A method and system for point cloud classification in substation scenarios
By using KP-FCNN and point cloud classification constraint graph neural network to perform multi-level classification of point clouds in substation scenarios, the problem of poor point cloud segmentation effect in existing technologies is solved, and more accurate point cloud segmentation and classification are achieved, supporting the construction of individual models.
Patent Information
- Application Number
- CN202311662246.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-05
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-12-05
AI Technical Summary
Existing technologies perform poorly in point cloud data classification in substation scenarios with dense equipment and complex structures, failing to achieve multi-level classification and affecting the subsequent construction of individual models.
KP-FCNN is used to classify point clouds in substation scenarios into three levels: region, equipment, and component. Intermediate results are verified and corrected by a pre-set point cloud classification constraint graph neural network. Knowledge graph learning is performed by embedding knowledge graph information into a network to improve classification accuracy.
Multi-level classification of point clouds in substation scenarios was achieved, improving the accuracy of point cloud segmentation and providing better data support for subsequent individual model construction.
Smart Images

Figure CN117636043B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of point cloud classification technology, and in particular to a point cloud classification method and system for substation scenarios. Background Technology
[0002] Individual digital twin models of substations are fundamental for intelligent operation and maintenance, safety risk management, and other applications in substations, and are key to the automation and intelligent development of the power industry. Currently, the construction of individual digital twin models typically employs automated modeling methods based on scene-specific laser point clouds. However, due to the dense equipment and facilities and structural replication characteristics of power operation sites, only partial surface point clouds of the equipment and facilities can be obtained. To proceed with the subsequent construction of individual component-level models, it is necessary to classify the point cloud data.
[0003] Current point cloud data classification methods typically rely on spatial information and reflection intensity, using artificial intelligence to extract and classify point cloud features. However, this approach performs poorly in substation scenarios with dense equipment and repetitive structures, and it cannot adequately consider multiple levels such as equipment and components. Therefore, a multi-level substation point cloud classification method is needed. Summary of the Invention
[0004] This invention provides a method and system for point cloud classification in substation scenarios, achieving more accurate multi-level point cloud classification in substations.
[0005] To address the aforementioned technical problems, embodiments of the present invention provide a point cloud classification method for substation scenes, including:
[0006] Obtain the scene point cloud to be classified in the substation;
[0007] The point cloud of the scene to be classified is labeled with region type, device type and component type in sequence using KP-FCNN to obtain intermediate classification results;
[0008] Based on a preset point cloud classification constraint graph neural network, the intermediate classification results are subjected to type labeling verification and correction to obtain scene point cloud segmentation results; wherein, the preset point cloud classification constraint graph neural network is obtained by training a knowledge graph information embedding network combined with a graph neural network; the knowledge graph information embedding network is obtained by using the R-GCN model to learn knowledge graph triples of the relationship between substation areas, equipment and components.
[0009] Understandably, compared to existing technologies, the method provided by this invention uses KP-FCNN to sequentially label scene point clouds according to three levels: region, device, and component. It then fuses the attributes of the three types of substation scene point clouds after classification at these three levels, giving the same substation scene point cloud a hierarchical attribute of "region-device-component". Subsequently, a graph neural network embedded with a knowledge graph network is used to verify and correct the intermediate classification results, further improving the accuracy of point cloud classification. This solves the shortcomings of current point cloud segmentation technologies in multi-scale, multi-level substation scenarios, such as poor point cloud segmentation results and lack of hierarchical classification, providing data support for subsequent single-unit model construction or other applications.
[0010] Furthermore, the step of sequentially labeling the scene point cloud to be classified with region type, device type, and component type using KP-FCNN to obtain intermediate classification results specifically includes:
[0011] The point cloud region type of the scene point cloud to be classified is classified by KP-FCNN to obtain the point cloud region classification result. Based on the point cloud region classification result, the region type of the point cloud data in the scene point cloud to be classified is labeled, and the point cloud of each region is obtained by density-based clustering algorithm.
