Heterogeneous spectrum and distributed intelligence combined valve hall equipment multi-dimensional defect identification method and equipment

Abstract

The invention relates to a valve hall equipment multidimensional defect identification method combining heterogeneous spectrum and distributed intelligence. The method comprises the following steps: acquiring a multispectral image and preprocessing the multispectral image; constructing a cloud edge collaboration system; constructing a multi-valve-hall knowledge graph dynamic adjustment network; constructing a network training model based on the joint features to obtain fusion features; and carrying out conjoint analysis on the fusion features to obtain a defect type, obtaining a solution corresponding to the defect type in the knowledge graph, and carrying out defect diagnosis and early warning on the valve hall equipment. According to the method, the defect detection efficiency and accuracy are improved by introducing the cloud edge collaborative architecture; problems existing in the multispectral image can be deeply identified; the detection result is adaptively optimized according to the specific application scene and the defect type through the dual networks of the multi-valve hall knowledge graph dynamic adjustment network and the joint feature extraction network, so that the detection precision is improved, and the flexibility and adaptability of the system are enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for defects in valve hall power equipment, and in particular to a method and device for multidimensional defect identification of valve hall equipment that combines heterogeneous spectroscopy and distributed intelligence. Background Technology

[0002] As the core facility of UHVDC transmission, the valve hall is subjected to high voltage and strong magnetic interference for a long time, which can easily lead to equipment defects in a multidimensional manner. The specific manifestations are mainly overheating, discharge, deformation and their coupling.

[0003] Existing detection technologies are insufficient to meet the needs of accurate diagnosis under complex operating conditions. The main problems include: First, for monitoring the operating status of valve hall equipment, traditional methods often rely on a single type of spectrum for defect identification, failing to effectively utilize the combined diagnostic information from visible, infrared, and ultraviolet heterogeneous spectral images. Second, reliance on a single server for centralized processing leads to excessive server load and limited real-time analysis computing power, making it difficult to meet the rapid response requirements of valve hall inspections. Third, existing methods primarily focus on single-dimensional analysis, lacking the ability to perform correlation analysis on the coupling of multidimensional defects. Fourth, existing technologies largely depend on traditional image feature recognition methods, failing to effectively identify high-dimensional features of multispectral images. Fifth, the correlation graph of "equipment parameters - defect features - failure consequences" has not been fully constructed, and industry knowledge and historical defect data have not been digitized; the correlation between multispectral and multimodal data has not been fully considered, and the combined features of images are lacking, making it difficult to form a comprehensive and accurate defect identification knowledge system. Sixth, the knowledge graph has not been dynamically fused, and network models have not been effectively utilized for dynamic adjustments to a large amount of valve hall equipment knowledge. Summary of the Invention

[0004] To address the challenge of identifying multidimensional defects in valve hall equipment under complex operating conditions, the primary objective of this invention is to provide a method for multidimensional defect identification in valve hall equipment that combines heterogeneous spectroscopy and distributed intelligence. This method can improve defect detection efficiency and accuracy, deeply identify problems in multispectral images, and enhance the flexibility and adaptability of the system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a multi-dimensional defect identification method for valve hall equipment combining heterogeneous spectroscopy and distributed intelligence, the method comprising the following sequential steps:

[0006] (1) The valve hall equipment is synchronously monitored. A fixed frequency clock is used to trigger the multispectral images simultaneously, acquire the multispectral images and perform preprocessing to obtain the preprocessed multispectral images;

[0007] (2) Construct a cloud-edge collaborative system to realize distributed intelligent computing;

[0008] (3) Construct a dynamic adjustment network for the knowledge graph of multiple valve halls. Input the detailed operating parameter information and historical defect data of the valve hall equipment into the dynamic adjustment network for the knowledge graph of multiple valve halls for training, obtain the training results, and generate a knowledge graph through the training results.

[0009] (4) Construct a network training model based on joint features. Input the preprocessed multispectral image into the network training model based on joint features for training to obtain fused features;

[0010] (5) Combine the multidimensional association rules in the knowledge graph to perform joint analysis on the fusion features to obtain the defect type, obtain the corresponding solution in the knowledge graph, and perform defect diagnosis and early warning for the valve hall equipment.

