Device familial defect automatic identification system based on spectral clustering and GNN

Abstract

The invention relates to an equipment familial defect automatic identification system based on spectral clustering and GNN. The system comprises a real-time data monitoring module, a graph model construction module, an adaptive learning module, a defect source tracing module, a predictive maintenance and early warning module, and a structured defect library and label management module. After equipment is classified by adopting a spectral clustering method, constructing a graph model for predicting familial defects of the equipment based on a graph neural network GNN; and aiming at the equipment with the equipment familial defects, constructing a defect association graph, and tracking the origin and the propagation path of the equipment familial defects. The system is connected with the hydroelectric industry internet, realizes real-time online detection of the familial defects of the equipment by virtue of high interconnection, expandability and visualization capability of an industrial internet platform, and effectively identifies root causes and propagation paths of the familial defects of the equipment.

Description

Technical Field

[0001] The present invention belongs to the field of hydropower industrial Internet data processing, and specifically relates to an automatic identification system for equipment family defects based on spectral clustering and GNN. Background Art

[0002] In the hydropower industry, the proper operation of equipment is crucial for ensuring safe production and improving economic efficiency. However, for various reasons, these equipment may develop familial defects—common defects or failures with similar characteristics that occur across a class or batch of products, equipment, or systems. These defects are often linked to common causes such as manufacturing, design, or material selection. These defects can lead to severe accidents such as widespread equipment downtime.

[0003] However, identifying family defects presents significant challenges. First, due to the unique nature of family defects, many manufacturers and suppliers may not proactively disclose relevant information, making defect identification particularly challenging. Furthermore, the hydropower industry boasts a wide variety of equipment, each with distinct operating characteristics and data. This leads to significant differences between family defects and a large number of them, further complicating defect identification.

[0004] Existing methods for identifying familial defects in equipment often rely heavily on feedback from operators and maintenance personnel, requiring a long period of time to uncover them. It can also be difficult to trace the root cause of these defects, making them difficult to prevent and resolve. Furthermore, while regular inspections and maintenance are effective ways to detect familial defects, this approach often requires significant human and material resources and can result in lengthy equipment downtime, significantly impacting production safety and efficiency. Summary of the Invention

[0005] The purpose of the present invention is to address the above-mentioned problems and provide an automatic identification system for equipment family defects based on spectral clustering and GNN. After classifying the equipment using the spectral clustering method, a graphical model for predicting equipment family defects is constructed based on the graph neural network GNN; for equipment with equipment family defects, a defect association graph is constructed to track the origin and propagation path of the equipment family defects; connected with the hydropower industrial Internet, relying on the high connectivity, scalability and visualization capabilities of the industrial Internet platform, real-time online detection of equipment family defects is realized, and the root cause and propagation path of equipment family defects are effectively identified.

[0006] The technical solution of the present invention is an automatic identification system for equipment family defects based on spectral clustering and GNN, including a real-time data monitoring module, a graph model construction module, an adaptive learning module, a defect source tracing module, a predictive maintenance and early warning module, a structured defect library and label management module, and a user feedback module.

[0007] The real-time data monitoring module obtains equipment data through the hydropower industrial Internet platform, transmits the collected equipment data to the database in real time, and stores the pre-processed data for use by the adaptive learning module and the defect source tracing module.

[0008] The graph model construction module adopts a spectral clustering method and combines it with a graph neural network (GNN) to construct a graph model, which is used to predict familial defects.

[0009] The adaptive learning module automatically updates and adjusts the graph model in the graph model construction module according to the new data output by the real-time data monitoring module.

[0010] The defect source tracing module is used to trace the origin and propagation path of equipment family defects.

[0011] The predictive maintenance and early warning module uses a graphical model to predict possible equipment family defects based on the real-time data of the equipment obtained by the real-time data monitoring module, and promptly issues early warnings for the predicted equipment family defects.

[0012] The structured defect library and label management module are used to store, manage and retrieve all information related to device family defects.

