Big data mining method and system applied to supply chain financial business

Abstract

The invention provides a big data mining method and system applied to supply chain financial businesses, and the method comprises the steps: firstly obtaining real-time heterogeneous data streams of a plurality of participants of a target supply chain network, covering structured orders, unstructured logistics texts and semi-structured settlement document data, carrying out the multi-modal feature fusion, and carrying out the real-time heterogeneous data streams, including the structured orders, the unstructured logistics texts and the semi-structured settlement document data, of a plurality of participants of the target supply chain network; the method comprises the following steps: generating a supply chain feature matrix containing static and dynamic features of entity nodes, constructing a supply chain time sequence diagram network based on the supply chain feature matrix, representing participants by nodes, representing transaction events with timestamps and transaction strength weights by edges, hierarchically aggregating features by using a pre-trained time sequence diagram convolutional network, extracting global transaction modes and local abnormal fluctuation features, and constructing a supply chain network; and finally, generating a supply chain risk conduction topological graph which comprises a risk propagation path and an intensity parameter, and triggering a real-time risk early warning signal so as to realize accurate insight and timely early warning of the supply chain financial risk.

Description

Technical Field

[0001] This application relates to the field of big data mining technology. Specifically, it relates to a big data mining method and system applied to supply chain finance business. Background Art

[0002] In the field of supply chain management, with the in-depth development of global trade and the rapid progress of information technology, supply chain finance business has gradually become a key link connecting capital flow, logistics, and information flow. Traditional supply chain finance risk assessment mainly relies on manual review, static financial statement analysis, and limited transaction history data. These methods are unable to cope effectively with the complex and changing supply chain environment. Especially in the big data era, the amount of data generated in the supply chain network has increased explosively, and the data types have become increasingly diverse. How to effectively integrate and utilize these massive and heterogeneous data resources to more accurately identify potential risks in the supply chain has become an urgent problem to be solved.

[0003] In the prior art, although there have been some attempts to apply big data technology to supply chain finance risk management, most of them focus on the analysis of single data types or simple data sets. For example, some methods only mine structured data, ignoring the rich information contained in unstructured and semi-structured data; some consider multiple data types, but lack effective data fusion means, resulting in insufficient information utilization and one-sided risk assessment results. In addition, existing risk assessment models often focus on static analysis and are difficult to capture the dynamic changes in the supply chain transaction process, especially the transaction pattern characteristics and local abnormal fluctuations across time windows, which limits the timeliness and accuracy of risk warning.

[0004] In terms of graph network analysis, although there have been studies attempting to construct supply chain network models to analyze the relationships between participants, most of these models ignore the timestamps of transaction events and the transaction intensity weights, and are unable to accurately reflect the temporal dynamics and transaction complexity of the supply chain network. Therefore, there are still great limitations in the prior art in identifying risk transmission paths and triggering warning signals. Summary of the Invention

[0005] In view of the problems mentioned above, in combination with the first aspect of this application, embodiments of this application provide a big data mining method applied to supply chain finance business. The method includes: Obtain real-time heterogeneous data streams of multiple participants in the target supply chain network, where the real-time heterogeneous data streams include structured order data, unstructured logistics text data, and semi-structured settlement document data; Perform multimodal feature fusion on the real-time heterogeneous data stream to generate a supply chain feature matrix in a unified coding format, where the supply chain feature matrix includes static attribute features and dynamic transaction behavior features of entity nodes; Based on the static attribute features and dynamic transaction behavior features in the supply chain feature matrix, construct a supply chain time-series graph network, where the nodes in the supply chain time-series graph network represent participating entity parties, and the edges represent transaction events, and the edges include timestamps and transaction intensity weights; Perform hierarchical feature aggregation on the supply chain time-series graph network through a pre-trained time-series graph convolutional network to extract global transaction pattern features and local abnormal fluctuation features across time windows; Generate a supply chain risk conduction topology graph according to the global transaction pattern features and local abnormal fluctuation features, where the supply chain risk conduction topology graph includes risk propagation paths and path strength parameters between nodes, and trigger real-time risk warning signals based on the path strength parameters.

[0006] On the other hand, an embodiment of the present application also provides a supply chain finance business system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, in the embodiment of the present application, by obtaining the real-time heterogeneous data streams of multiple participating parties in the target supply chain network, covering structured order data, unstructured logistics text data, and semi-structured settlement document data, a supply chain feature matrix in a unified coding format is generated, thereby organically combining the static attribute features and dynamic transaction behavior features of entity nodes, and potential relationships hidden between different data modalities can be mined. Then, a supply chain time-series graph network constructed based on the supply chain feature matrix, with nodes representing participating entity parties and edges representing transaction events, and timestamps and transaction intensity weights are assigned to the edges, can intuitively and accurately depict the dynamic transaction relationships among the participating parties in the supply chain over time. Next, use a pre-trained time-series graph convolutional network to perform hierarchical feature aggregation on the supply chain time-series graph network to extract global transaction pattern features and local abnormal fluctuation features across time windows, which can not only grasp the overall transaction pattern of the supply chain from a macro level, but also capture local subtle abnormal fluctuations, realizing multi-scale and accurate analysis of supply chain risks. Finally, generate a supply chain risk conduction topology graph including risk propagation paths and path strength parameters between nodes according to the extracted features, and trigger real-time risk warning signals based on the path strength parameters, thereby visualizing and quantifying complex supply chain risks, enabling decision-makers to quickly and clearly understand the risk propagation context and take targeted measures in a timely manner. Description of the Drawings

[0008] Figure 1 It is a schematic execution flow diagram of the big data mining method applied to the supply chain finance business provided by an embodiment of the present application.

[0009] Figure 2 It is a schematic hardware architecture diagram of the supply chain finance business system provided by an embodiment of the present application. Specific Embodiments

[0010] The present application will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 It is a schematic flow diagram of the big data mining method applied to the supply chain finance business provided by an embodiment of the present application. The big data mining method applied to the supply chain finance business will be introduced in detail below.

[0011] Step S110, obtain real-time heterogeneous data streams of multiple parties in the target supply chain network, where the real-time heterogeneous data streams include structured order data, unstructured logistics text data, and semi-structured settlement document data.

[0012] In this embodiment, taking the supply chain network of an electronic product as an example, it includes multiple parties such as manufacturers, suppliers, distributors, retailers, and logistics providers.

[0013] For manufacturers, the structured order data may include order information for ordering various raw materials (such as chips, circuit boards, etc.) from suppliers. For example, an order details that the model of the ordered chips is A123, the quantity is 1000 pieces, the order amount is 50,000 yuan, and the delivery cycle is 30 days. These order data are stored in tabular form, and each field has a clear definition and format, thus constituting structured order data.

[0014] Regarding the unstructured logistics text data, the logistics provider will generate some text records during the transportation process. For example, when transporting a batch of electronic products, the driver may record "Encountered heavy rain during transportation, the goods may be delayed in arrival. Currently, the transportation vehicle is a van, and the vehicle number is 12345". Information such as "heavy rain", "goods delayed in arrival", and "van" are key information extracted from the unstructured logistics text, and these texts have no fixed format and the content is relatively free.

[0015] Semi-structured settlement document data can be invoices or receipts provided by suppliers to manufacturers. For example, in an invoice, in addition to containing clear settlement amounts (such as 50,000 yuan) and payment methods (such as bank transfer), there may also be some nested tags, like the validity identification of the invoice (represented by a specific code or stamp), and information such as the recent change frequency of this settlement account (such as 3 fund in-and-outs in the past month). Such information has a certain structure (such as clear fields like amount and payment method), and there are also some nested or less regular parts (such as validity identification and account change frequency), thus constituting semi-structured settlement document data. Thus, the above various real-time heterogeneous data streams are obtained from these different types of participants.

[0016] Step S120, perform multi-modal feature fusion on the real-time heterogeneous data stream to generate a supply chain feature matrix in a unified coding format, where the supply chain feature matrix contains static attribute features and dynamic transaction behavior features of entity nodes.

[0017] Continuing with the example of the electronics product supply chain. For structured order data, perform field parsing to extract features. For example, from an order from a manufacturer to a supplier, the order amount of 50,000 yuan reflects the scale of the transaction, the delivery cycle of 30 days reflects information related to time cost, and the product category of chips is related to the product types in the supply chain. The above information is combined to generate a first-dimensional feature vector, such as [50000, 30, chip type code].

[0018] For unstructured logistics text data, perform keyword extraction and semantic role annotation. In the previously mentioned logistics record "Encountered heavy rain during transportation, the goods may be delayed in arrival. Currently, the transportation vehicle is a van, and the vehicle number is 12345", extract "heavy rain" as the description of the logistics delay event, "van" as the type of transportation tool, and then according to the enterprise's internal abnormal status coding system (assuming that heavy rain corresponds to the code 1001), generate a second-dimensional feature vector, such as [heavy rain description, van, 1001].

[0019] For semi-structured settlement document data, perform nested tag parsing. Information such as the change frequency of the settlement account in the invoice (such as 3 fund in-and-outs in the past month), the payment method (bank transfer), and the bill validity identification (assuming the stamp validity identification is 1) are combined to generate a third-dimensional feature vector, such as [3, bank transfer, 1].

[0020] Then, the feature vectors in the above three dimensions are concatenated tensorially to obtain a high-dimensional feature tensor. Since there may be feature redundancy in this feature tensor, for example, there may be some information that is repeated in different dimensions or has too high a correlation, through dimensionality reduction processing (such as algorithms like principal component analysis), these redundant features are eliminated, and finally a supply chain feature matrix in a unified coding format is generated. This matrix contains both static attribute features of entity nodes such as manufacturers and suppliers (such as static information like enterprise scale and industry affiliation can be reflected indirectly from order data and settlement document data), and dynamic transaction behavior features (such as the frequency of orders and the delay situation of logistics).

[0021] Step S130: Based on the static attribute features and dynamic transaction behavior features in the supply chain feature matrix, construct a supply chain time-series graph network. The nodes in the supply chain time-series graph network represent participating entity parties, and the edges represent transaction events. The edges contain timestamps and transaction intensity weights.

