Supply chain inventory pledge dynamic risk control method and system based on multi-source data fusion

Abstract

The application discloses a supply chain inventory pledge dynamic risk control method and system based on multi-source data fusion, relates to the technical field of supply chain financial risk control, and comprises the following steps: separating a contract residual error tensor into a contract cooperative offset component, generating a local abnormality matrix, calculating a local abnormality amplitude and a continuous abnormality proportion, constructing a dynamic residual error state vector based on the product of the two, and generating a dynamic comprehensive risk value; the application extracts the contract cooperative offset component, generates a local abnormality matrix for the residual error after deducting the contract cooperative offset component, and constructs a dynamic residual error state vector in combination with the local abnormality amplitude and the continuous abnormality proportion, so that the dynamic comprehensive risk value reflects an abnormal state that deviates from the common change of the contract layer and continuously exists in time, avoids misidentifying contract layer synchronous fluctuations as batch local abnormalities, and also avoids submerging batch local continuous abnormalities in contract overall fluctuations, thereby improving the hierarchy, accuracy and interpretability of the risk identification result.

Description

Technical Field

[0001] This invention relates to the field of supply chain finance risk control technology, and in particular to a dynamic risk control method and system for supply chain inventory pledge based on multi-source data fusion.

[0002] Inventory pledging has become an important way to alleviate corporate financing constraints and improve the liquidity of inventory assets. In existing technologies, risk management around inventory pledging has gradually formed a data-driven risk control path based on warehouse supervision, logistics tracking, transaction document verification, financing balance monitoring, and market price assessment. It combines multi-source business data such as warehouse ledgers, inventory records, order invoices, cash flow, credit limits, and market conditions. It is gradually evolving from single-point verification to multi-source heterogeneous data collaborative analysis, continuous time-series monitoring, and dynamic early warning, in order to promptly detect risk factors such as inventory distortion, missing documents, abnormal retention, value decline, and credit occupancy during business operations.

[0003] Existing supply chain inventory pledge dynamic risk control methods based on multi-source data fusion directly judge abnormal results in a single sampling period. Although they can detect risk factors such as inventory distortion, missing documents, logistics delays, abnormal payments, price declines, or credit occupancy, they lack a unified organization of multiple batches and multiple residual channels within the continuous contract monitoring window, and a method to first separate common changes at the contract level and then identify local continuous anomalies in batches. This leads to easy confusion between synchronous fluctuations at the contract level and local continuous anomalies in batches, affecting the hierarchy, accuracy, and stability of anomaly source identification. Summary of the Invention

[0004] Firstly, this invention provides a dynamic risk control method for supply chain inventory pledging based on multi-source data fusion, including:

[0005] By acquiring warehousing data, transaction data, logistics data, and financing data that are time-aligned under the same business identifier, a pledge detection unit can be constructed.

[0006] Based on the effective inventory observation set and financing support constraints in the pledge detection unit, the batch-consistent pledgeable quantity is determined, a contract continuous monitoring window is set, and a batch residual channel is constructed based on the batch-consistent pledgeable quantity. Within the contract continuous monitoring window, the batch residual channel is organized into a contract residual tensor according to the sampling time sequence.

[0007] A batch business graph is constructed based on the business characteristics in the pledge detection unit, a residual channel graph is constructed based on the batch residual channel, a two-sided constraint projection is applied to the contract residual tensor, and the contract collaborative offset component is extracted.

[0008] Separate the contract collaborative offset component from the contract residual tensor to generate a local anomaly matrix, calculate the local anomaly magnitude and the proportion of persistent anomalies, and construct a dynamic residual state vector based on the product of the two to generate a dynamic comprehensive risk value.

[0009] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, the step of performing bilateral constraint projection on the contract residual tensor and extracting the contract collaborative offset component includes:

[0010] For each valid pledge batch of the same contract, extract the business description vector, calculate the business distance based on any two business description vectors, calculate the business coupling weight based on the business distance, construct the business coupling matrix and perform normalization processing to obtain the batch normalized coupling matrix.

[0011] The contract residual tensor is expanded in terms of time and batch direction. The values ​​of the residual channels within the window are organized into vectors. The absolute value of the Pearson correlation coefficient between any two residual channel vectors is calculated to obtain the channel coupling weight. The channel coupling matrix is ​​constructed and normalized to obtain the channel normalization matrix.

[0012] The contract residual tensor is used to extract the batch channel residual slice matrix according to the sampling period. The batch channel residual slice matrix is ​​then multiplied by the batch normalized coupling matrix and then multiplied by the channel normalized matrix to obtain the coupled batch channel residual slice matrix. The matrix is ​​then reassembled according to the sampling period and expanded along the time direction to obtain the contract residual matrix.

[0013] Singular value decomposition is performed on the contract residual matrix to obtain a diagonal matrix, a left singular vector, and a right singular vector. Based on the diagonal matrix, the median of the singular values ​​and the median of the absolute deviation of the singular values ​​are calculated. The singular values ​​are then filtered, and the contract collaborative offset matrix is ​​reconstructed by combining the left and right singular vectors.

[0014] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, the step of separating the contract collaborative offset component from the contract residual tensor to generate a local anomaly matrix includes:

[0015] Subtract the contract collaboration offset matrix from the contract residual matrix to obtain the remaining residual matrix. Perform column processing on the remaining residual matrix to generate the local anomaly matrix.

[0016] The column processing includes, for any column, extracting the elements of each row of the column, calculating the absolute deviation of the median, and retaining the element if the absolute value of the difference between the element and the median of the column is greater than the corresponding absolute deviation of the median; otherwise, setting the element to zero.

[0017] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, the step of calculating the amplitude of local anomalies and the proportion of persistent anomalies, and constructing a dynamic residual state vector based on their product, includes:

[0018] For the batch residual channels in the local anomaly matrix, the absolute values ​​of the element values ​​are extracted in channel order to obtain the local anomaly amplitude.

[0019] Within the contract continuous monitoring window, the duration of the batch residual channel is counted along the time direction, and the duration is divided by the contract continuous monitoring window length to obtain the percentage of continuous anomalies in the batch residual channel.

[0020] Multiply the local anomaly amplitude of each channel by the corresponding proportion of continuous anomalies to obtain the dynamic state variables of the batch residual channels, and then splice the first and last values ​​to obtain the dynamic residual state vector.

[0021] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, wherein: the generation of dynamic comprehensive risk value includes:

[0022] Based on the dynamic residual state vector, the data reliability risk value, authenticity risk value, price coverage risk value, and performance continuity risk value are calculated to construct a risk value set;

[0023] The risk value set is summed and averaged to obtain a dynamic comprehensive risk value. The dynamic comprehensive risk value is compared with a preset risk threshold to obtain a risk level, and corresponding dynamic risk control actions are generated according to the risk level.

