A power grid construction risk hidden danger identification method based on multi-source spatial data

Abstract

This invention discloses a method for identifying power grid construction risks and hazards based on multi-source spatial data, comprising: Step 1: collecting multi-source spatial data of the power grid construction area; Step 2: performing standardization processing; Step 3: performing multimodal spatial feature modeling and semantic enhancement fusion through an improved SegFormer network to generate a fused feature tensor and a hazard area mask map; Step 4: constructing a risk level prediction model to determine risk levels and generate a risk level label map; Step 5: performing spatial overlay analysis to extract the risk level distribution and hazard area segments corresponding to each tower location, generating a tower location risk assessment table; Step 6: performing visualization analysis and displaying the results on a user interaction terminal; Step 7: incrementally updating the improved SegFormer network and the risk level prediction model. This invention improves the accuracy and intelligence level of power grid construction risk identification based on multi-source spatial data.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power systems and intelligent analysis of industrial spatial data, and in particular to a method for identifying potential risks and hazards in power grid construction based on multi-source spatial data. Background Technology

[0002] With the rapid advancement of power transmission and transformation projects in complex terrain areas, safety hazards induced by terrain disturbance, vegetation destruction, and geological disasters are frequently occurring during power grid construction, placing higher demands on site selection and risk early warning. Existing power grid site selection support systems largely rely on single remote sensing image analysis or expert judgment for hazard identification, but in practical applications, the following problems are commonly encountered: The lack of unified coordinate systems and large resolution differences among multi-source spatial data leads to high spatial registration errors, making it difficult to achieve high-precision layer overlay and consistency analysis. The semantic information fusion of remote sensing images, geological layers, and environmental factors is insufficient, and traditional feature extraction methods are unable to simultaneously characterize the multidimensional features of terrain anomalies, vegetation disturbance, and high-risk geological areas, resulting in serious problems of missed detection or misjudgment of potential hazard areas. Risk level classification methods mostly use fixed thresholds or manual rules, which are difficult to adapt to the power grid construction scenarios with strong spatial heterogeneity and complex geographical environments. Risk analysis results are often disconnected from the spatial distribution of tower sites and lack a direct correspondence with the actual construction site layout, making it difficult to provide accurate guidance for engineering planning.

[0003] Therefore, how to provide a method for identifying power grid construction risks and hazards based on multi-source spatial data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a method for identifying potential risks in power grid construction based on multi-source spatial data. This invention fully integrates remote sensing image data, digital elevation model data, tower location spatial vector data, geological disaster risk factor data, and normalized vegetation index data to construct an improved SegFormer network and risk level prediction model oriented towards spatial feature modeling. It describes in detail the process of identifying potential hazard areas, determining risk levels, assessing tower location risks, and spatial visualization in power grid construction areas. It has the advantages of high spatial perception accuracy, strong robustness in hazard identification, and intuitive risk assessment results.

[0005] A method for identifying potential risks in power grid construction based on multi-source spatial data according to an embodiment of the present invention includes the following steps: Step 1: Collect multi-source spatial data of the power grid construction area, and perform coordinate unification and spatial registration on the spatial data; Step 2: Standardize the multi-source spatial data to generate a fused layer tensor; Step 3: Input the fused layer tensor into the improved SegFormer network to perform multimodal spatial feature modeling and semantic enhancement fusion. The improved SegFormer network introduces modal grouping, elevation prior and alternating attention mechanism to generate fused feature tensor, and generates a hazard area mask map through multi-scale modeling and threshold judgment. Step 4: Construct a risk level prediction model, and based on the fusion feature tensor and the hidden danger area mask map, determine the risk level status of each spatial location in the power grid construction area and generate a risk level label map; Step 5: Perform spatial overlay analysis on the risk level label map, the hidden danger area mask map, and the tower site spatial vector map layer to extract the risk level distribution and hidden danger area segments corresponding to each tower site and generate a tower site risk assessment table; Step Six: Visualize and analyze the tower site risk assessment form, risk level label map, and hidden danger area mask map, and display them on the user interaction terminal; Step 7: Collect user interaction feedback data and incrementally update the improved SegFormer network and risk level prediction model.

[0006] Optionally, step one specifically includes: The multi-source spatial data includes remote sensing image data, digital elevation model data, tower site spatial vector data, geological hazard risk factor data, and normalized vegetation index data; The coordinate system of the multi-source spatial data is unified by the projection transformation method based on EPSG encoding, the spatial reference is unified by the affine transformation method, and the layer space clipping operation is performed by the vector polygon mask to generate remote sensing image layer, digital elevation model layer, tower site spatial vector layer, geological hazard layer and vegetation index layer.