[0013] The user feedback module transmits the device defects reported by users and the device feature data related to the device defects to the database, and stores them in the structured defect library after preprocessing.

[0014] Preferably, the graph model construction module specifically includes: The device classification submodule collects the data output by the real-time data detection module, generates a similarity matrix between devices based on the similarity of the device data, and uses a spectral clustering method to classify similar devices into the same category; The family defect prediction submodule uses the graph neural network GNN to predict the family defects of the equipment. Based on the results of spectral clustering, it generates a family defect prediction submodule for each equipment category. i Constructing a subgraph ,in is a node set of the subgraph, where the nodes represent the corresponding devices. For the i The edge set of the subgraph of the class device. The edges in the subgraph are used to represent the relationship between devices.

[0015] The graph model construction module is used to build nodes The eigenvector of Perform iterative calculations. In each iterative calculation, the feature of each node is updated as a function of its own and its neighbors' features. The iterative calculation formula is:

[0016] in, Representation node The neighbor set of u express The nodes in It is l The weight matrix of times, is the activation function, 、 Respectively represent l sequence l +1 node calculated by iterative calculation The eigenvector of .

[0017] After multiple iterations, each device will obtain the latest feature vector, and the output layer of the graph neural network (GNN) will predict device family defects based on the latest feature vector of each device.

[0018] Furthermore, the output layer of the graph neural network (GNN) predicts device family defects based on the latest feature vector of each device, specifically including: Using the weight matrix of the output layer The latest feature vector of the device is converted. The conversion formula is:

[0019] Where, Indicates the converted device v The eigenvector of For equipment v The eigenvector of represents the bias term.

[0020] The activation function of the output layer is based on Calculate the probability that device v has a family defect, which is calculated as:

[0021] Where, Representation device v The probability of familial defects.

[0022] like , then determine the device v There is a family defect, otherwise, the device v No familial defects.

[0023] Furthermore, during the training process of the graph neural network GNN, the output of the graph neural network GNN is Devices with training datasets vThe true labels of whether there are familial defects are compared and the loss is calculated using the binary cross entropy loss function:

[0024] Where, Representation device v Whether there is a true label of familial defects, log( ) represents the logarithmic function.

[0025] Furthermore, the graph model in the graph model construction module is automatically updated and adjusted according to the new data output by the real-time data monitoring module. is the current graph model, where are the parameters of the graphical model, is the input data of the graphical model, namely the device feature data matrix.

[0026] The goal of updating and adjusting the graph model in the graph model building module is to update , minimize the loss of the graph model on new data, update The calculation formula is: ; in, Represents the new device feature data matrix, where each row represents a device and each column represents a feature. express The corresponding familial defect label vector, is the loss function; Represents the parameters of the updated graphical model; argmin is the minimum index function.

[0027] After calculating the new parameters of the graph model, the graph model construction module uses the new device feature data to and existing data, update the similarity matrix between devices S ; Use the updated S , update the structure and weights of the graph neural network GNN.

[0028] Furthermore, the defect source tracing module determines the equipment, components and timestamps related to the family defect based on the data obtained by the real-time data monitoring module and the user feedback module.

[0029] Construct defect association maps for each known familial defect ,in Represents a set of defective nodes, where defective nodes represent defective devices; is the defect edge set, , Representation device and equipment Defect associations between Based on the similarity between devices, supply chain information and timestamps, the defect edges of the defect association graph are Assign defect weights , Representation device and equipment The possibility of defect propagation between them is determined by using the Dijkstra shortest path algorithm to find the defect edge with the largest defect weight, and the source of the defect is determined based on the defect edge.

[0030] Preferably, when the automatic identification system for device family defects detects a new device family defect or receives device defects reported by users, the structured defect library and label management module stores the device defect information and device feature data in the structured defect library. The structured defect library includes not only the description information and feature data of the device defect, but also the related devices, the time when the defect occurred, and the scope of the defect impact.