[0024] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, the step of constructing batch residual channels and organizing the batch residual channels into contract residual tensors according to the sampling time sequence within the contract continuous monitoring window includes:

[0025] Set up a continuous contract monitoring window and construct a batch residual channel based on the batch-consistent pledgeable quantity;

[0026] Based on the time dimension, batch dimension, and residual channel dimension, the batch residual channels of each sampling period within the contract continuous monitoring window are expanded and arranged to generate the contract residual tensor.

[0027] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, the determination of the batch-consistent pledgeable quantity based on the effective inventory observation set and financing support constraints in the pledge detection unit includes:

[0028] Extract inventory observations from the pledge detection unit, including book inventory quantity, physical inventory quantity, inventory observations calculated based on turnover, and document-supported inventory observations.

[0029] The inventory observations are filtered to obtain the effective inventory observation set. The arithmetic mean of the effective inventory observation set is then calculated to obtain the batch-consistent inventory quantity.

[0030] By comparing the batch-consistent inventory quantity with the document-supported inventory observation quantity, extracting the minimum value between the two, and applying financing support constraints, the batch-consistent pledgeable quantity is obtained.

[0031] As a preferred embodiment of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion described in this invention, the step of acquiring warehousing data, transaction data, logistics data, and financing data aligned to the same business identifier and time to construct a pledge detection unit includes:

[0032] Multiple sources of raw data are acquired in units of sampling period and arranged horizontally to construct a pledge detection unit for that sampling period;

[0033] The multi-source raw data includes warehousing data, logistics data, transaction data, and financing data.

[0034] Secondly, this invention provides a dynamic risk control system for supply chain inventory pledging based on multi-source data fusion, including:

[0035] The batch-consistent pledgeable quantity determination module is used to extract the book inventory quantity, physical inventory quantity, inventory observations calculated from circulation, and inventory observations supported by documents. It constructs an effective inventory observation set and performs robust processing. Combining the batch status coefficient, the upper limit of financing support quantity, and the effective pledge batch screening results, it determines the batch-consistent pledgeable quantity.

[0036] The contract continuous monitoring window and nine batch residual channel construction module are used to determine the contract continuous monitoring window based on the sampling period corresponding to the most recent valid inventory count, and to construct the following residuals within the window: book inventory deviation residual, inventory count inventory deviation residual, circulating inventory deviation residual, insufficient document coverage residual, logistics abnormal residual, insufficient payment fulfillment residual, value coverage residual, credit occupancy residual, and data abnormal residual.

[0037] The batch business graph and residual channel graph construction module is used to construct a batch business graph based on the business description vector of the pledged batch, and to construct a residual channel graph based on the correlation between the nine types of batch residual channels;

[0038] The Contract Collaborative Offset and Local Persistent Anomaly Separation Module is used to perform dual-graph business coupling of batch business graph and residual channel graph on the contract residual tensor, extract the contract layer collaborative offset part, and separate the local anomaly and persistent anomaly parts from the remaining residual.

[0039] The module for constructing dynamic residual state vectors and risk value sets is used to form dynamic residual state vectors based on the magnitude of local anomalies and the proportion of persistent anomalies, and to construct risk value sets.

[0040] The Dynamic Comprehensive Risk Value and Risk Level Handling Module is used to merge the risk value set to obtain a dynamic comprehensive risk value, classify risk levels, and generate dynamic risk control actions.

[0041] The beneficial effects of this invention are as follows: By performing time-series alignment and joint processing on warehousing data, logistics data, transaction data, and financing data under a unified sampling period, this invention determines the batch-consistent collateralizable quantity at the batch level, constructs a contract residual tensor at the contract level, and applies bilateral constraint processing to the contract residual tensor by combining the batch business graph and the residual channel graph. First, the contract collaborative offset component is extracted, and then a local anomaly matrix is ​​generated from the remaining residual after deducting the contract collaborative offset component. By combining the local anomaly amplitude and the proportion of persistent anomalies, a dynamic residual state vector is constructed. Thus, the dynamic comprehensive risk value reflects the abnormal state that deviates from the common changes at the contract level and persists in time. This avoids misidentifying synchronous fluctuations at the contract level as batch local anomalies and also avoids submerging batch local persistent anomalies in the overall contract fluctuations, thereby improving the hierarchy, accuracy, and interpretability of the risk identification results. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of the supply chain inventory pledge dynamic risk control method based on multi-source data fusion in Example 1;

[0044] Figure 2 This is a schematic diagram of the supply chain inventory pledge dynamic risk control system based on multi-source data fusion in Example 1. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a dynamic risk control method for supply chain inventory pledging based on multi-source data fusion, including the following steps:

[0047] S1. Under the same business identifier, construct a pledge detection unit by time-aligned warehousing data, transaction data, logistics data, and financing data;

[0048] Specifically, by acquiring warehousing data, transaction data, logistics data, and financing data under the same business identifier and through time alignment, a pledge detection unit is constructed, including:

[0049] The API interface is used to synchronously acquire data in sampling periods ranging from 1 to 24 hours. Multiple sources of raw data from each sampling period will be used to pledge contracts. The pledge batch and the business entity are set as business identifiers, and the same business identifier is extracted at the same sampling time. The multi-source raw data are arranged horizontally to obtain the first... A pledge detection unit for each sampling cycle;

[0050] The above business identifiers are generated using a joint identifier method. They are considered to be the same business identifier only when the pledge contract number, the pledged goods batch number, and the business entity number are all the same.

[0051] By using the above-mentioned joint identifier generation method, it can be ensured that the original data from multiple sources correspond to the same pledge batch under the same contract when arranged horizontally, thus avoiding mismatch between different business records caused by using only a single field for matching.

[0052] If the above is a certain type of data In the If any sampling period has missing data, then extract the corresponding data from the sampling period with no missing data. For example, the first indivual, To satisfy the condition that the data is the smallest positive integer, satisfying the condition that... , for the Missing values ​​in each sampling period are filled in, if consecutive If any data point is missing in any sampling period and cannot be filled in forward (meaning the data is missing in all forward sampling periods), then the anomaly flag for that data is set to 1; otherwise, it is set to 0. The corresponding batch of pledged goods marked as anomaly 1 is then assigned to this flag. Mark as an abnormal data state;

[0053] The above The maximum number of backtracking sampling periods allowed for backward completion, such as 1, 2, or 3, is used to limit data completion to only those caused by short-term missed sampling, interface delays, or single communication interruptions, avoiding the replacement of the current sampling period's true observations with long-term outdated data. Greater than If so, then directly set the exception flag to 1;

[0054] The aforementioned multi-source raw data includes warehousing data, logistics data, transaction data, and financing data;

[0055] The aforementioned warehousing data includes inventory quantity, storage location information, inbound time, outbound time, and inventory count records;

[0056] The above logistics data includes records of abnormal delays;

[0057] The aforementioned transaction data includes the number of orders, invoice information, amount received, and timing of receipt.