[0007] Optionally, step two specifically includes: The remote sensing image layer is normalized by channel linear normalization of the red, green and blue band pixel values, and the normalized remote sensing image data is spatially resampled to unify the resolution and generate a remote sensing image raster layer. The slope factor layer is calculated based on the digital elevation model layer. Specifically, the slope value is calculated at each pixel location using a third-order finite difference algorithm. The slope factor layer is then normalized using a channel linear normalization method to generate a slope raster layer. Perform a vector-to-raster conversion operation on the spatial vector layer of tower locations. Specifically, use the nearest neighbor mapping method to map each tower location to the corresponding raster cell to generate a tower location raster layer. Spatially resample the geological hazard layer and the vegetation index layer, unify the resolution, and perform scale alignment and map boundary consistency processing to generate a geological hazard raster layer and a vegetation index raster layer. The remote sensing image raster layer, slope raster layer, tower location raster layer, geological disaster raster layer, and vegetation index raster layer are stitched together according to the channel dimension to form a fused layer tensor. The dimensions of the fused layer tensor include image height, image height and width, and number of channels.

[0008] Optionally, step three specifically includes: The fused layer tensor is divided into several modal channel groups according to channels. The modal channel groups include remote sensing image modal channel group, elevation structure modal channel group, risk labeling modal channel group, vegetation disturbance modal channel group, and tower location modal channel group. Semantic modeling is performed on each modal channel group through an independent Transformer encoding path to generate a modal semantic feature map. Each Transformer encoding path includes a convolutional downsampling layer, a patch construction module, and a MixTransformer encoding module, specifically: Digital elevation model data is introduced and normalized using the channel linear normalization method to generate an elevation raster layer. The elevation raster layer and modal channel group are size-compressed by a 2D convolution with a kernel of 5×5, a stride of 2, and a symmetric zero-filling method. The compression tensor of elevation raster and modal channel group are then generated by nonlinear mapping with batch normalization and GELU activation function. The elevation raster compression tensor and the modal compression tensor are divided into several patch blocks according to the same window size. Each patch block is flattened into a one-dimensional vector, linearly mapped by MLP, and projected onto a preset dimension to generate elevation embedding vector and patch feature vector. Learnable weight coefficients are introduced, and the elevation embedding vector is multiplied element-wise with the learnable weight coefficients and then vector-added with the Patch feature vector to generate a hybrid Patch feature vector. The mixed patch feature vectors corresponding to the current modality channel group are combined into a mixed patch feature vector sequence and input into the Mix Transformer encoding module to generate a modality semantic feature map. The modal semantic feature maps generated by all Transformer encoding paths are batch normalized and then channel-wise concatenated and fused to generate a fused feature tensor. The fused feature tensor is input in parallel to three convolutional branches, and 1×1 convolution, 3×3 convolution and 5×5 convolution operations are performed respectively to generate local perceptual feature maps at three scales. The local perceptual feature maps at three scales are fused element-wise by addition to generate an enhanced convolutional feature map. The enhanced convolutional feature map is processed by 1×1 convolution to perform channel compression mapping, and nonlinear mapping is performed through two MLP layers to generate semantic fusion feature maps. The semantic fusion feature map is upsampled to be consistent with the spatial size of the original image, and a probability map of the hidden danger area is generated by the Sigmoid activation function. Set a hazard threshold, and judge the hazard for each pixel in the hazard area probability map. If the probability value corresponding to each pixel is less than the hazard threshold, set the pixel value to 0; otherwise, set the pixel value to 1. Then, use the hazard area probability map after hazard judgment as the hazard area mask map.

