Industrial chain disturbance identification method and detection system based on GAT and AE

The GAT-AE model improves industrial chain disturbance detection by dynamically learning neighbor nodes and reconstructing structures and attributes, addressing low accuracy and timeliness issues in existing methods, ensuring efficient and accurate interruption identification.

CN118710082BActive Publication Date: 2024-12-03JINAN UNIVERSITY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410994955.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-12-03
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

Existing industrial chain interruption detection methods suffer from low accuracy and poor timeliness due to the use of manual experience and statistical approaches, inability to uniformly detect interruptions, and the complexity of graph data analysis, leading to inefficient risk prevention and delayed response to disturbances.

Method used

An industrial chain disturbance identification method and system based on Graph Attention Network (GAT) and Autoencoder (AE), utilizing an attribute network diagram to input adjacency and node parameters into a trained GAT-AE disturbance detection model for real-time interruption detection, dynamically learning neighbor nodes and reconstructing structures and attributes to improve detection accuracy.

Benefits of technology

Enhances the accuracy and timeliness of industrial chain disturbance identification by capturing structural and attribute characteristics, reducing the need for labeled data, and enabling robust detection of interruptions in complex networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118710082B_ABST
    Figure CN118710082B_ABST
Patent Text Reader

Abstract

The present application relates to an industrial chain disturbance identification method and detection system based on GAT and AE, the method comprising: obtaining an attribute network diagram of a target industrial chain to be detected, the attribute network diagram comprising adjacency parameters corresponding to all enterprise nodes on the target industrial chain and node parameters corresponding to each enterprise node, the adjacency parameters being used to characterize whether there is a supply relationship between enterprise nodes, and the node parameters being used to characterize production status information of each enterprise node; inputting all adjacency parameters corresponding to the target industrial chain and node parameters corresponding to all the enterprise nodes into a GAT‑AE disturbance detection model for interruption detection, and obtaining disturbance parameters corresponding to all enterprise nodes, the disturbance parameters comprising a disturbance index corresponding to the corresponding enterprise node; obtaining a corresponding disturbance index from the disturbance parameters corresponding to each enterprise node, and determining an industrial chain disturbance identification result based on a judgment result of whether the disturbance index is less than a disturbance index threshold.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of industrial chain collaboration technology, and in particular to an industrial chain disturbance identification method and detection system based on GAT and AE. Background Art

[0002] The disruption of the industrial chain often stems from the frequent occurrence of emergencies, the incorrect implementation of lean and just-in-time systems, and the complexity of the supply structure. Therefore, the industrial chain is inevitably disturbed by demand fluctuations, production / logistics time uncertainty, and unexpected accidents, which in turn lead to supply interruptions, delivery delays, and other problems, which in severe cases produce ripple effects.

[0003] In the related technologies, for the interruption detection of the industrial chain, the existing node interruption detection adopts manual experience and statistical methods to make judgments. Due to the low judgment accuracy and long judgment time, the interruption detection efficiency of the industrial chain is low and the risks caused by the interruption of the industrial chain cannot be effectively prevented. At the same time, in the related technologies, due to the suddenness and diversity of emergencies, as well as the uncertain impact intensity and impact range, it is impossible to quickly and effectively identify the interruption of the industrial chain. Moreover, because the disturbed industrial chain has network attribute characteristics, when the graph data generated by the industrial chain is analyzed and detected using the existing two-dimensional data analysis technology, the interruption generated by the industrial chain cannot be uniformly detected, and the accuracy of interruption detection is low and the timeliness is poor.

[0004] Currently, there are problems in the related technologies such as the detection methods and analysis technologies for industrial chain interruptions that are unable to uniformly detect interruptions, have low detection accuracy, and poor timeliness, and no effective solution has yet been proposed. Summary of the invention

[0005] The embodiments of the present application provide an industrial chain disturbance identification method and detection system based on GAT and AE, as well as a device and a storage medium, so as to at least solve the problem that the detection method and analysis technology of industrial chain interruption in the related technology cannot uniformly detect the interruption, has low detection accuracy and poor timeliness.

[0006] In a first aspect, an embodiment of the present application provides an industrial chain disturbance identification method based on GAT and AE, comprising: obtaining an attribute network diagram of a target industrial chain to be detected, wherein the attribute network diagram includes adjacency parameters corresponding to all enterprise nodes on the target industrial chain and node parameters corresponding to each enterprise node, the adjacency parameters are used to characterize whether there is a supply relationship between the enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node; all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all the enterprise nodes are input into a constructed GAT-AE disturbance detection model for interruption detection, and the disturbance corresponding to all the enterprise nodes is obtained. Parameters, wherein the disturbance parameters include the disturbance index corresponding to the corresponding enterprise node, the disturbance index is used to characterize the possibility of the enterprise node causing an interruption in the target industrial chain, and the GAT-AE disturbance detection model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD; from the corresponding disturbance parameters of each of the enterprise nodes, the corresponding disturbance index is obtained, and according to the judgment result of whether the disturbance index is less than the disturbance index threshold, the industrial chain disturbance identification result is determined, wherein the industrial chain disturbance identification result includes the enterprise node interrupted in the target industrial chain and the corresponding node parameters.

[0007] In a second aspect, an embodiment of the present application provides a detection system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for identifying industrial chain disturbances based on GAT and AE as described in the first aspect are implemented.

[0008] In a third aspect, an embodiment of the present application provides a storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the industrial chain disturbance identification method based on GAT and AE as described in the first aspect above are implemented.

[0009] Compared with the related art, the industrial chain disturbance identification method and detection system based on GAT and AE provided in the embodiment of the present application adopts the method of obtaining the attribute network diagram of the target industrial chain to be detected, wherein the attribute network diagram includes the adjacency parameters corresponding to all enterprise nodes on the target industrial chain and the node parameters corresponding to each enterprise node, the adjacency parameters are used to characterize whether there is a supply relationship between the enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node; all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all the enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all the enterprise nodes are obtained, the disturbance parameters include the disturbance index corresponding to the corresponding enterprise node, and the disturbance index is used to characterize the interruption caused by the enterprise node on the target industrial chain. The GAT-AE disturbance detection model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD; from the corresponding disturbance parameters of each enterprise node, the corresponding disturbance index is obtained, and the industrial chain disturbance identification result is determined according to the judgment result of whether the disturbance index is less than the disturbance index threshold, and the industrial chain disturbance identification result includes the interrupted enterprise node and the corresponding node parameters on the target industrial chain, which solves the problem that the detection method and analysis technology of the interruption of the industrial chain in the related technology cannot uniformly detect the interruption, the detection accuracy is low and the timeliness is poor, and the GAT-AE detection model is used to dynamically learn neighbor nodes, reconstruct structures and reconstruct attributes of the attribute network graph, so as to better capture the structure and attribute characteristics of the attribute network graph and improve the accuracy of detecting interrupted nodes. At the same time, the deep learning algorithm is used to train the GAT-AE target detection model to improve the effect and robustness of industrial chain disturbance identification. In addition, structural reconstruction and attribute reconstruction are used to detect abnormal information inherent in the attribute network graph of the industrial chain, and the structure and attributes of the attribute network graph are learned in an unsupervised manner, which reduces the difficulty of training large-scale attribute networks and the dilemma of lack of labeled data, and achieves the beneficial effect of accurately detecting interruptions in industrial chain nodes in the case of unlabeled data and complex attribute network graphs.

[0010] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0012] Figure 1It is a hardware structure block diagram of a terminal of the industrial chain disturbance identification method based on GAT and AE in an embodiment of the present application;

[0013] Figure 2 is a flow chart of an industrial chain disturbance identification method based on GAT and AE according to an embodiment of the present application;

[0014] Figure 3 is a block diagram of a GAT-AE disturbance detection model of a preferred embodiment of the present application;

[0015] Figure 4 It is a structural block diagram of an industrial chain disturbance identification device based on GAT and AE according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for ordinary technicians in the field related to the contents disclosed in the present application, some changes such as design, manufacturing or production based on the technical contents disclosed in the present application are only conventional technical means, and should not be understood as insufficient contents disclosed in the present application.

[0017] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0018] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantitative limitation and may represent the singular or plural. The terms "including", "comprising", "having" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include unlisted steps or units, or may also include other steps or units inherent to these processes, methods, products or devices. The "multiple links" involved in this application refer to links greater than or equal to two. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" may represent: A exists alone, A and B exist at the same time, and B exists alone. The terms "first", "second", "third" and the like involved in this application are only used to distinguish similar objects and do not represent a specific ordering of objects.