[0022] In the scenario of the electronic product supply chain, extract information from the dynamic transaction behavior features. For example, in the transaction between a manufacturer and a supplier, the trading counterparty identifier of the manufacturer is the enterprise identifier of the supplier, the transaction timestamp records the specific time of each order placement (such as May 10, 2023), and the transaction amount is 50,000 yuan.

[0023] Based on the industry category in the static attribute features (for example, the manufacturer belongs to the electronic manufacturing industry and the supplier belongs to the electronic component supply industry) and the historical credit rating (assuming the credit rating of the manufacturer is A and that of the supplier is B), the initial association weight between participating entity parties can be calculated. For example, according to the internal algorithm of the enterprise, the initial association weight of enterprises in the same industry with similar credit ratings may be relatively high, while that of enterprises in different industries or with a large credit rating gap may be relatively low.

[0024] Taking 1 day as the preset time granularity interval, divide the transaction timestamps into consecutive time windows. Within each time window, combine the initial association weight and the transaction amount to calculate the comprehensive transaction intensity weight between trading counterparties. For example, within a certain day, the manufacturer placed a large order with the supplier. Since they are in the same industry and have relatively close credit ratings, the initial association weight is relatively high, and coupled with the large order amount, the comprehensive transaction intensity weight between them on this day will be very high.

[0025] Based on these comprehensive trading intensity weights, weighted trading edges are generated, and the trading edges of the same participating entity (such as a manufacturer) in different time windows (different dates) are connected in a time series. For example, if the manufacturer has a transaction with a supplier on May 10th and another transaction on May 15th, these two trading edges are connected in chronological order to form a supply chain time-series graph network with static attribute associations. In this supply chain time-series graph network, the nodes include participating entities such as manufacturers, suppliers, distributors, retailers, and logistics providers, and the edges represent the trading events between them, and the time stamps (the time when the transaction occurs) and trading intensity weights (reflecting the importance or frequency of the transaction, etc.) are marked on the edges.

[0026] Step S140, perform hierarchical feature aggregation on the supply chain time-series graph network through a pre-trained time-series graph convolutional network to extract global trading pattern features and local abnormal fluctuation features across time windows.

[0027] In the time-series graph network of the electronic product supply chain, for the initial layer operation of the time-series graph convolutional network. Within each time window (such as every day), neighborhood node feature sampling is performed on the trading edges. For example, for the node of the manufacturer, its neighborhood nodes may be suppliers, logistics providers, etc., and the features of these neighborhood nodes (such as the supply capacity of the supplier, the transportation efficiency of the logistics provider, etc.) are sampled to extract the local trading pattern of the current time window (the current day). For example, if the manufacturer smoothly obtains raw materials from the supplier on the current day and the logistics transportation is not delayed, this constitutes a local trading pattern.

[0028] In the middle layer, the local trading patterns of adjacent time windows (such as consecutive days) are spliced using a sliding window. Assume that in this embodiment, the time window from May 10th to May 15th is considered. The local trading patterns of each day are spliced together to generate trading trend features in the time dimension. For example, if the raw material supply of the manufacturer gradually increases and the logistics transportation efficiency is stable within these 6 days, this is a trading trend feature.

[0029] In the final layer, attention weight allocation is performed on the trading trend features in the time dimension. If there is a situation where the supplier delays delivery on one of these 6 days, this day is regarded as an abnormal time window. Through the attention mechanism, a higher weight is given to this abnormal time window to highlight its feature contribution degree, and the abnormal time window feature is obtained.

[0030] Then, these local transaction patterns, transaction trend features, and abnormal time window features are fused through a gating mechanism. For example, obtain the time window feature vector of the local transaction pattern (including the transaction-related feature vectors of each day), the sliding window concatenated tensor of the transaction trend feature (including the feature tensor concatenated over 6 days), and the attention weight vector of the abnormal time window feature (highlighting the weight vector of the abnormal days). Initialize the trainable gating parameter set of the gating fusion unit (including the feature selection weight matrix and the time decay coefficient vector). Concatenate the time window feature vector and the sliding window concatenated tensor across dimensions to generate a time-expanded feature block, and perform an outer product operation between the attention weight vector and the time-expanded feature block to generate a joint feature tensor. Input the joint feature tensor into the gating fusion unit, perform a linear transformation on the joint feature tensor through the feature selection weight matrix to generate a candidate gating signal basis, and use the time decay coefficient vector to scale the candidate gating signal basis in the time dimension to generate a normalized gating signal. Perform a non-linear activation process on the normalized gating signal to generate a feature selection gating value and a time decay gating value, and weight the joint feature tensor according to the feature selection gating value in the channel dimension, and at the same time adjust the feature intensity between time windows of the weighted channel features according to the time decay gating value to generate a dynamically fused feature tensor. Decompose the dynamically fused feature tensor into low-frequency components and high-frequency components, perform average pooling on the low-frequency components along the time dimension to generate global transaction pattern features (reflecting the overall transaction pattern, such as the overall stability of supply and transportation over a period of time), and perform peak detection within the time window on the high-frequency components to generate local abnormal fluctuation features (such as abnormal situations like sudden supply interruptions or transportation delays).

[0031] Step S150: Generate a supply chain risk conduction topology map based on the global transaction pattern features and the local abnormal fluctuation features. The supply chain risk conduction topology map includes the risk propagation paths and path strength parameters between nodes, and triggers a real-time risk warning signal based on the path strength parameters.

[0032] Based on the electronic product supply chain scenario, extract the transaction event time series of the participating entity from the global transaction pattern features. For example, information such as the transaction sequence and time interval between the manufacturer and suppliers, logistics providers, etc. over a period of time. At the same time, identify a set of risk source nodes with abnormal transaction intensity fluctuations based on the local abnormal fluctuation features. Assume that within a period of time, the supplier has experienced multiple delivery delays and unstable supply capacity, then the supplier is identified as a risk source node.

[0033] According to the node identifiers in the risk source node set (the enterprise identifiers of suppliers) and the time series of transaction events, traverse the set of adjacent nodes (such as manufacturers, logistics providers, etc.) in the supply chain time-series graph network that are associated with transaction events with the risk source nodes, and generate a set of candidate risk propagation paths. For example, the supply path from the supplier to the manufacturer, the goods handover path from the supplier to the logistics provider, etc. are all candidate risk propagation paths.

[0034] Based on the transaction intensity weights on each candidate risk propagation path in the set of candidate risk propagation paths (such as the weight determined according to the past transaction amount and frequency, etc. on the supply path from the supplier to the manufacturer) and the local abnormal fluctuation characteristics of the corresponding time window (such as the abnormal fluctuation situation during the time period corresponding to the supplier's delivery delay), calculate the correlation degree between the fluctuation variance of the transaction intensity weights on each candidate risk propagation path and the abnormal characteristics of the adjacent time window. For example, if during the period of the supplier's delivery delay, the transaction intensity weight between it and the manufacturer fluctuates greatly and is highly correlated with the abnormal characteristic of the delivery delay.

[0035] According to the weighted summation result of the correlation degree between the fluctuation variance and the abnormal characteristics, generate the path strength parameter for each candidate risk propagation path. Suppose the path strength parameter of the supply path from the supplier to the manufacturer is calculated to be 0.8. If the preset propagation threshold is 0.5, then this path is retained. Eliminate the candidate risk propagation paths whose path strength parameters are lower than the preset propagation threshold.

[0036] Arrange the candidate risk propagation paths in the retained set of candidate risk propagation paths in descending order according to the path strength parameters (such as the path strength parameter of the path from the supplier to the manufacturer is 0.8, and the path strength parameter of the path from the supplier to the logistics provider is 0.6), and topologically connect the overlapping nodes (such as the supplier) between the candidate risk propagation paths to generate a supply chain risk conduction topology graph containing multi-level cascaded risk propagation paths. For example, the supply problem of the supplier may further affect the distributor through the manufacturer, forming a multi-level cascaded risk propagation path.

[0037] According to the real-time transaction status data of each node in the supply chain risk conduction topology graph, verify the node connectivity of the multi-level cascaded risk propagation paths. For example, whether the current production status of the manufacturer is affected by the supply problem of the supplier. If there is an impact, dynamically correct the path strength parameter based on the verification result (such as adjusting the path strength parameter of the path from the supplier to the manufacturer to 0.9).

[0038] Overlay and map the dynamically corrected path strength parameters with the node attribute features of the supply chain time series diagram network (such as the production capacity of manufacturers, the transportation capacity of logistics providers, etc.) to generate a version update record of the supply chain risk conduction topology diagram with arrows indicating the direction of risk propagation. For example, clearly identify the arrows indicating the direction of risk propagation from suppliers to manufacturers in the diagram and record the changes in the path strength parameters.

[0039] Adjust the feature aggregation weight coefficients of the time series diagram convolutional network based on the time series change trend of the path strength parameters in the version update record. If it is found that a certain path strength parameter continues to rise, it may be necessary to increase the aggregation weight for the features related to that path. At the same time, synchronously update the node risk exposure value labels of the supply chain risk conduction topology diagram (for example, due to supply problems from suppliers, the risk exposure value of the manufacturer changes from low to high).

[0040] When it is detected that the path strength parameter exceeds the preset risk threshold (assumed to be 0.9), obtain the real-time transaction data of the participating entity on the corresponding risk propagation path (such as suppliers and manufacturers). For example, obtain the current inventory situation of suppliers, the current raw material demand situation of manufacturers, etc.

[0041] Extract incremental features from the real-time transaction data, such as extracting new features such as the recent urgent demand degree for raw materials from the manufacturer's production plan, and match the results of incremental feature extraction with the feature patterns of historical risk events. Assume that there was a historical risk event of production interruption at the manufacturer due to insufficient supply from the supplier, and compare the current features (such as tight inventory of suppliers, urgent demand for raw materials by manufacturers) with the feature patterns of historical risk events to obtain the similarity matching result.

