Engineering supply chain risk conduction analysis system and method based on graph neural network

Abstract

The invention provides an engineering supply chain risk conduction analysis system based on a graph neural network and a method thereof. The method comprises the following steps: acquiring supply chain operation data from multi-level supply chain nodes; constructing a graph structure model between supply chain nodes according to the operation data, and identifying and quantifying an association relationship between the nodes; simulating the risk event occurrence of the supply chain node on the graph structure model; based on a graph structure model, calculating the probability and influence value of risk propagation from an occurrence node to other nodes, and identifying a main risk propagation path; using a Motif anomaly detection module to detect an abnormal sub-graph mode appearing in a risk propagation process in the graph structure model; performing representation learning and association prediction on the graph structure model by using a self-supervised graph neural network, and mining potential hidden risk association; by constructing a multi-level supply chain association map and introducing risk disturbance simulation, the system can quantify the conduction rule of the risk in a supply chain network from a global perspective, and can automatically identify hidden high-risk links and key propagation paths.

Description

Technical Field

[0001] The present invention relates to the technical field of supply chain risk analysis, and in particular to an engineering supply chain risk conduction analysis system and method based on a graph neural network. Background Art

[0002] The supply chain of large-scale engineering projects interweaves material suppliers, manufacturers, subcontractors, logistics providers, and other parties, forming a hierarchical yet interconnected network. While surface risks such as delivery delays, quality defects, and cost fluctuations can be identified through ERP and contract audits, it's difficult to discern the chain reactions triggered by hidden connections such as equity, guarantees, and financial transactions.

[0003] Once hidden connections are triggered, risks can rapidly spread across layers along unmonitored channels, potentially impacting multiple processes within minutes or hours, posing serious challenges to project schedules and cost management. Manual verification of heterogeneous information, such as purchase orders, financial statements, public opinion data, and IoT monitoring, is not only time-consuming and labor-intensive, but also prone to errors and omissions.

[0004] Due to its advantages in graph structure data mining, graph neural networks can map supply chain nodes and relationships into graphs, and automatically capture high-order associations and abnormal patterns through iterative aggregation of nodes and edges.

[0005] In recent years, academics have demonstrated from various perspectives that graph learning methods are urgently needed to address the hierarchical diffusion and implicit connections within construction supply chains. The study "Risk Study of Prefabricated Building Supply Chains from an Information Diffusion Perspective" first used a questionnaire to extract 19 core risk factors and constructed a corresponding adjacency matrix to depict risk transmission channels, but failed to further introduce deep graph learning algorithms to quantify multi-hop effects. Subsequently, Yan Bin et al., in their empirical analysis of prefabricated building green supply chain risks using the ISM-MICMAC model, hierarchically decomposed 14 risks, but remained at a static structural decomposition level and failed to capture dynamic probability evolution. Guo Wenping's weighting system based on ANPs also lacked real-time assessment of hidden equity chains and multi-level transmission chains. In contrast, international research has applied GNNs to supply chain default prediction, demonstrating that graph models can significantly improve the accuracy of identifying high-order implicit connections. U.S. Patent US11599840B2 applied GNNs to the financial sector, successfully discovering implicit connections within lending networks and significantly improving the accuracy of risk identification.

[0006] This principle is particularly valuable in complex engineering supply chain scenarios, such as those for super-high-rise buildings, rail transit, municipal pipeline networks, cloud data centers, and large-scale IoT integration. Such projects are often spearheaded by a core general contractor, extending downward to dozens of first-tier specialized subcontractors (steel structures, curtain walls, intelligent electromechanical systems, etc.), then to second-tier material manufacturers (steel plates, specialty glass, PLC control cabinets), and even to third-tier service teams for galvanizing, anti-corrosion, and hoisting. Beyond the contractual supply chain, these seemingly independent nodes are interwoven with hidden connections such as shared shareholders, guarantee chains, common environmental policies, and shared logistics platforms. If a second-tier supplier stops supplying due to production restrictions, the risk can be amplified along the path of "raw materials-first-tier subcontractor-general contractor"; if multiple subcontractors are subject to the same financial guarantee chain, a single-point default may instantly affect the entire process network; monitoring only direct partners is not enough to gain insight into the overall situation. It is necessary to use methods such as graph neural networks to unify multi-level and cross-domain explicit and implicit relationships into a dynamic graph, and quantify the risk propagation probability and potential impact of each node in real time, so as to accurately warn and formulate mitigation measures before the risk is traced back from the secondary and tertiary links to the core enterprise.

