Supply Chain Risk Prediction Method and System Based on Multi-Source Heterogeneous Data Analysis
By analyzing multi-source heterogeneous data, the intensity of risk propagation and resonance effect in the supply chain network are calculated, and a logistic regression model is constructed. This solves the problem of inaccurate linear propagation in supply chain risk prediction and achieves accurate prediction of order fulfillment risks and precise location of nodes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO STATE RES SOFTWARE TECH
- Filing Date
- 2026-07-02
- Publication Date
- 2026-07-31
AI Technical Summary
Existing supply chain risk prediction technologies cannot accurately depict the nonlinear propagation of risks between upstream and downstream nodes, and ignore the homogeneous resonance effect of supplier nodes within the same link and the cumulative resonance across links, leading to an increased probability of order fulfillment failure.
Using a multi-source heterogeneous data analysis method, a supply chain network is constructed by calculating the basic risk value of supplier nodes and the weight of directed edges. Combined with the benchmark risk propagation coefficient, diffusion coefficient and propagation intensity, the transmission risk is calculated, and a logistic regression model is constructed for risk prediction.
It achieves accurate prediction of supply chain order fulfillment risks and precise location of risk nodes, accurately depicts the nonlinear propagation law of risks, captures the homogeneous resonance effect within the same link, and solves the problem of inaccurate risk prediction in existing technologies.
Smart Images
Figure CN122491949A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of supply chain risk analysis technology, and in particular to a supply chain risk prediction method and system based on multi-source heterogeneous data analysis. Background Technology
[0002] During the operation of the supply chain platform, each order corresponds to a supply chain link consisting of multiple fulfillment links and multiple supplier nodes. The transmission of risks between upstream and downstream nodes of the supply chain is not a simple linear superposition, but exhibits complex nonlinear propagation characteristics. At the same time, supplier nodes in the same link are susceptible to the influence of the same industry factors, resulting in the resonance of risks from the same source, which accumulates and amplifies layer by layer along the supply chain link, leading to a significant increase in the probability of order fulfillment failure.
[0003] Existing supply chain risk prediction technologies suffer from several shortcomings: they fail to accurately depict the nonlinear propagation of risks between upstream and downstream nodes, often employing linear propagation methods with fixed coefficients, which are disconnected from the actual characteristics of supply chain risk transmission. Furthermore, they neglect the homogeneous resonance effect of supplier nodes within the same link and the cumulative resonance across links, calculating only the risk of a single node in isolation. This fails to capture the cascading effects of common industry risks and cannot meet the needs of supply chain platforms for accurate prediction and early intervention of order fulfillment risks. To address these issues, this application proposes a supply chain risk prediction method and system based on multi-source heterogeneous data analysis to improve the accuracy of risk prediction. Summary of the Invention
[0004] To address the issues of inaccurate linear propagation and neglect of the same-source resonance effect in existing supply chain risk prediction technologies, this application provides a supply chain risk prediction method and system based on multi-source heterogeneous data analysis.
[0005] Firstly, this application provides a supply chain risk prediction method based on multi-source heterogeneous data analysis, employing the following technical solution: The supply chain risk prediction method based on multi-source heterogeneous data analysis includes: obtaining the multi-link supply chain corresponding to the order, collecting and preprocessing multi-source heterogeneous data, supply transaction data and performance result tags of historical orders; calculating the basic risk value of supplier nodes and the weight of directed edges, and constructing the supply chain network. Following the progressive order from the initial stage to the final stage of the supply chain, perform full-chain risk propagation calculations for each group of adjacent upstream and downstream stages; and calculate the baseline risk propagation coefficient by combining the number of directed edges and the total number of nodes in adjacent stages. For any downstream node, the upstream nodes that are associated with it by a directed edge are taken as associated upstream nodes, and the diffusion coefficient is calculated based on the number of associated upstream nodes. The propagation intensity is calculated by combining the baseline risk propagation coefficient, diffusion coefficient, and directed edge weights between related upstream and downstream nodes; common risks are calculated based on the distribution entropy and mean of the basic risk values of supplier nodes in the upstream link; upstream transmission risks are calculated based on the basic risk values of related upstream nodes, the directed edge weights, and the propagation intensity; and upstream transmission risks and common risks are integrated to obtain the upstream comprehensive risk. Based on the diffusion coefficient, the effective risk of the downstream node is calculated by nonlinear weighted fusion of the upstream comprehensive risk and the basic risk value of the downstream node. Based on the full-link risk propagation calculation results, a training sample set is constructed by extracting fixed-dimensional global risk statistical features, and a logistic regression model is trained. The fixed-dimensional global risk statistical features of the order to be predicted are input into the trained logistic regression model, and the total global supply chain risk of the order to be predicted is output.
