Transformer substation scene three-dimensional semantic segmentation method fusing local geometry and global context

Abstract

The invention discloses a transformer substation scene three-dimensional semantic segmentation method fusing local geometry and global context, and the method comprises the steps: obtaining the three-dimensional point cloud data of a transformer substation scene, and constructing a semantic segmentation data set; a three-dimensional semantic segmentation model is constructed, and fusion features integrating local geometric information, global context information and neighborhood information are extracted through a down-sampling module, a position coding module, a local-global fusion module and a neighborhood propagation unit in sequence; integrating the multi-scale fusion features through an up-sampling propagation module, and outputting a semantic category prediction result by using a semantic segmentation head; designing a loss function, and training the model; and loading the trained three-dimensional semantic segmentation model to realize refined segmentation of the substation scene. According to the method, the understanding capability of the complex three-dimensional structure of the transformer substation is effectively enhanced, the segmentation precision of refined power equipment parts is remarkably improved, and reliable technical support is provided for intelligent operation and maintenance of the transformer substation.

Description

Technical Field

[0001] This invention belongs to the field of substation scene point cloud data processing technology, specifically involving a three-dimensional semantic segmentation method for substation scenes that integrates local geometry and global context. Background Technology

[0002] With the widespread adoption of high-precision sensors such as LiDAR, structured light, and depth cameras, 3D point clouds have become an important data format for characterizing the geometry and semantics of the real world. Semantic segmentation of point clouds can provide point-by-point semantic labels for target recognition, scene understanding, and 3D reconstruction, widely serving scenarios such as autonomous driving, mobile robotics, industrial inspection, augmented / mixed reality, and medical imaging. The core task of 3D semantic segmentation is to learn the mapping from point cloud features and spatial relationships to semantic labels, enabling the model to understand the semantic meaning of each point at a fine-grained level.

[0003] In recent years, deep learning-based point cloud segmentation methods have developed rapidly, mainly including three categories: projection-based (mapping point clouds into 2D images or multiple views), voxel-based (regularized voxel meshes), and point-based (modeling directly on the point set). Among these, point-based methods are further divided into convolutional and Transformer paradigms: the former relies on local aggregation and convolution operators, possessing efficient and robust local feature extraction capabilities; the latter relies on attention mechanisms, excelling at modeling long-range dependencies and global context. Both types of methods have achieved excellent performance on general indoor and outdoor datasets.

[0004] However, in industrial scenarios such as substations, point clouds often exhibit strong structural heterogeneity and complex noise: equipment shapes are diverse and scales vary greatly, supporting components and cables generate numerous slender structures and occlusions, point density is uneven and reflections are missing, making it difficult to simultaneously capture local geometric details and long-range semantic dependencies. Existing methods either overemphasize local geometry and fail to adequately characterize the global context, or suffer from high global modeling costs and insufficient segmentation accuracy for fine-grained industrial components; meanwhile, custom operators still face engineering bottlenecks in terms of data type, dimension alignment, and efficient neighborhood aggregation. Therefore, there is an urgent need for a 3D semantic segmentation technology that balances local geometric sensitivity with global context modeling capabilities and is engineering-applicable, in order to improve the fine-grained segmentation accuracy and inference efficiency of key components in complex substation scenarios. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a 3D semantic segmentation method for substation scenes that integrates local geometry and global context. The aim is to enhance the semantic segmentation model's ability to understand the complex 3D structure of substations, improve the segmentation accuracy of refined power equipment components, and achieve more accurate semantic segmentation of substation scene point cloud data compared to existing technologies.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution: A 3D semantic segmentation method for substation scenes that integrates local geometry and global context includes the following steps: S1. Obtain 3D point cloud data of the substation scene and preprocess the obtained 3D point cloud data; then, according to the type of power equipment, perform category labeling and region division on the preprocessed 3D point cloud data to construct a 3D semantic segmentation dataset of the substation scene including training set and test set. S2. Construct a three-dimensional semantic segmentation model based on a multi-scale recursive architecture backbone network. The three-dimensional semantic segmentation model is divided into multiple different scale stages. Within each scale stage, the model sequentially passes through a downsampling module, a position encoding module, a local-global fusion module, and a neighborhood propagation unit to extract fusion features that integrate local geometric information, global context information, and neighborhood information from the input data. By integrating the fusion features output from multiple scale stages through a multi-level upsampling propagation module, the semantic category prediction result for each point is output using a semantic segmentation head; S3. Design a loss function and use an adaptive learning rate optimizer to train the 3D semantic segmentation model using 3D point cloud data in the training set, and optimize the model parameters using the gradient descent algorithm. S4. Load the trained 3D semantic segmentation model, perform forward inference on the 3D point cloud data in the test set, and output the semantic category label of each point to achieve refined 3D semantic segmentation of the substation scene.