[0019] The method embodiment provided in this embodiment can be executed in a terminal, a computer or a similar computing device. Taking running on a terminal as an example, Figure 1 1 is a hardware structure block diagram of a terminal of the industrial chain disturbance identification method based on GAT and AE in an embodiment of the present application. Figure 1 As shown, the terminal may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Optionally, the terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is for illustration only and does not limit the structure of the above terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0020] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the industrial chain disturbance identification method based on GAT and AE in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories can be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0021] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0022] This embodiment provides a method for identifying industrial chain disturbances based on GAT and AE running on the above terminal. Figure 2 is a flow chart of a method for identifying industrial chain disturbances based on GAT and AE according to an embodiment of the present application, such as Figure 2 As shown, the process includes the following steps:

[0023] Step S201, obtain the attribute network diagram of the target industrial chain to be detected, wherein the attribute network diagram includes adjacency parameters corresponding to all enterprise nodes on the target industrial chain and node parameters corresponding to each enterprise node, the adjacency parameters are used to characterize whether there is a supply relationship between enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node.

[0024] In this embodiment, the attribute network diagram is real-time enterprise data collected through the information system of each enterprise node in the industrial chain. Each enterprise node in the industrial chain can input the enterprise data into the information system through the Internet of Things (IoT) device or manual uploading. In this embodiment, IoT devices with communication capabilities are selected, such as temperature sensors, humidity sensors, pressure sensors, Arduino and special manufacturing equipment. At the same time, the data collected by IoT devices and manually recorded are pre-processed, such as data cleaning and format conversion. After that, the pre-processed data is sent to the IoT device based on the preset communication protocol (such as Wi-Fi, Bluetooth, Zigbee, LoRa, NB-IoT). The processed data is transmitted to the enterprise's information system in real time, and is saved, processed, visualized and monitored in JSON, XML, and CSV data formats; subsequently, the enterprise's data is screened and aggregated locally to reduce the amount of data transmission, and the key data is transmitted to the industrial chain through data transmission protocols such as MQTT, HTTP / HTTPS, CoAP, and AMQP. SSL / TLS encryption technology can be used during the transmission process to ensure data security; in this embodiment, when the corresponding processing system is a cloud platform, and after the cloud platform receives the data (attribute network diagram), the data will be clarified, filtered and preprocessed. Specifically, a stream processing framework (for example, Apache Kafka and Apache Flink) is used to implement real-time data processing and analysis, and a batch processing framework (for example, Apache Hadoop and Apache Spark) is used to analyze historical data.

[0025] In this embodiment, the attribute network diagram uploaded by each enterprise node to the industrial chain includes adjacency parameters representing the structural data between the enterprise nodes and node parameters representing the attribute data of each enterprise node, wherein the adjacency parameters can be represented by an adjacency matrix, which is a square matrix used to represent the connection relationship between the enterprise nodes in the attribute network diagram. It should be noted that in the upstream and downstream enterprise relationships in the industrial chain, each node represents an enterprise, and each edge represents a certain relationship between enterprises. In the adjacency matrix, each enterprise node corresponds to a row and a column in the adjacency matrix. The elements in the matrix represent the relationship between the two enterprises. The relationship between the two enterprise nodes includes supply relationship, customer relationship, and cooperation relationship. In this embodiment, it specifically refers to the supply relationship. In this embodiment, the corresponding adjacency matrix is ​​a binary adjacency matrix. In the adjacency matrix, if there is a relationship between enterprise node i and enterprise node j, the matrix element is 1 if the value is true, otherwise it is 0.

[0026] In this embodiment, the node parameters are the attribute data of each enterprise node. The attribute data of the enterprise node include key indicators reflecting the operation status, production situation and environment of the enterprise node, such as: basic information of the enterprise node, operation status, production data, environmental detection, energy consumption data, logistics data, and safety data. The basic information includes enterprise ID, enterprise name, enterprise type (for example, manufacturing enterprise, supplier, distributor, retailer, etc.), and enterprise location; the operation status includes operation status (for example, normal operation of the enterprise, shutdown of the enterprise, and failure of the enterprise), production capacity (indicating the production capacity of the enterprise, for example, the number of production lines, the number of equipment, and production capacity of the enterprise), and equipment status (for example, operation, maintenance, and failure records of the equipment); production data includes production volume (production volume per unit time), product quality (qualified rate, defective rate of the product), and product quality (qualified rate, defective rate of the product). ), raw material inventory (current raw material inventory and its changes), finished product inventory (current finished product inventory and its changes); environmental detection includes ambient temperature, humidity, and air quality; energy consumption data includes water and electricity consumption; logistics data includes logistics in and out, and upstream and downstream information of the supply chain; safety data includes safety incident records and personnel safety; in this embodiment, node parameters are saved in the form of vectors / lists, and the attribute content of node parameters is stored in numerical values. For example: there are 3 enterprise nodes, enterprise node A produces 10 products without quality problems; enterprise node B produces 5 products with quality problems; enterprise node C produces 20 products without quality problems, then the node parameters in the network attribute graph are represented by the following list / vector, h=[[10,0], [5,1], [20,0]].

[0027] Step S202, all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all enterprise nodes are obtained, wherein the disturbance parameters include the disturbance index corresponding to the corresponding enterprise node, and the disturbance index is used to characterize the possibility of the enterprise node causing an interruption in the target industrial chain. The GAT-AE disturbance detection model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD.

[0028] In this embodiment, GAT-AE is a model based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD, and is unsupervisedly trained based on a data set of historical attribute network graphs, wherein unsupervised training means that the label information of the attribute network graph is not required, wherein the data set collects attribute network graphs of different time periods and different industrial chain structures; in this embodiment, the attribute network graph collected from the industrial chain is inferred and detected through the GAT-AE perturbation detection model to obtain the perturbation parameters corresponding to all enterprise nodes.

[0029] Step S203, obtaining the corresponding disturbance index from the disturbance parameters corresponding to each enterprise node, and determining the industrial chain disturbance identification result based on the judgment result of whether the disturbance index is less than the disturbance index threshold, wherein the industrial chain disturbance identification result includes the interrupted enterprise nodes and corresponding node parameters on the target industrial chain.

[0030] In this embodiment, after the GAT-AE disturbance detection model obtains the disturbance parameters corresponding to all enterprise nodes, the disturbance index of each enterprise node (characterizing the risk of interruption of the corresponding enterprise node) is determined for the disturbance parameters corresponding to each enterprise node, and then the difference between the preset disturbance index threshold and the current disturbance index is used to determine whether the corresponding enterprise node is an interrupted enterprise node or a normal node; in this embodiment, the detection result aims to detect the interrupted enterprise nodes in the target industrial chain, and adjust the interrupted enterprise nodes, while maintaining the status quo for normal enterprise nodes, so that the entire target industrial chain maintains a stable supply state, and thus efficiently implements production.

[0031] Through the above steps S201 to S203, an attribute network diagram of the target industrial chain to be detected is obtained, and the attribute network diagram includes adjacency parameters corresponding to all enterprise nodes on the target industrial chain and node parameters corresponding to each enterprise node. The adjacency parameters are used to characterize whether there is a supply relationship between enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node; all adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all enterprise nodes are obtained. The disturbance parameters include the disturbance index corresponding to the corresponding enterprise node. The disturbance index is used to characterize the possibility of an enterprise node causing an interruption in the target industrial chain. The GAT-AE disturbance detection model is based on The neural network model trained by the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD; the corresponding disturbance index is obtained from the disturbance parameters corresponding to each enterprise node, and the industrial chain disturbance identification result is determined according to the judgment result of whether the disturbance index is less than the disturbance index threshold. The industrial chain disturbance identification result includes the interrupted enterprise nodes and corresponding node parameters on the target industrial chain, which solves the problem that the detection method and analysis technology of the industrial chain interruption in the related technology cannot uniformly detect the interruption, has low detection accuracy and poor timeliness. The GAT-AE detection model is used to dynamically learn neighbor nodes, reconstruct structures and reconstruct attributes of the attribute network graph, so as to better capture the structure and attribute characteristics of the attribute network graph and improve the accuracy of detecting interrupted nodes. At the same time, the deep learning algorithm is used to train the GAT-AE target detection model to improve the effect and robustness of industrial chain disturbance identification. In addition, structural reconstruction and attribute reconstruction are used to detect abnormal information inherent in the attribute network graph of the industrial chain, and the structure and attributes of the attribute network graph are learned in an unsupervised manner, which reduces the difficulty of training large-scale attribute networks and the dilemma of lack of labeled data, and achieves the beneficial effect of accurately detecting interruptions in industrial chain nodes in the case of unlabeled data and complex attribute network graphs.

[0032] It should be noted that the disturbance detection model of the embodiment of the present application consists of three parts: a graph attention network encoder GAT, a structure reconstruction decoder SRE and an attribute reconstruction decoder ARD. In this embodiment, the attribute network graph of an industrial chain is input into the constructed GAT-AE disturbance detection model. After being input into the model, the attribute network graph first enters the GAT to aggregate information on the attribute network graph and outputs an embedding representation vector Embedding. After that, the Embedding is respectively input into the SRE and ARD to reconstruct the network structure and attributes. Finally, the disturbance parameters of the enterprise nodes of the target industrial chain are output (represented by the interruption sorting list DRL). Through the disturbance parameters of all enterprise nodes, it is determined whether the enterprise nodes are interrupted.