[0009] Optionally, the Mix Transformer encoding module includes four Transformer encoding layers, performing alternating attention operations on the mixed patch feature vector sequence, specifically including: The first and third Transformer encoders perform multi-head self-attention operations within a local window on the input feature vector sequence, while the second and fourth Transformer encoders perform pooling-enhanced attention operations on the input feature vector sequence. The pooling-enhanced attention operation specifically involves: The input feature vector sequence is processed by two-dimensional average pooling and two-dimensional max pooling to obtain the average pooling feature vector sequence and the max pooling feature vector sequence, respectively. The average pooling feature vector sequence and the max pooling feature vector sequence are concatenated along the channel dimension, and a channel transformation is performed through a one-dimensional convolution layer to generate a pooling fusion vector sequence. The pooling-sensing fusion vector sequence is fused with the input feature vector sequence through element-wise addition to form an enhanced feature vector sequence. The enhanced feature vector sequence is subjected to self-attention operation to generate the output feature vector sequence of the current layer's Transformer encoder; The input to the first layer Transformer encoder is the sequence of mixed patch feature vectors, and the input to the (n+1)th layer Transformer encoder is the sequence of output feature vectors of the nth layer Transformer encoder, where n takes the values ​​1, 2, and 3. The output feature vector sequence of the 4th layer Transformer encoder is used as the modality coding feature vector sequence, and the modality coding feature vector sequence is reshaped into a two-dimensional spatial structure to generate a modality semantic feature map.

[0010] Optionally, step four specifically includes: The fused feature tensor is upsampled to the same spatial size as the original image through bilinear interpolation to generate a spatially aligned feature tensor. Copy the mask guide map along the channel dimension to form a guide mask tensor with the same channel dimension as the spatially aligned feature tensor. The spatial alignment feature tensor and the guiding mask tensor are multiplied element-wise at the channel level to generate the guiding fusion feature tensor. The guided fusion feature tensor is transformed by a 3×3 convolution and a ReLU activation function to generate an intermediate risk feature tensor; the intermediate risk feature tensor is channel compressed and mapped by a 1×1 convolution and then transformed into a risk level prediction vector for each pixel position by a Softmax classifier. The risk level prediction vector is a three-dimensional vector, which corresponds to the prediction probability values ​​of the current pixel position belonging to the low risk level, the medium risk level, and the high risk level, respectively. The category index corresponding to the maximum predicted probability value in the risk level prediction vector of each pixel location is used as the risk label, and a risk level label map is constructed; the risk labels are 0: low risk, 1: medium risk and 2: high risk.

[0011] Optionally, step five specifically includes: Based on the spatial coordinates of each tower site in the tower site spatial vector layer, a spatial window with a set side length is constructed in the risk level label map with the spatial coordinates as the center. Extract the risk labels corresponding to all pixels within the spatial window, and count the number of pixels with risk labels of 0, 1, and 2 respectively, and record them as the number of low-risk pixels, medium-risk pixels, and high-risk pixels respectively. Calculate the proportions of low-risk, medium-risk, and high-risk pixels to the total number of pixels in the spatial window, respectively, to obtain the low-risk proportion, medium-risk proportion, and high-risk proportion, and then form the risk level distribution probability vector of the current tower site in order. Obtain the area that intersects with the spatial window in the hazard area mask image, and take it as the hazard area segment of the current tower site. Count the number of pixels with a mask value of 1 in the hazard area segment as the number of hazard area pixels of the current tower site. The spatial coordinates of the current tower site, the probability vector of risk level distribution, and the number of pixels in the hidden danger area are sequentially used to construct the tower site risk record vector. Organize the risk record vectors of all tower sites into a tower site risk assessment table by row.

[0012] Optionally, step six specifically includes: Project the tower site risk assessment table onto geospatial coordinates to generate a tower site risk level distribution layer; The mask image of the potential hazard area is overlaid on the remote sensing image layer to form a spatial mask layer of the potential hazard area. The risk level label map is subjected to pseudo-color mapping to generate a risk level distribution layer, wherein 0 is mapped to green, 1 to yellow, and 2 to red in the pseudo-color mapping. The tower location risk visualization layer, the hidden danger area spatial masking layer, and the risk level distribution layer are aligned and overlaid based on spatial coordinates to generate a spatial risk visualization layer set, which is then output to the user interaction terminal for display.

[0013] Optionally, the user interaction feedback data includes risk level label correction information, hazard area boundary adjustment information, and supplementary hazard area annotations; By fusing user interaction feedback data with multi-source spatial data, an interaction feedback training sample set is constructed; Incremental updates are performed on the improved SegFormer network and the risk level prediction model based on the interactive feedback training sample set.

[0014] The beneficial effects of this invention are: This invention proposes a method for identifying power grid construction risks and hazards based on multi-source spatial data, which effectively addresses the limitations of existing methods in expressing the spatial distribution of hazards, refining risk level discrimination, and visualizing results. First, remote sensing image data, digital elevation model data, tower location spatial vector data, geological hazard risk factor data, and normalized vegetation index data of the power grid construction area are collected. Coordinate unification, spatial registration, and preprocessing operations are then performed to generate a structurally consistent fusion layer tensor, providing a data foundation for feature modeling and risk identification.