[0006] The calculation of the supplier node's basic risk value and directed edge weights includes: Extract the qualification, performance and compliance indicators from the multi-source heterogeneous data, and calculate the basic risk value by weighting and summing the qualification, performance and compliance indicators based on preset weight coefficients; For any pair of adjacent supplier nodes with an actual supply relationship in the supply chain network, the upstream and downstream nodes are determined according to the direction of the directed edge. The actual purchase amount between adjacent nodes is extracted, and the ratio of the actual purchase amount to the total purchase amount of the downstream node is calculated. The ratio is then normalized and used as the weight of the directed edge.
[0007] Optionally, the calculation method for the benchmark risk propagation coefficient includes: The total number of directed edges between adjacent upstream and downstream links is counted. The product of the total number of upstream nodes and the total number of downstream nodes is calculated, and the product is used as the theoretical maximum total number of directed edges. The ratio of the total number of directed edges to the theoretical maximum total number of directed edges is calculated. The ratio is then reverse-mapped to obtain a mapping value. Based on the preset base value of the propagation coefficient and the mapping value, the benchmark risk propagation coefficient is calculated.
[0008] Optionally, the propagation intensity can be calculated using the following methods: Calculate the product of the diffusion coefficient and the directed edge weights between the associated upstream and downstream nodes, and perform a nonlinear mapping on the product to obtain the mapping result; The product of the mapping result and the baseline risk propagation coefficient is used as the propagation strength between the associated upstream and downstream nodes.
[0009] Optionally, the calculation method for the common risks includes: Calculate the mean of the basic risk values of all upstream nodes in the upstream link; calculate the risk distribution entropy of the basic risk values of all upstream nodes in the upstream link; perform a nonlinear mapping on the risk distribution entropy to obtain the mapping value, and use the product of the mapping value and the mean as the common risk of the upstream link.
[0010] Optionally, the calculation method for the upstream comprehensive risk includes: The upstream transmission risk is obtained by summing the products of the basic risk value of each associated upstream node, the directed edge weight between the associated upstream node and the downstream node, and the propagation strength. The sum of the common risks and the upstream transmission risks is calculated, and the sum is nonlinearly constrained and mapped based on an exponential function to obtain the upstream comprehensive risk.
[0011] Optionally, the effective risk can be calculated in the following ways: The first safety margin is obtained by inverse mapping of the basic risk value of the downstream node; the second safety margin is obtained by inverse mapping of the upstream comprehensive risk of the downstream node; the diffusion coefficient is used as an exponential adjustment factor to calculate the weighted product of the first safety margin and the second safety margin to obtain the joint safety margin; the joint safety margin is inversely mapped, and the mapping result is used as the effective risk of the downstream node.
[0012] Optionally, in the end-to-end risk propagation calculation, the effective risk of the downstream node calculated in the current stage is used as the basic risk value of the corresponding node in the new upstream stage in the next round for iterative updates; the construction of the training sample set includes: extracting the statistical features of the effective risk of the entire chain and the statistical features of the effective risk of the end stage, constructing a fixed-dimensional supply chain statistical feature vector, constructing training samples based on the supply chain statistical feature vector of any historical order and its fulfillment result label, and forming a training sample set by combining the training samples of each historical order.
[0013] Optionally, the training logistic regression model includes: The training sample set is divided into a training set and a test set according to a preset ratio; Using the cross-entropy loss function as the optimization objective, gradient descent combined with cross-validation and parameter optimization strategies is used to iteratively train the model's hyperparameters until the preset maximum number of iterations is reached, or the loss value converges and training is terminated.
[0014] Secondly, this application provides a supply chain risk prediction system based on multi-source heterogeneous data analysis, employing the following technical solution: A supply chain risk prediction system based on multi-source heterogeneous data analysis includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the supply chain risk prediction method based on multi-source heterogeneous data analysis described above.
[0015] This application has the following technical advantages: 1. This application aims to solve the problems of inaccurate linear propagation and neglect of the same source resonance effect in supply chain risk prediction technology, and provides a supply chain risk prediction method and system based on multi-source heterogeneous data analysis. Through standardized preprocessing of multi-source heterogeneous data, calculation of nonlinear risk propagation across the entire chain, and construction of interpretable prediction models, it can achieve accurate prediction of supply chain order fulfillment risks and precise location of risk nodes.
