collect financial business data, which includes transaction flow data, account balance data, fund transfer data, counterparty data, merchant data, equipment data, geographical location data, contract and invoice data, reconciliation discrepancy data, and compliance list data. The financial business data is then aggregated according to data source identifier, subject identifier, event identifier, and timestamp to generate a set of original event records. Preprocessing is performed on the original event record set, including deduplication, missing value handling, outlier handling, time alignment, and correlation fusion. A unified event generation process is performed on the preprocessed set of original event records. Field merging and conflict resolution are performed on the original event records with the same event association key to form unified financial event data. Feature version identifiers are generated for the unified financial event data.

3. The digital financial management method based on deep learning according to claim 1, characterized in that, The process of forming the flow mark index table includes: Construct an entity graph structure based on unified financial event data, and generate an edge relationship set according to the entity relationships in the unified financial event data; The edge relationship set is processed by pulse sequence chain folding. The unified financial event data is sliced ​​according to the period to obtain the periodic sequence. Based on the periodic sequence, the edge records with the same entity pair identifier are concatenated in time order to obtain the pulse sequence chain. The entity pair identifier is the identifier obtained by concatenating the start node identifier and the end node identifier of an edge in the edge relationship set in order. The pulse sequence chain is connected by folding axis according to the periodic boundary. The folding axis connection is to connect the first edge record and the last edge record in adjacent periods to generate a cross-period folding axis edge and write it into the edge relationship set. Flow mark profile processing is performed on the edge relationship set. For each edge, the event sequence associated with the edge is extracted to generate an event trajectory record. The event trajectory records are arranged in chronological order to form a trajectory vector. The trajectory vector is sequentially encoded to generate a flow mark identifier. The flow mark identifier is associated with the edge identifier and written into an index structure to generate a flow mark index table.

4. The deep learning-based digital financial management method according to claim 1, characterized in that, The obtained structure field map includes: An improved Markov random field model is constructed. Based on the entity graph structure, a variable topological self-consistent domain, an entropy path bifurcation latent flow module, and a self-evolving roll-up graph are set up. The variable topological self-consistent domain is set up to form a set of domains on the set of nodes in the undirected graph. The entropy path bifurcation latent flow module is set up to add a set of latent flow nodes and a set of latent flow edges in each domain. The self-evolving roll-up graph is set up to establish a roll-up graph sequence on the time order of unified financial event data. Perform variable topology self-consistent domain processing, generate an initial domain set based on the edge relationship set and assign a domain identifier to each node, perform iterative merging and iterative splitting on the initial domain set according to the intra-domain connectivity determination to obtain a self-consistent domain set, and perform topology update on the intra-domain edge set within the self-consistent domain set; The entropy path bifurcation latent flow module is executed to generate a set of latent flow nodes in each self-consistent domain set and establish a set of latent flow edges with the nodes in the domain. The node information entropy is calculated based on the marginal probability distribution of the node state variables. Based on the rate of change of the node information entropy, bifurcation and convergence operations are performed on the set of latent flow edges. Perform self-evolving roll-up chart processing, write unified financial event data into the current chart in chronological order and generate the current chart identifier, mark the charts that arrive earlier than the current chart identifier as historical charts and generate historical chart identifiers, perform stability determination on historical charts, roll up the historical charts that are determined to be stable, and unfold the current chart. The improved Markov random field model is trained by extracting a training sample set from unified financial event data based on sample identifiers and generating a supervision label set. The potential function set parameters are updated using a pseudo-likelihood training method. In each training round, the corresponding training sample subset is selected based on the current curtain diagram identifier and the conditional probability of the node state variable is calculated. The pseudo-likelihood value is calculated based on the conditional probability. The potential function set parameters are updated based on the pseudo-likelihood value, and a structure field map is generated after training is completed.

5. The deep learning-based digital financial management method according to claim 1, characterized in that, The set of node risk evidence calculated based on node event sequences and structural field maps includes: Based on unified financial event data and rollup snapshots, determine the event subset corresponding to the current rollup chart, and merge the event subsets according to node identifiers to generate a node event sequence; A set of node risk evidence is generated based on the node attribute set of node event sequence and structural field map, and the structural field map is written back; Periodic alignment and differential calculation are performed on the node event sequence to generate an anomaly recording table. Cross-curtain accumulation and merging are performed on low-frequency weak signal segments to generate a fine-grained trace library. Low-frequency weak signal segments are event subsequences whose event count does not exceed a preset count within the time span. Cross-curtain accumulation and merging involves splicing low-frequency weak signal segments with the same node identification and event type identification in different curtain diagrams in chronological order and generating a trace summary, which is then written into the fine-grained trace library.