[0011] Step (1) specifically includes the following steps:

[0012] (1a) Multiple ultraviolet, infrared and visible light heterogeneous spectral sensors are arranged inside the valve hall to ensure coverage of critical equipment;

[0013] (1b) A fixed-frequency clock triggering mechanism is adopted to trigger the heterogeneous spectral sensor to avoid errors caused by time misalignment;

[0014] (1c) Preprocessing the multispectral images acquired by the ultraviolet, infrared and visible light heterogeneous spectral sensors, the preprocessing including Gaussian filtering to remove noise and gamma correction.

[0015] Step (2) specifically includes the following steps:

[0016] (2a) First, deploy a high-performance server cluster in the cloud for knowledge graph association analysis and multimodal feature fusion; configure lightweight computing nodes on the edge side to be responsible for real-time preprocessing of multispectral data and deployment of cloud models;

[0017] (2b) Intelligent scheduling of computing load is achieved through a dynamic task allocation strategy. The dynamic task allocation strategy specifically refers to:

[0018] (2b1) Set up a task listener to automatically listen for input tasks, extract task attributes and perform preliminary analysis to determine the type of task;

[0019] (2b2) ​​If the task is determined to be lightweight, the resource status of the edge node is checked first. When the resource load of the edge node is lower than the preset threshold, the task is directly sent to the edge node for execution. During the execution of the task on the edge node, the system monitors the task progress and the resource usage of the edge node in real time. If the task execution is hindered due to resource shortage, the task is migrated to the cloud in time.

[0020] (2b3) For complex tasks, the complex tasks are directly included in the cloud scheduling queue. The cloud scheduling center splits and allocates tasks based on the task priority, the type of resources required, and the current resource load of the cloud server cluster.

[0021] (2c) Cloud-edge communication uses differentiated transmission protocols to ensure efficient data flow.

[0022] Step (3) specifically includes the following steps:

[0023] (3a) First, collect detailed operating parameter information of the valve hall equipment, including basic information marked by the equipment at the factory, parameter information generated during operation, and historical maintenance records;

[0024] (3b) Review and analyze historical defect data, including defect type, location of occurrence, cause of defect and remediation measures;

[0025] (3c) Deploy a multi-valve hall knowledge graph dynamic adjustment network, namely BERT-Bi-LSTM-CRF network, in the cloud, and train it on detailed operating parameter information and historical defect data of valve hall equipment to obtain training results. The training results include equipment parameters, defect characteristics, and failure consequences. The multi-valve hall knowledge graph dynamic adjustment network includes an input layer, an output layer, a Bi-LSTM intermediate layer, and a CRF output layer. A BERT pre-training layer is introduced between the input layer and the Bi-LSTM intermediate layer.

[0026] (3d) Construct nodes and relationships in the Neo4j graph library through training results, take equipment parameters as nodes in the graph, and take defect features and failure consequences as the association between nodes to generate a knowledge graph that comprehensively reflects the knowledge of equipment defects.

[0027] Step (4) specifically includes the following steps:

[0028] (4a) Label the preprocessed multispectral images with single and joint defect features, and the labeled multispectral images form the initial dataset;

[0029] (4b) The initial dataset is randomly cropped and flipped to form an augmented dataset;

[0030] (4c) Input the three spectral images in the augmented dataset into the network training model based on joint features, i.e., the improved YOLOv8 network, for training. The improved YOLOv8 network specifically refers to: changing the backbone network of the YOLOv8 network to three parallel channels to form three parallel backbone network branches, which serve as YOLOv8 single-spectral feature extraction layers; adding CBAM modules after the three parallel backbone network branches respectively. The CBAM modules consist of channel attention and spatial attention.

[0031] (4d) The last layer of single-spectral features extracted from the three spectral images through the backbone network branches are enhanced by the CBAM module, and then spliced ​​in the feature fusion layer to obtain deep-level defect features;

[0032] (4e) Input the deep defect features into the classifier for defect type training and judgment, thereby achieving accurate identification of equipment defects.

[0033] Step (5) specifically includes the following steps:

[0034] (5a) The multi-valve hall knowledge graph is dynamically adjusted from the cloud and the network training model based on joint features is sent to the edge, and multiple edge devices are intelligently scheduled.

[0035] (5b) Input the image to be detected into the corresponding edge. If the load on the edge is light, the joint feature is obtained and the information of the device to be detected is recorded. If the load on the edge is too heavy, the defect identification task corresponding to the image to be detected is assigned to other edges and the judgment is made through other edges.

[0036] (5c) Use the high-performance Neo4j graph database to retrieve equipment and feature information and defects, detect corresponding solutions, and feed the solutions back to the operation and maintenance personnel to assist them in making decisions.