[0012] The point cloud device type is classified for each region using KP-FCNN to obtain the point cloud device classification result. Based on the point cloud device classification result, the device type of the point cloud data in the scene point cloud to be classified is labeled, and the point cloud of each device is obtained by density-based clustering algorithm.
[0013] KP-FCNN is used to classify the point cloud component types of each device's point cloud to obtain point cloud component classification results. Based on the point cloud component classification results, the component types of the point cloud data in the scene point cloud to be classified are labeled, and the point cloud of each component is obtained by density-based clustering algorithm.
[0014] The scene point cloud labeled with region type, device type, and component type is used as the intermediate classification result.
[0015] Understandably, the method provided by this invention classifies scene point clouds into three levels—region, device, and component—using KP-FCNN. Then, it fuses the attributes of the three types of substation scene point clouds after these three-level classifications, giving each substation scene point cloud a hierarchical attribute of "region-device-component." Furthermore, inputting the point cloud into KP-FCNN at different levels reduces the network size and enables rapid type labeling of the scene point clouds to be classified.
[0016] Furthermore, before classifying the scene point cloud to be classified using KP-FCNN, the method further includes: performing point cloud enhancement and ground point filtering on the scene point cloud to be classified.
[0017] It is understood that the method provided by this invention enhances and filters the scene point cloud before classification, thereby laying a working foundation for subsequent classification processing while preserving geometric features and improving the accuracy of subsequent classification.
[0018] Furthermore, the step of performing type labeling verification and correction on the intermediate classification results based on a preset point cloud classification constraint graph neural network to obtain scene point cloud segmentation results specifically includes:
[0019] The intermediate classification result is converted into a graph structure classification result; wherein, each node in the graph structure classification result is a point cloud data in the intermediate classification result, each node attribute is the type label of the corresponding point cloud data, and each edge is the hierarchical relationship of the type labels between two point cloud data in the intermediate classification result.
[0020] Using the graph structure classification result as input, the attributes of each node in the graph structure are verified and corrected through a preset point cloud classification constraint graph neural network to obtain the scene point cloud segmentation result.
[0021] It is understood that the method provided by the present invention converts the intermediate classification results into a graph structure, with the nodes in the graph structure corresponding to the point cloud data and the edges in the graph structure corresponding to the hierarchical relationship between different point cloud data, which facilitates the subsequent verification and error correction of the classification results through a preset point cloud classification constraint neural network.
[0022] Furthermore, the knowledge graph information embedding network is obtained by learning the knowledge graph through knowledge triples of the relationships between substation areas, equipment, and components using the R-GCN model, specifically including:
[0023] Obtain equipment management information and design information of substation equipment and facilities as structured data;
[0024] Obtain substation equipment management and maintenance information, substation equipment maintenance manual information, and substation inspection work guide information as text data;
[0025] Based on at least one of the structured data and text data, a knowledge triplet of the relationship between substation area, equipment and components is obtained;
[0026] Using the R-GCN model, a knowledge graph information embedding network is obtained by using knowledge triples of the relationships between the substation area, equipment, and components.
[0027] It is understood that the method provided by this invention collects data related to the substation system in structured and textual form, and converts it into knowledge triples of the relationship between substation areas, equipment and components, which serve as the data source for constructing a knowledge graph. This provides an embedding network for knowledge graph information for graph neural networks, assisting in verification and error correction.
[0028] Accordingly, embodiments of the present invention also provide a substation scene point cloud classification system, including:
[0029] The scene point cloud acquisition module is used to acquire scene point clouds to be classified in the substation;
[0030] The type classification module is used to sequentially label the point cloud of the scene to be classified with region type, device type and component type using KP-FCNN to obtain intermediate classification results;
[0031] The verification and correction module is used to perform type label verification and correction on the intermediate classification results according to the preset point cloud classification constraint graph neural network to obtain the scene point cloud segmentation result; wherein, the preset point cloud classification constraint graph neural network is obtained by training a knowledge graph information embedding network combined with a graph neural network; the knowledge graph information embedding network is obtained by learning the knowledge graph triples of the relationship between substation areas, equipment and components through the R-GCN model.