[0033] It should be noted that in the embodiment of the present application, after the detection of the industrial chain is completed, the detection results are streamed to the server in real time and displayed in real time on the server's web page; in this embodiment, the real-time streaming transmission uses Python's Flask framework and UDP protocol to transmit the node data that has completed the interruption detection in real time; in the real-time data transmission process, the display effect is optimized through icon highlighting and push alerts, for example: the icon of the node that detects the interruption is highlighted, and the information is pushed to other companies with supply relationships for early warning; in this embodiment, visualization tools are also used to display real-time data and analysis results.

[0034] It should be further explained that the embodiment of the present application adopts a deep learning algorithm to replace the traditional manual or statistical methods, so as to achieve a detection effect that exceeds the traditional method and has stronger robustness. In addition, the algorithm is an end-to-end network, which reduces the difficulty of algorithm design and training and has high flexibility in industrial applications. At the same time, in order to solve the difficulties of attribute network structure, attribute learning and few labels, an unsupervised GAT-AE detection algorithm is adopted, which can realize accurate detection of nodes in the case of insufficient label data and complex attribute network graph structure.

[0035] In some embodiments, all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all enterprise nodes are obtained, which is achieved by the following steps:

[0036] Step 21, based on the adjacency parameters corresponding to all enterprise nodes and the node parameters corresponding to each enterprise node, GAT is used to iterate the adjacency matrix calculation and information aggregation operations until an information vector corresponding to each enterprise node is generated.

[0037] In this embodiment, after the attribute network diagram corresponding to the target industrial chain is input into the GAT-AE disturbance detection model, information aggregation is first performed through GAT to generate a corresponding information vector.

[0038] Step 22, through SRE, the information vector corresponding to each enterprise node is structurally reconstructed to generate the reconstructed adjacency parameters corresponding to each enterprise node, and through ARD, the information vector corresponding to each enterprise node is attribute reconstructed to generate the reconstructed node parameters corresponding to each enterprise node.

[0039] In this embodiment, after information aggregation generates information vectors, the information vectors are respectively input into SRE and ARD to reconstruct the network structure and attributes, and the reconstructed adjacency parameters and reconstructed node parameters corresponding to the enterprise nodes of the target industrial chain are output.

[0040] Step 23, after determining the first deviation value between each reconstructed adjacency parameter and the corresponding adjacency parameter and determining the second deviation value between each reconstructed node parameter and the corresponding node parameter, weight the first deviation value and the second deviation value corresponding to each enterprise node to generate a disturbance parameter corresponding to each enterprise node.

[0041] In this embodiment, after performing structural reconstruction and attribute reconstruction through SRE and ARD respectively, and generating reconstructed adjacency parameters and reconstructed node parameters, the deviation between the reconstructed parameters and the corresponding parameters is determined, and then the disturbance parameters corresponding to each enterprise node are determined, that is, the risk of interruption of the enterprise node is determined.

[0042] In this embodiment, it is necessary to use the reconstruction adjacency parameter and refactor node parameters To calculate the disturbance parameters, the calculation formula of the disturbance parameters is as follows:

[0043]

[0044] in is the adjacent vector The i-th row vector of Node parameters The information vector of the i-th enterprise node, represents the perturbation parameters of the i-th enterprise node, that is, the dissimilarity between the reconstructed result (reconstructed adjacency parameters and reconstructed node parameters) and the original / true result (adjacency parameters and node parameters), represents the corresponding weighted weight, It means to subtract the vector x from the vector y and find the square of the norm.

[0045] Through the above steps, based on the adjacency parameters corresponding to all enterprise nodes and the node parameters corresponding to each enterprise node, GAT is used to iterate the adjacency matrix calculation and information aggregation operations until the information vector corresponding to each enterprise node is generated; through SRE, the information vector corresponding to each enterprise node is structurally reconstructed to generate the reconstructed adjacency parameters corresponding to each enterprise node, and through ARD, the information vector corresponding to each enterprise node is attribute reconstructed to generate the reconstructed node parameters corresponding to each enterprise node; after determining the first deviation value between each reconstructed adjacency parameter and the corresponding adjacency parameter and determining the second deviation value between each reconstructed node parameter and the corresponding node parameter, the first deviation value and the second deviation value corresponding to each enterprise node are weighted to generate the disturbance parameter corresponding to each enterprise node, so as to detect the disturbance parameters generated by the enterprise nodes on the target industrial chain based on the attribute network diagram, and then determine whether the enterprise node is interrupted.

[0046] In some embodiments, step 21 is implemented by using GAT to iterate adjacency matrix calculation and information aggregation operations based on adjacency parameters corresponding to all enterprise nodes and node parameters corresponding to each enterprise node until an information vector corresponding to each enterprise node is generated, which is implemented by the following steps:

[0047] Step 211, determine the first weight parameter and the preset first attention mechanism parameter corresponding to the neural network layer for the current adjacency matrix calculation and information aggregation operation of GAT, wherein the first weight parameter is used to characterize the weight matrix of the corresponding neural network layer.

[0048] Step 212, traverse the adjacent enterprise nodes associated with each enterprise node from all enterprise nodes, and determine the adjacent enterprise node group consisting of the enterprise node and an adjacent enterprise node, and determine the first node parameter corresponding to the target enterprise node in all adjacent enterprise node groups, the target enterprise node includes the enterprise node in the corresponding adjacent enterprise node group and one of the adjacent enterprise nodes, the first node parameter is generated by information aggregation of the historical adjacent parameters generated by all adjacent enterprise node groups associated with the enterprise node in the previous iteration operation and the historical node parameters corresponding to all adjacent enterprise nodes associated with the enterprise node.

[0049] Step 213, according to the first node parameters, first weight parameters and first attention mechanism parameters of the two target enterprise nodes corresponding to each adjacent enterprise node group associated with the corresponding enterprise node, vector concatenation, matrix inversion and power operation are performed to obtain the first relationship value corresponding to each adjacent enterprise node group, and after determining the first total relationship value corresponding to all adjacent enterprise node groups associated with the corresponding enterprise node according to the first relationship value corresponding to each adjacent enterprise node group, based on the ratio of the first relationship value corresponding to the target adjacent enterprise node group to the first total relationship value, generate the correlation coefficient corresponding to the current adjacency matrix calculation of the target adjacent enterprise node group, wherein the adjacency parameters include the correlation coefficient, the first relationship value is used to characterize the correlation value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group, the target adjacent enterprise node group includes one of all adjacent enterprise node groups associated with the corresponding enterprise node, and the correlation coefficient is used to characterize the degree of correlation between the corresponding two enterprise nodes;

[0050] Step 214, based on the correlation coefficient, the first weight parameter and the first node parameter corresponding to each target adjacent enterprise node group, information aggregation is performed to generate a current information vector corresponding to the current information aggregation operation of the enterprise node corresponding to the target adjacent enterprise node group.

[0051] Step 215, after the number of steps of generating the corresponding correlation coefficient and generating the corresponding current information vector has reached a preset number of times, the current information vector generated for the last time is used as the information vector corresponding to the corresponding enterprise node.

[0052] In this embodiment, the GAT-AE disturbance detection model first calculates the adjacency matrix of the industrial chain , the adjacency matrix Each element of is calculated by the relationship coefficient between the two enterprise nodes. If there is no supply relationship between the two enterprise nodes, the value is 0. If there is a supply relationship between the two enterprise nodes, the specific value is calculated using the following relationship coefficient calculation formula:

[0053]

[0054]

[0055] in, represents the correlation coefficient between enterprise node i and enterprise node j in the first neural network layer of GAT, represents the initial first node parameter of enterprise node i, N i represents the number of adjacent enterprise nodes of enterprise node i, W (1) represents the first weight parameter of the first neural network layer of GAT, a (1) represents the first attention mechanism parameter of the first neural network layer of GAT, || represents the vector concatenation operation, T represents the matrix inversion, LeakyReLU(x)=max( ,x), is a constant, and 0< <1,exp represents exponentiation.

[0056] After calculating the industry chain adjacency matrix Afterwards, the adjacency matrix is ​​aggregated using GAT and the first node parameters of each enterprise node The information of the enterprise node i is used to generate the current information vector of the first neural network layer of GAT, which is also the corresponding first node parameter. In this embodiment, the adjacency matrix of the industrial chain is calculated twice, that is, the correlation coefficient of the first neural network layer of GAT is generated respectively. Correlation coefficient with the second neural network layer of GAT , and the corresponding information aggregation is also performed twice, that is, based on the original node parameters of enterprise node i Calculate the nodes separately The first node parameters of the first neural network layer in GAT and Enterprise Node The first node parameters of the second neural network layer in GAT , specifically,

[0057]

[0058]

[0059]

[0060] in, represents the correlation coefficient between enterprise node i and enterprise node j in the second neural network layer of GAT, N represents the first node parameter of enterprise node i in the second neural network layer of GAT, i represents the number of adjacent enterprise nodes of enterprise node i, W (2) represents the first weight parameter of the second neural network layer of GAT, a (2) represents the first attention mechanism parameter of the second layer of the GAT neural network layer, || represents the vector concatenation operation, T represents the matrix inversion, LeakyReLU(x)=max( ,x), is a constant, and 0< <1, exp means power operation, ReLU(x)=max(0,x), is the i-node vector of the intermediate vector embedding B. In this embodiment, the purpose of information aggregation is to identify deeper information of these nodes to help determine the state of each node and provide information support for subsequent determination of whether a node has occurred.

