Converged agent provider risk intelligence monitoring method and system

Abstract

The application discloses a supplier risk intelligent monitoring method and system of fusion intelligent agent, and belongs to the technical field of risk intelligent monitoring; the method specifically comprises the following steps: obtaining supplier business data, constructing a supplier risk cause-effect diagram, and obtaining a supplier graph signal set; based on the supplier graph signal set, a supplier risk monitoring evolution equation is constructed, and a supplier risk scenario field is obtained; based on the supplier risk scenario field, a fusion intelligent agent risk pathway hierarchical diagram is constructed, and the supplier risk is monitored; the application improves the accuracy of the supplier risk monitoring result, accelerates the identification response speed of the hidden risk and the cumulative risk, and enhances the stability of the supplier risk monitoring.

Description

Technical Field

[0001] This invention belongs to the field of intelligent risk monitoring, specifically relating to a method and system for intelligent risk monitoring of suppliers with integrated intelligent agents. Background Technology

[0002] Within the group's electronic procurement platform, the entire supplier lifecycle management process involves multi-source business data, including supplier onboarding, qualification changes, transaction behavior, abnormal procurement, and external regulatory feedback. This data is scattered across different business systems and modules, manifesting as heterogeneous event flows across stages and timeframes. Risk signals exhibit weak correlations and fragmented characteristics, making it difficult to form a continuous and interpretable description of risk evolution from a unified perspective. This increases the complexity of supplier risk identification and dynamic monitoring.

[0003] In existing technologies, supplier risk monitoring methods are typically based on scoring models, threshold rules, or static statistical analysis, using weighted summaries or single-point early warning processing for abnormal qualifications, transaction fluctuations, and blacklist information. However, these methods lack the ability to characterize the temporal correlations and causal propagation paths between multi-source business events, making it difficult to continuously identify and address hidden and cumulative risks. Furthermore, in the multi-strategy collaborative analysis involving integrated intelligent agents, existing solutions cannot construct a unified cross-module event flow driven by the integrated intelligent agents, nor can they automatically identify and classify complex supplier risk scenarios, resulting in poor accuracy and stability in supplier risk monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a supplier risk intelligent monitoring method and system with integrated intelligent agents. By constructing a supplier risk causal graph and generating a graph signal set, combining graph frequency domain decomposition and risk evolution deduction to form a continuous risk scenario field, and constructing a risk path hierarchical graph of integrated intelligent agents to achieve collaborative judgment, the system realizes dynamic evolution modeling of supplier risk and joint identification of multiple intelligent agents, thereby improving the accuracy and stability of monitoring.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A supplier risk intelligent monitoring method integrating intelligent agents includes:

[0007] Obtain supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set;

[0008] Based on the supplier graph signal set, a supplier risk monitoring evolution equation is constructed to obtain the supplier risk scenario field. The supplier risk monitoring evolution equation is obtained by performing graph frequency domain decomposition on the supplier graph signal set, extracting long-term risk trend features and short-term abnormal risk features, performing risk evolution mode decomposition on the multidimensional risk monitoring state vector, and performing risk-time slice extrapolation on the low-dimensional risk monitoring state vector sequence.

[0009] Based on the supplier risk scenario, a risk path hierarchy map of the integrated intelligent agent is constructed to monitor supplier risks.

[0010] Specifically, acquire supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set, including:

[0011] Obtain supplier business data, divide the supplier business data into risk time slices, generate business scenario keys, and obtain a candidate set of supplier business events;

[0012] Risk sensitivity joint discrimination is performed on each supplier business candidate event in the supplier business event candidate set, and normalized coding is performed to obtain the supplier risk event set;

[0013] Based on the supplier risk event set, risk nodes are generated through node mapping, and supplier risk relationship edges are established between each risk node to obtain a preliminary graph of multiple supplier risk relationships. The risk nodes include supplier nodes, event nodes, and scenario nodes.

[0014] Based on the initial graph of multiple relationships of supplier risks, causal determination and merging are performed on the risk relationship edges of each supplier, and risk nodes are aggregated to obtain the supplier risk causal graph.

[0015] Based on the supplier risk causal graph, a time series of risk characteristics corresponding to each supplier node is constructed to obtain the supplier graph signal set.

[0016] Specifically, based on the supplier risk causal graph, a time series of risk characteristics corresponding to each supplier node is constructed to obtain the supplier graph signal set, including:

[0017] Based on the supplier risk causal graph, the category and intensity distribution of each event node adjacent to each supplier node in each risk time slice are statistically analyzed to obtain the supplier risk neighborhood sequence.

[0018] Based on the supplier risk neighborhood sequence, the category and intensity distribution within each risk time slice are aggregated into a multidimensional risk feature vector, and time-series alignment is performed to obtain the risk feature time series corresponding to each supplier.

[0019] The risk characteristic time series is mapped to the supplier risk monitoring signal sequence corresponding to each supplier node, and graph smoothing and anomaly retention processing are performed to obtain the supplier graph signal set.

[0020] Specifically, based on the supplier graph signal set, a supplier risk monitoring evolution equation is constructed to obtain the supplier risk scenario field, including:

[0021] Based on the supplier graph signal set, graph frequency domain multi-scale decomposition is performed to extract long-term risk trend features and short-term abnormal risk features, resulting in a multi-dimensional risk monitoring state vector sequence.

[0022] Based on the multidimensional risk monitoring state vector sequence, a supplier risk monitoring evolution equation is constructed, and risk-by-risk time slice extrapolation is performed to obtain the risk monitoring state trajectory of each supplier in each risk time slice.

[0023] Based on the risk monitoring status trajectory, the risk monitoring status trajectory values ​​of each supplier in each risk time slice are extracted, and the risk monitoring status trajectory values ​​are spliced ​​together in the same time slice to obtain a risk monitoring status field arranged according to the risk time slice.

[0024] Based on the risk monitoring state field, the risk monitoring state trajectory values ​​of adjacent risk time slices in the risk monitoring state field are interpolated and smoothed to obtain the supplier risk scenario field.

[0025] Specifically, based on the supplier graph signal set, multi-scale decomposition in the graph frequency domain is performed to extract long-term risk trend features and short-term abnormal risk features, resulting in a multi-dimensional risk monitoring state vector sequence, including:

[0026] Based on the supplier graph signal set, the risk monitoring signal sequences of each supplier are uniformly preprocessed and time-aligned to obtain the supplier risk monitoring signal matrix.

[0027] Based on the supplier risk causal graph, a supplier risk adjacency operator and a graph Laplacian operator are constructed, and the supplier risk monitoring signal matrix is ​​mapped to the graph frequency domain to obtain the supplier risk graph spectrum.

[0028] The supplier risk map spectrum is divided into frequency bands, and the signal components are reconstructed to obtain low-frequency risk components and high-frequency risk components;

[0029] Based on low-frequency and high-frequency risk components, long-term risk trend features and short-term abnormal risk features are extracted, and feature aggregation is performed within each risk time slice to obtain a multi-dimensional risk monitoring state vector sequence.

[0030] Specifically, based on low-frequency and high-frequency risk components, long-term risk trend features and short-term abnormal risk features are extracted, and feature aggregation is performed within each risk time slice to obtain a multi-dimensional risk monitoring state vector sequence, including:

[0031] Based on low-frequency and high-frequency risk components, the low-frequency and high-frequency risk component values ​​of each supplier in each risk time slice are organized to construct the supplier risk time slice feature tensor.

[0032] Based on the supplier risk time slice feature tensor, a segmented low-rank decomposition is performed with risk time slice as the segment unit to obtain long-term risk trend features and residual sub-tensor.

[0033] Based on the residual subtensor, a risk anomaly energy function is constructed, and the risk anomaly energy function is minimized and reconstructed to obtain short-term anomaly risk characteristics.

[0034] Within each risk time slice, the long-term risk trend characteristics and short-term abnormal risk characteristics are normalized and spliced ​​together to obtain a multi-dimensional risk monitoring state vector sequence.

[0035] Specifically, based on the multidimensional risk monitoring state vector sequence, a supplier risk monitoring evolution equation is constructed, and a risk-by-risk time-slice extrapolation is performed to obtain the risk monitoring state trajectory of each supplier in each risk time-slice, including:

[0036] Risk evolution mode decomposition is performed on the multidimensional risk monitoring state vector sequence to extract the supplier risk evolution fundamental mode;

[0037] A low-dimensional risk evolution subspace is constructed based on the supplier risk evolution basic model, and the multi-dimensional risk monitoring state vector is projected to obtain a low-dimensional risk monitoring state vector sequence.

[0038] Clustering and segmenting of low-dimensional risk monitoring state vector sequences to identify risk evolution state types;

[0039] Based on the risk evolution state type, the temporal evolution relationship between low-dimensional risk monitoring state vectors is determined, and parameter fitting is performed to obtain the supplier risk monitoring evolution equation.

[0040] Based on the supplier risk monitoring evolution equation, the supplier risk monitoring state trajectory is obtained by performing risk-by-risk time slice extrapolation on the low-dimensional risk monitoring state vector sequence.

[0041] Specifically, based on supplier risk scenarios, a risk path hierarchy map of the integrated intelligent agent is constructed to monitor supplier risks, including:

[0042] The supplier risk monitoring rules are analyzed, a set of integrated intelligent agents is constructed, and a corresponding set of risk scenario feature path templates is generated.

[0043] Risk scenario fragments are extracted from the supplier risk scenario field and matched with the risk scenario feature path template set to obtain the supplier risk path set of the fusion agent.

[0044] Based on the supplier risk causal graph, causal time sequence correction is performed on the risk paths corresponding to different fusion intelligent agents to obtain a set of corrected risk paths;

[0045] Based on the set of corrected risk paths, a hierarchical risk path diagram of the fusion agent is constructed to identify overlapping risk paths, and path rewriting and risk assessment are performed to obtain a set of supplier risk assessment results.

[0046] Based on the supplier risk assessment results set, the dominant risk sources and collaborative risk sources are identified, and joint assessment and risk classification are performed to obtain the supplier risk monitoring results.

[0047] Specifically, a risk path hierarchy map of the fusion agent is constructed based on the corrected risk path set to identify overlapping risk paths, and path rewriting and risk assessment are performed to obtain a set of supplier risk assessment results, including:

[0048] Based on the supplier risk cause-effect diagram, the common preceding risk events of different fusion intelligent agent risk paths are identified, and the risk paths are corrected for their time and location.

[0049] The corrected risk paths are mapped as path nodes. Time continuity edges are established within the same fused intelligent body. Path intersection edges are established between risk paths of different fused intelligent bodies that have common preceding risk events and whose corrected times overlap, resulting in a risk path layering graph of the fused intelligent body.

[0050] Based on the risk path hierarchy map of the fused intelligent agent, the overlapping risk paths are extracted and the dominant and cooperative paths are identified.

[0051] Overlapping path segments in the dominant and collaborative paths are merged, non-overlapping path segments in the collaborative paths are connected to the dominant path, and risk assessment is performed to obtain a set of supplier risk assessment results.

[0052] A supplier risk intelligent monitoring system integrating intelligent agents includes:

[0053] The data acquisition module is used to acquire supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set;

[0054] The risk evolution module, based on the supplier graph signal set, constructs the supplier risk monitoring evolution equation to obtain the supplier risk scenario field;

[0055] The risk monitoring module, based on the supplier risk scenario, constructs a risk path hierarchy diagram of the integrated intelligent agent to monitor supplier risks.

[0056] Compared with existing technologies, the beneficial effects of this invention include: constructing a supplier risk causal graph based on supplier business data and extracting a supplier graph signal set; performing multi-scale decomposition and risk evolution mode analysis on the supplier graph signal set in the graph frequency domain, constructing a supplier risk monitoring evolution equation, and deriving a continuously evolving supplier risk scenario field; parsing supplier risk monitoring rules on the supplier risk scenario field, generating a fusion intelligent agent and corresponding risk scenario feature path templates, constructing a risk path hierarchical graph of the fusion intelligent agent, and obtaining supplier risk monitoring results through path rewriting and collaborative judgment by the fusion intelligent agent. This solves the problem in existing technologies of difficulty in sequentially associating cross-module, multi-type business events under the unified drive of a fusion intelligent agent, and automatically identifying and classifying complex supplier risk scenarios, thereby improving the accuracy of supplier risk monitoring results, accelerating the identification and response speed of hidden and cumulative risks, and enhancing the stability of supplier risk monitoring. Attached Figure Description

[0057] Figure 1 Flowchart of the supplier risk intelligent monitoring method for the fusion intelligent agent provided by the present invention;

[0058] Figure 2 This invention provides a schematic diagram of the risk feature evolution mapping.

[0059] Figure 3 A flowchart for generating the risk responsibility allocation results of the fusion intelligent agent provided by this invention;

[0060] Figure 4 The structural diagram of the supplier risk intelligent monitoring system for the fusion intelligent agent provided by the present invention. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0062] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0063] Example 1

[0064] Please see Figures 1-3 The present invention provides an embodiment of a supplier risk intelligent monitoring method for integrated intelligent agents, comprising the following specific steps:

[0065] Step S1: Obtain supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set.

[0066] The specific steps of step S1 are as follows:

[0067] Step S101: Obtain supplier business data, divide the supplier business data into risk time slices, generate business scenario keys, and obtain a candidate set of supplier business events.

[0068] In this embodiment, supplier registration information, qualification information, quotation and transaction information, order and goods receipt settlement information, and external regulatory feedback information recorded in the electronic procurement platform are obtained. The supplier-related business records in each business system are extracted and cleaned, and the original records in different business systems are uniformly converted into supplier business data containing supplier identification, business occurrence time, business type, procurement method, material category, procurement project identification, bid section identification, and quantity and amount information.

