A knowledge graph-based intelligent identification system for supply chain risks

Abstract

This invention relates to the field of big data technology in industrial production, and discloses a knowledge graph-based intelligent supply chain risk identification system. The system includes a data processing module, a basic knowledge graph construction module, a disturbance feature generation module, a latent dependency identification module, a shadow topology verification module, a risk propagation identification module, and a graph state management module. It constructs a basic knowledge graph and execution disturbance sequences based on time-series execution data, forms a candidate dependency set based on explicit relation edges, and determines the degree of disturbance association within an allowable delay set to generate candidate virtual dependency nodes and candidate latent dependency edges. By comparing the propagation error between the shadow knowledge graph and the basic knowledge graph, it determines the amount of propagation error improvement, and accordingly writes into or removes candidate topologies. Based on the updated basic knowledge graph, it outputs the risk propagation results. This invention can reduce missed and false identifications caused by missing propagation paths and erroneous associations, and improve the accuracy of supply chain risk identification.

Description

Technical Field

[0001] This invention relates to the field of big data technology in industrial production, and in particular to a supply chain risk intelligent identification system based on knowledge graphs. Background Technology

[0002] With the informatization of production, quality inspection, material warehousing and logistics, data such as material arrival time, process status, equipment status, quality inspection results, logistics node status and transportation route are continuously generated during the execution of the supply chain. When material delivery delays, quality abnormalities, process abnormalities, equipment downtime or logistics delays occur, the relevant abnormalities may affect other supply execution nodes through material, process or logistics relationships, thereby creating supply chain risks.

[0003] Currently, supply chain risk identification typically employs methods such as rule-based threshold judgment, statistical analysis, or knowledge graph analysis. Among these methods, knowledge graph analysis primarily establishes nodes and relationship edges based on the supply execution object, material object, process object, and logistics object, as well as the registered material relationships, process routes, and logistics paths. After detecting node anomalies, other potentially affected nodes are analyzed along existing relationship edges to identify nodes and propagation paths that may be affected by the risk.

[0004] However, existing supply chain knowledge graphs are mainly built based on registered objects and relationships. When multiple supply execution nodes jointly depend on processing links, equipment resources, or logistics resources that are not recorded in the existing system, the graph lacks corresponding common dependent nodes and relationship edges, which can easily lead to missing risk propagation paths and missed identification of affected nodes. At the same time, the same abnormal factor may manifest at different nodes with delays due to process handling, logistics transmission, or execution buffering. Comparing only the state data at the same time location can easily miss relationships with time differences. Directly writing the relationships inferred from state similarity into the knowledge graph may allow incorrect relationships to participate in subsequent propagation calculations, causing irrelevant nodes to be misidentified, thereby reducing the accuracy and reliability of risk identification. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge graph-based intelligent identification system for supply chain risks, so as to at least solve one of the above-mentioned problems existing in the prior art.

[0006] Specifically, this is achieved through the following technical solution:

[0007] A knowledge graph-based intelligent supply chain risk identification system includes:

[0008] The data processing module is used to acquire execution data and generate time-series execution data;

[0009] The basic knowledge graph construction module is used to generate supply execution nodes based on the time-series execution data, and to establish explicit relationship edges based on the registered material relationships, process relationships and logistics relationships in order to construct the basic knowledge graph;

[0010] The disturbance feature generation module is used to generate an execution disturbance sequence based on the state deviation of each supply execution node relative to the historical baseline.

[0011] The implicit dependency identification module is used to form a candidate dependency set based on the explicit relation edges, determine the degree of perturbation association between supply execution nodes in the candidate dependency set within the allowable delay set, and generate candidate virtual dependency nodes and candidate implicit dependency edges based on the degree of perturbation association.

[0012] The shadow topology verification module is used to add the candidate virtual dependency nodes and candidate implicit dependency edges to the copy of the basic knowledge graph to form a shadow knowledge graph.

[0013] The risk propagation identification module is used to input the same risk trigger data into the basic knowledge graph and the shadow knowledge graph within the verification window, execute risk propagation according to the same propagation rules, obtain a first prediction result and a second prediction result, and output the risk propagation result based on the updated basic knowledge graph.

[0014] The shadow topology verification module is also used to determine the actual perturbation result based on the actual execution data collected in the verification window, determine the propagation error of the first prediction result and the second prediction result relative to the actual perturbation result, and determine the propagation error improvement amount based on the difference between the propagation error of the basic knowledge graph and the shadow knowledge graph.

[0015] The graph state management module is used to write the candidate virtual dependency nodes and candidate implicit dependency edges into the basic knowledge graph or remove them from the shadow knowledge graph based on the propagation error improvement amount.

[0016] The data processing module generates time-series execution data, specifically including:

[0017] The execution data is grouped into objects based on the supply execution object identifier, and the grouped execution data is batch-associated based on the execution batch identifier;

[0018] The associated execution data is converted to a unified time base and arranged according to a preset evaluation cycle to form time-series execution data corresponding to each supply execution object;

[0019] When any supply execution object lacks the valid data required to form a state deviation in any assessment period, an invalid flag is set for that assessment period so that the corresponding data is not involved in the determination of the degree of subsequent disturbance correlation.

[0020] The basic knowledge graph construction module constructs the basic knowledge graph, specifically including:

[0021] A supply execution node is generated based on the supply execution object identifier in the time-series execution data;

[0022] Based on the material object identifier, process object identifier, and logistics object identifier recorded in the time-series execution data, corresponding associated nodes are generated;

[0023] Based on the registered material correspondence, process routes, and logistics paths, establish explicit relationship edges between the supply execution nodes and their corresponding associated nodes;

[0024] For multiple time-series execution data with the same object identifier, the status data in the multiple time-series execution data is written to or updated in the node attribute or time-series attribute of the corresponding graph node, without generating the corresponding graph node repeatedly.

[0025] The perturbation feature generation module generates an execution perturbation sequence, specifically including:

[0026] The deviation value of the original status is determined based on the execution status of the supply execution nodes in each evaluation period;

[0027] The historical benchmark and historical fluctuation scale are determined based on the execution data of the supply execution nodes within the historical window.

[0028] Based on the original state deviation, historical benchmark, and historical fluctuation scale, the standardized perturbation value for each evaluation period is determined, and the standardized perturbation value is arranged according to the evaluation period to form an execution perturbation sequence;

[0029] Based on the standardized perturbation value and the perturbation activation condition corresponding to the perturbation type, the effective perturbation period for perturbation correlation calculation is determined.

[0030] The evaluation period with the invalidation mark is not involved in the correlation calculation between different execution perturbation sequences.

[0031] The implicit dependency identification module forms a candidate dependency set, specifically including:

[0032] Based on the material association, process association, and logistics association represented by the explicit relation edges, determine the structural association degree between supply execution node pairs;

[0033] Supply execution node pairs whose structural correlation reaches a preset structural threshold are identified as candidate relationships;

[0034] A candidate relationship graph is formed based on the candidate relationships, and multiple candidate supply execution nodes that are interconnected in the candidate relationship graph are assigned to the same candidate dependency set, so as to limit the determination range of the degree of disturbance association to the same candidate dependency set.

[0035] The implicit dependency identification module determines the degree of perturbation correlation within the allowable delay set, specifically including:

[0036] Based on the process processing time, logistics delivery time, execution buffer time, or historical abnormal delivery time between candidate supply execution nodes, determine the set of allowable delays consisting of multiple candidate delays;

[0037] For each candidate delay, the execution perturbation sequence of one candidate supply execution node is shifted in time relative to the execution perturbation sequence of another candidate supply execution node;

[0038] The sequence differences are determined within the evaluation period when both execution perturbation sequences after time shift have valid data and the corresponding perturbation values ​​meet the corresponding perturbation activation conditions.

[0039] When the number of effective perturbation pairing cycles reaches a preset lower limit, the candidate delay that minimizes the sequence difference is determined as the optimal delay, and the degree of perturbation correlation between the two candidate supply execution nodes is determined based on the corresponding minimum sequence difference.

[0040] The implicit dependency identification module generates candidate virtual dependency nodes and candidate implicit dependency edges, specifically including:

[0041] The implicit dependency confidence of the candidate dependency set is determined based on the degree of perturbation association and structural association of each candidate supply execution node pair within the candidate dependency set.

[0042] When the implicit dependency confidence reaches the candidate threshold and the number of candidate supply execution nodes that meet the effective perturbation association conditions reaches a preset number, candidate virtual dependency nodes are generated to represent unregistered common dependencies, and candidate implicit dependency edges are established between each candidate supply execution node and the candidate virtual dependency node.

[0043] The candidate implicit dependency edge includes an evidence aggregation edge pointing from the candidate supply execution node to the candidate virtual dependency node and a risk propagation edge set in reverse; the edge weight of the candidate implicit dependency edge is determined according to the implicit dependency confidence and the degree of perturbation correlation between the corresponding candidate supply execution node and other nodes in the set.

[0044] The node delay offset of each candidate supply execution node is determined based on the optimal delay of each candidate supply execution node pair, and the propagation delay when risk propagation occurs through the candidate virtual dependency node is determined based on the node delay offset.

[0045] The shadow topology verification module forms a shadow knowledge graph, specifically including:

[0046] For each candidate virtual dependency node, copy all or part of the graph structure related to the verification of that candidate virtual dependency node in the basic knowledge graph;

[0047] The candidate virtual dependency node and candidate implicit dependency edge are added to the copied graph structure to form a shadow knowledge graph corresponding to the candidate virtual dependency node. The graph structure used as the comparison benchmark has the same original nodes, effective relation edges, node temporal states, propagation parameters, and evaluation node set as the shadow knowledge graph. The topological difference between the two is the current candidate virtual dependency node and candidate implicit dependency edge.