[0016] 2. This application quantifies the global risk transmission intensity of adjacent links through a benchmark risk propagation coefficient, and calculates the propagation intensity between nodes by combining the diffusion coefficient and directed edge weights, accurately depicting the nonlinear propagation law of risk; it introduces the calculation of common risks in links to capture the homogeneous resonance effect of suppliers within the same link, solving the defect of existing technologies that ignore common risks; it calculates the effective risk of downstream nodes through a nonlinear fusion formula, realizing the coupling amplification of upstream comprehensive risk and the basic risk of the node itself, accurately reflecting the overall risk level of the node; through iterative calculation of risk across the entire chain, it realizes the layer-by-layer cumulative transmission of risk along the supply chain, ensuring that the risk set of the end links can fully carry the risk information of the entire chain; it extracts fixed-dimensional global risk statistical features to construct a sample set, avoiding the feature misalignment problem caused by the inconsistency of the number and identity of end nodes in different orders. Attached Figure Description
[0017] Figure 1 This is a flowchart of steps S1-S5 in the supply chain risk prediction method based on multi-source heterogeneous data analysis in the embodiments of this application. Detailed Implementation
[0018] This application discloses a supply chain risk prediction method based on multi-source heterogeneous data analysis, referring to... Figure 1 This includes steps S1-S5.
[0019] S1: Based on the supply chain platform, obtain the multi-stage supply chain corresponding to the order, collect multi-source heterogeneous data of multiple historical orders, supply transaction data and performance result tags, perform data preprocessing, calculate the basic risk value of supplier nodes and directed edge weights, and build a supply chain network.
[0020] Each order on the supply chain platform corresponds to a dedicated supply chain, which consists of suppliers and supply relationships for each stage required to complete the material supply, processing, and circulation of the order. For each order, its corresponding dedicated supply chain is used as the target supply chain. Based on the order's Bill of Materials (BOM) trajectory, the target supply chain is divided into progressive stages (i.e., from the initial material supply stage to the final fulfillment stage, each stage is divided sequentially) according to the order of material flow. The supplier set corresponding to each stage is then defined (i.e., all suppliers involved in the material supply for the order included in each stage).
[0021] Furthermore, by combining the actual flow direction of the order materials, a supply chain network corresponding to the target supply chain is constructed. Each supplier in the target supply chain is assigned as a node. Based on the direction of material flow from upstream suppliers to downstream suppliers, directed edges are constructed between corresponding nodes of adjacent links. These directed edges are specifically used to represent the exclusive supply relationship between upstream supplier nodes and downstream supplier nodes, clarifying the direction of material flow and the corresponding entities. At the same time, the set of supply chain nodes corresponding to each link is determined, thus completing the construction of the supply chain network.
[0022] Data from multiple historical orders on the supply chain platform is collected. Based on the supply chain network corresponding to each historical order, the collection scope includes: multi-source heterogeneous data of any node in the supply chain corresponding to each historical order, supply transaction data between adjacent supplier nodes, and the final fulfillment result label of each historical order. Specifically, the qualification information (supplier qualification level), fulfillment data (historical order fulfillment rate, default rate, number of complaints), and industry compliance data (number of industry regulatory penalty records, environmental compliance level) of each supplier node are collected. All of the above data are collected from the historical database of the supply chain platform, comprehensively representing the operational and fulfillment capabilities of the supplier nodes themselves. The actual purchase amount between two adjacent supplier nodes is collected through the bill of materials trajectory of the corresponding historical orders, and the total purchase amount of each supplier node is obtained. The fulfillment results of each historical order are collected through the order fulfillment records of the supply chain platform. If the order completes all material delivery and fulfillment requirements within the preset delivery period, the fulfillment result of the order is recorded as successful; otherwise, the fulfillment result of the order is recorded as unsuccessful.
[0023] For non-numerical data such as qualification levels and environmental compliance levels in multi-source heterogeneous data, numerical conversion is performed, mapping different levels to corresponding values (dividing the range into [0,1] based on the total number of levels, such as five qualification levels, mapped from high to low as 1.0, 0.8, 0.6, 0.4, 0.2); missing values and outliers in multi-source heterogeneous data are handled—missing values are filled with the median of the corresponding dimension data (to avoid the mean being affected by extreme values); outliers are identified using the 3σ principle, and outliers exceeding [μ-3σ, μ+3σ] (μ is the mean of the data in this dimension, and σ is the standard deviation) are truncated to the boundary value of this range to avoid the outlier data causing deviations in subsequent calculations; for all cleaned data (including quantifiable indicators such as performance, operation, and finance), corresponding Min-Max normalization processing is performed to map all values to the [0,1] range, eliminating the influence of different units on subsequent fusion calculations.