[0069] The supplier business data undergoes processes including anomaly removal, duplicate record merging, missing field marking, and time format standardization. Specifically, business records lacking supplier identification, business occurrence time, or business type, and which cannot be completed using related fields from the same business record, are removed. For duplicate business records with identical supplier identification, business occurrence time, and business type, only the record with the most recent business update time is retained.

[0070] The risk time granularity is determined based on the actual occurrence interval of supplier business events and the allowable response time for risk monitoring. Specifically, the determination method is as follows: based on the historical confirmed risk evolution process, historical supplier business events are grouped according to supplier identifier and business scenario key. Within each group, events are arranged from earliest to latest according to their occurrence time. The occurrence time of the next business event is subtracted from the occurrence time of the previous business event to obtain the time interval between adjacent risk events. These intervals are then aggregated to form a sample set of adjacent risk event time intervals.

[0071] The time intervals of adjacent risk events are arranged in ascending order, and the cumulative proportion of each time interval in the entire sample is calculated. The coverage ratio of the main business change process is used to retain the majority of continuous risk events while eliminating a small number of tail intervals formed by long-term unrelated business. The coverage ratio is determined based on the empirical distribution of historical time intervals of adjacent risk events. Specifically, the method is as follows: different cumulative proportions are used to extract time interval samples, and the upper limit of the corresponding time interval and the integrity rate of the retained risk evolution chain are compared. The previous cumulative proportion, where the increase in the integrity rate of the risk evolution chain slows down while the upper limit of the time interval increases significantly, is selected as the coverage ratio. For example, in the embodiments of this application, the coverage ratio of the main business change process is preferably 85% to 95%.

[0072] The time interval corresponding to reaching the aforementioned coverage ratio is determined as the upper limit of the main business change time interval, and adjacent risk events less than or equal to the upper limit of the time interval are taken as continuous business change samples. The median of the time intervals corresponding to the continuous business change samples is calculated, and the median is used as the basic risk time granularity.

[0073] The timeliness of risk monitoring requires quantification using the maximum monitoring response time. The specific determination method is as follows: For historically confirmed risk events, extract the time of occurrence of the first risk-sensitive business event and the time of confirmation of the corresponding risk status. Use the time difference between the time of risk status confirmation and the time of occurrence of the first risk-sensitive business event as the historical risk response time, obtaining a sample set of risk response times. Arrange the risk response times in ascending order, and use the quantile value of the response time that covers the main confirmed risk events as the maximum monitoring response time. The response coverage ratio is set according to the monitoring system's requirement to promptly identify most risk events, and is verified using historical data playback. For example, in this embodiment, the response coverage ratio is preferably 75% to 90%.

[0074] Compare the basic risk time granularity with the maximum monitoring response time. If the basic risk time granularity is not greater than the maximum monitoring response time, the basic risk time granularity is used as the risk time granularity; if the basic risk time granularity is greater than the maximum monitoring response time, the maximum monitoring response time is used as the risk time granularity. Round the risk time granularity to the smallest time unit used by the system.

[0075] Based on the determined risk time granularity, the business occurrence time in the supplier business data is mapped onto a time axis, the continuous time axis is divided into ordered risk time slices, and the supplier business data is classified according to the risk time slice to which the business occurrence time of each supplier business data belongs, thus obtaining the supplier business data set corresponding to each risk time slice.

[0076] The data is based on a combination of fields from the supplier's business data, including procurement project identifier, bid section identifier, material category, procurement method, and business type. A combined code is generated according to a fixed field order of procurement project identifier, bid section identifier, material category, procurement method, and business type, and this combined code serves as the business scenario key.

[0077] Supplier business data is aggregated based on supplier identifier, risk time slice, and business scenario key. Business data with the same supplier identifier, risk time slice, and business scenario key are grouped together, and each group of supplier business data is used as a candidate supplier business event to obtain a candidate set of supplier business events.

[0078] Step S102: Perform risk-sensitive joint discrimination on each supplier business candidate event in the supplier business event candidate set, and perform normalized coding to obtain the supplier risk event set.

[0079] In this embodiment, a set of risk-sensitive fields is determined based on the supplier's historical risk event records. First, according to the supplier identifier and risk confirmation time of the historical risk event, supplier business records within a major business change time interval prior to the risk confirmation time are extracted as the risk association observation interval; at the same time, non-risk observation intervals are extracted from historical periods without confirmed risk results, with the same time length.

[0080] For any candidate business field, the number of risk observation intervals in which the business field exhibits abnormal changes within each risk association observation interval is counted. This number is then divided by the total number of risk association observation intervals to obtain the risk association frequency of the business field. For status fields, a field change is determined when the field value changes to enabled, disabled, frozen, invalid, or its level changes. For continuous numerical fields, the normal value range is determined based on the historical value distribution within non-risk observation intervals. A field anomaly is determined when the field value exceeds the normal value range.

[0081] The association filtering criteria are determined based on the risk association frequency distribution of all candidate business fields. Specifically, the risk association frequency of each candidate business field is arranged in ascending order, and the upper quartile of the risk association frequency is taken as the initial association filtering threshold. Verification is performed using historical risk observation intervals and non-risk observation intervals. If the fields filtered using this threshold cannot cover the main historical risk events, the association filtering threshold is gradually reduced until the proportion of historical risk events covered by the filtered fields reaches the coverage proportion of the main business change process determined in step S101. For example, in this embodiment, the risk association frequency corresponding to the association filtering threshold is preferably between 0.25 and 0.45.

[0082] Candidate business fields whose risk association frequency reaches the association screening threshold are identified as risk-sensitive fields. These risk-sensitive fields include qualification information fields, status change fields, quotation and transaction price fields, emergency procurement indicator fields, and external regulatory feedback fields.

[0083] Based on risk-sensitive fields, the magnitude of qualification changes, frequency of status changes, degree of price deviation, density of emergency procurement, and intensity of adverse external regulatory records are extracted to obtain risk feature vectors corresponding to each supplier's business candidate events.

[0084] For the quote and transaction price fields, the short period is determined based on the risk time granularity and the historical rate of price anomaly formation. Specifically, the method is to count the number of risk time slices experienced from the first occurrence of price deviation to the confirmation of the risk status for historically confirmed transaction anomalies, and take the median of this number as the short period length. For example, in this embodiment, the short period preferably covers 2 to 4 risk time slices.

[0085] The historical trading range is determined based on the historical valid quotes and transaction prices corresponding to the current business scenario key. Starting from the valid trading records prior to the current risk time slice, the historical sample window is expanded backwards. For each expanded risk time slice, the historical median price and the historical absolute median price difference are recalculated. Expansion stops when the rate of change of the median price and the rate of change of the absolute median price difference before and after continuous expansion are both lower than the statistical stability threshold. The time range corresponding to the current sample window is then used as the historical trading statistical period. The statistical stability threshold is determined based on the natural fluctuation range of the median and absolute median price difference during historical normal trading periods, taking the high quantile of the rate of change of the statistics during normal periods as the statistical stability threshold. For example, in this embodiment, the statistical stability threshold is preferably 3% to 8%.

[0086] Based on the historical trading statistical period, the median price and the absolute median price difference are calculated. The absolute difference between the current quoted or traded price and the median price is divided by the absolute median price difference to obtain the price deviation multiple. The significant price deviation threshold is determined based on the distribution of price deviation multiples during historical normal trading periods. Specifically, the price deviation multiples during normal trading periods are arranged from smallest to largest, and the highest quantile value that covers the main normal price fluctuations is taken as the significant price deviation threshold. For example, in this embodiment, the significant price deviation threshold is preferably 2.5 to 4.0 times the historical absolute median price difference. When the absolute median price difference is 0, the historical price interquartile range is used instead of the absolute median price difference.

[0087] For similar supplier price levels, other suppliers' valid quotations and transaction prices are screened according to the same material category, procurement method, and risk time slice. The median of these valid prices is calculated as the price benchmark for similar suppliers. When the deviation of the current quotation or transaction price from the historical transaction range or the price benchmark for similar suppliers reaches a significant price deviation threshold, and the deviation persists for a short period, the transaction risk sensitivity condition is determined to be met.

[0088] For the qualification information field, the qualification risk sensitivity condition is determined to be met when the qualification validity status changes from valid to invalid, the qualification level changes to a lower level, or the remaining validity period is lower than the low percentile of the historical renewal lead time for the corresponding qualification type. For the status change field, the status risk sensitivity condition is determined to be met when there is a freeze, suspension, or abnormal status switch. For the emergency procurement identifier field, the number of emergency procurements by the same supplier within the current major business change time interval is counted and compared with the high percentile of the number of emergency procurements by the same supplier within the same length window during the historical normal period. If it exceeds the high percentile, the procurement behavior risk sensitivity condition is determined to be met. For the external regulatory feedback field, the external regulatory risk sensitivity condition is determined to be met when there are structured and marked abnormal regulatory results in the records.

[0089] Each risk characteristic value is normalized. For risk characteristics with higher values ​​indicating higher risk, the current risk characteristic value is subtracted from the minimum value in historical normal samples, and then divided by the difference between the maximum and minimum values ​​in historical normal samples. For state-type risk characteristics, a high-risk state is defined when the corresponding risk sensitivity condition is met, and a normal state is defined when it is not met. The normalized risk characteristic values ​​are limited to 0 to 1.

[0090] The weights of risk features are determined based on the degree of correlation between each risk feature and historical risk results. The specific determination method is as follows: Calculate the proportion of anomalies for each risk feature in the historical risk observation period and the proportion of anomalies in the non-risk observation period. Subtract the proportion of anomalies in the non-risk observation period from the proportion of anomalies in the risk observation period to obtain the correlation increment of that risk feature. Only correlation increments greater than 0 are retained. Divide the correlation increment of each risk feature by the sum of the correlation increments of all retained risk features to obtain the corresponding risk feature weight, ensuring that the sum of all risk feature weights equals 1.

[0091] Each normalized risk feature value is multiplied by its corresponding risk feature weight, and the resulting contribution values ​​of each risk feature are summed to obtain the basic risk trigger strength.

[0092] For supplier business candidate events that simultaneously meet multiple risk-sensitive conditions, a combined enhancement is performed based on the number of conditions met. The specific enhancement method is as follows: first, subtract 1 from the number of conditions met, then multiply this number by the combined enhancement coefficient, and add the result to 1 to obtain the combined enhancement factor; then, multiply the basic risk trigger strength by the combined enhancement factor, and take 1 if the result is greater than 1 to obtain the risk trigger strength.

[0093] The combined enhancement coefficient is determined based on historically confirmed risk events. Specifically, the determination method is as follows: values ​​are sequentially selected from the candidate combined enhancement coefficient range; risk triggering replays of historical supplier business events are performed; the risk event identification accuracy and recall rate are calculated respectively; and the combined enhancement coefficient with the highest overall identification effect is selected. For example, in this embodiment, the combined enhancement coefficient is preferably between 0.05 and 0.20.

[0094] Supplier risk event metadata records are generated by encoding the supplier identifier, risk event type, risk time slice identifier, business scenario key, and risk trigger intensity in a fixed field order. The risk event type, risk time slice identifier, and business scenario key are uniformly encoded, and the risk trigger intensity is maintained within a normalized range of 0 to 1. The encoded supplier risk events are then aggregated to obtain a supplier risk event set.

[0095] Step S103: Based on the supplier risk event set, generate risk nodes through node mapping, and establish supplier risk relationship edges between each risk node to obtain the initial graph of supplier risk multi-relationship.

[0096] In this embodiment, based on the supplier risk event set, each supplier risk event in the supplier risk event set is parsed, and the supplier set, risk event set, and business scenario set participating in each supplier risk event are determined according to fields such as supplier identifier, risk time slice identifier, and business scenario key.

[0097] Based on each supplier in the supplier set, a supplier node is generated. Based on each supplier risk event in the risk event set, an event node is generated. Based on each business scenario key in the business scenario set, a scenario node is generated. The supplier node, event node, and scenario node are then uniformly classified as risk nodes.

[0098] Based on the supplier identifier and risk event identifier in each supplier risk event record in the supplier risk event set, establish business participation relationship edges between supplier nodes and event nodes. Based on the business scenario key in each supplier risk event record, establish scenario affiliation relationship edges between event nodes and scenario nodes. Based on the supplier nodes that appear together in the same risk time slice and the same business scenario key in the supplier risk event set, establish common risk scenario relationship edges between supplier nodes.

[0099] By integrating the business participation relationship edges of suppliers and risk events, the scenario affiliation relationship edges of risk events and business scenarios, and the common risk scenario relationship edges of suppliers, a preliminary multi-relationship graph of supplier risks is constructed, consisting of risk nodes and supplier risk relationship edges.

[0100] Step S104: Based on the initial graph of multiple relationships of supplier risks, perform causal determination and merging on the risk relationship edges of each supplier, and aggregate the risk nodes to obtain the supplier risk causal graph.

[0101] In this embodiment, based on the initial graph of multiple relationships of supplier risks, the risk node types, risk time slice order and business scenario keys of each supplier risk relationship edge are parsed to obtain candidate risk node pairs.

[0102] For any pair of candidate risk nodes with a temporal relationship, the risk node that occurs earlier is identified as the preceding risk node, and the risk node that occurs later is identified as the subsequent risk node. The causal observation time range adopts the upper limit of the time interval of major business changes determined in step S101.

[0103] The frequency of occurrence of preceding and subsequent events is expressed as the proportion of occurrence of conditions. The specific calculation process is as follows: count the total number of occurrences of risk events corresponding to preceding risk nodes in historical data, count the number of preceding risk events that occur within the causal observation time range corresponding to subsequent risk nodes, and divide the latter by the former to obtain the frequency of occurrence of subsequent risk nodes.