[0048] During the verification process, the baseline topology of the basic knowledge graph is kept unchanged due to the current candidate virtual dependency node, and the newly added execution data in the verification window is synchronized to update the temporal state of the corresponding nodes in the two graphs.

[0049] The data range used to generate candidate virtual dependency nodes is set as the generation window, and the data range used to verify candidate virtual dependency nodes is set as the verification window. The verification window is located after the generation window and does not overlap with the generation window.

[0050] The risk propagation identification module performs risk propagation, and the shadow topology verification module compares propagation errors, specifically including:

[0051] Within the verification window, the same risk trigger data is input into the basic knowledge graph and the shadow knowledge graph;

[0052] Risk propagation is performed according to the same relational edge direction, edge weight usage rules, propagation delay usage rules, maximum propagation level, and termination conditions to obtain the first and second prediction results.

[0053] The actual disturbance node and actual disturbance time are determined based on the actual execution data collected within the verification window;

[0054] Under the same set of evaluation nodes and verification weights, the propagation errors of the first and second prediction results relative to the actual perturbation results are determined respectively, and the amount of propagation error improvement is determined based on the difference between the propagation errors of the two spectra.

[0055] The graph state management module updates the state of candidate virtual dependency nodes and candidate implicit dependency edges, specifically including:

[0056] When the propagation error improvement amount of the same candidate virtual dependency node reaches the improvement threshold within a preset number of consecutive valid verification windows, the candidate virtual dependency node and the candidate implicit dependency edge are written into the basic knowledge graph, and the candidate virtual dependency node and the candidate implicit dependency edge are synchronously updated to the valid state.

[0057] When there is insufficient valid verification data in the verification window, the verification window is not counted in the number of consecutive valid verification windows, and the candidate status of the candidate virtual dependency node is maintained.

[0058] When only a portion of the propagation error improvement in the most recent preset number of consecutive valid verification windows reaches the improvement threshold, the candidate state of the candidate virtual dependency node is maintained, and rolling verification is performed using subsequent valid verification windows.

[0059] When the propagation error improvement amount of the same candidate virtual dependency node does not reach the improvement threshold within a preset number of consecutive valid verification windows, the candidate virtual dependency node and the candidate implicit dependency edge are removed from the shadow knowledge graph and updated to an invalid state.

[0060] For virtual dependency nodes and implicit dependency edges that have been written into the basic knowledge graph, if they fail to meet the improvement conditions required to maintain an effective state within a preset number of subsequent consecutive valid verification windows, they will be stopped from participating in formal risk propagation and updated to a pending verification state or a historical failure state.

[0061] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0062] 1. This invention forms a candidate dependency set based on the explicit relation edges in the basic knowledge graph, and limits the correlation analysis between execution perturbation sequences to supply execution nodes with structural correlations. This can reduce invalid data comparisons between unrelated nodes and reduce the possibility of accidental state changes forming erroneous correlations.

[0063] 2. This invention performs time shifting and difference comparison on the execution disturbance sequences of different supply execution nodes within the allowable delay set, which can identify the disturbance correlation formed by the delay manifestation of the same abnormal factor on different supply execution nodes, and reduce the omission of correlation caused by the time misalignment of abnormal states.

[0064] 3. This invention generates candidate virtual dependency nodes and candidate implicit dependency edges based on the degree of perturbation association between supply execution nodes within the candidate dependency set. This enables common processing links, equipment resources, or logistics resources that are not explicitly registered in the basic knowledge graph to participate in risk propagation calculation in the form of virtual topology, thereby reducing the loss of propagation paths and the omission of affected nodes caused by the lack of common dependency relationships.

[0065] 4. Before writing candidate virtual dependency nodes into the basic knowledge graph, this invention first verifies them in the shadow knowledge graph. The validity of the candidate virtual dependency nodes is determined based on the difference in propagation error between the basic knowledge graph and the shadow knowledge graph relative to the actual perturbation results. This can reduce the entry of unverified speculative relationships into the basic knowledge graph, reduce the possibility of irrelevant nodes being misidentified, and improve the accuracy and reliability of risk propagation identification results. Attached Figure Description

[0066] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 A flowchart illustrating the module composition and process interaction of the intelligent supply chain risk identification system provided in this embodiment of the invention.

[0068] Figure 2 This is a schematic diagram of multi-source execution data time alignment and invalid cycle marking provided in an embodiment of the present invention;

[0069] Figure 3 A schematic diagram illustrating the determination of standardized perturbation values ​​and perturbation activation periods provided in an embodiment of the present invention;

[0070] Figure 4 This is a schematic diagram illustrating the joint screening of structural correlation evidence and temporal perturbation evidence provided in an embodiment of the present invention;

[0071] Figure 5 This is a schematic diagram illustrating the comparison of candidate delay distances and the determination of the optimal delay provided in an embodiment of the present invention;

[0072] Figure 6 A schematic diagram comparing the propagation error of the basic knowledge graph and the shadow knowledge graph provided in this embodiment of the invention;

[0073] Figure 7 This is a schematic diagram illustrating the continuous verification and risk propagation results provided in an embodiment of the present invention, wherein, Figure 7 (a) is a schematic diagram comparing the propagation error of continuous effective verification windows. Figure 7 (b) is a schematic diagram of the risk intensity distribution of each supply execution node. Detailed Implementation

[0074] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. The embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0075] Example:

[0076] This embodiment provides a knowledge graph-based intelligent supply chain risk identification system. The system can be deployed on a server, server cluster, or supply execution data processing platform with data processing capabilities, and can receive execution data generated by production execution systems, quality inspection systems, warehouse management systems, transportation management systems, equipment controllers, or positioning terminals through database interfaces, message interfaces, or file interfaces.

[0077] like Figure 1 As shown, the system includes a data processing module, a basic knowledge graph construction module, a perturbation feature generation module, a latent dependency identification module, a shadow topology verification module, a risk propagation identification module, and a graph state management module.

[0078] The data processing module collects, batch-associates, and unifies the time base of the acquired execution data representing the status of material supply, production and processing, quality inspection, equipment operation, or logistics execution, forming time-series execution data, which is then provided to the basic knowledge graph construction module and the perturbation feature generation module respectively. The basic knowledge graph construction module constructs a basic knowledge graph based on the time-series execution data and the registered material relationships, process relationships, and logistics relationships. The perturbation feature generation module generates an execution perturbation sequence based on the state deviation of each supply execution node relative to the historical baseline. The implicit dependency identification module combines the explicit relationships and structural constraints provided by the basic knowledge graph with the execution perturbation sequence to generate candidate virtual dependency nodes and candidate implicit dependency edges, which are then provided to the shadow topology verification module.

[0079] The shadow topology verification module adds candidate virtual dependency nodes and candidate implicit dependency edges to a copy of the basic knowledge graph to form a shadow knowledge graph. The risk propagation identification module receives the basic knowledge graph, the shadow knowledge graph, and the same risk trigger data, and performs risk propagation according to the same propagation rules to obtain a first prediction result and a second prediction result. The shadow topology verification module determines the actual perturbation result based on the actual execution data collected in the verification window, and compares the actual perturbation result with the first prediction result and the second prediction result to determine the amount of propagation error improvement. The graph state management module performs write, hold, or undo processing on the candidate virtual dependency nodes and candidate implicit dependency edges according to the amount of propagation error improvement. Among them, the verified and valid candidate virtual dependency nodes and candidate implicit dependency edges are written into the basic knowledge graph to participate in subsequent formal risk propagation.

[0080] For example, multiple production and processing units may actually share a transfer conveyor line, shared equipment, or logistics resources that are not recorded in the existing system. When an anomaly occurs in this shared resource, each production and processing unit may experience delays in material arrival, process waiting, or completion delays due to differences in preceding inventory, process cycle time, or transportation buffer. If the basic knowledge graph lacks this shared resource and its connections, then when propagating risk along existing material relationships, process relationships, or logistics relationships, it may only be possible to identify some of the affected nodes.

[0081] Therefore, this embodiment first uses explicit relationships in the basic knowledge graph to limit the scope of candidate analysis, then combines perturbation changes within the allowed delay set to generate candidate virtual dependency topologies, and uses the actual perturbation results in the subsequent verification window to verify whether the candidate virtual dependency topologies can reduce propagation errors. This not only supplements risk propagation paths formed by unregistered joint processing links, equipment resources, or logistics resources, reducing the omission of affected nodes, but also prevents candidate relationships formed solely by accidental state similarity from entering the basic knowledge graph without verification, thereby reducing the possibility of irrelevant nodes being incorrectly propagated and incorrectly identified.

[0082] Understandably, a supply execution object refers to an actual execution object that can independently collect execution status data and participate in material, process, or logistics relationships. It can be a material supply unit, production and processing unit, quality inspection unit, or logistics execution unit. The supply execution object identifier is a unified object identifier used in the original business system to identify the execution object. The supply execution node is a graph node generated by the basic knowledge graph construction module based on the supply execution object identifier. Both can use the same unified coding, but it is called the supply execution object identifier during the data processing stage and the supply execution node identifier after the graph is constructed.

[0083] In this embodiment, when the data processing module generates time-series execution data, it first collects the execution records according to the supply execution object identifier, and then associates the records generated in the procurement, warehousing, production, quality inspection and logistics links of the same batch of materials, the same production task or the same transportation task according to the execution batch identifier.

[0084] The execution record may include at least the supply execution object identifier, execution batch identifier, event time, status value, data source and validity status, and may also include one or more of the following: material object identifier, process object identifier, logistics object identifier, planned time, actual time, quality inspection value, process execution duration, equipment downtime, logistics dwell time and transportation route information.