[0007] Furthermore, step S1 specifically includes: S11. Use LiDAR to acquire three-dimensional point cloud data of the substation scene, and then preprocess the three-dimensional point cloud data, including noise reduction, filtering and normalization. S12. Based on the appearance, functional attributes and spatial distribution of the objects, the preprocessed point cloud data is divided into non-power equipment point clouds and power equipment point clouds, and semantic category labels are set for each. The power equipment point clouds are further divided into small equipment point clouds, medium equipment point clouds and large equipment point clouds according to the size of the power equipment and the number of point clouds contained therein. S13. Combine spatial distribution and semantic category to divide the scene into several region blocks, divide the training set and test set according to a preset ratio, and ensure that both the training set and the test set must have point clouds of all categories.

[0008] Furthermore, in step S2, the downsampling module and the position encoding module perform hierarchical downsampling and neighborhood grouping on the input data, specifically as follows: In the downsampling module, the input point set is processed according to the pre-calculated downsampling index. Hierarchical downsampling is performed to obtain the downsampling point set. ;right Each center point In the location encoding module, the center point is constructed using the K-nearest neighbor index. neighborhood For the center point Each neighboring point Calculate the center point With neighboring points relative position vector and neighborhood features splicing forms splicing features ,in This indicates a splicing operation.

[0009] Further, in step S2, the local-global fusion module extracts and fuses local geometric features and global contextual information in parallel from the output of the location encoding module, specifically as follows: The local-global fusion module is centered on the point. initial features and the output of the position encoding module For input, first... The neighborhood prior is obtained by MLP transformation through an embedded network and max pooling along the neighborhood dimension. Neighborhood Prior With initial features The center point is obtained by fusion. Enhanced features The formula is expressed as: ; ; Enhanced features First, the residual is connected through an MLP to obtain the residual enhancement feature. ; then residual enhancement features Input into the local branch and the global branch respectively; In local branches, First, feature transformation is performed through a shared linear layer to obtain the transformed features. Next, perform a grouping operation to obtain the center point. K-nearest neighbor set The neighborhood features after transformation of all corresponding neighborhood points ; for the transformed neighborhood features Perform max pooling operation, then combine with Subtraction is performed, and finally, batch normalization is used to obtain the local geometric features. The formula is expressed as: ; ; ; in, Indicates a linear layer. Indicates batch normalization, Indicates max pooling; In the global branch, residual enhancement features are applied to all points. The characteristic matrix formed First, perform layer normalization to obtain Then, it is input into the Mamba state space module to obtain the global context feature matrix. The formula is expressed as: ; ; For each center point Its global context features That is, a matrix The Line characteristics; Local geometric features With global context features First, the data is spliced ​​along the channels, and then sequentially processed through linear layers and batch normalization to form a fused feature. fusion features Then, with residual enhancement features The output features of the local-fusion module are obtained by adding the residuals. The formula is expressed as: ; ; ; in, This represents a concatenated feature that combines local geometric features with global contextual features. For batch normalization operations, It is a linear layer.