[0058] The above financing data includes credit line, pledge ratio, financing balance, and overdue status.

[0059] S2. Based on the effective inventory observation set and financing support constraints in the pledge detection unit, determine the batch-consistent pledgeable quantity, set the contract continuous monitoring window, construct the batch residual channel based on the batch-consistent pledgeable quantity, and organize the batch residual channel into a contract residual tensor according to the sampling time sequence within the contract continuous monitoring window.

[0060] Specifically, based on the effective inventory observation set and financing support constraints in the collateral detection unit, the batch-consistent collateralizable quantity is determined, including:

[0061] For each batch of pledged goods under the pledge agreement The inventory observations are extracted from the pledge detection unit, including book inventory quantity, physical inventory quantity, inventory observations calculated from turnover, and document-supported inventory observations.

[0062] The above-mentioned book inventory quantity is directly read from the inventory quantity in the warehouse data;

[0063] The inventory quantities mentioned above are directly read from the inventory records in the warehouse data;

[0064] The above-mentioned inventory observations for estimated turnover are derived from the first The batch-consistent inventory quantity for each sampling period is obtained by adding the inbound quantity and subtracting the outbound quantity. The batch-consistent inventory quantity is set to the book inventory quantity during the initial sampling period.

[0065] The above document supports inventory observation, which is set based on the minimum of the quantity supported by the order and the quantity supported by the invoice. This represents the quantity of goods that have been signed but not yet shipped in the order information and the quantity of goods that have been confirmed as received by the invoice unit in the invoice information.

[0066] For inventory observations, a valid inventory observation set is defined based on unlabeled data. This includes defining the book inventory quantity, physical inventory quantity, and turnover-estimated inventory observations as the valid inventory observation set when the warehouse data is 0; setting the document-supported inventory observation as the valid inventory observation set when the transaction data is 0; and directly setting the inventory observation as the valid inventory observation set when both the warehouse data and transaction data are not 0.

[0067] The above data showing 0 for both storage and transaction is 0, indicating that in the first... In the sampling period, if this type of data is in the first sampling period... If a sampling period has an original value or can be padded, the data is marked as 1; otherwise, the data is marked as 0, which does not mean that the actual value of this type of data is 0.

[0068] By defining business data separately from data source availability, we can avoid misjudging true zero values ​​as invalid observations.

[0069] Median extraction is performed on the effective inventory observation set. This involves calculating the absolute value of the data in the effective inventory observation set minus the median, extracting the median from the absolute value, obtaining the median absolute deviation, and then limiting the data in the effective inventory observation set to obtain the inventory observation quantity. The formula is as follows:

[0070] ;

[0071] in, For inventory observation, The median extracted from the effective inventory observation set. For symbolic functions, Set to 1 when the value is greater than 0. Set to -1 when less than 0, and set to 0 when equal to 0. For data in the effective inventory observation set, This represents the absolute deviation of the median.

[0072] Pull inventory observations that deviate too far from the robust center back into the current effective dispersion range;

[0073] The arithmetic mean of inventory observations in the effective inventory observation set is used to obtain the batch-consistent inventory quantity. This quantity is then filtered to obtain the batch-consistent pledgeable quantity, representing the batch of pledged goods. In the The amount that can be pledged at each sampling period represents the actual existence of goods and corresponding transaction documents;

[0074] The pledgeable quantity for the above batch consistency is the minimum of the document-supported inventory observation and the batch-consistent inventory quantity.

[0075] pledge contract In the The total inventory is obtained by summing up the number of valid pledge batches for each sampling period, based on the consistent pledgeable quantities of each batch.

[0076] The above-mentioned valid pledge batches refer to pledge batches that have not been marked as having abnormal data status, have not been marked as frozen by the business system, have not exceeded the warehouse age, and have no environmental anomalies.

[0077] The aforementioned valid pledge batches are used to calculate the pledgeable amount consistent with the batch, the total pledgeable amount consistent with the contract, and to participate in the construction of the business graph for subsequent batches;

[0078] When constructing nine batch residual channels within the contract continuous monitoring window, the corresponding residuals are calculated separately for each batch of pledged goods entering the contract continuous monitoring window under the same business identifier. Among them, the pledged goods batches marked as data abnormality are not included in the batch-consistent pledgeable quantity and the contract-consistent pledgeable total quantity, but their abnormality mark is retained for the calculation of data abnormality residuals. For this type of pledged goods batch, only its abnormality mark is included in the data abnormality residual, and its inventory observation, document support quantity, repayment amount, contract-consistent security pledge value and financing amount are not included in the calculation of other residual channels.

[0079] The above-mentioned absence of environmental anomalies refers to obtaining the alarm status of the batch's storage location through the API interface;

[0080] The above-mentioned "not exceeding the storage age" refers to the calculation sampling time. The actual storage age is obtained by the time difference between the storage time and the storage entry time, and the actual storage age is less than or equal to the preset storage age threshold.

[0081] The aforementioned preset storage age threshold refers to the shelf life of the pledged goods obtained through the API interface, the industry standard storage period, or the longest storage time of similar historical goods that have not suffered quality damage during the pledge period. The shortest longest storage time among the three is selected and set as the preset storage age threshold. If the industry standard storage period and the longest storage time of similar historical goods that have not suffered quality damage during the pledge period cannot be obtained, the shelf life is set as the preset storage age threshold.

[0082] The aforementioned historical similar goods refer to historical pledged goods with the same material code, the same specification grade, and the same storage area. When the number of historical samples that fully meet the above three conditions is insufficient, the scope of similar goods shall be determined according to the same material code and the same specification grade.

[0083] The aforementioned range of similar items determined by the same material code and the same specification level is only used to determine the longest historical inventory duration of the same type and the preset inventory age threshold, and is not used for subsequent business distance calculation and batch business map construction.

[0084] The above-mentioned abnormal retention records refer to records in the logistics data where the location of the same pledge batch has not been updated for two or more consecutive sampling periods, and there is no change in the status of receipt, warehousing, outbound or waybill closure, or records where the location has been updated but the actual retention time is greater than the preset transit time threshold. When any of the conditions are met, the abnormal representation corresponding to the sampling period is set to 1; otherwise, it is set to 0.