[0024] Establish a three-tiered risk assessment system encompassing qualification, performance, and compliance dimensions. The qualification dimension includes one indicator: supplier qualification level. The performance dimension includes three indicators: historical order fulfillment rate, historical default rate, and number of historical complaints. The compliance dimension includes two indicators: the number of industry regulatory penalty records and environmental compliance level. In total, there are six quantitative indicators.
[0025] Furthermore, based on supply chain industry risk management experience, the weight coefficients corresponding to the above six indicators are manually preset, and the sum of the weight coefficients of all indicators is 1. The weight coefficients of the six indicators can be set as follows: qualification level 0.15, historical order fulfillment rate 0.25, historical default rate 0.20, historical complaint count 0.10, number of industry regulatory penalty records 0.15, and environmental compliance level 0.15. The size of the weight coefficient is positively correlated with the degree of influence of the indicator on the supplier's performance risk, that is, the greater the influence of the indicator on the performance risk, the larger its weight coefficient. The normalized data of each dimension of a supplier node is multiplied by the corresponding weight coefficient, and the sum is obtained to obtain the basic risk value of the supplier node. The basic risk value ranges from [0,1] and is specifically used to characterize the supplier node's self-risk. The inherent performance risk level of the supplier is as follows: the higher the basic risk value, the greater the probability of the supplier defaulting on its performance. For any group of adjacent supplier nodes with actual supply relationship in the supply chain network, the upstream and downstream nodes are determined according to the direction of the directed edge (the starting point of the directed edge is the upstream node and the ending point is the downstream node). The actual purchase amount between the nodes in this group is extracted, and the proportion of the actual purchase amount between the nodes in this group to the total purchase amount of the downstream nodes is calculated. The proportion is subjected to global Min-Max normalization and used as the weight of the directed edge between the nodes in this group (the value range is [0,1]). This represents the tightness of the supply relationship between a group of upstream and downstream nodes. The larger the edge weight, the higher the supply contribution of the upstream node to the downstream node and the stronger the business binding relationship between the two.
[0026] S2: Following the progressive order from the initial link to the final link in the supply chain, perform full-link risk propagation calculations for each group of adjacent upstream and downstream links. Combine the baseline risk propagation coefficient, propagation intensity, and common risks to calculate the effective risk of nodes between adjacent links.
[0027] Taking any historical order as the target order, retrieve its corresponding multi-stage supply chain. Starting with the first stage, simulate the risk propagation process stage by stage, and calculate the effective risk of each supplier node in the supply chain. Select any pair of adjacent stages in the supply chain, defining the earlier stage as the upstream stage and the later stage as the downstream stage. For example, taking the first stage (upstream stage) and the second stage (downstream stage) as examples, perform the following steps to complete the risk propagation calculation: Based on the multi-stage supply chain of the target order obtained in step S1, the supplier node sets corresponding to the upstream stage (first stage) and the downstream stage (second stage) are extracted respectively, and defined as the upstream node set. and downstream node set Nodes within the set are defined as upstream and downstream nodes, respectively. A baseline risk propagation coefficient is calculated based on the directed edge connections between adjacent links. Specifically, the connection relationships between the upstream and downstream node sets of all adjacent links in the target order supply chain are statistically analyzed. For each pair of adjacent links, the upstream and downstream nodes are traversed: if there is an actual supply transaction between two nodes, it is recorded as a directed edge. All connections meeting the criteria are summarized to obtain the set of directed edges between the upstream and downstream node sets, and the total number of directed edges in this set is counted. Simultaneously, calculate the theoretical maximum number of directed edges in the fully connected state of adjacent links, which is the product of the total number of nodes in the upstream link and the total number of nodes in the downstream link. Then, the baseline risk transmission coefficient was calculated.
[0028] Specifically, the formula for calculating the benchmark risk transmission coefficient can be: in, This refers to the baseline risk transmission coefficient between upstream and downstream links; These refer to the total number of nodes in the upstream node set and the downstream node set, respectively. This refers to the total number of directed edges between upstream and downstream links; This refers to a preset base value for the propagation coefficient, set based on industry risk transmission experience values. This prevents the baseline risk propagation coefficient from being too small, rendering subsequent propagation intensity calculations meaningless. In this embodiment, we take... ; The baseline risk propagation coefficient is used to characterize the global baseline strength of risk transmission between adjacent links, and the total number of directed edges. The larger, through This reflects that upstream risks can be diverted and transmitted through multiple paths, and the overall impact of an anomaly at a single upstream node on downstream processes is diluted, corresponding to the baseline risk propagation coefficient between the two processes. The smaller.