[0042] Adjust the weight allocation of the path strength parameters according to the similarity matching result. If the similarity is very high, the weight of the path strength parameter may be further increased. And generate a warning signal including the risk level (such as high risk), the impact range (such as may cause the entire production chain to interrupt) and the coping strategy (such as finding alternative suppliers).

[0043] Push the warning signal to the terminal device corresponding to the target participating entity, triggering interface visual warning (such as a red warning box popping up on the production management system interface of the manufacturer) and automated transaction interception instructions (such as pausing further payment to the supplier until the supply problem is resolved).

[0044] Based on the above steps, in the embodiment of the present application, real-time heterogeneous data streams of multiple parties in the target supply chain network are obtained, including structured order data, unstructured logistics text data, and semi-structured settlement document data. Thereby, a supply chain feature matrix with a unified coding format is generated, so as to organically combine the static attribute features and dynamic transaction behavior features of entity nodes, and potential relationships hidden between different data modalities can be mined. Then, based on the supply chain feature matrix, a supply chain time-series graph network is constructed, with nodes representing participating entity parties, edges representing transaction events, and time stamps and transaction intensity weights assigned to the edges, which can intuitively and accurately depict the dynamic transaction relationships among various participating parties in the supply chain over time. Next, the pre-trained time-series graph convolutional network is used to perform hierarchical feature aggregation on the supply chain time-series graph network, extracting global transaction pattern features and local abnormal fluctuation features across time windows, which can not only grasp the overall transaction pattern of the supply chain from a macroscopic level but also capture local subtle abnormal fluctuations, realizing multi-scale and accurate analysis of supply chain risks. Finally, based on the extracted features, a supply chain risk conduction topology graph including the risk propagation paths between nodes and the path strength parameters is generated, and real-time risk warning signals are triggered based on the path strength parameters, thereby visualizing and quantifying complex supply chain risks, enabling decision-makers to quickly and clearly understand the risk propagation context and take targeted measures in a timely manner.

[0045] In a possible implementation manner, step S120 includes: Step S121, perform field parsing on the structured order data, extract features such as order amount, delivery cycle, and product category, and generate a first-dimensional feature vector.

[0046] In this embodiment, from the aspect of structured order data, the order information between the manufacturer and the supplier constitutes the structured order data. For example, the manufacturer orders chips for producing smart phones from the supplier, with an order amount of 100,000 yuan, a delivery cycle set to 45 days, and the product category being a specific model of high-end chips. By performing field parsing on such structured order data and extracting these key features of order amount, delivery cycle, and product category according to the established algorithms and data structure requirements, a first-dimensional feature vector is generated. This first-dimensional feature vector may be represented in a specific format, such as [100,000, 45, high-end chip model code], which can accurately reflect important order-related information in subsequent analyses and provide basic information about transaction scale, time limit, and product type for the construction of the entire supply chain features.

[0047] Step S122, perform keyword extraction and semantic role annotation on the unstructured logistics text data, extract logistics delay event descriptions, transportation tool types, and abnormal status codes, and generate a second-dimensional feature vector.

[0048] For unstructured logistics text data, logistics providers generate a large number of text records during the transportation process. When transporting chips required by a manufacturer, the logistics text records may contain information such as: "Due to road construction, the transport vehicle was delayed on the way, and it is expected that the chips will arrive 3 days late. The transport vehicle is a refrigerated truck. According to the enterprise's internal logistics exception coding system, the corresponding code for this delay is 2001". When processing such unstructured logistics text data, through keyword extraction and semantic role annotation techniques. First, "road construction" is identified as a description of the logistics delay event, which can intuitively reflect the abnormal situation and its reasons in the logistics transportation process; "refrigerated truck" is determined as the type of transport vehicle, and different types of transport vehicles may affect the transportation conditions, costs, and transportation efficiency of goods; "2001" is used as the exception status code, which is an identifier for the enterprise to uniformly manage logistics exceptions. Based on the results of these extractions and annotations, a second-dimensional feature vector is generated, such as [road construction - delayed 3 days, refrigerated truck, 2001]. This second-dimensional feature vector brings key information about the logistics link to the supply chain feature matrix, including the logistics delay situation that may affect the production plan and information related to the transport vehicle.

[0049] Step S123: Parse the nested tags of the semi-structured settlement document data, extract the settlement account change frequency, payment method distribution, and bill validity identifier, and generate a third-dimensional feature vector.

[0050] For example, the invoice issued by a supplier to a manufacturer is a type of semi-structured settlement document. In addition to conventional information such as the clear amount (e.g., 100,000 yuan), the invoice also contains information related to some nested tags. For example, the settlement account has 5 fund changes in the past three months, the payment method is electronic transfer, and there are specific stamps and codes on the invoice indicating the bill validity. By parsing the nested tags of these semi-structured settlement document data, features such as the settlement account change frequency, payment method distribution, and bill validity identifier can be extracted, and a third-dimensional feature vector is generated, such as [5, electronic transfer, valid identifier code]. This third-dimensional feature vector reflects the relevant characteristics of the settlement link, including information such as the stability of capital flow and the effectiveness of settlement.

[0051] Step S124: Tensor splice the first-dimensional feature vector, the second-dimensional feature vector, and the third-dimensional feature vector, and after eliminating feature redundancy through dimensionality reduction processing, generate the supply chain feature matrix in the unified coding format.

[0052] Finally, the first-dimensional feature vectors, second-dimensional feature vectors, and third-dimensional feature vectors respectively generated from the structured order data, unstructured logistics text data, and semi-structured settlement document data are tensor concatenated to form a high-dimensional tensor containing more comprehensive information. However, there may be a problem of feature redundancy in this high-dimensional tensor. For example, in some cases, there may be a certain correlation between the order amount and the change frequency of the settlement account, and this correlation may lead to overemphasis or repeated calculation of some features in subsequent analysis. To solve this problem, dimensionality reduction processing techniques such as the principal component analysis algorithm are adopted. By re-evaluating and combining the features in the tensor, those redundant features are eliminated, and finally a supply chain feature matrix in a unified coding format is generated. This supply chain feature matrix comprehensively and refinedly contains the static attribute features of entity nodes (such as some inherent attributes of an enterprise can be reflected from data such as orders and settlement documents) and dynamic transaction behavior features (such as the transaction process of an order, the transportation situation of logistics, etc.).

[0053] In a possible implementation manner, step S130 includes: Step S131, extracting the counterparty identifier, transaction timestamp, and transaction amount of the participating entity from the dynamic transaction behavior features.

[0054] For example, in the transaction interaction between a manufacturer and a supplier, the manufacturer can clearly identify its counterparty identifier, which is the unique enterprise identifier of the supplier. Each transaction has an exact transaction timestamp. For example, on June 15, 2023, the manufacturer placed a chip purchase order with the supplier, and this date constitutes the transaction timestamp. At the same time, the transaction amount of this order is 100,000 yuan. These information accurately describe the basic elements of each transaction and are important data points for constructing the supply chain network relationship.

[0055] Step S132, calculating the initial association weight between the participating entities based on the industry category and historical credit rating in the static attribute features.

[0056] In this electronic product supply chain, the manufacturer belongs to the electronic manufacturing industry, and the supplier belongs to the electronic component supply industry. They are in the upstream and downstream relationship of the same industrial chain. The historical credit rating of the manufacturer is A-level, indicating that it has a good credit record in past transactions; the historical credit rating of the supplier is B-level, although it is also a relatively good rating, there is a certain gap compared with the manufacturer. According to the association weight calculation algorithm formulated within the enterprise, since they are in the same industry and the credit ratings are relatively close, the calculated initial association weight between them is 0.6. This initial association weight reflects the basic association degree between the two parties based on industry attributes and historical credit without considering the specific transaction situation.

[0057] Step S133: Divide the transaction timestamp into consecutive time windows at a preset time granularity interval, and calculate the comprehensive transaction intensity weight between counterparty entities within each time window by combining the initial association weight and the transaction amount.

[0058] Taking 1 week as the preset time granularity interval, divide the transaction timestamp into consecutive time windows. Within each time window, calculate the comprehensive transaction intensity weight between counterparty entities by combining the initial association weight and the transaction amount. For example, within a certain week in June 2023, the manufacturer placed a chip order worth 100,000 yuan with the supplier. Considering the previously calculated initial association weight of 0.6, according to the calculation formula for the comprehensive transaction intensity weight (which may be a complex formula considering factors such as the association weight, the proportion of the transaction amount in the total business volume, etc.), the calculated comprehensive transaction intensity weight between the manufacturer and the supplier within this week is 0.8. This comprehensive transaction intensity weight accurately reflects the closeness and importance of the transaction between the two parties within a specific time window.

[0059] Step S134: Generate weighted transaction edges based on the comprehensive transaction intensity weight, and perform time-series connection on the transaction edges of the same participating entity in different time windows to form a supply chain time-series graph network with static attribute associations.

[0060] In the supply chain time-series graph network, taking the manufacturer and the supplier as an example, the transaction edge between them is assigned a weight of 0.8, indicating the importance degree of this transaction relationship within this time window. Then, perform time-series connection on the transaction edges of the same participating entity in different time windows. For example, the manufacturer had a transaction with the supplier in the first week of June, and the weight of the transaction edge was 0.8; and there was another transaction in the third week of June, with the weight of the transaction edge being 0.7. Connect these two transaction edges in different time windows in chronological order to form a supply chain time-series graph network with static attribute associations. In this network, each participating entity (such as manufacturers, suppliers, distributors, retailers, and logistics providers, etc.) serves as a node, and their transaction relationships are connected by weighted transaction edges, and these edges contain the timestamp information of the transaction occurrence, thus comprehensively reflecting the transaction relationships between all links in the entire electronic product supply chain and their changes over time.

[0061] In a possible implementation manner, step S140 includes: Step S141: At the initial layer of the time-series graph convolutional network, sample the neighborhood node features of the transaction edges within each time window to extract the local transaction pattern of the current time window.