6. The digital financial management method based on deep learning according to claim 1, characterized in that, The output posterior risk result includes: Iterative posterior estimation of the structural field graph is performed based on generalized approximate message passing inference. The node set and edge relationship set of the current curtain graph are extracted from the structural field graph. The observation vector is extracted and the state vector to be estimated is generated based on the node risk evidence set. A linear combination relationship is established such that the linear combination result vector is equal to the product of the linear combination matrix and the state vector to be estimated. An observation relationship is established such that the conditional probability of each element in the observation vector is determined by the corresponding linear combination result vector element. Perform rotation, projection, and replane processing. In each iteration, perform orthogonal rotation and projection on the variable-side message vector and the observation-side message vector respectively, and perform inverse transformation and write back. Orthogonal rotation is to multiply the message vector with the orthogonal matrix to obtain the rotated message vector. Projection is to multiply the rotated message vector with the projection matrix to obtain the projected message vector. Inverse transformation and write back is to multiply the projected message vector with the transpose of the orthogonal matrix to obtain the write-back message vector and replace the original message vector. Perform random cyclotron aftershock channel processing, record the residual vector of each iteration and calculate the residual energy, generate a trigger sequence based on the residual energy and determine the set of trigger iteration numbers, inject historical residual information into the update term of the observation end in the iteration corresponding to the set of trigger iteration numbers and perform attenuation write-back, stop the iteration when the L2 norm of the difference vector between two adjacent iterations of the state vector to be estimated is zero, and output the posterior risk result.

7. The deep learning-based digital financial management method according to claim 1, characterized in that, The set of generated actions includes: A corroborating dataset is generated based on the posterior risk results, the flow mark index table, and the node risk evidence set. The posterior risk results are expanded according to the node identifier and the edge identifier to obtain the risk item set. The trajectory item set is retrieved from the flow mark index table according to the flow mark pattern identifier. The evidence item set is retrieved from the node risk evidence set according to the node identifier. The risk item set, trajectory item set, and evidence item set are associated and spliced ​​to generate the corroborating dataset. Perform matrix cross-verification on the verification dataset to generate a consistency verification matrix, and calculate the consistency value for each matrix element in the consistency verification matrix; A set of actions is generated based on the consistency verification matrix and the set of compliance rules. A rule judgment value is generated for each compliance rule. An action instruction is generated based on the rule judgment value. An action identifier, an action object identifier, a trigger basis identifier, and a timestamp are generated for each action instruction. Perform a chain of responsibility encapsulation and archiving on the action set to generate a chain of responsibility snapshot and write it to the archive library, and calculate a check value for the chain of responsibility snapshot.

8. The digital financial management method based on deep learning according to claim 1, characterized in that, The obtained feedback samples and verification results include: Receive the execution results and review conclusions of the action set, and generate a feedback sample set according to the action identifier and action object identifier; Based on the feedback sample identifier set, the node identifiers and edge identifiers of the structure field map are located. Update entries are generated and written into the node attribute set and edge relationship set according to the execution status identifier and the review conclusion identifier. Update snapshots are generated and written into the rollup snapshot based on the update entries. Perform mirror verification of the reconciliation trajectory on the updated snapshot and the chain of responsibility snapshot in the archive to generate mirror reconciliation results. The mirror reconciliation results consist of a line-by-line comparison result, which is a Boolean value obtained by verifying the consistency of fields between the updated entry corresponding to the same action identifier and the action instruction in the chain of responsibility snapshot.

9. A deep learning-based digital financial management system, executing the deep learning-based digital financial management method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data preprocessing module is used to collect financial business data and complete preprocessing to generate unified financial event data; The entity graph construction module is used to generate an entity graph structure from unified financial event data, forming a flow trace index table. The structure field generation module is used to construct an improved Markov random field model and generate structure field maps. The risk evidence generation module is used to calculate and write back the node risk evidence set based on the node event sequence and structural field map, and generate anomaly and trace results. The posterior inference module is used to complete iterative posterior estimation based on generalized approximate message passing and output the posterior risk result. The compliance decision archiving module is used to perform matrix cross-verification and responsibility chain encapsulation archiving based on the results of subsequent risk assessments, and generate a set of actions. The mirror reconciliation feedback module is used to perform mirror comparison on the action set to generate mirror reconciliation results and output feedback samples and verification results.

Images