[0010] Furthermore, in step S2, the neighborhood propagation unit performs neighborhood information diffusion and feature update on the output features of the local-fusion module within the local neighborhood, specifically as follows: For the output characteristics of the local-fusion module The neighborhood propagation unit first performs feature transformation on the feature through a shared linear layer to obtain the transformed feature. Next, perform a grouping operation to obtain the center point. K-nearest neighbor set The neighborhood features after transformation of all corresponding neighborhood points Finally, the neighborhood difference is constructed. The formula is expressed as: ; ; Neighborhood difference After nonlinear activation, average pooling is performed along the neighborhood dimension, as expressed by the formula: ; in, It is a non-linear activation function. The neighborhood difference after average pooling; The neighborhood difference after average pooling is then subjected to a series of nonlinear transformations to obtain the feature update value. The formula is expressed as: ; in, For batch normalization, Linear layer; Final feature update With input The residuals are summed to obtain the output of the neighborhood propagation unit. The formula is expressed as: Output of the neighborhood propagation unit That is, the fusion feature output at each scale stage.

[0011] Furthermore, the specific process described in step S2, which involves integrating the fusion features output from multiple scale stages through a multi-level upsampling propagation module and using a semantic segmentation head to output the semantic category prediction result for each point, is as follows: The multi-level upsampling propagation module has one fewer level than the scale stages. Each upsampling propagation module maps the fused features output from the current scale stage back to the previous scale through a pre-saved inverse index. The output of each upsampling propagation module, together with the fused features output from the previous scale stage, serves as the input for the next level. This process is repeated layer by layer until the output features are restored to their original resolution, resulting in multi-scale integrated features. In the semantic segmentation head, the multi-scale integrated features are sequentially subjected to discarding, normalization, and linear classification mapping, ultimately outputting the semantic category prediction result for each point.

[0012] Furthermore, in step S3, the cross-entropy loss function is used, and the specific calculation process is as follows: The predicted output logits and the true class label target of the 3D semantic segmentation model are converted into probability distributions using a softmax function. and Next, find the predicted probability corresponding to the true class and calculate its negative logarithm to obtain the cross-entropy loss, expressed by the formula: ; in, To predict the probability distribution of categories, This represents the true class probability distribution; n represents the number of classes. i Indicates the first i There are several categories.

[0013] Furthermore, the present invention also discloses an electronic device comprising a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to, while running the computer program, execute a 3D semantic segmentation method for substation scenes that integrates local geometry and global context, as described above.

[0014] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute a substation scene three-dimensional semantic segmentation method that integrates local geometry and global context as described above.

[0015] Based on the above technical solution, the present invention has at least the following beneficial effects: This invention addresses local structural details and long-range spatial dependencies. The 3D semantic segmentation model employs a multi-scale recursive architecture as its backbone network, achieving multi-scale feature extraction through hierarchical downsampling and neighborhood grouping. At each scale stage, a local-global fusion module is set up to model neighborhood geometric features and long-range spatial dependencies, while a neighborhood propagation unit is established to achieve cross-neighborhood information diffusion and feature updates. The 3D semantic segmentation model constructed in this invention, through an effective fusion mechanism of local geometry and global context, coupled with dynamic updates of neighborhood information, significantly improves the model's feature extraction capability for equipment at different scales in substation scenarios. It also significantly enhances the semantic segmentation accuracy of complex point cloud data in substation scenarios, solving the technical problem of insufficient segmentation accuracy in existing 3D semantic segmentation methods when dealing with refined components in substation scenarios. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a three-dimensional semantic segmentation method for substation scenes that integrates local geometry and global context, as proposed in this invention. Figure 2 This is a schematic diagram of the three-dimensional semantic segmentation network framework constructed in this invention; Figure 3This is a schematic diagram of the local-global fusion module structure in the three-dimensional semantic segmentation network constructed in this invention; Figure 4 This is a schematic diagram of the neighborhood propagation unit structure in the three-dimensional semantic segmentation network constructed in this invention. Figure 5 This is a schematic diagram of the three-dimensional semantic segmentation result of the point cloud data in the substation dataset of this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-5 The present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] Although the steps in this invention are arranged by reference numerals, this is not intended to limit the order of the steps. Unless the order of the steps is explicitly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" as used herein refers to and covers any and all possible combinations of one or more of the associated listed items.