[0007] However, it still has many technical defects in the engineering supply chain scenario: first, it only constructs a static relationship diagram, which cannot reflect the multi-level and real-time dynamics under multi-source heterogeneous data; second, it lacks the simulation of the transmission time series and the quantification of the propagation probability of risk events; third, it does not introduce disturbance simulation and motif-level anomaly detection, making it difficult to locate the vulnerable points of the link structure; at the same time, it does not use self-supervised comparative learning to explore potential associations in unlabeled data. Summary of the Invention

[0008] The purpose of the present invention is to provide an engineering supply chain risk transmission analysis system and method based on graph neural network to solve the problems in the existing technology such as insufficient identification of implicit correlation risks and difficulty in quantifying risk transmission laws.

[0009] The overall architecture of the system includes multiple functional modules, which work together to provide end-to-end risk analysis support from data collection, risk identification to result visualization.

[0010] The present invention provides an engineering supply chain risk transmission analysis system based on graph neural network, which includes the following contents:

[0011] First, the multi-layer supply chain data collection module is used to collect operational data from nodes at all levels of the supply chain in real time or periodically. This operational data includes: order fulfillment status, delivery cycle, inventory levels, financial indicators, contract fulfillment history, corporate equity structure, management relationships, and external public opinion information.

[0012] By collecting data from different levels of nodes, including core enterprises, first-, second-, and third-tier suppliers, and major customers, we build a rich risk analysis database. This module can connect to internal enterprise ERP / SCM systems, IoT devices, and external public data sources to achieve comprehensive acquisition and preprocessing of structured and unstructured data.

[0013] Preferably, order and inventory data are extracted from the ERP system, corporate equity relationships are obtained from the industrial and commercial database, and negative reports on suppliers are captured from the news and public opinion platform.

[0014] The data acquisition module also cleans, aligns and standardizes the collected data to ensure the accuracy of subsequent analysis.

[0015] Secondly, the multi-layer supply chain structure model building module* builds a graph model of the supply chain network based on the collected multi-source data.

[0016] The graph model is formally represented as a directed graph G = (V, E), where the node set V represents the enterprise entities in the supply chain (including core enterprises and suppliers and customers at all levels), and the edge set E represents the relationship between nodes.

[0017] The associated relationships include both direct supply and demand relationships, including supply contract links, and implicit business relationships, including common shareholders, senior executives, and guarantee relationships.

[0018] The model building module automatically identifies the association strength between supply chain nodes through data fusion and relational reasoning, and generates a multi-level supply chain association graph.

[0019] Let A represent the adjacency matrix of the graph, then A ij >0 indicates that there is a certain degree of correlation between node i and node j, and the value reflects the strength of the relationship.

[0020] In view of the hierarchical characteristics of the supply chain, the model building module can label each edge with a hierarchical difference value ΔL ij , represents the hierarchical distance between nodes i and j;

[0021] For example, the core enterprise has ΔL=1 with its first-tier suppliers, ΔL=2 with its second-tier suppliers, and so on;

[0022] This module supports dynamic updates of the network model: when new supplier relationships or node status changes are detected, the graph structure and associated weights are updated in a timely manner to ensure that risk analysis is based on the latest global supply chain map.

[0023] Based on the constructed supply chain diagram, the present invention is provided with a risk disturbance simulation module and a risk propagation path analysis module.

[0024] The risk disturbance simulation module is designed to simulate the impact of a risk event occurring at a certain node in the supply chain on the entire supply chain, and is used to quantify the risk transmission law;

[0025] This module can select one or more nodes on the graph G to apply perturbations;

[0026] Preferably, simulate a scenario where a supplier fails to deliver, defaults, or goes bankrupt, and track the spread of this disturbance in the network.

[0027] The risk propagation path analysis module uses graph algorithms to calculate the probability and impact of disturbances being transmitted from the occurrence node to other nodes along different paths.