[0085] When different data sources use different local codes for the same supply execution object or the same execution batch, the local codes can be converted into a unified supply execution object identifier and execution batch identifier according to a preset object mapping table. If the logistics terminal, production execution system, or warehouse management system uses different time bases, all time records can be uniformly converted to the same time base. When there is a fixed time offset, the offset can be determined by the time records of the same execution event in multiple data sources, and the corresponding timestamps can be corrected. After completing object collection, batch association, and time correction, the execution records are arranged according to preset evaluation cycles such as hours, shifts, days, weeks, or execution batches to form time-series execution data corresponding to each supply execution object.

[0086] When any supply execution object lacks the valid data required to form a status deviation within any evaluation period, the data processing module sets an invalid mark for that evaluation period. The invalid evaluation period retains its time position but is not included in subsequent numerical comparisons. For example... Figure 2 As shown, after batch correlation and time correction, the multi-source event records are projected onto a unified evaluation period axis. Periods lacking valid data only retain the time position and are marked as invalid. It can be seen that the multi-source records have a unified object, batch and time reference before entering the disturbance correlation calculation, and invalid periods are not treated as zero disturbances for comparison, thus avoiding erroneous correlation caused by time misalignment and common missing data.

[0087] The invalid flag is used to distinguish between "no valid data obtained" and "no node disturbance", avoiding common missing data being mistakenly judged as consistent state by the same fill value. For data with different reliability levels, data weights can also be set according to the degree of automatic collection, record integrity and historical verification results, so that reliable data can make a higher contribution in subsequent calculations.

[0088] Based on the aforementioned time-series execution data, the basic knowledge graph construction module further generates supply execution nodes according to the supply execution object identifiers in the time-series execution data, and generates corresponding material nodes, process nodes, and logistics nodes according to the material object identifiers, process object identifiers, and logistics object identifiers, thereby establishing explicit relationship edges based on the registered material correspondences, process routes, and logistics paths; for multiple time-series execution data with the same object identifier, graph nodes are not created repeatedly, but the current state is written into the node attributes separately, and the time-series attributes are updated according to the evaluation cycle.

[0089] Understandably, in this embodiment, node attributes can record relatively stable information such as object type, unified identifier, current status, data source, and effective status; time-series attributes can save change information such as actual arrival time, actual completion time, quality status, dwell time, or downtime according to the evaluation cycle; relationship edge attributes can record relationship type, relationship direction, effective time interval, relationship status, and relationship reliability; and only relationship edges that are in the current effective state participate in the current risk propagation.

[0090] For example, the underlying knowledge graph used when verifying the current candidate virtual dependency node can be represented as:

[0091] ;

[0092] In the formula, This represents the basic knowledge graph used when verifying the current candidate virtual dependency node, and does not include the candidate virtual dependency node or its candidate implicit dependency edges; Represents the current set of valid nodes; This represents the set of valid relation edges currently participating in formal risk propagation. The set of valid relation edges includes explicit relation edges formed based on the relationships between registered objects, as well as valid implicit dependency edges that have been verified and written into the basic knowledge graph. Represents a collection of node attributes; Represents the set of attributes of the relation edge.

[0093] In this embodiment, during the initial construction phase of the basic knowledge graph, only nodes and explicit relationship edges that can be directly determined by execution data, material master data, process route tables, or logistics path tables are written. During system operation, virtual dependency nodes and implicit dependency edges that are verified as valid by the shadow knowledge graph can be further written into the basic knowledge graph, while candidate virtual dependency nodes and candidate implicit dependency edges that are currently to be verified are not written into the basic knowledge graph before being confirmed as valid.

[0094] As a preferred implementation, the disturbance feature generation module determines the original state deviation value based on the execution status of the supply execution node in each evaluation period, and determines the historical benchmark and historical fluctuation scale based on the effective execution data of the node in the historical window.

[0095] Different types of disturbances can correspond to different execution anomaly mechanisms. For example, time offset is used to reflect delivery or completion delays, quality deviation is used to reflect changes in test results relative to the qualified state, logistics dwell deviation is used to reflect congestion at transportation nodes, and process execution deviation is used to reflect increased process time or abnormal equipment shutdown.

[0096] Therefore, when an evaluation cycle contains multiple valid execution batches, the supply execution node... The time offset disturbance value can be expressed as:

[0097] ;

[0098] In the formula, Indicates the supply execution node During the evaluation period The time offset perturbation value within; Indicates the type of time offset disturbance; Indicates the evaluation period Internal and supply execution nodes The associated set of valid execution batches; Represents any execution batch in the set of valid execution batches; Indicates the execution batch The actual time; Indicates the execution batch The base time; Indicates the execution batch Non-negative weights; This represents a denominator stability parameter greater than zero, used to reduce the sensitivity of the calculation results to small data changes when the sum of the effective batch weights is small.

[0099] It should be added that when When the data is not empty and the sum of the weights of each valid execution batch is greater than zero, the time offset perturbation value is calculated according to the above formula; while when... Empty, or If the sum of the weights of all valid execution batches within a set is zero but not empty, the time offset disturbance value is not calculated, and the corresponding evaluation period is marked as invalid.

[0100] This weighted average can reduce the amplified impact of a single sporadic batch on the overall state of the node; the maximum value operation only retains the positive delay, avoiding early arrival or early completion to offset the delay.

[0101] For quality deviation, logistics dwell time deviation, and process execution deviation, the corresponding original state deviation values ​​can be constructed based on the non-conformance rate, rework rate, the degree to which the test value exceeds the acceptable range, the difference between the actual dwell time and the benchmark time, the increase in process time, or the abnormal downtime of equipment.

[0102] At the same time, in order to make different nodes and different disturbance indicators have comparable dimensions, the supply execution node... During the evaluation period Internal targeting of disturbance type The standardized perturbation value can be expressed as:

[0103] ;

[0104] In the formula, Indicates the supply execution node During the evaluation period Internal targeting of disturbance type The standardized perturbation value; This represents the corresponding deviation from the original state. Indicates the supply execution node For the type of disturbance Historical benchmarks; Indicates the corresponding historical fluctuation scale; This represents the lower limit protection parameter for fluctuation scales greater than zero.

[0105] As a preferred implementation, the historical benchmark can be determined using the mean or median of historical valid data, and the historical fluctuation scale can be determined using the standard deviation or median absolute deviation. The historical window used to determine the historical benchmark and historical fluctuation scale is located before the current evaluation period and contains no less than a preset number of valid historical samples. When the number of valid historical samples is insufficient, the historical window can be extended; if the extended historical window still cannot obtain enough valid historical samples, the current disturbance type is marked as temporarily uncalculated in the corresponding evaluation period, and standardized disturbance values ​​lacking statistical basis are not used in candidate relation generation.

[0106] The above standardization process converts the original deviations of different dimensions into the degree of deviation relative to the normal state of the node itself, and avoids abnormal values ​​by protecting the lower bound parameter of the fluctuation scale from being generated when the historical fluctuation is zero or too small.

[0107] Furthermore, in order to uniformly organize the standardized disturbance values ​​of the same supply execution node under different disturbance types, the standardized disturbance values ​​of the same supply execution node within the same evaluation period can be combined into an execution disturbance vector according to a pre-set disturbance type order that is consistent for each supply execution node.

[0108] Specifically, the execution perturbation vector can be represented as:

[0109] ;

[0110] In the formula, Indicates the supply execution node During the evaluation period The execution perturbation vector within; This indicates the number of disturbance types that participate in the calculation according to a preset order; This indicates the type of perturbation in the perturbation vector. The corresponding standardized perturbation value.

[0111] The system also sets validity markers for each component in the perturbation vector. Perturbation type components that do not have valid data are marked as invalid and excluded in subsequent association calculations for the corresponding perturbation type, instead of being filled with zero values. This avoids mistaking missing data for nodes not being perturbed, or forming false associations due to multiple nodes being missing data at the same vector position.

[0112] To eliminate false high correlations formed when two nodes remain in a normal state for an extended period, this embodiment further sets a perturbation activation condition, and the perturbation activation flag corresponding to each perturbation component can be represented as:

[0113] ;

[0114] In the formula, Supply execution node During the evaluation period Internal targeting of disturbance type The perturbation activates the marker.

[0115] As can be seen from the above formula, for unidirectional adverse disturbances such as delivery delays, increased process time, equipment downtime, and logistics delays, when season Both excessively high and excessively low values ​​may constitute abnormal two-way disturbances, when... season ,in, Indicates the type of disturbance The corresponding perturbation activation threshold; while in other cases, let .

[0116] It should be noted that the validity flag indicates whether the corresponding perturbation component has data available for calculation, while the perturbation activation flag indicates whether the perturbation component with valid data has met the corresponding anomalous activation condition. For evaluation periods with valid data but not meeting the perturbation activation condition, when they subsequently enter the corresponding historical window, they can be used as normal valid samples to update the historical benchmark and historical fluctuation scale, but not as valid perturbation evidence for generating unregistered common dependencies.

[0117] like Figure 3 As shown, valid data within the historical window is used to determine the historical baseline and fluctuation scale. Evaluation periods where the standardized perturbation value exceeds the corresponding positive or bidirectional activation threshold are identified as valid perturbation periods. Normal fluctuations that do not exceed the threshold can still be used to update the historical baseline, but are not considered as evidence of implicit dependence. The attached figure indicates that the system compares activated anomalous changes rather than long-term normal states, thereby reducing spurious correlations caused by normal resonances or co-missing states.

[0118] Based on the above execution disturbance sequence, the implicit dependency identification module first restricts the comparison range of supply execution nodes according to the explicit relationship edges in the basic knowledge graph in order to determine which supply execution nodes have a common structural basis of relying on the same processing link, equipment resources or logistics resources, and then performs time delay correlation analysis on the disturbance changes between the corresponding supply execution nodes.

