A water transportation risk control method based on a knowledge graph

Abstract

The application discloses a water transportation risk control method based on a knowledge graph, belongs to the technical field of water transportation supply chain risk control and knowledge graph data processing, and aims at solving the problems that water transportation supply chain multi-source performance data is fragmented, and abnormal link and fund risk transmission paths are difficult to identify. The application builds performance evidence anchor points, subject relationship graphs and performance event evidence graphs, determines the propagation gate state by using existence support channels and consistency contradiction channels, and performs risk contribution deduction and local recalculation according to evidence version dependency indexes, so that the technical effects of distinguishing between business nonoccurrence and evidence contradiction, inhibiting risk mispropagation and reducing the update recalculation range are achieved.

Description

Technical Field

[0001] This invention relates to the field of water transport supply chain risk control and knowledge graph data processing, and in particular to a water transport risk control method based on knowledge graphs. Background Technology

[0002] The fulfillment process of the water transport supply chain involves multiple types of data, including orders, waybills, vessel trajectories, port operation events, transport documents, payment records, invoices, insurance, and claims. Existing systems typically store this data separately according to business departments or information systems. Different data sources lack unified semantics in terms of subject identification, business objects, event times, and document fields, making it difficult to stably correlate information corresponding to the same order, voyage, or cargo batch.

[0003] Existing risk identification methods often rely on field rules or independent anomaly scores in a single data table. This makes it difficult to organize the physical facts of the trajectory, document assertions, and financial records into the same performance context, and also makes it difficult to determine whether anomalies are propagated along directed relationships between performance events, documents, invoices, payments, and insurance. For changes in evidence such as document revisions, payment reversals, invoice cancellations, policy amendments, and trajectory re-transmissions, existing methods may continue to use old evidence to generate risky results, or recalculate the entire dataset, leading to false propagation and an expanded scope of processing.

[0004] Therefore, a water transport risk control method is needed to address the shortcomings of existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based risk control method for water transport. Addressing the problems in existing technologies such as inconsistent semantics among multi-source data in the water transport supply chain, difficulty in aligning subjects and fulfilling parties, the inability to form a traceable propagation chain for anomalous evidence, and the difficulty in incrementally updating risk results after evidence correction, this invention proposes a technical solution that organizes subject relationship graphs and performance event evidence graphs using performance evidence anchor points. Within the same anchor point, it generates trajectory physical facts, evidence assertions, and typified contradictory evidence, and uses existence support channels and consistency contradiction channels to jointly determine the state of candidate propagation edges. Simultaneously, it establishes a dependency index between evidence version, contradictory edges, propagation gate states, and risk path contributions, and revoks old contributions and partially recalculates the associated subgraph when evidence changes. This invention possesses the technical effects of distinguishing between transactions that have not occurred and transactions that have occurred but with contradictory evidence, restricting the direction of risk message propagation, maintaining traceable evidence paths, and reducing mispropagation and duplicate calculations caused by corrected data.

[0006] This invention provides a knowledge graph-based water transport risk control method, comprising:

[0007] S1. Acquire multi-source business data of the water transport supply chain and perform unified semantic mapping and entity alignment. Based on the alignment results, generate performance evidence anchor points, subject relationship diagrams and performance event evidence diagrams to obtain anchored diagram data.

[0008] S2. Generate trajectory physical facts and evidence assertions based on the anchor point graph data. Within the same performance evidence anchor point, detect conflicts according to the reference performance partial order and calculate multidimensional consistency residuals to obtain typified contradictory evidence and performance anomaly subgraphs.

[0009] S3. Align the performance anomaly subgraph with the subject relationship graph, establish existence support channels and consistency contradiction channels for candidate directed propagation edges respectively, and output the business existence status, necessary evidence completeness and contradiction evidence vector pointing to the target node; determine the propagation gate state based on the output and event timing reachability to obtain the gated propagation graph;

[0010] S4. Generate an initial risk message based on the typified contradictory evidence associated with the source node. In the gated propagation graph, the risk message is only transmitted along the activated candidate directed propagation edges. Combined with historical performance characteristics, abnormal performance chains, risk subject association groups, affected nodes, and traceable evidence paths are obtained. At the same time, a dependency index is established between evidence version, contradictory edge, propagation gate state, and risk path contribution.

[0011] S5. In response to evidence changes, based on the invalidated evidence of the dependent index and the corresponding risk path contribution, only the associated performance evidence anchor subgraph is recalculated to obtain the updated propagation gate state, risk value, and evidence difference.

[0012] Optionally, S1 includes: order information, waybills, vessel AIS tracks, BeiDou tracks, port events, transport documents, payment records, invoices, insurance policies, claims records, and historical performance records;

[0013] Map fields from various data sources to a unified semantic type for subjects, business objects, performance events, documents, fund records, and insurance records, and merge entities pointing to the same object based on subject identifier, vessel identifier, document identifier, time proximity relationship, and port location relationship;

[0014] The performance evidence anchor is generated by combining the identifiers of order, waybill, shipment batch, vessel, voyage, port of origin and destination, time window and settlement stage;

[0015] The subject relationship graph is formed by combining the subject nodes, subject identifiers, control, transaction and historical performance relationships. The performance event evidence graph is formed by combining the performance event nodes, trajectory physical facts, assertion nodes and the relationship between the nodes and the performance evidence anchor points.

[0016] Optionally, S2 includes: generating physical facts of berthing, departure, stay, segment and yaw trajectories based on the time, location, speed, heading and spatial relationship with port geofence of continuous trajectory points;

[0017] Convert the field values ​​in the transport documents into assertion nodes, and record the data source, evidence version, effective time, expiration time, content summary and credibility for the assertion nodes;

[0018] The reference performance partial order is constructed according to the sequential constraints between performance event types. Events in the same performance evidence anchor point are compared with the reference performance partial order to detect missing, reversed, and time-out events. Consistency residuals for time and space, cargo volume, amount, tax amount, and insurance coverage are calculated. The detection results and consistency residuals are encoded as the typified contradictory evidence according to the conflict type.

[0019] Furthermore, the consistency residuals are dimensionless, and the dimensionless residuals are compared with the residual boundaries determined based on historical normal performance samples of the same route, cargo type and settlement stage.

[0020] Write the residuals that exceed the residual boundaries, the corresponding residual directions, the associated evidence identifiers, and the target node identifiers into the contradictory evidence vector;

[0021] When there is a reachable path from a preceding event type to a subsequent event type in the reference performance partial order and the time order of the corresponding observed events is reversed, the corresponding event pair is marked as an inverse conflict.

[0022] For event pairs that are not comparable in the reference performance partial order, no reverse order marking is made;

[0023] The time interval exceeding the historical normal performance duration quantile value of the corresponding event type is marked as a timeout conflict, and the reverse order conflict and timeout conflict are written into the contradictory evidence vector.

[0024] Optionally, S3 includes: using the directed relationship between the performance event node, document node, invoice node, payment node, and insurance node as the candidate directed propagation edge;

[0025] The existence support channel performs existence verification on records from independent data sources that are associated with the same performance evidence anchor point, and outputs the business existence status and the completeness of necessary evidence.

[0026] The consistency contradiction channel encodes the typified contradiction evidence according to the contradiction type, contradiction pointing node, residual direction, and residual magnitude, and outputs the contradiction evidence vector;

[0027] When the business exists in the state of "has occurred", the completeness of the necessary evidence satisfies the completeness condition of the corresponding edge type, the contradictory evidence vector points to the target node of the candidate directed propagation edge and the target node is reachable from the source node in time sequence, the candidate directed propagation edge is set to the active state.

[0028] When the business exists in a state of pending supplementary evidence, the completeness of necessary evidence does not meet the completeness condition, contradictory evidence is missing, or time sequence information is missing, the candidate directed propagation edge is set to a state of pending supplementary evidence.

[0029] When the service exists in a state of not occurring, the contradictory evidence vector is determined not to point to the target node, or the target node is determined to be unreachable in time sequence, the candidate directed propagation edge is set to a blocked state.

[0030] Furthermore, the existence support channel is configured with a set of necessary evidence types for different candidate directed propagation edge types. The evidence types that have been obtained and aligned with entities within the same performance evidence anchor point are intersected with the set of necessary evidence types. The ratio of the number of evidence types in the intersection to the number of evidence types in the set of necessary evidence types is used as the completeness of the necessary evidence.

[0031] Independent data sources refer to evidence generated independently by different entities or isolated original collection systems, and there is no copying, transfer, or derivation relationship between the evidence records;

[0032] When the completeness of the necessary evidence is one and the evidence in the set of necessary evidence types comes from at least two independent data sources, output the status of the business that has occurred.

[0033] If there are no records to confirm that a transaction has occurred, output the status of the non-occurring transaction.

[0034] In other cases, output the status of the business requiring supplementary certification;

[0035] Furthermore, the event timing reachability is determined based on the source node event time, the target node event time, and the reference fulfillment partial order of the candidate directed propagation edge;

[0036] The contradictory evidence vector pointing to the target node means that the target node identifier in the contradictory evidence vector is consistent with the target node identifier of the candidate directed propagation edge, and the residual direction or partial order conflict indicates that the risk impact is transmitted along the direction of the candidate directed propagation edge.

[0037] For candidate directed propagation edges that are in the pending supplementary evidence status, record the type of missing evidence and the deadline for supplementary evidence; for candidate directed propagation edges that are in the blocked status, record the reason for the blockage.

[0038] When missing evidence is added, the version of the evidence corresponding to the reason for the blockage changes, or the deadline for supplementing evidence expires, the propagation gate state is redefined, and when the deadline for supplementing evidence expires and the missing evidence is still not added, the corresponding candidate directed propagation edge is set to the blocked state.

[0039] Optionally, S4 includes: generating the initial risk message based on the residual magnitude and partial order conflict type in the typified contradictory evidence associated with the source node;

[0040] The ratio of the number of target-oriented evidence types obtained to the number of target-oriented evidence types required for the corresponding contradiction type is used as the completeness of the contradictory evidence. The average credibility of the assertion nodes associated with the contradictory evidence vector is used as the credibility of the data source. The evidence time decay factor is obtained by monotonically decreasing mapping based on the time difference between the evaluation time and the effective time of the assertion node.

[0041] The product of the completeness of the contradictory evidence, the credibility of the data source, and the evidence time decay factor is used as the propagation gating coefficient, and the risk message transmitted from the source node to the target node is adjusted by the propagation gating coefficient.

[0042] A baseline risk value is generated based on the historical performance characteristics of the entity corresponding to the target node. The risk increments formed by each risk message received by the target node are summed and added to the baseline risk value to obtain the risk value of the target node.

[0043] The target node that is reached by the activated propagation path and receives at least one risk message is identified as the affected node, and the semantic type of the affected node as a funding node or insurance node is retained.

[0044] The abnormal performance chain and traceable evidence path are generated according to the directed propagation edge sequence through which the risk message passes, and the risk subject association group is formed based on the nodes of the shared subject relationship edge in the abnormal performance chain.