[0062] For example, in the initial layer of the temporal graph convolutional network, taking the supply chain temporal graph network composed of participants such as manufacturers, suppliers, distributors, retailers, and logistics providers as an example, the transaction edges within each time window reflect the transaction relationships among the participants during a specific period. For each time window, neighborhood node feature sampling is performed to extract the local transaction patterns of the current time window. For example, considering a one-week time window, within a certain week, there are transaction edges between the manufacturer node and the supplier and the logistics provider. Looking at the transaction edge from the manufacturer to the supplier, it may involve features such as the supply quantity and supply price of chips; looking at the transaction edge from the manufacturer to the logistics provider, it may involve features such as transportation cost and transportation time. For the supplier node, its neighborhood nodes may include raw material producers, etc., involving features such as the stability of raw material supply and raw material price fluctuations. Sampling the relevant features of these neighborhood nodes and aggregating them obtain the local transaction pattern for that week. For example, within this week, the manufacturer received sufficient and on-time chip supplies from the supplier, and the logistics transportation proceeded normally. These transaction-related situations together constitute the local transaction pattern of the current time window, which reflects the local state of transactions among all links in the supply chain within this week.

[0063] Step S142, in the middle layer, splice the local transaction patterns of adjacent time windows using a sliding window to generate transaction trend features in the time dimension.

[0064] Suppose we consider a time window of 5 consecutive weeks. In the first week, the manufacturer had sufficient raw material supply, normal production, and the products were successfully delivered to the distributor; in the second week, there were slight quality fluctuations in the raw materials supplied by the supplier but it did not affect production; in the third week, the logistics transportation was delayed by one day due to weather but it did not cause a major impact; in the fourth week, the production efficiency increased and the product delivery was advanced; in the fifth week, all links were operating normally. Splice the local transaction patterns of these 5 weeks in chronological order using a sliding window to form a spliced tensor containing the transaction features of multiple time windows. This spliced tensor can reflect the transaction trend of the entire supply chain over a relatively long period. For example, the overall trend from raw material supply, production to product delivery is basically stable but there are some small fluctuations, such as supply quality fluctuations and transportation delays. These information are all integrated into the transaction trend features in the time dimension.

[0065] Step S143, in the final layer, assign attention weights to the transaction trend features in the time dimension to obtain abnormal time window features to highlight the feature contribution degree of abnormal time windows.

[0066] During the aforementioned five - week period, the raw material quality fluctuations in the second week and the transportation delays in the third week are abnormal situations. Through the attention mechanism, higher weights are assigned to these abnormal time windows. For example, the raw material quality fluctuations in the second week may affect quality inspection and other subsequent production processes, and the transportation delays in the third week may lead to a temporary shortage of inventory for distributors. These abnormal situations are of great significance in the transaction trends of the entire supply chain. Through the assignment of attention weights, the characteristics of these abnormal time windows can be more prominently reflected in subsequent analyses, enhancing the attention to abnormal situations, and thus obtaining the characteristics of abnormal time windows.

[0067] Step S144, fuse the local transaction pattern, transaction trend characteristics, and abnormal time window characteristics through a gating mechanism to generate the global transaction pattern characteristics and local abnormal fluctuation characteristics.

[0068] In a possible implementation manner, step S144 includes: Step S1441, obtain the time - window feature vector of the local transaction pattern, the sliding - window concatenated tensor of the transaction trend characteristics, and the attention - weight vector of the abnormal time window characteristics, and initialize the trainable gating parameter set of the gating fusion unit. The trainable gating parameter set includes a feature - selection weight matrix and a time - decay coefficient vector.

[0069] The time - window feature vector of the local transaction pattern may contain specific feature values of the transactions of each participant within each week; the sliding - window concatenated tensor of the transaction trend characteristics contains the tensor information after integrating the transaction characteristics of each week within five consecutive weeks; the attention - weight vector of the abnormal time window characteristics clarifies the importance weights of abnormal weeks such as the second week and the third week. At the same time, initialize the trainable gating parameter set of the gating fusion unit, which includes a feature - selection weight matrix and a time - decay coefficient vector. These parameters are given default values initially. For example, the initial value of the feature - selection weight matrix may be set based on past experience or predefined rules, and the initial value of the time - decay coefficient vector is also set according to the preliminary estimation of the impact of time factors in the supply chain.

[0070] Step S1442, perform cross - dimensional concatenation on the time - window feature vector and the sliding - window concatenated tensor to generate a time - extended feature block, and perform an outer - product operation between the attention - weight vector and the time - extended feature block to generate a joint feature tensor.

[0071] For example, the specific transaction feature values of each week are concatenated with the tensor information integrated over 5 weeks, so that the new feature block contains not only the detailed information of a single time window, but also the trend information within a longer time range. Then, the attention weight vector and the time-expanded feature block are subjected to a matrix outer product operation to generate a joint feature tensor, which synthesizes multi-faceted information such as the importance weight of the abnormal time window, the specific transaction features of a single time window, and the transaction trend features over a long time.

[0072] Step S1443: Input the joint feature tensor into a gated fusion unit, perform a linear transformation on the joint feature tensor through the feature selection weight matrix to generate a candidate gated signal basis, and use the time decay coefficient vector to scale the candidate gated signal basis in the time dimension to generate a normalized gated signal.

[0073] For example, according to the weight coefficients in the feature selection weight matrix, perform a linear combination of the elements in the joint feature tensor to obtain a new tensor as the candidate gated signal basis. Then use the time decay coefficient vector to scale the candidate gated signal basis in the time dimension to generate a normalized gated signal. This step takes into account the influence of time factors on different transaction features, and adjusts the proportional relationship of the candidate gated signal basis in the time dimension through the time decay coefficient vector, so that the generated gated signal is more in line with the actual situation of the transaction features changing with time in the supply chain.

[0074] Step S1444: Perform a non-linear activation process on the normalized gated signal to generate a feature selection gated value and a time decay gated value, and perform channel weighting on the joint feature tensor according to the feature selection gated value, and at the same time adjust the feature intensity between time windows of the weighted channel features according to the time decay gated value to generate a dynamic fusion feature tensor.

[0075] For example, by using a non-linear activation function such as ReLU, convert the normalized gated signal into a feature selection gated value and a time decay gated value. Perform channel weighting on the joint feature tensor according to the feature selection gated value, which means adjusting the weights of different channels (which can be understood as different types of transaction features) in the joint feature tensor according to the gated value to highlight some important channel features. At the same time, adjust the feature intensity between time windows of the weighted channel features according to the time decay gated value. For example, for the features of an earlier time window, if the time decay gated value is large, its intensity in the overall features may be reduced, thus generating a dynamic fusion feature tensor.

[0076] Step S1445: Decompose the dynamic fusion feature tensor into low-frequency components and high-frequency components. Perform average pooling on the low-frequency components along the time dimension to generate global transaction pattern features, and perform peak detection within a time window on the high-frequency components to generate local abnormal fluctuation features.

[0077] The low-frequency components represent relatively stable transaction features over a long time range, such as the overall supply level, average transportation time, etc.; the high-frequency components contain fluctuation features within a short time, such as a sudden supply interruption or transportation delay within a certain week. Perform average pooling on the low-frequency components along the time dimension to generate global transaction pattern features. For example, average features such as supply level and transportation time in the low-frequency components over a 5-week time range to obtain a feature that can represent the overall transaction pattern of the supply chain throughout the time period, such as stable overall supply and average transportation time within a certain range. Perform peak detection within a time window on the high-frequency components to generate local abnormal fluctuation features. For example, detect the peaks corresponding to the raw material quality fluctuation in the second week and the transportation delay in the third week in the high-frequency components, and these peaks accurately reflect the local abnormal fluctuation situation.

[0078] Step S1446: According to the decomposition error of the dynamic fusion feature tensor, reversely adjust the feature selection weight matrix and the time decay coefficient vector to generate an updated set of trainable gating parameters to iteratively optimize the gating signal generation process.

[0079] Since there may be errors when decomposing the dynamic fusion feature tensor into low-frequency components and high-frequency components, this error reflects the difference between the decomposition process and the actual transaction features. Through the backpropagation algorithm, adjust the weight coefficients in the feature selection weight matrix and the coefficients in the time decay coefficient vector according to this error. For example, if the decomposition error indicates that some features in the low-frequency components are overemphasized or some fluctuation features in the high-frequency components are missed, then adjust the coefficients in the feature selection weight matrix and the time decay coefficient vector accordingly, so that the next gating signal generation process is more accurate, thereby continuously optimizing the process of fusing features through the gating mechanism, improving the accuracy and effectiveness of the global transaction pattern features and local abnormal fluctuation features, and better reflecting the transaction situation in the electronic product supply chain.

[0080] In a possible implementation manner, step S150 includes: Step S151: Extract the transaction event time series of the participating entity from the global transaction pattern features, and identify a set of risk source nodes with abnormal transaction intensity fluctuations based on the local abnormal fluctuation features.

[0081] In the electronic product supply chain, the characteristics of the global transaction mode reflect the overall transaction situation. For example, from the perspective of the manufacturer, the time series of transaction events may include the chronological order and intervals of events such as placing an order with a supplier at a specific time, arranging product transportation with a logistics provider, and delivering products to a distributor. The local abnormal fluctuation characteristics can highlight some abnormal situations, such as a large fluctuation in the quantity of chips supplied by a supplier within a certain period, or a sudden increase in the transportation cost of a logistics provider. Based on these local abnormal fluctuation characteristics, the supplier with unstable supply or the logistics provider with abnormal transportation costs is identified as a risk source node with abnormal transaction intensity fluctuations. Suppose that within a certain period, Supplier A has a huge fluctuation in the quantity of chip supply due to its own production equipment failure, and Logistics Provider B has a significant increase in transportation costs due to a transportation route adjustment. Then Supplier A and Logistics Provider B form a set of risk source nodes.