[0019] like Figure 1 As shown, this invention illustrates a 3D semantic segmentation method for substation scenes that integrates local geometry and global context, specifically including the following steps: S1. Obtain 3D point cloud data of the substation scene and preprocess the obtained 3D point cloud data; then, according to the type of power equipment, perform category labeling and region division on the preprocessed 3D point cloud data to construct a 3D semantic segmentation dataset of the substation scene including training set and test set. In a preferred embodiment, step S1 specifically includes: S11. Use LiDAR to acquire three-dimensional point cloud data of the substation scene, and then preprocess the three-dimensional point cloud data, including noise reduction, filtering and normalization. S12. Based on the appearance, functional attributes, and spatial distribution of the objects, the preprocessed point cloud data is divided into non-power equipment point clouds (in this application, this part of the point cloud data is divided into 7 categories: busbars, power lines, overhead lines, debris, buildings, ground, and walls) and power equipment point clouds, and semantic category labels are set for each. The power equipment point clouds are further divided into small equipment point clouds, medium equipment point clouds, and large equipment point clouds based on the size of the power equipment and the number of point clouds it contains. S13. Combining spatial distribution and semantic categories, the scene is divided into several region blocks, and training and test sets are divided according to a preset ratio, ensuring that both the training and test sets contain point clouds of all categories. In this embodiment, the substation scene is first divided into 6 regions of different sizes. Region 5 is used as the test set, which contains 16 region blocks of different sizes that include all categories; the other regions are used as the training set, containing a total of 29 region blocks of different sizes that include all categories. This division of the test and training sets ensures the comprehensiveness and objectivity of subsequent 3D semantic segmentation training and testing.

[0020] S2. Construct a three-dimensional semantic segmentation model based on a multi-scale recursive architecture backbone network. The three-dimensional semantic segmentation model is divided into multiple different scale stages. Within each scale stage, the model sequentially passes through a downsampling module, a position encoding module, a local-global fusion module, and a neighborhood propagation unit to extract fusion features that integrate local geometric information, global context information, and neighborhood information from the input data. By integrating the fusion features output from multiple scale stages through a multi-level upsampling propagation module, the semantic category prediction result for each point is output using a semantic segmentation head; The three-dimensional semantic model constructed in this application is as follows: Figure 2 As shown, and in a preferred embodiment, the modules in the three-dimensional semantic model are specifically as follows: The downsampling module and the position encoding module perform hierarchical downsampling and neighborhood grouping on the input data, specifically as follows: In the downsampling module, the input point set is processed according to the pre-calculated downsampling index. Hierarchical downsampling is performed to obtain the downsampling point set. The downsampling module expands the network's receptive field, reduces computational cost, and preserves the geometric skeleton. Each center point In the location encoding module, the center point is constructed using the K-nearest neighbor index. neighborhood For the center point Each neighboring point Calculate the center point With neighboring points relative position vector and neighborhood features splicing forms splicing features ,in This indicates a splicing operation.

[0021] In this embodiment, geometric priors are embedded into features using positional encoding with translation-invariant relative coordinates, thereby achieving joint spatial-feature modeling.

[0022] The structure of the local-global fusion module is as follows: Figure 3As shown, this module extracts and fuses local geometric features and global contextual information in parallel from the output of the positional encoding module. This achieves complementary advantages between local details and global context, providing richer feature data for accurate segmentation of complex scenes. Specifically: The local-global fusion module is centered on the point. initial features and the output of the position encoding module For input, first... The neighborhood prior is obtained by MLP transformation through an embedded network and max pooling along the neighborhood dimension. Neighborhood Prior With initial features The center point is obtained by fusion. Enhanced features The formula is expressed as: ; ; Enhanced features First, the residual is connected through an MLP to obtain the residual enhancement feature. ; then residual enhancement features Input into the local branch and the global branch respectively; In local branches, First, feature transformation is performed through a shared linear layer to obtain the transformed features. Next, perform a grouping operation to obtain the center point. K-nearest neighbor set The neighborhood features after transformation of all corresponding neighborhood points ; for the transformed neighborhood features Perform max pooling operation, then combine with Subtraction is performed, and finally, batch normalization is used to obtain the local geometric features. The formula is expressed as: ; ; ; in, Indicates a linear layer. Indicates batch normalization, Indicates max pooling; In the global branch, residual enhancement features are applied to all points. The characteristic matrix formed First, perform layer normalization to obtain Then, it is input into the Mamba state space module to obtain the global context feature matrix. The formula is expressed as: ; ; For each center point Its global context features That is, a matrix The Line characteristics; Local geometric features With global context features First, the data is spliced ​​along the channels, and then sequentially processed through linear layers and batch normalization to form a fused feature. fusion features Then, with residual enhancement features The output features of the local-fusion module are obtained by adding the residuals. The formula is expressed as: ; ; ; in, This represents a concatenated feature that combines local geometric features with global contextual features. For batch normalization operations, It is a linear layer.