[0045] Optionally, S5 includes: when document revision, payment reversal, invoice red-ink reversal, policy amendment, or trajectory retransmission occurs, setting the expiration time of the evidence before the change to the effective time of the change and generating a new version of the evidence after the change;

[0046] Based on the dependency index, determine the conflicting edges, propagation gate states, and traceable evidence paths that use the evidence before the change, and deduct the risk path contribution corresponding to the evidence before the change.

[0047] Using the affected performance evidence anchor points as boundaries, re-execute event conflict detection, multidimensional consistency calculation, and propagation gate state determination, and generate early warning interpretation records using the updated risk values ​​and the differences in the content summaries of evidence before and after the change;

[0048] Furthermore, the dependency index uses the evidence version identifier as the first index key, records the contradictory edge identifier generated by the evidence version, the propagation gate identifier controlled by the contradictory edge, the traceable evidence path identifier using the propagation gate, and the risk increment of the evidence version to each traceable evidence path, with the risk increment serving as the risk path contribution.

[0049] The risk value of a path is obtained by adding the risk path contributions belonging to the same traceable evidence path, and the risk value of the target node is obtained by adding the cumulative risk values ​​of paths reaching the same target node and then adding them to the baseline risk value.

[0050] When the evidence version becomes invalid, the risk path contribution is deducted from the cumulative risk value of the corresponding path and the risk value of the target node according to the dependency index, and only the nodes and edges that have data dependency relationship with the corresponding performance evidence anchor point are added to the recalculation set.

[0051] After the evidence is changed, the differences in evidence version identifiers, content summaries, transmission gate status, and risk values ​​before and after the change will be written into the warning interpretation record.

[0052] The beneficial effects of this invention are:

[0053] 1. By unifying semantic mapping, entity alignment, and performance evidence anchors, order, trajectory, document, fund, and insurance data are organized into subject relationship graphs and performance event evidence graphs, enabling subject relationships and performance events across data sources to be associated in the same graph data structure.

[0054] 2. By outputting the business existence status and contradictory evidence pointing to the target node through the existence support channel and consistency contradiction channel respectively, and combining the event timing reachability to determine the propagation gate status, the risk message is only transmitted along the directed edge that meets the conditions, thereby reducing the false propagation of risks caused by missing, delayed or inconsistent evidence.

[0055] 3. By using the dependency index of evidence version, contradictory edges, propagation gate status and risk path contribution, the contribution of old evidence is deducted and the anchor point subgraph of related performance evidence is partially recalculated when evidence is revised or retransmitted. This can keep the update results interpretable and reduce the recalculation scope when data is corrected. Attached Figure Description

[0056] 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:

[0057] Figure 1 This is the overall flowchart of the water transport risk control method based on knowledge graphs of the present invention.

[0058] Figure 2This is a flowchart of the S3 gated propagation graph construction of the present invention. Detailed Implementation

[0059] 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.

[0060] refer to Figures 1-2 A knowledge graph-based water transport risk control method includes:

[0061] S1. Acquire multi-source business data of the water transport supply chain and perform unified semantic mapping and entity alignment. Based on the alignment results, generate performance evidence anchor points, subject relationship diagrams and performance event evidence diagrams to obtain anchored diagram data.

[0062] S2. Generate trajectory physical facts and evidence assertions based on the anchor point graph data. Within the same performance evidence anchor point, detect conflicts according to the reference performance partial order and calculate multidimensional consistency residuals to obtain typified contradictory evidence and performance anomaly subgraphs.

[0063] S3. Align the performance anomaly subgraph with the subject relationship graph, establish existence support channels and consistency contradiction channels for candidate directed propagation edges respectively, and output the business existence status, necessary evidence completeness and contradiction evidence vector pointing to the target node; determine the propagation gate state based on the output and event timing reachability to obtain the gated propagation graph;

[0064] S4. Generate an initial risk message based on the typified contradictory evidence associated with the source node. In the gated propagation graph, the risk message is only transmitted along the activated candidate directed propagation edges. Combined with historical performance characteristics, abnormal performance chains, risk subject association groups, affected nodes, and traceable evidence paths are obtained. At the same time, a dependency index is established between evidence version, contradictory edge, propagation gate state, and risk path contribution.

[0065] S5. In response to evidence changes, based on the invalidated evidence of the dependent index and the corresponding risk path contribution, only the associated performance evidence anchor subgraph is recalculated to obtain the updated propagation gate state, risk value, and evidence difference.

[0066] In this specific embodiment, S1 includes:

[0067] This specific implementation method is executed by the water transport risk control platform. The platform takes a water transport business from Wuhan Port to Shanghai Port, with steel as the cargo type and a planned cargo volume of 8200 tons, as the processing object throughout S1 to S5. The access records are organized according to the data source identifier, original record identifier, collection time, business time, subject identifier, vessel identifier, document identifier, order identifier, waybill identifier, cargo batch identifier, voyage identifier, origin and destination ports, settlement stage, and original content summary. The order system, transportation management system, automatic identification system for vessels, Beidou terminal, port operation system, electronic document system, payment system, invoice system, insurance system, claims system, and historical performance database submit incremental records through read-only interfaces. The platform uses the combination of the original system serial number and the data source identifier as the idempotent key. For records that arrive repeatedly, only the receiving batch number is updated without duplicate map creation. Records with empty business times are written to the pending verification queue and do not participate in the entity merging in this cycle.

[0068] The semantic mapping module reads the water transport semantic dictionary version SEM-WATER-2026.08, mapping carriers, shippers, consignees, invoice issuers, payers, and insurers as subject types; orders, waybills, cargo batches, vessels, voyages, and ports as business object types; berthing, departure, stay, loading and unloading, payment, invoicing, insurance, and claims as performance event types; and bills of lading, loading lists, receipts, invoices, insurance policies, and endorsements as documents, fund records, or insurance records. A record in the semantic dictionary has source system, source field, target type, and conversion rule as fields. The source system is the transportation management system, the source field is the carrier entity code field, the target type is the carrier entity identifier, and the conversion rule is to remove spaces and convert to the uppercase form of the unified social credit code. This dictionary is constructed from the field directories of each source system and confirmed historical records. A new version is generated when the error rate of adding, deleting, or converting source fields exceeds 0.5%.

[0069] The entity alignment module first performs full-width and half-width character unification, administrative division suffix normalization, and character noise reduction on the entity name. For the vessel name, it binds the Chinese vessel name to the MMSI identifier. Time is uniformly converted to Beijing time millisecond timestamps, and port location is uniformly converted to WGS-84 coordinates. Then, within the same entity type, it generates deterministic candidates based on entity identifier, vessel identifier, or document identifier. If a deterministic identifier is missing, supplementary candidates are generated based on time proximity and port location proximity. When a non-conflicting entity's unified social credit code, vessel MMSI, or document number precisely matches the unique standard entity, the identifier evidence is directly attached and saved, and it does not enter the similarity threshold branch. Similarity scores are only calculated when a deterministic identifier is missing or not matched. For joint business record candidates that simultaneously have entity, vessel, and document fields... and Calculate entity alignment score , Indicates candidate records and The dimensionless entity alignment score, with values ​​ranging from 0 to 1. This indicates the record index in the source record table to be aligned in the current entity alignment batch, with a value range from the first to the last record in that table. Representation and Recording The entity index in the candidate standard entity set of the same entity type, with a value range of all members of the candidate set. and These are all discrete indices rather than physical quantities, therefore units are not applicable. The value is 1 if the subject identifiers match and are identical, otherwise it is 0. The value is 1 if the ship identification is consistent and the identification is consistent, otherwise it is 0. The value is 1 if the document identifiers match and are consistent, otherwise it is 0. This represents the similarity of business times, which decays linearly relative to a 24-hour window based on the time difference and is truncated to 0 to 1. The similarity of port locations is expressed as a linear decay over a 5 km relative spatial distance window and truncated to 0 to 1. , , , and All are dimensionless quantities; the entity alignment configuration table ENTITY-ALIGN-2026.08 stores applicable indicators and original weights according to four categories: entity, vessel, document, and joint business records. Pure entity records only retain entity identifier, time, and location indicators; pure vessel records only retain vessel identifier, time, and location indicators; and pure document records only retain document identifier, time, and location indicators. Before calculation, inapplicable indicators are deleted, and the original weights of the retained indicators are normalized to a weight sum of 1 according to their proportion to the total retained weights. Automatic alignment is rejected and written to the configuration repair queue when the configuration is missing, multiple rows are hit, or the total retained weights are not positive. The weights are calibrated by the most recent 180 days of manually confirmed entities on the training set according to the comprehensive precision and completeness indicators and are fixed in this configuration version.

[0070] The entity alignment module sorts all candidates for the same source record in descending order of score. When scores are the same, they are sorted in lexicographical order of the candidate standard entity identifiers, but this does not automatically eliminate the uncertainty of ties. ENTITY-ALIGN-2026.08 also records a dimensionless uniqueness interval of 0.05. Only when the highest-scoring candidate is unique, the highest score is greater than or equal to 0.85, and the difference between the highest and second-highest scores is greater than or equal to 0.05, will the platform automatically attach the source record to that unique highest-scoring standard entity and retain the source record identifier; if there is only one candidate, the second-highest score is considered 0. When the highest scores are tied, the difference between the highest and second-highest scores is less than 0.05, two or more candidates simultaneously reach 0.85 but do not meet the uniqueness interval, or the highest score is greater than or equal to 0.65 but less than 0.85, all relevant candidates are written to the entity table to be reviewed, and a temporary entity identifier is used in the drawing for this period, without attaching any candidate; a new standard entity is created when the highest score is less than 0.65 or the candidate set is empty. When a conflict occurs between a deterministic entity identifier, a vessel MMSI identifier, or a document number, the deterministic identifier is not overridden by name similarity. Instead, an identifier conflict record is established and the corresponding relationship edges are isolated. Before a graph transaction is committed, it is verified that a source record has at most one currently valid standard entity affiliation based on the source system, source record identifier, and entity alignment configuration version. If multiple valid affiliations are found, the record is rolled back and transferred to pending review. This ensures that entity merging can be traced back to the source record, entity alignment configuration version, and semantic dictionary version that participated in the calculation.

[0071] The anchor point generation module performs SHA-256 hashing on the normalized strings of order identifier, waybill identifier, cargo batch identifier, vessel standard entity identifier, voyage identifier, port of departure, port of destination, planned performance window, and settlement stage, and encodes the first 128 bits as the performance evidence anchor point identifier. The anchor point records generated by this bill of lading business include order ORD-WHSH-017, waybill WB-017, cargo batch LOT-STEEL-01, vessel MMSI 413000017, voyage VOY-202608, port of departure Wuhan Port, port of destination Shanghai Port, planned performance window from 00:00 on August 1, 2026 to 24:00 on August 8, 2026, and settlement after arrival in the settlement stage. If the order is split into multiple cargo batches or the settlement stage is switched, sub-anchor points are generated respectively and associated with the parent anchor point with the parent anchor point identifier field.