[0028] Specifically, define the risk from node N i Propagate to node N j The propagation probability P ij , its calculation can comprehensively consider factors such as the direct connection strength between nodes and the hierarchical distance, and is calculated using the following formula:

[0029] P ij =αexp(-λΔL ij )·S ij ,

[0030] Where 0≤P ij ≤1, α is the propagation probability coefficient constant, S ij Represents node N i With N j The strength of the association between them (given by the edge weight of the supply chain graph), λ is the hierarchical influence coefficient (a positive value used to control the inhibitory effect of hierarchical differences on the probability of risk propagation), ΔL ij For node N i With N j The level difference between them.

[0031] Through the above model, the contribution of nodes farther away in the hierarchy (larger ΔL) to risk transmission is attenuated, and nodes with closer connections have a higher probability of risk transmission.

[0032] Based on the calculated propagation probability matrix P, the risk propagation path analysis module further evaluates the risk impact value of each node affected by the initial risk source.

[0033] Preferably, define node N i When a risk occurs, another node N j Impact value I j|i for:

[0034]

[0035] where R i Represents node N iThe initial risk intensity (such as the severity of the risk event, which can be set as R i =1 means risk occurs, 0 means no risk occurs), sum up all the i to N j All possible propagation paths of the , the propagation probability of each edge (u→v) on the path is P uv express.

[0036] The influence value measures the node N i The risk is cumulatively transmitted to N through various channels j The possible size of the impact.

[0037] In particular, when i=j, path(i→i) is considered as a path with a length of 0, and its own influence value is specified as I i|i =R i .

[0038] By calculating the I of all i to j j|i And take the largest one, we can get node N j The comprehensive risk impact value I under the action of a certain initial risk source j .

[0039] Based on this, the risk propagation path analysis module identifies the main propagation paths from the initial risk node to key nodes such as core enterprises, as well as high-risk vulnerable points. The system can automatically mark paths with high propagation probability and large cumulative impact as key risk links, providing a basis for subsequent abnormal pattern recognition and decision support.

[0040] Based on the above risk analysis, the present invention also introduces a Motif anomaly detection module to discover abnormal patterns in supply chain risk propagation.

[0041] Motif refers to a specific subgraph structure pattern that often appears in the network (such as a triangular closed loop, a star structure, etc.).

[0042] In the supply chain scenario, different Motif patterns may correspond to different risk transmission mechanisms. For example, a "triangular closed loop" relationship means that the three parties are suppliers to each other or have circular dependence.

[0043] The anomaly detection module first enumerates the Motif subgraph patterns in the supply chain association graph and counts the occurrence frequency or feature distribution of various patterns during normal operation as a benchmark.

[0044] Then, the risk propagation path generated by the above disturbance simulation is matched with the typical Motif pattern, and the difference between the actual pattern frequency and the baseline expectation is compared.

[0045] If certain Motif patterns are significantly abnormal (too high or too low) during risk transmission, it is determined that the corresponding correlation relationship has abnormal signs.

[0046] For example, a type of related triangle that is rarely seen under normal circumstances appears frequently in a certain risk event, which may mean a hidden common cause failure chain.

[0047] The Motif anomaly detection module outputs a list of abnormal subgraphs and their corresponding anomaly scores to help locate unconventional risk propagation pathways.

[0048] In addition, the system of the present invention includes a self-supervised graph neural network analysis module, which uses advanced machine learning technology to further improve the depth and accuracy of risk transmission relationship mining.

[0049] This module is built on the supply chain association graph and uses graph neural networks to perform feature representation learning (embedding) for each node. Unlike supervised learning, the self-supervised GNN module does not require manual labeling of risk events and learns patterns from the data itself by designing self-supervised tasks.

[0050] Preferably, the system adopts the contrastive learning method:

[0051] Randomly mask or perturb a part of the nodes or edges in the graph, and then let GNN predict the masked information to explore the potential correlation features between the nodes.

[0052] In the specific implementation, a multi-layer graph convolutional network is used to transmit and aggregate neighbor node information.

[0053] For each layer l, node N i The update formula of the representation vector is as follows:

[0054]

[0055] in is the set of neighboring nodes of node i, d i and d j are the degrees of nodes i and j respectively, W (l) is the trainable weight matrix of the lth layer, and σ is the nonlinear activation function.