[0061] In this embodiment, after performing two information aggregations, the corresponding generated current information vector will be used as the final information vector corresponding to the corresponding enterprise node, that is, As the information vector corresponding to the corresponding enterprise node.

[0062] Through the above steps 211 to 215, GAT is used to calculate the correlation coefficients of the graph attention network twice and to perform information aggregation twice, thereby generating information vectors corresponding to each enterprise node.

[0063] In some embodiments, step 22 of reconstructing the information vector corresponding to each enterprise node by SRE to generate the reconstructed adjacency parameter corresponding to each enterprise node includes the following steps:

[0064] Step 221, determine the second weight parameter and the preset second attention mechanism parameter corresponding to the neural network layer for structural reconstruction of SRE, wherein the second weight parameter is used to characterize the weight matrix of the neural network layer corresponding to SRE.

[0065] In this embodiment, the generation of the corresponding correlation coefficient and the generation of the corresponding current information vector are repeated until the final information vector is generated. Then, using SRE, and based on the information vector Before performing structural reconstruction, the second weight parameter and the second attention mechanism parameter corresponding to the neural network layer for structural reconstruction by SRE are first determined.

[0066] Step 222, according to the information vectors of the two target enterprise nodes corresponding to each corresponding adjacent enterprise node group, the second weight parameter and the second attention mechanism parameter, vector concatenation, matrix inversion and power operation are performed to obtain the second relationship value corresponding to each corresponding adjacent enterprise node group, and after determining the second total relationship value corresponding to all adjacent enterprise node groups associated with the corresponding enterprise node according to the second relationship value corresponding to each adjacent enterprise node group, a reconstruction correlation coefficient corresponding to the structural reconstruction of the target adjacent enterprise node group is generated based on the ratio of the second relationship value corresponding to the target adjacent enterprise node group to the second total relationship value, wherein the second relationship value is used to characterize the correlation value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group, and the reconstruction correlation coefficient is used to characterize the degree of correlation between the corresponding two enterprise nodes after the structural reconstruction.

[0067] In this embodiment, after determining the second weight parameter and the second attention mechanism parameter, based on the information vector Reconstruct the corresponding structure, that is, use the information vector Information reconstruction / reconstruction correlation coefficient of the predicted industrial chain , the operations of its structural reconstruction are as follows:

[0068] in, represents the reconstruction correlation coefficient of enterprise node i and enterprise node j in the neural network layer corresponding to SRE, represents the information vector generated by GAT for enterprise node i, N i represents the number of adjacent enterprise nodes of enterprise node i, W (3) Represents the second weight parameter corresponding to the neural network layer for SRE structural reconstruction, a (3) represents the second attention mechanism parameter corresponding to the neural network layer for SRE structural reconstruction, || represents the vector concatenation operation, T represents the matrix inversion, LeakyReLU(x)=max( ,x), is a constant, and 0< <1,exp represents exponentiation.

[0069] Step 223, based on the reconstruction correlation coefficient, the second weight parameter and the corresponding information vector corresponding to each target adjacent enterprise node group, information aggregation is performed to generate corresponding structural vector information, and the structural vector information is decoded through SRE to generate the reconstruction adjacent parameters corresponding to each enterprise node, wherein the structural vector information is used to characterize the information vector generated by the structural reconstruction of the enterprise nodes corresponding to the corresponding target adjacent enterprise node group.

[0070] In this embodiment, when generating the reconstruction correlation coefficient Afterwards, the correlation coefficient was reconstructed based on and the corresponding information vector Perform information aggregation to generate structural vector information In this embodiment, the following formula is used to calculate :

[0071]

[0072] in, Represents the structural vector information of enterprise node i, represents the reconstruction correlation coefficient of enterprise node i and enterprise node j in the neural network layer corresponding to SRE, W (3) ReLU(x) = max(0,x) represents the second weight parameter corresponding to the neural network layer for SRE structural reconstruction. is the i-node vector of the intermediate vector embedding B.

[0073] In this embodiment, when generating structure vector information Afterwards, SRE is used to analyze the structure vector information Decode and generate the reconstruction adjacency parameters corresponding to each enterprise node , decoded according to the following formula: , Z = , T represents matrix inversion, and sigmoid represents the activation function of the neural network corresponding to SRE.

[0074] Through the above steps 221 to 223, the information vector generated by SRE and the graph attention network encoder is realized. Perform structural reconstruction to generate corresponding structural vector information , providing parameters that characterize the adjacency relationship in order to determine the disturbance parameters corresponding to each enterprise node.

[0075] In some embodiments, step 23 reconstructs the attributes of the information vector corresponding to each enterprise node through ARD to generate the reconstructed node parameters corresponding to each enterprise node, including the following steps:

[0076] Step 231, determine the target neural network layer for attribute reconstruction by ARD, and determine the third weight parameter corresponding to the corresponding target neural network layer and the preset third attention mechanism parameter, wherein the third weight parameter is used to characterize the weight matrix of the corresponding target neural network layer, and the target network neural layer includes a first neural network layer and a second neural network layer.

[0077] In this embodiment, before using ARD to reconstruct attributes, the number of neural network layers used by ARD to reconstruct attributes, as well as the weight parameters and attention mechanism parameters corresponding to each neural network layer are first determined. In this embodiment, ARD will perform two correlation coefficient calculations and two information aggregations to match the attribute reconstruction. It can be understood that the target neural network layer for ARD to reconstruct attributes includes a first neural network layer and a second neural network layer, and the weight parameters corresponding to the first neural network layer and the second neural network layer are both defined as third weight parameters, and the attention mechanism parameters corresponding to the first neural network layer and the second neural network layer are both defined as third attention mechanism parameters, but the specific parameter values ​​will be set according to different neural network layers.

[0078] Step 232, according to the information vectors of the two target enterprise nodes corresponding to each corresponding adjacent enterprise node group, the third weight parameters and the third attention mechanism parameters respectively corresponding to the first neural network layer, perform vector concatenation, matrix inversion and power operation corresponding to the first neural network layer to obtain the third relationship value corresponding to each corresponding adjacent enterprise node group in the first neural network layer, and after determining the third total relationship value corresponding to the first neural network layer of all adjacent enterprise node groups associated with the corresponding enterprise node according to the third relationship value corresponding to each adjacent enterprise node group, based on the ratio of the second relationship value corresponding to the target adjacent enterprise node group to the third total relationship value, generate the first correlation coefficient generated by the attribute reconstruction of the target adjacent enterprise node group in the first neural network layer, wherein the third relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group in the first neural network layer, and the first correlation coefficient is used to characterize the degree of correlation between the corresponding two enterprise nodes after the attribute reconstruction in the first neural network layer.

[0079] In this embodiment, after determining the third weight parameter and the third attention mechanism parameter corresponding to the first neural network layer, based on the information vector Reconstruct the attributes of enterprise nodes, that is, use information vector Reconstruct / predict the first correlation coefficient of each enterprise node in the industrial chain , the attribute reconstruction operation is as follows:

[0080]

[0081] in, represents the first correlation coefficient between enterprise node i and enterprise node j in the first neural network layer, represents the information vector generated by GAT for enterprise node i, N i represents the number of adjacent enterprise nodes of enterprise node i, represents the third weight parameter corresponding to the first neural network layer, represents the third attention mechanism parameter corresponding to the first neural network layer, || represents the vector concatenation operation, T represents the matrix inversion, LeakyReLU(x)=max( ,x), is a constant, and 0< <1,exp represents exponentiation.

[0082] Step 233, after performing information aggregation based on the first correlation coefficient, the third weight parameter and the corresponding information vector corresponding to each target adjacent enterprise node group to generate the first attribute vector information, perform corresponding attribute reconstruction and information aggregation based on the first attribute vector information, the third weight parameter and the third attention mechanism parameter corresponding to the second neural network layer respectively to generate the second attribute vector information, and use the second attribute vector information as the reconstructed node parameter corresponding to each enterprise node, wherein the first attribute vector information and the second attribute vector information respectively represent the information vector generated by the attribute reconstruction of the enterprise node corresponding to the corresponding target adjacent enterprise node group in the corresponding target neural network layer.

[0083] In this embodiment, when generating the first correlation coefficient Afterwards, based on the first correlation coefficient and the corresponding information vector Perform information aggregation to generate the first attribute vector information In this embodiment, the following formula is used to calculate

[0084]

[0085] in, Represents the first attribute vector information of enterprise node i, represents the first correlation coefficient between enterprise node i and enterprise node j in the first neural network layer, ReLU(x)=max(0,x) represents the third weight parameter corresponding to the first neural network layer. is the i-node vector of the intermediate vector embedding B.