[0029] Statistics A downstream node is defined as an upstream node with which there is a directed edge connection to its upstream node set. The set of associated upstream nodes corresponding to this downstream node is further constructed as follows: Count the total number of associated upstream nodes within this set, and calculate the nth node based on the negative correlation mapping relationship using the total number of associated upstream nodes. The diffusion coefficient of each downstream node characterizes the diffusion characteristics of risk diversion, and the diffusion coefficient is negatively correlated with the number of associated upstream nodes.
[0030] Specifically, the formula for calculating the diffusion coefficient can be: in, It refers to the first The diffusion coefficient of each downstream node ranges from (0,1]. It refers to the first The total number of nodes in the set of associated upstream nodes of each downstream node, with a value range of [0,+∞); Diffusion coefficient here Used to characterize a single downstream node After absorbing upstream risks, the average risk intensity transmitted outwards, and the number of associated upstream nodes. The larger the value, the more upstream alternative suppliers the downstream node can choose from. The higher the degree to which upstream risks are diverted and diluted when propagating to that downstream node, the weaker the directional risk propagation strength on a single directed edge. (Corresponding to the...) Diffusion coefficient of each downstream node The smaller.
[0031] Based on the benchmark risk propagation coefficient between upstream and downstream links Combining the above calculations, the first The diffusion coefficient of the downstream node and the diffusion coefficient of the downstream node. The first associated upstream node and the first The directed edge weights of each downstream node are used to calculate the propagation strength between two nodes, quantifying the risk transmission strength from a single upstream node to a downstream node when there is a directed edge connection.
[0032] Specifically, no. The first associated upstream node and the first The formula for calculating the propagation intensity of a downstream node can be: in, It refers to the first The first associated upstream node and the first The propagation strength between downstream nodes; This refers to the baseline risk transmission coefficient between upstream and downstream links; It refers to the first The diffusion coefficient of each downstream node; It refers to the first The first associated upstream node and the first The weight of the directed edges between each downstream node; By weighting directed edges using the diffusion coefficient, the risk diversion effect in scenarios with multiple suppliers at downstream nodes is quantified. When a downstream node has multiple upstream suppliers (small diffusion coefficient), even if a single directed edge has a large weight, its product will be diluted accordingly, thus reducing the propagation strength of that edge. The larger the diffusion coefficient and the larger the weight of the directed edge at the downstream node, the stronger the propagation, which fully conforms to the business rule that "the stronger the supply binding, the less risk diversion, and the higher the propagation strength." At the same time, an exponential mapping is implemented using the sigmoid function, and the mapping result is combined with the benchmark risk propagation coefficient between the upstream and downstream links corresponding to the two nodes. Quantify the propagation strength between two nodes .
[0033] For multiple suppliers in the upstream segment of the supply chain, they are susceptible to the combined impact of common external factors such as policy adjustments, raw material supply, industry regulation, and market fluctuations, resulting in synchronous, systemic, and shared risks. These risks are not individual risks of a single supplier, but rather a resonant risk across the entire supply chain, which can be transmitted downstream in batches. Common risks are quantified by using the risk distribution entropy and risk mean of the upstream segment—extracting the basic risk values corresponding to all upstream nodes in the upstream segment of the target order, summing all basic risk values to define the total risk, further calculating the proportion of the basic risk value of each upstream node to the total risk, defining it as the risk probability of that upstream node; obtaining the risk probability distribution of upstream nodes, and then calculating the risk distribution entropy based on the Shannon entropy formula; simultaneously, calculating the mean of the basic risk values of all upstream nodes in the upstream segment. Further calculations will be made of the common risks in this process.
[0034] Specifically, the formula for calculating common risks can be: in, This refers to common risks in the upstream process; This refers to the average risk of upstream nodes; This refers to the risk distribution entropy in the upstream segment; It refers to the first stage in the upstream process. The probability of risk for each upstream node; This refers to the total number of upstream nodes in the upstream process; Risk distribution entropy is calculated by assessing the uniformity of risk distribution within the upstream segment. A higher risk distribution entropy indicates that the basic risk values of suppliers within the upstream segment are closer and the risk distribution is more uniform. This suggests that most suppliers are simultaneously impacted by the same factors, making it easier to form common industry risks. The average risk value within the upstream segment... The larger the risk distribution entropy, the higher the risk distribution entropy. The larger the value, the greater the common risks in the upstream links. The two are multiplied together to represent the overall common risks at the link level.
[0035] Based on the The upstream transmission risk is obtained by weighting and summing the risks transmitted from each associated upstream node to the downstream node, using the basic risk of the upstream node, the weight of the directed edge between upstream and downstream nodes, and the intensity of risk propagation between nodes as calculation factors.