[0119] Nodes lacking material, process, or logistics connections are not directly considered as common dependency candidates even if they experience short-term synchronization anomalies.

[0120] In this embodiment, the supply execution node will be... With supply execution nodes The combination of nodes formed is called a supply execution node pair, and is denoted as . Therefore, the supply execution node is... The structural correlation can be expressed as:

[0121] ;

[0122] And the weights of each structural dimension satisfy:

[0123] ;

[0124] In the formula, Indicates supply execution node pair The degree of structural correlation; Indicates the similarity of material properties; Indicates the similarity of process paths; Indicates the degree of overlap in logistics routes; , and These represent the non-negative weights of the corresponding structural dimensions. , and All are normalized to the range of 0 to 1, therefore It is also in the range of 0 to 1.

[0125] Material attribute similarity can be determined based on the similarity of material categories and specifications, as well as the overlap ratio of supply sources; process path similarity can be determined based on the overlap of common key processes, process sequences, or shared outsourcing links; logistics path overlap can be determined based on the overlap ratio of common logistics nodes, transit warehouses, transportation routes, or distribution areas. The weights of each structural dimension can be set according to the specific supply execution scenario. For example, for scenarios with high risks associated with shared process equipment, the weight corresponding to process path similarity can be increased; for scenarios with frequent regional logistics disruptions, the weight corresponding to logistics path overlap can be increased.

[0126] When the structural correlation of a node pair reaches a preset structural threshold, a candidate relationship is established between the two corresponding supply execution nodes. This forms a candidate relationship graph from all candidate relationships, and candidate dependency sets can be formed based on the connected components in the candidate relationship graph. To prevent a single bridging node from merging multiple nodes with weak internal correlations into the same set, each candidate dependency set can be required to reach a preset minimum edge density or minimum average structural correlation. For connected components that do not meet the constraints, they can be further divided into multiple candidate dependency sets based on the distribution of candidate relationship edges. Entering the same candidate dependency set only indicates that the corresponding supply execution node has the structural basis for further perturbation correlation calculations; it does not indicate that their common dependency relationship has been confirmed.

[0127] Therefore, the candidate dependency set only limits the scope of subsequent calculations and preserves the structural basis between candidate nodes, without directly confirming common dependencies.

[0128] In some embodiments, for candidate supply execution nodes within the same candidate dependency set, the implicit dependency identification module can determine an allowable delay set consisting of multiple candidate delays based on process processing time, logistics delivery time, execution buffer time, or historical abnormal delivery time.

[0129] The same anomaly may be delayed in appearing at different points due to process transfer, inventory consumption or logistics buffer, and comparing only the same assessment period may easily miss the true correlation.

[0130] For example, in this embodiment, node pairs The allowable delay set can also be represented as:

[0131] ;

[0132] In the formula, Represents a node With nodes The set of allowable delays between; to This represents the candidate delay converted into the number of evaluation cycles; This indicates the number of candidate delays.

[0133] Allowable delay set The set is a finite non-empty set, where candidate delays are determined based on the possible anomaly propagation time during actual supply execution. By pre-defining the range of candidate delays, it is possible to avoid searching for the best match within an arbitrarily long time range, thereby reducing the possibility of forcibly aligning two anomalies that are too far apart and lack a real propagation basis.

[0134] When performing latency correlation calculations, the allowable latency set is considered. Each candidate delay in Supply execution node The execution perturbation sequence relative to the supply execution node The execution perturbation sequence is time-aligned, and perturbation data at corresponding time positions are compared. In this embodiment, positive delay is defined as the supply delay of the execution node. The disturbance leads the supply execution node in time. The disturbance; when verification of the opposite direction is required, the order of the two supply execution nodes can be swapped and calculated separately.

[0135] In candidate delay Below, the node pairs target the disturbance type. The effective set of perturbation pairing periods can be represented as:

[0136] ;

[0137] In the formula, Indicates candidate delay The effective perturbation pairing period set is used to limit the evaluation period participating in the sequence difference calculation. By requiring both supply execution nodes to have valid data and meet the perturbation activation conditions within the corresponding evaluation period, spurious similarities formed by common missing data or common normal states can be eliminated.

[0138] And the node pair in the candidate delay The following is about the type of disturbance The constrained time delay distance can be expressed as:

[0139] ;

[0140] In the formula, Indicates candidate delay Weighted sequence differences; Indicates the evaluation period The corresponding non-negative effective data weights; The denominator is a stable parameter that is greater than zero. The restricted delay distance is calculated according to the above formula only when the number of elements in the effective perturbation pairing period set is not less than the preset lower limit and the sum of the weights corresponding to each effective perturbation pairing period is greater than zero; otherwise, the restricted delay distance corresponding to the candidate delay is not generated.

[0141] Based on the above, the constrained delay distance is calculated by averaging the squared difference after alignment and according to the effective weight. On the one hand, this can eliminate the influence of the positive and negative directions of the difference, and on the other hand, it can avoid the loss of comparability due to the different summation scales of nodes with different effective sample numbers. That is, the smaller the distance, the closer the abnormal change patterns of the two nodes are under the corresponding candidate delay.

[0142] The set of candidate time delays that satisfies the effective perturbation pairing condition is denoted as The optimal delay can then be expressed as:

[0143] ;

[0144] The corresponding minimum constrained delay distance is:

[0145] ;

[0146] In the formula, Represents the set of allowed delays The system can form a candidate delay set with effective limited delay distance, wherein the number of effective perturbation pairing cycles corresponding to each candidate delay reaches a preset lower limit, and the sum of the corresponding effective data weights is greater than zero. Represents node pairs For the type of disturbance The optimal latency; This represents the minimum constrained delay distance corresponding to the optimal delay.

[0147] When multiple candidate delays achieve the same minimum distance, the candidate delay with the smaller absolute value is selected first; if the absolute values ​​of the candidate delays are still the same, they are selected according to a preset fixed order to ensure that the same input yields a definite calculation result. And when... If empty, then the optimal delay is not applied. and minimum constrained time delay distance Define it.

[0148] Since a smaller minimum constrained delay distance indicates that the abnormal change patterns of two supply execution nodes under the corresponding disturbance type are more similar, this embodiment can map the minimum constrained delay distance to a temporal disturbance correlation degree to facilitate a unified representation of the correlation strength between nodes and to provide input for subsequent threshold comparison and comprehensive calculation of multiple disturbance types. The expression is as follows:

[0149] ;

[0150] In the formula, Indicates supply execution node pair In disturbance type The correlation degree of temporal perturbation on; The distance decay parameter represents a value greater than zero; while the temporal perturbation correlation... The value of is in the range of 0 to 1. When the minimum restricted delay distance is zero, the temporal disturbance correlation degree is one. As the minimum restricted delay distance increases, the temporal disturbance correlation degree continuously decreases. Therefore, the minimum restricted delay distances corresponding to different supply execution node pairs can be converted into a uniform scale of correlation strength, so that the supply execution node pairs with smaller minimum restricted delay distances have higher temporal disturbance correlation degrees.

[0151] When the same supply execution node has valid disturbance pairing data under multiple disturbance types, the temporal disturbance correlation degree corresponding to each disturbance type can be further weighted and synthesized to form a comprehensive temporal correlation degree. This comprehensive temporal correlation degree can be expressed as:

[0152] ;

[0153] In the formula, Represents node pairs A set of perturbation types with sufficiently effective perturbation pairing data; Indicates the type of disturbance The corresponding positive weights; when If the value is empty, no comprehensive temporal correlation degree is generated.

[0154] like Figure 4 As shown, by partitioning the structural correlation degree and temporal perturbation correlation degree of node pairs using structural threshold and evidence threshold, it can be seen that only node pairs that are simultaneously above both thresholds and have both structural basis and consistency of abnormal changes are taken as valid evidence, while those that only meet a single threshold are excluded. This confirms that the proposed scheme suppresses the false correlation of irrelevant nodes through dual screening of structure and temporal.

[0155] Each perturbation type participates in the synthesis only when there is sufficient effective paired data, thus accommodating situations where data types between nodes are not completely consistent while aggregating evidence of multiple perturbations.

[0156] It should be noted that the aforementioned comprehensive temporal correlation degree is used to characterize the correlation strength reflected by different disturbance types, while the optimal delay determined for different disturbance types may not be consistent. To unify the delay information corresponding to different disturbance types as a basis for determining the delay between supply execution nodes and to provide a foundation for subsequently determining the propagation delay of candidate virtual dependency topologies, the implicit dependency identification module can further perform a weighted comprehensive calculation of the constrained delay distances corresponding to each disturbance type under each candidate delay, thereby obtaining the cross-disturbance type comprehensive distance, the expression of which is:

[0157] ;

[0158] In the formula, Indicates candidate delay The set of perturbation types that satisfy the valid perturbation pairing conditions; Indicates the node pair in the candidate delay The overall distance across perturbation types.

[0159] Supply execution node pair The overall optimal delay can be expressed as:

[0160] ;

[0161] In the formula, Represents the set of allowed delays It can form a candidate delay set for the integrated distance across disturbance types, that is, under the corresponding candidate delay, there is at least one disturbance type that satisfies the effective disturbance pairing condition; Indicates the supply execution node Relative to supply execution node The optimal overall latency is determined, and the following is agreed upon. Represents a node Disturbance leading node .

[0162] When multiple candidate delays correspond to the same minimum cross-disturbance type composite distance, the candidate delay with the smaller absolute value is selected first; if the absolute values ​​of the candidate delays are still the same, they are selected according to a preset fixed order; when When empty, the corresponding optimal overall latency is uncertain. Based on Figure 5 As shown, different supply execution node pairs in the same candidate dependency set have different sequence differences under each candidate delay. The lowest point of the curve corresponds to the optimal delay of each node pair. This scheme can compensate for the abnormal manifestation time difference caused by process transfer, inventory buffer and logistics transportation by selecting the position of minimum difference within the pre-defined allowable delay set. It also provides a basis for determining the subsequent node delay offset and propagation delay, and can still correctly align and identify the association with time misalignment.