[0037] Another object of the present invention is to provide an electronic device comprising:

[0038] Processor; and

[0039] A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the multidimensional defect identification method for valve hall equipment combining heterogeneous spectroscopy and distributed intelligence as described above.

[0040] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the multidimensional defect identification method for valve hall equipment combining heterogeneous spectroscopy and distributed intelligence as described above.

[0041] As can be seen from the above technical solution, the beneficial effects of this invention are as follows: First, by introducing a cloud-edge collaborative architecture, the limitations of insufficient computing power and slow response during peak periods in traditional single-service systems are improved, thereby increasing the efficiency and accuracy of defect detection. Second, by extracting single-spectral features from multispectral images and extracting joint spectral features from fused images, problems existing in multispectral images can be identified at a deeper level. Third, by designing a domain knowledge graph, industry knowledge and defect generation mechanisms are digitally represented, and the Neo4j high-efficiency graph database provides support for fault source identification, revealing the causes of defects more deeply and achieving a user-friendly approach for maintenance personnel. Fourth, by using a dual network of dynamic adjustment network and joint feature extraction network based on a multi-valve hall knowledge graph, the detection results are adaptively optimized according to specific application scenarios and defect types. This combination of dual networks not only improves detection accuracy but also enhances the flexibility and adaptability of the system, providing strong support for the intelligent operation and maintenance of valve hall equipment. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a flowchart of the multispectral image acquisition and preprocessing method in this invention;

[0044] Figure 3 This is a schematic diagram of knowledge graph generation in this invention;

[0045] Figure 4 This is a system architecture diagram of the present invention. Detailed Implementation

[0046] like Figure 1 , Figure 4 As shown, a multidimensional defect identification method for valve hall equipment combining heterogeneous spectroscopy and distributed intelligence is presented. This method includes the following sequential steps:

[0047] (1) The valve hall equipment is synchronously monitored. A fixed frequency clock is used to trigger the multispectral images simultaneously, acquire the multispectral images and perform preprocessing to obtain the preprocessed multispectral images;

[0048] (2) Construct a cloud-edge collaborative system to realize distributed intelligent computing; the cloud-edge collaborative system can effectively distribute computing tasks to powerful servers in the cloud and lightweight computing nodes at the edge to reduce the load on a single server and improve the speed and real-time performance of data processing.

[0049] (3) Construct a dynamic adjustment network for the knowledge graph of multiple valve halls. Input the detailed operating parameter information and historical defect data of the valve hall equipment into the dynamic adjustment network for the knowledge graph of multiple valve halls for training, obtain the training results, and generate a knowledge graph through the training results.

[0050] (4) Construct a network training model based on joint features. Input the preprocessed multispectral image into the network training model based on joint features for training to obtain fused features;

[0051] (5) Combine the multidimensional association rules in the knowledge graph to perform joint analysis on the fusion features to obtain the defect type, obtain the corresponding solution in the knowledge graph, and perform defect diagnosis and early warning for the valve hall equipment.

[0052] like Figure 2 As shown, step (1) specifically includes the following steps:

[0053] (1a) Arrange multiple ultraviolet, infrared and visible light heterogeneous spectral sensors inside the valve hall to ensure coverage of key equipment; the key equipment refers to equipment that undertakes core functions such as power conversion, voltage control and current transmission, and whose failure has serious consequences.

[0054] (1b) A fixed-frequency clock triggering mechanism is adopted to trigger the heterogeneous spectral sensor to avoid errors caused by time misalignment;

[0055] (1c) Preprocessing the multispectral images acquired by the ultraviolet, infrared and visible light heterogeneous spectral sensors, the preprocessing including Gaussian filtering to remove noise and gamma correction.

[0056] Step (2) specifically includes the following steps:

[0057] (2a) First, deploy a high-performance server cluster in the cloud for knowledge graph association analysis and multimodal feature fusion; configure lightweight computing nodes on the edge side to be responsible for real-time preprocessing of multispectral data and deployment of cloud models;

[0058] (2b) Intelligent scheduling of computing load is achieved through a dynamic task allocation strategy. The dynamic task allocation strategy specifically refers to:

[0059] (2b1) Set up a task listener to automatically listen to the input tasks, extract task attributes and perform preliminary analysis to determine the type of the task; the task type is divided into lightweight tasks and complex tasks; lightweight tasks refer to tasks that require less server resources and computing power, have lower training intensity, or have less detection data during execution, while complex tasks are the opposite.