[0072] The entity relationship diagram is stored using an attribute graph. Each entity node records at least the entity node identifier, entity type, unified social credit code, name standard value, evidence source set, and historical performance profile version. Entity identifier edges, control edges, transaction edges, and historical performance relationship edges record the relationship edge identifier, source entity, target entity, relationship type, effective time, expiration time, evidence version set, and confidence status, respectively. The platform retains relationship edges only when at least one currently valid record supports the relationship. When there are conflicting registration records in the control relationship, all evidence versions are retained, and the relationship edge is marked as pending judgment. Control edges pending judgment are not used for subsequent determination of risk entity association groups.

[0073] The performance event evidence diagram and the subject relationship diagram share standard entity identifiers. The performance event node records the event node identifier, anchor point identifier, event type, event time, port location, cargo volume, amount, and status. The trajectory physical fact node records the fact node identifier, trajectory segment identifier, fact type, start and end time, spatial range, and generation rule version. The assertion node records the assertion node identifier, field name, field value, unit, data source, evidence version, effective time, expiration time, content summary, and credibility. The event nodes, trajectory physical facts, assertion nodes, and performance evidence anchor points are connected through anchor point attribution relationships, fact support relationships, field assertion relationships, and source derivation relationships. Each relationship edge stores the source record identifier and processing batch number, enabling subsequent contradiction detection to locate the original evidence.

[0074] The graph writing module writes standard entities, performance evidence anchors, subject relationship graphs, and performance event evidence graphs within a single transaction. It establishes a composite index using anchor identifiers, node types, and node identifiers, and an inverted index using evidence versions and business times. Before transaction commit, it checks that the anchor combination fields are complete, the nodes at both ends of the relationship edge exist, and the evidence version is not expired. Records that fail the check are written to the graph record rejection queue and the reason for rejection is recorded. After the check passes, it outputs a snapshot of the anchored graph data, which includes the anchor table, node table, relationship edge table, source record to standard entity mapping table, and semantic dictionary version, for S2 to read by anchor identifier.

[0075] In this specific embodiment, S2 includes:

[0076] The trajectory fact generation module reads the AIS and BeiDou trajectories corresponding to ship MMSI 413000017 in the S1 anchor point map data, and uniformly forms a trajectory point sequence that includes trajectory source, sampling time, longitude, latitude, ground speed, heading, positioning quality, and original record identifier. The AIS trajectory uses a target sampling interval of 30 seconds, and the BeiDou trajectory uses a target sampling interval of 60 seconds. When there are adjacent points from the two sources within 90 seconds, they are sorted according to the sampling time without averaging. When the time of consecutive points is reversed or the instantaneous position shift speed exceeds 35 kn, the point is marked as an anomaly and the original value is retained but not included in the physical fact calculation. When there is a continuous lack of measurement for more than 15 minutes, a new trajectory segment is formed.

[0077] The platform uses the approved WGS-84 polygon geofences of Wuhan Port and Shanghai Port to determine the execution points of trajectory points within the polygons. The berthing start condition is three consecutive trajectory points entering the fence with a ground speed not exceeding 1 kn, and the departure condition is three consecutive trajectory points leaving the fence with a ground speed exceeding 2 kn. The stay fact is defined as the cumulative duration of adjacent valid points not less than 30 minutes and the average speed not exceeding 0.5 kn. The valid trajectory segment between two adjacent berthing or departure facts is defined as the segment fact. When a trajectory point deviates from the centerline of the historical normal trajectory of the same route by more than 8 km and lasts for 10 minutes, a deviation fact is generated. The fence version, centerline version, threshold, and supported trajectory point identifiers are all written into the trajectory physical fact node.

[0078] The assertion generation module parses each field of bills of lading, loading orders, receipts, invoices, payment records, insurance policies, and endorsements, converting carrier, vessel, port of loading / unloading, loading / unloading time, cargo volume, tax-inclusive amount, tax amount, insurance amount, and policy validity period into assertion nodes. Each assertion node records the data source, source entity, original record identifier, evidence version, effective time, expiration time, content summary, field unit, and credibility. Credibility is determined by looking up three items in a table: source authentication level, signature verification result, and field completeness rate. One record in the source credibility table is a port operation system record that has passed electronic signature, with authentication level A, field completeness rate not less than 0.98, and credibility of 0.95. This table is calibrated by source grouping based on the manual verification results of the most recent 180 days and is updated when the monthly verification accuracy rate changes by more than 2 percentage points.

[0079] The reference performance partial order is stored using a directed acyclic graph. Nodes represent event types such as order confirmation, berthing at loading port, loading completion, departure from loading port, en route segment, berthing at unloading port, unloading completion, receipt, invoicing, payment, and insurance activation. Edges record the preceding event type, following event type, applicable route, cargo type, settlement stage, minimum normal duration, and maximum normal duration. The platform also reads an event necessity table of version EVT-REQ-2026.08, which is categorized by route, cargo type, settlement stage, insurance coverage, business type, and... The contract conditions are used as a joint lookup key, marking each event as required, not required, or inapplicable. If the contract stipulates that insurance is effective when the insurance is purchased, it is required; if not, it is inapplicable. The platform statistically analyzes the duration distribution of completed normal performance samples with the same route, cargo type, and consistent settlement stage in the past 12 months, using the 1st quantile and the 99th quantile as the duration boundary. When the sample size is less than 100, it is reverted to samples of the same route area and cargo type, and the revert level is recorded on the edge. When loop detection finds that a new constraint forms a closed loop, the constraint version is rejected.

[0080] For the same performance evidence anchor point, the platform first filters out the events that are required in the current anchor point according to the event necessity table, and then checks whether the required events are missing according to the reference performance partial order. Non-necessary or inapplicable events are not included in the missing event detection. If the event necessity record does not exist, multiple valid records are hit by the same key, or the contract conditions cannot be uniquely resolved, the configuration pending status is output, no missing event conflict is generated, and the anchor point is written to the configuration repair queue. If there is a reachable path from the preceding event type to the following event type in the partial order graph, but the observed business time is reversed, a reverse order conflict is generated. If the two event types are not comparable in the partial order graph, no reverse order mark is made. When the actual time interval exceeds the 99th percentile of the historical normal performance time of the corresponding event type, a timeout conflict is generated. Missing events that have not reached the planned completion time are only marked as under observation. Only necessary events that have reached the planned completion time but still have no independent evidence to support them generate missing event conflicts.

[0081] The multidimensional consistency module generates the original variances for spatiotemporal, cargo volume, amount, tax, and insurance coverage. The spatiotemporal variance consists of the shortest distance from the declared port to the track's berthing fence and the difference between the declared time and the actual time of the track. The cargo volume variance is the difference between the bill of lading cargo volume and the port loading / unloading volume. The amount variance is the difference between the invoice's tax-inclusive amount and the cumulative payment amount. The insurance coverage variance is the positive part of the claim claim loss amount minus the effective insurance amount. The tax variance is determined based on the tax rate parameter table of version TAX-WATER-2026.08. This table uses the invoice's goods or services tax category code, taxpayer type, transaction time, and applicable region as a joint key, recording the tax rate, price-tax caliber, and validity period. For invoices with a tax-exclusive amount, the reference tax amount is calculated by multiplying the tax-exclusive amount by the applicable tax rate. For invoices with a tax-inclusive amount, the reference tax amount is calculated by subtracting the tax-inclusive amount from the tax-inclusive amount, dividing by one, and adding the applicable tax rate. The tax difference is then calculated by subtracting the reference tax amount from the invoice tax amount. The tax rate table is published by the audited financial and tax parameter library and the version is locked at the start of batch S2. If a unique valid tax rate cannot be found, the price and tax caliber is missing, or there is a conflict in the tax rate records, the tax rate status is output as pending, no tax amount contradiction is generated, and the data is written to the parameter repair queue. Each difference quantity retains its original direction and unit. Distance is expressed in km, time in h, cargo quantity in t, and amount and insurance amount in yuan. It is forbidden to directly sum the original differences of different units.

[0082] Dimension The original difference is robustly dimensionless. , Representing dimensions Signed dimensionless uniform residuals This indicates a dimension index obtained from the dimension fields of the residual calibration table version RES-WHSH-2026.08. The member range only includes six categories: spatiotemporal distance, spatiotemporal time, cargo volume, amount, tax amount, and insurance coverage. Since this is a discrete category index rather than a physical quantity index, units are not applicable. Indicates the current anchor point in dimension The original differences are determined by the corresponding dimension. This represents the median of the original differences in this dimension among historical normal performance samples of the same route, cargo type, and settlement stage, and is consistent with... Same unit, This indicates that the robust scale is obtained by multiplying the interquartile range of the corresponding dimension of the historical sample by 0.7413 and is consistent with... Same unit, This indicates the smallest scale within the same unit to prevent division by zero, and takes the corresponding measurement resolution. This indicates that the value of the robust scale and the minimum scale is not less than the other term. Historical normal samples, robust scales, and minimum scales are all stored in the residual calibration table version RES-WHSH-2026.08.

[0083] The residual calibration table uses route, cargo type, settlement stage, and dimension as joint keys to record the median, robust scale, minimum scale, two-sided residual boundaries, sample size, validity period, and version. One cargo volume record has the following keys: Wuhan Port to Shanghai Port, steel, settlement upon arrival, and cargo volume. The robust scale is 18 t, the minimum scale is 1 t, the two-sided boundaries are -3.2 to +3.2, and the sample size is 1264. When the current residual is below the lower boundary or above the upper boundary, a numerical inconsistency is generated in the direction of negative or positive, respectively. No numerical inconsistency is generated when the residual is within the boundary. The calibration table is recalculated monthly using closed and undisputed normal performance anchor points and the old version is retained to support backtracking.

[0084] The contradiction coding module encodes missing, reversed order, timeout, spatial inconsistency, inconsistent cargo quantity, inconsistent amount, inconsistent tax amount, and insufficient insurance coverage as typified contradictory evidence. It records the contradictory evidence identifier, anchor point identifier, conflict type, source node identifier, target node identifier, set of associated evidence identifiers, set of evidence versions, residual dimension, residual direction, residual magnitude, residual boundary, partial order path, and generation rule version. Signed residuals that exceed the residual boundary, target node identifiers, and their associated evidence identifiers are simultaneously written into the contradictory evidence vector. Dimensions that do not exceed the boundary are not falsely labeled with zero values ​​to create contradictions.

[0085] When searching the residual calibration table, only the four joint keys of the anchor point—route, cargo type, settlement stage, and residual dimension—are used to precisely match the unique valid record at the evaluation time. Only when this precise hierarchical record exists and there are at least 100 closed, undisputed, and normally fulfilled samples for the same route, cargo type, and settlement stage, is the median, robust scale, minimum scale, and two-sided boundaries of this record allowed to be used to calculate residuals and generate numerical contradictions. Wide-level records for the same cargo type in the same flight area, the same cargo type across the entire flight area, or those with deleted settlement stages are only used for sample collection planning, quality control, and prior diagnosis; their boundaries must not be directly used for residual comparison or contradiction generation at the current anchor point. If the precise record does not exist, multiple rows are matched, the sample size is less than 100, or any of the four joint keys is missing, the dimension outputs a calibration insufficiency status and is written to the monthly sample replenishment queue, without generating numerical contradictions. Other dimensions with valid precise hierarchical records are still judged independently according to their respective boundaries.