[0163] The aforementioned delay alignment is used to compensate for the time difference in anomaly manifestation, and the candidate delay is constrained by the actual transmission duration to avoid forcibly associating independent anomalies that are too far apart.

[0164] Based on the comprehensive temporal correlation of the aforementioned supply execution node pairs, the implicit dependency identification module comprehensively evaluates temporal perturbation evidence and structural relationship evidence at the candidate dependency set level. Candidate Dependency Set The temporal correlation degree and structural correlation degree of a set can be expressed as follows:

[0165] ;

[0166] ;

[0167] In the formula, Indicates the first A set of candidate dependencies; Represents the candidate dependency set An unordered set of node pairs that possesses an effective comprehensive temporal correlation degree; Represents a set The number of node pairs in the data; and only if The set association degree is calculated when the preset minimum number of node pairs is reached.

[0168] When records of known common disturbance events are available, the implicit dependency identification module can also match known common disturbance events, such as regional logistics interruptions, downtime of registered shared equipment, or abnormal unified data collection, with the disturbance occurrence time intervals of each supply execution node. When the matching degree reaches a preset interpretation threshold, the corresponding disturbance interval is marked as interpreted data, and the weight of the data in the interval is reduced in the candidate association calculation, or it is not used as evidence for generating candidate virtual dependency nodes. The processing does not delete the original execution records, nor does it affect the use of the interval data in the basic state records and subsequent evaluation of actual disturbance results.

[0169] The comprehensive discriminant of the candidate dependency set adopts This is to avoid interference with standardized perturbation values. Confusion, which can be represented as:

[0170] ;

[0171] In the formula, Represents the candidate dependency set The comprehensive discrimination metric; Indicates the temporal correlation degree of a set; Indicates the degree of association in the set structure; Indicates the amount of information explained by a known event; , and These represent the non-negative weights of the corresponding evaluation quantities; This represents the discrimination bias parameter; when a record of a known common disturbance event cannot be obtained, it can be set as follows: .

[0172] The comprehensive discriminant can then be further transformed into the implicit dependency confidence of the candidate dependency set based on the following formula:

[0173] ;

[0174] In the formula, Indicates the first Implicit dependency confidence of each candidate dependency set; Represented by natural constant An exponential function with base 0. The value of is greater than zero and less than one, and it follows... As the value increases, the confidence level also increases. By using the above logical mapping, the comprehensive discriminant value with different numerical ranges can be converted into a confidence level within a unified range, which facilitates comparison with candidate thresholds.

[0175] When the implicit dependency confidence reaches the candidate threshold, and the number of candidate supply execution nodes in the candidate dependency set that can form an effective comprehensive temporal correlation with at least one other candidate supply execution node reaches a preset number, the system generates candidate virtual dependency nodes to represent un-explicitly registered common dependencies. This candidate virtual dependency node records the set of associated supply execution nodes, implicit dependency confidence, generation window, basic knowledge graph version, candidate state, and optimal latency information for node pairs.

[0176] Furthermore, for the supply execution node that intends to establish candidate implicit dependency edges with candidate virtual dependency nodes... Its node association support is adopted express:

[0177] ;

[0178] In the formula, Indicates candidate virtual dependency nodes With supply execution nodes The support of node associations between them; Represents the candidate dependency set In and supply execution nodes A set of other supply execution nodes that form an effective comprehensive temporal correlation degree; This indicates the number of nodes in the set. And when... At that time, no calculation was performed. It is also not in the candidate virtual dependency node. With supply execution nodes Establish candidate implicit dependency edges between them.

[0179] Among them, candidate virtual dependency nodes With supply execution nodes The basic weights of candidate implicit dependency edges between them can be expressed as:

[0180] ;

[0181] Candidate implicit dependency edges include those from supply execution nodes. Pointing to candidate virtual dependency nodes Evidence aggregation edges, and candidate virtual dependency nodes Pointing to the supply execution node The risk propagation edge, the actual weights of the two types of edges can be expressed as follows:

[0182] ;

[0183] In the formula, Represents the basic weights of candidate implicit dependency edges; Indicates that the supply execution node Pointing to candidate virtual dependency nodes Evidence aggregation edge weights; Indicates the candidate virtual dependency node Pointing to the supply execution node Risk propagation edge weight; and Let represent the edge type coefficients for the two types of edges, and . .because and Both types of edge weights can be restricted to the range of 0 to 1.

[0184] Clearly, the evidence aggregation edge is used to transmit risk evidence already formed by the associated supply execution nodes to the candidate virtual dependency node, while the risk propagation edge is used to simulate the impact on other supply execution nodes after the common dependency object represented by the candidate virtual dependency node experiences an anomaly. When a candidate virtual dependency node is confirmed to be valid, its corresponding evidence aggregation edge and risk propagation edge are simultaneously updated to valid implicit dependency edges; when a candidate virtual dependency node is revoked or becomes invalid, the corresponding candidate implicit dependency edge simultaneously stops participating in propagation.

[0185] Furthermore, in order to convert the overall optimal latency between node pairs into a propagation latency that can be actually used in the candidate virtual dependency topology, this embodiment performs a set of candidate dependencies. Configure node delay offset for each supply execution node in the process. .

[0186] For example, the node delay offset can be determined by the following formula:

[0187] ;

[0188] In the formula, Represents the candidate dependency set The corresponding node delay offset vector; and These represent the supply execution nodes. and Relative to candidate virtual dependency nodes The time delay offset of the commonly dependent events; Represents a node Relative to node The optimal overall latency is determined, and the following is agreed upon. Represents a node Disturbance leading node ; As the weight of the node's time delay relationship; further constrained Used to eliminate the non-uniqueness caused by the simultaneous translation of all node delay offsets.

[0189] When the optimal delay of multiple node pairs can be fully satisfied by the same set of node delay offsets, the delay offset of each node can be directly determined; however, when the optimal delay of multiple node pairs has measurement errors or is not completely consistent, the above formula determines the set of node delay offsets that best fits the delay relationship of highly correlated node pairs by weighted minimization.

[0190] When the supply execution node During the evaluation period When a risk is triggered, it is detected via candidate virtual dependency nodes. To the supply execution node The effective propagation delay can be further expressed as:

[0191] ;

[0192] In the formula, Indicates from the supply execution node via candidate virtual dependency nodes Propagation to supply execution nodes Effective propagation delay; when At that time, the node is not predicted in reverse using the current triggering event. Risks occurring at an earlier time; thus, the optimal latency of node pairs is transformed into a propagation time parameter that the candidate virtual dependency topology can actually execute, rather than merely as an adjunct record of the candidate relationship.

[0193] The aforementioned candidate virtual dependency nodes aggregate structural evidence, temporal evidence, and propagation delay, and limit the risk diffusion caused by nodes with weaker evidence through edge weights.

[0194] After candidate virtual dependency nodes are generated, the shadow topology verification module can copy the graph structure related to the verification of the candidate node in the basic knowledge graph for each candidate virtual dependency node, and add the current candidate virtual dependency node and its candidate implicit dependency edges to the copied graph structure to form the corresponding shadow knowledge graph.

[0195] Among them, with the first The shadow knowledge graph corresponding to each candidate virtual dependency node can be represented as:

[0196] ;

[0197] In the formula, Indicates the relationship with the first Shadow knowledge graphs corresponding to each candidate dependency set; Indicates the current candidate virtual dependency node; Indicates the relationship with candidate virtual dependency nodes The corresponding set of candidate implicit dependency edges; The node attributes that represent candidate virtual dependency nodes; This represents the edge attributes of candidate implicit dependency edges.

[0198] In some embodiments, the shadow topology verification module can employ either full graph verification or local subgraph verification. When full graph verification is used, both the base knowledge graph and the shadow knowledge graph use the complete graph as the propagation scope. When local subgraph verification is used, a baseline subgraph related to the current candidate virtual dependency node is extracted from the base knowledge graph, and this baseline subgraph is used as the basis for calculating the first prediction result. Then, the current candidate virtual dependency topology is added to a copy of the same baseline subgraph to form a shadow subgraph.

[0199] Since the baseline graph and the shadow knowledge graph should have the same original nodes, effective relation edges, node temporal states, risk triggering data, propagation parameters, maximum propagation level, and evaluation node set, the only topological difference that needs to be evaluated between the two is the current candidate virtual dependency node and its candidate implicit dependency edge.

[0200] Furthermore, the baseline topology of the basic knowledge graph remains unchanged during the verification process; that is, the current candidate virtual dependency nodes and candidate implicit dependency edges are not written into the basic knowledge graph. New execution data added within the verification window is still synchronously used to update the temporal attributes and risk triggering status of the corresponding nodes in both the basic and shadow knowledge graphs. Therefore, unconfirmed relationships will not affect the formal risk output, but both graphs still use the same latest node status for comparison.

[0201] For example, the data range used to generate candidate virtual dependency nodes is set as the generation window, and the data range used to verify candidate virtual dependency nodes is set as the verification window; and the generation window and the verification window satisfy the following expression:

[0202] ;

[0203] In the formula, This represents a non-empty generation window consisting of several consecutive evaluation periods; This represents a non-empty verification window following the generation window; the two windows do not overlap, and the last evaluation period of the generation window is earlier than the first evaluation period of the verification window. When either window is empty or the two windows overlap, the validity verification of the current candidate virtual dependency node is not performed.