[0023] The structure of the neighborhood propagation module is as follows: Figure 4 As shown, the neighborhood propagation unit in this module diffuses neighborhood information and updates features of the output features of the local-fusion module within the local neighborhood, thereby smoothing features and optimizing the fineness of segmentation boundaries. Specifically: For the output characteristics of the local-fusion module The neighborhood propagation unit first performs feature transformation on the feature through a shared linear layer to obtain the transformed feature. Next, perform a grouping operation to obtain the center point. K-nearest neighbor set The neighborhood features after transformation of all corresponding neighborhood points Finally, the neighborhood difference is constructed. The formula is expressed as: ; ; Neighborhood difference After nonlinear activation, average pooling is performed along the neighborhood dimension, as expressed by the formula: ; in, It is a non-linear activation function. The neighborhood difference after average pooling; The neighborhood difference after average pooling is then subjected to a series of nonlinear transformations to obtain the feature update value. The formula is expressed as: ; in, For batch normalization, Linear layer; Final feature update With input The residuals are summed to obtain the output of the neighborhood propagation unit. The formula is expressed as: Output of the neighborhood propagation unit That is, the fusion feature output at each scale stage.

[0024] Finally, this application uses a semantic segmentation head to output the semantic category prediction result for each point. The specific process is as follows: The multi-level upsampling propagation module has one fewer level than the scale stages. Each upsampling propagation module maps the fused features output from the current scale stage back to the previous scale through a pre-saved inverse index. The output of each upsampling propagation module, together with the fused features output from the previous scale stage, serves as the input for the next level. This process is repeated layer by layer until the output features are restored to their original resolution, resulting in multi-scale integrated features. In the semantic segmentation head, the multi-scale integrated features are sequentially subjected to discarding, normalization, and linear classification mapping, ultimately outputting the semantic category prediction result for each point.

[0025] The 3D semantic segmentation model designed in this embodiment combines relative position encoding, KNN edge pooling, and collaborative modeling of Mamba and neighborhood propagation units. While ensuring full integration of local geometry and global context, it significantly improves the segmentation accuracy of complex substation scenarios.

[0026] S3. Design a loss function and use an adaptive learning rate optimizer to train the 3D semantic segmentation model using 3D point cloud data in the training set, and optimize the model parameters using the gradient descent algorithm. In a preferred embodiment, the loss function in step S3 is the cross-entropy loss function, and the specific calculation process is as follows: The predicted output logits and the true class label target of the 3D semantic segmentation model are converted into probability distributions using a softmax function. and Next, find the predicted probability corresponding to the true class and calculate its negative logarithm to obtain the cross-entropy loss, expressed by the formula: ; in, To predict the probability distribution of categories, This represents the true class probability distribution; n represents the number of classes. i Indicates the first i There are several categories.

[0027] In this embodiment, the optimizer is AdamW, the learning rate is set to 0.004, and the weight decay is... During training, parameters are updated iteratively until the loss function of the semantic segmentation model reaches its maximum value. The optimization continues until convergence is achieved. Experiments have verified that setting the optimizer parameters in this way enables the model to achieve the best predictive performance.

[0028] S4. Load the trained 3D semantic segmentation model, perform forward inference on the 3D point cloud data in the test set, and output the semantic category label of each point to achieve refined 3D semantic segmentation of the substation scene.

[0029] Furthermore, the present invention also discloses an electronic device comprising a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to, while running the computer program, execute a 3D semantic segmentation method for substation scenes that integrates local geometry and global context, as described above.