[0104] Consistency in risk state changes is determined based on the direction of change in risk trigger intensity. For each set of consecutive risk events, the risk trigger intensity corresponding to the time slice before the preceding risk event and the subsequent risk event is read. When the subsequent risk trigger intensity increases relative to the risk trigger intensity before the preceding event, it is recorded as a risk enhancement direction; when it decreases, it is recorded as a risk weakening direction. The dominant change direction in historically confirmed risk evolution chains of the same type is statistically analyzed, and the number of valid event pairs in the current candidate risk node pair that are consistent with the dominant change direction is counted. This number is divided by the total number of valid event pairs in the candidate risk node pair to obtain the consistency of risk state changes.

[0105] The degree of business scenario association is determined based on the consistency of business scenario keys. The number of valid event pairs among candidate risk node pairs that share the same business scenario key, or whose material category, procurement method, and procurement project identifier all form a direct business association, is counted. This number is then divided by the total number of valid event pairs among candidate risk node pairs to obtain the degree of business scenario association.

[0106] The thresholds for the frequency of subsequent risk nodes, the consistency threshold for changes in risk status, and the correlation threshold for business scenarios are determined based on historically confirmed causal and non-causal risk relationship samples. Specifically, for each indicator, historical causal and non-causal relationship samples are arranged in ascending order of their corresponding indicator values. Different indicator values ​​are then used as candidate thresholds. The correct retention rate of causal relationships and the incorrect retention rate of non-causal relationships are calculated separately. The indicator value that best distinguishes between the two types of samples is selected as the corresponding threshold. For example, in this embodiment, the frequency of subsequent risk nodes is preferably between 0.60 and 0.80, the consistency threshold for changes in risk status is preferably between 0.70 and 0.90, and the correlation threshold for business scenarios is preferably between 0.60 and 0.85.

[0107] When the frequency of subsequent risk nodes, the consistency of risk state changes, and the degree of correlation with business scenarios of a candidate risk node pair all reach the corresponding thresholds, the corresponding supplier risk relationship edge is marked as a causal candidate edge. If any indicator fails to reach the corresponding threshold, the corresponding supplier risk relationship edge is marked as a non-causal edge. Therefore, in this embodiment, a higher frequency of occurrence specifically refers to the frequency of subsequent risk nodes reaching the aforementioned frequency threshold.

[0108] Causal candidate edges connecting the same pair of risk nodes and of the same relationship type are merged. The weights of the merged edges are determined based on three causal indicators. The frequency of subsequent risk node occurrences, consistency of risk state changes, and relevance to business scenarios are normalized to 0 to 1, and then the weights of the indicators are determined based on their ability to distinguish between historical causal and non-causal relationships. The specific weight determination method is as follows: calculate the difference between the average value of the three indicators in the historical causal relationship samples and the average value in the non-causal relationship samples, divide each difference by the sum of the three positive differences to obtain the corresponding indicator weight. Multiply the three normalized indicators by their corresponding indicator weights and sum them to obtain the weights of the synthesized causal relationship edges.

[0109] Remove non-causal edges. Aggregate event nodes that are under the same supplier, the same risk time slice, and the same business scenario key, and that have the same risk event type. The aggregated nodes are used as aggregated risk event nodes, and the synthetic causal relationship edges between each event node and the supplier node and scenario node before aggregation are maintained to form a supplier risk causal graph.

[0110] Step S105: Based on the supplier risk causal graph, construct the risk feature time series corresponding to each supplier node to obtain the supplier graph signal set.

[0111] The specific steps of step S105 are as follows:

[0112] Step S1051: Based on the supplier risk causal graph, statistically analyze the category and intensity distribution of each event node adjacent to each supplier node in each risk time slice to obtain the supplier risk neighborhood sequence.

[0113] In this embodiment, based on the supplier risk cause-effect graph, each supplier node in the graph is traversed. According to the risk time slice identifier of each risk event determined in the supplier risk event set, the risk time slice of each event node connected to each supplier node is determined. Event nodes that are adjacent to the same supplier node in the same risk time slice are grouped into the same supplier risk neighborhood unit, thus obtaining the supplier risk neighborhood unit set corresponding to each supplier and each risk time slice.

[0114] Based on the risk event types and risk triggering intensities of the event nodes contained in each supplier's risk neighborhood unit, the event nodes are classified, statistically analyzed, and their intensities are aggregated. The occurrence count, frequency, and combined triggering intensity of each risk event type in the corresponding supplier's risk neighborhood unit are calculated. The statistically obtained category distribution features and intensity distribution features are combined into the supplier's risk neighborhood feature vector for that risk time slice.

[0115] Obtain the time sequence of the risk time slices. Based on the time sequence of the risk time slices, arrange the risk neighborhood feature vectors of the same supplier in each risk time slice in sequence to form the supplier risk neighborhood sequence corresponding to that supplier. Then, summarize them to obtain the supplier risk neighborhood sequence set.

[0116] Step S1052: Based on the supplier risk neighborhood sequence, aggregate the category and intensity distribution within each risk time slice into a multidimensional risk feature vector, and perform time-series alignment to obtain the risk feature time series corresponding to each supplier.

[0117] In this embodiment, based on the supplier risk neighborhood sequence, the risk neighborhood feature vectors of each supplier at each risk time slice are analyzed. According to the risk event category coding rules, the occurrence frequency, occurrence intensity, and trigger intensity composite values ​​of various risk events in the risk neighborhood feature vector are normalized and rearranged in dimensions. The normalized statistics of various risk events are then concatenated according to a preset category order to construct a multi-dimensional risk feature vector. The risk event category coding rules are determined based on the risk event node types in the supplier risk causal graph. Specifically, the risk events are classified according to the business source and risk impact type corresponding to the risk event node, and a unique code identifier is assigned to each type of risk event. For example, qualification anomaly events, status anomaly events, transaction anomaly events, procurement anomaly events, and external regulatory anomaly events are divided into different risk event categories, and the category coding order is determined according to the chronological relationship in the business occurrence chain. The preset category order is determined based on the correlation between risk events and the evolution of supplier risks. Specifically, the method is to statistically analyze the sequential occurrence of different risk event categories in the historical risk event chain, arrange risk event categories with higher frequency of occurrence and continuous evolutionary relationship in adjacent positions, and form a risk event category splicing order.

[0118] Based on a preset global time axis for risk time slices, the risk time slice identifiers corresponding to each supplier are mapped to a unified time axis position. Suppliers lacking risk neighborhood feature vectors in some risk time slices are filled with zero-value multidimensional risk feature vectors, aligning the multidimensional risk feature vectors of each supplier in each risk time slice on the time axis. The global time axis for risk time slices is determined based on the time span of the supplier's business data and the cycle of risk event changes. Specifically, the determination method is as follows: statistically analyze the occurrence time range of each business event in the supplier's business data, determine the start and end times covering all supplier business events, and divide the time range into consecutively arranged risk time slices according to the aforementioned risk time granularity. The time identifiers corresponding to each risk time slice are arranged in chronological order to form the global time axis for risk time slices. For example, when the supplier's business data covers multiple consecutive months, the time slices can be divided according to the risk time granularity at the daily or weekly level, and the risk feature vectors of each supplier in the corresponding time slice are mapped to a unified time axis position.

[0119] Based on the temporal order of risk time slices, the multidimensional risk feature vectors of the same supplier in each risk time slice are arranged sequentially to obtain the risk feature time series corresponding to that supplier. These are then summarized to form a set of risk feature time series corresponding to each supplier.

[0120] Step S1053: Map the risk feature time series to the supplier risk monitoring signal series corresponding to each supplier node, and perform graph domain smoothing and anomaly retention processing to obtain the supplier graph signal set.

[0121] In this embodiment, based on the risk feature time series, the risk feature time series of each supplier is associated with the corresponding supplier node in the supplier risk causal graph according to the supplier identifier. The multidimensional risk feature vector in each risk time slice is used as the risk monitoring signal value of the corresponding supplier node in that risk time slice, thus obtaining the supplier risk monitoring signal sequence.

[0122] Based on the supplier nodes directly connected to the current supplier node in the supplier risk causal graph, a set of adjacent supplier nodes is obtained. The edge weights corresponding to each causal relationship edge are read, and the risk monitoring signal values ​​of adjacent supplier nodes in the same risk time slice are multiplied by the corresponding edge weights. All product results are summed and then divided by the sum of the corresponding edge weights to obtain the neighborhood weighted average of the current supplier node.

[0123] The graph domain smoothing ratio is determined based on the effect of suppressing local signal fluctuations under historical normal risk conditions and the effect of preserving abnormal amplitudes under historical abnormal risk conditions. Specifically, the method is as follows: multiple candidate smoothing ratios are set between 0 and 0.5, and graph domain smoothing is performed on historical normal samples and confirmed abnormal samples respectively; for normal samples, the degree of decrease in signal differences between adjacent supplier nodes before and after smoothing is calculated; for abnormal samples, the degree of preservation of the abnormal peak value after smoothing relative to the original abnormal peak value is calculated; the degree of decrease in local differences of normal samples and the degree of preservation of abnormal peak values ​​are normalized and then multiplied, and the candidate ratio with the largest product result is selected as the smoothing ratio. Limiting the candidate range to 0 to 0.5 is to ensure that the influence of the current supplier node's own signal in the fusion result is not lower than that of the neighboring information. For example, in this embodiment, the smoothing ratio is preferably between 0.15 and 0.35.

[0124] The neighborhood weighted average value is multiplied by the smoothing ratio, and the original risk monitoring signal value of the current supplier node is multiplied by the ratio after subtracting the smoothing ratio from 1. The two results are then added together to obtain the smoothed risk monitoring signal value in the graph domain.

[0125] The difference between the original risk monitoring signal value of the current supplier node and the risk monitoring signal value after graph domain smoothing is calculated, and the absolute value of the difference is taken as the abnormal deviation value.

[0126] The anomaly detection threshold is determined based on the empirical distribution of anomalous deviation values ​​under historical normal risk conditions. Specifically, the threshold is determined by arranging the anomalous deviation values ​​corresponding to historical normal samples in ascending order, and selecting the highest quantile value that covers the main fluctuations in normal deviations. For example, in this embodiment, the quantile ratio used to determine the anomaly detection threshold is preferably 95% to 99%.

[0127] When the abnormal deviation value exceeds the anomaly detection threshold, the difference between the original risk monitoring signal value and the graph smoothing result is taken as the abnormal component and superimposed back onto the graph smoothed risk monitoring signal value; when the abnormal deviation value is not greater than the anomaly detection threshold, the graph smoothing result remains unchanged. The processing results of each supplier node are summarized to obtain the supplier graph signal set.

[0128] Step S2: Based on the supplier graph signal set, construct the supplier risk monitoring evolution equation to obtain the supplier risk scenario field.

[0129] The specific steps of step S2 are as follows:

[0130] Step S201: Based on the supplier graph signal set, perform graph frequency domain multi-scale decomposition to extract long-term risk trend features and short-term abnormal risk features, and obtain a multi-dimensional risk monitoring state vector sequence.

[0131] The specific steps of step S201 are as follows:

[0132] Step S2011: Based on the supplier graph signal set, perform unified preprocessing and time alignment on the risk monitoring signal sequences of each supplier to obtain the supplier risk monitoring signal matrix.

[0133] In this embodiment, based on the supplier graph signal set, the risk monitoring signal sequences of each supplier in the supplier graph signal set are analyzed. According to the sampling time identifier and sampling value of each supplier risk monitoring signal sequence on the risk time slice, abnormal sampling points, missing sampling points and uneven sampling time in each supplier risk monitoring signal sequence are identified.

[0134] Based on preset data preprocessing rules, abnormal sampling points are removed, missing sampling points are filled by interpolation, and the sampled values ​​are normalized. The data preprocessing rules include abnormal sampling point determination rules, missing sampling point filling rules, and amplitude normalization rules. The abnormal sampling point determination rules are set based on the fluctuation range of sampled values ​​in historical normal supplier risk monitoring signal sequences. Specifically, the method is to statistically analyze the average and standard deviation of sampled values ​​for each risk time slice under normal conditions, and determine sampling points that deviate from the average by more than a preset multiple of the standard deviation as abnormal sampling points. The missing sampling point filling rules are determined based on the time interval between adjacent valid sampling points. Specifically, when the time interval between adjacent valid sampling points does not exceed a preset continuous missing range, the sampled value at the missing position is calculated using the linear change relationship of adjacent valid sampling points; when it exceeds the preset continuous missing range, the historical trend value of the adjacent risk time slice is used for filling. The amplitude normalization rules are set based on the value range of each risk characteristic component. Specifically, the method is to statistically analyze the maximum and minimum values ​​of each risk characteristic component in historical supplier risk monitoring signals, and convert each sampled value to a uniform numerical range according to the proportional relationship between the maximum and minimum values.

[0135] Based on the global time axis of the risk time slice, the sampling time of each supplier's risk monitoring signal sequence after preprocessing is converted into a unified time format and mapped to the time axis to obtain the time and amplitude of each adjacent sampling point. According to the time and amplitude of the adjacent sampling points, the sampling points that are not at the sampling position of the global time axis are resampled and interpolated so that each supplier's risk monitoring signal sequence has a consistent sampling position on the global time axis of the risk time slice.

[0136] Based on the time sequence of the global time axis of the risk time slice and the supplier identifier, the sampled values ​​of the risk monitoring signal sequences of each supplier in the same risk time slice are arranged in the column dimension, and each risk time slice is arranged in the row dimension, to construct a supplier risk monitoring signal matrix with the risk time slice as the row index and the supplier and its risk monitoring signal components as the column index.