[0204] Understandably, in this embodiment, the generation window and the verification window are separated to prevent the use of the same set of perturbation data to both generate candidate relationships and prove their validity. Candidate virtual dependency nodes can only be established based on historical evidence in the generation window, and their propagation capability is then tested using the actual perturbation results that occur afterward, thereby reducing spurious improvements caused by overfitting to historical noise and ex-post interpretation.

[0205] In a preferred implementation, after entering the verification window, the risk propagation identification module inputs the same risk trigger data into the basic knowledge graph and the shadow knowledge graph respectively, and performs risk propagation according to the same relation edge direction, edge weight usage rules, propagation delay usage rules, maximum propagation level and termination conditions to obtain the first prediction result and the second prediction result.

[0206] In this embodiment, the propagation verification does not pre-input all anomalies that have occurred within the verification window into the graph. Instead, when a certain node first reaches the risk trigger condition, the data already obtained at that time is used as a common input for backward prediction, while the execution data of other nodes collected after the trigger period is used to form the actual disturbance result and calculate the propagation error.

[0207] In each verification event, the assessment period in which the risk triggering condition is first met within the verification window can be defined as the triggering period, and one or more supply execution nodes that have observed state deviations at that time can be defined as the set of triggering nodes. The basic knowledge graph and shadow knowledge graph only receive the same data that was obtained before or at the time of the triggering period, thus avoiding the use of the results to be verified to generate prediction results.

[0208] Therefore, for a risk-triggered node, its initial risk intensity can be expressed as:

[0209] ;

[0210] In the formula, Represents a node During the evaluation period The initial risk intensity before the virus has spread; Represents a node During the evaluation period A set of perturbation types containing valid data; Indicates the type of disturbance Non-negative risk weights; Indicates the standardized perturbation value; Indicates risk trigger bias; This represents a function that monotonically maps the calculation results to the range of zero to one. The initial risk intensity of non-triggering nodes at the start of this propagation can be set to 0.

[0211] For bidirectional anomalies, the standardized disturbance value can be converted into a non-negative anomaly intensity relative to the normal range before the initial risk intensity is formed. Changes in state that do not constitute the current risk direction, such as early arrival or shortened process time, can be excluded from being considered positive risk inputs. Therefore, the initial risk intensity remains consistent with the aforementioned disturbance activation conditions.

[0212] Furthermore, the target node The risk intensity can be updated layer by layer according to the following formula:

[0213] ;

[0214] In the formula, Represents a node In the Each level of transmission, assessment cycle The intensity of the risks that have already formed; Indicates pointing to the target node The set of valid predecessor nodes; Indicates the predecessor node Risk intensity after considering transmission delay; This represents the non-negative effective propagation delay of the relation edge or its corresponding virtual propagation path. Indicates the weight of the relation edge; Represents the relation type decay coefficient, and satisfies , The right side adopts This serves as the computational basis for the next propagation layer, preserving the target node risks already formed in the previous propagation layer.

[0215] Understandably, the above propagation formula aggregates the risk contributions of multiple precursor nodes by subtracting the product of the unpropagated risks of each precursor from 1. Any high-risk precursor can increase the risk of the target node, and the risk is further enhanced when multiple precursors act simultaneously, while the calculation result remains in the range of zero to one. The relation edge weight is used to reflect the strength of the association, the relation type decay coefficient is used to distinguish the propagation ability of material edge, process edge, logistics edge and implicit dependency edge, and the propagation delay is used to place the risk contribution into the corresponding prediction time position.

[0216] For example, when a certain predecessor node is When no effective risk intensity has been formed, the risk contribution of the predecessor node in this multiplication is recorded as zero, rather than stopping the entire prediction path because the actual execution data has not yet been generated. Therefore, when the propagation intensity is lower than the termination threshold, reaches the maximum propagation level, exceeds the maximum prediction time range, or the same path will repeatedly pass through the same node, the corresponding propagation path is stopped. Candidate virtual dependency nodes do not immediately return the risk to their direct predecessor nodes; nodes are not allowed to appear repeatedly in the same propagation path to avoid invalid loops formed between evidence aggregation edges and risk propagation edges.

[0217] Both the first and second prediction results can include predictions of affected nodes, risk intensity, disturbance time, and propagation path. After the verification window ends, based on the execution data actually collected after the triggering period, nodes whose standardized disturbance values ​​reach the corresponding actual disturbance judgment threshold are identified as actual disturbance nodes, and the evaluation period in which the threshold is first reached is identified as the actual disturbance time; nodes that completely lack valid data are not included in the evaluation.

[0218] To evaluate the risk propagation outcomes of the basic knowledge graph and the shadow knowledge graph separately, either the basic knowledge graph or the shadow knowledge graph is denoted as the graph to be evaluated. Among them, the map to be evaluated It can be the basic knowledge graph used as a comparison benchmark, or it can be the shadow knowledge graph formed after adding the current candidate virtual dependency nodes and candidate implicit dependency edges.

[0219] For each supply execution node participating in the evaluation within the verification window and possessing sufficient actual data, the evaluation is based on the node's actual disturbance label and the evaluation graph. Based on the predicted risk intensity and corresponding validation weights given for this node, the graph to be evaluated is determined. The node identification error is used to evaluate the map to be evaluated. The accuracy of identifying actual disturbance nodes.

[0220] The node identification error can be expressed as:

[0221] ;

[0222] In the formula, Indicates the graph to be evaluated Node identification error; This represents the set of nodes that participate in the evaluation and have sufficient actual data. Represents a node The actual disturbance label can be set to 1 when an actual disturbance occurs and 0 when no actual disturbance occurs. Representation of the spectrum Nodes within the evaluation interval The maximum predicted risk intensity is given. Represents a node Non-negative validation weights; This represents a denominator stability parameter greater than zero, determined based on the magnitude of the verification weights and the system's computational precision, and is less than the preset minimum effective weight. This is to reduce the impact of rounding errors in floating-point operations on the stability of node identification error calculation (for example, it can be lower than the preset minimum effective weight). Doubled The error is calculated only when the sum of the verification weights reaches the preset minimum effective weight; otherwise, the corresponding verification window is marked as having insufficient effective verification data.

[0223] Node identification error can evaluate both missed identification and false identification. For nodes that have actually been disturbed, if the prediction risk is too low, it will increase the error; while for nodes that have not been disturbed, if the prediction risk is too high, it will also increase the error. Therefore, the shadow knowledge graph will not necessarily achieve a better evaluation simply by expanding the prediction range.

[0224] When the system outputs the predicted disturbance time, the graph to be evaluated... The time prediction error can be expressed as:

[0225] ;

[0226] In the formula, Indicates the graph to be evaluated Time prediction error; This represents the set of nodes that actually experienced disturbances and have valid time records. Representation of the spectrum For nodes The normalized time error. The normalized time error can be determined by dividing the absolute difference between the predicted and actual disturbance times by the validation window length; when the actual disturbance node is not predicted, the corresponding... It is denoted as 1; when the actual set of disturbed nodes is empty, this item can be zero, but the corresponding verification window cannot be used alone as positive evidence for the candidate topology to pass verification. When the system does not output the predicted disturbance time, the weight of the time prediction error can be set to zero.

[0227] Furthermore, the map to be evaluated The overall propagation error can be expressed as:

[0228] ;

[0229] In the formula, Indicates the graph to be evaluated The overall propagation error; Indicates node identification error; Indicates the time prediction error; This represents the non-negative weight of the time prediction error. The node identification error, time prediction error, and overall propagation error of both the basic knowledge graph and the shadow knowledge graph are determined based on the same set of evaluation nodes, actual perturbation labels, verification weights, and verification time range. Any additional nodes predicted by the shadow knowledge graph that belong to the evaluation node set should be included in the misidentification evaluation based on whether or not a perturbation actually occurred.

[0230] No. The improvement in propagation error corresponding to each candidate virtual dependency node can be expressed as:

[0231] ;

[0232] In the formula, Indicates the first The amount of propagation error improvement corresponding to each candidate virtual dependency node; This represents the overall propagation error of the basic knowledge graph; This represents the overall propagation error of the corresponding shadow knowledge graph; This indicates that adding candidate virtual dependency topologies reduces propagation error; This indicates that the candidate topology did not improve the situation or introduced additional error propagation.

[0233] By employing homogeneous triggering, parallel propagation, and posterior error comparison, the validity of candidate virtual dependency nodes is no longer solely determined by perturbation similarity, but rather by their ability to reduce missed identification of affected nodes, misidentification of irrelevant nodes, and deviations in perturbation time prediction during subsequent real execution. Figure 6 As shown in the figure, under the same risk-triggered data and propagation rules, the propagation errors of the shadow knowledge graph for nodes such as processing, quality inspection, warehousing, and logistics are all lower than those of the basic knowledge graph, and the corresponding time prediction errors are also given. This figure indicates that the effectiveness of the candidate virtual dependency topology is not determined by the correlation itself, but rather by whether it provides a stable improvement to the node identification errors and time prediction errors of subsequent actual perturbation results.

[0234] Based on the above, after obtaining the improvement in propagation error, the graph state management module performs lifecycle management on candidate virtual dependency nodes and candidate implicit dependency edges. A candidate state indicates that evidence of generation exists but sufficient subsequent verification has not yet been obtained; a valid state indicates that the propagation error can be stably reduced in multiple consecutive valid verification windows; and a failed state indicates that the propagation result cannot be improved even after sufficient verification. Candidate nodes only exist in the shadow knowledge graph, while valid nodes and valid implicit dependency edges are written into the basic knowledge graph and participate in formal risk propagation. Failed nodes and their corresponding edges stop propagating but retain necessary historical records.

[0235] It should be noted that the continuous valid verification window referred to in this embodiment refers to the verification windows arranged in chronological order and all of which have sufficient valid evaluation data. Windows with insufficient valid evaluation data are not included in the number of continuous valid verification windows. The time sequence information of the window can be retained, but it is not used as the basis for determining whether the candidate relationship is valid or invalid.