[0032] Understandably, compared to existing technologies, the system provided by this invention uses a KP-FCNN network to sequentially construct multi-level point cloud segmentation results for substation site clouds, including regions, equipment, and components. After three levels of classification, the point clouds of the three types of substation scenarios are fused by attribute integration. Subsequently, a graph neural network with embedded knowledge graph network is used to verify and correct the intermediate classification results, further improving the accuracy of point cloud classification. This solves the shortcomings of current point cloud segmentation technologies in multi-scale, multi-level substation scenarios, such as poor point cloud segmentation results and lack of hierarchical classification, providing data support for subsequent single-unit model construction or other applications.
[0033] Furthermore, the type classification module specifically includes:
[0034] The regional point cloud classification submodule is used to classify the point cloud region type of the scene point cloud to be classified using KP-FCNN, obtain the point cloud region classification result, label the region type of the point cloud data in the scene point cloud to be classified according to the point cloud region classification result, and obtain the point cloud of each region through a density-based clustering algorithm.
[0035] The device point cloud classification submodule is used to classify the point cloud device type of each region's point cloud using KP-FCNN to obtain the point cloud device classification result. Based on the point cloud device classification result, the device type of the point cloud data in the scene point cloud to be classified is labeled, and the point cloud of each device is obtained through a density-based clustering algorithm.
[0036] The component point cloud classification submodule is used to classify the point cloud component types of each device's point cloud using KP-FCNN to obtain point cloud component classification results. Based on the point cloud component classification results, the component types of the point cloud data in the scene point cloud to be classified are labeled, and the point cloud of each component is obtained through a density-based clustering algorithm.
[0037] The intermediate classification result determination submodule is used to take the scene point cloud after labeling the region type, device type, and component type as the intermediate classification result.
[0038] Understandably, the system provided by this invention classifies scene point clouds into three levels—region, device, and component—using KP-FCNN. It then fuses the attributes of the three types of substation scene point clouds after these three-level classifications, giving each substation scene point cloud a hierarchical attribute of "region-device-component." Furthermore, inputting the point cloud into KP-FCNN at different levels reduces the network size and enables rapid type labeling of the scene point clouds to be classified.
[0039] Furthermore, before the regional point cloud classification submodule, there is also a data preprocessing submodule, which is used to perform point cloud enhancement and ground point filtering processing on the scene point cloud to be classified.
[0040] It is understood that the system provided by this invention enhances and filters the scene point cloud before classification, thereby laying a working foundation for subsequent classification processing while preserving geometric features and improving the accuracy of subsequent classification.
[0041] Furthermore, the verification and correction module specifically includes:
[0042] The graph structure conversion submodule is used to convert the intermediate classification result into a graph structure classification result; wherein, each node in the graph structure classification result is a point cloud data in the intermediate classification result, each node attribute is the type label of the corresponding point cloud data, and each edge is the hierarchical relationship of the type labels between two point cloud data in the intermediate classification result;
[0043] The node attribute correction submodule is used to take the graph structure classification result as input, and through a preset point cloud classification constraint graph neural network, to verify and correct the attributes of each node in the graph structure to obtain the scene point cloud segmentation result.
[0044] It is understood that the system provided by this invention converts intermediate classification results into a graph structure, with nodes in the graph structure corresponding to point cloud data and edges in the graph structure corresponding to hierarchical relationships between different point cloud data, which facilitates subsequent verification and error correction of classification results through a preset point cloud classification constraint neural network.
[0045] Furthermore, the knowledge graph information embedding network is obtained by learning the knowledge graph through knowledge triples of the relationships between substation areas, equipment, and components using the R-GCN model, specifically including:
[0046] The structured data acquisition submodule is used to acquire equipment management information and design information of substation equipment and facilities as structured data.
[0047] The text data acquisition submodule is used to acquire substation equipment management and maintenance information, substation equipment maintenance manual information, and substation inspection work guide information as text data.
[0048] The triplet determination submodule is used to obtain knowledge triples of the association between substation areas, equipment and components based on at least one of the structured data and text data.
[0049] The embedded network determination submodule is used to obtain the embedded network of knowledge graph information by using the knowledge triples of the relationship between the substation area, equipment and components through the R-GCN model.