[0137] Step S2012: Based on the supplier risk causal graph, construct the supplier risk adjacency operator and the graph Laplacian operator, and map the supplier risk monitoring signal matrix to the graph frequency domain to obtain the supplier risk graph spectrum.

[0138] In this embodiment, based on the supplier risk cause-effect graph, each supplier node in the supplier risk cause-effect graph is numbered to obtain a supplier node index set.

[0139] Based on the causal relationship edges between supplier nodes in the supplier risk causal graph and the preset weight allocation rules, the edge weights between supplier node pairs with direct causal relationships are determined. According to the arrangement of the supplier node index set in rows and columns, each supplier node is used as the row index node and column index node of the adjacency matrix. The edge weights corresponding to the causal relationship edges between any two supplier nodes are filled into the matrix elements at the intersection of the corresponding row index node and column index node. For supplier node pairs without direct causal relationship edges, the corresponding matrix elements are set to zero, resulting in the supplier risk adjacency matrix. The supplier risk adjacency matrix is ​​used as a matrix operator to represent the causal connection relationship between supplier nodes. This matrix operator is used to perform data transformation processing on the supplier risk monitoring signal matrix under node relationship constraints, serving as the supplier risk adjacency operator.

[0140] Based on the supplier risk adjacency matrix, the elements in each row of the supplier risk adjacency matrix are summed along the arrangement direction of the supplier node index set to obtain the node connection strength corresponding to each supplier node. The node connection strengths are then filled into the diagonal positions of the matrix to construct the supplier risk degree matrix. Based on the supplier risk degree matrix and the supplier risk adjacency matrix, a matrix difference operation is performed between the supplier risk degree matrix and the supplier risk adjacency matrix to obtain the supplier risk graph Laplace matrix. The supplier risk graph Laplace matrix is ​​used as the graph structure operator describing the differences in the node connection structure in the supplier risk causal graph, i.e., the supplier risk graph Laplace operator. Based on the supplier risk graph Laplace matrix, the matrix is ​​eigenvalued to solve for the eigenvalues ​​and eigenvectors that satisfy the characteristic relationship of the graph Laplace matrix. The eigenvectors are arranged in ascending order of their corresponding eigenvalues ​​and combined sequentially along the column direction to obtain the graph frequency basis matrix composed of the graph frequency basis vectors.

[0141] Based on the supplier risk monitoring signal matrix, the supplier node positions corresponding to each column in the supplier risk monitoring signal matrix are determined according to the supplier node index set. The risk monitoring signal sample values ​​corresponding to each supplier node are arranged according to the supplier node dimension. Based on the graph frequency basis vectors in each column of the graph frequency basis matrix, the correlation degree between each supplier node signal vector and each graph frequency basis vector in the supplier risk monitoring signal matrix is ​​calculated in turn. The correlation degree is used as the projection coefficient on the corresponding graph frequency basis vector. According to the arrangement order of the graph frequency basis vectors in the graph frequency basis matrix, the projection coefficients are arranged to obtain the supplier risk graph spectrum representing the response intensity of the supplier risk monitoring signal on different graph frequency components.

[0142] Step S2013: Divide the spectrum of the supplier risk map into frequency bands and reconstruct the signal components to obtain low-frequency risk components and high-frequency risk components.

[0143] In this embodiment, based on the eigenvalue arrangement order corresponding to the supplier risk graph Laplacian matrix, the graph frequency basis vectors are arranged in ascending order of their corresponding eigenvalues. The smaller the eigenvalue, the lower the graph frequency basis vector corresponds to the lower-order graph frequency with more gradual changes between supplier nodes, and the larger the eigenvalue, the higher-order graph frequency with more obvious local changes between supplier nodes.

[0144] For each graph frequency, the corresponding graph spectral coefficient is squared, and the squared values ​​for all risk time slices are summed to obtain the graph spectral energy for that frequency. The graph spectral energies of all graph frequencies are summed to obtain the total graph spectral energy.

[0145] The graph spectral energy is accumulated sequentially according to the graph frequency in ascending order of eigenvalues, and the accumulated graph spectral energy is divided by the total graph spectral energy to obtain the cumulative energy percentage.

[0146] The low-frequency cumulative energy threshold is determined based on the trend reconstruction effect and anomaly retention effect in the historical supplier risk map signal. Specifically, the determination method is as follows: multiple candidate cumulative energy ratios are set, and low-frequency reconstruction and high-frequency reconstruction are performed on each candidate ratio; the error between the low-frequency reconstruction result and the historical long-term trend is calculated, and the retention rate of historically confirmed local anomalies by the high-frequency reconstruction result is also calculated; among the candidate ratios that can prevent local anomalies from being significantly transferred to the low-frequency component, the cumulative energy ratio with the smallest low-frequency trend reconstruction error is selected as the low-frequency cumulative energy threshold. For example, in this embodiment, the low-frequency cumulative energy threshold is preferably between 0.85 and 0.95.

[0147] Starting from the lowest-order graph frequency, the graph spectral energy is accumulated sequentially. The graph frequency order at which the accumulated energy percentage first reaches the low-frequency accumulated energy threshold is determined as the low-frequency cutoff order. Graph frequencies below or equal to the low-frequency cutoff order are classified as low-frequency graph bands, and graph frequencies above the low-frequency cutoff order are classified as high-frequency graph bands.

[0148] The low-frequency risk component is obtained by retaining the graph spectral coefficients in the low-frequency graph band, setting the high-frequency graph spectral coefficients to zero, and performing inverse graph frequency domain mapping based on the graph frequency basis matrix. The high-frequency risk component is obtained by retaining the graph spectral coefficients in the high-frequency graph band, setting the low-frequency graph spectral coefficients to zero, and performing inverse graph frequency domain mapping.

[0149] like Figure 2 As shown, step S2014: Based on the low-frequency risk component and the high-frequency risk component, extract the long-term risk trend features and the short-term abnormal risk features, and perform feature aggregation within each risk time slice to obtain a multi-dimensional risk monitoring state vector sequence.

[0150] The specific steps of step S2014 are as follows:

[0151] Step S20141: Based on the low-frequency risk component and the high-frequency risk component, organize the low-frequency risk component value and the high-frequency risk component value of each supplier in each risk time slice to construct the supplier risk time slice feature tensor.

[0152] In this embodiment, based on low-frequency and high-frequency risk components, the low-frequency and high-frequency risk components are parsed. The association between supplier nodes, risk time slices, and risk component types is established according to the supplier node identifiers and risk time slice identifiers corresponding to the low-frequency and high-frequency risk components. The node index position of each supplier node in the supplier risk causal graph is determined based on the supplier node index set. The arrangement position of each risk time slice in a unified time axis is determined based on the risk time slice index set. Based on the component type identifiers of the low-frequency and high-frequency risk components, the corresponding risk component values ​​are associated with the low-frequency component position and the high-frequency component position, respectively, to obtain the low-frequency risk component value and high-frequency risk component value corresponding to each supplier in each risk time slice.

[0153] Based on a pre-defined supplier index set and risk time slice index set, the low-frequency and high-frequency risk component values ​​of each supplier within each risk time slice are aligned and organized. The supplier index is used as the first dimension, the risk time slice index as the second dimension, and the types of low-frequency and high-frequency risk components as the third dimension. According to the combination relationship of the first, second, and third dimension indices, the corresponding risk component values ​​are filled into the corresponding data element positions to construct a supplier risk time slice feature tensor. The supplier index set is determined based on the set of supplier nodes in the supplier risk causal graph. Specifically, the supplier nodes are arranged according to their node numbers in the supplier risk causal graph to form the supplier index set. The risk time slice index set is determined based on the time slice arrangement order in the global time axis of the risk time slices. Specifically, the risk time slices are numbered according to their chronological order on a unified time axis to form the risk time slice index set.

[0154] Step S20142: Based on the supplier risk time slice feature tensor, perform segmented low-rank decomposition with risk time slice as the segment unit to obtain long-term risk trend features and residual sub-tensors.

[0155] In this embodiment, based on the supplier risk time slice feature tensor, continuous risk phase intervals are divided according to the risk time slice sequence. The phase length is determined based on the number of continuous time slices in which the historical supplier risk characteristics maintain the same direction of change. Specifically, the determination method is as follows: count the lengths of time slices in the historical supplier risk characteristics that maintain the same direction of change, take the median of the length distribution as the basic phase length, and convert it into the corresponding number of risk time slices according to the risk time granularity.

[0156] For any risk stage interval, extract the corresponding tensor block from the supplier risk time slice feature tensor, expand the supplier dimension and risk component type dimension into feature dimensions, and keep the risk time slice dimension as the time dimension to obtain the risk time slice feature matrix.

[0157] This embodiment employs truncated singular value decomposition (SVD) to perform low-rank decomposition on the risk time slice feature matrix. The specific decomposition process is as follows: SVD is performed on the risk time slice feature matrix to obtain a left singular vector matrix, a sequence of singular values, and a right singular vector matrix, which are then arranged in descending order of singular values.

[0158] Squaring each singular value individually and summing the results of squaring all singular values ​​gives the total singular value energy. Starting from the largest singular value, summing the results of squaring the singular values ​​sequentially and dividing the cumulative result by the total singular value energy gives the percentage of cumulative singular value energy.

[0159] The cumulative energy threshold corresponding to the rank is determined based on the reconstruction stability of historical normal risk trends and the ability of the remaining components to distinguish abnormal risks. Specifically, the determination method is as follows: candidate ranks are determined using different candidate cumulative energy ratios; the long-term risk trend matrix is ​​reconstructed separately; and the trend reconstruction error of normal risk samples and the remaining energy of confirmed abnormal risk samples are calculated. The candidate cumulative energy ratio with the lowest normal sample trend reconstruction error and the largest difference between the remaining energy of abnormal samples and the remaining energy of normal samples is selected as the rank cumulative energy threshold. For example, in this embodiment, the rank cumulative energy threshold is preferably between 0.90 and 0.95.

[0160] Starting with the largest singular value, singular values ​​are selected. The number of singular values ​​included when the cumulative energy percentage first reaches the rank cumulative energy threshold is determined as the low-rank decomposition rank. The left singular vector, singular value, and right singular vector corresponding to the rank are retained, and matrix reconstruction is performed to obtain the long-term risk trend matrix of the current risk stage interval.

[0161] The long-term risk trend matrix is ​​restored to a trend subtensor according to the expansion relationship of the original risk time slice feature matrix, and then spliced ​​together according to the time order of the risk stage intervals to obtain the long-term risk trend features.

[0162] Subtract the trend feature value of the corresponding position in the long-term risk trend feature from the original feature value of each position in the supplier risk time slice feature tensor to obtain the remaining feature values. Then, rearrange them according to the supplier dimension, risk time slice dimension, and risk component type dimension to obtain the remaining subtensor.

[0163] Step S20143: Based on the residual subtensor, construct the risk anomaly energy function, and minimize and reconstruct the risk anomaly energy function to obtain the short-term anomaly risk characteristics.

[0164] In this embodiment, based on the residual subtensor, the residual feature values ​​of each supplier at each risk time slice and each risk component type are used as anomalous observations, and anomalous estimation variables are established according to the same dimension as the residual subtensor.

[0165] The anomaly estimation variables are initially set as the residual feature values ​​at the corresponding locations. A first energy term is constructed based on the anomaly observations and the anomaly estimation variables. The specific calculation process is as follows: subtract the corresponding anomaly estimation variable from the anomaly observation at each location, square the resulting difference, and sum the squared results for all locations to obtain the first energy term, which is used to constrain the deviation between the anomaly reconstruction result and the original residual features.

[0166] The second energy term is constructed based on the adjacency relationship of risk time slices. The specific calculation process is as follows: for the same supplier and the same risk component type, the abnormal estimated variables corresponding to adjacent risk time slices are interpolated, and the absolute value of the difference is taken; the absolute differences corresponding to all adjacent risk time slices are summed to obtain the second energy term, which is used to suppress the frequent jumps of abnormal estimated variables on the time axis without basis, while retaining short-term abnormal segments with clear boundaries.

[0167] A third energy term is constructed based on the supplier risk causal graph. For supplier node pairs with direct causal relationships, the abnormal estimated variables corresponding to the two supplier nodes are interpolated at the same risk time slice and the same risk component type, and the difference is squared. The squared result is multiplied by the weight of the corresponding causal relationship edge, and the results corresponding to all causal relationship edges are accumulated to obtain the third energy term, which is used to constrain the abnormal synergy between supplier nodes with risk correlation.

[0168] Before constructing the risk anomaly energy function, the initial magnitudes of the first, second, and third energy terms are calculated using the initial values ​​of the anomaly estimation variables, and the corresponding initial magnitudes are used to scale the three energy terms so that the three energies are within a comparable range in the initial state.

[0169] The weight of the first energy term is set to 1 as a baseline weight. The weights of the second and third energy terms are determined based on historically confirmed abnormal risk samples. Specifically, the weights of the second and third energy terms are varied within the candidate weight range. Abnormal reconstruction is performed on historical normal and abnormal samples. The accuracy and recall of abnormal sample identification, as well as the false positive rate of normal samples, are calculated. The weight combination with the highest overall identification effect is selected. For example, in this embodiment, the weight of the second energy term is preferably between 0.10 and 0.40, and the weight of the third energy term is preferably between 0.05 and 0.30.

[0170] The normalized first, second, and third energy terms are multiplied by their corresponding weights, and the three results are summed to obtain the risk anomaly energy function.

[0171] This embodiment employs the alternating direction multiplier method to minimize the risk anomaly energy function. For the absolute difference between adjacent time slices in the second energy term, a time difference auxiliary variable is introduced, allowing the anomaly estimation variable and the time difference auxiliary variable to be updated separately. In each iteration, firstly, the time difference auxiliary variable is fixed, and the anomaly estimation variable is solved based on the first energy term, the third energy term, and the time difference constraint; then, the anomaly estimation variable is fixed, and the time difference auxiliary variable is updated with a soft threshold based on the second energy term; finally, the multiplier variable is updated based on the deviation between the time difference result of the anomaly estimation variable and the time difference auxiliary variable.