[0030] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute a substation scene three-dimensional semantic segmentation method that integrates local geometry and global context as described above.

[0031] Based on the method and the constructed three-dimensional semantic segmentation model proposed in this invention, the semantic segmentation results for the substation scene are as follows: Figure 5 The figure shows the prediction results processed by the 3D semantic segmentation method proposed in this embodiment. As can be seen from the figure, this invention effectively enhances the model's understanding of the complex 3D structure of substations by integrating local geometric perception and global context modeling, providing reliable technical support for intelligent operation and maintenance of substations. It solves the technical problem of insufficient accuracy of existing 3D semantic segmentation methods when facing complex point cloud data in substation scenarios, achieving accurate semantic segmentation of substation point cloud data.

[0032] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0033] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0034] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A 3D semantic segmentation method for substation scenes that integrates local geometry and global context, characterized in that, Specifically, the following steps are included: S1. Obtain 3D point cloud data of the substation scene and preprocess the obtained 3D point cloud data; then, according to the type of power equipment, perform category labeling and region division on the preprocessed 3D point cloud data to construct a 3D semantic segmentation dataset of the substation scene including training set and test set. S2. Construct a three-dimensional semantic segmentation model based on a multi-scale recursive architecture backbone network. The three-dimensional semantic segmentation model is divided into multiple different scale stages. Within each scale stage, the model sequentially passes through a downsampling module, a position encoding module, a local-global fusion module, and a neighborhood propagation unit to extract fusion features that integrate local geometric information, global context information, and neighborhood information from the input data. By integrating the fusion features output from multiple scale stages through a multi-level upsampling propagation module, the semantic category prediction result for each point is output using a semantic segmentation head; S3. Design a loss function and use an adaptive learning rate optimizer to train the 3D semantic segmentation model using 3D point cloud data in the training set, and optimize the model parameters using the gradient descent algorithm. S4. Load the trained 3D semantic segmentation model, perform forward inference on the 3D point cloud data in the test set, and output the semantic category label of each point to achieve refined 3D semantic segmentation of the substation scene.

2. The three-dimensional semantic segmentation method for substation scenes integrating local geometry and global context as described in claim 1, characterized in that, Step S1 specifically includes: S11. Use LiDAR to acquire three-dimensional point cloud data of the substation scene, and then preprocess the three-dimensional point cloud data, including noise reduction, filtering and normalization. S12. Based on the appearance, functional attributes and spatial distribution of the objects, the preprocessed point cloud data is divided into non-power equipment point clouds and power equipment point clouds, and semantic category labels are set for each. The power equipment point clouds are further divided into small equipment point clouds, medium equipment point clouds and large equipment point clouds according to the size of the power equipment and the number of point clouds contained therein. S13. Combine spatial distribution and semantic category to divide the scene into several region blocks, divide the training set and test set according to a preset ratio, and ensure that both the training set and the test set must have point clouds of all categories.

3. The three-dimensional semantic segmentation method for substation scenes integrating local geometry and global context as described in claim 1, characterized in that, In step S2, the downsampling module and the position encoding module perform hierarchical downsampling and neighborhood grouping on the input data, specifically as follows: In the downsampling module, the input point set is processed according to the pre-calculated downsampling index. Hierarchical downsampling is performed to obtain the downsampling point set. ;right Each center point In the location encoding module, the center point is constructed using the K-nearest neighbor index. neighborhood For the center point Each neighboring point Calculate the center point With neighboring points relative position vector and neighborhood features splicing forms splicing features ,in This indicates a splicing operation.