[0082] Step S152: According to the node identifiers in the set of risk source nodes and the time series of transaction events, traverse the set of adjacent nodes in the supply chain time series graph network that are associated with the risk source nodes through transaction events, and generate a set of candidate risk propagation paths.

[0083] For example, for the risk source node Supplier A, its node identifier is the unique supplier number. In the supply chain time series graph network, the adjacent node associated with Supplier A through transaction events is the manufacturer. Since the manufacturer depends on Supplier A for chip supply, a candidate risk propagation path is formed from Supplier A to the manufacturer. Similarly, for Logistics Provider B, its adjacent nodes may include the manufacturer and the retailer. Since Logistics Provider B is responsible for transporting the manufacturer's products to the retailer, additional candidate risk propagation paths are formed from Logistics Provider B to the manufacturer and from Logistics Provider B to the retailer. These paths together form a set of candidate risk propagation paths.

[0084] Step S153: Based on the transaction intensity weights on each candidate risk propagation path in the set of candidate risk propagation paths and the local abnormal fluctuation characteristics of the corresponding time window, calculate the correlation degree between the variance of the transaction intensity weights on each candidate risk propagation path and the abnormal characteristics of the adjacent time window.

[0085] Taking the candidate risk propagation path from Supplier A to the manufacturer as an example, the transaction intensity weight may be determined based on factors such as past order amounts and order frequencies. Suppose the previous transaction intensity weight was 0.8. During the period when the supply quantity of Supplier A fluctuated, the transaction intensity weight changed and may have decreased to 0.6. Calculate the variance of this fluctuation in the transaction intensity weight, while considering the correlation degree of abnormal feature in adjacent time windows. For example, the greater the fluctuation in the supply quantity, the more obvious the abnormal supply features in adjacent time windows, and the higher the correlation degree between the two. If the fluctuation in the supply quantity seriously affects the manufacturer's production plan, then this correlation degree will be very high. Through a specific algorithm, calculate by comprehensively considering the variance of the fluctuation and the correlation degree of abnormal features.

[0086] Step S154, generate the path strength parameter for each candidate risk propagation path according to the weighted summation result of the variance of the fluctuation and the correlation degree of abnormal features, and eliminate the candidate risk propagation paths whose path strength parameters are lower than the preset propagation threshold.

[0087] Suppose for the candidate risk propagation path from Supplier A to the manufacturer, the weighted summation result of the variance of the fluctuation and the correlation degree of abnormal features is 0.7, and the set preset propagation threshold is 0.5. Then the path strength parameter of this path is 0.7 and it will be retained. For some other candidate risk propagation paths, if the weighted summation result is lower than 0.5, they will be eliminated. For example, for a certain path from Logistics Provider B to the retailer, the weighted summation result is 0.4. Since it is lower than the preset propagation threshold, this path will not be included in the subsequent analysis.

[0088] Step S155, sort the candidate risk propagation paths in the set of retained candidate risk propagation paths in descending order of the path strength parameter, and topologically connect the overlapping nodes among the candidate risk propagation paths to generate a supply chain risk conduction topology graph containing multi-level cascading risk propagation paths.

[0089] For example, sort the retained candidate risk propagation paths according to the path strength parameter. For example, the path strength parameter of the path from Supplier A to the manufacturer is 0.7, and the path strength parameter of the path from Logistics Provider B to the manufacturer is 0.6. First, consider the path with a high path strength. Since both Supplier A and Logistics Provider B are connected to the manufacturer, the manufacturer is the overlapping node. Topologically connect these paths to construct a supply chain risk conduction topology graph. For example, the supply problem of Supplier A may affect the distributor through the manufacturer, and the transportation cost problem of Logistics Provider B may also affect the manufacturer and then affect the retailer, forming multi-level cascading risk propagation paths.

[0090] Step S156: Based on the real-time transaction status data of each node in the supply chain risk conduction topology graph, verify the node connectivity of the multi-level cascading risk propagation path, and dynamically correct the path strength parameter based on the verification result.

[0091] In the actual operation of the electronic product supply chain, each node has real-time transaction status data. For example, the raw material inventory level and production progress of the manufacturer, the sales data of the retailer, etc. If the manufacturer can cope with the supply fluctuations of Supplier A due to sufficient inventory, then the actual impact of the risk propagation path from Supplier A to the manufacturer may not be as large as previously estimated. According to this situation, dynamically correcting the path strength parameter may adjust the path strength parameter from Supplier A to the manufacturer from 0.7 to 0.6.

[0092] Step S157: Superimpose and map the dynamically corrected path strength parameter with the node attribute characteristics of the supply chain time series graph network to generate a version update record of the supply chain risk conduction topology graph with risk propagation direction arrows.

[0093] For example, the manufacturer has node attribute characteristics such as production capacity and product quality. Superimpose and map the adjusted path strength parameter of 0.6 from Supplier A to the manufacturer with these node attribute characteristics of the manufacturer. For example, if the manufacturer has strong production capacity, it may have a certain buffering effect on risk propagation, which is reflected in the generated version update record as a risk propagation direction arrow pointing from Supplier A to the manufacturer, and at the same time record the change of the path strength parameter and its association with the node attribute characteristics of the manufacturer.

[0094] Step S158: Based on the time series change trend of the path strength parameter in the version update record, adjust the feature aggregation weight coefficient of the time series graph convolutional network, and synchronously update the node risk exposure value label of the supply chain risk conduction topology graph.

[0095] For example, if the path strength parameter from Supplier A to the manufacturer shows a gradually decreasing trend over a period of time, this indicates that the risk propagation ability of this path is weakening. According to this trend, adjust the feature aggregation weight coefficient related to this path in the time series graph convolutional network. For example, if the aggregation weight of the features related to the supply of Supplier A was previously high, it can be appropriately reduced now. At the same time, synchronously update the node risk exposure value label of the supply chain risk conduction topology graph. The manufacturer's risk exposure value was previously high due to the supply fluctuation risk of Supplier A. As the path strength parameter decreases, the manufacturer's risk exposure value can be adjusted from high risk to medium risk, accurately reflecting the risk status of the node in the supply chain.

[0096] In a possible implementation manner, step S150 further includes: Step S159: When it is detected that the path strength parameter exceeds the preset risk threshold, obtain the real-time transaction data of the participating entity on the corresponding risk propagation path.

[0097] For example, in the previously constructed supply chain risk conduction topology map, the path risk threshold from Supplier A to the manufacturer is set to 0.8. At a certain moment, due to the exacerbation of the production equipment failure of Supplier A, its supply capacity further decreases, resulting in the path strength parameter from Supplier A to the manufacturer reaching 0.9, exceeding the preset risk threshold. At this time, it is necessary to obtain the real-time transaction data of Supplier A and the manufacturer on this risk propagation path. For Supplier A, data such as its current chip inventory quantity, the operating status of production equipment, and the number of orders being executed need to be obtained; for the manufacturer, data such as the immediate demand quantity of chips, the time that the chips in inventory can maintain production, and the busyness of the production line need to be obtained.

[0098] Step S1510: Extract incremental features from the real-time transaction data, and perform similarity matching between the incremental feature extraction results and the feature patterns of historical risk events to obtain a similarity matching result.

[0099] For example, from the real-time transaction data of Supplier A, incremental features such as a sharp decline in chip inventory and order delivery delays caused by production equipment failures may be extracted; from the real-time transaction data of the manufacturer, incremental features such as the imminent depletion of chip inventory and the upcoming shutdown of the production line due to lack of materials may be extracted. In the enterprise's risk event database, there are historical risk events such as the manufacturer's production suspension caused by insufficient supply from a supplier. These historical risk events have their specific feature patterns. For example, previously, a supplier's raw material supply interruption led to the manufacturer's production suspension within a specific time. At that time, the inventory consumption speed of the manufacturer, the sequence of production line shutdowns, etc. all constituted the feature pattern. Compare the currently extracted incremental features with these feature patterns of historical risk events in detail. If the inventory decline speed of the current Supplier A and the imminent depletion of the manufacturer's inventory are highly similar to the historical events, the similarity matching result will show a high similarity; if only some features are similar, it is a medium similarity; if there are almost no similarities, it is a low similarity.

[0100] Step S1511: Adjust the weight allocation of the path strength parameter according to the similarity matching result, and generate a warning signal including the risk level, the scope of influence, and the coping strategy.

[0101] If the similarity matching result is a high similarity, indicating that the current situation is extremely similar to a serious risk event in history, then it is necessary to increase the weight allocation of the path strength parameter because the risk is likely to expand rapidly in this case. For example, adjust the path strength parameter from 0.9 to 0.95 for the path from Supplier A to the manufacturer. Based on this adjusted path strength parameter and the current transaction data situation, generate a warning signal. Since the path strength parameter is high and similar to a historical serious risk event, the risk level is determined to be high risk. In terms of the scope of influence, considering that the manufacturer may stop production due to a shortage of chips, which in turn affects the product supply to the distributor and ultimately the sales of the retailer, the scope of influence is from Supplier A to the entire downstream manufacturers, distributors, and retailers. The coping strategy can be to notify the manufacturer to immediately find alternative suppliers and at the same time notify the distributor and retailer to adjust their sales plans to cope with possible product shortages.

[0102] Step S1512, push the warning signal to the terminal device corresponding to the target participating party entity, triggering an interface visual alarm and an automated transaction interception instruction.

[0103] For example, the terminal device of the manufacturer may be the operation interface of the production management system, and the terminal devices of the distributor and retailer may be the interfaces of the inventory management and sales management systems. After pushing the warning signal to the production management system interface of the manufacturer, the system will trigger an interface visual alarm. For example, a red warning box will pop up on the interface, showing "High risk: Severe shortage of chip supply from Supplier A, may lead to production stoppage, affecting the entire supply chain. Please immediately find alternative suppliers." At the same time, for the automated transaction system between the manufacturer and Supplier A, if there are unfinished payment or new order transactions, an automated transaction interception instruction will be triggered to suspend the payment to Supplier A or the placement of new orders until the risk is alleviated or resolved. For the distributor and retailer, their terminal devices will also receive a similar warning signal, triggering an alarm on their inventory management and sales management system interfaces, and they can adjust their inventory management and sales plans according to the coping strategy in the warning signal to reduce the impact of the risk.