[0086] In this embodiment, the first attribute vector information is generated Afterwards, based on the first attribute vector information Reconstruct the attributes of the enterprise node, that is, use the first attribute vector information Generate the second correlation coefficient corresponding to each enterprise node , the attribute reconstruction operation is as follows:

[0087]

[0088] in, represents the second correlation coefficient between enterprise node i and enterprise node j in the second neural network layer, Represents the first attribute vector information corresponding to enterprise node i, N i represents the number of adjacent enterprise nodes of enterprise node i, represents the third weight parameter corresponding to the second neural network layer, represents the third attention mechanism parameter corresponding to the second neural network layer, || represents the vector concatenation operation, T represents the matrix inversion, LeakyReLU(x)=max( ,x), is a constant, and 0< <1,exp represents exponentiation.

[0089] In this embodiment, when generating the second correlation coefficient Afterwards, based on the second correlation coefficient and the first attribute vector information Perform information aggregation to generate the second attribute vector information In this embodiment, the following formula is used to calculate

[0090]

[0091] in, The second attribute vector information of the output is also used as the reconstruction node parameter , Or the prediction result of the initial node parameter h of each enterprise node, if and If the two data sets are very similar, it means that the probability of the risk of disruption in the current industrial chain may be lower.

[0092] Through the above steps 231 to 233, the information vector generated by the ARD and the graph attention network encoder is realized. Perform attribute reconstruction to generate corresponding reconstruction node parameters , providing parameters that characterize node attributes in order to determine the disturbance parameters corresponding to each enterprise node.

[0093] In some embodiments, according to the result of judging whether the disturbance index is less than the disturbance index threshold, the result of identifying the disturbance of the industrial chain is determined by the following steps:

[0094] Step 31, determining whether the disturbance index is less than a disturbance index threshold.

[0095] Step 32, when it is determined that the disturbance index is not less than the disturbance index threshold, the corresponding enterprise node is determined to be an industrial chain interruption node, wherein the industrial chain disturbance identification result includes the industrial chain interruption node.

[0096] Step 33: When it is determined that the disturbance index is less than the disturbance index threshold, the corresponding enterprise node is determined to be a normal node.

[0097] Through the above steps, it is determined whether the disturbance index is less than the disturbance index threshold; when it is determined that the disturbance index is not less than the disturbance index threshold, the corresponding enterprise node is determined to be an industrial chain interruption node, wherein the industrial chain disturbance identification result includes the industrial chain interruption node; when it is determined that the disturbance index is less than the disturbance index threshold, the corresponding enterprise node is determined to be a normal node, thereby realizing the judgment of the interruption of the enterprise node on the target industrial chain.

[0098] In some of the embodiments, after determining the result of the industrial chain disturbance identification, the following steps are further implemented: in the attribute network diagram, the industrial chain interruption nodes are marked according to a preset marking format, and the attribute network diagram is updated.

[0099] In this embodiment, after determining the status of each enterprise node (interrupted or normal), the corresponding icon position of the interrupted enterprise node in the target industrial chain is highlighted in a preset highlight format (for example: the interrupted node icon becomes a red glowing style), and the information of the interrupted node is shared with the entire industrial chain for early warning.

[0100] In some embodiments, before using GAT to iterate adjacency matrix calculation and information aggregation operations based on adjacency parameters corresponding to all enterprise nodes and node parameters corresponding to each enterprise node, the following steps are further implemented:

[0101] Step 41, determining the parameter value corresponding to the adjacency parameter, and judging whether the parameter value is a target parameter value, wherein the target parameter value is used to indicate that there is no supply relationship between the corresponding enterprise nodes.

[0102] Step 42, when it is determined that the parameter value is the target parameter value, the adjacency parameter and the node parameter are used as the corresponding reconstructed adjacency parameter and the reconstructed node parameter respectively, and it is determined that the corresponding enterprise node is currently an interrupted node in the industrial chain.

[0103] Figure 3is a block diagram of the GAT-AE disturbance detection model of the preferred embodiment of the present application, refer to Figure 3 The following is an explanation of the training process of the GAT-AE disturbance detection model of the preferred embodiment of the present application. The training includes the following steps:

[0104] Step 1: Collect the dataset of historical attribute network graph from the industrial chain.

[0105] In this embodiment, the data set includes multiple attribute network diagrams, each of which includes node numbers, node attributes, and adjacency matrices, wherein the node number is used to uniquely number each enterprise node; the node attribute is the characteristic data of each enterprise node, which is stored in the form of a vector, and each element in the vector represents the characteristics of the node, for example, the type and location coordinates of the enterprise node, production status information, logistics status information, storage status information, surrounding emergency information, etc. The vectors corresponding to the characteristic data constitute an attribute matrix of "the number of all nodes × the length of each node vector"; the adjacency matrix represents the connection relationship between nodes, which is represented by a matrix of "the number of nodes × the number of nodes", and the adjacency matrix represents the connection relationship between nodes. The elements in the matrix are represented by 01 codes, where 1 indicates that there is a connection relationship (edge) between the enterprise node indexed by the column and the enterprise node indexed by the row. For example, 1 indicates that there is a logistics relationship, a supply relationship, and information transmission between the enterprise node indexed by the column and the enterprise node indexed by the row. In this embodiment, the data of the attribute network graph is processed using a python interface, and specifically a third-party open source module torch_geometric is used to process the attribute network graph, and the data set of the attribute network graph is converted into an iterator that can extract data. The data extraction iterator is used to extract the data set in various forms, for example, extracting one or more attribute network graphs in the data set in sequence / randomly, which is convenient for subsequent training.

[0106] Step 2: After preparing the data set of attribute network graphs and their data extraction iterators, each attribute network graph in the data set trains the GAT-AE perturbation detection model, taking the extraction of one attribute network graph from the data extraction iterator as an example:

[0107] Step 2.1, obtain an attribute network graph, and input the attribute network graph into the GAT-AE disturbance detection model. The attribute network graph first enters the attribute network encoding module ANE of the GAT-AE disturbance detection model, and the encoding module is composed of a multi-layer GAT network.

[0108] Step 2.2, the attribute network graph first enters the first layer of GAT network, and the first layer of GAT network performs the first information aggregation on the attribute network graph, wherein the corresponding information aggregation refers to the use of the attention mechanism to dynamically learn the attribute information of the neighbor nodes of each node, that is, the attribute information of the neighbor nodes is aggregated into the node in the form of weighted sum using the weights learned by the attention mechanism, and the neighbor nodes refer to other nodes connected to the node (obtained from the adjacency matrix in the attribute network graph), and the attribute information refers to the attribute matrix of "the number of all nodes × the length of the node attribute vector" composed of the attributes of each node. GAT can capture the characteristics and structure of the attribute network graph; in addition, GAT can learn the weights between different nodes, making the model more flexible and applicable to different types of attribute network graphs; after the attribute network graph enters the first GAT network, the intermediate vector of each enterprise node is output through the ReLU function, and the length of the intermediate vector is preset, and the corresponding intermediate vector constitutes an intermediate matrix of "the number of nodes × the length of the intermediate vector", which is the final output of the first GAT.

[0109] Step 2.3, input the intermediate matrix into the second GAT network, and the second GAT network performs another information aggregation. The corresponding information aggregation is the same as the first GAT. In this embodiment, the second GAT network performs another information aggregation based on the first GAT information aggregation, so that each enterprise node aggregates the attribute information of a wider range of related nodes, which helps ANE to better learn the structure and attribute information of the entire attribute network graph; in this embodiment, the corresponding structure refers to the connection relationship between enterprise nodes, that is, the adjacency matrix. According to the pre-set requirements, more GAT networks can be added after the second GAT network to achieve the purpose of sufficient information aggregation; after the last GAT network, the embedding vectors of all enterprise nodes are output. The embedding vector refers to the mapping of the industrial chain node that has learned the neighbor node information to a vector representation of a low-dimensional space. The vector dimension is preset. The embedding vector is a low-dimensional space vector representation rich in attribute network structure and attribute information after multiple GAT information aggregations. The vector can provide support for subsequent calculation operations. The embedding vectors of all enterprise nodes constitute an embedding matrix of "all node numbers × embedding vector length", which is also the output of the ANE module.

[0110] In step 2.4, the output of ANE (corresponding embedding matrix) is input into the structure reconstruction decoder SRE module and the attribute reconstruction decoder ARD module in parallel.

[0111] In this embodiment, the SRE module reconstructs the adjacency matrix of the current attribute network graph according to the embedded matrix information, and the ARD module reconstructs all node attributes of the current attribute network graph according to the embedded matrix information, that is, reconstructs the node attribute matrix.

[0112] In this embodiment, the SDR module specifically operates as follows:

[0113] a. Input the embedding matrix into a new GAT network, and after the information aggregation operation (refer to step 2.2), output an intermediate matrix of "all node numbers × preset vector length". The structure and attribute information of the attribute network graph are aggregated again on the basis of the previous GAT of this intermediate matrix;

[0114] b. Perform a square operation on the intermediate matrix to obtain a "node number × node number" matrix, and input the matrix into the corresponding sigmoid function to map the elements of the matrix between 0 and 1. Finally, a reconstructed adjacency matrix is ​​output. The reconstructed adjacency matrix is ​​a fitted version of the adjacency matrix of the real attribute network graph.