[0031] Preferably, the structured defect library and label management module also includes a label management submodule for adding, modifying or deleting labels for equipment defects. The labels are set based on multiple factors such as the type, severity and source of the equipment defects. The labels are used to quickly retrieve specific types of equipment defects.

[0032] Preferably, the structured defect library and label management module also includes a defect feedback submodule, which is used for the maintenance team or experts to provide feedback information on the data in the structured defect library; when further classifying existing equipment defects, the defect feedback submodule provides the function of editing, adding or modifying labels.

[0033] The above-mentioned method of the automatic identification system of equipment family defects based on spectral clustering and GNN includes the following steps: S1. The collected device data is transmitted to the database in real time, the device data is preprocessed using a data preprocessing component, and the preprocessed device data is stored; S2. After classifying the devices using the spectral clustering method, a graph neural network (GNN) is used to construct a graph model of the devices. The graph model is used to predict family defects of the devices. S3. Based on the new data of the equipment obtained by the real-time data monitoring module, update and adjust the graph model in the graph model construction module to adapt the graph model to the new operating conditions and changes of the equipment; S4. Based on real-time equipment data, a graphical model is used to predict equipment family defects and timely issue warnings for the predicted equipment family defects; S5. For devices with family defects, a defect correlation diagram is constructed to track the origin and propagation path of the family defects.

[0034] Compared with the prior art, the present invention has the following beneficial effects: 1) The present invention constructs a graphical model for predicting equipment family defects and a defect association graph for tracing the origins and propagation paths of equipment family defects. By integrating with the industrial Internet platform, real-time online detection of equipment family defects is achieved, improving the accuracy of equipment family defect detection, reducing false positives and missed positives, and effectively identifying the root causes and propagation paths of equipment family defects. By integrating dual-channel analysis of real-time data streams (time dimension) and equipment association graphs (spatial dimension), the prediction accuracy is significantly improved compared to single-dimensional methods.

[0035] 2) This invention adopts a joint modeling strategy of spectral clustering (explicit topology) and GNN (implicit topology), breaking through the cognitive limitations of traditional single-mode representation through the topological complementarity mechanism. This design significantly improves the defect recognition rate of the model in cross-vendor and cross-model equipment clusters, especially in the scenario of new heterogeneous equipment access, showing a generalization capability that is significantly superior to traditional methods.

[0036] 3) The present invention introduces a parameter update formula for the graphical model, realizing online incremental learning of model parameters, breaking through the bottleneck of traditional GNN static training. This enables the device family defect detection model of the present invention to have dynamic learning and adaptability, and continuously learns and optimizes based on new data and user feedback, ensuring the continuous improvement of defect detection accuracy.

[0037] 4) This invention constructs a defect association graph and uses the Dijkstra algorithm to search for the maximum defect weight path, combining timestamps with supply chain information to achieve efficient and accurate location of the defect source and effectively identify the root cause and propagation path of family defects in equipment.

[0038] 5) The user feedback module provided by the present invention is combined with the structured defect library and the hydropower industrial Internet platform to form a "data-model-knowledge" closed loop. Through dynamic labeling, the autonomous evolution of the equipment defect knowledge library of the hydropower industrial Internet platform is achieved.

[0039] 6) Through deep integration with the Industrial Internet platform, the present invention has achieved technological breakthroughs and synergistic efficiency in the entire defect detection process. Through docking with the Industrial Internet platform, the real-time data of the equipment is synchronized with the SCADA system and MIS system in seconds, thereby improving the coverage of detection data. The defect detection results are visualized in combination with the Industrial Internet platform, and the family defects, sources and propagation paths of the equipment are mapped to the digital twin objects of the Industrial Internet platform, so that operation and maintenance personnel can quickly locate the source of the fault and take targeted measures. The equipment defect library is structured and labeled, and synchronized with the equipment defect knowledge base of the hydropower industrial Internet platform to form a quality traceability chain for the entire life cycle of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below with reference to the accompanying drawings and examples.