[0085] Based on price volatility and the market price of the pledged assets Calculate the security price of a contract The formula is:

[0086] ;

[0087] The system retrieves alarm fields via API and extracts abnormal retention records from multi-source raw data. If a temperature and humidity record is marked as abnormal, the alarm value is set to 1; otherwise, it is 0. Similarly, if abnormal retention records exist in the logistics data, the abnormal value is set to 1; otherwise, it is 0. A batch status coefficient is defined. The formula is:

[0088] ;

[0089] in, To indicate an alarm, This indicates an anomaly;

[0090] The above batch status coefficient A value of 1 indicates that the batch has neither abnormal storage environment nor abnormal logistics delays; otherwise, a value of 0 means that batches with obvious abnormal conditions will no longer be included in the calculation of the contract security pledge value.

[0091] Based on credit limit and pledge ratio Calculate the upper limit of the amount of financing support. This represents the maximum amount of collateral that can be supported under the credit limit, based on the collateral ratio and the safety price. The formula is:

[0092] ;

[0093] in, To prevent tiny positive numbers with a denominator of 0, such as ;

[0094] If the batch-consistent pledgeable amount is greater than or equal to the maximum pledgeable amount, then the maximum pledgeable amount will be set as the batch-consistent pledgeable amount and used in place of the original batch-consistent pledgeable amount for subsequent operations.

[0095] Furthermore, a batch residual channel is constructed, and within the contract continuous monitoring window, the batch residual channel is organized into a contract residual tensor according to the sampling time sequence, including:

[0096] Using single-period results for scoring cannot distinguish between short-term disturbances and persistent anomalies, nor can it differentiate between overall contract coordination offsets and local batch anomalies. Therefore, a continuous contract monitoring window is constructed, with the following formula:

[0097] ;

[0098] in, For the continuous monitoring window length of the contract, For pledge contract The sampling period number corresponding to the most recent valid inventory count;

[0099] In inventory pledge risk control scenarios, multiple pledge batches under the same pledge contract may simultaneously deviate on multiple residual channels due to unified inventory delays, unified accounting lags, unified price adjustments, or unified credit reductions. A single pledge batch may also continuously deviate on a few residual channels due to local inbound / outbound anomalies, missing documents, logistics delays, or local performance anomalies. If scoring is directly based on the abnormal results of a single sampling period, or if the abnormal results of all batches under the same contract are simply averaged, it is difficult to distinguish between common fluctuations at the contract level and local continuous anomalies at the batch level. It is easy to misidentify synchronous deviations at the contract level as local anomalies, and it is also easy to submerge local continuous anomalies in the overall fluctuations of the contract.

[0100] The process of taking the maximum value within the contract is to extract batch local anomalies from the remaining residuals after deducting the contract collaborative offset component. The retained dynamic state quantity does not correspond to synchronous fluctuations at the contract level, but to batch local anomalies that deviate from the common changes at the contract level and persist within the contract continuous monitoring window. This avoids repeatedly including the common changes at the contract level in the dynamic comprehensive risk value within the contract, while retaining sensitivity to the most unfavorable local anomaly state.

[0101] Within the contract continuous monitoring window, nine types of batch residual channels are constructed based on the batch-consistent pledgeable quantity, and the batch business relationship is linked with the residual channel relationship. These are then applied to the same contract residual tensor to first extract contract co-offsets and then separate local persistent anomalies, thereby improving the accuracy of anomaly attribution and the stability of risk identification results.

[0102] For any sampling period within the contract continuous monitoring window The consistent pledgeable quantities of each batch are summed to obtain the total consistent pledgeable quantity of the contract. The consistent pledgeable quantity and the contract security price are then multiplied, and the products are summed to obtain the contract security pledge value.

[0103] For any sampling period The book inventory quantity, physical inventory quantity, estimated inventory quantity, and document-supported inventory quantity are summed to obtain the total book inventory, total physical inventory, estimated inventory quantity, and document-supported inventory quantity.

[0104] Nine types of batch residual channels were constructed, including book inventory deviation residuals. Inventory deviation residuals Inventory turnover deviation residual Insufficient document coverage and residuals abnormal residuals in logistics Insufficient payment and residual amount Value Coverage Residual Credit occupancy residual and abnormal residuals The formula is:

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109]

[0110] in, The formulas for calculating the deviation residuals of book inventory, physical inventory, and circulating inventory are consistent. The total amount that can be pledged is consistent with the contract. This includes the total book inventory, the total physical inventory, and the total inventory estimated through turnover. For the number of documents supported, For the residual due to insufficient document coverage, The amount due is... This represents the actual amount received. To address the shortfall in payment collection and contract fulfillment, To cover residuals with value, To ensure the security of the pledged value in accordance with the contract, For the financing balance, For credit line, Residuals are used for credit granting;

[0111] The aforementioned logistics anomaly residuals are obtained by summing the products of the total amount that can be pledged in accordance with the contract and the anomaly representation.

[0112] The above-mentioned abnormal data residual is the summation average of the abnormal data markers of the multi-source original data corresponding to each batch of pledged goods within the contract continuous monitoring window. Among them, the abnormal data residual calculation is only retained for the batch of pledged goods that is marked as having an abnormal data status but is not included in the valid pledge batch.

[0113] To ensure that the residual channels of the nine batches can be compared horizontally and integrated in the same contract window, the residuals of all batches are uniformly expressed as the ratio of deviation to the corresponding benchmark quantity in a dimensionless manner.

[0114] For cases where the denominator may be 0 or close to 0, a positive stability term ε is added after the corresponding benchmark quantity to prevent division by zero and suppress excessive amplification of the residual value by the smallest denominator. Its value is set to a positive number that is much smaller than the normal range of the benchmark quantity, based on the statistical magnitude of the corresponding benchmark quantity.

[0115] All nine batch residuals were converted into non-negative dimensionless deviations with the same meaning. The larger the residual value, the higher the degree of deviation of the current pledged goods batch in the corresponding risk dimension. This provides a unified measurement basis for the subsequent construction of the three-dimensional contract residual tensor, implementation of dual-graph constraint projection, and formation of dynamic residual state vector.

[0116] Define a three-dimensional contract residual tensor, where the first dimension represents time, the second dimension represents batch, and the third dimension represents the nine batch residual channels.

[0117] Under the same contract, not all batches are completely independent, and the residuals are not isolated from each other. After obtaining the contract residual tensor, dual graph constraints are constructed based on the business similarity between batches and the business linkage between residual channels, so that the separated collaborative offsets and local anomalies have business interpretability.