[0172] The penalty parameter of the alternating direction multiplier method is adjusted based on the balance between the original and dual residuals. When the original residual is consistently greater than the dual residual, the penalty parameter is increased; when the dual residual is consistently greater than the original residual, the penalty parameter is decreased; and when the two types of residuals are of the same order of magnitude, the penalty parameter remains unchanged. Therefore, this embodiment does not use a fixed gradient adjustment step size to avoid convergence instability caused by manually setting the gradient step size.

[0173] The iteration termination threshold is determined based on the relative change ratio between two adjacent rounds of historical normal risk samples when the energy function tends to stabilize. Specifically, the threshold is determined by performing a complete iteration on multiple sets of historical normal risk samples, extracting the relative energy change ratio after entering the stable phase, and taking the high quantile of the change ratio distribution as the iteration termination threshold. For example, in this embodiment, the iteration termination threshold is preferably between 0.0001 and 0.001. Iteration stops when the relative change ratio of the energy function of two consecutive rounds of risk anomalies is not greater than the iteration termination threshold, and both the original residual and the dual residual have entered a stable range.

[0174] The absolute values ​​of the converged anomaly estimates for each risk component type are summed and divided by the number of the corresponding risk component types to obtain the anomaly intensity component for each supplier in each risk time slice.

[0175] The anomaly determination threshold is determined based on the empirical distribution of the anomaly intensity components under historical normal risk conditions. Specifically, the anomaly intensity components of normal samples are arranged in ascending order, and the highest quantile value is taken as the anomaly determination threshold. For example, in this embodiment, the quantile ratio used to determine the anomaly determination threshold is preferably 95% to 99%.

[0176] The number of risk time slices in which the abnormality intensity component is continuously higher than the abnormality judgment threshold is counted. The number of consecutive risk time slices is used as the abnormality persistence component. The abnormality intensity component and the abnormality persistence component are combined to obtain the short-term abnormality risk characteristics.

[0177] Step S20144: Normalize and splice the long-term risk trend characteristics and short-term abnormal risk characteristics within each risk time slice to obtain a multi-dimensional risk monitoring state vector sequence.

[0178] In this embodiment, based on long-term risk trend characteristics, short-term abnormal risk characteristics, supplier index sets, and risk time slice index sets, the long-term risk trend characteristic components of each supplier in each risk time slice are organized. According to the risk source type and changing effect corresponding to each characteristic component in the long-term risk trend characteristics, the long-term risk trend characteristic components are arranged according to the risk evolution order to obtain the corresponding long-term risk trend characteristic vector. The risk evolution order is determined based on the risk change process reflected by each characteristic component in the long-term risk trend characteristics. Specifically, the method is as follows: the long-term risk trend characteristic components are arranged according to the order of risk status from basic change trend to comprehensive change trend, forming the long-term risk trend characteristic dimension order.

[0179] Based on the supplier index set and risk time slice index set, the short-term abnormal risk feature components of each supplier in each risk time slice are sorted out. According to the risk performance of the abnormality intensity component and the abnormality duration component in the short-term abnormal risk features, the abnormality intensity component is arranged before the abnormality duration component to obtain the corresponding short-term abnormal risk feature vector.

[0180] Based on the long-term risk trend feature vector, the short-term abnormal risk feature vector, and the normalization rule, the amplitude of the feature components in each feature vector is transformed. The normalization rule is determined according to the value range and scale of change of different risk feature components. The specific determination method is as follows: statistically analyze the maximum value, minimum value, and range of change of each risk feature component in historical supplier risk monitoring data, and transform the corresponding feature values ​​to a unified numerical range according to the value range of each feature component, so that risk feature components from different sources and with different dimensions are comparable, thus obtaining the normalized long-term risk trend feature vector and the normalized short-term abnormal risk feature vector.

[0181] Based on the normalized long-term risk trend feature vector and the normalized short-term abnormal risk feature vector, according to the formation process of risk features, the normalized long-term risk trend feature vector of the same supplier in the same risk time slice is arranged first, followed by the normalized short-term abnormal risk feature vector. The feature components in the two types of normalized feature vectors are sequentially connected to obtain a multi-dimensional risk monitoring state vector. Based on the supplier index set and the risk time slice index set, the multi-dimensional risk monitoring state vectors of each supplier in each risk time slice are arranged in chronological order to obtain a multi-dimensional risk monitoring state vector sequence. Step S202: Based on the multi-dimensional risk monitoring state vector sequence, a supplier risk monitoring evolution equation is constructed, and a risk-time slice-by-risk deduction is performed to obtain the risk monitoring state trajectory of each supplier in each risk time slice.

[0182] The specific steps of step S202 are as follows:

[0183] Step S2021: Perform risk evolution mode decomposition on the multidimensional risk monitoring state vector sequence and extract the supplier risk evolution fundamental mode.

[0184] In this embodiment, based on the multidimensional risk monitoring state vector sequence, the multidimensional risk monitoring state vectors of each supplier in each risk time slice are arranged in chronological order according to the supplier identifier to obtain the supplier risk monitoring state matrix.

[0185] This embodiment employs a dynamic mode decomposition algorithm for risk evolution mode decomposition. For any supplier risk monitoring state matrix, the state vectors of each risk time slice (excluding the last risk time slice) are sequentially used as preceding state columns to form a preceding state matrix; the state vectors of each risk time slice (excluding the first risk time slice) are sequentially used as following state columns to form a following state matrix, such that each column in the preceding state matrix and the same column in the following state matrix constitute a set of adjacent risk time slice state transition samples.

[0186] Perform singular value decomposition on the preceding state matrix to obtain a left singular vector matrix, a sequence of singular values, and a right singular vector matrix. Arrange the singular values ​​in descending order and calculate the cumulative proportion of the squared value of each singular value to the sum of the squares of all singular values.

[0187] The dynamic mode decomposition (DMD) truncation rank is determined based on the state transition reconstruction error. Specifically, it is determined by: obtaining candidate truncation ranks using different cumulative singular value energy ratios, reconstructing the subsequent state matrix, and calculating the mean squared error between the predicted state and the actual subsequent state; selecting the minimum cumulative energy ratio that allows the mean squared error to enter the stable interval as the DMD truncation threshold. For example, in this embodiment, the preferred DMD truncation threshold is 0.90 to 0.98.

[0188] Preserve the left singular vector, singular value, and right singular vector corresponding to the truncated rank. First, transpose the left singular vector matrix, then multiply it with the subsequent state matrix, then multiply it with the preserved right singular vector matrix, and finally multiply it with the inverse of the diagonal matrix formed by the preserved singular values ​​to obtain the low-dimensional state transition operator.

[0189] The low-dimensional state transition operator is subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors. The subsequent state matrix is ​​multiplied by the retained right singular vector matrix, then multiplied by the inverse of the retained singular value diagonal matrix, and finally multiplied by the eigenvectors of the low-dimensional state transition operator to obtain the candidate risk evolution modes.

[0190] Based on the multidimensional risk monitoring state vector of the first risk time slice, the initial mode coefficients of each candidate risk evolution mode are solved by least squares error. According to the eigenvalues ​​of the low-dimensional state transition operator, the state expansion of each candidate risk evolution mode is performed along the risk time slice to obtain the modal response of each candidate risk evolution mode in each risk time slice.

[0191] The modal responses of each candidate risk evolution mode across all risk time slices are squared and summed to obtain the modal contribution of the corresponding candidate risk evolution mode. The candidate risk evolution modes are then arranged in descending order of modal contribution, and the cumulative modal contribution ratio is calculated.

[0192] The cumulative contribution threshold of the evolutionary fundamental mode is determined based on the reconstruction error of the historical risk state sequence. Specifically, the method is as follows: The number of retained candidate risk evolutionary modes is increased sequentially, and the historical multidimensional risk monitoring state vector sequence is reconstructed using the retained modes. When the rate of decrease in state reconstruction error slows significantly after further increasing the number of modes, the cumulative mode contribution ratio corresponding to the previous number of modes is determined as the cumulative contribution threshold of the evolutionary fundamental mode. For example, in this embodiment, the cumulative contribution threshold of the evolutionary fundamental mode is preferably between 0.85 and 0.95.

[0193] Candidate risk evolution modes are selected in descending order of modal contribution until the cumulative modal contribution ratio reaches the cumulative contribution threshold of the evolutionary baseline mode. The selected candidate risk evolution modes are then used as the supplier risk evolution baseline modes.

[0194] Step S2022: Construct a low-dimensional risk evolution subspace based on the supplier risk evolution fundamental model, and project the multi-dimensional risk monitoring state vector to obtain a low-dimensional risk monitoring state vector sequence.

[0195] In this embodiment, based on the supplier risk evolution fundamental model set, each supplier risk evolution fundamental model is represented as a risk change direction vector in the multidimensional risk monitoring state vector space. The fundamental models are sorted according to their contribution level, with those contributing more prominently. The basis vectors corresponding to each fundamental model are then combined along the column direction according to this sorting result to construct a risk evolution basis vector matrix. The order of the basis vectors is determined based on the contribution level of each supplier risk evolution fundamental model to the multidimensional risk monitoring state changes. Specifically, the method is as follows: the contribution percentage of each fundamental model in step S2021 is calculated, and the corresponding basis vectors are arranged in descending order of contribution percentage to form the risk evolution basis vector matrix.

[0196] Based on the risk evolution basis vector matrix, the set of all vectors that each column basis vector can represent is determined as a low-dimensional risk evolution subspace, so that any risk state vector in this low-dimensional risk evolution subspace can be generated by multiple risk evolution basis vectors in different combination ratios.

[0197] Based on the multidimensional risk monitoring state vector sequence, supplier index set, and risk time slice index set, the multidimensional risk monitoring state vector of each supplier in each risk time slice is analyzed, and the multidimensional risk monitoring state vector of any supplier in any risk time slice is taken as the risk state vector to be projected.

[0198] Based on the risk state vector to be projected and the risk evolution basis vector matrix, a combination relationship is established between the risk state vector to be projected and the risk evolution basis vectors. The risk state vector to be projected is represented as a reconstructed vector obtained by weighting the risk evolution basis vectors according to their corresponding low-dimensional coordinate coefficients. According to the variation range of the low-dimensional coordinate coefficients, the combination ratios corresponding to each basis vector are adjusted so that the difference between the reconstructed vector and the risk state vector to be projected gradually decreases.

[0199] The specific solution process is as follows: Initialize the low-dimensional coordinate coefficients corresponding to each risk evolution basis vector; perform a weighted combination of each risk evolution basis vector based on the current low-dimensional coordinate coefficients to obtain the reconstructed risk state vector; calculate the corresponding dimensional difference between the reconstructed risk state vector and the original multidimensional risk monitoring state vector to obtain the reconstruction deviation in each feature dimension; summarize the reconstruction deviations in each feature dimension to obtain the reconstruction error corresponding to the current coordinate coefficients; adjust the low-dimensional coordinate coefficients according to the direction of the reconstruction error change so that the reconstructed risk state vector obtained in the next combination is closer to the original multidimensional risk monitoring state vector; repeat the above process until the reconstruction error reaches the preset convergence condition, and use the finally obtained low-dimensional coordinate coefficient vector as the low-dimensional risk monitoring state vector of the corresponding supplier in the corresponding risk time slice. The convergence condition is determined based on the error change amplitude in the historical risk state projection process. The specific determination method is: statistically analyze the error change ratio of normal risk state samples in the continuous coordinate update process, and use the change range corresponding to the error change stabilizing as the convergence condition.

[0200] Based on low-dimensional risk monitoring state vectors, the low-dimensional risk monitoring state vectors of the same supplier in different risk time slices are concatenated in chronological order according to the arrangement order of the supplier index set and the risk time slice index set, to obtain the low-dimensional risk monitoring state vector sequence corresponding to each supplier.

[0201] Step S2023: Cluster and segment the low-dimensional risk monitoring state vector sequence to identify the risk evolution state type.

[0202] In this embodiment, based on the low-dimensional risk monitoring state vector sequence, supplier index set, and risk time slice index set, the low-dimensional risk monitoring state vectors of the same supplier in different risk time slices are arranged in chronological order to obtain a low-dimensional risk monitoring state matrix with suppliers corresponding to rows and risk time slices corresponding to columns. The low-dimensional risk monitoring state vectors in each column are used as continuous risk evolution trajectory points on a unified time axis.

[0203] Based on the low-dimensional risk monitoring state matrix, feature distances are calculated for the low-dimensional risk monitoring state vectors corresponding to each risk time slice. The degree of state similarity is determined based on the spatial distance between state vectors in the low-dimensional risk evolution subspace. Combined with the temporal continuity between adjacent risk time slices, cluster analysis is performed on the low-dimensional risk monitoring state vectors. Specifically, the clustering process is as follows: the distance between any two low-dimensional risk monitoring state vectors corresponding to any two risk time slices is calculated; state vectors with smaller distances and falling within the range of adjacent risk time slices are identified as state points with similar risk evolution characteristics; based on the similarity relationship between state points, low-dimensional risk monitoring state vectors with similar risk evolution characteristics are grouped into the same risk evolution state cluster, resulting in an initial set of risk evolution state clusters covering each supplier's risk time slice.

[0204] Based on the initial set of risk evolution state clusters, and according to the preset segmentation criteria, low-dimensional risk monitoring state vectors belonging to the same risk evolution state cluster and continuously distributed on the risk time slice are merged on the time axis to form multiple time-continuous typical risk evolution state segments, and discrete risk evolution state identifiers are assigned to each typical risk evolution state segment.