[0036] Specifically, the formula for calculating upstream transmission risk can be: in, It refers to the first The risk of upstream transmission to downstream nodes; It refers to the first The first associated upstream node and the first The propagation strength between downstream nodes; It refers to the first The first associated upstream node and the first The weight of the directed edges between each downstream node; It refers to the first The basic risk value of each associated upstream node; It refers to the set of associated upstream nodes of the j-th downstream node; Upstream transmission risk Characterized the first The sum of independent risks transmitted from each associated upstream node to a downstream node based on the node's basic risk, directed edge weights, and propagation strength reflects the degree of targeted impact of individual anomalies of associated upstream nodes on downstream nodes.
[0037] After obtaining the upstream transmission risk, the upstream transmission risk is compared with the common risks of the upstream links calculated above. By performing nonlinear fusion, the upstream comprehensive risk is calculated, which comprehensively characterizes the overall risk level of the transmission from the upstream link to the downstream node.
[0038] Specifically, the formula for calculating upstream comprehensive risk can be: in, It refers to the first The comprehensive upstream risks of each downstream node; This refers to common risks in the upstream process; It refers to the first The risk of upstream transmission to downstream nodes.
[0039] The saturation mapping constructed using exponential functions A nonlinear constraint mapping is applied to the sum of common risks and upstream transmission risks. This function has saturation constraint properties: for upstream transmission risks... and common risks ,when When the value is small, the compressed value is almost equal to the original value (preserving linearity); as the combined value increases, the growth rate gradually slows down, eventually leading to a greater overall risk upstream. It approaches 1 infinitely. Therefore, The range of values converges to [0,1), ensuring comparability with the underlying risk.
[0040] In actual supply chain operations, the risks of downstream nodes mainly fall into two categories: one is upstream comprehensive risk, which is an uncertain risk whose magnitude is affected by the risk propagation process in the upstream links and has transmission uncertainty; the other is the inherent risk of the downstream node itself, which is a deterministic risk. A nonlinear fusion formula based on the product of safety margins (where safety margin is the "safety probability" corresponding to a certain risk, i.e., the probability that the risk has not occurred) is constructed. This formula is used to nonlinearly fuse the upstream comprehensive risk and the inherent risk of the downstream node, ultimately calculating the... Effective risk of each downstream node.
[0041] Specifically, the formula for calculating effective risk can be: in, This refers to the effective risk of the j-th downstream node; It refers to the first The basic risk value of each downstream node; It refers to the first The diffusion coefficient of each downstream node; It refers to the first The comprehensive upstream risks of each downstream node; For the The basic risk value of each downstream node is mapped in reverse ( ) to obtain the first The first safety margin corresponding to the inherent risks of each downstream node, for the first... The upstream comprehensive risk of each downstream node is mapped in reverse. The second safety margin corresponding to the upstream comprehensive risk is obtained, and the product of the two is... The joint security margin, representing both upstream security and its own security, is adjusted using a diffusion coefficient as an exponential modifier to account for the contribution ratio of both. A larger value indicates fewer upstream nodes associated with the downstream node, a higher dependence on upstream supply, and a greater weight of the upstream comprehensive risk on the node's effective risk. The effective risk is represented as "1 - joint security margin." When two risks exist simultaneously, the two security margins are doubly compressed, naturally achieving a "1+1>2" coupling amplification effect. In one embodiment, for a node with... , For a power function of an exponent, when the base is less than 1, the larger the exponent, the smaller the result of the power function. Therefore, the upstream comprehensive risk... The larger, The smaller, The closer to 1, Less than , making The larger.
[0042] S3: Based on the full-link risk iterative propagation calculation, the effective risks of downstream nodes are used as the input for the next round of iteration to generate the effective risk set of the end link.
[0043] Supply chain risks are characterized by cascading transmission and cumulative effects. Risks in upstream segments can propagate downstream along the supply chain, necessitating iterative risk calculations across the entire supply chain. For each historical order, the aforementioned step S2, a layer-by-layer nonlinear risk propagation calculation process, is repeated sequentially. Following the order of the supply chain from the first link to the final link, risk iterative calculations are performed on each group of adjacent upstream and downstream links until the effective risk values of all supplier nodes at the final fulfillment stage of the supply chain are calculated. This yields the set of effective risk values for the final link corresponding to that historical order. This set comprehensively carries the risk superposition information of the entire supply chain, fully reflecting the overall supply chain risk status corresponding to that order.
[0044] Specifically, to achieve the layer-by-layer transmission and accumulation of risks along the supply chain, a risk iteration propagation rule is specifically set up during the iterative calculation process. The effective risks of each downstream node in the downstream link calculated in the previous round of adjacent links are used as the input risk values of the corresponding supplier nodes in the new upstream links in the next round. This directly replaces the original basic risk value of the node and participates in the entire calculation process of the benchmark risk propagation coefficient, node propagation intensity, common risks, and effective risks of downstream nodes in the next adjacent link. The above iterative updates are performed layer by layer along the supply chain to ensure that the risks at each level can be completely transmitted to the next level, achieving the accumulation of risks without omissions throughout the entire chain.