[0118] S3. Construct a batch business graph based on the business characteristics in the pledge detection unit, construct a residual channel graph based on the batch residual channel, perform bilateral constraint projection on the contract residual tensor, and extract the contract collaborative offset component.

[0119] Specifically, a two-sided constrained projection is applied to the contract residual tensor to extract the contractual cooperative offset components, including:

[0120] Batch business diagrams and residual channel diagrams differ in their target objects and source methods. Batch business diagrams represent the horizontal coupling relationships between different batches of pledged goods under the same contract due to similarities in material codes, specifications, storage areas, and business activities. They reflect which changes between different batches should occur simultaneously.

[0121] Residual channel diagrams are used to characterize the vertical linkages between different residual sources such as inventory, documents, logistics, fulfillment, financing, and data quality, reflecting which changes between different deviation channels should be addressed together.

[0122] If the above two types of relationships are mixed into the same graph structure, it is easy to confuse the batch proximity relationship with the channel linkage relationship, causing the common fluctuations of multiple adjacent batches to be misidentified as channel anomalies, or causing the local peaks of a single channel to be misidentified as contract layer collaborative changes.

[0123] Batch business graph and residual channel graph are used to carry horizontal batch coupling and vertical channel linkage respectively. Double-sided constraint projection is implemented on the same contract residual tensor. First, the common offset of the contract layer is strengthened, and then isolated deviations that are not supported by both graphs are retained. This provides a direct structural basis for separating the contract collaborative offset from local persistent anomalies in the future.

[0124] The contractual collaborative offset component is used to characterize the synchronous offset of multiple batches and multiple residual channels caused by changes in the same pledge contract, such as unified inventory delay, unified accounting lag, unified price adjustment, unified credit reduction, or other contract layers. This offset reflects the common changes in the contract layers and is not directly used as a source of local anomalies. After deducting the corresponding contractual collaborative offset component from the contract residual tensor, the remaining part is used to characterize the local deviation of batches that deviate from the common changes in the contract layers, thereby providing an input basis for the subsequent generation of local anomaly matrix and the calculation of the proportion of continuous anomalies.

[0125] Extract business description vectors, including inventory age percentage. Abnormal dwell frequency Environmental alarm frequency and the proportion of circulation fluctuations ;

[0126] The above-mentioned inventory age ratio refers to the actual inventory age divided by the sum of the preset inventory age threshold and the smallest positive number;

[0127] The above-mentioned abnormal retention frequency refers to the number of times abnormal retention records are used divided by the length of the contract continuous monitoring window;

[0128] The above-mentioned environmental alarm frequency refers to the number of times environmental alarms are used divided by the contract continuous monitoring window length;

[0129] The formula for the above-mentioned turnover fluctuation ratio is:

[0130] ;

[0131] in, For the quantity received into the warehouse, For the quantity shipped out, The amount that can be pledged is consistent with the batch.

[0132] For any two pledge batches, if the material code, specification grade, and storage area are all the same, the business distance is calculated based on the business description vector. If any one of the material code, specification grade, or storage area is different, the business distance is defined as infinite, and a business coupling weight is generated. The formula is as follows:

[0133] ;

[0134] ;

[0135] in, For pledge batch and pledge batches Business distance, For business coupling weights;

[0136] A business coupling matrix is ​​constructed based on the business coupling weights, and then transformed to obtain the batch-normalized coupling matrix, as shown in the formula:

[0137] ;

[0138] ;

[0139] in, For business coupling matrix, For batch normalized coupling matrix, This is a degree matrix, where the diagonal elements are the sum of business coupling weights, and the off-diagonal elements are 0. For batches;

[0140] To illustrate the business linkage relationship between different residual channels, a residual channel coupling diagram is then constructed within the window;

[0141] Expand the contract residual tensor according to time and batch, and then expand the first batch of the contract residual tensor. All data from each residual channel is defined as a vector. ;

[0142] For any two residual channels and The Pearson correlation coefficient is calculated using the Pearson correlation coefficient formula for the vectors, and the absolute value of the Pearson correlation coefficient is taken to obtain the channel coupling weight. Construct the residual channel coupling matrix Then, the transformation is performed to obtain the channel normalization matrix;

[0143] Batch business diagrams represent the horizontal coupling between batches of the same material, specification, and region, while residual channel diagrams represent the vertical linkage between inventory, documents, fulfillment, and financing support. These two types of coupling relationships have different origins and functions in actual business and cannot be represented by a single diagram. Using two diagrams to represent the two types of business relationships separately is more in line with the actual scenario. Changes that truly belong to contract collaborative offsets usually meet two characteristics at the same time, including consistency between adjacent batches and linkage between business linkage channels. By first creating business coupling in two diagrams, these collaborative changes can be enhanced in advance, while isolated deviations that are not supported by batch adjacency relationships and channel linkage relationships are retained for easy separation later.

[0144] Extract time slices from the contract residual tensor to obtain the corresponding batch channel residual slice matrix. The rows of the matrix are batches, the columns are residual channels, and the matrix elements are the current sampling period, the effective pledge batch, the sampling period within the window, and the residual channel.

[0145] Matrix multiplication is performed on the batch normalized coupling matrix, the channel normalized matrix, and the batch channel residual slice matrix to obtain the coupled batch channel residual slice matrix. The coupled batch channel residual slice matrix is ​​then reorganized in ascending order of sampling period to obtain the contract residual tensor after the dual-graph service coupling, which is defined as the coupled residual tensor.

[0146] The matrix multiplication mentioned above refers to processing the batch channel residual slice matrix corresponding to each sampling period within the contract continuous monitoring window separately, first multiplying it on the left with the batch normalized coupling matrix, and then multiplying it on the right with the channel normalized matrix;

[0147] In inventory pledge risk control scenarios, if multiple similar batches under the same contract shift synchronously in inventory, documents and financing support channels, such changes are more likely to be contract-level coordinated shifts. However, if a batch deviates from its adjacent batches, or a channel changes abruptly without being accompanied by responses from other channels, it is more likely to be a local anomaly. Coupling the business between the two graphs first can lay a direct foundation for separating coordinated shifts into local anomalies in the future.

[0148] Expanding the coupled residual tensor along the time and channel directions yields the congruent residual matrix, where rows are batches and every 9 columns correspond to 9 residual channels;

[0149] Singular value decomposition is performed on the contract residual matrix to obtain a diagonal matrix, a left singular vector, and a right singular vector.