[0041] Figure 1 Schematic diagram of an automatic identification system for equipment family defects based on spectral clustering and GNN according to an embodiment of the present invention.

[0042] Figure 2 Schematic diagram of the process of the device family defect identification method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] Example 1 like Figure 1 As shown in the figure, the automatic identification system for equipment family defects based on spectral clustering and GNN includes a real-time data monitoring module, a graph model construction module, an adaptive learning module, a defect source tracing module, a predictive maintenance and early warning module, a structured defect library and label management module, and a user feedback module.

[0044] The real-time data monitoring module transmits the collected equipment data to the database in real time and stores the pre-processed data for use by the adaptive learning module and the defect source tracing module.

[0045] In the embodiment, the real-time data monitoring module is connected to the hydropower industrial Internet platform, and is connected to the sensor group, SCADA system and MIS system through the industrial Internet platform to obtain data for equipment family defect detection.

[0046] The graphical model construction module adopts the spectral clustering method and combines it with the graph neural network GNN to construct a graphical model, which is used to predict familial defects.

[0047] The graph model construction module specifically includes: The device classification submodule collects the data output by the real-time data detection module, generates a similarity matrix between devices based on the similarity of the device data, and uses a spectral clustering method to classify similar devices into the same category; The family defect prediction submodule uses the graph neural network GNN to predict the family defects of the equipment. Based on the results of spectral clustering, it generates a family defect prediction submodule for each equipment category. i Constructing a subgraph ,in is a node set of the subgraph, where the nodes represent the corresponding devices. For the i The edge set of the subgraph of the class device. The edges in the subgraph are used to represent the relationship between devices.

[0048] For Node The eigenvector of Perform iterative calculations. In each iterative calculation, the feature of each node is updated as a function of its own and its neighbors' features. The iterative calculation formula is:

[0049] in, Representation node The neighbor set of u express The nodes in It is l The weight matrix of times, is the activation function, 、 Respectively represent l sequence l +1 node calculated by iterative calculation The eigenvector of .

[0050] After multiple iterations, each device will obtain the latest feature vector, and the output layer of the graph neural network (GNN) will predict device family defects based on the latest feature vector of each device.

[0051] The output layer of the graph neural network (GNN) predicts device family defects based on the latest feature vector of each device, specifically including: Using the weight matrix of the GNN output layer The latest feature vector of the device is converted. The conversion formula is:

[0052] Where, Indicates the converted device v The eigenvector of For equipment v The eigenvector of represents the bias term.

[0053] The activation function of the output layer is based on Calculate the probability that device v has a family defect, which is calculated as:

[0054] Where, Representation device v The probability of familial defects.

[0055] like , then determine the device v There is a family defect, otherwise, the device v No familial defects.

[0056] During the training process of the graph neural network GNN, the output of the graph neural network GNN is Devices with training datasets v The true labels of whether there are familial defects are compared and the loss is calculated using the binary cross entropy loss function:

[0057] Where, Representation device v Whether there is a true label of familial defects, log( ) represents the logarithmic function.

[0058] The adaptive learning module automatically updates and adjusts the graph model in the graph model construction module according to the new data output by the real-time data monitoring module.

[0059] set up is the current graph model, where are the parameters of the graphical model, is the input data of the graphical model, namely the device feature data matrix.

[0060] The goal of updating and adjusting the graph model in the graph model building module is to update , minimize the loss of the graph model on new data, update The calculation formula is: ;. in, Represents the new device feature data matrix, where each row represents a device and each column represents a feature. express The corresponding familial defect label vector, is the loss function; Represents the parameters of the updated graphical model; argmin is the minimum index function.

[0061] After calculating the new parameters of the graph model, the graph model construction module uses the new device feature data to and existing data, update the similarity matrix between devices S ; Use the updated S , update the structure and weights of the graph neural network GNN.

[0062] The defect source tracing module is used to track the origin and propagation path of equipment family defects.