[0060] (2b2) ​​If the task is determined to be lightweight, the resource status of the edge node is checked first. When the resource load of the edge node is lower than the preset threshold, the task is directly sent to the edge node for execution. During the execution of the task on the edge node, the system monitors the task progress and the resource usage of the edge node in real time. If the task execution is hindered due to resource shortage, the task is migrated to the cloud in time.

[0061] (2b3) For complex tasks, the complex tasks are directly included in the cloud scheduling queue. The cloud scheduling center splits and allocates tasks based on the task priority, the type of resources required, and the current resource load of the cloud server cluster.

[0062] (2c) Cloud-edge communication uses differentiated transmission protocols to ensure efficient data flow.

[0063] The valve hall has a large equipment area and many important devices. Relying on a single server is difficult to meet the timely response of all important devices. Therefore, multiple servers forming multiple edge terminals can effectively handle the situation and support cross-edge terminal processing to meet the needs of high-efficiency processing.

[0064] like Figure 3 As shown, step (3) specifically includes the following steps:

[0065] (3a) First, collect detailed operating parameter information of the valve hall equipment from multiple sources such as the equipment historical operation database and equipment nameplate. The detailed operating parameter information includes the basic information marked by the equipment at the factory, the parameter information generated during operation, and historical maintenance records.

[0066] (3b) Review and analyze historical defect data, including defect type, location, cause and repair measures; this information helps to identify the patterns and characteristics of equipment defects and provides experience, data and standard support for defect identification.

[0067] (3c) Deploy a multi-valve hall knowledge graph dynamic adjustment network in the cloud, train it on detailed operating parameter information and historical defect data of the valve hall equipment, and obtain training results. The training results include equipment parameters, defect features, and failure consequences. The multi-valve hall knowledge graph dynamic adjustment network includes an input layer, an output layer, a Bi-LSTM intermediate layer, and a CRF output layer. A BERT pre-training layer is introduced between the input layer and the Bi-LSTM intermediate layer. The input data of the input layer is character vectors and additional features as training text. The Bi-LSTM intermediate layer models the input text sequence, and the CRF output layer outputs the recognition results. The BERT pre-training layer is a pre-training mechanism that can predict randomly masked words based on contextual semantic information, and can better learn contextual content features. Then, joint extraction and knowledge aggregation are performed to obtain the required information.

[0068] (3d) Construct nodes and relationships in the Neo4j graph library through training results, take equipment parameters as nodes in the graph, and take defect features and failure consequences as the association between nodes to generate a knowledge graph that comprehensively reflects the knowledge of equipment defects.

[0069] Because the number of valve chambers can change dynamically, and the equipment information within each chamber also changes with actual usage, the cloud-deployed model continuously optimizes itself based on input data, thereby constantly refining the map. Furthermore, as new defect data is continuously added, the map will be constantly updated to reflect the latest equipment status and defect information. Simultaneously, by integrating knowledge from other domains, the accuracy and practicality of the map are further improved.

[0070] The cloud distributes the trained knowledge graph to the edge, which can then choose whether to iterate based on the monitored area's device information, or choose to feed it back to the cloud for iteration based on its own load. For example... Figure 3 As shown, this involves sorting out information on valve hall equipment and historical defect information, as well as planning the relevant field standards. This results in a complete domain knowledge graph, which forms a knowledge base that can characterize valve hall equipment defects, providing a reliable basis for valve hall equipment identification and early warning.

[0071] Step (4) specifically includes the following steps:

[0072] (4a) Fully consider the scenario of multidimensional defect coupling, and annotate the preprocessed multispectral images with single defect features and joint features. The annotated multispectral images form the initial dataset. It is necessary to refer to the defect types and corresponding features defined in the knowledge graph to ensure the consistency and standardization of the annotation.

[0073] (4b) The initial dataset is randomly cropped and flipped to form an augmented dataset; ensuring that there are corresponding images in multiple angles, thereby expanding to form images that can simulate different perspectives and segments in real scenes;

[0074] (4c) Input the three spectral images from the augmented dataset into the network training model based on joint features, i.e., the improved YOLOv8 network, for training. The improved YOLOv8 network specifically refers to: changing the backbone network of the YOLOv8 network to three parallel channels, i.e., copying one backbone network to become three backbone networks, forming three parallel backbone network branches, which serve as YOLOv8 single-spectral feature extraction layers; adding CBAM modules after the three parallel backbone network branches respectively. The CBAM module consists of channel attention and spatial attention. The YOLOv8 network has efficient target detection performance. By extracting features from the input ultraviolet, infrared, and visible light spectral images, it can accurately identify single defect features in the image.