[0015] Secondly, an improved SegFormer network is used to perform multimodal spatial feature modeling and semantic enhancement fusion on the fused layer tensor. The fused layer tensor is modally grouped, and cross-modal spatial terrain features, vegetation disturbance features, and geological disaster risk-related features are extracted through independent Transformer encoding paths, outputting modal semantic feature maps. The Transformer encoding path introduces digital elevation model data as a structural prior, guiding the encoder to focus on areas with abrupt terrain changes and high elevation differences, enhancing the perception ability of geological disaster-prone areas. Alternating attention is introduced in the Mix Transformer encoding module, enabling the network to capture potential intermodal cooperative relationships and regional spatial dependencies, effectively improving the improved SegFormer network's ability to distinguish weakly textured potential hazard areas and blurred boundary areas.

[0016] Furthermore, the risk level prediction model combines a fusion feature tensor with a hazard area mask map, introducing a mask-guided mechanism to focus on areas with significant risks. This guides the network to generate spatially aligned intermediate risk feature tensors, and a Softmax classifier is used to determine the risk level status of each spatial location within the power grid construction area, generating a high-resolution and highly discriminative risk level label map. Further, by combining a tower site spatial vector layer and performing spatial overlay analysis, the model accurately extracts the risk level distribution and hazard area range around each tower site, forming a structured tower site risk assessment table. This provides a quantitative risk assessment basis for power grid construction and maintenance. By performing a visualization mapping on the risk level label map, hazard area mask map, and tower site risk assessment table, a spatial risk visualization layer set is formed, enabling an intuitive representation of the spatial distribution of construction risks.

[0017] In summary, this invention not only achieves refined identification and classification of spatial risks and hazards in power grid construction areas, but also provides risk assessment results for engineering practice through spatial overlay analysis and visualization. It has the advantages of accurate data fusion, effective feature extraction, refined risk identification, intuitive result expression, and sustainable model optimization. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a method for identifying potential risks in power grid construction based on multi-source spatial data, as proposed in this invention. Figure 2 This is a flowchart of the improved SegFormer network structure in a power grid construction risk and hazard identification method based on multi-source spatial data proposed in this invention; Figure 3 This is a flowchart of risk level prediction and tower location assessment in a power grid construction risk and hazard identification method based on multi-source spatial data proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A method for identifying potential risks in power grid construction based on multi-source spatial data includes the following steps: Step 1: Collect multi-source spatial data of the power grid construction area, and perform coordinate unification and spatial registration on the spatial data; Step 2: Standardize the multi-source spatial data to generate a fused layer tensor; Step 3: Input the fused layer tensor into the improved SegFormer network to perform multimodal spatial feature modeling and semantic enhancement fusion. The improved SegFormer network introduces modal grouping, elevation prior and alternating attention mechanism to generate fused feature tensor, and generates a hazard area mask map through multi-scale modeling and threshold judgment. Step 4: Construct a risk level prediction model, and based on the fusion feature tensor and the hidden danger area mask map, determine the risk level status of each spatial location in the power grid construction area and generate a risk level label map; Step 5: Perform spatial overlay analysis on the risk level label map, the hidden danger area mask map, and the tower site spatial vector map layer to extract the risk level distribution and hidden danger area segments corresponding to each tower site and generate a tower site risk assessment table; Step Six: Visualize and analyze the tower site risk assessment form, risk level label map, and hidden danger area mask map, and display them on the user interaction terminal; Step 7: Collect user interaction feedback data and incrementally update the improved SegFormer network and risk level prediction model.

[0021] In this embodiment, step one specifically includes: The multi-source spatial data includes remote sensing image data, digital elevation model data, tower site spatial vector data, geological hazard risk factor data, and normalized vegetation index data; The digital elevation model data is used to calculate the corresponding terrain elevation values ​​and spatial terrain grids; The spatial vector data of the tower location includes the tower location number, the coordinates of the tower base center, and the tower type. The normalized vegetation index data is used as a vegetation cover factor layer. Multi-source spatial data are subjected to coordinate system unification using an EPSG-based projection transformation method, transforming all spatial layers to the same geographic projection coordinate system; spatial benchmark unification is achieved through affine transformation, correcting geographic location offset errors between layers; and layer spatial clipping is performed using vector polygon masks to retain spatial data within the construction area, generating remote sensing image layers, digital elevation model layers, tower site spatial vector layers, geological hazard layers, and vegetation index layers.