[0086] The S2 batch log also saves the trajectory fence version, the reference performance partial order version, the source credibility table version, and the residual calibration table version. When any version is switched during operation, the current anchor point uses the version locked at the beginning of the batch to complete the calculation. The new version only applies to the next batch. This version locking rule ensures that trajectory facts, evidence assertions, partial order conflicts, and consistency residuals in the same performance anomaly subgraph have a consistent parametric time section and provides input for S5 to locally recalculate according to the original version.

[0087] Centered on the anchor point of the promissory note performance evidence, the platform forms a performance anomaly subgraph by including event nodes, trajectory physical facts, assertion nodes, corresponding typified contradictory evidence, and their SUPPORTS, CONTRADICTS, and TARGETS relationships among the event nodes involved in conflict judgment. Each contradictory relationship edge stores the original difference, calibration table version, and evidence version for recalculation. The output object includes the performance anomaly subgraph identifier, a list of typified contradictory evidence, contradictory evidence vectors, and a reference performance partial order version, and is passed to S3 for node alignment with the main relationship graph.

[0088] In this specific embodiment, S3 includes:

[0089] The gating graph construction module takes the performance anomaly subgraph output by S2 and the main relationship graph output by S1 as input. It aligns the graphs according to the shared standard entity identifier, performance evidence anchor point identifier, and evidence version. It reads the directed relationships allowed by the business definition between performance event nodes, document nodes, invoice nodes, payment nodes, and insurance nodes. The candidate directed propagation edges include at least the relationships of performance event pointing to document, document pointing to invoice, invoice pointing to payment, transportation event pointing to insurance, and claims pointing to payment. The edge records the candidate edge identifier, source node, target node, edge type, source event time, target event time, necessary evidence type set version, and current gate status.

[0090] The existence support channel reads the necessary evidence type set for each candidate edge type. A set of records for an invoice pointing to a payment edge includes three evidence types: valid invoice, corresponding contract or waybill, and bank payment record. A set of records for a transportation event pointing to an insurance edge includes four evidence types: valid insurance policy, insured shipment batch, insurance effective time, and corresponding voyage. The necessary evidence set is constructed from confirmed business processes and historical normal performance samples, using edge type and settlement stage as the lookup key. The version is REQ-2026.08, and it will be updated when process fields change or the misjudgment rate exceeds 1% for two consecutive weeks.

[0091] For candidate edges Calculate the completeness of necessary evidence , Indicate candidate edges The dimensionless completeness of necessary evidence, with values ​​ranging from 0 to 1. This represents the edge index in the candidate directed propagation edge table of the current performance evidence anchor point, and the traversal range is from the first to the last candidate edge in the table. Since this is a discrete index rather than a physical quantity, units are not applicable. This refers to the set of evidence types that have been obtained within the same performance evidence anchor point, aligned with entities, and are valid at the time of assessment. Indicate candidate edges The corresponding set of necessary evidence types, This represents the number of distinct evidence types in the intersection of two sets. Indicates the number of evidence types in the set of necessary evidence types; if the set of necessary evidence types is empty, the candidate edge is rejected and a configuration error is recorded, and zero is not used as the denominator.

[0092] The platform checks the independence of data sources based on the source system identifier, source entity identifier, original record identifier, and derived source identifier. Only when the source entities are different or the original collection systems are physically isolated, and there is no copying, transfer, or derivative chain between the evidence, are they counted as mutually independent sources. When the completeness of the necessary evidence is 1 and the evidence comes from at least two mutually independent data sources, the business existence status is set to "occurred". When there are no valid records to confirm the occurrence of the business, it is set to "not occurred". All other cases are set to "pending supplementary evidence". The types of evidence obtained, the types of missing evidence, and the number of independent sources are written into the existence support record.

[0093] The consistency contradiction channel reads the typified contradiction evidence from S2 and generates a fixed-length contradiction evidence vector according to the contradiction type, target node identifier, residual direction, residual absolute magnitude, correlation assertion confidence, and evidence version. The direction determination field adopts three states: matching, non-matching, and unknown. When the target node identifier is consistent with the candidate edge target node identifier and the residual direction or partial order conflict indicates that the influence is transmitted along the direction from the source node to the target node, it is set to matching. When only the target node identifier has been obtained but is inconsistent with the candidate edge target node, the evidence is only associated with the source node, or the propagation direction has been determined to be opposite, it is set to non-matching. When the target node identifier, pointing evidence, or direction field required for determination is missing, it is set to unknown and the missing item is recorded. Unknown is not used to replace non-matching.

[0094] The absolute magnitude of the residual in the contradictory evidence vector is the absolute value of the dimensionless residual of S2 and truncated between 0 and 10. The larger the value, the stronger the deviation from the historical normal boundary. The residual direction is stored in a negative, positive or no-direction field. The absolute magnitude cannot replace the propagation direction. The platform does not directly select candidate edges based on this magnitude. Instead, it first requires the target node identifier and direction to match simultaneously, and then the gate state predicate determines whether to activate. After the vector is written, the contradictory evidence identifier and the generation batch are saved for S4 to form the initial risk message.

[0095] Event time-series reachability is jointly determined by the event time of the candidate edge source node, the event time of the target node, and the S2 reference fulfillment partial order. The platform first checks whether there is a directed path from the source event type to the target event type in the partial order graph, and then checks whether the target event time is earlier than the source event time. When the partial order is reachable and the time order is consistent, it outputs reachable. When the partial order is incomparable or any event time is missing, it outputs unknown. When the partial order is clearly unreachable or the time order is opposite to the only reachable direction, it outputs unreachable. The partial order path and event time evidence identifier used are written into the reachability record.

[0096] The gate state determiner executes according to deterministic predicates. When the business existence status is "occurred", the completeness of necessary evidence reaches 1, the direction determination field is matched, and the temporal status is reachable, the candidate edge is set to active. When any of the following conditions are met: business existence status is "pending verification", necessary evidence completeness is less than 1, direction determination field is unknown, or temporal status is unknown, and there are no conditions for "business not occurred", "direction determined to be mismatched", or "temporal status clearly unreachable", it is set to "pending verification". When any of the following conditions are met: business existence status is "not occurred", "direction determination field mismatched", or "temporal status unreachable", it is set to "blocked". The determined blocking conditions take precedence over the pending verification conditions, but unknown directions are not blocking conditions. When the direction changes from unknown to matched after verification, the business existence status, completeness, and temporal status determination are recombined. If the target-pointing evidence cannot be obtained by the deadline for verification, it is switched to blocking according to the pending verification timeout rules. Only active edges are allowed to transmit risk messages by S4.

[0097] The higher the value of the completeness of necessary evidence, the more complete the evidence types are. The completeness is only used to compare with the completeness conditions of the edge type and is not used to rank the risk scores. In this specific implementation, the completeness condition is set to 1. When the completeness is insufficient, even if the contradiction amplitude exceeds the corresponding residual boundary, the candidate edge will not be activated. When any of the blocking conditions in the business existence status, direction matching field and time series reachability are met, the blocking will not be lifted due to the increase in completeness. Thus, the evidentiary sufficiency meaning of completeness is independent of the S4 risk intensity direction.

[0098] For edges requiring supplementary documentation, the platform records the set of missing evidence types, the initial time of supplementary documentation request, the deadline for supplementary documentation, and the time limit configuration version. The initial time of supplementary documentation request is the transaction commit time when the candidate edge first changes from a non-supplementary documentation status to a supplementary documentation status; repeated evaluations cannot be postponed. The supplementary documentation time limit configuration table in version EVIDENCE-DEADLINE-2026.08 uses the candidate edge type as the primary key, and adds route, cargo type, settlement stage, contract conditions, or business event type restriction keys when there are business differences, covering all candidate edge types allowed to enter the supplementary documentation status in S3: 48 hours from the initial time of supplementary documentation request for performance events to document edges, 72 hours from the initial time of supplementary documentation request for document edges to invoice edges, 72 hours from the initial time of supplementary documentation request for invoice edges to payment edges, 24 hours from the initial time of supplementary documentation request for transportation events to insurance edges, and 120 hours from the initial time of supplementary documentation request for claims to payment edges. h; When the contract, order, policy, or claims rules have given the corresponding business period, the deadline for supplementary certification is the earlier of the business period and the time calculated according to the configuration period. If the business period is earlier than the first time for supplementary certification, the edge is immediately written into the gate status recalculation queue. The configuration table records the period value, unit h, applicable limit key, effective start and end time, basis file and version line by line. The gate status transaction is uniquely matched according to the edge type and applicable limit key. If the configuration is missing, multiple lines are hit, the period is not positive, the unit is wrong, or there is no effective configuration for the newly added edge type, the gate status update of the edge is frozen, the original status is retained and written into the configuration repair queue. The default time limit without source is not used. For blocking edges, the platform records the specific blocking reasons that have not occurred, the contradiction does not point to the target, or the time sequence is unreachable. When missing evidence is added, the version of the evidence corresponding to the blocking reason changes, or the deadline for supplementary certification expires, the candidate edge identifier is written into the gate status recalculation queue. The edge for supplementary certification that has not been supplemented by the deadline is turned into a blocking edge.

[0099] The records for evidence to be supplemented are structured into four categories: missing necessary evidence type, insufficient number of independent sources, missing target field, and missing event time field. Multiple categories of reasons can be recorded simultaneously. Each category of reason stores the type or field to be supplemented, the current value, the required value, the initial time of supplementation, and the deadline for supplementation of the same candidate edge. When the deadline for supplementation expires, the state predicate of the entire department is re-executed: if any type of missing necessary evidence is still not obtained, it is switched to blocking and a necessary evidence missing timeout is written; if the number of independent sources is still lower than the edge type requirement, it is switched to blocking and an independent source insufficiency timeout is written; if the target is still unknown, it is switched to blocking and a target-pointing evidence missing timeout is written; if the event time is still missing or the time sequence status is still unknown, it is switched to blocking and a time sequence evidence missing timeout is written. If multiple conditions are not met simultaneously, the first one is selected as the main blocking reason in a fixed order of necessary evidence, independent source, target, and time sequence, and the list of remaining reasons is retained. Only when all pending items are met will the activation, pending certificate, or blocking be re-determined based on the updated business status, completeness, direction, and timing. The initial pending certificate time or the certificate deadline will not be extended.

[0100] The gate status update adopts version comparison and is submitted serially according to the candidate edge identifier. Only the last confirmed result is retained for the same edge in an evaluation batch. After the evidence is supplemented, the completeness of the necessary evidence is recalculated and the number of independent sources is checked again. After the evidence corresponding to the blocking cause is updated, the latest valid version is read again. The event of the deadline for supplementing evidence is executed according to the above four categories of items to be supplemented and the final state is unique. If the recalculation result is the same as the previous version, only the timestamp is updated and S4 repeated propagation is not triggered.