[0050] It is understood that the system provided by this invention collects data related to the substation system in structured and textual form, and converts it into knowledge triples of the relationship between substation areas, equipment and components, which serve as the data source for constructing a knowledge graph. This provides an embedding network for knowledge graph information for graph neural networks, assisting in verification and error correction. Attached Figure Description
[0051] Figure 1 : A flowchart illustrating the steps of a point cloud classification method for a substation scene provided in an embodiment of the present invention;
[0052] Figure 2 : This is an overall flowchart of a point cloud classification method for a substation scene provided in an embodiment of the present invention;
[0053] Figure 3 : A schematic diagram of the structure of a substation scene point cloud classification system provided in an embodiment of the present invention;
[0054] Figure 4 : This is a schematic diagram of the internal structure of the type classification module in a substation scene point cloud classification system provided in an embodiment of the present invention;
[0055] Figure 5 This is a schematic diagram of the internal structure of a type classification module, including a data preprocessing submodule, in a point cloud classification system for a substation scene provided in an embodiment of the present invention.
[0056] Figure 6 This is a schematic diagram of the internal structure of a verification and correction module in a substation scene point cloud classification system provided in an embodiment of the present invention.
[0057] Figure 7 This is a schematic diagram of the internal structure of a knowledge graph embedding network in a substation scene point cloud classification system provided in an embodiment of the present invention. Detailed Implementation
[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] In multi-scale, multi-level substation scenarios, hierarchical classification of substation point clouds requires processing large amounts of point cloud data, placing demands on both computing resources and algorithms. To improve the automation level of substations and provide data support for subsequent individual model construction, multi-level point cloud segmentation is necessary.
[0060] Example 1
[0061] Please refer to Figure 1 The flowchart of a point cloud classification method for a substation scene provided in an embodiment of the present invention includes steps S101-S103, and the specific steps are as follows.
[0062] S101: Obtain the scene point cloud to be classified in the substation.
[0063] S102: The point cloud of the scene to be classified is labeled with region type, device type and component type in sequence using KP-FCNN to obtain intermediate classification results.
[0064] It should be noted that before classifying the point cloud regions into types, the process includes point cloud enhancement and ground point filtering. This preprocessing ensures that the number of point clouds meets the requirements of the point cloud segmentation network, while also reducing the complexity of subsequent classification and improving classification accuracy.
[0065] In this embodiment, KP-FCNN is used to classify the point cloud regions of the scene to be classified, resulting in point cloud region classification results. Based on these classification results, the region types of the point cloud data in the scene to be classified are labeled. If a point in the original scene is not labeled, it is labeled with the same attribute as the nearest labeled point. For point clouds of the same category, a density-based DBSCAN clustering algorithm is used for segmentation, spatially clustering point clouds of the same category to obtain point clouds for each region.
[0066] In this embodiment, the point clouds of each region are first preprocessed, and then KP-FCNN is used to classify the device types of the point clouds in each region to obtain the device classification results. Based on the device classification results, the device types of the point cloud data in the scene to be classified are labeled. If a point is not labeled in the original scene, it is labeled with the same attribute as the nearest labeled point. For point clouds of the same category, a density-based DBSCAN clustering algorithm is used for segmentation, spatially clustering point clouds of the same category to obtain the point clouds for each device.
[0067] In this embodiment, the point clouds of each device are first processed. KP-FCNN is used to classify the point cloud component types of each device's point cloud, resulting in a point cloud component classification result. Based on this classification result, the component types of the point cloud data in the scene point cloud to be classified are labeled. If a point in the original scene is not labeled, it is labeled with the same attribute as the nearest labeled point. For point clouds of the same category, a density-based DBSCAN clustering algorithm is used for segmentation, spatially clustering point clouds of the same category to obtain the point clouds of each component.
[0068] In this embodiment, the scene point cloud labeled with region type, device type, and component type is used as the intermediate classification result.
[0069] S103: Based on the preset point cloud classification constraint graph neural network, the intermediate classification results are verified and corrected by type labeling to obtain the scene point cloud segmentation results.