[0205] Based on the time coverage of each typical risk evolution state segment, the risk evolution state identifier that each risk time slice falls into is taken as the risk evolution state type corresponding to that risk time slice, thus obtaining a sequence of risk evolution state types corresponding to each risk time slice.

[0206] Step S2024: Based on the risk evolution state type, determine the temporal evolution relationship between low-dimensional risk monitoring state vectors, and perform parameter fitting to obtain the supplier risk monitoring evolution equation.

[0207] In this embodiment, based on the low-dimensional risk monitoring state vector sequence and the risk evolution state type corresponding to each risk time slice, the low-dimensional risk monitoring state vectors corresponding to two adjacent risk time slices of the same supplier are paired. The low-dimensional risk monitoring state vector of the previous risk time slice is used as the state input, the low-dimensional risk monitoring state vector of the next risk time slice is used as the state output, and the risk evolution state type corresponding to the previous risk time slice is used as the state type identifier to obtain the supplier risk evolution sample set.

[0208] This embodiment constructs piecewise affine state transition functions for different risk evolution state types. The dimension of the low-dimensional risk evolution subspace is denoted as d. For any risk evolution state type, a d-row, d-column state transition matrix and a bias vector with d components are set. The low-dimensional risk monitoring state vector of the previous risk time slice is multiplied by the state transition matrix corresponding to the current risk evolution state type. Then, the resulting vector is added dimension-wise to the corresponding bias vector to obtain the predicted low-dimensional risk monitoring state vector for the next risk time slice.

[0209] Therefore, each risk evolution state type contains d squared state transition matrix parameters and d bias parameters. The parameters in the state transition matrix represent the degree to which each low-dimensional risk feature component propagates and couples from the current risk time slice to the next risk time slice, and the bias parameters represent the basic changes that do not depend on the current state value under the corresponding risk evolution state type.

[0210] During parameter initialization, the state transition matrix corresponding to each risk evolution state type is initialized to an identity matrix, and the bias vector is initialized to a zero vector, so that the initial model corresponds to the baseline state transition relationship where the states of adjacent risk time slices remain unchanged.

[0211] For any risk evolution state type, all state input vectors belonging to that state type are arranged in the order of samples to form an input matrix, and a constant component is added to the end of each input vector to uniformly solve the bias parameter; the corresponding state output vectors are arranged in the same order of samples to form an output matrix.

[0212] The state transition matrix and bias vector are solved using a least squares method with quadratic regularization constraints. The specific solution process is as follows: calculate the transpose of the input matrix; multiply the transpose of the input matrix with the input matrix; add regularization coefficients to the diagonal positions of the resulting matrix; then calculate the inverse of the resulting matrix; multiply this inverse matrix sequentially with the transpose of the input matrix and the output matrix to obtain a parameter matrix containing the state transition matrix parameters and bias parameters.

[0213] The regularization coefficient is determined using a time-series verification method. Specifically, the method involves maintaining the time sequence of historical risk evolution samples, using samples from earlier time periods for parameter fitting, and using samples from subsequent time periods for verification. Different candidate regularization coefficients are used to solve for the parameters, and the mean squared error of the one-step state prediction for the verification samples is calculated. The regularization coefficient with the smallest mean squared error is selected. For example, in this embodiment, when the low-dimensional risk monitoring state vector has been normalized, the regularization coefficient is preferably between 0.0001 and 0.01.

[0214] The parameters of the state transition matrix and bias vector corresponding to each risk evolution state type are fitted separately, using the risk evolution state type as the selection condition for the state transition function. After inputting the low-dimensional risk monitoring state vector of the previous risk time slice and the current risk evolution state type, the corresponding state transition matrix and bias vector are first selected according to the risk evolution state type, and then matrix multiplication and dimension-by-dimensional addition are performed to obtain the predicted low-dimensional risk monitoring state vector for the next risk time slice. The piecewise affine state transition functions corresponding to each risk evolution state type together constitute the supplier risk monitoring evolution equation.

[0215] Step S2025: Based on the supplier risk monitoring evolution equation, perform risk-time slice extrapolation on the low-dimensional risk monitoring state vector sequence to obtain the supplier risk monitoring state trajectory.

[0216] In this embodiment, based on the supplier index set and the risk time slice index set, the low-dimensional risk monitoring state vector sequence corresponding to each supplier is organized according to the time order of the risk time slice, and the low-dimensional risk monitoring state vector of each supplier in the initial risk time slice is extracted as the initial state set of the supplier risk monitoring evolution equation.

[0217] Based on the supplier risk monitoring evolution equation, the initial state set, and the risk time slice index set, the low-dimensional risk monitoring state vector of each supplier in adjacent risk time slices is extrapolated on a risk-time slice basis. In the current risk time slice, the low-dimensional risk monitoring state vector of the previous risk time slice and the risk evolution state type corresponding to the current risk time slice are input into the supplier risk monitoring evolution equation to calculate the predicted low-dimensional risk monitoring state vector of the current risk time slice.

[0218] Based on the predicted low-dimensional risk monitoring state vector within each risk time slice, range and continuity checks are performed on each low-dimensional feature component in the prediction result. When any low-dimensional feature component exceeds the effective change range of historical risk evolution states, or the change amplitude between adjacent risk time slices exceeds the change range during normal risk evolution, the corresponding predicted low-dimensional risk monitoring state vector is determined as an abnormal prediction state. According to the effective state change trend within adjacent risk time slices, the abnormal feature components in the abnormal prediction state are constrained and corrected, adjusting them to a numerical range that conforms to the continuity of risk evolution, resulting in the corrected predicted low-dimensional risk monitoring state vector. Specifically, the correction method involves statistically analyzing the change range of each low-dimensional feature component in the historical supplier risk monitoring state trajectory, determining the normal change interval of each low-dimensional feature component within adjacent risk time slices, and truncating or smoothing the excess portion according to the change trend of adjacent risk time slices when the prediction result exceeds the corresponding change interval. Based on the corrected predicted low-dimensional risk monitoring state vector, it is written to the corresponding supplier and risk time slice positions, updating the low-dimensional risk monitoring state vector sequence corresponding to each supplier.

[0219] Based on the updated low-dimensional risk monitoring state vector sequence, the low-dimensional risk monitoring state vectors of the same supplier in each risk time slice are sequentially concatenated in each supplier dimension according to the order of the risk time slice index set, so as to obtain the risk monitoring state trajectory of each supplier in each risk time slice.

[0220] Step S203: Based on the risk monitoring status trajectory, extract the risk monitoring status trajectory values ​​of each supplier in each risk time slice, and splice the risk monitoring status trajectory values ​​in the same time slice to obtain a risk monitoring status field arranged according to the risk time slice.

[0221] In this embodiment, based on the risk monitoring status trajectory of each supplier in each risk time slice, the supplier index set, and the risk time slice index set, the low-dimensional risk monitoring status vector of each supplier in each risk time slice is extracted one by one. Each low-dimensional risk monitoring status vector is used as the risk monitoring status trajectory value of the corresponding supplier and the corresponding risk time slice position, and a set of risk monitoring status trajectory values ​​covering all suppliers and all risk time slices is constructed.

[0222] Based on the risk monitoring status trajectory value set and the risk time slice index set, within the same risk time slice, the risk monitoring status trajectory values ​​of each supplier are concatenated according to the order of the supplier index set to form a risk monitoring status arrangement vector expanded in the supplier dimension. Then, according to the order of the risk time slice index set, the risk monitoring status arrangement vectors corresponding to each risk time slice are arranged sequentially to construct a risk monitoring status field that is orderly distributed in the supplier dimension and the risk time slice dimension.

[0223] Step S204: Based on the risk monitoring state field, interpolate and smooth the risk monitoring state trajectory values ​​of adjacent risk time slices in the risk monitoring state field to obtain the supplier risk scenario field.

[0224] In this embodiment, based on the risk monitoring state field and the risk time slice index set, the risk monitoring state trajectory values ​​are arranged according to the chronological order of the risk time slices at each supplier dimension. The time intervals corresponding to adjacent risk time slices are determined, and the risk monitoring state trajectory values ​​at both ends of adjacent risk time slices are extracted as endpoint state values ​​of the local time interpolation interval. Based on the time interval between adjacent risk time slices and the changes in risk monitoring state trajectory values, the time positions requiring refinement are determined, and continuously arranged intermediate time points are generated between adjacent risk time slices according to a preset time interpolation interval. Based on the time position relationship between the risk monitoring state trajectory values ​​at both ends of adjacent risk time slices and the intermediate time points, interpolation calculations are performed on the risk monitoring state trajectory values ​​at each intermediate time point, ensuring a continuous transition of state changes between adjacent risk time slices in chronological order, resulting in a set of risk monitoring state trajectory values ​​containing the original risk time slices and intermediate time points. The time interpolation interval is determined based on the rate of change of the risk monitoring status trajectory value and the time span between risk time slices. Specifically, the determination method is as follows: statistically analyze the state change amplitude between adjacent time slices in the historical risk monitoring status trajectory. When the change amplitude between adjacent time slices is large, the interpolation interval is reduced to increase the number of sampling points in the state change process; when the state change amplitude is small, the interpolation interval is increased to reduce redundant interpolation points.

[0225] Based on the risk monitoring status trajectory value set, the original risk time slices and intermediate time points are arranged in chronological order across each supplier dimension. The risk monitoring status trajectory values ​​at each time point are smoothed to obtain a risk monitoring status distribution result that changes continuously across the supplier and time dimensions. This risk monitoring status distribution result is then used as the supplier risk scenario.

[0226] exist Figure 2 In the left rectangular diagram, the yellow outer rectangle represents the projection of the supplier risk time slice feature tensor, organized in terms of supplier dimension, risk time slice dimension, and risk component type dimension, onto a two-dimensional plane. The horizontal direction represents the risk time slice direction, and the vertical direction represents the expansion direction of supplier and risk component type.

[0227] Figure 2 In the rectangular diagram on the left, the light yellow vertical and horizontal grid lines represent the discretization of the risk time slice and risk components into 18 grid cells, which correspond to the discrete sampling positions of the low-frequency risk component values ​​and high-frequency risk component values ​​of each supplier in each risk time slice.

[0228] Figure 2In the left rectangle, the three clusters of orange crosses, blue dots, and green triangles represent the feature value clusters of different suppliers or different risk components in the supplier risk time slice feature tensor over six consecutive risk time slices. A group of points of the same color represents the set of low-frequency and high-frequency risk component values ​​of a certain supplier within a consecutive risk time slice. The differences in shape and color between the three clusters represent the differences in the distribution of different suppliers or different risk patterns over the time slices.

[0229] Figure 2 The left rectangle in the middle points to Figure 2 The black arrow in the middle circle represents the process of taking the supplier risk time slice feature tensor as input, using rectangular piecewise low-rank decomposition and risk anomaly energy function minimization to reconstruct the original time slice feature tensor onto the feature space of long-term risk trend features and short-term abnormal risk features.

[0230] Figure 2 In the central circular diagram, the outer thick orange circle represents the overall feature activity range formed in the feature space after segmenting the risk time slice feature matrix of each risk stage interval into a low-rank decomposition, which is the joint feature space of long-term risk trend features and short-term abnormal risk features.

[0231] Figure 2 In the middle circular diagram, the inner blue dashed circle represents the baseline range composed of long-term risk trend characteristics, that is, the amplitude range of the long-term risk trend matrix in this space after the low-rank decomposition and reconstruction of each risk stage interval.

[0232] Figure 2 In the diagram, four black arrows emanating from the center of the central circular diagram represent four main directions of risk change after short-term risk anomalies are superimposed on the long-term risk trend. The arrows point to different characteristic change patterns, which correspond to the directions of the base matrices of the main change patterns at each stage. The length of the arrow represents the overall intensity in that direction; the longer the arrow, the greater the contribution of that pattern to the supplier risk time slice feature tensor. The arrows pass through the dashed circle and extend to the outer ring, indicating that short-term abnormal risk features are superimposed on the long-term risk trend.

[0233] Figure 2 In the diagram, the black arrow pointing from the central circular diagram to the right-hand colored closed loop represents the normalization and splicing of the two types of features within each risk time slice after obtaining the long-term risk trend characteristics and short-term abnormal risk characteristics, thus forming a continuous multi-dimensional risk monitoring state vector sequence.

[0234] Figure 2The colored closed loop on the right represents the sequence of multi-dimensional risk monitoring state vectors connected end to end along the time axis according to the time order of the risk time slices, forming a loop outline. That is, the overall outline of the evolution trajectory of each supplier's risk monitoring status in the feature space within a complete observation period. The entire colored closed arc represents a complete risk monitoring state vector sequence. The arc segments of different colors indicate that the state patterns after splicing long-term risk trend characteristics and short-term abnormal risk characteristics are different in different time intervals, reflecting the stage differences in the normalization splicing of long-term risk trend characteristics and short-term abnormal risk characteristics within each risk time slice.

[0235] Step S3: Based on the supplier risk scenario, construct a risk path hierarchy map of the integrated intelligent agent to monitor supplier risks.

[0236] The specific steps of step S3 are as follows:

[0237] Step S301: Parse the supplier risk monitoring rules, construct a set of integrated intelligent agents, and generate a set of corresponding risk scenario feature path templates.

[0238] In this embodiment, the supplier risk monitoring rules used in the historical risk monitoring process are obtained, and the applicable risk types, risk event categories, risk time ranges, risk triggering conditions, risk change directions and risk level output conditions in each supplier risk monitoring rule are parsed to obtain a set of structured representations of supplier risk monitoring rules.