[0045] S4: Extract global risk statistical features with fixed dimensions across the entire link to construct a training sample set, and use a logistic regression binary classification prediction model to complete model training and hyperparameter saving.
[0046] Because the number and identities of suppliers at the end of the supply chain vary across different historical orders, the effective risk of a single end node is not directly used as the model input feature. Instead, fixed-dimensional global risk statistical features are extracted from the risk data of the entire supply chain corresponding to the order to form a feature vector of uniform length.
[0047] Specifically, global statistical features of the entire supply chain risk are extracted, including effective risk statistical features of the entire supply chain and effective risk statistical features of the end links. The effective risk statistical features of the entire supply chain include the maximum, mean, median, 75th percentile, and 25th percentile of the effective risk values of all nodes in the entire supply chain, totaling 5 statistical features. The effective risk statistical features of the end links include the maximum and mean of the effective risk set of the end links, totaling 2 statistical features. The above two types of statistical features are concatenated in order to form a feature vector with a fixed dimension of 7, which is defined as the supply chain statistical feature vector.
[0048] Furthermore, the supply chain statistical feature vector corresponding to each historical order is associated one by one with the fulfillment result label of the order (fulfillment success = 0, fulfillment failure = 1) to construct the training sample corresponding to the historical order. The training samples corresponding to all historical orders together constitute a complete training sample set. Each training sample contains a 7-dimensional supply chain statistical feature vector and a binary classification label.
[0049] Logistic regression is used as the supply chain risk prediction model. The model input is a supply chain statistical feature vector of uniform length, and the output is a binary probability of order fulfillment success and failure. The probability of order fulfillment failure is defined as the total global supply chain risk of that order.
[0050] Specifically, the constructed training sample set is randomly divided into a training set and a test set in an 8:2 ratio; the parameters of the logistic regression model are initialized (including weight coefficients initialized to a zero vector, bias term initialized to 0, learning rate set to 0.01, maximum number of iterations set to 1000, and L2 regularization coefficient set to 0.001), using the cross-entropy loss function as the loss metric to quantify the deviation between the model's predicted output and the actual order fulfillment results, and gradient descent is used for iterative optimization; the hyperparameters such as learning rate, regularization coefficient, and maximum number of iterations are optimized using grid search combined with five-fold cross-validation, and the average validation accuracy is used as the selection metric for hyperparameter combinations. The above gradient descent iterative optimization process is repeated until the preset maximum number of iterations or the loss value converges; after training, the optimal hyperparameter combination and the trained model parameters are saved for subsequent real-time risk prediction.
[0051] S5: For orders to be predicted, reuse the data preprocessing and end-to-end risk propagation calculation process, input the trained logistic regression model, output the total global risk of the supply chain and realize the source of risk nodes.
[0052] For orders to be predicted, based on the bill of materials trajectory of the order, the same process described above is followed to complete data collection and standardization, construct the supply chain network, calculate the nonlinear risk propagation layer by layer across the entire chain, and extract the 7-dimensional supply chain statistical feature vector of the order to be predicted according to the feature extraction rules in step S4. The feature vector is then input into the logistic regression model trained in step S4. The model outputs the probability of failure of the order to be predicted. This probability is defined as the total global risk of the supply chain of the order to be predicted. The larger the value, the higher the probability of order failure.
[0053] This application also discloses a supply chain risk prediction system based on multi-source heterogeneous data analysis, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the supply chain risk prediction method based on multi-source heterogeneous data analysis according to this application is implemented.
[0054] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0055] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), etc., or any other medium that can be used to store desired information and can be accessed by an application program, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.