[0056] After multi-layer propagation, the high-dimensional representation h of each node can be obtained i , capturing their structural positions and relationship patterns in the supply chain network.

[0057] Then, a self-supervised loss function is designed, for example, making the representations of nodes with close distances more similar in the vector space, while the representations of random unrelated nodes are more different, prompting the model to learn meaningful representations.

[0058] The GNN model trained in this way can be used for link prediction (i.e., predicting potential associations that may exist in the graph but are not explicitly recorded) and node anomaly scoring (evaluating the risk anomaly of a node based on its deviation from the group distribution).

[0059] The introduction of the self-supervised GNN module enables the system to continuously improve its risk insight capabilities from unlabeled data and discover deep patterns that are missed by traditional algorithms.

[0060] Finally, the system is equipped with a risk radar heat map display module to visualize the above analysis results, making it easier for decision makers to intuitively understand the risk distribution of the supply chain.

[0061] This module can convert key risk indicators (such as node risk impact value I j , path propagation probability, etc.) are mapped into intuitive heat maps or radar maps.

[0062] In summary, the technical solution of this invention covers the entire process from data acquisition, model building, risk simulation, pattern recognition, to result presentation. Through the close coordination of various modules, it achieves in-depth analysis of the hidden risk transmission relationships in the engineering supply chain.

[0063] The beneficial effects achieved by the present invention using the above method and system are as follows:

[0064] 1. By constructing a multi-level supply chain association map and introducing risk perturbation simulation, this system can quantify the transmission patterns of risks in the supply chain network from a global perspective. Compared with traditional point-by-point monitoring methods, this invention can automatically identify hidden high-risk links and key transmission paths, enabling enterprises to focus on potential risk sources in advance and improve the timeliness and accuracy of early warnings.

[0065] 2. This system innovatively integrates motif pattern anomaly detection and self-supervised graph neural network technology to conduct in-depth research on supply chain risks. Motif detection can discover unconventional risk propagation patterns, while self-supervised GNN extracts risk correlation features from massive amounts of unlabeled data. The combination of the two greatly improves the ability to identify hidden risk relationships and the adaptability of the model, providing insights into complex supply chains that go beyond human experience.

[0066] 3. This system provides an intuitive radar chart / heat map interface to display risk analysis results. Combined with the decision support module, it provides targeted risk mitigation recommendations. By graphically presenting high-risk nodes, important paths, and abnormal patterns, managers can effectively interpret risk distribution. Based on the analysis results, the system also provides specific recommendations, such as replacing high-risk suppliers and increasing inventory, to help companies develop effective supply chain optimization and risk mitigation strategies, comprehensively improving the robustness of their supply chains. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the system architecture of the present invention;

[0068] Figure 2 It is a schematic diagram of the Motif anomaly detection process of the present invention;

[0069] Figure 3 It is a heat map diagram of the risk propagation probability of the present invention. DETAILED DESCRIPTION

[0070] The engineering supply chain risk transmission analysis system and method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0071] like Figure 1 As shown, the system of the present invention generally includes multiple functional modules, and each module realizes the identification and analysis of supply chain risks through information flow connection. In specific implementation, the multi-layer supply chain data acquisition module 101 is used to collect structured and unstructured data from supply chain participants. For example, in the supply chain of an engineering project, the core enterprise can provide project procurement plans and progress data, the first-tier supplier uploads the production delivery progress, and the second-tier and third-tier suppliers feedback the inventory and raw material procurement status through the interface. In addition, the data acquisition module 101 also obtains the industrial and commercial registration information of the supplier company, the shareholder and senior executive related information, and the news media reports on related companies (such as financial risks, business anomalies, etc.) through the Internet. All these data are pre-processed in module 101 and stored in the graph database in a unified format to provide a basis for subsequent analysis.

[0072] The supply chain structure model building module 102 retrieves the cleaned node information from the database and automatically identifies the various association relationships between the nodes.