[0115] In this embodiment, the ARD module specifically operates as follows:

[0116] a. Input the embedding matrix into two new GAT networks in turn. After two information aggregation operations (refer to step 2.2), a reconstructed node attribute matrix (corresponding to the node parameters) of "the number of all nodes × the preset vector length" is output.

[0117] Step 3: According to the reconstructed adjacency matrix and the reconstructed node attribute matrix, the loss function of the GAT-AE perturbation detection model is calculated in the form of a two-norm. The loss function takes into account the difference between the true adjacency matrix and the reconstructed adjacency matrix and the difference between the true node attribute matrix and the reconstructed node attribute matrix, and uses a weight To balance these two differences, the specific formula is as follows. The training goal is to minimize the following function:

[0118]

[0119] in, is the true adjacency matrix, is the reconstructed adjacency matrix, is the real node attribute matrix, is the reconstructed node attribute matrix, is the Frobenius norm of the matrix. In this embodiment, the norm is 2.

[0120] Step 4: Calculate the disturbance parameter (corresponding to the interruption index) of each node. The disturbance parameter refers to the index of the possible interruption of the corresponding enterprise node. The larger the index, the greater the risk of interruption. The formula for the interruption index of each enterprise node is as follows: ,in and Respectively The real adjacency vector and reconstructed adjacency vector of enterprise nodes, and Respectively The real attribute vector and the reconstructed attribute vector of each node. It should be noted that the interruption index of each enterprise node is arranged in descending order to form an interruption sorting list.

[0121] Step 5: According to the loss function in step 3, the GAT-AE perturbation detection model is trained. The training process is completed by the python interface and torch_geometric.nn.GATConv in the torch_geometric third-party module. The hyperparameters are tuned using grid search, the learning rate is defined as 0.01~0.00001, and the Adam optimizer is used.

[0122] Step 6: Continuously extract the attribute network graph in the iterator and repeat steps 2 to 5 until the loss function converges to a low-level stable state.

[0123] Step 7: After the GAT-AE disturbance detection model training is completed, an interruption threshold is set.

[0124] In this embodiment, the interruption threshold (corresponding to the disturbance index threshold) represents the intersection point between the interruption node and the normal node. A node greater than the interruption threshold is detected as an interruption node, and a node less than the interruption threshold is detected as a normal node.

[0125] In some optional implementations, the interruption threshold is determined according to the ROC-AUC indicator, and the specific steps are as follows:

[0126] Step 1: Draw the ROC curve of the rate of change indicator of the current GAT-AE disturbance detection model, and set a binary classification data, that is, the label list of the real state (interruption or normal) of each enterprise node and the interruption ranking list of each enterprise node predicted by the model. Use the label list and interruption ranking list to calculate different interruption thresholds 𝑡( ) under the true positive rate (True Positive Rate, referred to as TPR) and false positive rate (False Positive Rate, referred to as FPR), and draw the ROC curve with FPR as the x-axis and TPR as the y-axis.

[0127] Step 2: Calculate the area of ​​the lower right space of the ROC curve to obtain the AUC (Area Under Curve) indicator.

[0128] In this embodiment, the closer the AUC is to 1, the higher the detection accuracy of the GAT-AE disturbance detection model is. Calculate, where and They respectively represent the true positive rate and false positive rate of the change of the interruption threshold 𝑡; in this embodiment, the values ​​and their AUCs in the interruption sorting list are traversed from small to large, and the value with the largest AUC is set as the final interruption threshold.

[0129] This embodiment also provides an industrial chain disturbance identification device based on GAT and AE, which is used to implement the above embodiments and preferred implementations, and will not be repeated here. As used below, the terms "module", "unit", "subunit", etc. can implement a combination of software and / or hardware for a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0130] Figure 4 is a structural block diagram of an industrial chain disturbance identification device based on GAT and AE according to an embodiment of the present application, such as Figure 3 As shown, the device includes an acquisition module 41, a detection module 42 and a processing module 43, wherein:

[0131] The acquisition module 41 is used to obtain the attribute network diagram of the target industrial chain to be detected, wherein the attribute network diagram includes the adjacency parameters corresponding to all enterprise nodes on the target industrial chain and the node parameters corresponding to each enterprise node, the adjacency parameters are used to characterize whether there is a supply relationship between enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node.

[0132] The detection module 42 is coupled to the acquisition module 41 and is used to input all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all the enterprise nodes into the constructed GAT-AE disturbance detection model for interruption detection, and obtain the disturbance parameters corresponding to all the enterprise nodes, wherein the disturbance parameters include the disturbance index corresponding to the corresponding enterprise node, and the disturbance index is used to characterize the possibility of the enterprise node causing an interruption in the target industrial chain. The GAT-AE disturbance detection model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD;

[0133] The processing module 43 is coupled to the detection module 42, and is used to obtain the corresponding disturbance index from the disturbance parameters corresponding to each enterprise node, and determine the industrial chain disturbance identification result based on the judgment result of whether the disturbance index is less than the disturbance index threshold, wherein the industrial chain disturbance identification result includes the interrupted enterprise nodes and corresponding node parameters on the target industrial chain.

[0134] Through the GAT and AE-based industrial chain disturbance identification device of the embodiment of the present application, an attribute network diagram of the target industrial chain to be detected is obtained, and the attribute network diagram includes adjacency parameters corresponding to all enterprise nodes on the target industrial chain and node parameters corresponding to each enterprise node. The adjacency parameters are used to characterize whether there is a supply relationship between enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node; all adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all enterprise nodes are obtained. The disturbance parameters include the disturbance index corresponding to the corresponding enterprise node, and the disturbance index is used to characterize the possibility of the enterprise node generating an interruption in the target industrial chain. The GAT-AE disturbance detection model The model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD; the corresponding disturbance index is obtained from the disturbance parameters corresponding to each enterprise node, and the industrial chain disturbance identification result is determined based on the judgment result of whether the disturbance index is less than the disturbance index threshold. The industrial chain disturbance identification result includes the interrupted enterprise nodes and corresponding node parameters on the target industrial chain, which solves the problem that the detection method and analysis technology of the interruption of the industrial chain in the related technology cannot uniformly detect the interruption, the detection accuracy is low and the timeliness is poor. The GAT-AE detection model is used to dynamically learn neighbor nodes, reconstruct structures and reconstruct attributes of the attribute network graph, so as to better capture the structure and attribute characteristics of the attribute network graph and improve the accuracy of detecting interrupted nodes. At the same time, the deep learning algorithm is used to train the GAT-AE target detection model to improve the effect and robustness of industrial chain disturbance identification. In addition, structural reconstruction and attribute reconstruction are used to detect abnormal information inherent in the attribute network graph of the industrial chain, and the structure and attributes of the attribute network graph are learned in an unsupervised manner, which reduces the difficulty of training large-scale attribute networks and the dilemma of lack of labeled data, and achieves the beneficial effect of accurately detecting interruptions in industrial chain nodes in the case of unlabeled data and complex attribute network graphs.

[0135] In some embodiments, the detection module 42 further includes:

[0136] An iteration unit, for iterating adjacency matrix calculation and information aggregation operations using GAT based on adjacency parameters corresponding to all enterprise nodes and node parameters corresponding to each enterprise node, until an information vector corresponding to each enterprise node is generated;

[0137] The reconstruction unit is coupled to the iteration unit and is used to reconstruct the structure of the information vector corresponding to each enterprise node through SRE to generate the reconstructed adjacency parameter corresponding to each enterprise node, and to reconstruct the attribute of the information vector corresponding to each enterprise node through ARD to generate the reconstructed node parameter corresponding to each enterprise node;

[0138] The generation unit is coupled to the reconstruction unit and is used to weight the first deviation value and the second deviation value corresponding to each enterprise node after determining the first deviation value between each reconstructed adjacent parameter and the corresponding adjacent parameter and determining the second deviation value between each reconstructed node parameter and the corresponding node parameter, so as to generate the disturbance parameter corresponding to each enterprise node.