[0075] (4d) The last layer of single-spectral features extracted from the three spectral images through the backbone network branches are enhanced by the CBAM module, and then spliced ​​in the feature fusion layer to obtain deep-level defect features. This fusion method can make full use of the complementarity of different spectral images in reflecting equipment defects, and improve the accuracy and robustness of defect identification.

[0076] (4e) Input the deep defect features into the classifier for defect type training and judgment, thereby achieving accurate identification of equipment defects.

[0077] Step (5) specifically includes the following steps:

[0078] (5a) The multi-valve hall knowledge graph is dynamically adjusted from the cloud and the network training model based on joint features is sent to the edge, and multiple edge devices are intelligently scheduled.

[0079] (5b) Input the image to be detected into the corresponding edge. If the edge has a light load, obtain the joint features and record the information of the device to be detected. The edge has a light load, which means that the edge server's CPU, memory, and disk space are all in a low load state. If the edge has a heavy load, assign the defect identification task corresponding to the image to be detected to other edges and make a judgment through other edges. The edge has a heavy load, which means that the edge server is in a high load state and the CPU, memory, and disk space resources have been occupied in large quantities.

[0080] (5c) Use the high-performance Neo4j graph database to retrieve equipment and feature information and defects, detect corresponding solutions, and feed the solutions back to the operation and maintenance personnel to assist them in making decisions.

[0081] In summary, this invention improves upon the limitations of traditional single-service computing power and slow response during peak periods by introducing a cloud-edge collaborative architecture, thereby enhancing defect detection efficiency and accuracy. By extracting single-spectral features from multispectral images and extracting joint spectral features from fused images, it can deeply identify problems present in multispectral images. Through the design of a domain knowledge graph, industry knowledge and defect generation mechanisms are digitally represented, and the Neo4j high-efficiency graph database provides fault source support for feature identification, revealing defect causes more deeply and achieving a user-friendly approach for maintenance personnel. The dual network approach—a dynamic adjustment network based on the multi-valve hall knowledge graph and a joint feature extraction network—adaptively optimizes detection results according to specific application scenarios and defect types. This combination of networks not only improves detection accuracy but also enhances the system's flexibility and adaptability, providing strong support for the intelligent operation and maintenance of valve hall equipment.

[0082] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for multidimensional defect identification of valve hall equipment combining heterogeneous spectroscopy and distributed intelligence, characterized in that: The method includes the following steps in sequence: (1) The valve hall equipment is synchronously monitored. A fixed frequency clock is used to trigger the multispectral images simultaneously, acquire the multispectral images and preprocess them to obtain the preprocessed multispectral images. (2) Construct a cloud-edge collaborative system to realize distributed intelligent computing; (3) Construct a dynamic adjustment network for the knowledge graph of multiple valve halls. Input the detailed operating parameter information and historical defect data of the valve hall equipment into the dynamic adjustment network for the knowledge graph of multiple valve halls for training, obtain the training results, and generate a knowledge graph through the training results. (4) Construct a network training model based on joint features. Input the preprocessed multispectral image into the network training model based on joint features for training to obtain fused features; (5) Combine the multidimensional association rules in the knowledge graph to perform joint analysis on the fusion features to obtain the defect type, obtain the corresponding solution in the knowledge graph, and perform defect diagnosis and early warning for the valve hall equipment.

2. The method for multidimensional defect identification of valve hall equipment combining heterogeneous spectroscopy and distributed intelligence according to claim 1, characterized in that: Step (1) specifically includes the following steps: (1a) Multiple ultraviolet, infrared and visible light heterogeneous spectral sensors are arranged inside the valve hall to ensure coverage of critical equipment; (1b) A fixed-frequency clock triggering mechanism is adopted to trigger the heterogeneous spectral sensor to avoid errors caused by time misalignment; (1c) Preprocessing the multispectral images acquired by the ultraviolet, infrared and visible light heterogeneous spectral sensors, the preprocessing including Gaussian filtering to remove noise and gamma correction.