[0239] Based on the structured representation set of supplier risk monitoring rules, these rules are categorized according to risk type. Rules handling the same risk type are grouped into the same rule group, and each rule group is assigned to a fusion intelligent agent, resulting in a set of fusion intelligent agents. These fusion intelligent agents are used to handle at least three types of risks: qualification change risks, business behavior risks, and external regulatory risks.

[0240] Based on the supplier risk monitoring rules corresponding to each integrated intelligent agent, according to the risk time granularity determined in step S101, the risk time range in the rules is converted into the number of risk time slices, and the risk event category, risk intensity condition, risk change direction and risk evolution state type are arranged according to the triggering order of risk events in the rules, so as to obtain the rule feature sequence corresponding to each supplier risk monitoring rule.

[0241] Rule-based feature sequences belonging to the same fused intelligent agent and with the same order of risk event categories are aggregated. The risk event category, risk intensity condition, risk change direction, and risk evolution state type corresponding to each risk time slice are used as pathway template units. These pathway template units are connected according to the order of the risk time slices to form risk scenario feature pathway templates. The risk scenario feature pathway templates corresponding to each fused intelligent agent are then summarized to obtain a risk scenario feature pathway template set.

[0242] Each risk scenario feature path template retains the corresponding fusion agent identifier, risk event category sequence, number of risk time slices, and risk level output conditions, which are used for subsequent matching of risk scenario segments in the supplier risk scenario field.

[0243] Step S302: Extract risk scenario fragments based on the supplier risk scenario field and match them with the risk scenario feature path template set to obtain the supplier risk path set of the fusion agent.

[0244] In this embodiment, based on the supplier risk scenario, the risk monitoring status trajectory value and risk evolution status type of each supplier in continuous risk time slices are extracted according to the supplier identifier and arranged in chronological order of risk time slices.

[0245] For any risk scenario feature path template, the length of the sliding window is determined according to the number of risk time slices contained in the template. The sliding window is moved step by step in the supplier risk scenario field according to the moving interval of one risk time slice. The continuous risk time slices covered by each sliding window, as well as the corresponding risk monitoring status trajectory value and risk evolution status type, are used as candidate risk scenario segments.

[0246] Based on the risk monitoring status trajectory values ​​of adjacent risk time slices in the candidate risk scenario segments, the change direction of the status trajectory value of the later risk time slice relative to the status trajectory value of the previous risk time slice is compared one by one to obtain the risk status change direction sequence; the risk evolution status types corresponding to each risk time slice are arranged in chronological order to obtain the risk evolution status type sequence.

[0247] The sequence of risk state change directions is compared position by position with the template change directions in the risk scenario feature path template. The number of positions with the same direction is divided by the total number of template change positions to obtain the change direction matching degree. The sequence of risk evolution state types is compared position by position with the risk evolution state types in the template. The number of positions with the same type is divided by the total number of template risk time slices to obtain the state type matching degree.

[0248] The weights for the change direction matching degree and the state type matching degree are determined based on historically confirmed template matching samples. Specifically, for historically confirmed matching segments and non-matching segments, matching is performed using only the change direction matching degree and only the state type matching degree, respectively, calculating the risk segment identification F1 value corresponding to each of the two individual indicators. The F1 value corresponding to the change direction matching degree is divided by the sum of the two F1 values ​​to obtain the change direction matching degree weight; the F1 value corresponding to the state type matching degree is divided by the sum of the two F1 values ​​to obtain the state type matching degree weight, ensuring the sum of the two weights is 1. For example, in this embodiment, when the two features have similar distinguishing abilities, the two matching degree weights are preferably between 0.40 and 0.60.

[0249] The template matching degree is obtained by multiplying the matching degree of the direction of change with the corresponding weight, multiplying the matching degree of the state type with the corresponding weight, and adding the two product results.

[0250] The template matching threshold is determined based on the template matching degree of historically confirmed matching segments and non-matching segments. Specifically, the determination method is as follows: the template matching degrees appearing in historical samples are sequentially used as candidate thresholds; for each candidate threshold, the accuracy and recall of risk scenario segment identification are statistically analyzed, the corresponding F1 score is calculated, and the candidate threshold with the largest F1 score is selected as the template matching threshold. For example, in this embodiment, the template matching threshold is preferably between 0.70 and 0.85.

[0251] When the template matching degree of a candidate risk scenario segment reaches the template matching threshold, the candidate risk scenario segment is determined as a supplier risk path segment that matches the corresponding risk scenario feature path template.

[0252] The risk intensity of a risk pathway segment is determined based on the risk monitoring status trajectory values ​​covered by the segment. The specific calculation process is as follows: the absolute values ​​of the risk monitoring status trajectory values ​​of each risk time slice within the segment are processed and converted to 0 to 1 according to the normalization scale adopted in step S20144; the normalized absolute values ​​corresponding to all risk time slices and all effective risk status components are accumulated, and then divided by the number of effective status values ​​participating in the accumulation to obtain the risk intensity of the corresponding supplier risk pathway segment.

[0253] The supplier identifier, fusion agent identifier, risk scenario feature path template identifier, start risk time slice, end risk time slice, and risk intensity are associated with the corresponding supplier risk path segments, and all matching results are summarized to obtain the fusion agent supplier risk path set.

[0254] Step S303: Based on the supplier risk causal graph, perform causal time-series correction on the risk paths corresponding to different fusion agents to obtain a set of corrected risk paths.

[0255] In this embodiment, based on the risk path set of the fusion agent supplier, risk paths belonging to the same supplier but with different fusion agent identifiers are extracted according to the supplier identifier to obtain the risk path group to be corrected.

[0256] For any risk path, based on the sequence of risk event categories and corresponding risk time slices associated with that risk path, we search for event nodes belonging to the same supplier, with the same risk event category, and corresponding to the same risk time slice in the supplier's risk causality graph. If multiple event nodes correspond to the same risk event category, we select event nodes from the candidate event nodes that can form a continuous causal relationship with the previous risk event node based on the causal edges between the event nodes, and connect them according to the order of the risk event category sequence to obtain the sequence of risk event nodes corresponding to the risk path.

[0257] For different fusion agent risk paths in the risk path group to be corrected, starting from the first risk event node corresponding to each risk path, we traverse backward along the causal relationship edge in the supplier risk causal graph to extract the set of preceding risk events corresponding to each risk path, and perform intersection processing on the set of preceding risk events of different risk paths to obtain a common set of preceding risk events.

[0258] When the candidate set of common preceding risk events contains multiple common preceding risk events, the number of causal relationship edges between each common preceding risk event and the first risk event node of each risk path is counted. The number of causal relationship edges corresponding to the same common preceding risk event is accumulated, and the common preceding risk event with the smallest accumulated result is determined as the common preceding risk event of the corresponding risk path group.

[0259] Obtain the risk time slice where the common preceding risk event is located and the risk time slice where the first risk event node of each risk path is located. Use the time slice difference between the risk time slice where the first risk event node of each risk path is located and the risk time slice where the common preceding risk event is located as the time offset of the corresponding risk path.

[0260] The original time slice sequence number of each risk time slice in the risk path is subtracted from the corresponding time offset to align the first risk event node of different fused intelligent agent risk paths according to the unified time reference corresponding to the common preceding risk event, while keeping the relative time interval between risk time slices within the same risk path unchanged, thus obtaining the corrected risk path. This process is repeated for all risk paths to obtain the corrected risk path set.

[0261] Through the above-mentioned causal timing correction, risk pathways that originate from the same prior risk event but have different formation times in different business modules due to qualification information, transaction behavior, procurement behavior or external regulatory information can be transformed to a unified time benchmark according to the common causal source.

[0262] Step S304: Construct a risk path hierarchy map of the fusion agent based on the corrected risk path set, identify overlapping risk paths, and perform path rewriting and risk assessment to obtain a set of supplier risk assessment results.

[0263] The specific steps of step S304 are as follows:

[0264] Step S3041: Based on the supplier risk cause-effect diagram, identify the common preceding risk events of different fusion agent risk paths, and perform time location correction on the risk paths.

[0265] In this embodiment, based on the set of correction risk paths, the common preceding risk events and time offsets associated with each correction risk path in step S303 are read, and the correction risk paths are grouped according to the supplier identifier.

[0266] For different fusion agent risk paths corresponding to the same supplier, the common antecedent risk events corresponding to each risk path are identified according to the supplier risk causal graph. When different fusion agent risk paths are associated with the same common antecedent risk event, the risk paths are classified into the same causal source group.

[0267] Based on the corrected start risk time slice and corrected end risk time slice of each corrected risk path, the correction time interval of each risk path under a unified time reference is determined, and the common pre-risk event identifier, the corrected start risk time slice and the corrected end risk time slice are associated with the corresponding corrected risk path for subsequent determination of the time overlap relationship between risk paths of different fusion agents.

[0268] For different fused agent risk pathways where common preceding risk events cannot be identified in the supplier risk causal graph, they are not assigned to the same causal source group, and the original correction time position of each risk pathway is maintained.

[0269] Step S3042: Map the corrected risk paths to path nodes, establish time continuation edges within the same fused intelligent body, and establish path crossing edges between risk paths of different fused intelligent bodies that have common preceding risk events and overlapping corrected times, to obtain a risk path layering graph of the fused intelligent body.

[0270] In this embodiment, based on the set of correction risk paths, each correction risk path is mapped to a path node, and the supplier identifier, fusion agent identifier, common pre-risk event identifier, correction time interval, and corresponding risk event node sequence are used as path node attributes.

[0271] Path nodes are grouped according to supplier identifier and fused agent identifier. For path nodes belonging to the same supplier and the same fused agent, they are sorted according to the corrected start risk time slice. When the corrected end risk time slice of the previous path node is adjacent or consecutive to the corrected start risk time slice of the next path node, a time continuation edge is established between the two path nodes, pointing from the previous path node to the next path node.

[0272] For path nodes belonging to the same supplier but with different agent identifiers, compare the corresponding common pre-risk event identifiers and correction time intervals. When the common pre-risk events corresponding to two path nodes are the same, and the two correction time intervals have at least one common risk time slice, the two path nodes are identified as cross-agent associated path nodes, and a path crossing edge is established between the two path nodes.

[0273] Based on the graph structure organization of all path nodes, time continuation edges, and path intersection edges, the path nodes connected by time continuation edges within the same fused agent are organized into a time continuation layer, and the path nodes connected by path intersection edges between different fused agents are organized into an agent intersection layer, thus obtaining a risk path layer graph for fused agents.

[0274] Step S3043: Based on the risk pathway hierarchy map of the fused intelligent agent, extract the overlapping risk pathways and determine the dominant and cooperative pathways.

[0275] In this embodiment, based on the risk path hierarchy graph of the fused intelligent agent, path nodes that are connected by path intersection edges and have the same common preceding risk event identifier are extracted and grouped according to the common preceding risk event identifier to obtain the same source risk path group.

[0276] For any group of risk pathways with the same origin, the correction time intervals of each risk pathway in the group are compared. Risk pathways with overlapping correction time intervals are identified as overlapping risk pathways with the same origin. Based on the sequence of risk event nodes associated with each overlapping risk pathway with the same origin, risk event nodes that are covered by the corresponding risk pathway after the common preceding risk event are extracted.

[0277] The risk event nodes covered by all risk pathways within the same overlapping risk pathway group are deduplicated and summarized to obtain the causal risk event set corresponding to that overlapping risk pathway group. The number of risk event nodes covered by each risk pathway is counted, and the count is divided by the total number of risk event nodes in the causal risk event set to obtain the causal event coverage of the corresponding risk pathway.

[0278] The risk pathway with the largest coverage of causal events was identified as the dominant pathway, and the remaining risk pathways in the same homologous overlapping risk pathway group were identified as cooperating pathways.

[0279] When multiple risk paths have the same maximum causal event coverage, the number of causal edges traversed between the first risk event node and the common preceding risk event in each risk path is compared, and the risk path with fewer causal edges is identified as the dominant path. If the number of causal edges is still the same, the risk intensity of the corresponding risk paths is compared, and the risk path with higher risk intensity is identified as the dominant path.

[0280] By using the coverage of causal events as the basis for determining the dominant pathway, the dominant pathway is given priority to retain risk information that covers the causal evolution process of the same risk more completely, thus avoiding the omission of subsequent causal risk processes caused by selecting the dominant pathway only based on the intensity of local risks.

[0281] Step S3044: Merge overlapping path segments in the dominant and collaborative paths, connect non-overlapping path segments in the collaborative paths to the dominant path, and perform risk assessment to obtain a set of supplier risk assessment results.

[0282] In this embodiment, based on the dominant and cooperating pathways determined in step S3043, the risk pathway segments contained in each pathway are read in the corrected risk time slice order.

[0283] For path segments in the dominant and cooperative pathways that share the same common preceding risk event and correspond to the same causal propagation location, these path segments are identified as overlapping path segments. The corresponding path segments in the dominant pathway are retained, and the fusion agent identifier corresponding to the cooperative pathway is associated with the retained path segments, thus completing the merging of overlapping path segments.

[0284] For overlapping pathway segments, the risk intensity of the dominant pathway and the risk intensity of the co-pathway are not directly added together to avoid repeated amplification of the same risk source. For the same corrected risk time slice, the risk intensity of the dominant pathway segment and the risk intensity of the associated co-pathway segment are compared, and the larger risk intensity is taken as the merged risk intensity of that risk time slice.

[0285] For pathway segments in the collaborative pathway that do not overlap causally with the dominant pathway, they are identified as non-overlapping pathway segments. Based on the supplier risk causality graph, the preceding and subsequent causal positions of the risk events corresponding to the non-overlapping pathway segments are determined, and the non-overlapping pathway segments are connected to the dominant pathway according to the direction of the causal relationship edges. After connection, all pathway segments are rearranged according to the corrected risk time slice order to obtain the rewritten risk pathway.