[0070] In this embodiment, the step of performing type label verification and correction on the intermediate classification results according to the preset point cloud classification constraint graph neural network to obtain the scene point cloud segmentation result specifically includes: converting the intermediate classification results into graph structure classification results; wherein, each node in the graph structure classification result is a point cloud data in the intermediate classification result, each node attribute is the type label of the corresponding point cloud data, and each edge is the hierarchical relationship of the type labels between two point cloud data in the intermediate classification result; taking the graph structure classification result as input, and verifying and correcting the attributes of each node in the graph structure through the preset point cloud classification constraint graph neural network to obtain the scene point cloud segmentation result.
[0071] As a preferred embodiment, this example represents the undirected graph using an adjacency matrix plus a node feature matrix. The adjacency matrix is a two-dimensional matrix of size NxN, where N is the number of nodes in the graph. Each element A[i][j] indicates whether there is a connection (edge) between node i and node j, represented by 1 or 0. The node feature matrix is an NxD matrix, where N is the number of nodes in the graph and D is the dimension of the node features. Each row corresponds to a node, and each column corresponds to a feature.
[0072] As a preferred approach, a graph neural network is used to examine the substation scene point cloud, which is assigned classification attributes for regions, equipment, and components. If any point in the substation scene point cloud is incorrect, the corresponding point in the substation scene point cloud is corrected based on the knowledge graph.
[0073] In this embodiment, the preset point cloud classification constraint graph neural network is obtained by training a knowledge graph information embedding network combined with a graph neural network; the knowledge graph information embedding network is obtained by using the R-GCN model to learn the knowledge triples of the relationship between substation areas, equipment and components.
[0074] It should be noted that the graph neural network can be GraphSAGE (Graph Sample and Aggregated).
[0075] It should be noted that the knowledge graph information embedding network is obtained by learning the knowledge triples of the relationships between substation areas, equipment, and components through the R-GCN model. Specifically, this includes: acquiring equipment management information and design information of substation equipment and facilities as structured data; acquiring substation equipment management and maintenance information, substation equipment maintenance manual information, and substation inspection work guide information as text data; obtaining knowledge triples of the relationships between substation areas, equipment, and components based on at least one of the structured data and text data; and obtaining the knowledge graph information embedding network through the knowledge triples of the relationships between substation areas, equipment, and components using the R-GCN model.
[0076] As a preferred approach, the structured data is processed in a rule-based manner to obtain knowledge triples of the relationships between substation areas, equipment, and components; based on the text data, natural language processing technology is used to extract knowledge triples of the relationships between substation areas, equipment, and components; all knowledge triples of the relationships between substation areas, equipment, and components are stored together, and duplicate knowledge triples are deleted.
[0077] For the preferred option, please refer to Figure 2This is an overall flowchart of a substation scene point cloud classification method provided in an embodiment of the present invention. As can be seen, after preprocessing, the substation scene point cloud is sequentially labeled with region, equipment, and component types using KP-FCNN. A corresponding graph structure is constructed for the substation scene point cloud with region, equipment, and component labeled attributes. The input graph structure is then verified and corrected using a graph neural network with a knowledge graph, GraphSAGE, to obtain the point cloud segmentation result.
[0078] The method provided by this invention uses KP-FCNN to sequentially label scene point clouds according to three levels: region, device, and component. After classification by these three levels, the three types of substation scene point clouds are fused to give the same substation scene point cloud a hierarchical attribute of "region-device-component". Subsequently, a graph neural network with embedded knowledge graph network is used to verify and correct the intermediate classification results, further improving the accuracy of point cloud classification. This method solves the shortcomings of current point cloud segmentation technology in multi-scale and multi-level substation scenarios, which has poor point cloud segmentation effect and lack of hierarchical classification. It provides data support for subsequent single-unit model construction or other applications.
[0079] Example 2
[0080] Please refer to Figure 3 This is a schematic diagram of the structure of a substation scene point cloud classification system provided in an embodiment of the present invention. The substation scene point cloud classification system includes a scene point cloud acquisition module 201, a type classification module 202, and a verification and correction module 203.