[0104] In a possible implementation manner, the method further includes: Step S210, deploy distributed feature collection agents in the supply chain time series diagram network, and the distributed feature collection agents are embedded in the transaction systems of the participating party entities.

[0105] Taking the manufacturer's trading system as an example, feature collection agents are embedded in various subsystems related to supply chain transactions, such as its procurement module, production management module, sales module, etc. For the procurement module, when the manufacturer sends an order request to the supplier, receives the supplier's quotation, confirms the order, etc., the feature collection agent embedded therein can monitor these operations in real time. In the production management module, when events such as production plan adjustment, raw material input into production, and finished product output occur, the feature collection agent can also capture relevant information. In the sales module, operations such as product shipment, sales order processing, and customer payment collection are also within the monitoring scope of the feature collection agent. Similarly, such feature collection agents are also deployed in the trading systems of participating entity parties such as suppliers, distributors, retailers, and logistics providers. For example, operations such as production plan adjustment, raw material procurement, and shipment to the manufacturer by the supplier will be monitored by the feature collection agent in its trading system.

[0106] Step S220, through the feature collection agent, real-time capture the operation logs and data stream change events of the trading system, and generate an incremental feature update request.

[0107] For the manufacturer, when a new chip procurement order is added in the procurement module, changes in information such as the chip model, quantity, and price in the order will be recorded in the operation log, and at the same time, this also constitutes a data stream change event. After the feature collection agent detects these changes, it will generate an incremental feature update request according to the preset rules. The incremental feature update request contains various information changes brought by the new order. For example, an increase in the order amount may affect relevant features such as the transaction intensity weight between the manufacturer and the supplier. On the supplier side, if new raw materials are put into production on the production line, this operation is recorded, which may affect features such as its supply capacity, thus generating an incremental feature update request containing this information.

[0108] Step S230, dynamically update the supply chain feature matrix according to the incremental feature update request, and trigger the online incremental training of the temporal graph convolutional network.

[0109] Step S240, update the path strength parameters of the supply chain risk conduction topology graph based on the network parameters after incremental training.

[0110] Specifically, after the above online incremental training, the parameters of the temporal graph convolutional network are optimized. These optimized parameters will affect the assessment of the supply chain risk conduction. For example, in the relationship between a manufacturer and a supplier, if the chip quality requirements in a new order cause the network to re-evaluate their relationship, it may be found that the supplier's ability to meet the new quality requirements is insufficient, thus increasing the possibility of risk propagation. According to this change, the path strength parameter from the supplier to the manufacturer will be updated accordingly, such as adjusting the original path strength parameter from 0.8 to 0.85 to reflect this new risk situation, so that the supply chain risk conduction topology map can adapt to the new changes in the supply chain in a timely manner and more accurately reflect the risk propagation situation.

[0111] Among them, step S230 includes: Step S231, retaining the intermediate feature representations and gradient distribution data generated during the historical training process.

[0112] When the incremental feature update request of the manufacturer arrives, information such as the new order amount and the new delivery cycle involved will be integrated into the supply chain feature matrix. If the number of chips in the new order increases significantly, it may change the weight relationship between the order amount and the commodity category in the matrix, thus realizing the dynamic update of the supply chain feature matrix. This update will trigger the online incremental training of the temporal graph convolutional network.

[0113] For triggering the online incremental training of the temporal graph convolutional network, first, retain the intermediate feature representations and gradient distribution data generated during the historical training process. During the previous training process, the temporal graph convolutional network has processed a large amount of transaction data and generated intermediate feature representations, which reflect information such as the transaction patterns and relationship strengths between different parties. For example, the long-term transaction pattern between the manufacturer and the supplier is reflected in the intermediate feature representations, and the gradient distribution data records information such as the direction and amplitude of the adjustment of the network parameters during the training process.

[0114] Step S232, comparing the new data stream corresponding to the incremental feature update request with the historical intermediate feature representations to identify the feature distribution offset area.

[0115] For example, assume that some special requirements (such as higher requirements for chip quality) in the manufacturer's new chip purchase order have not appeared in previous transactions. Then, this part of the information will show a feature distribution offset when compared with the historical intermediate feature representations. This offset may be concentrated in the feature area related to chip quality. For example, in the previous intermediate feature representations, the features related to chip quality may only account for a small proportion, but the special requirements of the new order make the importance of this feature prominent, thus identifying this feature distribution offset area.

[0116] Step S233: Insert an adaptive normalization layer in the temporal graph convolutional network to perform local parameter adjustment on the feature distribution offset region.

[0117] For example, for the identified feature distribution offset region related to chip quality, insert an adaptive normalization layer. This adaptive normalization layer can adjust the network parameters related to this region according to the chip quality requirement information in the new data stream. For example, if a new order requires higher chip quality, the adaptive normalization layer may increase the weights of network parameters related to chip quality detection results, supplier quality control capabilities, etc., to better adapt to the new feature distribution.

[0118] Step S234: Use the elastic weight consolidation algorithm to freeze the network parameters of the non-offset region and only update the partial network layers related to the feature distribution offset region.

[0119] For example, in the temporal graph convolutional network, except for the feature distribution offset region related to chip quality, if there are no obvious changes in the feature distributions of other regions such as transaction amount and delivery cycle, use the elastic weight consolidation algorithm to freeze the network parameters of these regions. This can avoid unnecessary interference to the already stable feature regions during the update process, and at the same time only update the partial network layers related to the feature distribution offset region related to chip quality, such as updating the network layer parameters related to chip quality detection and supplier quality assurance measure evaluation, so as to achieve efficient and targeted online incremental training.

[0120] In a possible implementation manner, the method further includes: Step S310: Deploy a risk simulation calculation engine in the supply chain risk conduction topology graph. The risk simulation calculation engine receives the stress test scenario parameters input from the outside.

[0121] For example, the stress test scenario parameters input from the outside may be set as a 30% delay in the raw material supply of the supplier, a 50% increase in the transportation cost of the logistics provider, and a sudden 40% decrease in the market demand of the manufacturer. These stress test scenario parameters are designed to simulate the stress faced by the supply chain in extreme or adverse situations.

[0122] Step S320: Modify the path strength parameters according to the stress test scenario parameters, simulate the risk propagation process of the supply chain network under the preset stress conditions, and record the risk exposure values of each participating entity and the number of system cascade failure nodes during the simulated risk propagation process, and generate a stress test report.

[0123] For example, for the supply path from the supplier to the manufacturer, due to a 30% delay in raw material supply, the path strength parameter (assumed to be 0.8) originally calculated based on normal transactions will be modified according to the preset adjustment rules and may be reduced to 0.5 to reflect the negative impact of the supply delay on the relationship between the two parties. During the simulation of the risk propagation process, due to insufficient raw material supply, the inventory level of the manufacturer continues to decline, resulting in an increased risk of production stagnation, and its risk exposure value rises from a low level under normal circumstances to a relatively high level. The transportation cost of the logistics provider increases by 50%, which may affect the cooperation stability between it and the manufacturer and the retailer, and its own risk exposure value also increases accordingly. If the retailer is unable to meet market demand due to insufficient product supply from the manufacturer, it may lead to a significant decline in the retailer's sales performance and even face difficulties in capital turnover, which may become a system cascade failure node. During the entire simulation process, the risk exposure values of each participating entity and the number of system cascade failure nodes such as the retailer are recorded in detail, and finally a stress test report is generated.

[0124] Step S330, optimize the training strategy of the temporal graph convolutional network and the risk warning threshold configuration based on the stress test report.

[0125] In a possible implementation manner, step S330 includes: Step S331, analyze the data of the simulated risk propagation process recorded in the stress test report, extract the risk exposure values of each participating entity and the number of system cascade failure nodes, generate a set of key stress test indicators, and identify the participating entities with risk exposure values exceeding the preset exposure threshold according to the distribution range of the risk exposure values in the set of key stress test indicators, and generate a list of risk entity identifiers.

[0126] In the stress test report, the risk exposure value of the manufacturer reaches 0.7 (assuming the range of the risk exposure value is 0 - 1) under the pressure of supply delay and demand decline, the risk exposure value of the logistics provider is 0.6, and the risk exposure value of the retailer is 1 because it becomes a system cascade failure node, and the number of system cascade failure nodes is 1. A set of key stress test indicators is generated based on these data. Then, according to the distribution range of the risk exposure values in the set of key stress test indicators, the participating entities with risk exposure values exceeding the preset exposure threshold (assumed to be 0.6) are identified, and a list of risk entity identifiers is generated. In this example, the risk exposure values of the manufacturer, the logistics provider, and the retailer all exceed the preset exposure threshold, so they are all included in the list of risk entity identifiers.

[0127] Step S332: Based on the list of risk entity identifiers, traverse the transaction event time series of each risk entity in the supply chain time series graph network, extract the fluctuation characteristics of the transaction intensity weights of the risk entities under preset stress conditions, align the fluctuation characteristics of the transaction intensity weights with the number of system cascade failure nodes in the stress test key index set in a time window, and generate a fluctuation characteristic sequence with failure node marks.

[0128] Taking the manufacturer as an example, under the stress conditions of supply delay and demand decline, the transaction intensity weight between it and the supplier fluctuates from 0.8 to 0.5, and the transaction intensity weight between it and the logistics provider fluctuates from 0.7 to 0.4. These fluctuation characteristics of the transaction intensity weights are aligned with the number of system cascade failure nodes (1) in the time window. For example, if the supply delay occurs from January to February, the demand decline occurs from February to March, and the retailer becomes a system cascade failure node in March, then the fluctuation characteristics of the transaction intensity weights are associated with the number of system cascade failure nodes in this time sequence to generate a fluctuation characteristic sequence with failure node marks, which can accurately reflect the relationship between the changes in transaction relationships and system failures under stress conditions.