[0139] In some embodiments, the iteration unit is also used to determine the first weight parameter and the preset first attention mechanism parameter corresponding to the neural network layer of the GAT for the current adjacency matrix calculation and information aggregation operation, wherein the first weight parameter is used to characterize the weight matrix of the corresponding neural network layer; from all enterprise nodes, traverse the adjacent enterprise nodes associated with each enterprise node, and determine the adjacent enterprise node group consisting of the enterprise node and an adjacent enterprise node, and determine the first node parameter corresponding to the target enterprise node in all adjacent enterprise node groups, wherein the target enterprise node includes the enterprise node in the corresponding adjacent enterprise node group and one of the adjacent enterprise nodes, and the first node parameter is generated by information aggregation of the historical adjacency parameters generated by all adjacent enterprise node groups associated with the enterprise node in the previous iteration operation and the historical node parameters corresponding to all adjacent enterprise nodes associated with the enterprise node; according to the first node parameters, first weight parameters and first attention mechanism parameters of the two target enterprise nodes corresponding to each adjacent enterprise node group associated with the corresponding enterprise node, vector splicing, matrix inversion and power operation are performed to obtain the corresponding each adjacent enterprise node The first relationship value corresponding to the group, and after determining the first total relationship value corresponding to all adjacent enterprise node groups associated with the corresponding enterprise node according to the first relationship value corresponding to each adjacent enterprise node group, based on the ratio of the first relationship value corresponding to the target adjacent enterprise node group to the first total relationship value, generate the corresponding correlation coefficient for the current adjacency matrix calculation of the target adjacent enterprise node group, wherein the adjacency parameter includes the correlation coefficient, the first relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group, the target adjacent enterprise node group includes one of all adjacent enterprise node groups associated with the corresponding enterprise node, and the correlation coefficient is used to characterize the degree of correlation between the corresponding two enterprise nodes; based on the correlation coefficient corresponding to each target adjacent enterprise node group, the first weight parameter and the first node parameter, perform information aggregation to generate the current information vector corresponding to the current information aggregation operation of the enterprise node corresponding to the target adjacent enterprise node group; after the number of times of repeatedly executing the steps of generating the corresponding correlation coefficient and generating the corresponding current information vector reaches the preset number, the current information vector generated for the last time is used as the information vector corresponding to the corresponding enterprise node.

[0140] In some of the embodiments, the reconstruction unit is also used to perform vector concatenation, matrix inversion and power operation according to the information vectors of the two target enterprise nodes corresponding to each corresponding adjacent enterprise node group, the second weight parameter and the second attention mechanism parameter, to obtain the second relationship value corresponding to each corresponding adjacent enterprise node group, and after determining the second total relationship value corresponding to all adjacent enterprise node groups associated with the corresponding enterprise node according to the second relationship value corresponding to each adjacent enterprise node group, generate the reconstruction relationship corresponding to the structural reconstruction of the target adjacent enterprise node group based on the ratio of the second relationship value corresponding to the target adjacent enterprise node group to the second total relationship value. number, wherein the second relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group, and the reconstruction correlation coefficient is used to characterize the degree of correlation between the two corresponding enterprise nodes after structural reconstruction; based on the reconstruction correlation coefficient, the second weight parameter and the corresponding information vector corresponding to each target adjacent enterprise node group, information aggregation is performed to generate corresponding structural vector information, and the structural vector information is decoded through SRE to generate the reconstruction adjacency parameters corresponding to each enterprise node, wherein the structural vector information is used to characterize the information vector generated by the structural reconstruction of the enterprise nodes corresponding to the corresponding target adjacent enterprise node group.

[0141] In some of the embodiments, the reconstruction unit is also used to determine the target neural network layer for attribute reconstruction by ARD, and determine the third weight parameter and the preset third attention mechanism parameter corresponding to the corresponding target neural network layer, wherein the third weight parameter is used to characterize the weight matrix of the corresponding target neural network layer, and the target network neural layer includes a first neural network layer and a second neural network layer; according to the information vectors of the two target enterprise nodes corresponding to each corresponding adjacent enterprise node group, the third weight parameters and the third attention mechanism parameters respectively corresponding to the first neural network layer, vector concatenation, matrix inversion and power operation corresponding to the first neural network layer are performed to obtain the third relationship value corresponding to each corresponding adjacent enterprise node group in the first neural network layer, and after determining the third total relationship value corresponding to the first neural network layer of all adjacent enterprise node groups associated with the corresponding enterprise node according to the third relationship value corresponding to each adjacent enterprise node group, a target is generated based on the ratio of the second relationship value corresponding to the target adjacent enterprise node group to the third total relationship value. The first correlation coefficient generated by the attribute reconstruction of the adjacent enterprise node group in the first neural network layer, wherein the third relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group in the first neural network layer, and the first correlation coefficient is used to characterize the degree of correlation between the two corresponding enterprise nodes after the attribute reconstruction in the first neural network layer; after information aggregation is performed based on the first correlation coefficient, the third weight parameter and the corresponding information vector corresponding to each target adjacent enterprise node group to generate the first attribute vector information, the corresponding attribute reconstruction and information aggregation are performed according to the first attribute vector information, the third weight parameter and the third attention mechanism parameter corresponding to the second neural network layer respectively, to generate the second attribute vector information, and the second attribute vector information is used as the reconstruction node parameter corresponding to each enterprise node, wherein the first attribute vector information and the second attribute vector information respectively characterize the information vector generated by the attribute reconstruction of the enterprise node corresponding to the corresponding target adjacent enterprise node group in the corresponding target neural network layer.

[0142] In some embodiments, the processing module 43 is also used to determine whether the disturbance index is less than the disturbance index threshold; when it is determined that the disturbance index is not less than the disturbance index threshold, the corresponding enterprise node is determined to be an industrial chain interruption node, wherein the industrial chain disturbance identification result includes the industrial chain interruption node; when it is determined that the disturbance index is less than the disturbance index threshold, the corresponding enterprise node is determined to be a normal node.

[0143] In some of the embodiments, after the processing module 43 determines the industrial chain disturbance identification result, the industrial chain disturbance identification device based on GAT and AE is also used to mark the industrial chain interruption nodes in the attribute network diagram according to a preset marking format and update the attribute network diagram.

[0144] In some of the embodiments, before using GAT to iterate the adjacency matrix calculation and information aggregation operations based on the adjacency parameters corresponding to all enterprise nodes and the node parameters corresponding to each enterprise node, the industrial chain disturbance identification device based on GAT and AE is also used to determine the parameter value corresponding to the adjacency parameter, and judge whether the parameter value is the target parameter value, wherein the target parameter value is used to characterize that there is no supply relationship between the corresponding enterprise nodes; when it is judged that the parameter value is the target parameter value, the adjacency parameter and the node parameter are respectively used as the corresponding reconstructed adjacency parameter and the reconstructed node parameter, and it is determined that the corresponding enterprise node is currently an industrial chain interruption node.

[0145] This embodiment further provides a detection system, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0146] Optionally, the detection system may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0147] Optionally, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0148] S1, obtain the attribute network diagram of the target industrial chain to be detected. The attribute network diagram includes the adjacency parameters corresponding to all enterprise nodes in the target industrial chain and the node parameters corresponding to each enterprise node. The adjacency parameters are used to characterize whether there is a supply relationship between enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node.

[0149] S2, input all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all enterprise nodes into the constructed GAT-AE perturbation detection model for interruption detection, and obtain the perturbation parameters corresponding to all enterprise nodes. The perturbation parameters include the perturbation index corresponding to the corresponding enterprise node. The perturbation index is used to characterize the possibility of enterprise nodes causing interruptions in the target industrial chain. The GAT-AE perturbation detection model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD.

[0150] S3, obtain the corresponding disturbance index from the disturbance parameters corresponding to each enterprise node, and determine the industrial chain disturbance identification result based on the judgment result of whether the disturbance index is less than the disturbance index threshold. The industrial chain disturbance identification result includes the interrupted enterprise nodes and corresponding node parameters on the target industrial chain.

[0151] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be described in detail here.

[0152] In addition, in combination with the industrial chain disturbance identification method based on GAT and AE in the above embodiment, the embodiment of the present application can provide a storage medium for implementation. The storage medium stores a computer program; when the computer program is executed by the processor, any one of the industrial chain disturbance identification methods based on GAT and AE in the above embodiment is implemented.