[0056] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A supply chain risk prediction method based on multi-source heterogeneous data analysis, characterized in that, include: Obtain multi-stage supply chain data corresponding to orders, collect multi-source heterogeneous data of historical orders, supply transaction data and fulfillment result tags, and preprocess them; Calculate the basic risk value of supplier nodes and the weight of directed edges to construct the supply chain network; Following the progressive order from the initial stage to the final stage of the supply chain, perform full-chain risk propagation calculations for each group of adjacent upstream and downstream stages; and calculate the baseline risk propagation coefficient by combining the number of directed edges and the total number of nodes in adjacent stages. For any downstream node, the upstream nodes that are associated with it by a directed edge are taken as associated upstream nodes, and the diffusion coefficient is calculated based on the number of associated upstream nodes. The propagation intensity is calculated by combining the baseline risk propagation coefficient, diffusion coefficient, and directed edge weights between related upstream and downstream nodes; common risks are calculated based on the distribution entropy and mean of the basic risk values of supplier nodes in the upstream link. The upstream transmission risk is calculated based on the basic risk value of the associated upstream node, the directed edge weight, and the propagation strength. The upstream transmission risk is then integrated with the common risk to obtain the upstream comprehensive risk. Based on the diffusion coefficient, the effective risk of the downstream node is calculated by nonlinear weighted fusion of the upstream comprehensive risk and the basic risk value of the downstream node. Based on the full-link risk propagation calculation results, a training sample set is constructed by extracting fixed-dimensional global risk statistical features, and a logistic regression model is trained. The fixed-dimensional global risk statistical features of the order to be predicted are input into the trained logistic regression model, and the total global supply chain risk of the order to be predicted is output.
2. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The calculation of the supplier node's basic risk value and directed edge weights includes: Extract the qualification, performance and compliance indicators from the multi-source heterogeneous data, and calculate the basic risk value by weighting and summing the qualification, performance and compliance indicators based on preset weight coefficients; For any pair of adjacent supplier nodes with an actual supply relationship in the supply chain network, the upstream and downstream nodes are determined according to the direction of the directed edge. The actual purchase amount between adjacent nodes is extracted, and the ratio of the actual purchase amount to the total purchase amount of the downstream node is calculated. The ratio is then normalized and used as the weight of the directed edge.
3. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The calculation method for the benchmark risk transmission coefficient includes: The total number of directed edges between adjacent upstream and downstream links is counted. The product of the total number of upstream nodes and the total number of downstream nodes is calculated, and the product is used as the theoretical maximum total number of directed edges. The ratio of the total number of directed edges to the theoretical maximum total number of directed edges is calculated. The ratio is then reverse-mapped to obtain a mapping value. Based on the preset base value of the propagation coefficient and the mapping value, the benchmark risk propagation coefficient is calculated.
4. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The calculation method for the propagation intensity includes: Calculate the product of the diffusion coefficient and the directed edge weights between the associated upstream and downstream nodes, and perform a nonlinear mapping on the product to obtain the mapping result; The product of the mapping result and the baseline risk propagation coefficient is used as the propagation strength between the associated upstream and downstream nodes.
5. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The calculation methods for the common risks include: Calculate the mean of the basic risk values of all upstream nodes in the upstream link; calculate the risk distribution entropy of the basic risk values of all upstream nodes in the upstream link; perform a nonlinear mapping on the risk distribution entropy to obtain the mapping value, and use the product of the mapping value and the mean as the common risk of the upstream link.
6. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The calculation method for the upstream comprehensive risk includes: The upstream transmission risk is obtained by summing the products of the basic risk value of each associated upstream node, the directed edge weight between the associated upstream node and the downstream node, and the propagation strength. The sum of the common risks and the upstream transmission risks is calculated, and the sum is nonlinearly constrained and mapped based on an exponential function to obtain the upstream comprehensive risk.
7. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The calculation methods for the effective risk include: The first safety margin is obtained by inverse mapping of the basic risk value of the downstream node; the second safety margin is obtained by inverse mapping of the upstream comprehensive risk of the downstream node; the diffusion coefficient is used as an exponential adjustment factor to calculate the weighted product of the first safety margin and the second safety margin to obtain the joint safety margin; the joint safety margin is inversely mapped, and the mapping result is used as the effective risk of the downstream node.
8. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, In the full-link risk propagation calculation, the effective risk of the downstream node calculated in the current stage is used as the basic risk value of the corresponding node in the new upstream stage in the next round for iterative updates. The construction of the training sample set includes: extracting effective risk statistical features of the entire chain and effective risk statistical features of the end link, constructing a fixed-dimensional supply chain statistical feature vector, constructing training samples based on the supply chain statistical feature vector of any historical order and its fulfillment result label, and forming a training sample set by combining the training samples of each historical order.
9. The supply chain risk prediction method based on multi-source heterogeneous data analysis according to claim 1, characterized in that, The trained logistic regression model includes: The training sample set is divided into a training set and a test set according to a preset ratio; Using the cross-entropy loss function as the optimization objective, gradient descent combined with cross-validation and parameter optimization strategies is used to iteratively train the model's hyperparameters until the preset maximum number of iterations or the loss value converges and training is terminated.
10. A supply chain risk prediction system based on multi-source heterogeneous data analysis, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the supply chain risk prediction method based on multi-source heterogeneous data analysis according to any one of claims 1-9.