[0150] Extract the singular values ​​from the diagonal of the diagonal matrix, extract the median of the singular values, calculate the median absolute deviation of the singular values, retain the corresponding patterns of the singular values ​​that meet the conditions, and generate the contractual collaborative offset matrix. The formula is as follows:

[0151] ;

[0152] ;

[0153] ;

[0154] in, The median of the absolute deviations of the singular values. It is a singular value. The median, For the maximum singular value, It is the minimum singular value. The number of singular values. For the first A singular value, For contractual coordination offset matrix, To preserve the diagonal matrix formed by the singular values ​​that satisfy the conditions, and Let them be left singular vectors and right singular vectors. For transpose, The number of modes to retain;

[0155] The contract collaborative offset matrix represents the dominant offset structure carried by multiple batches, multiple sampling periods, and multiple channels. This type of offset is usually not a single batch anomaly, but more like a synchronization deviation at the contract level, such as unified inventory delay, unified accounting lag, unified price adjustment lag, etc. The determined retention mode of the median of singular values ​​and the median of the absolute deviation of singular values ​​is adopted instead of manually setting the rank value in advance, because the number of batches, window length and fluctuation intensity of different contracts are different. The number of collaborative offset modes should not be fixed and must be directly determined by the current contract window data.

[0156] If the contract coordination offset matrix is ​​directly output as the source of risk, it is easy to misjudge the synchronous offset formed by multiple batches of pledged goods under the same pledge contract within the same sampling period due to the influence of the same business conditions as local anomalies. The sequential processing method of first extracting the contract coordination offset and then performing local anomaly analysis on the remaining residuals ensures that the anomaly results retained later no longer correspond to the common changes at the contract level, but correspond to the batch local anomalies that deviate from the common changes at the contract level.

[0157] By extracting common offsets first and then local anomalies, we can avoid misjudging uniform fluctuations at the contract level as batch-level risks and avoid local spike anomalies being diluted by average in the common pattern, thus making the results of local anomalies closer to the real source of risk.

[0158] In supply chain pledge risk control, the truly high-risk aspects are often not the instantaneous spikes in a single sampling period, but rather the anomalies that continuously occur in the same batch within consecutive sampling periods. After extracting the contract collaborative offset matrix, the parts of the remaining residuals that truly deviate from the common change patterns of the contracts are extracted, and the anomalies with time persistence are further retained to obtain the local persistent anomaly parts.

[0159] S4. Separate the contract collaborative offset component from the contract residual tensor, generate a local anomaly matrix, calculate the local anomaly magnitude and the proportion of persistent anomalies, construct a dynamic residual state vector based on the product of the two, and generate a dynamic comprehensive risk value.

[0160] Specifically, the contractual collaborative offset components are separated from the contractual residual tensor to generate a local anomaly matrix, including:

[0161] Subtracting the contractual collaborative offset matrix from the contractual residual matrix yields the residual matrix. For any column of the residual matrix, the median of that column is extracted, and the absolute deviation of the median is calculated. A local anomaly matrix is ​​then defined. The formula is:

[0162] ;

[0163] ;

[0164] in, This represents the absolute deviation of the median. For column indexes, The remaining residual matrix is ​​the first Line 1 Column elements, For the number of valid pledge batches, The median, For the local anomaly matrix, the first... Line 1 The elements of the column.

[0165] Furthermore, the magnitude of local anomalies and the proportion of persistent anomalies are calculated, and a dynamic residual state vector is constructed based on their product, including:

[0166] Define column numbers for the 9 types of batch residual channels, with the starting column number being... The terminating column number is The element values ​​of the local anomaly matrix are extracted according to the column number of the 9 batch residual channels, and the absolute value of the element values ​​is taken to obtain the local anomaly amplitude.

[0167] Abnormal amplitudes in the current sampling period alone are insufficient to reflect dynamic risks. In real-world scenarios, deviations in a single sampling period may originate from system synchronization delays, document entry time differences, inventory count timing differences, or market data refresh lags, and do not necessarily constitute a persistent risk.

[0168] Calculate the continuous duration of the 9 types of anomalies within the window and convert it into the percentage of continuous anomalies to indicate whether the anomalies are continuous;

[0169] The duration of a persistent anomaly is defined by the following formula:

[0170] ;

[0171] in, For the duration of the anomaly, The number of sampling periods during which the continuous anomaly persists. For the first Local abnormal amplitude of the class;

[0172] The percentage of continuous anomalies is obtained by dividing the length of the continuous anomaly by the length of the contract continuous monitoring window.

[0173] The key to persistent anomalies is not how big the anomaly is, but how long the anomaly lasts. The proportion of the duration to the window length can directly reflect the extent to which this type of anomaly occupies the entire continuous monitoring window.

[0174] A contract may contain multiple valid pledge batches, and dynamic risk control focuses more on the most significant and unfavorable abnormal state in the current contract. Therefore, this step does not simply average all batches, but uses the product of the abnormal amplitude and the continuous proportion, and takes the maximum value in the contract to form a 9-dimensional consistent dynamic residual state vector.

[0175] For all valid pledged batches, the product of the local anomaly amplitude and the proportion of continuous anomalies in each of the nine batch residual channels is calculated. The maximum product is selected and set as the dynamic residual state quantity. The nine dynamic residual state vectors are spliced ​​together to obtain the dynamic residual state vector.

[0176] Dynamic risk control focuses more on the most significant and persistent abnormal batches in the contract, rather than the dilution result after overall averaging. Multiplying the abnormal amplitude by the persistence ratio can automatically weaken single peaks of abnormalities that are not persistent, and also prevent abnormalities that are long-lasting but have very small amplitudes from being over-amplified.

[0177] Furthermore, a dynamic composite risk value is generated, including:

[0178] The dynamic state quantity of insufficient payment in the residual state vector is directly defined as the performance continuity risk value, and the dynamic state quantity of data anomaly in the residual state vector is directly defined as the data reliability risk value.

[0179] Based on the residual state vector, a set of risk values ​​is constructed, including the true risk value, the performance continuity risk value, the data reliability risk value, and the price coverage risk value, as shown in the formula:

[0180] ;

[0181] ;

[0182] in, This represents the true risk value. This is the sum of the dynamic state quantities corresponding to book inventory, physical inventory, circulating inventory, document coverage, and logistics anomalies in the residual state vector. To cover the risk value in terms of price, The sum of the value-covering dynamic state quantity and the credit-occupying dynamic state quantity;

[0183] If the authenticity risk deteriorates significantly in any aspect, such as large inventory discrepancies or persistent logistical anomalies, it may directly affect the authenticity and controllability of the pledged assets. The root mean square formula is more sensitive to larger outliers and is more in line with the characteristic that the weakness of authenticity risk determines the overall situation.