[0022] In this embodiment, step two specifically includes: The remote sensing image layer is normalized by channel linear normalization of the red, green and blue band pixel values, and the normalized remote sensing image data is spatially resampled to unify the resolution and generate a remote sensing image raster layer. The slope factor layer is calculated based on the digital elevation model layer. Specifically, the slope value is calculated at each pixel location using a third-order finite difference algorithm. The slope factor layer is then normalized using a channel linear normalization method to generate a slope raster layer. Perform a vector-to-raster conversion operation on the spatial vector layer of tower locations. Specifically, use the nearest neighbor mapping method to map each tower location to the corresponding raster cell to generate a tower location raster layer. Spatially resample the geological hazard layer and the vegetation index layer, unify the resolution, and perform scale alignment and map boundary consistency processing to generate a geological hazard raster layer and a vegetation index raster layer. The remote sensing image raster layer, slope raster layer, tower location raster layer, geological disaster raster layer, and vegetation index raster layer are stitched together according to the channel dimension to form a fused layer tensor. The dimensions of the fused layer tensor include image height, image height and width, and number of channels.

[0023] In this embodiment, step three specifically includes: The fused layer tensor is divided into several modal channel groups according to channels. The modal channel groups include remote sensing image modal channel group, elevation structure modal channel group, risk labeling modal channel group, vegetation disturbance modal channel group, and tower location modal channel group. In this invention, the fused layer tensor is divided into several modal channel groups according to channels, and semantic modeling is performed on each modal channel group through an independent Transformer encoding path. This helps to fully explore the heterogeneous features of different data modalities in terms of spatial distribution, texture structure, and semantic information. Each modal encoding path can extract differentiated features for information such as remote sensing images, elevation structures, vegetation disturbance, and tower locations, avoiding interference and information dilution between multimodal features, thereby improving the ability of the improved SegFormer network to identify hidden danger areas and the accuracy of risk level discrimination.

[0024] Semantic modeling is performed on each modal channel group through an independent Transformer encoding path to generate a modal semantic feature map. Each Transformer encoding path includes a convolutional downsampling layer, a patch construction module, and a MixTransformer encoding module, specifically: Digital elevation model data is introduced and normalized using the channel linear normalization method to generate an elevation raster layer. The elevation raster layer and modal channel group are size-compressed by a 2D convolution with a kernel of 5×5, a stride of 2, and a symmetric zero-filling method. The compression tensor of elevation raster and modal channel group are then generated by nonlinear mapping with batch normalization and GELU activation function. The elevation raster compression tensor and the modal compression tensor are divided into several patch blocks according to the same window size. Each patch block is flattened into a one-dimensional vector, linearly mapped by MLP, and projected onto a preset dimension to generate elevation embedding vector and patch feature vector. Learnable weight coefficients are introduced, and the elevation embedding vector is multiplied element-wise with the learnable weight coefficients and then vector-added with the Patch feature vector to generate a hybrid Patch feature vector. In this invention, digital elevation model data is introduced as an independent channel to participate in modal coding processing. This can effectively capture the terrain undulations, slope changes and structural elevation differences in the power grid construction area, and fuse them with the semantic features of each modality. This can enhance the spatial identification ability of the improved SegFormer network for geological disaster hazards such as landslides and ground fissures, and improve the three-dimensional perception ability of hazard detection and the spatial accuracy of risk level judgment.

[0025] The mixed patch feature vectors corresponding to the current modality channel group are combined into a mixed patch feature vector sequence and input into the Mix Transformer encoding module to generate a modality semantic feature map. The modal semantic feature maps generated by all Transformer encoding paths are batch normalized and then channel-wise concatenated and fused to generate a fused feature tensor. The fused feature tensor is input in parallel to three convolutional branches, and 1×1 convolution, 3×3 convolution and 5×5 convolution operations are performed respectively. The convolution stride is 1 and the padding method is symmetric zero padding, generating local perceptual feature maps at three scales. The local perceptual feature maps at three scales are fused element-wise by addition to generate an enhanced convolutional feature map. The enhanced convolutional feature map is subjected to channel compression mapping using 1×1 convolution, and nonlinear mapping is performed through two layers of MLP to generate a semantic fusion feature map. The activation function in the MLP is GELU. The semantic fusion feature map is upsampled to be consistent with the spatial size of the original image, and a probability map of the hidden danger area is generated by the Sigmoid activation function. Set a hazard threshold, and judge the hazard for each pixel in the hazard area probability map. If the probability value corresponding to each pixel is less than the hazard threshold, set the pixel value to 0; otherwise, set the pixel value to 1. Then, use the hazard area probability map after hazard judgment as the hazard area mask map.