[0101] Before committing a graph transaction, each candidate edge is checked to ensure it has a unique source node and target node, that the existence support record and contradictory channel record are valid, and that the gate status is only active, pending verification, or blocked. If the check fails, the current batch update of the candidate edge is revoked, and the previous valid version is retained. Edges that pass the check are indexed together by anchor point identifier and candidate edge identifier, which facilitates S5 to locate gate status and controlled path from change evidence.

[0102] The existence support record stores the set of evidence version identifiers, independent source groups, and supporting fields actually read when the business has occurred or a conclusion is pending. The direction determination record stores the set of evidence version identifiers read when the target judgment is made. The time-series reachability record stores the set of evidence version identifiers for the comparison of the time of the participating events and the reference partial order mapping. All three sets exclude corroborating evidence that did not participate in the current judgment and are fixed with the gate state version. The candidate edge record takes the union of the above three sets of judgment evidence versions to form the gate state judgment evidence version set, and stores it separately from the set of versions that generate contradictory evidence. When any judgment evidence version is missing, invalid, or its source independence changes, even if the version does not generate contradictory edges, the corresponding existence support record, direction determination record, time-series reachability record, and candidate edge must be written into the gate state recalculation queue.

[0103] The gating propagation graph retains all candidate directed propagation edges but distinguishes between active, pending verification, and blocked states using the gate state field. Node records the source graph identifier and semantic type, while edge records the existence support record identifier, contradictory evidence vector identifier, temporal reachability record identifier, gate state judgment evidence version set, gate state version, judgment time, missing evidence, and blocking reason. After the graph transaction is committed, a snapshot of the gating propagation graph and the occurrence status table of the bill business are output. S4 only queries edges with an active gate state, while pending verification and blocked edges are retained for incremental recalculation in S5.

[0104] In this specific embodiment, S4 includes:

[0105] The risk propagation module reads the S3 gated propagation graph and the S2 typified conflict evidence, and establishes a risk message for each conflict source node. The message records the risk message identifier, source node, conflict type, target, residual absolute magnitude, residual boundary excess, partial order conflict level, and evidence version set. The risk propagation parameter table version RISK-PROP-2026.08 provides a dimensionless baseline strength for all eight types of conflicts: missing 0.45, reverse order 0.70, timeout 0.55, spatial inconsistency 0.60, cargo quantity inconsistency 0.65, amount inconsistency 0.70, tax amount inconsistency 0.70, and insufficient insurance coverage 0.75. The values ​​range from 0 to 1. If the conflict type is not in the enumeration or no unique valid parameter can be found, the risk message is rejected and a configuration error is recorded. If the conflict does not point to a candidate edge target node, no message is created for that target.

[0106] Risk information Calculate the initial risk intensity , This indicates a risk message. The dimensionless initial risk intensity, with a larger value indicating a higher risk of source conflict. This represents the message index in the risk message table of the currently running batch, and the traversal range is from the first to the last message in the table. Since this is a discrete index rather than a physical quantity, units are not applicable. Indicates from risk information The conflict type field obtains category symbols, and its member range includes only eight categories: missing, reverse order, timeout, spatial inconsistency, quantity inconsistency, amount inconsistency, tax inconsistency, and insufficient insurance coverage. Since it is a discrete category rather than a physical quantity, units are not applicable. Indicates by The dimensionless baseline strength obtained from version RISK-PROP-2026.08 ranges from 0 to 1; there are inconsistencies in the values ​​for space, cargo volume, value, tax, and insurance coverage. Take the non-negative difference obtained by subtracting the absolute value of the corresponding boundary from the absolute value of the residual, divide it by the absolute value of the corresponding boundary, and truncate it to 0 to 1. If the boundary value is zero, refuse to generate a message; for missing contradictions... The value is set to 0.70. For inversion contradictions, the values ​​are 0.60, 0.80, and 1.00 respectively, based on the inversion path crossing one, two, and at least three necessary events. For timeout contradictions, the values ​​are 0.60, 0.80, and 1.00 respectively, based on the proportion exceeding the maximum normal duration: not exceeding 0.25, exceeding 0.25 but not exceeding 0.50, and exceeding 0.50. The mapping and baseline strength used are both from version RISK-PROP-2026.08. It is a dimensionless quantity and its value ranges from 0 to 1. This indicates that the product will be limited to 0 to 1; this risk message will not be generated if the residual is not out of bounds and there are no non-numerical conflicts.

[0107] For each risk message and its path consisting of activated edges Calculate risk path contribution , This indicates a risk message. Along the path The resulting dimensionless cumulative risk path contribution ranges from 0 to 1. This indicates the message index in the risk message table of the currently running batch, and its range includes all messages in that table. This represents all valid paths whose active path indices are obtained by enumerating the gated propagation graph from the current anchor point, and whose range is from one to six edges. Representing a path The activation candidate edge indices are traversed sequentially, and the range includes all edges of the path. , and These are all discrete indices rather than physical quantities, therefore units are not applicable. Indicates the initial risk intensity. This represents the completeness of dimensionless contradictory evidence, with values ​​ranging from 0 to 1. This represents the arithmetic mean of the confidence levels of dimensionless association assertion nodes, with a value ranging from 0 to 1. This represents a dimensionless evidence time decay factor with a value ranging from 0 to 1. All four factors only apply to messages. Multiply by one, Representing an edge The dimensionless risk propagation weight has a value range of 0 to 1. Indicates path The weights of all edges in the table are producted sequentially; the target-pointing evidence type configuration table TARGET-EVIDENCE-2026.08 uses the contradiction type and the semantic type of the target node as the joint primary key, and adds a limit key according to the route, cargo type or settlement stage with business differences, to record the required set of target-pointing evidence types, the effective start and end time and the configuration version; The numerator only counts the number of different types of evidence obtained within the same performance evidence anchor point, valid at the evaluation time, with consistent target node identification, and excluding duplicates, transfers, or derivatives. The denominator takes the number of different types of evidence required in the unique and valid configuration row. If the configuration is missing, the hit is not unique, the set is empty, or the denominator is not positive, the corresponding risk path contribution is rejected and written to the parameter repair queue. Only the credibility of the above valid and deduplicated associated assertion nodes is averaged. When the set of associated assertion nodes is empty, path contribution is also rejected and assertion is missing. The default value of an empty set is not used. The configuration version is written to the risk message, the running batch and the dependency index.

[0108] Evidence time decay factor calculate, This indicates a risk message. The dimensionless evidence of the time decay factor, This represents the index of discrete messages in the risk message table of the currently running batch, and the traversal range is all messages in the table. Since it is not a physical quantity, the unit is not applicable. Indicates assessment time and risk information The assertion uses a non-negative time difference between the effective times of the nodes, and the unit is days (d). The half-life corresponding to the contradiction type is expressed in days (d) and must be strictly greater than 0. This represents the natural exponential function. Represents the natural logarithm function; the set of assertion nodes used and the computation The set of valid and deduplicated association assertion nodes used is exactly the same; when the set contains multiple nodes, nodes that have expired at the evaluation time, have not yet taken effect, or have missing effective time fields are first removed, and then the node with the earliest effective time is selected for calculation. If the effective times are the same, the first one is selected according to the lexicographical order of the assertion node identifier, and the selected assertion node identifier and effective time are written into the risk message and dependency index. If the set is empty after removal, the node's effective time is later than the evaluation time, or a non-negative time difference cannot be formed, the corresponding risk path contribution is rejected and an assertion time error is recorded, and the default time is not used. When loading the risk propagation parameter table, the half-life unit and positive value are checked at the same time. Risk path contributions to the message will be rejected if the value is missing, zero, or negative, and parameter configuration errors will be recorded. Negative, zero, or unsourced default values ​​will not be used for fallback. The half-life for payment and invoice-related conflicts will be 90 days, and the half-life for trajectory and port event-related conflicts will be 30 days. d; The edge type weight configuration table of version RISK-PROP-2026.08 uses edge type as the primary key and adds route, cargo type, settlement stage, and validity period key when there are business differences. It covers all candidate edge types that S3 allows to enter the propagation. The weight of the edge from performance event to document is 0.80, the weight of the edge from document to invoice is 0.80, the weight of the edge from invoice to payment is 0.85, the weight of the edge from transportation event to insurance is 0.75, and the weight of the edge from claims to payment is 0.80. All weights are dimensionless and range from 0 to 1. Configure the source sample version, calibration date, effective start and end time and configuration version of the line record. The propagation transaction is uniquely matched according to the edge type and applicable key. When the configuration is missing, multiple rows are hit, the weight is out of bounds, or the newly added edge type has no effective configuration, the edge cannot be activated or participate in the path enumeration, and the corresponding object is written to the parameter repair queue. The above half-life and edge weight are calibrated by the historical confirmed risk sample through time segmentation validation set, and are updated every quarter or when the risk hit rate drops by more than 3 percentage points.

[0109] Risk information Calculate the propagation gating coefficient , This indicates a risk message. The dimensionless propagation gating coefficient, with values ​​ranging from 0 to 1. This represents the index of discrete messages in the risk message table of the currently running batch, and the traversal range is all messages in the table. Since it is not a physical quantity, the unit is not applicable. This indicates the completeness of the aforementioned dimensionless contradictory evidence, with a value ranging from 0 to 1. This represents the average confidence level of the aforementioned dimensionless correlation assertion nodes, with a value ranging from 0 to 1. This represents a dimensionless evidence time decay factor, ranging from 0 to 1. Larger values ​​for the three factors indicate more complete evidence pointing to the target, a more reliable data source, and evidence closer to the evaluation time, respectively. Therefore, a larger product allows for a stronger risk message to be transmitted. When any factor is zero, the risk path contribution of that message on the currently active edge is zero. The edge type risk transmission weight only adjusts the transmission ratio of different business relationships without changing the aforementioned monotonic direction. The platform will... The risk message is written to the record and read directly by the path contribution calculator, as mentioned above. The calculation uses the already saved Equivalent substitution product That is, first with take Then multiply by each of the items in the path in sequence. You may not board again at the same time. , or Path records and dependency indexes are stored simultaneously. And its three source factors, verification before transaction commit. Consistent with the three-factor product at six decimal places, if inconsistent, the path is rejected and entered into the risk message recalculation queue, thereby enabling the propagation gating coefficient to participate in the contribution of each downstream risk path, the target risk value, and the S5 local recalculation;

[0110] When the typified conflict evidence is a numerical residual out-of-bounds error, the message generator selects the invoice, payment, or insurance target node pointed to by the conflict according to the out-of-bounds direction, and forms the initial risk intensity with the out-of-bounds ratio; when the conflict is a reverse-order conflict, the subsequent event node with reference to the partial order of performance is used as the target and the reverse-order baseline intensity is used; when the conflict is missing or timed out, the message is only established when the planned completion time has been reached and the semantic type of the target node is clear. When the target node is missing or still in the observation period, no propagable message is established and the record to be evaluated is retained.