[0063] Based on the data obtained from the real-time data monitoring module and the user feedback module, the devices, components and timestamps related to the familial defect are determined.

[0064] Construct defect association maps for each known familial defect ,in Represents a set of defective nodes, where defective nodes represent defective devices; is the defect edge set, , Representation device and equipment Defect associations between Based on the similarity between devices, supply chain information and timestamps, the defect edges of the defect association graph are Assign defect weights , Representation device and equipment .The possibility of defect propagation between them is determined by using the Dijkstra shortest path algorithm to find the defect edge with the largest defect weight, and the source of the defect is determined based on the defect edge.

[0065] The predictive maintenance and early warning module uses a graphical model to predict possible equipment family defects based on the real-time data of the equipment obtained by the real-time data monitoring module, and issues timely early warnings for the predicted equipment family defects.

[0066] The structured defect library and tag management module is used to store, manage and retrieve all information related to device family defects.

[0067] The structured defect library and tag management module includes a tag management submodule and a defect feedback submodule.

[0068] The tag management submodule is used to add, modify or delete tags for equipment defects. Tags are set based on multiple factors such as the type, severity and source of the equipment defect. The tags are used to quickly retrieve equipment defects of a specific type.

[0069] The defect feedback submodule is used for the maintenance team or experts to provide feedback information on the data in the structured defect library; when further classifying existing equipment defects, the defect feedback submodule provides the function of editing, adding or modifying tags.

[0070] The user feedback module transmits the device defects reported by users and the device feature data related to the device defects to the database, and stores them in the structured defect library after preprocessing.

[0071] When the automatic identification system for device family defects detects a new device family defect or receives device defects reported by users, the structured defect library and label management module stores the device defect information and device feature data in the structured defect library. The structured defect library includes not only the description information and feature data of the device defect, but also the related devices, the time when the defect occurred, and the scope of the defect impact.

[0072] In the embodiment, the user feedback module and the structured defect library are combined with the hydropower industrial Internet platform to form a "data-model-knowledge" closed loop. Through online real-time detection of equipment family defects, identification of defect root causes, structured management of the equipment defect library and dynamic labeling of tags, as well as synchronization with the equipment defect knowledge base of the hydropower industrial Internet platform, the autonomous evolution of the equipment defect knowledge base of the hydropower industrial Internet platform is achieved.

[0073] The implementation results show that the combination of this invention and the hydropower industrial Internet platform has formed a quality traceability chain for the entire life cycle of the equipment, reducing the defect reproduction rate by 82% and increasing the accuracy of spare parts replacement to 57.3%, producing synergistic benefits that significantly exceed the expected technology superposition.

[0074] Example 2 like Figure 2 As shown, the method of the automatic identification system of equipment family defects based on spectral clustering and GNN includes the following steps: S1. The collected device data is transmitted to the database in real time, the device data is preprocessed using a data preprocessing component, and the preprocessed device data is stored.

[0075] S2. After classifying the devices using the spectral clustering method, a graph model of the devices is constructed using the graph neural network (GNN). The graph model is used to predict family defects of the devices.

[0076] Based on the results of spectral clustering, for each device category i Constructing a subgraph ,in is a node set of the subgraph, where the nodes represent the corresponding devices. For the i The edge set of the subgraph of the class device. The edges in the subgraph are used to represent the relationship between devices.

[0077] For Node The eigenvector of Perform iterative calculations. In each iterative calculation, the feature of each node is updated as a function of its own and its neighbors' features. The iterative calculation formula is:

[0078] in, Representation node The neighbor set of u express The nodes in It is l The weight matrix of times, is the activation function, 、 Respectively represent l sequencel +1 node calculated by iterative calculation The eigenvector of After multiple iterations, each device will obtain the latest feature vector, and the output layer of the graph neural network (GNN) will predict device family defects based on the latest feature vector of each device.