[0184] The aforementioned true risk value represents the overall deviation of the inventory book, physical count, circulation, document and logistics status from the consistent inventory status, and whether these deviations are in a state of continuous increase. The performance continuity risk value represents the state of insufficient payment and credit occupancy. The price coverage risk value represents whether the financing balance can be effectively covered by the current contract consistent and secure pledge value. The data reliability risk value represents whether the data source itself participating in the risk control calculation has continuous missing and abnormalities.

[0185] All four types of risk values ​​are input only as the dynamic residual state vector generated in the preceding steps. They do not rely on manual assignment of levels to a single business event, thus ensuring that the calculation chain of the dynamic comprehensive risk value is closed and that the output source is unique.

[0186] Since the dynamic residual state vector is derived from the residual after deducting the contract collaborative offset component, and is formed by combining the local abnormal amplitude and the proportion of continuous abnormality, the dynamic comprehensive risk value does not directly reflect the common changes in the contract layer itself, but reflects the comprehensive deviation of abnormal states that deviate from the common changes in the contract layer and persist in time in terms of inventory, performance, price coverage and data status.

[0187] After the two-stage compression of the aforementioned nine types of residuals and four types of risks, the dimensions of each risk value have been unified into dimensionless state values, and each type of risk has been internally amplified or retained in the previous step. The risk value set is summed and averaged to obtain the dynamic comprehensive risk value.

[0188] The dynamic comprehensive risk value is classified into risk levels, and dynamic risk control actions are set. The formula is as follows:

[0189] ;

[0190] in, Risk level, , as well as As a risk threshold, satisfying , This is a dynamic comprehensive risk value;

[0191] The aforementioned risk thresholds are set based on a quantile threshold method. Supply chain inventory pledge risk is typically not a sudden, single-point problem, but rather accumulates gradually with factors such as inventory anomalies, logistics delays, price declines, and delayed payments. Therefore, using distribution quantiles to classify risk levels better reflects the gradual evolution of risk. For example, by obtaining the N most recent historical pledge business samples that have been settled, released from pledge, or had their risks resolved through an API interface (N, for example, 100), their dynamic comprehensive risk values ​​are extracted and sorted in ascending order to obtain a historical risk value sequence. Percentiles are then extracted from this historical risk value sequence; for example, the 50th, 75th, and 90th percentiles are set as... , as well as If the number of historical pledge business samples is less than N, then all historical pledge business samples are obtained. This can avoid mismatches caused by fixed thresholds under different types of goods, different contract sizes and different operating stages. It can also ensure that the same risk control system maintains a consistent risk output scale in different business scenarios, reducing false alarms and false negatives.

[0192] When the number of historical pledge transactions that have been settled, released from pledge, or disposed of through risk disposal is zero, , as well as The values ​​were set to 0.25, 0.50, and 0.75 respectively. After obtaining historical pledge business samples, the values ​​were then analyzed according to the aforementioned 50th percentile, 75th percentile, and 90th percentile. , as well as Update;

[0193] The aforementioned dynamic risk control actions are control instructions automatically generated by the risk control system based on the risk level. For example, when the risk level is 1, routine monitoring is carried out; when the risk level is 2, the sampling frequency is increased, manual review is conducted, and inventory is reviewed; when the risk level is 3, new pledges are frozen, the pledge discount rate is increased, and on-site verification is initiated; when the risk level is 4, withdrawals are suspended, additional collateral is added, and the contingency plan is activated.

[0194] First, a multi-caliber inventory observation set is constructed, including book inventory, physical inventory, inventory calculated based on turnover, and inventory supported by documents. Then, batch-consistent inventory quantities are obtained under validity identification constraints. Next, a contract residual tensor is constructed based on nine types of batch residuals. Using batch business graphs and residual channel graphs, contract-level collaborative offsets and local persistent anomalies are separated. The amplitude of local anomalies and the proportion of persistent anomalies are compressed into dynamic residual state quantities, further forming four types of risk values ​​and dynamic comprehensive risk values. This avoids misidentifying single spikes as persistent risks and avoids misidentifying common fluctuations in the contract layer as local anomalies, thereby enhancing the stability, interpretability, and practical executability of risk identification results under different data source distortion conditions.

[0195] This embodiment also provides a dynamic risk control system for supply chain inventory pledging based on multi-source data fusion, including:

[0196] The batch-consistent pledgeable quantity determination module is used to extract the book inventory quantity, physical inventory quantity, inventory observations calculated from circulation, and inventory observations supported by documents. It constructs an effective inventory observation set and performs robust processing. Combining the batch status coefficient, the upper limit of financing support quantity, and the effective pledge batch screening results, it determines the batch-consistent pledgeable quantity.

[0197] The contract continuous monitoring window and nine batch residual channel construction module are used to determine the contract continuous monitoring window based on the sampling period corresponding to the most recent valid inventory count, and to construct the following residuals within the window: book inventory deviation residual, inventory count inventory deviation residual, circulating inventory deviation residual, insufficient document coverage residual, logistics abnormal residual, insufficient payment fulfillment residual, value coverage residual, credit occupancy residual, and data abnormal residual.

[0198] The batch business graph and residual channel graph construction module is used to construct a batch business graph based on the business description vector of the pledged batch, and to construct a residual channel graph based on the correlation between the nine types of batch residual channels;

[0199] The Contract Collaborative Offset and Local Persistent Anomaly Separation Module is used to perform dual-graph business coupling of batch business graph and residual channel graph on the contract residual tensor, extract the contract layer collaborative offset part, and separate the local anomaly and persistent anomaly parts from the remaining residual.

[0200] The module for constructing dynamic residual state vectors and risk value sets is used to form dynamic residual state vectors based on the magnitude of local anomalies and the proportion of persistent anomalies, and to construct risk value sets.

[0201] The Dynamic Comprehensive Risk Value and Risk Level Handling Module is used to merge the risk value set to obtain a dynamic comprehensive risk value, classify risk levels, and generate dynamic risk control actions.

[0202] In summary, this invention performs time-series alignment and joint processing on warehousing data, logistics data, transaction data, and financing data under a unified sampling period. It determines the batch-consistent collateralizable quantity at the batch level, constructs a contract residual tensor at the contract level, and applies bilateral constraints to the contract residual tensor by combining the batch business graph and the residual channel graph. First, it extracts the contract collaborative offset component, then generates a local anomaly matrix from the remaining residual after deducting the contract collaborative offset component. Combining the local anomaly amplitude and the proportion of persistent anomalies, it constructs a dynamic residual state vector. This ensures that the dynamic comprehensive risk value reflects anomalies that deviate from the common changes at the contract level and persist over time. This avoids misidentifying synchronous fluctuations at the contract level as local batch anomalies and also prevents persistent local batch anomalies from being submerged in overall contract fluctuations, thus improving the hierarchy, accuracy, and interpretability of the risk identification results.