[0111] Risk messages and propagation edges employ a two-level deterministic binding. The first hop only allows the selection of active edges where the source node identifier equals the message source node, the target node identifier equals the message target, and the candidate edge's saved contradiction evidence vector identifier or source contradiction evidence identifier is completely consistent with the message source contradiction identifier. If any field is not satisfied, the edge will not enter the message's first hop set. After a message reaches the first target node, it is not regenerated or rewritten based on other contradictions at that node. Instead, the original risk message identifier, original source contradiction identifier, and original message-level factor are retained, and the currently arrived node is used as the next hop source node to continue propagating along downstream business relationships. The routing configuration table in version RISK-ROUTE-2026.08 uses the original contradiction type, current node semantic type, next edge type, and next target node semantic type as a composite key, recording the allowed propagation flag, maximum remaining depth, effective start and end times, and configuration version. Subsequent hops only allow the selection of edges where the source node equals the current node, the gate state is active, the time is not reversed, and the edge uniquely matches the allowed propagation configuration. If the configuration is missing, multiple rows are matched, or the allowed flag is negative, the branch is terminated, and other contradictory messages associated with that edge are not used. For each path, write the risk message identifier, the original source contradiction identifier, the matching first-hop contradiction evidence vector identifier, the gate status identifier, the route configuration version, and the edge identifier hop by hop, and write these fields into the dependency index.

[0112] The risk propagation queue uses the source event time, path depth, and risk message identifier as sorting keys. Messages with earlier source event times are processed first. Messages with the same time are processed in ascending order of path depth. If they are still the same, they are processed in lexicographical order of risk message identifier. After retrieving a message, the queue enumerates outgoing edges that are active in the current gate state version and whose target event time is not earlier than the source event time, based on the first or subsequent hop binding conditions mentioned above. If the outgoing edge set is empty, the current node is recorded as the path endpoint. If the outgoing edge state or routing configuration changes before the propagation transaction is committed, the branch is canceled and re-enqueued with the new gate state and new routing configuration.

[0113] For each new path generated, the platform immediately verifies that its anchor point identifier is consistent with the source message anchor point, the first and last nodes of adjacent edges are continuous, the version of the cited evidence is valid at the time of evaluation, and the cumulative path contribution does not exceed the parent path contribution. If any condition is not met, the candidate path is deleted and written into the propagation rejection record. Only when all conditions are met is the path contribution written into the target node's summary buffer. After the target risk value is calculated, the buffer is cleared and the path contribution is transferred to the dependency index.

[0114] The propagator traverses the active edges in the gated propagation graph in ascending order of event time. The same risk message must not pass through the same candidate edge identifier repeatedly. The maximum path length is set to 6 edges. After a message reaches an intermediate node, the cumulative contribution of the parent path is multiplied by the edge type weight of the next active edge. The initial risk intensity, completeness of contradictory evidence, credibility of data source, or evidence time decay factor are not multiplied again. This is consistent with the definition that the message-level factor is multiplied only once in the risk path contribution formula. The parent risk message identifier, candidate edge identifier, and cumulative path contribution are recorded. The branch is terminated when an edge to be supplemented, a blocking edge, an edge that has been visited, or the maximum path length is exceeded. Multiple source messages are allowed to reach the same target node, but their respective path identifiers are retained.

[0115] When the same target node receives multiple risk messages, the platform first deduplicates them by source conflict evidence identifier, evidence version, and complete edge sequence, and then accumulates the risk path contribution in ascending order by path identifier. When two paths have the same source evidence and the same edge sequence but different evidence versions, the version with the latest effective time is retained. When the path contributions are exactly the same, the priority selection is not performed, but evidence paths from different sources are retained respectively, and the sum is uniformly calculated by the target risk value formula.

[0116] The historical performance characteristics module constructs baseline characteristics based on the target entity's closed performance business over the past 180 days, including on-time arrival rate, document consistency rate, payment overdue rate, invoice red-inking rate, insurance claim rate, and confirmed anomaly rate. Each ratio is calculated with the number of valid business tickets as the denominator and smoothed using Bayesian smoothing with 20 tickets. The platform maps each ratio to a risk direction of 0 to 1, sums them according to the fixed weights of the risk control parameter version HIS-2026.08, and truncates them to 0 to 1 to obtain the baseline risk value. When the sample is less than 20 tickets, the overall prior of the same entity type and the same flight area is used, and the fallback source is recorded in the baseline source field. The baseline characteristics are updated offline daily, and the version that was in effect at the time of assessment is used in this propagation.

[0117] For the target node Calculate risk value , Represents the target node The dimensionless risk value, with a larger value indicating a higher risk. This indicates the node index in the target node table where the current anchor point is affected by the risk message, and the traversal range is all nodes in that table. Since this is a discrete index rather than a physical quantity, units are not applicable. Represents the target node The corresponding entity's dimensionless baseline risk value, ranging from 0 to 1, is derived from the historical performance profile version of HIS-2026.08, which was in effect at the time of assessment. Indicates arrival at the target node The set of effective risk paths, Represents a set The path indices are traversed in ascending order by path identifier, and the range includes all members of the set. Since this is a discrete index rather than a physical quantity, units are not applicable. Representing a path The cumulative risk path contribution is a non-negative dimensionless number. Represents a set Contribution of all effective risk paths in Summing in ascending order by path identifier This means that the summation value will be limited to 0 to 1; when the same contradictory evidence reaches the same goal through the exact same edge sequence, only the path with the most recent version of the evidence will be retained to avoid double counting;

[0118] The platform will mark target nodes that arrive at and receive positive risk path contributions from at least one active propagation path as affected nodes. When the node risk value is greater than or equal to 0.70, it will be written into the high-risk queue; when it is greater than or equal to 0.40 and less than 0.70, it will be written into the attention queue; when it is less than 0.40, only the profile will be updated. When the affected node is a payment node or an insurance node, its desensitized value of the semantic type of the fund node or insurance node, account or policy identifier and associated entity identifier will be retained, and it will not be rewritten as an ordinary performance event node due to risk propagation.

[0119] The system writes the abnormal performance chain and traceable evidence path according to the candidate directed propagation edges through which each risk message passes. The path record includes source contradictory evidence, evidence version, gate state version, edge identifier, node identifier, unilateral contribution, and cumulative contribution in sequence. The system performs a union-query set on the subject nodes of the shared control, transaction, or historical performance relationship edges in the abnormal performance chain and forms a risk subject association group. The subject relationship edges to be judged do not participate in the merging. The association group records the group identifier, member subjects, shared relationship edges, highest risk node, and generation batch.

[0120] The dependency index uses the evidence version identifier as the first index key and the contradictory edge identifier, propagation gate identifier, traceable evidence path identifier, and target node identifier as ordered fields. It records the risk increment of each evidence version to each path as the risk path contribution. The risk path contribution of the same path is summarized as the path cumulative risk value. The path cumulative risk value reaching the same target node and the baseline risk value together form the target risk value. All addends are saved to six decimal places and recalculated with the same propagation parameter version to ensure that S5 can be accurately revoked according to the evidence version.

[0121] The calculation of the completeness of contradictory evidence further limits the set of evidence types to be obtained: the platform first obtains the set of types from evidence that is within the same performance evidence anchor point, is valid at the time of evaluation, has the same target node identifier, and has excluded duplicates, transfers, and derivatives, and then only retains members of the set of evidence types required by the target pointing to the unique valid configuration line of TARGET-EVIDENCE-2026.08; evidence types outside the configuration set are not allowed to enter the numerator. The numerator is the number of different types found in the intersection of the obtained set and the desired set, and the denominator is the number of different types found in the desired set. Therefore, the absence of any desired type cannot be compensated for by irrelevant types; verification is performed after calculation. If the value is between 0 and 1, or if the result is out of bounds, the required set is empty, or the denominator is not positive, the path contribution will be rejected and written to the parameter repair queue.

[0122] The Historical Performance Profile Parameter Table HIS-2026.08 uses subject type, flight area, and assessment time as unique search keys to record the number of prior successful tickets, the number of prior failed tickets, risk direction, dimensionless weight, effective start and end time, and version of six characteristics. The number of prior successful and failed tickets for on-time arrival is set to 18 and 2, respectively; the document consistency rate is set to 19 and 1, respectively; the payment overdue rate, invoice red-inking rate, insurance claim rate, and confirmed anomaly rate are all set to 1 and 19, respectively; and the prior strength of all six items is 20 tickets. Each smoothing ratio is calculated with the sum of historical successful votes and prior successful votes as the numerator, and the sum of valid business votes and prior successful and failed votes as the denominator. On-time arrival rate and document consistency rate are positive indicators, with the risk characteristic calculated as 1 minus the smoothing ratio. The other four indicators are negative indicators, with the risk characteristic directly calculated as the smoothing ratio. The weights of the six indicators are 0.20, 0.20, 0.20, 0.15, 0.10, and 0.15 respectively, and the sum of the weights must be 1. The platform calculates the weighted sum in this order and truncates it to 0 to 1 only after the summation is complete. When the number of subject samples is less than 20, the prior smoothing method described above will still be applied, and prior lines of the same subject type and flight area will be used as the only rollback line. If there are missing configurations, multiple hits, negative prior counts, out-of-bounds weights, weight sums not equal to 1, or risk directions not within the positive and negative enumerations, the subject risk update will be frozen and written to the profile parameter repair queue. The configuration version, rollback level, and six smoothing ratios used will be considered. Write the runtime batch and dependent indexes;

[0123] For a risk message that relies on multiple versions of evidence, the platform first takes the union of contradictory evidence and the set of evidence versions actually used in the risk message with the set of gate state determination evidence versions stored in each active edge within the path, forming a complete set of dependent evidence versions contributing to the risk path. Then, it generates a unique risk path contribution identifier using the risk message identifier, the complete candidate edge sequence, the sorted complete set of dependent evidence versions, the gate state version, and the risk propagation parameter version. The contribution is saved only once in the risk path contribution table. For each evidence version key in the complete dependency evidence version set, the dependency index only stores references to the same contribution identifier, corresponding existence support record, direction determination record, timing reachability record, and propagation gate identifier, without copying contribution values ​​to form new addable items. Both the cumulative path value and the target risk value are deduplicated by contribution identifier before aggregation. When S5 queries by any failed evidence version, it also first deduplicates by contribution identifier, then sets the unique contribution record as pending recalculation and deducts it only once. When multiple dependency versions fail simultaneously in the same transaction, only one deduction item is generated. Other versions that are not failed but depend on the contribution retain index references and are marked as waiting for new contributions. The transaction log records the complete dependency evidence version set, contribution identifier, number of references before deduplication, number of contributions after deduplication, and actual deduction value. Before committing, it verifies that the contribution identifiers in the target aggregation, index references, and deduction set are all unique.