[0236] When the propagation error improvement of the same candidate virtual dependency node reaches the improvement threshold within a consecutive preset number of valid verification windows, the graph state management module writes the candidate virtual dependency node and its candidate implicit dependency edge into the basic knowledge graph, and synchronously updates the candidate virtual dependency node, evidence aggregation edge and risk propagation edge to a valid state.

[0237] When only a portion of the propagation error improvement amounts in the most recent set of valid verification windows reach the improvement threshold, the graph state management module maintains the candidate state of the candidate virtual dependency node and the candidate implicit dependency edge, and continues to perform rolling judgments after obtaining new valid verification windows; when none of the propagation error improvement amounts in the most recent set of valid verification windows reach the improvement threshold, the graph state management module removes the candidate virtual dependency node and its candidate implicit dependency edge from the shadow knowledge graph and updates them to the invalid state.

[0238] For valid virtual dependency nodes and valid implicit dependency edges already written into the basic knowledge graph, the system continues to verify their continued validity using subsequent consecutive valid verification windows. When the improvement in propagation error corresponding to these nodes is lower than the maintenance threshold within a preset number of consecutive valid verification windows, or when they continuously produce a negative improvement to the formal risk propagation result, the graph state management module stops the virtual dependency node and its implicit dependency edges from participating in formal risk propagation and updates them to a pending verification state or a historical invalid state according to preset state rules. If new evidence that meets the candidate generation conditions is subsequently obtained, the candidate virtual dependency topology is regenerated and verified according to the new generation window and the new version of the basic knowledge graph, instead of directly restoring the original state.

[0239] The aforementioned thresholds and weight parameters can be calibrated offline based on the confirmed positive and negative samples and the corresponding actual perturbation results, and can be updated periodically provided that the evaluation criteria remain consistent.

[0240] To further illustrate the operation of the above technical solution, the following uses three supply execution nodes as an example to explain the data transmission relationship between each processing step.

[0241] In this example, the three supply execution nodes correspond to the incoming material receiving unit, processing unit, and quality inspection unit within the same production area, respectively. The evaluation cycle is one hour. The basic knowledge graph has recorded the explicit relationships between the three supply execution nodes and their respective material nodes, process nodes, and logistics nodes, but it has not recorded a transit and handling resource that the three supply execution nodes actually use together.

[0242] The data processing module unifies the coding of local objects in the warehouse management system, production execution system, and quality inspection system, and associates the warehousing records, start records, completion records, and inspection records corresponding to the same production task according to the execution batch identifier. After correcting the time records in different systems, it forms time-series execution data according to an hourly evaluation cycle. For evaluation cycles that lack valid records, an invalid mark is set, and zero values ​​are not used to fill in the gaps.

[0243] The disturbance feature generation module determines the corresponding original state deviation value based on the material arrival time, process waiting time, and detection start time, and completes the standardization process based on the valid data of each supply execution node within the historical window; only the evaluation cycle that meets the corresponding disturbance activation condition is used as valid disturbance evidence for subsequent delay correlation calculation.

[0244] The implicit dependency identification module determines the structural correlation degree of each supply execution node pair based on the existing material, process, and logistics relationships among the three supply execution nodes. Since the structural correlation degree of each supply execution node pair reaches the preset structural threshold, the three supply execution nodes are classified into the same candidate dependency set.

[0245] Within the generation window, the implicit dependency identification module performs time shifting and sequence difference comparison on the effective perturbation sequences of each supply execution node, determining that the perturbation of the receiving unit leads the processing unit by two evaluation cycles, the perturbation of the quality inspection unit by three evaluation cycles, and the perturbation of the processing unit leads the quality inspection unit by one evaluation cycle. Based on the optimal time delay relationship between the above node pairs, the node time delay offsets of the receiving unit, processing unit, and quality inspection unit relative to the candidate virtual dependency events can be determined as zero, two evaluation cycles, and three evaluation cycles, respectively.

[0246] When the implicit dependency confidence determined by the set temporal correlation degree and set structural correlation degree reaches the candidate threshold and the corresponding node quantity condition is met, the implicit dependency identification module generates candidate virtual dependency nodes to represent unregistered transit resources, and establishes evidence aggregation edges and risk propagation edges between the candidate virtual dependency nodes and the corresponding supply execution nodes. The edge weight of each candidate implicit dependency edge is determined according to the implicit dependency confidence degree and the node correlation support degree of the corresponding supply execution node, and the corresponding propagation delay is determined according to the node delay offset of each supply execution node.

[0247] After entering the verification window, if the receiving unit first reaches the risk triggering condition, the risk propagation identification module inputs the same risk triggering data into the basic knowledge graph and the shadow knowledge graph, and performs risk propagation according to the same propagation rules. The shadow knowledge graph predicts the delay risk of the processing unit and the quality inspection unit in the subsequent corresponding evaluation period based on the candidate virtual dependency topology and node delay offset.

[0248] The shadow topology verification module determines the actual perturbation node and actual perturbation time based on the execution data subsequently collected within the verification window, and calculates the propagation error of the basic knowledge graph and the shadow knowledge graph relative to the actual perturbation result, respectively. When the propagation error improvement of the same candidate virtual dependency node reaches the improvement threshold within a consecutive preset number of valid verification windows, the candidate virtual dependency node and its candidate implicit dependency edges are written into the basic knowledge graph and updated to a valid state. Subsequently, the risk propagation identification module can use this valid virtual dependency node to supplement the unregistered common resource propagation paths in the original basic knowledge graph.

[0249] As Figure 7 As shown in (a), within multiple consecutive valid verification windows, the propagation error of the shadow knowledge graph is consistently lower than that of the basic knowledge graph, and the improvement in propagation error in each window meets the corresponding conditions before the candidate virtual dependency topology is updated to a valid state; the accidental improvement in a single window is insufficient to complete the write operation. Figure 7 As shown in (b), after the candidate topology is written, the system can output the risk intensity distribution of each supply execution node according to the evaluation cycle, so that the risk gradually appears at different time positions along the verified propagation path. Thus, the map update is completed based on continuous posterior verification and can transform implicit dependencies into temporal propagation results that can be used for formal risk identification.

[0250] When the logistics route is adjusted, the transit and handling resources are replaced, or the common dependencies between supply execution nodes change, if the effective virtual dependency node no longer meets the improvement conditions required to maintain the effective state within the subsequent consecutive effective verification window, it will stop participating in the formal risk propagation and be updated to the pending verification state or the historical failure state. In this way, the system completes the closed-loop processing from missing relationship discovery, candidate topology generation, subsequent propagation verification, formal writing to failure exit.

[0251] Time-series execution data can be stored in a time-series database, while basic knowledge graphs and shadow knowledge graphs can be stored in a graph database. The two can be located in independent graph spaces or distinguished in the same graph database by graph type, candidate identifier, and status field. When there are a large number of nodes, they can be processed in parallel by partitioning according to material category, key process, production area, or logistics area, and a corresponding shadow graph or shadow subgraph can be configured for each candidate virtual dependency node.

[0252] In some embodiments, a knowledge graph-based intelligent identification method for supply chain risks is also provided, the method comprising:

[0253] Step 1: Process the execution data representing the supply execution status into time-series execution data arranged according to the evaluation cycle, and construct a basic knowledge graph including supply execution nodes and explicit relationship edges based on the time-series execution data and the registered material relationships, process relationships, and logistics relationships. Step 2: Generate an execution perturbation sequence for each supply execution node based on its state deviation relative to a historical baseline. Step 3: Form a candidate dependency set based on the explicit relationship edges, and compare the time-shift differences of the execution perturbation sequences within the candidate dependency set and under a preset allowable delay set to determine the perturbation associations between supply execution nodes. Based on this, generate a candidate virtual dependency topology to represent unregistered common dependencies. The candidate virtual dependency topology includes candidate virtual dependency nodes and candidate implicit dependencies. Step 4: Add the candidate virtual dependency topology to a copy of the basic knowledge graph to form a shadow knowledge graph; Step 5: Input the same risk trigger data into the basic knowledge graph and the shadow knowledge graph in the verification window, and perform risk propagation according to the same propagation rules to obtain a first prediction result and a second prediction result; Step 6: Determine the actual perturbation result based on the actual execution data collected in the verification window, compare the actual perturbation result with the first prediction result and the second prediction result respectively, determine the propagation error improvement amount; and perform write, keep or revoke operations on the candidate virtual dependency topology according to the propagation error improvement amount, and only write the verified and valid candidate virtual dependency topology into the basic knowledge graph to participate in subsequent risk propagation.

[0254] It should be noted that each step of the above method corresponds to the processing of the aforementioned data processing module, basic knowledge graph construction module, perturbation feature generation module, implicit dependency identification module, shadow topology verification module, risk propagation identification module, and graph state management module. The specific implementation method, calculation formula, and boundary processing of each step can be referred to the description of the corresponding module above, and will not be repeated here.

[0255] Furthermore, the aforementioned modules can be implemented by program units running on the same processor, or deployed on interconnected data processing nodes, graph computing nodes, and graph storage nodes. The system can be implemented by an electronic device including a processor and memory, wherein the processor executes the computer program stored in the memory to implement the aforementioned processing procedures. Further, the aforementioned computer program can be stored in a computer-readable storage medium, and the computer program stored on the computer-readable storage medium, when executed by the processor, implements the aforementioned processing procedures. The computer-readable storage medium can be a read-only memory, random access memory, magnetic disk, optical disk, or flash memory, or other non-volatile or volatile storage media.