[0081] The scene point cloud acquisition module 201 is used to acquire the scene point cloud to be classified in the substation.
[0082] The type classification module 202 is used to sequentially label the scene point cloud to be classified with region type, device type and component type using KP-FCNN to obtain intermediate classification results;
[0083] The verification and correction module 203 is used to perform type label verification and correction on the intermediate classification results according to the preset point cloud classification constraint graph neural network to obtain the scene point cloud segmentation result; wherein, the preset point cloud classification constraint graph neural network is obtained by training a knowledge graph information embedding network combined with a graph neural network; the knowledge graph information embedding network is obtained by learning the knowledge graph triples of the relationship between substation area, equipment and components through the R-GCN model.
[0084] In this embodiment, please refer to Figure 4The diagram shows the internal structure of a type classification module in a substation scene point cloud classification system provided in an embodiment of the present invention. The type classification module specifically includes: a region point cloud classification submodule 301, a device point cloud classification submodule 302, a component point cloud classification submodule 303, and an intermediate classification result determination submodule 304.
[0085] The regional point cloud classification submodule 301 is used to classify the point cloud region type of the scene point cloud to be classified using KP-FCNN, obtain the point cloud region classification result, label the region type of the point cloud data in the scene point cloud to be classified according to the point cloud region classification result, and obtain the point cloud of each region through a density-based clustering algorithm.
[0086] The device point cloud classification submodule 302 is used to classify the point cloud device type of each region's point cloud using KP-FCNN to obtain the point cloud device classification result. Based on the point cloud device classification result, the device type of the point cloud data in the scene point cloud to be classified is labeled, and the point cloud of each device is obtained through a density-based clustering algorithm.
[0087] The component point cloud classification submodule 303 is used to classify the point cloud component types of each device's point cloud using KP-FCNN to obtain point cloud component classification results. Based on the point cloud component classification results, the component types of the point cloud data in the scene point cloud to be classified are labeled, and the point cloud of each component is obtained through a density-based clustering algorithm.
[0088] The intermediate classification result determination submodule 304 is used to take the scene point cloud after labeling the region type, device type and component type as the intermediate classification result.
[0089] In this embodiment, please refer to Figure 5 This is a schematic diagram of the internal structure of a type classification module including a data preprocessing submodule in a substation scene point cloud classification system provided by an embodiment of the present invention. Before the area point cloud classification submodule 301, it also includes a data preprocessing submodule 300, which is used to perform point cloud enhancement and ground point filtering processing on the scene point cloud to be classified.
[0090] In this embodiment, please refer to Figure 6 This is a schematic diagram of the internal structure of a verification and correction module in a substation scene point cloud classification system provided in an embodiment of the present invention. The verification and correction module 203 specifically includes: graph structure conversion submodules 401 and 402.
[0091] The graph structure conversion submodule 401 is used to convert the intermediate classification result into a graph structure classification result; wherein, each node in the graph structure classification result is a point cloud data in the intermediate classification result, each node attribute is the type label of the corresponding point cloud data, and each edge is the hierarchical relationship of the type labels between two point cloud data in the intermediate classification result.
[0092] The node attribute correction submodule 402 is used to take the graph structure classification result as input, and through a preset point cloud classification constraint graph neural network, to verify and correct the attributes of each node in the graph structure to obtain the scene point cloud segmentation result.
[0093] In this embodiment, please refer to Figure 7 This is a schematic diagram of the internal structure of a knowledge graph embedding network in a substation scene point cloud classification system provided in an embodiment of the present invention. The knowledge graph embedding network is obtained by learning the knowledge graph through knowledge triples of the relationship between substation areas, equipment and components using the R-GCN model. Specifically, it includes: a structured data acquisition submodule 501, a text data acquisition submodule 502, a triple determination submodule 503, and an embedding network determination submodule 504.
[0094] The structured data acquisition submodule 501 is used to acquire equipment management information and design information of substation equipment and facilities as structured data.