[0124] This step outputs an abnormal performance chain list, a risk subject association group list, an affected node list, a traceable evidence path list, a target node risk list, and a dependency index snapshot. It also writes the assessment time, anchor point identifier, reference performance partial order version, necessary evidence set version, risk propagation parameter version, and historical performance profile version into the running batch record. These outputs are used to generate the current warning, and S5 can locate objects that need to be invalidated, deducted, or partially recalculated when revising documents, reversing payments, reversing invoices, amending policies, or re-transmitting tracking data.

[0125] In this specific embodiment, S5 includes:

[0126] The evidence change monitoring module subscribes to document revision, payment reversal, invoice red-ink reversal, insurance policy amendment, and trajectory retransmission events. Each change event includes at least a change event identifier, change type, original evidence version identifier, new original record identifier, change effective time, and source system. The platform uses the change event identifier as an idempotent key, sets the expiration time of the original evidence version to the change effective time in the same graph transaction, calculates the new content summary, and generates a monotonically increasing new evidence version. The original version node and source derivation relationship are retained without deletion, thus enabling the assertions, trajectory facts, and their sources before and after the change to be traceable.

[0127] After the new evidence version is generated, the platform first performs an upstream key field difference determination, comparing the semantic type, subject definiteness identifier, vessel MMSI, document number, order identifier, waybill identifier, cargo batch identifier, voyage identifier, port of departure, port of destination, planned performance window, and settlement stage between the old and new versions item by item. If none of the above fields have changed, the original standard entity and the original performance evidence anchor point attribution are retained, and partial recalculation is performed along the existing S2 to S4. If any field changes, the original attribution cannot be directly inherited, but the SEM-WATER-2026.08 semantic mapping, ENTITY-ALIGN-2026.08 entity alignment, and anchor point combination field hashing of S1 are re-executed for the new version. In the same version change transaction, the attribution relationship between the original evidence version and the old standard entity and the old anchor point is set with an expiration time, the old relationship record is retained, and old anchor point cancellation tasks and new anchor point recalculation tasks are created separately. The old anchor point cancellation task is performed by deduplicating the path contribution of the original version based on the contribution identifier according to the dependency index, and then re-executing S2 to S4 with the remaining valid evidence of the old anchor point; the new anchor point recalculation task generates trajectory facts or assertions from the new attribution record, and then performs S2 contradiction detection, S3 gate status determination and S4 risk propagation. If the entity alignment enters the pending review stage, the deterministic identifier conflicts, the anchor point combination field is incomplete, or there are multiple new anchor point candidates, the attribution is not migrated or a new warning is issued, the relevant warning updates of the old anchor point and all new candidate anchor points are frozen and written to the manual review queue, and the queue records the change event identifier, the old entity and the old anchor point, all new candidate entities and anchor points, unresolved fields and evidence versions; after manual confirmation, the above dual-task process continues with the confirmation result;

[0128] The incremental updater queries the S4 dependency index using the version identifier of the failed evidence to obtain the conflict edge identifier, the propagation gate identifier controlled by the conflict edge, the traceable evidence path identifier using the propagation gate, the target node identifier, and the risk path contribution identifier reference. The query results are first deduplicated according to the risk path contribution identifier. Even if the same contribution identifier is hit by multiple version indexes in the failed evidence set, only one corresponding per-path risk path contribution is obtained, only one path and gate status is set to pending recalculation, and only one deduction item is formed. When the index item is missing, the referenced contribution identifier does not exist, or the same contribution identifier corresponds to multiple contribution records, a full graph scan or deduction is not performed. Instead, the corresponding performance evidence anchor point is written to the index repair queue and the warning update of the anchor point is frozen. When the index is complete and the contribution identifier is unique, the old running batch number, the number of references before deduplication, and the number of contributions after deduplication are recorded in the transaction log.

[0129] For each affected target node Transient risk value after deducting evidence of failed reorganization , This represents the dimensionless transient risk value obtained solely from the reconstruction of the baseline risk value of the same version and contributions from non-failed paths. This indicates the node indexes in the affected target node table obtained from this index-dependent query, and the range includes all nodes in that table. Since this is a discrete index rather than a physical quantity, units are not applicable. Indicates the target node saved in the original running batch. The corresponding dimensionless baseline risk value for the subject, with a value ranging from 0 to 1. This represents the set of paths in the original path set that remain valid after excluding all dependent invalid evidence versions and have been deduplicated according to the source contradictory evidence, evidence version, and complete edge sequence. Represents a set The path indices are traversed in ascending order by path identifier, and the range includes all members of the set. Since this is a discrete index rather than a physical quantity, units are not applicable. Indicates the path to be saved based on the index. The non-negative dimensionless cumulative contribution, Represents a set Contribution of all non-invalid and deduplicated paths in Summing in ascending order by path identifier This means that the result will be limited to 0 to 1 only after the baseline value and all retained contributions have been summed without truncation; This indicates that the evidence was truncated to the target node between 0 and 1 before the change. Dimensionless risk value; the platform does not use the already truncated value. Reverse deduction: when there are multiple failed paths, deduplication is performed first and then exclusion is performed. All contributions are retained to six decimal places and rounded off after summation.

[0130] The recalculation set is bounded by the affected performance evidence anchor points. First, it adds the changed evidence nodes, assertion nodes with source-derived or field-assertive relationships, and trajectory physical facts. Then, it adds conflicting edges generated by this version through dependency indexes, as well as existence support records, direction determination records, temporal reachability records, propagation gates, backtrackable paths using these gates, and target nodes referenced when this version is only used as gate state determination evidence. Even if the changed evidence does not generate conflicting edges, as long as it appears in the gate state determination evidence version set, S3's support, direction, temporal, and gate state determinations must be re-executed, and S4 recalculation must be performed on existing contributions controlled by that gate. The main relationship graph is read only by reference and does not extend to other anchor points, unless the same evidence version is referenced by multiple anchor points simultaneously. In this case, an independent recalculation task is created for each referenced anchor point, with task fields recording the anchor point identifier, change event identifier, node set, edge set, and old run batch number.

[0131] When the platform receives an invoice reversal event with a negative amount of 8.2 million yuan, it first invalidates the original invoice evidence version. Then, it uses the dependency index to locate the invoice-generated amount contradiction edge, the invoice-to-payment propagation gate, and the two evidence paths leading to the payment node. If there is no currently valid invoice after the reversal, but there are still bills of lading, receipts, or payment records to support that the transaction has occurred and the deadline for supplementary documentation has not expired, the corresponding transaction status is changed to pending supplementary documentation, and the record shows the missing valid invoice and the deadline for supplementary documentation. If there are no valid records to confirm that the transaction has occurred, it is changed to "not occurred". If the direction is mismatched, the timing is clearly unreachable, or a valid invoice has not been added by the deadline for supplementary documentation, the corresponding edge is changed to "blocked". The contribution of the old path is deducted according to the index-preserved value before recalculation. Other trajectories, insurance, and payment paths that do not depend on the invoice are not added to the recalculation set, thereby verifying the local boundary, deduction order, and gate status update rules.

[0132] The local calculator re-executes S2's trajectory fact or assertion generation, event missing / reverse timeout detection, and multidimensional consistency residual calculation on the recalculated set, and then executes S3's existence support, contradictory target pointing, event temporal reachability, and gate state determination; the triggering conditions for S4 local propagation include newly generated active edges, active edges whose gate state or edge attributes have changed, contradictory amplitudes of risk messages or changes in any message-level factor, and any unique risk path contribution identifier being set to be recalculated due to the invalidation of the dependent evidence version. The unchanged state of candidate edges does not constitute a reason for skipping. For each contribution identifier to be recalculated, if the updated source contradiction still exists and every edge in the original complete candidate edge sequence is active in the new gate state version, then the updated initial risk intensity, contradiction evidence completeness, correlation assertion confidence, time decay factor, and complete dependency evidence version set are used along the original edge sequence, and the baseline and risk propagation parameter versions locked in the original batch are recalculated to generate a new unique risk path contribution identifier and contribution record; if the source contradiction disappears, the target direction changes, or any edge in the original sequence becomes pending verification or blocked, then the old contribution is not reconstructed, and local paths are enumerated separately according to the new target direction or active edge. Contradictory edges, propagation gates, and traceable paths that have not entered the recalculation set retain their original versions and original risk contributions, while edges pending verification and blocked edges are only updated in terms of status, reason, and deadline without transmitting risk messages;

[0133] The platform uses the same baseline risk version and risk propagation parameter version as the original batch, performs an untruncation summation of the baseline risk value of the same version, the contribution of all valid paths that have not failed, and the contribution of new paths obtained from local propagation, and only truncates it to 0 to 1 when generating updated risk values ​​externally; if the parameter version has been changed, this change will still use the old version to complete the evidence difference calculation, and a separate parameter change batch will be established to avoid mixing evidence changes and model parameter changes into the same early warning interpretation; when the new path and the retained path have the same source contradictory evidence, evidence version, and complete edge sequence, only the one with the latest effective time is retained, the empty path set is treated as zero contribution, and the larger the updated risk value, the higher the risk;

[0134] The evidence difference module compares the standardized content summary, validity period, field values, units, and source identifiers of evidence before and after the change, forming field-level added, removed, and changed sets, and compares the propagation gate status and target node risk value before and after the change; for the target node... Record risk value differences , Represents the target node The difference in signed dimensionless risk values, This indicates the node index in the affected target node table that is being recalculated, and the traversal range includes all nodes in that table. Since this is a discrete index rather than a physical quantity, units are not applicable. This represents the dimensionless risk value after evidence changes and partial recalculation. This represents the dimensionless risk value before the change; a positive value indicates an increase in risk, a negative value indicates a decrease in risk, and zero indicates that the risk remains unchanged.

[0135] The warning interpretation record uses the change event identifier and anchor point identifier as a joint key, and records the evidence version before the change, the evidence version after the change, the difference in content summary, the invalid contradictory edge, the newly added or updated contradictory edge, the difference in the propagation gate status, the contribution of the deducted risk path, the contribution of the newly added risk path, the risk value before and after the change, and the difference in risk value. When there is still evidence of other business transactions after the invoice is reversed and the deadline for supplementing the certificate has not expired, the interpretation record records the invoice to payment edge as pending supplementation and lists the missing valid invoice, the deadline for supplementation, and the old path identifier. Only when the business has not occurred, the direction is mismatched, the timing is clearly unreachable, or the deadline for supplementation has expired and there is still a missing valid invoice, the edge is recorded as blocked, and the unique blocking condition and the corresponding risk reduction amount are recorded. Business personnel can trace back the original evidence along the path identifier.

[0136] Before the transaction is committed, the updater verifies that the total amount deducted from the old contribution is consistent with the total amount of the dependent index, the sum of the new contribution is consistent with the new path record, each propagation path contains only active edges and the recalculated objects are all affected anchors. If the verification fails, the local update is rolled back and the original warning result is retained. If the verification passes, the new evidence version, gate state version, risk path, target risk value, dependent index and warning explanation record are atomically written, and the change event is marked as processed. If the same change event is received repeatedly, the existing update batch is returned directly.