[0073] In this embodiment, module 102 first establishes a basic supply chain graph based on direct business relationships: for example, a directed edge connects node A (the core enterprise) and node B (its first-tier supplier), representing that B supplies goods to A. Simultaneously, the algorithm traverses the equity relationship data. If it finds that node C and node B are controlled by the same parent company, an edge between C and B is added to the graph, with the relationship type marked as "shared control." In a similar manner, module (102) integrates data from multiple sources, adding all explicit and implicit relationships to the graph model.

[0074] Once the model is built, the system generates a graph that shows all relevant supply chain entities and their relationships. Each node in the graph includes attributes (such as company size and location), and each edge includes relationship attributes (such as frequency of collaboration and equity ownership) and a weight (which quantifies the strength of the relationship).

[0075] For pairs of nodes at different levels, the system calculates the level difference ΔL and stores it in the edge attribute for use in subsequent risk propagation calculations. In this embodiment, if the core enterprise is at level 0, the first-tier supplier is at level 1, and the second-tier supplier is at level 2, then ΔL = 2 between the core enterprise and the second-tier supplier.

[0076] The model building module 102 has online update capabilities: when a change in the supply chain structure is detected, such as the addition of a new supplier, the module automatically updates the graph database to keep the risk analysis model synchronized with reality.

[0077] After completing the supply chain map, enter the risk analysis stage.

[0078] The risk disturbance simulation module 103 and the risk propagation path analysis module of the present invention work together to evaluate the transmission impact of risks in the network.

[0079] In this embodiment, the user may select a high-risk node on the interface as a starting point to trigger the disturbance simulation.

[0080] For example, in a simulation, the secondary supplier node 304 (corresponding number is shown in Figure 1 ) default occurs, the module 103 will change the initial risk level of node 304 to R 304 Set to 1 (other nodes are 0) and pass this information to the risk propagation analysis module.

[0081] The risk propagation path analysis module calculates the propagation probability P from node 304 to other nodes based on the edge weights and level differences in the graph model. 304→j (j represents other nodes in the network).

[0082] Using formula P ij =αexp(-λΔL ij )S ij Calculations are performed, where parameters α and λ can be tuned based on historical data.

[0083] In this embodiment, since 304 is a second-tier supplier, it has direct supply relationships with some first-tier suppliers (ΔL=1 and S 304,302 Higher), so P 304→302 The calculated result will be larger. However, there is a layer of suppliers (ΔL=2) between 304 and the core enterprise. Even if there is an indirect relationship, the probability of transmission will be significantly reduced due to the layer attenuation.

[0084] After obtaining the segment-by-segment propagation probability, the system further enumerates 304 all possible paths to other nodes.

[0085] For example, 304→302→301 (core enterprise) is a two-segment path.

[0086] Module 103 calculates the cumulative propagation probability ΠP=P along this path 304→302 ×P 302→301 , and considering the initial risk intensity R 304 , get the impact value I of 304 on 301 301|304 The module will perform similar calculations on the path from 304 to each node, and add up the impact values of all paths to get the total impact I j .

[0087] Calculations in this embodiment show that the impact value of node 304 on the core enterprise 301 is approximately 0.4, while it has essentially no impact on another secondary supplier 305 (different branch paths are independent, and 305 is not on any path of 304).

[0088] Based on this, the analysis module produces a list of affected nodes sorted in descending order of impact value.

[0089] In the list, supplier 302 is labeled as a high-risk node due to its direct connection to 304 and its connection to the core enterprise. Core enterprise 301 ranks second, while irrelevant supplier 305 has an impact value close to 0. This result is consistent with intuition: 304's default primarily impacts nodes closely related to its business.

[0090] Next, the Motif anomaly detection module 104 performs pattern analysis on the above risk propagation process.

[0091] In the embodiment, the system pre-defines and counts common Motif patterns in the supply chain network and their normal occurrence frequencies. For example, the average occurrence rate of a "triangle closed loop" is 5%, which is often composed of a core enterprise-supplier-subcontractor.

[0092] After the default event of node 304 occurs, module 104 finds that the frequency of occurrence of triangular closed-loop structures in the subgraph involved in the propagation path increases to 10%, which is an abnormality.

[0093] Specifically, 304 connects back to the core enterprise through two different first-tier suppliers, forming a closed triangle (304-302-301-304). This motif, while not normally active, played a significant role in this risk event. This suggests to the enterprise that a previously unnoticed circular dependency structure amplified the risk impact.