[0095] The text data acquisition submodule 502 is used to acquire substation equipment management and maintenance information, substation equipment maintenance manual information, and substation inspection work guide information as text data.
[0096] The triplet determination submodule 503 is used to obtain a knowledge triplet of the association relationship between substation area, equipment and components based on at least one of the structured data and text data.
[0097] The embedded network determination submodule 504 is used to obtain the embedded network of knowledge graph information by using the knowledge triplet of the relationship between the substation area, equipment and components through the R-GCN model.
[0098] The system provided by this invention uses a KP-FCNN network to sequentially construct multi-level point cloud segmentation results for substation sites, including regions, equipment, and components. After three levels of classification, the point clouds of the three types of substation scenarios are fused by attribute integration. Subsequently, a graph neural network with embedded knowledge graph network is used to verify and correct the intermediate classification results, further improving the accuracy of point cloud classification. This solves the shortcomings of current point cloud segmentation technology in multi-scale, multi-level substation scenarios, such as poor point cloud segmentation effect and lack of hierarchical classification, and provides data support for subsequent single-unit model construction or other applications.
[0099] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.
Claims
1. A point cloud classification method for a substation scene, characterized in that, include: Obtain the scene point cloud to be classified in the substation; The point cloud of the scene to be classified is labeled with region type, device type and component type in sequence using KP-FCNN to obtain intermediate classification results; Based on a preset point cloud classification constraint graph neural network, the intermediate classification results are subjected to type labeling verification and correction to obtain scene point cloud segmentation results; wherein, the preset point cloud classification constraint graph neural network is obtained by training a knowledge graph information embedding network combined with a graph neural network; the knowledge graph information embedding network is obtained by learning the knowledge graph triples of the relationship between substation areas, equipment and components through the R-GCN model; Specifically, after classifying the point cloud region type of the scene point cloud to be classified using KP-FCNN to obtain the point cloud region classification result, the region type is labeled. After classifying the point cloud device types for each region using KP-FCNN to obtain the point cloud device classification results, the device types are then labeled. After classifying the point cloud components for each device using KP-FCNN and obtaining the component classification results, the component types are labeled.
2. The substation scene point cloud classification method as described in claim 1, characterized in that, The process involves sequentially labeling the scene point cloud to be classified with region type, device type, and component type using KP-FCNN to obtain intermediate classification results, specifically including: Based on the point cloud region classification results, the region type of the point cloud data in the scene point cloud to be classified is labeled, and the point cloud of each region is obtained by using a density-based clustering algorithm. Based on the point cloud device classification results, the device type of the point cloud data in the scene point cloud to be classified is labeled, and the point cloud of each device is obtained by using a density-based clustering algorithm. Based on the point cloud component classification results, the component types of the point cloud data in the scene point cloud to be classified are labeled, and the point cloud of each component is obtained by density-based clustering algorithm. The scene point cloud labeled with region type, device type, and component type is used as the intermediate classification result.
3. The substation scene point cloud classification method as described in claim 2, characterized in that, Before classifying the scene point cloud to be classified using KP-FCNN, the method further includes: performing point cloud enhancement and ground point filtering on the scene point cloud to be classified.
4. The substation scene point cloud classification method as described in claim 1, characterized in that, The step of performing type labeling verification and correction on the intermediate classification results based on a preset point cloud classification constraint graph neural network to obtain scene point cloud segmentation results specifically includes: The intermediate classification result is converted into a graph structure classification result; wherein, each node in the graph structure classification result is a point cloud data in the intermediate classification result, each node attribute is the type label of the corresponding point cloud data, and each edge is the hierarchical relationship of the type labels between two point cloud data in the intermediate classification result. Using the graph structure classification result as input, the attributes of each node in the graph structure are verified and corrected through a preset point cloud classification constraint graph neural network to obtain the scene point cloud segmentation result.