[0026] In this embodiment, the Mix Transformer encoding module includes four Transformer encoding layers, which perform alternating attention operations on the mixed Patch feature vector sequence, specifically including: The first and third Transformer encoders perform multi-head self-attention operations within a local window on the input feature vector sequence, while the second and fourth Transformer encoders perform pooling-enhanced attention operations on the input feature vector sequence. The pooling-enhanced attention operation specifically involves: The input feature vector sequence is processed by two-dimensional average pooling and two-dimensional max pooling to obtain the average pooling feature vector sequence and the max pooling feature vector sequence, respectively. The average pooling feature vector sequence and the max pooling feature vector sequence are concatenated along the channel dimension, and a channel transformation is performed through a one-dimensional convolution layer to generate a pooling fusion vector sequence. The pooling-sensing fusion vector sequence is fused with the input feature vector sequence through element-wise addition to form an enhanced feature vector sequence. The enhanced feature vector sequence is subjected to self-attention operation to generate the output feature vector sequence of the current layer's Transformer encoder; In this invention, an alternating attention mechanism is employed, alternately embedding multi-head self-attention and pooling-enhanced attention within a local window into the Mix Transformer encoding module. This effectively combines local spatial detail perception with global semantic aggregation capabilities. Local self-attention can finely model structural edges and texture variations within the same modality, while pooling-enhanced attention strengthens the response of key regions by introducing global statistical features. This helps improve the improved SegFormer network's ability to represent spatial hazard features at multiple scales and to make regional comparison judgments, thereby enhancing the accuracy and robustness of hazard region identification.

[0027] The input to the first layer Transformer encoder is the sequence of mixed patch feature vectors, and the input to the (n+1)th layer Transformer encoder is the sequence of output feature vectors of the nth layer Transformer encoder, where n takes the values ​​1, 2, and 3. The output feature vector sequence of the 4th layer Transformer encoder is used as the modality coding feature vector sequence, and the modality coding feature vector sequence is reshaped into a two-dimensional spatial structure to generate a modality semantic feature map.

[0028] In this embodiment, step four specifically includes: The fused feature tensor is upsampled to the same spatial size as the original image through bilinear interpolation to generate a spatially aligned feature tensor. Copy the mask guide map along the channel dimension to form a guide mask tensor with the same channel dimension as the spatially aligned feature tensor. The spatial alignment feature tensor and the guiding mask tensor are multiplied element-wise at the channel level to generate the guiding fusion feature tensor. The guided fusion feature tensor is transformed by a 3×3 convolution and a ReLU activation function to generate an intermediate risk feature tensor; the intermediate risk feature tensor is channel compressed and mapped by a 1×1 convolution and then transformed into a risk level prediction vector for each pixel position by a Softmax classifier. The risk level prediction vector is a three-dimensional vector, which corresponds to the prediction probability values ​​of the current pixel position belonging to the low risk level, the medium risk level, and the high risk level, respectively. The category index corresponding to the maximum predicted probability value in the risk level prediction vector of each pixel location is used as the risk label, and a risk level label map is constructed; the risk labels are 0: low risk, 1: medium risk and 2: high risk.

[0029] In this embodiment, step five specifically includes: Based on the spatial coordinates of each tower site in the tower site spatial vector layer, a spatial window with a set side length is constructed in the risk level label map with the spatial coordinates as the center. Extract the risk labels corresponding to all pixels within the spatial window, and count the number of pixels with risk labels of 0, 1, and 2 respectively, and record them as the number of low-risk pixels, medium-risk pixels, and high-risk pixels respectively. Calculate the proportions of low-risk, medium-risk, and high-risk pixels to the total number of pixels in the spatial window, respectively, to obtain the low-risk proportion, medium-risk proportion, and high-risk proportion, and then form the risk level distribution probability vector of the current tower site in order. Obtain the area that intersects with the spatial window in the hazard area mask image, and take it as the hazard area segment of the current tower site. Count the number of pixels with a mask value of 1 in the hazard area segment as the number of hazard area pixels of the current tower site. The spatial coordinates of the current tower site, the probability vector of risk level distribution, and the number of pixels in the hidden danger area are sequentially used to construct the tower site risk record vector. Organize the risk record vectors of all tower sites into a tower site risk assessment table by row.