[0094] In addition to the triangle, the module also detected an abnormal "star" motif in a certain supply chain link. A secondary supplier affected multiple first-tier suppliers at the same time, suggesting that the secondary supplier was a single point of failure.

[0095] All these abnormal patterns, along with their risk indicators, are output as abnormality reports for decision makers’ reference.

[0096] The self-supervised graph neural network analysis module 105 is used in this embodiment to further improve risk identification.

[0097] The system uses the GNN model trained offline on historical data to generate a risk feature vector for each node in the current supply chain graph, and performs link prediction and node anomaly score calculation as needed.

[0098] When the default event 304 occurs, module 105 performs a forward propagation on the network to obtain the embedded representation of the core enterprise 301 and its related nodes.

[0099] The results show that two node pairs that were not identified in the original graph were assigned high association scores by the model: one is between the second-tier supplier 304 and another second-tier supplier 306, where the model believes there is a hidden association (perhaps both rely on the same third-party supplier); the other is between the first-tier supplier 302 and supplier 303, where the model determines that they may be affected by the same external factors (such as geopolitical risks).

[0100] These potential risk associations given by GNN can prompt risk managers to conduct further investigation and verification to prevent the spread of risks on unknown links.

[0101] The self-supervised graph neural network (GNN) described in this embodiment adopts a three-layer GraphSAGE framework.

[0102] The model has 3 layers, L = 3, with hidden layer dimensions set sequentially from 128 to 128 to 64, and an output node representation dimension of 64. Mean Aggregator is used as the aggregation function for each layer, with ReLU activation and BatchNormalization introduced after adjacent message aggregation. A dropout rate of 0.2 is used between layers to suppress overfitting and stabilize gradients. To ensure effective deep feature transfer, residual connections are added between the input and output of each layer.

[0103] To learn potential node associations without manual annotation, the system randomly generates two views of the supply chain association graph G: one in which 10% of the edges are randomly permuted or deleted, and the other in which ±10% Gaussian noise is added to the node attributes. These are used to construct positive and negative sample pairs, employing the InfoNCE contrastive loss:

[0104]

[0105] where h (1) ,h (2) is the embedding vector of node v in two views, with a temperature coefficient of τ = 0.5. The model achieves unsupervised feature alignment by maximizing the similarity of embeddings of the same node in different views and minimizing the similarity with the embeddings of other nodes.

[0106] The Adam optimizer was used in the training phase, with an initial learning rate of 1×10-3, β parameters (0.9, 0.999), and weight decay of 5×10 -4 The batch size is the entire graph (Graph Batch), and the training is performed for 200 epochs. The early stopping patience is set to 20. When the validation set loss does not decrease significantly within 20 consecutive epochs, the training is terminated to prevent overfitting.

[0107] The model was trained on a single NVIDIA A100 GPU (40GB of video memory), taking an average of approximately 25 minutes per epoch. During the inference phase, only forward propagation is performed. For medium-sized graphs with 10,000 to 50,000 nodes, the average latency is less than 200ms, meeting the real-time early warning requirements of engineering projects. The model embedding results serve as input for node anomaly scoring and potential association prediction in the 105 modules.

[0108] In addition to the contrast loss, we can also add the edge reconstruction binary cross entropy loss The comprehensive loss is:

[0109]

[0110] The value of λ is between 0.1 and 0.3, which takes into account both structural reconstruction and recognition and discrimination capabilities.

[0111] This embodiment is compared with the traditional method:

[0112] In order to verify the technical advantages of the present invention in multi-level implicit risk identification and real-time response, this embodiment selects three methods for comparison: the first is the traditional rule engine method, which is based on preset thresholds and manual inspections and can only capture explicit single-hop risks; the second is the multi-factor single-hop evaluation method, which uses a multi-factor model to quantify the risks of directly cooperating nodes, but does not consider implicit multi-hop transmission; the third is the GNN+Motif-Radar method described in the present invention, which integrates multi-level graph models, self-supervised graph neural networks and Motif anomaly detection, and supports the automatic identification of explicit, implicit and multi-hop risks.