3. The method for multidimensional defect identification of valve hall equipment combining heterogeneous spectroscopy and distributed intelligence according to claim 1, characterized in that: Step (2) specifically includes the following steps: (2a) First, deploy a high-performance server cluster in the cloud for knowledge graph association analysis and multimodal feature fusion; configure lightweight computing nodes on the edge side to be responsible for real-time preprocessing of multispectral data and deployment of cloud models; (2b) Intelligent scheduling of computing load is achieved through a dynamic task allocation strategy. The dynamic task allocation strategy specifically refers to: (2b1) Set up a task listener to automatically listen for input tasks, extract task attributes and perform preliminary analysis to determine the type of task; (2b2) ​​If the task is determined to be lightweight, the resource status of the edge node is checked first. When the resource load of the edge node is lower than the preset threshold, the task is directly sent to the edge node for execution. During the execution of the task on the edge node, the system monitors the task progress and the resource usage of the edge node in real time. If the task execution is hindered due to resource shortage, the task is migrated to the cloud in time. (2b3) For complex tasks, the complex tasks are directly included in the cloud scheduling queue. The cloud scheduling center splits and allocates tasks based on the task priority, the type of resources required, and the current resource load of the cloud server cluster. (2c) Cloud-edge communication uses differentiated transmission protocols to ensure efficient data flow.

4. The method for multidimensional defect identification of valve hall equipment combining heterogeneous spectroscopy and distributed intelligence according to claim 1, characterized in that: Step (3) specifically includes the following steps: (3a) First, collect detailed operating parameter information of the valve hall equipment, including basic information marked by the equipment at the factory, parameter information generated during operation, and historical maintenance records; (3b) Review and analyze historical defect data, including defect type, location of occurrence, cause of defect and remediation measures; (3c) Deploy a multi-valve hall knowledge graph dynamic adjustment network in the cloud, train it on detailed operating parameter information and historical defect data of valve hall equipment, and obtain training results. The training results include equipment parameters, defect characteristics and failure consequences. The multi-valve hall knowledge graph dynamic adjustment network includes an input layer, an output layer, a Bi-LSTM intermediate layer and a CRF output layer. A BERT pre-training layer is introduced between the input layer and the Bi-LSTM intermediate layer. (3d) Construct nodes and relationships in the Neo4j graph library through training results, take equipment parameters as nodes in the graph, and take defect features and failure consequences as the association between nodes to generate a knowledge graph that comprehensively reflects the knowledge of equipment defects.

5. The method for multidimensional defect identification of valve hall equipment combining heterogeneous spectroscopy and distributed intelligence according to claim 1, characterized in that: Step (4) specifically includes the following steps: (4a) Label the preprocessed multispectral images with single and joint defect features, and the labeled multispectral images form the initial dataset; (4b) The initial dataset is randomly cropped and flipped to form an augmented dataset; (4c) Input the three spectral images in the augmented dataset into the network training model based on joint features, i.e., the improved YOLOv8 network, for training. The improved YOLOv8 network specifically refers to: changing the backbone network of the YOLOv8 network to three parallel channels to form three parallel backbone network branches, which serve as YOLOv8 single-spectral feature extraction layers; adding CBAM modules after the three parallel backbone network branches respectively. The CBAM modules consist of channel attention and spatial attention. (4d) The last layer of single-spectral features extracted from the three spectral images through the backbone network branches are enhanced by the CBAM module, and then spliced ​​in the feature fusion layer to obtain deep-level defect features; (4e) Input the deep defect features into the classifier for defect type training and judgment, thereby achieving accurate identification of equipment defects.

6. The method for multidimensional defect identification of valve hall equipment combining heterogeneous spectroscopy and distributed intelligence according to claim 1, characterized in that: Step (5) specifically includes the following steps: (5a) The multi-valve hall knowledge graph is dynamically adjusted from the cloud and the network training model based on joint features is sent to the edge, and multiple edge devices are intelligently scheduled. (5b) Input the image to be detected into the corresponding edge. If the load on the edge is light, the joint feature is obtained and the information of the device to be detected is recorded. If the load on the edge is too heavy, the defect identification task corresponding to the image to be detected is assigned to other edges and the judgment is made through other edges. (5c) Use the high-performance Neo4j graph database to retrieve equipment and feature information and defects, detect corresponding solutions, and feed the solutions back to the operation and maintenance personnel to assist them in making decisions.

7. An electronic device, comprising: processor; as well as A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the multidimensional defect identification method for valve hall equipment combining heterogeneous spectroscopy and distributed intelligence as described in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the multidimensional defect identification method for valve hall equipment combining heterogeneous spectroscopy and distributed intelligence as described in any one of claims 1-6.

Images