[0030] In this embodiment, step six specifically includes: Project the tower site risk assessment table onto geospatial coordinates to generate a tower site risk level distribution layer; The mask image of the potential hazard area is overlaid on the remote sensing image layer to form a spatial mask layer of the potential hazard area. The risk level label map is subjected to pseudo-color mapping to generate a risk level distribution layer, wherein 0 is mapped to green, 1 to yellow, and 2 to red in the pseudo-color mapping. The tower location risk visualization layer, the hidden danger area spatial masking layer, and the risk level distribution layer are aligned and overlaid based on spatial coordinates to generate a spatial risk visualization layer set, which is then output to the user interaction terminal for display.

[0031] In this embodiment, the user interaction feedback data includes risk level label correction information, hazard area boundary adjustment information, and supplementary hazard area labeling; By fusing user interaction feedback data with multi-source spatial data, an interaction feedback training sample set is constructed; Incremental updates are performed on the improved SegFormer network and the risk level prediction model based on the interactive feedback training sample set.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a power transmission project in a high-altitude mountainous area. The goal was to identify areas around tower sites where construction safety hazards were caused by topographic, geological, and environmental disturbances, and to provide quantitative risk classification support for construction site selection and project management.

[0033] The construction area features dramatic terrain variations, including canyons, landslides, steep slopes, and intersecting rivers. The project team collected data including: 2-meter resolution remote sensing imagery, a 30-meter resolution DEM elevation map, a 1:10000 geological hazard risk factor map, the latest tower site layout vector map (124 tower sites in total), and an NDVI vegetation index map. All layers underwent coordinate system transformation, map cropping, spatial registration, and preprocessing to form a fused layer tensor. An improved SegFormer network was used to model the fused layer tensor, automatically extracting landslide accumulation areas, vegetation destruction patches, steep slopes, and potential collapse edge areas within the construction area, generating a hazard area mask map. A mask-guided mechanism was introduced to construct a risk level prediction network, labeling and predicting the risk level at each spatial location. This prediction was further overlaid with the tower site spatial vector map layer for analysis, outputting the risk level distribution probability and hazard pixel percentage for each tower site.

[0034] To verify the effectiveness of the present invention, the method of the present invention is compared and analyzed with the following methods: (1) SegFormer method: The standard SegFormer network is used to model the tensor of the fused layer, without introducing modal grouping, structural prior and alternating attention mechanism. (2) U-Net remote sensing segmentation method: The standard U-Net structure is used, and only remote sensing image data is used as input to perform end-to-end semantic segmentation and output risk level map and hidden danger area map. (3) Traditional remote sensing supervised classification method: The SVM model is used, based on the artificial feature vector composed of remote sensing image data, digital elevation model data, tower site spatial vector data, geological disaster risk factor data and normalized vegetation index data, supervised training and hidden danger classification are performed based on pixel-level labels, and the hidden danger area and tower site risk level are output. The experimental results are shown in Table 1.

[0035] Table 1. Performance Comparison of Different Methods in Power Grid Construction Risk and Hazard Identification Task

[0036] As shown in Table 1, the method of this invention outperforms the comparative methods in multiple metrics. Regarding the average IoU in hazard areas, the method of this invention achieves 0.782, significantly higher than SegFormer's 0.684, U-Net remote sensing segmentation's 0.635, and traditional remote sensing supervised classification's 0.472, indicating that this invention has stronger discriminative ability in identifying the boundaries of spatial hazard areas. In terms of F1 score and hazard identification recall, the method of this invention achieves 0.841 and 0.884 respectively, showing significant improvements over other methods, verifying that the method of this invention provides a more comprehensive and accurate identification of spatial terrain anomalies, vegetation disturbance, and geological disaster areas.