[0113] The comparative experiment was conducted in a real engineering supply chain online environment, with a total of approximately 12,000 supply chain nodes and approximately 68,000 associated edges.

[0114] Evaluation indicators include: risk detection accuracy (the ratio of correctly identified positive samples to excluded negative samples) and average response time (the delay from the triggering of a risk event to the completion of the propagation path and the output of the abnormal pattern). The experimental results are shown below.

[0115]

[0116] The experimental results show that:

[0117] The detection accuracy rate increased from 78.5% of the traditional rules to 92.7%, an increase of 14.2 percentage points;

[0118] The average response time was shortened from 5.2 minutes to 0.9 minutes, a reduction of 82.7%, significantly meeting the real-time warning needs;

[0119] The method of the present invention has obvious advantages in capturing the hidden risk transmission chain of second-tier and third-tier suppliers, and can provide more forward-looking and accurate risk warnings and mitigation suggestions for engineering projects.

[0120] The risk radar heat map display module 106 presents the above analysis in a visual form.

[0121] like Figure 3 As shown in the figure, the system generates a 6×6 risk propagation probability matrix heat map (node numbers 1-6 correspond to several actual supply chain companies). In the heat map, for the 304 default event (304 corresponds to number 4), you can see that the 4th row, 1st column and the 4th row, 2nd column are bright red, indicating that the risk propagation probability of node 4 to node 1 (core enterprise) and node 2 (first-tier supplier) is high; the lighter colors of the other columns in the 4th row indicate that the risk is gradually attenuated. Managers use heat maps to Figure 1 It is clear that the risk is mainly transmitted from 304 to node 2 and then to node 1.

[0122] In addition to the heat map, the system interface also provides a radar chart showing the scores of the core enterprises' main supply chain risk dimensions; as well as a list of high-risk paths, highlighting paths such as 304-302-301. Users can interactively click on a path to obtain detailed data, such as the P value of each edge. ij Value and cumulative impact value I path .

[0123] In this embodiment, the decision maker identifies 304 as a high-risk supplier that requires focused management, and notes the importance of the 302-301 path.

[0124] Finally, with the assistance of the decision support module, enterprises can formulate corresponding risk mitigation measures.

[0125] Based on the analysis, the module recommends that the core enterprise take the following measures for supplier 304: Since 304 has a significant impact on 302 and 301, it is recommended to add a redundant supplier in the supply chain where 304 is located to disperse the risk; require supplier 302 to increase the frequency of monitoring of 304's supply and reserve safety stock in advance when necessary; at the same time, investigate the potential connection between 304 and 306 to prevent related risks.

[0126] These targeted suggestions are automatically generated by the system based on the analysis results and presented to users, greatly improving the scientific nature and timeliness of supply chain risk management decisions.

[0127] From the description of the above embodiments, it can be seen that the system and method of the present invention can effectively identify and quantify the implicit risk transmission relationship in the supply chain network, making up for the shortcomings of traditional technologies.

[0128] The combination of deep learning models based on graph neural networks and graph algorithms makes risk analysis of complex supply chains efficient and intelligent.

[0129] The system can be used for pre-emptive risk detection and early warning, as well as for traceability analysis and emergency decision support after a risk event occurs. It is of great significance for improving the risk resistance of the engineering supply chain.

[0130] The protection scope of the present invention is not limited to the above embodiments. Without departing from the spirit of the present invention, ordinary technicians in the technical field can also make various modifications or equivalent substitutions, which should also be regarded as the protection scope of the present invention.

Claims

1. An engineering supply chain risk transmission analysis system based on graph neural network, characterized by: include: Multi-layer supply chain data collection module, used to collect supply chain operation data from multi-layer supply chain nodes; A multi-layer supply chain structure model building module is used to build a graph structure model between supply chain nodes based on the operation data, and identify and quantify the relationship between each node; A risk disturbance simulation module, used to simulate the occurrence of risk events at supply chain nodes on the graph structure model; A risk propagation path analysis module is used to calculate the probability and impact value of risk propagation from the occurrence node to other nodes based on the graph structure model, and identify the main risk propagation paths; Motif anomaly detection module, used to detect abnormal subgraph patterns that appear in the risk propagation process in the graph structure model; A self-supervised graph neural network analysis module is used to perform representation learning and association prediction on the graph structure model and to mine potential hidden risk associations; Risk radar heat map display module, used to visualize the node risk level and risk propagation path in the supply chain network; and decision support modules to generate recommended solutions for supply chain risk mitigation and optimization based on the analysis results.