[0256] The above specific embodiments are used to illustrate the technical solution and implementation process of the present invention, and are not intended to limit the present invention. For those skilled in the art, corresponding adjustments can be made to the data source, calculation parameters, storage method, module deployment method, or propagation evaluation method without departing from the technical concept of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A knowledge graph-based intelligent identification system for supply chain risks, characterized in that, include: The data processing module is used to acquire execution data and generate time-series execution data; The basic knowledge graph construction module is used to generate supply execution nodes based on the time-series execution data, and to establish explicit relationship edges based on the registered material relationships, process relationships and logistics relationships in order to construct the basic knowledge graph; The disturbance feature generation module is used to generate an execution disturbance sequence based on the state deviation of each supply execution node relative to the historical baseline. The implicit dependency identification module is used to form a candidate dependency set based on the explicit relation edges, determine the degree of perturbation association between supply execution nodes in the candidate dependency set within the allowable delay set, and generate candidate virtual dependency nodes and candidate implicit dependency edges based on the degree of perturbation association. The shadow topology verification module is used to add the candidate virtual dependency nodes and candidate implicit dependency edges to the copy of the basic knowledge graph to form a shadow knowledge graph. The risk propagation identification module is used to input the same risk trigger data into the basic knowledge graph and the shadow knowledge graph within the verification window, execute risk propagation according to the same propagation rules, obtain a first prediction result and a second prediction result, and output the risk propagation result based on the updated basic knowledge graph. The shadow topology verification module is also used to determine the actual perturbation result based on the actual execution data collected in the verification window, determine the propagation error of the first prediction result and the second prediction result relative to the actual perturbation result, and determine the propagation error improvement amount based on the difference between the propagation error of the basic knowledge graph and the shadow knowledge graph. The graph state management module is used to write the candidate virtual dependency nodes and candidate implicit dependency edges into the basic knowledge graph or remove them from the shadow knowledge graph based on the propagation error improvement amount.

2. The knowledge graph-based intelligent supply chain risk identification system as described in claim 1, characterized in that, The data processing module generates time-series execution data, specifically including: The execution data is grouped into objects based on the supply execution object identifier, and the grouped execution data is batch-associated based on the execution batch identifier; The associated execution data is converted to a unified time base and arranged according to a preset evaluation cycle to form time-series execution data corresponding to each supply execution object; When any supply execution object lacks the valid data required to form a state deviation in any assessment period, an invalid flag is set for that assessment period so that the corresponding data is not involved in the determination of the degree of subsequent disturbance correlation.

3. The knowledge graph-based intelligent supply chain risk identification system as described in claim 1, characterized in that, The basic knowledge graph construction module constructs the basic knowledge graph, specifically including: A supply execution node is generated based on the supply execution object identifier in the time-series execution data; Based on the material object identifier, process object identifier, and logistics object identifier recorded in the time-series execution data, corresponding associated nodes are generated; Based on the registered material correspondence, process routes, and logistics paths, establish explicit relationship edges between the supply execution nodes and their corresponding associated nodes; For multiple time-series execution data with the same object identifier, the status data in the multiple time-series execution data is written to or updated in the node attribute or time-series attribute of the corresponding graph node, without generating the corresponding graph node repeatedly.

4. The knowledge graph-based intelligent supply chain risk identification system as described in claim 2, characterized in that, The perturbation feature generation module generates an execution perturbation sequence, specifically including: The deviation value of the original status is determined based on the execution status of the supply execution nodes in each evaluation period; The historical benchmark and historical fluctuation scale are determined based on the execution data of the supply execution nodes within the historical window. Based on the original state deviation, historical benchmark, and historical fluctuation scale, the standardized perturbation value for each evaluation period is determined, and the standardized perturbation value is arranged according to the evaluation period to form an execution perturbation sequence; Based on the standardized perturbation value and the perturbation activation condition corresponding to the perturbation type, the effective perturbation period for perturbation correlation calculation is determined. The evaluation period with the invalidation mark is not involved in the correlation calculation between different execution perturbation sequences.

5. The knowledge graph-based intelligent supply chain risk identification system as described in claim 4, characterized in that, The implicit dependency identification module forms a candidate dependency set, specifically including: Based on the material association, process association, and logistics association represented by the explicit relation edges, determine the structural association degree between supply execution node pairs; Supply execution node pairs whose structural correlation reaches a preset structural threshold are identified as candidate relationships; A candidate relationship graph is formed based on the candidate relationships, and multiple candidate supply execution nodes that are interconnected in the candidate relationship graph are assigned to the same candidate dependency set, so as to limit the determination range of the degree of disturbance association to the same candidate dependency set.

6. The knowledge graph-based intelligent supply chain risk identification system as described in claim 5, characterized in that, The implicit dependency identification module determines the degree of perturbation correlation within the allowable delay set, specifically including: Based on the process processing time, logistics delivery time, execution buffer time, or historical abnormal delivery time between candidate supply execution nodes, determine the set of allowable delays consisting of multiple candidate delays; For each candidate delay, the execution perturbation sequence of one candidate supply execution node is shifted in time relative to the execution perturbation sequence of another candidate supply execution node; The sequence differences are determined within the evaluation period when both execution perturbation sequences after time shift have valid data and the corresponding perturbation values ​​meet the corresponding perturbation activation conditions. When the number of effective perturbation pairing cycles reaches a preset lower limit, the candidate delay that minimizes the sequence difference is determined as the optimal delay, and the degree of perturbation correlation between the two candidate supply execution nodes is determined based on the corresponding minimum sequence difference.

7. The knowledge graph-based intelligent supply chain risk identification system as described in claim 6, characterized in that, The implicit dependency identification module generates candidate virtual dependency nodes and candidate implicit dependency edges, specifically including: The implicit dependency confidence of the candidate dependency set is determined based on the degree of perturbation association and structural association of each candidate supply execution node pair within the candidate dependency set. When the implicit dependency confidence reaches the candidate threshold and the number of candidate supply execution nodes that meet the effective perturbation association conditions reaches a preset number, candidate virtual dependency nodes are generated to represent unregistered common dependencies, and candidate implicit dependency edges are established between each candidate supply execution node and the candidate virtual dependency node. The candidate implicit dependency edge includes an evidence aggregation edge pointing from the candidate supply execution node to the candidate virtual dependency node and a risk propagation edge set in reverse; the edge weight of the candidate implicit dependency edge is determined according to the implicit dependency confidence and the degree of perturbation correlation between the corresponding candidate supply execution node and other nodes in the set. The node delay offset of each candidate supply execution node is determined based on the optimal delay of each candidate supply execution node pair, and the propagation delay when risk propagation occurs through the candidate virtual dependency node is determined based on the node delay offset.

8. The knowledge graph-based intelligent supply chain risk identification system as described in claim 7, characterized in that, The shadow topology verification module forms a shadow knowledge graph, specifically including: For each candidate virtual dependency node, copy all or part of the graph structure related to the verification of that candidate virtual dependency node in the basic knowledge graph; The candidate virtual dependency node and candidate implicit dependency edge are added to the copied graph structure to form a shadow knowledge graph corresponding to the candidate virtual dependency node. The graph structure used as the comparison benchmark has the same original nodes, effective relation edges, node temporal states, propagation parameters, and evaluation node set as the shadow knowledge graph. The topological difference between the two is the current candidate virtual dependency node and candidate implicit dependency edge. During the verification process, the baseline topology of the basic knowledge graph is kept unchanged due to the current candidate virtual dependency node, and the newly added execution data in the verification window is synchronized to update the temporal state of the corresponding nodes in the two graphs. The data range used to generate candidate virtual dependency nodes is set as the generation window, and the data range used to verify candidate virtual dependency nodes is set as the verification window. The verification window is located after the generation window and does not overlap with the generation window.

9. The knowledge graph-based intelligent supply chain risk identification system as described in claim 8, characterized in that, The risk propagation identification module performs risk propagation, and the shadow topology verification module compares propagation errors, specifically including: Within the verification window, the same risk trigger data is input into the basic knowledge graph and the shadow knowledge graph; Risk propagation is performed according to the same relational edge direction, edge weight usage rules, propagation delay usage rules, maximum propagation level, and termination conditions to obtain the first and second prediction results. The actual disturbance node and actual disturbance time are determined based on the actual execution data collected within the verification window; Under the same set of evaluation nodes and verification weights, the propagation errors of the first and second prediction results relative to the actual perturbation results are determined respectively, and the amount of propagation error improvement is determined based on the difference between the propagation errors of the two spectra.

10. The knowledge graph-based intelligent supply chain risk identification system as described in claim 9, characterized in that, The graph state management module updates the state of candidate virtual dependency nodes and candidate implicit dependency edges, specifically including: When the propagation error improvement amount of the same candidate virtual dependency node reaches the improvement threshold within a preset number of consecutive valid verification windows, the candidate virtual dependency node and the candidate implicit dependency edge are written into the basic knowledge graph, and the candidate virtual dependency node and the candidate implicit dependency edge are synchronously updated to the valid state. When there is insufficient valid verification data in the verification window, the verification window is not counted in the number of consecutive valid verification windows, and the candidate status of the candidate virtual dependency node is maintained. When only a portion of the propagation error improvement in the most recent preset number of consecutive valid verification windows reaches the improvement threshold, the candidate state of the candidate virtual dependency node is maintained, and rolling verification is performed using subsequent valid verification windows. When the propagation error improvement amount of the same candidate virtual dependency node does not reach the improvement threshold within a preset number of consecutive valid verification windows, the candidate virtual dependency node and the candidate implicit dependency edge are removed from the shadow knowledge graph and updated to an invalid state. For virtual dependency nodes and implicit dependency edges that have been written into the basic knowledge graph, if they fail to meet the improvement conditions required to maintain an effective state within a preset number of subsequent consecutive valid verification windows, they will be stopped from participating in formal risk propagation and updated to a pending verification state or a historical failure state.

Images