[0079] The output layer of the graph neural network (GNN) predicts device family defects based on the latest feature vector of each device, specifically including: Using the weight matrix of the output layer The latest feature vector of the device is converted. The conversion formula is:

[0080] Where, Indicates the converted device v The eigenvector of For equipment v The eigenvector of represents the bias term; The activation function of the output layer is based on Calculate the probability that device v has a family defect, which is calculated as:

[0081] Where, Representation device v The probability of familial defects.

[0082] like , then determine the device v There is a family defect, otherwise, the device v No familial defects.

[0083] During the training process of the graph neural network GNN, the output of the graph neural network GNN is Devices with training datasets v The true labels of whether there are familial defects are compared and the loss is calculated using the binary cross entropy loss function:

[0084] Where, Representation device v Whether there is a true label of familial defects, log( ) represents the logarithmic function.

[0085] S3. Based on the new data of the equipment obtained by the real-time data monitoring module, update and adjust the graph model in the graph model construction module to adapt the graph model to the new operating conditions and changes of the equipment; set up is the current graph model, where are the parameters of the graphical model, The input data of the graphical model is the device feature data matrix; The goal of updating and adjusting the graph model in the graph model building module is to update , minimize the loss of the graph model on new data, update The calculation formula is: ; in, Represents the new device feature data matrix, where each row represents a device and each column represents a feature. express The corresponding familial defect label vector, is the loss function; Represents the parameters of the updated graphical model; argmin is the minimum index function.

[0086] S4. Based on the equipment data in the database, a graphical model is used to predict equipment family defects, and early warnings are issued in a timely manner for the predicted equipment family defects.

[0087] S5. For devices with family defects, a defect correlation diagram is constructed to track the origin and propagation path of the family defects.

[0088] Identify devices, parts, and timestamps associated with family-specific defects based on real-time device-related data and user feedback data.

[0089] Construct defect association maps for each known familial defect ,in Represents a set of defective nodes, where defective nodes represent defective devices; is the defect edge set, , Representation device and equipment Defect associations between Based on the similarity between devices, supply chain information and timestamps, the defect edges of the defect association graph are Assign defect weights , Representation device and equipment The possibility of defect propagation between them is determined by using the Dijkstra shortest path algorithm to find the defect edge with the largest defect weight, and the source of the defect is determined based on the defect edge.

[0090] S6. Store, manage and retrieve information data related to device family defects.

Claims

1. Automatic identification system of equipment family defects based on spectral clustering and GNN, characterized by: Includes the following modules: The real-time data monitoring module transmits the collected equipment data to the database in real time and stores the pre-processed data for use by the adaptive learning module and the defect source tracing module; A graphical model construction module uses a spectral clustering method combined with a graph neural network (GNN) to construct a graphical model for predicting familial defects. The adaptive learning module automatically updates and adjusts the graph model in the graph model construction module based on the new data output by the real-time data monitoring module; Defect source tracing module, used to track the origin and propagation path of equipment family defects; The predictive maintenance and early warning module uses a graphical model to predict equipment family defects based on the real-time data of the equipment obtained by the real-time data monitoring module, and issues timely early warnings for the predicted equipment family defects; The structured defect library and tag management module is used to store, manage and retrieve all information related to device family defects.

2. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 1 is characterized in that: The graph model construction module specifically includes: The device classification submodule collects the data output by the real-time data monitoring module, generates a similarity matrix between devices based on the similarity of the device data, and uses a spectral clustering method to classify similar devices into the same category; The family defect prediction submodule uses the graph neural network GNN to predict the family defects of the equipment. Based on the results of spectral clustering, it generates a family defect prediction submodule for each equipment category. i Constructing a subgraph ,in is a node set of the subgraph, where the nodes represent the corresponding devices. For the i The edge set of the subgraph of the class device. The edges in the subgraph are used to represent the relationship between devices.

3. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 2 is characterized in that: The graph model construction module is used to build nodes The eigenvector of Perform iterative calculations. In each iterative calculation, the feature of each node is updated as a function of its own and its neighbors' features. The iterative calculation formula is: ； in, Representation node The neighbor set of u express The nodes in It is l The weight matrix of times, is the activation function, 、 Respectively represent l sequence l +1 node calculated by iterative calculation The eigenvector of After multiple iterations, the latest feature vector of each device is obtained, and the output layer of the graph neural network (GNN) predicts device family defects based on the latest feature vector of each device.

4. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 3 is characterized in that: In the graph model building module, the output layer of the graph neural network (GNN) predicts device family defects based on the latest feature vector of each device, specifically including: Using the weight matrix of the GNN output layer The latest feature vector of the device is converted. The conversion formula is: ； Where, Indicates the converted device v The eigenvector of For equipment v The eigenvector of represents the bias term; The activation function of the output layer is based on Calculate the probability that device v has a family defect, which is calculated as: ； Where, Representation device v The probability of familial defects.

5. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 4 is characterized in that: In the adaptive learning module, the graph model in the graph model construction module is automatically updated and adjusted according to the new data output by the real-time data monitoring module. is the current graph model, where are the parameters of the graphical model, The input data of the graphical model is the device feature data matrix; The goal of updating and adjusting the graph model in the graph model building module is to update , minimize the loss of the graph model on new data, update The calculation formula is: ； in, Indicates new device characteristic data, express The corresponding familial defect label vector, is the loss function; Represents the parameters of the updated graphical model; argmin is the minimum index function.

6. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 5 is characterized in that: After calculating the new parameters of the graph model, the graph model construction module uses the new device feature data to and existing data, update the similarity matrix between devices S ; Use the updated S , update the structure and weights of the graph neural network GNN.

7. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 2, 3, 4, 5 or 6, characterized in that: The defect source tracing module determines the equipment, components and timestamps related to the family defect based on the data obtained by the real-time data monitoring module and the user feedback module; Build a defect correlation graph for each device family defect ,in Represents a set of defective nodes, where defective nodes represent defective devices; is the defect edge set, , Representation device and equipment Defect associations between Based on the similarity between devices, supply chain information and timestamps, the defect edges of the defect association graph are Assign defect weights , Representation device and equipment The possibility of defect propagation between them is determined by using the Dijkstra shortest path algorithm to find the defect edge with the largest defect weight, and the source of the defect is determined based on the defect edge.

8. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 7 is characterized in that: When the device family defect automatic identification system detects a new device family defect or receives device defects reported by users, the structured defect library and label management module stores the device defect information and device feature data in the structured defect library. The structured defect library includes not only the description information and feature data of the device defect, but also the related devices, the time when the defect occurred, and the scope of the defect impact.

9. The automatic identification system for equipment family defects based on spectral clustering and GNN according to claim 2, 3, 4, 5, 6 or 8, characterized in that: The structured defect library and label management module also includes a label management submodule for adding, modifying or deleting labels for device defects. The labels are set based on multiple factors such as the type, severity and source of the device defects. The labels are used to quickly retrieve specific types of device defects.

10. The method of the automatic identification system of equipment family defects based on spectral clustering and GNN according to claim 2, 3, 4, 5, 6 or 8, characterized in that: The following steps are involved: S1. Transmit the collected device data to the database in real time, pre-process the device data, and store the pre-processed device data; S2. After classifying the devices using the spectral clustering method, a graph model of the devices is constructed using a graph neural network (GNN). The graph model is used to predict family defects of the devices. S3. Based on the new data of the equipment obtained by the real-time data monitoring module, update and adjust the graph model in the graph model construction module to adapt the graph model to the new operating conditions and changes of the equipment; S4. Based on real-time equipment data, a graphical model is used to predict equipment family defects and timely issue warnings for the predicted equipment family defects; S5. For devices with family defects, a defect correlation diagram is constructed to track the origin and propagation path of the family defects. S6. Store, manage and retrieve information data related to device family defects.

Images