[0129] Step S333: According to the fluctuation characteristic sequence with failure node marks, construct an adversarial training sample set reflecting the correlation between stress conditions and node failures. The adversarial training sample set includes normal transaction mode samples and stress perturbation samples.

[0130] For example, select some transaction data from the normal transaction period of the manufacturer as normal transaction mode samples, such as the transaction data during the period of normal supply and stable demand. At the same time, use the transaction data under stress perturbations such as supply delay and demand decline in the stress test as stress perturbation samples. These samples can comprehensively reflect the state of the supply chain under normal and stress conditions, thus constructing an adversarial training sample set.

[0131] Step S334: Perform incremental adversarial training on the time series graph convolutional network through the adversarial training sample set, and adjust the feature aggregation weight coefficients related to the risk exposure value in the time series graph convolutional network.

[0132] For example, during the incremental adversarial training process, the normal transaction mode samples and stress perturbation samples are alternately input into the time series graph convolutional network. When the stress perturbation sample is input, the network will try to adjust according to the previously learned normal transaction mode to adapt to the stress situation. For example, if the risk exposure value of the manufacturer in the stress perturbation sample increases due to supply delay, the network will adjust the feature aggregation weight coefficients related to the manufacturer's supply, such as increasing the weights of features related to the supply stability of the supplier and the raw material inventory level, to more accurately evaluate the risk.

[0133] Step S335: Extract the updated path strength parameters output by the temporal graph convolutional network after incremental adversarial training, perform regression fitting on the updated path strength parameters and the risk exposure values in the set of stress test key metrics to generate a dynamic mapping relationship between the path strength parameters and the risk exposure values, and based on the dynamic mapping relationship, divide the critical values of the path strength parameters corresponding to different risk exposure value intervals to generate a segmented risk warning threshold configuration table.

[0134] For example, assume that after incremental adversarial training, when the manufacturer has a supply delay, the updated path strength parameter is 0.45 and its risk exposure value is 0.7. By performing regression fitting on multiple such sample data, the functional relationship between the path strength parameter and the risk exposure value is obtained. Based on this relationship, the critical values of the path strength parameters corresponding to different risk exposure value intervals are divided. For example, when the risk exposure value is in the range of 0 - 0.3, the critical value of the path strength parameter is 0.8; when the risk exposure value is in the range of 0.3 - 0.6, the critical value of the path strength parameter is 0.6; when the risk exposure value is in the range of 0.6 - 1, the critical value of the path strength parameter is 0.4. Organize these critical values into a segmented risk warning threshold configuration table.

[0135] Step S336: Based on the critical values in the segmented risk warning threshold configuration table, update the triggering conditions for real-time risk warning signals of the supply chain risk conduction topology graph, and deploy the updated triggering conditions to the risk warning signal generation module.

[0136] For example, in the supply chain risk conduction topology graph, for the path from the manufacturer to the distributor, if the previous risk warning threshold was that a warning was triggered when the path strength parameter was lower than 0.7, according to the new segmented risk warning threshold configuration table, when the risk exposure value is in the range of 0.6 - 1, the critical value of the path strength parameter is 0.4, then the triggering condition is updated to trigger a warning when the path strength parameter is lower than 0.4. Deploy this updated triggering condition to the risk warning signal generation module to make the generation of warning signals more in line with the actual risk situation.

[0137] Step S337: Inject the stress perturbation samples in the adversarial training sample set into the supply chain temporal graph network, verify the suppression effect of the updated temporal graph convolutional network on the number of system cascade failure nodes, generate a network optimization verification result, and based on the suppression effect index in the network optimization verification result, adjust the injection ratio of the stress perturbation samples in the adversarial training sample set to generate a mixed training sample ratio parameter that balances the normal transaction mode and the stress perturbation.

[0138] For example, inject pressure perturbation samples into the supply chain time series graph network and observe the change in the number of system cascade failure nodes under the updated time series graph convolutional network. If after injecting the pressure perturbation samples, the number of system cascade failure nodes decreases from the original 1 to 0.5 (here it is assumed that decimals are allowed to represent the degree of suppression effect), this indicates that the updated network has a certain effect on suppressing system cascade failures. According to this suppression effect index, if the effect is good, the injection ratio of pressure perturbation samples in the adversarial training sample set can be appropriately increased, for example, from the original 30% to 40%, so as to generate a mixed training sample ratio parameter that balances normal transaction patterns and pressure perturbations.

[0139] Step S338: Based on the mixed training sample ratio parameter, retrain the time series graph convolutional network periodically, synchronously update the critical values in the segmented risk warning threshold configuration table, and timestamp bind the updated segmented risk warning threshold configuration table with the version update record of the supply chain risk conduction topology graph to generate a threshold configuration tracking chain with historical version backtracking ability.

[0140] For example, according to the newly generated mixed training sample ratio parameter (such as 60% for normal transaction pattern samples and 40% for pressure perturbation samples), retrain the time series graph convolutional network regularly. During the retraining process, due to the adjustment of the network, the relationship between the path strength parameter and the risk exposure value may change, so the critical values in the segmented risk warning threshold configuration table are synchronously updated. Timestamp bind the updated segmented risk warning threshold configuration table with the version update record of the supply chain risk conduction topology graph. For example, an update was made on August 1, 2023, and this timestamp is bound to the updated configuration table. In this way, a threshold configuration tracking chain with historical version backtracking ability is generated, which is convenient for viewing the threshold configuration under different versions and its impact on risk assessment.

[0141] Step S339: Dynamically adjust the pressure perturbation sample generation strategy of the adversarial training sample set according to the version change trend in the threshold configuration tracking chain to form a training strategy update mechanism with closed-loop feedback optimization.

[0142] For example, if it is found in the threshold configuration tracking chain that as time goes by, after increasing the injection ratio of pressure perturbation samples in a certain stage, the accuracy of risk warning has improved, but overfitting occurs subsequently, then adjust the pressure perturbation sample generation strategy according to this trend. For example, the generation method of pressure perturbation samples can be adjusted, the parameter range or type of pressure perturbation can be changed to avoid overfitting, so as to continuously optimize the training strategy of the time series graph convolutional network and make it better adapt to the risk assessment requirements in the supply chain.

[0143] Figure 2 FIG. Figure 2 shows a hardware structure diagram of a supply chain finance business system 100 for implementing the above-mentioned big data mining method applied to supply chain finance business provided by an embodiment of the present application. As Figure 2 shown in FIG. Figure 2 , the supply chain finance business system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0144] In a possible design, the supply chain finance business system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the supply chain finance business system 100 may be a distributed system). In some embodiments, the supply chain finance business system 100 may be local or remote. For example, the supply chain finance business system 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. Alternatively, the supply chain finance business system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the supply chain finance business system 100 may be implemented on a supply chain finance business system. By way of example only, the supply chain finance business system may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc., or any aggregation thereof.

[0145] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the supply chain finance business system 100 uses to execute or use to complete the exemplary methods described in the present application.

[0146] In a specific implementation process, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, so that the processors 110 may execute the big data mining method applied to supply chain finance business in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected via the bus 130, and the processors 110 may be used to control the transceiver actions of the communication unit 140.

[0147] For the specific implementation process of the processor 110, reference may be made to the respective method embodiments executed by the above supply chain finance business system 100. The implementation principles and technical effects are similar, and will not be elaborated herein in this embodiment.

[0148] In addition, an embodiment of the present application further provides a readable storage medium, in which computer-executable instructions are set. When a processor runs the computer-executable instructions, the above-mentioned big data mining method applied to supply chain finance business is implemented.

[0149] It should be noted that, in order to simplify the description of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof. Similarly, it should be noted that, in order to simplify the description of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof.

Claims

1. A big data mining method applied to supply chain finance business, characterized in that: The method comprises: Acquire real-time heterogeneous data streams of multiple participants in a target supply chain network, wherein the real-time heterogeneous data streams include structured order data, unstructured logistics text data, and semi-structured settlement document data; Perform multimodal feature fusion on the real-time heterogeneous data stream to generate a supply chain feature matrix in a unified coding format, wherein the supply chain feature matrix includes static attribute features and dynamic transaction behavior features of entity nodes; Based on the static attribute characteristics and dynamic transaction behavior characteristics in the supply chain feature matrix, a supply chain timing graph network is constructed, wherein the nodes in the supply chain timing graph network represent the participating entities, the edges represent the transaction events, and the edges contain timestamps and transaction intensity weights; The supply chain time series graph network is hierarchically aggregated through a pre-trained time series graph convolutional network to extract global transaction pattern features and local abnormal fluctuation features across time windows; According to the global transaction pattern characteristics and local abnormal fluctuation characteristics, a supply chain risk transmission topology map is generated. The supply chain risk transmission topology map includes risk propagation paths and path strength parameters between nodes, and a real-time risk warning signal is triggered based on the path strength parameters.

2. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: The multi-modal feature fusion of the real-time heterogeneous data stream to generate a supply chain feature matrix in a unified coding format includes: Performing field analysis on the structured order data, extracting order amount, delivery cycle and commodity category features, and generating a first dimension feature vector; Extract keywords and annotate semantic roles on the unstructured logistics text data, extract logistics delay event descriptions, transportation tool types, and abnormal state codes, and generate a second-dimensional feature vector; Performing nested tag parsing on the semi-structured settlement document data, extracting the settlement account change frequency, payment method distribution and bill validity identifier, and generating a third-dimensional feature vector; The first dimensional feature vector, the second dimensional feature vector and the third dimensional feature vector are tensor-concatenated, and feature redundancy is eliminated through dimensionality reduction processing to generate the supply chain feature matrix in the unified coding format.

3. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: The supply chain timing graph network is constructed based on the static attribute characteristics and dynamic transaction behavior characteristics in the supply chain characteristic matrix, including: Extracting the counterparty identification, transaction occurrence timestamp and transaction amount of the participating entity from the dynamic transaction behavior characteristics; Calculate the initial association weights between the participating entities based on the industry categories and historical credit ratings in the static attribute features; The transaction occurrence timestamp is divided into continuous time windows at intervals of preset time granularity, and the comprehensive transaction intensity weight between the transaction counterparties is calculated by combining the initial association weight and the transaction amount in each time window; The weighted transaction edges are generated according to the comprehensive transaction intensity weights, and the transaction edges of the same participating entity in different time windows are connected in time series to form a supply chain timing graph network with static attribute associations.

4. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: The pre-trained time-series graph convolutional network is used to perform hierarchical feature aggregation on the supply chain time-series graph network to extract global transaction pattern features and local abnormal fluctuation features across time windows, including: In the initial layer of the temporal graph convolutional network, the neighborhood node features are sampled for the transaction edges in each time window to extract the local transaction pattern of the current time window; In the middle layer, the local transaction patterns of adjacent time windows are joined by sliding windows to generate transaction trend features in the time dimension; In the final layer, attention weights are allocated to the transaction trend features in the time dimension to obtain abnormal time window features to highlight the feature contribution of the abnormal time window; The local trading pattern, trading trend characteristics and abnormal time window characteristics are integrated through a gating mechanism to generate the global trading pattern characteristics and local abnormal fluctuation characteristics; The step of fusing the local trading pattern, trading trend characteristics and abnormal time window characteristics through a gating mechanism to generate the global trading pattern characteristics and local abnormal fluctuation characteristics includes: Obtain the time window feature vector of the local transaction pattern, the sliding window splicing tensor of the transaction trend feature, and the attention weight vector of the abnormal time window feature, and initialize the trainable gating parameter set of the gated fusion unit, wherein the trainable gating parameter set includes a feature selection weight matrix and a time attenuation coefficient vector; Cross-dimensionally concatenate the time window feature vector and the sliding window concatenation tensor to generate a time extension feature block, and perform a matrix outer product operation on the attention weight vector and the time extension feature block to generate a joint feature tensor; The joint feature tensor is input into the gating fusion unit, the joint feature tensor is linearly transformed by the feature selection weight matrix to generate a candidate gating signal basis, and the candidate gating signal basis is time-scaled by the time attenuation coefficient vector to generate a normalized gating signal; Performing nonlinear activation processing on the normalized gating signal to generate a feature selection gating value and a time attenuation gating value, and performing channel weighting on the joint feature tensor according to the feature selection gating value, and adjusting the feature intensity between time windows of the weighted channel features according to the time attenuation gating value to generate a dynamic fusion feature tensor; Decomposing the dynamic fusion feature tensor into a low-frequency component and a high-frequency component, performing average pooling processing on the low-frequency component along the time dimension to generate a global transaction pattern feature, and performing peak detection within the time window on the high-frequency component to generate a local abnormal fluctuation feature; According to the decomposition error of the dynamic fusion feature tensor, the feature selection weight matrix and the time attenuation coefficient vector are reversely adjusted to generate an updated set of trainable gating parameters to iteratively optimize the gating signal generation process.

5. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: Generating a supply chain risk transmission topology map according to the global transaction pattern characteristics and the local abnormal fluctuation characteristics includes: Extracting the transaction event time series of the participant entity from the global transaction pattern feature, and identifying the risk source node set with abnormal transaction intensity fluctuation based on the local abnormal fluctuation feature; According to the node identifiers in the risk source node set and the transaction event time series, traverse the set of adjacent nodes in the supply chain timing graph network that are associated with the risk source node and the transaction event, and generate a set of candidate risk propagation paths; Based on the transaction intensity weight on each candidate risk propagation path in the candidate risk propagation path set and the local abnormal fluctuation characteristics of the corresponding time window, the correlation between the fluctuation variance of the transaction intensity weight on each candidate risk propagation path and the abnormal characteristics of the adjacent time window is calculated; Generate a path strength parameter for each candidate risk propagation path based on the weighted sum of the volatility variance and the degree of correlation of the abnormal feature, and eliminate candidate risk propagation paths whose path strength parameters are lower than a preset propagation threshold; The candidate risk propagation paths in the retained candidate risk propagation path set are arranged in descending order according to the path strength parameter, and the overlapping nodes between the candidate risk propagation paths are topologically connected to generate a supply chain risk transmission topology graph containing multi-layer cascade risk propagation paths; Verify the node connectivity of the multi-layer cascade risk transmission path according to the real-time transaction status data of each node in the supply chain risk transmission topology diagram, and dynamically modify the path strength parameter based on the verification result; The dynamically corrected path strength parameter is superimposed and mapped with the node attribute characteristics of the supply chain time series graph network to generate a version update record of the supply chain risk transmission topology graph with risk propagation direction arrows; Based on the temporal change trend of the path strength parameter in the version update record, the feature aggregation weight coefficient of the temporal graph convolutional network is adjusted, and the node risk exposure value label of the supply chain risk transmission topology map is synchronously updated.

6. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: The triggering of a real-time risk warning signal based on the path strength parameter includes: When it is detected that the path strength parameter exceeds a preset risk threshold, real-time transaction data of the participant entities on the corresponding risk propagation path is obtained; Performing incremental feature extraction on the real-time transaction data, and performing similarity matching between the incremental feature extraction result and the feature pattern of the historical risk event to obtain a similarity matching result; Adjust the weight distribution of the path strength parameter according to the similarity matching result, and generate an early warning signal including risk level, impact scope and response strategy; The warning signal is pushed to the terminal device corresponding to the target participant entity, triggering an interface visual alarm and an automated transaction interception instruction.

7. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: The method further comprises: Deploying a distributed feature collection agent in the supply chain timing graph network, wherein the distributed feature collection agent is embedded in a transaction system of a participant entity; The feature collection agent captures the operation log and data flow change events of the transaction system in real time, and generates an incremental feature update request; Dynamically updating the supply chain feature matrix according to the incremental feature update request, and triggering online incremental training of the temporal graph convolutional network; Update the path strength parameters of the supply chain risk transmission topology graph based on the network parameters after incremental training; The triggering of online incremental training of the temporal graph convolutional network includes: Preserve the intermediate feature representation and gradient distribution data generated during the historical training process; Comparing the new data stream corresponding to the incremental feature update request with the historical intermediate feature representation to identify the feature distribution offset area; Inserting an adaptive normalization layer into the temporal graph convolutional network to perform local parameter adjustment on the feature distribution offset region; The elastic weight solidification algorithm is used to freeze the network parameters in the non-shifted area, and only some network layers related to the feature distribution shifted area are updated.

8. The big data mining method applied to supply chain finance business according to claim 1 is characterized in that: The method further comprises: Deploy a risk simulation calculation engine in the supply chain risk transmission topology diagram, wherein the risk simulation calculation engine receives externally input stress test scenario parameters; Modify the path strength parameter according to the stress test scenario parameter, simulate the risk propagation process of the supply chain network under preset stress conditions, record the risk exposure value of each participant entity and the number of system cascading failure nodes in the simulated risk propagation process, and generate a stress test report; The training strategy and risk warning threshold configuration of the timing graph convolutional network are optimized based on the stress test report.

9. The big data mining method applied to supply chain finance business according to claim 8 is characterized in that: The optimizing the training strategy and risk warning threshold configuration of the time series graph convolutional network based on the stress test report includes: Parse the simulated risk propagation process data recorded in the stress test report, extract the risk exposure value of each participant entity and the number of system cascading failure nodes, generate a stress test key indicator set, and identify the participant entities whose risk exposure values ​​exceed the preset exposure threshold according to the risk exposure value distribution interval in the stress test key indicator set, and generate a risk entity identification list; Based on the risk entity identification list, the transaction event time series of each risk entity in the supply chain timing graph network is traversed to extract the transaction intensity weight fluctuation characteristics of the risk entity under preset stress conditions, and the transaction intensity weight fluctuation characteristics are aligned with the number of system cascading failure nodes in the stress test key indicator set in a time window to generate a fluctuation feature sequence with failure node marks; According to the fluctuation feature sequence with failed node marks, an adversarial training sample set reflecting the correlation between pressure conditions and node failure is constructed, wherein the adversarial training sample set includes normal trading mode samples and pressure disturbance samples; Performing incremental adversarial training on the time series graph convolutional network through the adversarial training sample set, and adjusting the feature aggregation weight coefficient related to the risk exposure value in the time series graph convolutional network; Extract the updated path strength parameters output by the time-series graph convolutional network after incremental adversarial training, perform regression fitting on the updated path strength parameters and the risk exposure values ​​in the stress test key indicator set, generate a dynamic mapping relationship between the path strength parameters and the risk exposure values, and divide the path strength parameter critical values ​​corresponding to different risk exposure value intervals according to the dynamic mapping relationship, and generate a segmented risk warning threshold configuration table; Based on the critical value in the segmented risk warning threshold configuration table, updating the real-time risk warning signal triggering condition of the supply chain risk transmission topology diagram, and deploying the updated triggering condition to the risk warning signal generation module; Injecting the stress disturbance samples in the adversarial training sample set into the supply chain timing graph network, verifying the suppression effect of the updated timing graph convolutional network on the number of system cascading failure nodes, generating network optimization verification results, and adjusting the injection ratio of stress disturbance samples in the adversarial training sample set according to the suppression effect index in the network optimization verification results, generating a mixed training sample ratio parameter that balances the normal transaction mode and stress disturbance; Based on the mixed training sample ratio parameter, the time series graph convolutional network is periodically retrained, and the critical value in the segmented risk warning threshold configuration table is synchronously updated, and the updated segmented risk warning threshold configuration table is timestamped and bound to the version update record of the supply chain risk transmission topology map to generate a threshold configuration tracking chain with historical version backtracking capability; According to the version change trend in the threshold configuration tracking chain, the pressure perturbation sample generation strategy of the adversarial training sample set is dynamically adjusted to form a closed-loop feedback optimized training strategy update mechanism.

10. A supply chain financial business system, characterized in that: The supply chain finance business system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to implement the big data mining method applied to supply chain finance business as described in any one of claims 1 to 9 above.

Images