5. The substation scene point cloud classification method as described in claim 1, characterized in that, The knowledge graph information embedding network is obtained by learning the knowledge graph from knowledge triples relating substation areas, equipment, and components using the R-GCN model, specifically including: Obtain equipment management information and design information of substation equipment and facilities as structured data; Obtain substation equipment management and maintenance information, substation equipment maintenance manual information, and substation inspection work guide information as text data; Based on at least one of the structured data and text data, a knowledge triplet of the relationship between substation area, equipment and components is obtained; Using the R-GCN model, a knowledge graph information embedding network is obtained by using knowledge triples of the relationships between the substation area, equipment, and components.
6. A point cloud classification system for substation scenes, characterized in that, include: The scene point cloud acquisition module is used to acquire scene point clouds to be classified in the substation; The type classification module is used to sequentially label the point cloud of the scene to be classified with region type, device type and component type using KP-FCNN to obtain intermediate classification results; The verification and correction module is used to verify and correct the intermediate classification results by type labeling based on a preset point cloud classification constraint graph neural network, so as to obtain the scene point cloud segmentation result; wherein, the preset point cloud classification constraint graph neural network is obtained by training a knowledge graph information embedding network combined with a graph neural network; the knowledge graph information embedding network is obtained by learning the knowledge graph triples of the relationship between substation area, equipment and components through the R-GCN model; The type classification module specifically includes: The regional point cloud classification submodule is used to classify the point cloud region type of the scene point cloud to be classified using KP-FCNN, and then to label the region type after obtaining the point cloud region classification result; The device point cloud classification submodule is used to classify the point cloud device types of each region using KP-FCNN, and then label the device types after obtaining the point cloud device classification results. The component point cloud classification submodule is used to classify the point cloud component types of each device using KP-FCNN, and then label the component types after obtaining the point cloud component classification results.
7. The substation scene point cloud classification system as described in claim 6, characterized in that: The regional point cloud classification submodule is used to label the regional type of point cloud data in the scene point cloud to be classified according to the point cloud regional classification results, and obtain the point cloud of each region through a density-based clustering algorithm. The device point cloud classification submodule is used to label the device type of the point cloud data in the scene point cloud to be classified according to the point cloud device classification results, and obtain the point cloud of each device through a density-based clustering algorithm. The component point cloud classification submodule is used to label the component type of the point cloud data in the scene point cloud to be classified according to the point cloud component classification results, and obtain the point cloud of each component through a density-based clustering algorithm. The intermediate classification result determination submodule is used to take the scene point cloud after labeling the region type, device type, and component type as the intermediate classification result.
8. The substation scene point cloud classification system as described in claim 7, characterized in that, Before the regional point cloud classification submodule, the system further includes a data preprocessing submodule, which is used to perform point cloud enhancement and ground point filtering on the scene point cloud to be classified.
9. The substation scene point cloud classification system as described in claim 6, characterized in that, The verification and correction module specifically includes: The graph structure conversion submodule is used to convert the intermediate classification result into a graph structure classification result; wherein, each node in the graph structure classification result is a point cloud data in the intermediate classification result, each node attribute is the type label of the corresponding point cloud data, and each edge is the hierarchical relationship of the type labels between two point cloud data in the intermediate classification result; The node attribute correction submodule is used to take the graph structure classification result as input, and through a preset point cloud classification constraint graph neural network, to verify and correct the attributes of each node in the graph structure to obtain the scene point cloud segmentation result.
10. The substation scene point cloud classification system as described in claim 6, characterized in that, The knowledge graph information embedding network is obtained by learning the knowledge graph from knowledge triples relating substation areas, equipment, and components using the R-GCN model, specifically including: The structured data acquisition submodule is used to acquire equipment management information and design information of substation equipment and facilities as structured data. The text data acquisition submodule is used to acquire substation equipment management and maintenance information, substation equipment maintenance manual information, and substation inspection work guide information as text data. The triplet determination submodule is used to obtain knowledge triples of the association between substation areas, equipment and components based on at least one of the structured data and text data. The embedded network determination submodule is used to obtain the embedded network of knowledge graph information by using the knowledge triples of the relationship between the substation area, equipment and components through the R-GCN model.
Citation Information
Patent Citations
Intelligent segmentation and identification method and system of transformer substation and storage medium
CN110414351A
Point cloud data splicing method based on graph residual neural network fusion
CN111028151A