[0137] This step ultimately outputs a partial snapshot of the updated gated propagation graph, a target node risk table, an evidence version dependency index, and a warning explanation record. It also records the actual number of recalculated nodes, the number of recalculated edges, and the total number of nodes and edges in the original graph as statistics of the processing range. The platform publishes risk values, risk rise and fall directions, affected funds or insurance nodes, evidence differences, and traceable paths to the risk control terminal. The performance evidence anchors, propagation gate status, and risk path contributions that are not dependent on this evidence version remain unchanged.

[0138] 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.

[0139] This invention transforms multi-source performance data into graph data organized with performance evidence anchors through unified semantic mapping and entity alignment. Then, it generates typified contradictory evidence by referencing performance partial order and multidimensional consistency residuals, enabling the influence relationships between subject associations, performance anomalies, and funding and insurance nodes to be identified and traced by the same processing chain.

[0140] This invention addresses the characteristics of delayed, supplementary, and corrected evidence in the water transport supply chain by adding an existence support channel and a consistency conflict channel to the propagation structure, and constraining risk messages with propagation gate states. At the same time, it utilizes evidence version dependency indexes to revoke the path contribution of invalid evidence and limits the local recalculation set, thereby reducing the amount of false propagation and full graph update processing caused by old evidence while maintaining the traceability of evidence.

Claims

1. A knowledge graph-based water transport risk control method, characterized in that, include: S1. Acquire multi-source business data of the water transport supply chain and perform unified semantic mapping and entity alignment. Based on the alignment results, generate performance evidence anchor points, subject relationship diagrams and performance event evidence diagrams to obtain anchored diagram data. S2. Generate trajectory physical facts and evidence assertions based on anchor point graph data. Within the same performance evidence anchor point, detect conflicts according to the reference performance partial order and calculate multidimensional consistency residuals to obtain typified contradictory evidence and performance anomaly subgraphs. S3. Align the performance anomaly subgraph with the subject relationship graph, establish existence support channels and consistency contradiction channels for candidate directed propagation edges respectively, and output the business existence status, necessary evidence completeness and contradiction evidence vector pointing to the target node; determine the propagation gate state based on the output and event time-series reachability to obtain the gated propagation graph; S4. Generate initial risk messages based on the typified contradictory evidence associated with the source node. In the gated propagation graph, risk messages are only transmitted along the activated candidate directed propagation edges. Combined with historical performance characteristics, abnormal performance chains, risk subject association groups, affected nodes, and traceable evidence paths are obtained. At the same time, a dependency index is established between evidence version, contradictory edges, propagation gate status, and risk path contribution. S5. In response to evidence changes, based on the invalidation of the dependent index and the replacement of the evidence, and deducting the corresponding risk path contribution, only the associated performance evidence anchor subgraph is recalculated to obtain the updated propagation gate state, risk value, and evidence difference.

2. The knowledge graph-based water transport risk control method according to claim 1, characterized in that, S1 includes: Acquire orders, waybills, vessel AIS trajectories, BeiDou trajectories, port events, transport documents, payment records, invoices, insurance policies, claims records, and historical performance records; map fields from each data source to a unified semantic type for entities, business objects, performance events, documents, fund records, and insurance records; and merge entities pointing to the same object based on entity identifiers, vessel identifiers, document identifiers, temporal proximity relationships, and port location relationships; generate the performance evidence anchor points using a combination of identifiers for orders, waybills, cargo batches, vessels, voyages, ports of origin and destination, time windows, and settlement stages; The subject relationship graph is formed by combining the subject nodes, subject identifiers, control, transaction and historical performance relationships. The performance event evidence graph is formed by combining the performance event nodes, trajectory physical facts, assertion nodes and the relationship between the nodes and the performance evidence anchor points.

3. The knowledge graph-based water transport risk control method according to claim 1, characterized in that, S2 includes: Based on the time, location, speed, heading, and spatial relationship with the port geofence, the physical facts of the trajectory are generated for berthing, departure, stop, segment, and deviation. The field values ​​in the transport documents are converted into assertion nodes, and the data source, evidence version, effective time, expiration time, content summary and credibility are recorded for the assertion nodes. The reference performance partial order is constructed according to the sequential constraints between performance event types. Events in the same performance evidence anchor point are compared with the reference performance partial order to detect missing, reversed and timeout events, and the consistency residuals of time and space, cargo volume, amount, tax and insurance coverage are calculated. The detection results and consistency residuals are encoded into the typed contradictory evidence according to the conflict type.

4. The knowledge graph-based water transport risk control method according to claim 1, characterized in that, S3 includes: using the directed relationships between performance event nodes, document nodes, invoice nodes, payment nodes, and insurance nodes as candidate directed propagation edges; verifying the existence of records from independent data sources associated with the same performance evidence anchor point using the existence support channel, and outputting the business existence status and necessary evidence completeness; encoding the typed contradictory evidence according to the contradiction type, contradiction pointing node, residual direction, and residual magnitude using the consistency contradiction channel, and outputting the contradictory evidence vector; when the business existence status is "occurred" and the necessary evidence completeness is "full"... When the integrity condition of the corresponding edge type is met, the contradictory evidence vector points to the target node of the candidate directed propagation edge, and the target node is temporally reachable from the source node, the candidate directed propagation edge is set to an active state; when the business existence state is pending supplementary evidence, the necessary evidence integrity does not meet the integrity condition, the contradictory pointing evidence is missing, or the temporal information is missing, the candidate directed propagation edge is set to a pending supplementary evidence state; when the business existence state is not occurred, the contradictory evidence vector is determined not to point to the target node, or the target node is determined to be temporally unreachable, the candidate directed propagation edge is set to a blocked state.

5. The knowledge graph-based water transport risk control method according to claim 1, characterized in that, S4 includes: The initial risk message is generated based on the residual magnitude and partial order conflict type in the typified contradictory evidence associated with the source node; The ratio of the number of target-oriented evidence types obtained to the number of target-oriented evidence types required for the corresponding contradiction type is used as the completeness of the contradictory evidence. The average credibility of the assertion nodes associated with the contradictory evidence vector is used as the credibility of the data source. The evidence time decay factor is obtained by monotonically decreasing mapping based on the time difference between the evaluation time and the effective time of the assertion node. The product of the completeness of the contradictory evidence, the credibility of the data source, and the evidence time decay factor is used as the propagation gating coefficient. The propagation gating coefficient is used to adjust the risk message transmitted from the source node to the target node. A baseline risk value is generated based on the historical performance characteristics of the entity corresponding to the target node. The risk increments formed by each risk message received by the target node are summed and added to the baseline risk value to obtain the risk value of the target node. The target node that arrives from the activated propagation path and receives at least one risk message is identified as the affected node, and the semantic type of the affected node as a funding node or insurance node is retained; the abnormal performance chain and the traceable evidence path are generated according to the order of the directed propagation edges through which the risk message passes, and the risk subject association group is formed according to the nodes sharing the subject relationship edge in the abnormal performance chain.

6. The knowledge graph-based water transport risk control method according to claim 1, characterized in that, S5 includes: When document revision, payment reversal, invoice red-ink reversal, policy amendment, or trajectory retransmission occurs, the expiration time of the evidence before the change is set as the effective time of the change, and a new evidence version of the evidence after the change is generated; the contradictory edges, propagation gate states, and traceable evidence paths using the evidence before the change are determined based on the dependency index, and the risk path contribution corresponding to the evidence before the change is deducted; Using the affected performance evidence anchor points as boundaries, the event conflict detection, multidimensional consistency calculation, and propagation gate state determination are re-executed, and the early warning interpretation record is generated using the updated risk value and the difference in the content summary of the evidence before and after the change.

7. The knowledge graph-based water transport risk control method according to claim 4, characterized in that, The existence support channel configures a set of necessary evidence types for different candidate directed propagation edge types. The necessary evidence type set is intersected with the evidence types already obtained and aligned with entities within the same performance evidence anchor point. The ratio of the number of evidence types in the intersection to the number of evidence types in the necessary evidence type set is used as the completeness of the necessary evidence. Independent data sources refer to evidence generated independently by different source entities or isolated original acquisition systems, with no copying, transfer, or derivation relationship between evidence records. When the completeness of the necessary evidence is one and the evidence in the necessary evidence type set comes from at least two independent data sources, the existing business status is output. When there are no records to confirm the occurrence of the business, the existing business status is output. In other cases, output the status of the business that needs to be supplemented with a certificate.

8. The knowledge graph-based water transport risk control method according to claim 3, characterized in that, The consistency residuals are dimensionless, and the dimensionless residuals are compared with the residual boundaries determined based on historical normal performance samples of the same route, cargo type and settlement stage. Write the residuals that exceed the residual boundary, the corresponding residual direction, the associated evidence identifier, and the target node identifier into the contradictory evidence vector; when there is a reachable path from the preceding event type to the following event type in the reference performance partial order and the time order of the corresponding observed events is reversed, mark the corresponding event pair as reverse order conflict; For event pairs that are not comparable in the reference performance partial order, no reverse order marking is made; The time interval exceeding the historical normal performance duration quantile value of the corresponding event type is marked as a timeout conflict, and the reverse order conflict and timeout conflict are written into the contradictory evidence vector.

9. The knowledge graph-based water transport risk control method according to claim 4, characterized in that, The event time reachability is determined based on the source node event time, target node event time, and reference performance partial order of the candidate directed propagation edge; the contradictory evidence vector pointing to the target node means that the target node identifier in the contradictory evidence vector is consistent with the target node identifier of the candidate directed propagation edge, and the residual direction or partial order conflict indicates that the risk impact is transmitted along the direction of the candidate directed propagation edge; for candidate directed propagation edges in the pending evidence supplementation state, the missing evidence type and the evidence supplementation deadline are recorded, and for candidate directed propagation edges in the blocked state, the blocking reason is recorded; when missing evidence is supplemented, the evidence corresponding to the blocking reason undergoes a version change, or the evidence supplementation deadline expires, the propagation gate state is re-determined, and when the evidence supplementation deadline expires and the missing evidence has not been supplemented, the corresponding candidate directed propagation edge is set to the blocked state.

10. The knowledge graph-based water transport risk control method according to claim 6, characterized in that, The dependency index uses the evidence version identifier as the first index key, and records the contradictory edge identifier generated by the evidence version, the propagation gate identifier controlled by the contradictory edge, the traceable evidence path identifier using the propagation gate, and the risk increment of the evidence version to each traceable evidence path, with the risk increment serving as the risk path contribution. The risk value of a path is obtained by adding the risk path contributions belonging to the same traceable evidence path, and the risk value of the target node is obtained by adding the cumulative risk values ​​of paths reaching the same target node and then adding them to the baseline risk value. When the evidence version becomes invalid, the risk path contribution is deducted from the cumulative risk value of the corresponding path and the risk value of the target node according to the dependency index, and only the nodes and edges that have data dependency relationship with the corresponding performance evidence anchor point are added to the recalculation set. After the evidence is changed, the differences in evidence version identifiers, content summaries, transmission gate status, and risk values ​​before and after the change will be written into the warning interpretation record.

Images