2. The system according to claim 1, wherein: The operating data collected by the multi-layer supply chain data collection module includes order fulfillment status, delivery cycle, inventory level, fulfillment history, corporate relationship information and external public opinion data, etc., and the data is cleaned and standardized and then stored in the graph database to support the association relationship identification of the graph structure model construction module.

3. The system according to claim 1, wherein: The multi-layer supply chain structure model construction module constructs a supply chain association graph containing node entities and multiple types of association edges by integrating the business relationships and enterprise association relationships of core enterprises, suppliers at all levels and customer nodes; each association edge is attached with a weight to represent the node association strength, and a hierarchical difference value is attached to represent the hierarchical distance between nodes. The graph structure model supports dynamic updates to reflect changes in the supply chain structure.

4. The system according to claim 1, wherein: The risk propagation path analysis module includes: wherein the risk propagation path analysis module includes: a risk propagation probability calculation unit for calculating the risk from the occurrence node N based on the association weights and hierarchical differences between the nodes in the graph structure model i Propagate to target node N j The propagation probability P ij ; Risk impact measurement unit, used to calculate the risk impact value I of the risk transferred from the occurrence node to each target node according to the cumulative effect of the propagation probability along the supply chain path j ; and a multi-level risk path identification unit, used to identify the main risk propagation path from the occurrence node to the predetermined key node based on the risk impact value.

5. The system according to claim 1, wherein: The risk disturbance simulation module is used to accept the risk event node selected or preset by the user, set the initial risk level of the node to a predetermined value and trigger the risk propagation path analysis module to perform calculations; the risk disturbance simulation module also supports applying risk disturbances to multiple nodes at the same time to simulate the interactive propagation effect of risks in complex scenarios.

6. The system according to claim 1, wherein: The Motif anomaly detection module pre-stores several common subgraph patterns in the supply chain network and their normal frequency thresholds. After the risk disturbance simulation and propagation analysis are completed, the subgraph patterns involved in the actual risk propagation are extracted and compared with the normal threshold. If the frequency of occurrence of a subgraph pattern exceeds the threshold range, it is judged as an anomaly, and a corresponding anomaly score and anomaly pattern report are generated.

7. The system according to claim 1, wherein: The self-supervised graph neural network analysis module uses a graph neural network model to perform embedded representation learning on the supply chain association graph, and trains model parameters through self-supervised strategies such as contrastive learning or attribute reconstruction; the module identifies potential associations between supply chain nodes without manual labeling, and calculates risk anomaly metrics for each node to assist in discovering high-risk nodes or relationships that are not identified by traditional rule-based methods.

8. The system according to claim 1, wherein the risk radar heat map display module includes an interactive visual interface for displaying the risk indicators and risk transmission relationships of each supply chain node; the interface highlights the scores of key nodes under different risk dimensions through radar charts, displays the risk transmission probability between nodes through a heat map matrix, and supports users to select specific nodes or paths to obtain detailed risk data.

9. The system according to claim 1, wherein: The decision support module generates mitigation strategy recommendations for high-risk nodes or links based on the results of the risk propagation path analysis module and the anomaly detection module.

10. A method for analyzing engineering supply chain risk transmission based on graph neural network, characterized in that: Include: A. Collect the operation data of nodes at each level, clean it and store it in the graph database to generate the initial node relationship graph; B. Integrate business and equity information to build a multi-layer supply chain relationship graph; calculate the weight and level difference for each edge; C. Select a target node in the graph to inject the risk state and trigger the propagation calculation; D. Enumerate all propagation paths, calculate the cumulative probability and impact value, and locate the path that has the greatest impact on the key node; E. Perform Motif comparison to compare the actual subgraph with the benchmark distribution and generate anomaly alarms; F. Use self-supervised graph neural networks to generate node risk vectors to supplement potential association prediction and anomaly scoring; G. Display risk indicators and main paths using radar charts / heat maps, and automatically output mitigation suggestions.

Images