[0286] For each risk time slice in the rewritten risk pathway, the corresponding merged risk intensity is read; for risk time slices formed by non-overlapping pathway segments, the original risk intensity of the non-overlapping pathway segment is directly read to obtain the risk intensity time series of the rewritten risk pathway.

[0287] The risk level classification threshold is determined based on historically labeled supplier risk pathway samples. The specific method is as follows: Supplier risk pathway samples are grouped according to historical risk levels, and the risk intensity corresponding to each sample is read. For two adjacent risk levels, the risk intensities appearing in the samples of both levels are sequentially used as candidate level thresholds. The classification accuracy of each candidate threshold for the two levels is calculated, and the data is balanced according to the number of samples in each level. The risk intensity with the highest balanced classification accuracy is selected as the level threshold between two adjacent risk levels. For multiple risk levels, adjacent level thresholds are determined sequentially from low to high risk level, ensuring that each level threshold increases with the risk level.

[0288] For example, when the risk intensity has been normalized to 0 to 1, and a three-level classification of low risk, medium risk, and high risk is adopted, the threshold for the low-risk to medium-risk boundary obtained according to the above historical sample discrimination method is preferably 0.35 to 0.50, and the threshold for the medium-risk to high-risk boundary is preferably 0.65 to 0.80. These values ​​are only used as reference ranges for the embodiments obtained according to the above determination method, and are not intended as fixed limitations.

[0289] The risk intensity corresponding to each risk time slice is compared with the risk level threshold. When the risk intensity is lower than the first level threshold, it is determined as the lowest risk level; when the risk intensity is between two adjacent level thresholds, it is determined as the corresponding intermediate risk level; when the risk intensity reaches the highest level threshold, it is determined as the highest risk level.

[0290] Based on the supplier identifier and risk time slice, the risk level, the fusion agent identifier corresponding to the dominant path, the fusion agent identifier corresponding to the collaborative path, and the corresponding risk intensity are associated, and the supplier risk assessment results are summarized to obtain a set of results.

[0291] Step S305: Identify the dominant risk sources and collaborative risk sources based on the supplier risk assessment result set, and perform joint assessment and risk classification to obtain the supplier risk monitoring results.

[0292] In this embodiment, based on the supplier risk assessment result set, the dominant path fusion agent identifier and the collaborative path fusion agent identifier corresponding to each risk assessment result are read according to the supplier identifier and risk time slice.

[0293] The fusion agent that forms the main body of the rewritten risk path and is identified as the dominant path in step S3043 is identified as the dominant risk source of the corresponding supplier in the corresponding risk time slice. Other fusion agents whose path segments are merged into the dominant path, or whose non-overlapping path segments are connected to the dominant path and participate in risk assessment, are identified as collaborative risk sources.

[0294] Based on the dominant risk source and synergistic risk source corresponding to each risk time slice, the risk assessment result corresponding to the dominant pathway is read as the basic risk level, and the risk intensity and risk level output conditions of the pathway segment corresponding to each synergistic risk source are read respectively to determine whether the synergistic risk source has an effect on improving the basic risk level.

[0295] The specific judgment method is as follows: The risk monitoring status trajectory value and risk intensity corresponding to the collaborative pathway segment are matched with the risk level output conditions in the corresponding risk scenario characteristic pathway template. When the risk level corresponding to the collaborative pathway segment is lower than or equal to the basic risk level, the basic risk level remains unchanged. When the risk level corresponding to the collaborative pathway segment is higher than the basic risk level, and the collaborative pathway segment forms a continuous causal relationship with the dominant pathway through the causal relationship edge in the supplier risk causal graph, the basic risk level is raised to the risk level corresponding to that collaborative pathway segment. When multiple collaborative risk sources exist, the above judgment is performed sequentially, and the highest effective risk level obtained is taken as the joint risk level of the corresponding risk time slice.

[0296] The joint risk levels of each supplier are arranged according to the time sequence of the risk time slices to obtain the supplier risk level time series. The dominant risk source and the collaborative risk source corresponding to each risk time slice are then associated with the supplier risk level time series to form the supplier risk monitoring results.

[0297] The supplier risk monitoring results include at least the supplier identifier, risk time slice, corresponding risk level, dominant risk source, and collaborative risk source, which are used to characterize the risk status and corresponding risk sources of the supplier at different risk time slices.

[0298] Example 2

[0299] Please see Figure 4 The present invention provides an embodiment of a supplier risk intelligent monitoring system that integrates intelligent agents, including a data acquisition module, a risk evolution module, and a risk monitoring module.

[0300] The data acquisition module is used to acquire supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set.

[0301] The risk evolution module constructs a supplier risk monitoring evolution equation based on the supplier graph signal set to obtain the supplier risk scenario.

[0302] The risk monitoring module, based on the supplier risk scenario, constructs a hierarchical risk path diagram of the integrated intelligent agent to monitor supplier risks.

[0303] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0304] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A supplier risk intelligent monitoring method integrating intelligent agents, characterized in that, include: Obtain supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set; Based on the supplier graph signal set, a supplier risk monitoring evolution equation is constructed to obtain the supplier risk scenario field. The supplier risk monitoring evolution equation is obtained by performing graph frequency domain decomposition on the supplier graph signal set, extracting long-term risk trend features and short-term abnormal risk features, performing risk evolution mode decomposition on the multidimensional risk monitoring state vector, and performing risk-time slice extrapolation on the low-dimensional risk monitoring state vector sequence. Based on the supplier risk scenario, a risk path hierarchy map of the integrated intelligent agent is constructed to monitor supplier risks.

2. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 1, characterized in that, The process of acquiring supplier business data, constructing a supplier risk cause-and-effect graph, and obtaining a supplier graph signal set includes: Obtain supplier business data, divide the supplier business data into risk time slices, generate business scenario keys, and obtain a candidate set of supplier business events; Risk sensitivity joint discrimination is performed on each supplier business candidate event in the supplier business event candidate set, and normalized coding is performed to obtain the supplier risk event set; Based on the supplier risk event set, risk nodes are generated through node mapping, and supplier risk relationship edges are established between each risk node to obtain a preliminary graph of multiple supplier risk relationships. The risk nodes include supplier nodes, event nodes, and scenario nodes. Based on the initial graph of multiple relationships of supplier risks, causal determination and merging are performed on the risk relationship edges of each supplier, and risk nodes are aggregated to obtain the supplier risk causal graph. Based on the supplier risk causal graph, a time series of risk characteristics corresponding to each supplier node is constructed to obtain the supplier graph signal set.

3. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 2, characterized in that, The process involves constructing a time series of risk characteristics corresponding to each supplier node based on a supplier risk causal graph, resulting in a supplier graph signal set, including: Based on the supplier risk causal graph, the category and intensity distribution of each event node adjacent to each supplier node in each risk time slice are statistically analyzed to obtain the supplier risk neighborhood sequence. Based on the supplier risk neighborhood sequence, the category and intensity distribution within each risk time slice are aggregated into a multidimensional risk feature vector, and time-series alignment is performed to obtain the risk feature time series corresponding to each supplier. The risk characteristic time series is mapped to the supplier risk monitoring signal sequence corresponding to each supplier node, and graph smoothing and anomaly retention processing are performed to obtain the supplier graph signal set.

4. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 3, characterized in that, The supplier risk monitoring evolution equation is constructed based on the supplier graph signal set to obtain the supplier risk scenario field, including: Based on the supplier graph signal set, graph frequency domain multi-scale decomposition is performed to extract long-term risk trend features and short-term abnormal risk features, resulting in a multi-dimensional risk monitoring state vector sequence. Based on the multidimensional risk monitoring state vector sequence, a supplier risk monitoring evolution equation is constructed, and risk-by-risk time slice extrapolation is performed to obtain the risk monitoring state trajectory of each supplier in each risk time slice. Based on the risk monitoring status trajectory, the risk monitoring status trajectory values ​​of each supplier in each risk time slice are extracted, and the risk monitoring status trajectory values ​​are spliced ​​together in the same time slice to obtain a risk monitoring status field arranged according to the risk time slice. Based on the risk monitoring state field, the risk monitoring state trajectory values ​​of adjacent risk time slices in the risk monitoring state field are interpolated and smoothed to obtain the supplier risk scenario field.

5. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 4, characterized in that, The process involves performing multi-scale decomposition in the graph frequency domain based on the supplier graph signal set to extract long-term risk trend features and short-term abnormal risk features, resulting in a multi-dimensional risk monitoring state vector sequence, including: Based on the supplier graph signal set, the risk monitoring signal sequences of each supplier are uniformly preprocessed and time-aligned to obtain the supplier risk monitoring signal matrix. Based on the supplier risk causal graph, a supplier risk adjacency operator and a graph Laplacian operator are constructed, and the supplier risk monitoring signal matrix is ​​mapped to the graph frequency domain to obtain the supplier risk graph spectrum. The supplier risk map spectrum is divided into frequency bands, and the signal components are reconstructed to obtain low-frequency risk components and high-frequency risk components; Based on low-frequency and high-frequency risk components, long-term risk trend features and short-term abnormal risk features are extracted, and feature aggregation is performed within each risk time slice to obtain a multi-dimensional risk monitoring state vector sequence.

6. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 5, characterized in that, The process involves extracting long-term risk trend features and short-term abnormal risk features based on low-frequency and high-frequency risk components, and aggregating these features within each risk time slice to obtain a multi-dimensional risk monitoring state vector sequence, including: Based on low-frequency and high-frequency risk components, the low-frequency and high-frequency risk component values ​​of each supplier in each risk time slice are organized to construct the supplier risk time slice feature tensor. Based on the supplier risk time slice feature tensor, a segmented low-rank decomposition is performed with risk time slice as the segment unit to obtain long-term risk trend features and residual sub-tensor. Based on the residual subtensor, a risk anomaly energy function is constructed, and the risk anomaly energy function is minimized and reconstructed to obtain short-term anomaly risk characteristics. Within each risk time slice, the long-term risk trend characteristics and short-term abnormal risk characteristics are normalized and spliced ​​together to obtain a multi-dimensional risk monitoring state vector sequence.

7. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 6, characterized in that, The method involves constructing a supplier risk monitoring evolution equation based on a multidimensional risk monitoring state vector sequence, and performing risk-by-risk time-slice extrapolation to obtain the risk monitoring state trajectory of each supplier at each risk time-slice, including: Risk evolution mode decomposition is performed on the multidimensional risk monitoring state vector sequence to extract the supplier risk evolution fundamental mode; A low-dimensional risk evolution subspace is constructed based on the supplier risk evolution basic model, and the multi-dimensional risk monitoring state vector is projected to obtain a low-dimensional risk monitoring state vector sequence. Clustering and segmenting of low-dimensional risk monitoring state vector sequences to identify risk evolution state types; Based on the risk evolution state type, the temporal evolution relationship between low-dimensional risk monitoring state vectors is determined, and parameter fitting is performed to obtain the supplier risk monitoring evolution equation. Based on the supplier risk monitoring evolution equation, the supplier risk monitoring state trajectory is obtained by performing risk-by-risk time slice extrapolation on the low-dimensional risk monitoring state vector sequence.

8. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 7, characterized in that, The method of constructing a risk path hierarchy map for a fusion intelligent agent based on supplier risk scenarios to monitor supplier risks includes: The supplier risk monitoring rules are analyzed, a set of integrated intelligent agents is constructed, and a corresponding set of risk scenario feature path templates is generated. Risk scenario fragments are extracted from the supplier risk scenario field and matched with the risk scenario feature path template set to obtain the supplier risk path set of the fusion agent. Based on the supplier risk causal graph, causal time sequence correction is performed on the risk paths corresponding to different fusion intelligent agents to obtain a set of corrected risk paths; Based on the set of corrected risk paths, a hierarchical risk path diagram of the fusion agent is constructed to identify overlapping risk paths, and path rewriting and risk assessment are performed to obtain a set of supplier risk assessment results. Based on the supplier risk assessment results set, the dominant risk sources and collaborative risk sources are identified, and joint assessment and risk classification are performed to obtain the supplier risk monitoring results.

9. The supplier risk intelligent monitoring method for fused intelligent agents according to claim 8, characterized in that, The process involves constructing a hierarchical risk path diagram for the fusion agent based on the corrected risk path set, identifying overlapping risk paths, rewriting the paths, and determining the risks to obtain a set of supplier risk determination results, including: Based on the supplier risk cause-effect diagram, the common preceding risk events of different fusion intelligent agent risk paths are identified, and the risk paths are corrected for their time and location. The corrected risk paths are mapped as path nodes. Time continuity edges are established within the same fused intelligent body. Path intersection edges are established between risk paths of different fused intelligent bodies that have common preceding risk events and whose corrected times overlap, resulting in a risk path layering graph of the fused intelligent body. Based on the risk path hierarchy map of the fused intelligent agent, the overlapping risk paths are extracted and the dominant and cooperative paths are identified. Overlapping path segments in the dominant and collaborative paths are merged, non-overlapping path segments in the collaborative paths are connected to the dominant path, and risk assessment is performed to obtain a set of supplier risk assessment results.

10. A supplier risk intelligent monitoring system for a fusion intelligent agent, used to implement the supplier risk intelligent monitoring method for a fusion intelligent agent as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire supplier business data, construct a supplier risk cause-and-effect graph, and obtain a supplier graph signal set; The risk evolution module, based on the supplier graph signal set, constructs the supplier risk monitoring evolution equation to obtain the supplier risk scenario field; The risk monitoring module, based on the supplier risk scenario, constructs a risk path hierarchy diagram of the integrated intelligent agent to monitor supplier risks.

Images