4. The three-dimensional semantic segmentation method for substation scenes integrating local geometry and global context according to claim 3, characterized in that, In step S2, the local-global fusion module extracts and fuses local geometric features and global contextual information in parallel from the output of the location encoding module, specifically as follows: The local-global fusion module uses the center point initial features and the output of the position encoding module For input, first... The neighborhood prior is obtained by MLP transformation through an embedded network and max pooling along the neighborhood dimension. Neighborhood Prior With initial features The center point is obtained by fusion. Enhanced features The formula is expressed as: ； ； Enhanced features First, the residual is connected through an MLP to obtain the residual enhancement feature. ; Residual enhancement features Input into the local branch and the global branch respectively; In local branches, First, feature transformation is performed through a shared linear layer to obtain the transformed features. Next, perform a grouping operation to obtain the center point. K-nearest neighbor set The neighborhood features after transformation of all corresponding neighborhood points ; for the transformed neighborhood features Perform max pooling operation, then combine with Subtraction is performed, and finally, batch normalization is used to obtain the local geometric features. The formula is expressed as: ； ； ； in, Indicates a linear layer. Indicates batch normalization, Indicates max pooling; In the global branch, residual enhancement features are applied to all points. The characteristic matrix formed First, perform layer normalization to obtain Then, it is input into the Mamba state space module to obtain the global context feature matrix. The formula is expressed as: ； ； For each center point Its global context features That is, a matrix The Line characteristics; Local geometric features With global context features First, the data is spliced ​​along the channels, and then sequentially processed through linear layers and batch normalization to form a fused feature. fusion features Then, with residual enhancement features The output features of the local-fusion module are obtained by adding the residuals. The formula is expressed as: ； ； ； in, This represents a concatenated feature that combines local geometric features with global contextual features. For batch normalization operations, It is a linear layer.

5. The three-dimensional semantic segmentation method for substation scenes integrating local geometry and global context according to claim 3, characterized in that, In step S2, the neighborhood propagation unit performs neighborhood information diffusion and feature update on the output features of the local-fusion module within the local neighborhood, specifically as follows: For the output characteristics of the local-fusion module The neighborhood propagation unit first performs feature transformation on the feature through a shared linear layer to obtain the transformed feature. Next, perform a grouping operation to obtain the center point. K-nearest neighbor set The neighborhood features after transformation of all corresponding neighborhood points Finally, the neighborhood difference is constructed. The formula is expressed as: ； ； Neighborhood difference After nonlinear activation, average pooling is performed along the neighborhood dimension, as expressed by the formula: ； in, It is a non-linear activation function. The neighborhood difference after average pooling; The neighborhood difference after average pooling is then subjected to a series of nonlinear transformations to obtain the feature update value. The formula is expressed as: ； in, For batch normalization, Linear layer; Final feature update With input The residuals are summed to obtain the output of the neighborhood propagation unit. The formula is expressed as: Output of the neighborhood propagation unit That is, the fusion feature output at each scale stage.

6. The three-dimensional semantic segmentation method for substation scenes integrating local geometry and global context according to claim 1, characterized in that, The specific process described in step S2, which involves integrating the fusion features output from multiple scale stages through a multi-level upsampling propagation module and using a semantic segmentation head to output the semantic category prediction result for each point, is as follows: The multi-level upsampling propagation module has one fewer level than the scale stages. Each upsampling propagation module maps the fused features output from the current scale stage back to the previous scale through a pre-saved inverse index. The output of each upsampling propagation module, together with the fused features output from the previous scale stage, serves as the input for the next level. This process is repeated layer by layer until the output features are restored to their original resolution, resulting in multi-scale integrated features. In the semantic segmentation head, the multi-scale integrated features are sequentially subjected to discarding, normalization, and linear classification mapping, ultimately outputting the semantic category prediction result for each point.

7. A three-dimensional semantic segmentation method for substation scenes integrating local geometry and global context as described in claim 1, characterized in that, In step S3, the cross-entropy loss function is used, and the specific calculation process is as follows: The predicted output logits and the true class label target of the 3D semantic segmentation model are converted into probability distributions using a softmax function. and Next, find the predicted probability corresponding to the true class and calculate its negative logarithm to obtain the cross-entropy loss, expressed by the formula: ； in, To predict the probability distribution of categories, This represents the true class probability distribution; n represents the number of classes. i Indicates the first i There are several categories.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to, while running the computer program, execute a three-dimensional semantic segmentation method for substation scenes that integrates local geometry and global context as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute a method for three-dimensional semantic segmentation of a substation scene that integrates local geometry and global context, as described in any one of claims 1-7.

Images