[0153] Those skilled in the art should understand that the technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for identifying industrial chain disturbances based on graph attention network GAT and autoencoder AE, characterized in that: include: Obtaining an attribute network diagram of the target industrial chain to be detected, wherein the attribute network diagram includes adjacency parameters corresponding to all enterprise nodes on the target industrial chain and node parameters corresponding to each enterprise node, the adjacency parameters are used to characterize whether there is a supply relationship between the enterprise nodes, and the node parameters are used to characterize the production status information of each enterprise node; All the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all the enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all the enterprise nodes are obtained, wherein the disturbance parameters include the disturbance index corresponding to the corresponding enterprise node, and the disturbance index is used to characterize the possibility of the enterprise node generating an interruption on the target industrial chain. The GAT-AE disturbance detection model is a neural network model trained based on the attention network encoder GAT, the structure reconstruction decoder SRE and the attribute reconstruction decoder ARD; Obtaining the corresponding disturbance index from the disturbance parameters corresponding to each enterprise node, and determining the industrial chain disturbance identification result according to the judgment result of whether the disturbance index is less than the disturbance index threshold, wherein the industrial chain disturbance identification result includes the enterprise nodes interrupted on the target industrial chain and the corresponding node parameters; Among them, all the adjacency parameters corresponding to the target industrial chain and the node parameters corresponding to all the enterprise nodes are input into the constructed GAT-AE disturbance detection model for interruption detection, and the disturbance parameters corresponding to all the enterprise nodes are obtained, including: Based on the adjacency parameters corresponding to all the enterprise nodes and the node parameters corresponding to each of the enterprise nodes, using the GAT to iterate adjacency matrix calculation and information aggregation operations until an information vector corresponding to each of the enterprise nodes is generated; By means of the SRE, the information vector corresponding to each of the enterprise nodes is structurally reconstructed to generate reconstructed adjacency parameters corresponding to each of the enterprise nodes; and by means of the ARD, the information vector corresponding to each of the enterprise nodes is attribute reconstructed to generate reconstructed node parameters corresponding to each of the enterprise nodes; After determining a first deviation value between each of the reconstructed adjacency parameters and the corresponding adjacency parameter and determining a second deviation value between each of the reconstructed node parameters and the corresponding node parameter, weighting the first deviation value and the second deviation value corresponding to each of the enterprise nodes to generate the disturbance parameter corresponding to each of the enterprise nodes; Wherein, based on the adjacency parameters corresponding to all the enterprise nodes and the node parameters corresponding to each of the enterprise nodes, using the GAT to iterate adjacency matrix calculation and information aggregation operations until an information vector corresponding to each of the enterprise nodes is generated, including: Determine a first weight parameter and a preset first attention mechanism parameter corresponding to the neural network layer for the GAT to perform current adjacency matrix calculation and information aggregation operation, wherein the first weight parameter is used to characterize the weight matrix of the corresponding neural network layer; From all the enterprise nodes, traverse the adjacent enterprise nodes associated with each of the enterprise nodes, and determine an adjacent enterprise node group consisting of the enterprise node and one of the adjacent enterprise nodes, and determine a first node parameter corresponding to a target enterprise node in all the adjacent enterprise node groups, wherein the target enterprise node includes the enterprise node in the corresponding adjacent enterprise node group and one of the adjacent enterprise nodes, and the first node parameter is generated by aggregating information of historical adjacent parameters generated by all the adjacent enterprise node groups associated with the enterprise node in the previous iteration operation and historical node parameters corresponding to all the adjacent enterprise nodes associated with the enterprise node; According to the first node parameters, the first weight parameters and the first attention mechanism parameters of the two target enterprise nodes corresponding to each of the adjacent enterprise node groups associated with the corresponding enterprise node, vector concatenation, matrix inversion and power operation are performed to obtain the first relationship value corresponding to each of the adjacent enterprise node groups, and after determining the first total relationship value corresponding to all the adjacent enterprise node groups associated with the corresponding enterprise node according to the first relationship value corresponding to each of the adjacent enterprise node groups, based on the ratio of the first relationship value corresponding to the target adjacent enterprise node group to the first total relationship value, the correlation coefficient corresponding to the current adjacency matrix calculation of the target adjacent enterprise node group is generated, wherein the adjacency parameters include the correlation coefficient, the first relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group, the target adjacent enterprise node group includes one of all the adjacent enterprise node groups associated with the corresponding enterprise node, and the correlation coefficient is used to characterize the degree of correlation between the corresponding two enterprise nodes; Based on the correlation coefficient, the first weight parameter and the first node parameter corresponding to each target adjacent enterprise node group, information aggregation is performed to generate a current information vector corresponding to the current information aggregation operation performed by the enterprise node corresponding to the target adjacent enterprise node group; After the number of times of repeatedly executing the steps of generating the corresponding correlation coefficient and generating the corresponding current information vector reaches a preset number, the current information vector generated for the last time is used as the information vector corresponding to the corresponding enterprise node, wherein the information vector corresponding to each enterprise node is structurally reconstructed by the SRE to generate the reconstructed adjacency parameters corresponding to each enterprise node, including: Determine a second weight parameter and a preset second attention mechanism parameter corresponding to the neural network layer for structural reconstruction of the SRE, wherein the second weight parameter is used to characterize a weight matrix of the neural network layer corresponding to the SRE; According to the information vectors, the second weight parameters and the second attention mechanism parameters of the two target enterprise nodes corresponding to each corresponding adjacent enterprise node group, vector concatenation, matrix inversion and power operation are performed to obtain the second relationship value corresponding to each corresponding adjacent enterprise node group, and after determining the second total relationship value corresponding to all the adjacent enterprise node groups associated with the corresponding enterprise node according to the second relationship value corresponding to each adjacent enterprise node group, a reconstruction correlation coefficient corresponding to the structural reconstruction of the target adjacent enterprise node group is generated based on the ratio of the second relationship value corresponding to the target adjacent enterprise node group to the second total relationship value, wherein the second relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group, and the reconstruction correlation coefficient is used to characterize the degree of correlation between the two corresponding enterprise nodes after structural reconstruction; Based on the reconstruction correlation coefficient, the second weight parameter and the corresponding information vector corresponding to each of the target adjacent enterprise node groups, information aggregation is performed to generate corresponding structural vector information, and the structural vector information is decoded through the SRE to generate the reconstruction adjacency parameters corresponding to each of the enterprise nodes, wherein the structural vector information is used to characterize the information vector generated by the structural reconstruction of the enterprise nodes corresponding to the corresponding target adjacent enterprise node group.

2. The method according to claim 1, characterized in that: Through the ARD, the information vector corresponding to each enterprise node is attribute reconstructed to generate a reconstructed node parameter corresponding to each enterprise node, including: Determine the target neural network layer for attribute reconstruction by the ARD, and determine a third weight parameter and a preset third attention mechanism parameter corresponding to the corresponding target neural network layer, wherein the third weight parameter is used to characterize the weight matrix of the corresponding target neural network layer, and the target neural network layer includes a first neural network layer and a second neural network layer; According to the information vectors of the two target enterprise nodes corresponding to each corresponding adjacent enterprise node group, the third weight parameters and the third attention mechanism parameters respectively corresponding to the first neural network layer, vector concatenation, matrix inversion and power operation corresponding to the first neural network layer are performed to obtain the third relationship value corresponding to each corresponding adjacent enterprise node group in the first neural network layer, and after determining the third total relationship value corresponding to all the adjacent enterprise node groups associated with the corresponding enterprise node in the first neural network layer according to the third relationship value corresponding to each adjacent enterprise node group, based on the ratio of the second relationship value corresponding to the target adjacent enterprise node group to the third total relationship value, a first correlation coefficient generated by attribute reconstruction of the target adjacent enterprise node group in the first neural network layer is generated, wherein the third relationship value is used to characterize the association value of the two target enterprise nodes corresponding to the corresponding adjacent enterprise node group in the first neural network layer, and the first correlation coefficient is used to characterize the degree of correlation between the corresponding two enterprise nodes after attribute reconstruction in the first neural network layer; After information aggregation is performed based on the first correlation coefficient, the third weight parameter and the corresponding information vector corresponding to each of the target adjacent enterprise node groups to generate the first attribute vector information, corresponding attribute reconstruction and information aggregation are performed according to the first attribute vector information, the third weight parameter and the third attention mechanism parameter respectively corresponding to the second neural network layer to generate the second attribute vector information, and the second attribute vector information is used as the reconstructed node parameter corresponding to each of the enterprise nodes, wherein the first attribute vector information and the second attribute vector information respectively represent the information vectors generated by the attribute reconstruction of the enterprise nodes corresponding to the corresponding target adjacent enterprise node group in the corresponding target neural network layer.

3. The method according to claim 1, characterized in that According to the judgment result of whether the disturbance index is less than the disturbance index threshold, the industrial chain disturbance identification result is determined, including: Determining whether the disturbance index is less than the disturbance index threshold; In the case where it is determined that the disturbance index is not less than the disturbance index threshold, determining the corresponding enterprise node as an industrial chain interruption node, wherein the industrial chain disturbance identification result includes the industrial chain interruption node; When it is determined that the disturbance index is less than the disturbance index threshold, the corresponding enterprise node is determined to be a normal node.

4. The method according to claim 3, characterized in that After determining the result of identifying the disturbance in the industrial chain, the method further includes: in the attribute network diagram, marking the interruption node of the industrial chain according to a preset marking format, and updating the attribute network diagram.

5. The method according to claim 1, characterized in that Before performing iterations of adjacency matrix calculation and information aggregation operations using the GAT based on the adjacency parameters corresponding to all the enterprise nodes and the node parameters corresponding to each of the enterprise nodes, the method further includes: Determine a parameter value corresponding to the adjacency parameter, and determine whether the parameter value is a target parameter value, wherein the target parameter value is used to indicate that there is no supply relationship between the corresponding enterprise nodes; When it is determined that the parameter value is the target parameter value, the adjacency parameter and the node parameter are used as the corresponding reconstructed adjacency parameter and the reconstructed node parameter respectively, and it is determined that the corresponding enterprise node is currently an industrial chain interruption node.

6. A detection system, comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps of the industrial chain disturbance identification method based on graph attention network GAT and autoencoder AE as described in any one of claims 1 to 5.

7. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the industrial chain disturbance identification method based on graph attention network GAT and autoencoder AE as described in any one of claims 1 to 5.

Citation Information

Patent Citations

  • Industrial chain model generation method and system

    CN117634991A

  • Industrial chain risk assessment method of graph neural network

    CN118114989A