[0203] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic risk control method for supply chain inventory pledge based on multi-source data fusion, characterized by: include: By acquiring warehousing data, transaction data, logistics data, and financing data that are time-aligned under the same business identifier, a pledge detection unit can be constructed. Based on the effective inventory observation set and financing support constraints in the pledge detection unit, the batch-consistent pledgeable quantity is determined, a contract continuous monitoring window is set, and a batch residual channel is constructed based on the batch-consistent pledgeable quantity. Within the contract continuous monitoring window, the batch residual channel is organized into a contract residual tensor according to the sampling time sequence. The batch residual channels include book inventory deviation residual, physical inventory deviation residual, circulating inventory deviation residual, insufficient document coverage residual, logistics abnormal residual, insufficient payment fulfillment residual, value coverage residual, credit occupancy residual, and data abnormal residual. A batch business graph is constructed based on the business characteristics in the pledge detection unit, a residual channel graph is constructed based on the batch residual channel, a two-sided constraint projection is applied to the contract residual tensor, and the contract collaborative offset component is extracted. Separate the contract collaborative offset component from the contract residual tensor to generate a local anomaly matrix, calculate the local anomaly magnitude and the proportion of persistent anomalies, and construct a dynamic residual state vector based on the product of the two to generate a dynamic comprehensive risk value.

2. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 1, characterized in that: The step of performing a two-sided constrained projection on the contract residual tensor to extract the contractual cooperative offset component includes: For each valid pledge batch of the same contract, extract the business description vector, calculate the business distance based on any two business description vectors, calculate the business coupling weight based on the business distance, construct the business coupling matrix and perform normalization processing to obtain the batch normalized coupling matrix. The contract residual tensor is expanded in terms of time and batch direction. The values ​​of the residual channels within the window are organized into vectors. The absolute value of the Pearson correlation coefficient between any two residual channel vectors is calculated to obtain the channel coupling weight. The channel coupling matrix is ​​constructed and normalized to obtain the channel normalization matrix. The contract residual tensor is used to extract the batch channel residual slice matrix according to the sampling period. The batch channel residual slice matrix is ​​then multiplied by the batch normalized coupling matrix and then multiplied by the channel normalized matrix to obtain the coupled batch channel residual slice matrix. The matrix is ​​then reassembled according to the sampling period and expanded along the time direction to obtain the contract residual matrix. Singular value decomposition is performed on the contract residual matrix to obtain a diagonal matrix, a left singular vector, and a right singular vector. Based on the diagonal matrix, the median of the singular values ​​and the median of the absolute deviation of the singular values ​​are calculated. The singular values ​​are then filtered, and the contract collaborative offset matrix is ​​reconstructed by combining the left and right singular vectors.

3. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 2, characterized in that: The step of separating the contractual collaborative offset components from the contractual residual tensor to generate a local anomaly matrix includes: Subtract the contractual collaborative offset matrix from the contractual residual matrix to obtain the remaining residual matrix. Perform column processing on the remaining residual matrix to generate the local anomaly matrix. The column processing includes, for any column, extracting the elements of each row of the column, calculating the absolute deviation of the median, and retaining the element if the absolute value of the difference between the element and the median of the column is greater than the corresponding absolute deviation of the median; otherwise, setting the element to zero.

4. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 3, characterized in that: The calculation of local anomaly amplitude and persistent anomaly proportion, and the construction of a dynamic residual state vector based on their product, includes: For the batch residual channels in the local anomaly matrix, the absolute values ​​of the element values ​​are extracted in channel order to obtain the local anomaly amplitude. Within the contract continuous monitoring window, the duration of the batch residual channel is counted along the time direction, and the duration is divided by the contract continuous monitoring window length to obtain the percentage of continuous anomalies in the batch residual channel. Multiply the local anomaly amplitude of each channel by the corresponding proportion of continuous anomalies to obtain the dynamic state variables of the batch residual channels, and then splice the first and last values ​​to obtain the dynamic residual state vector.

5. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 4, characterized in that: The generation of the dynamic comprehensive risk value includes: Based on the dynamic residual state vector, the data reliability risk value, authenticity risk value, price coverage risk value, and performance continuity risk value are calculated to construct a risk value set; The risk value set is summed and averaged to obtain a dynamic comprehensive risk value. The dynamic comprehensive risk value is compared with a preset risk threshold to obtain a risk level, and corresponding dynamic risk control actions are generated according to the risk level.

6. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 5, characterized in that: The construction of batch residual channels, which organizes the batch residual channels into contract residual tensors according to the sampling time sequence within the contract continuous monitoring window, includes: Set up a continuous contract monitoring window and construct a batch residual channel based on the batch-consistent pledgeable quantity; Based on the time dimension, batch dimension, and residual channel dimension, the batch residual channels of each sampling period within the contract continuous monitoring window are expanded and arranged to generate the contract residual tensor.

7. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 6, characterized in that: The determination of batch-consistent pledgeable quantity based on the effective inventory observation set and financing support constraints in the pledge detection unit includes: Extract inventory observations from the pledge detection unit, including book inventory quantity, physical inventory quantity, inventory observations calculated from turnover, and document-supported inventory observations. The inventory observations are filtered to obtain the effective inventory observation set. The arithmetic mean of the effective inventory observation set is then calculated to obtain the batch-consistent inventory quantity. By comparing the batch-consistent inventory quantity with the document-supported inventory observation quantity, extracting the minimum value between the two, and applying financing support constraints, the batch-consistent pledgeable quantity is obtained.

8. The supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in claim 7, characterized in that: The process of acquiring warehousing data, transaction data, logistics data, and financing data under the same business identifier and constructing a pledge detection unit through time-aligned data includes: Multiple sources of raw data are acquired in units of sampling period and arranged horizontally to construct a pledge detection unit for that sampling period; The multi-source raw data includes warehousing data, logistics data, transaction data, and financing data.

9. A supply chain inventory pledge dynamic risk control system based on multi-source data fusion, based on the supply chain inventory pledge dynamic risk control method based on multi-source data fusion as described in any one of claims 1 to 8, characterized in that: include: Module for determining the amount of collateral that can be pledged in a consistent batch; Contract continuous monitoring window and nine batch residual channel construction module; Batch business diagram and residual channel diagram construction module; Contract collaboration offset and local persistent anomaly separation module; Module for constructing dynamic residual state vectors and risk value sets; Dynamic integrated risk value and risk level handling module.

Images