[0037] In terms of risk level classification accuracy, the method of this invention achieves 91.3%, a significant improvement compared to 84.5% for the SegFormer method and 80.9% for the U-Net remote sensing segmentation method. This indicates that the risk level prediction model designed in this invention has stronger robustness and accuracy in spatial location risk classification. Regarding the tower site risk assessment discrepancy rate and user interaction correction ratio, the method of this invention has 5.7% and 4.1% respectively, far lower than the comparative methods, demonstrating that the method of this invention is more stable and reliable in terms of tower site-level risk assessment and subsequent correction requirements.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying potential risks and hazards in power grid construction based on multi-source spatial data, characterized in that, Includes the following steps: Step 1: Collect multi-source spatial data of the power grid construction area, and perform coordinate unification and spatial registration on the spatial data; Step 2: Standardize the multi-source spatial data to generate a fused layer tensor; Step 3: Input the fused layer tensor into the improved SegFormer network to perform multimodal spatial feature modeling and semantic enhancement fusion. The improved SegFormer network introduces modal grouping, elevation prior and alternating attention mechanism to generate fused feature tensor, and generates a hazard area mask map through multi-scale modeling and threshold judgment. Step 4: Construct a risk level prediction model, and based on the fusion feature tensor and the hidden danger area mask map, determine the risk level status of each spatial location in the power grid construction area and generate a risk level label map; Step 5: Perform spatial overlay analysis on the risk level label map, the hidden danger area mask map, and the tower site spatial vector map layer to extract the risk level distribution and hidden danger area segments corresponding to each tower site and generate a tower site risk assessment table; Step Six: Visualize and analyze the tower site risk assessment form, risk level label map, and hazard area mask map, and display them on the user interaction terminal; Step 7: Collect user interaction feedback data and incrementally update the improved SegFormer network and risk level prediction model.

2. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, Step one specifically includes: The multi-source spatial data includes remote sensing image data, digital elevation model data, tower site spatial vector data, geological hazard risk factor data, and normalized vegetation index data; The coordinate system of the multi-source spatial data is unified by the projection transformation method based on EPSG encoding, the spatial reference is unified by the affine transformation method, and the layer space clipping operation is performed by the vector polygon mask to generate remote sensing image layer, digital elevation model layer, tower site spatial vector layer, geological hazard layer and vegetation index layer.

3. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, Step two specifically includes: The remote sensing image layer is normalized by channel linear normalization of the red, green and blue band pixel values, and the normalized remote sensing image data is spatially resampled to unify the resolution and generate a remote sensing image raster layer. The slope factor layer is calculated based on the digital elevation model layer. Specifically, the slope value is calculated at each pixel location using a third-order finite difference algorithm. The slope factor layer is then normalized using a channel linear normalization method to generate a slope raster layer. Perform a vector-to-raster conversion operation on the spatial vector layer of tower locations. Specifically, use the nearest neighbor mapping method to map each tower location to the corresponding raster cell to generate a tower location raster layer. Spatially resample the geological hazard layer and the vegetation index layer, unify the resolution, and perform scale alignment and map boundary consistency processing to generate a geological hazard raster layer and a vegetation index raster layer. The remote sensing image raster layer, slope raster layer, tower location raster layer, geological disaster raster layer, and vegetation index raster layer are stitched together according to the channel dimension to form a fused layer tensor. The dimensions of the fused layer tensor include image height, image height and width, and number of channels.

4. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, Step three specifically includes: The fused layer tensor is divided into several modal channel groups according to channels. The modal channel groups include remote sensing image modal channel group, elevation structure modal channel group, risk labeling modal channel group, vegetation disturbance modal channel group, and tower location modal channel group. Semantic modeling is performed on each modal channel group through an independent Transformer encoding path to generate a modal semantic feature map. Each Transformer encoding path includes a convolutional downsampling layer, a patch construction module, and a MixTransformer encoding module, specifically: Digital elevation model data is introduced and normalized using the channel linear normalization method to generate an elevation raster layer. The elevation raster layer and modal channel group are size-compressed by a 2D convolution with a kernel of 5×5, a stride of 2, and a symmetric zero-filling method. The compression tensor of elevation raster and modal channel group are then generated by nonlinear mapping with batch normalization and GELU activation function. The elevation raster compression tensor and the modal compression tensor are divided into several patch blocks according to the same window size. Each patch block is flattened into a one-dimensional vector, linearly mapped by MLP, and projected onto a preset dimension to generate elevation embedding vector and patch feature vector. Learnable weight coefficients are introduced, and the elevation embedding vector is multiplied element-wise with the learnable weight coefficients and then vector-added with the Patch feature vector to generate a hybrid Patch feature vector. The mixed patch feature vectors corresponding to the current modality channel group are combined into a mixed patch feature vector sequence and input into the Mix Transformer encoding module to generate a modality semantic feature map. The modal semantic feature maps generated by all Transformer encoding paths are batch normalized and then channel-wise concatenated and fused to generate a fused feature tensor. The fused feature tensor is input in parallel to three convolutional branches, and 1×1 convolution, 3×3 convolution and 5×5 convolution operations are performed respectively to generate local perceptual feature maps at three scales. The local perceptual feature maps at three scales are fused element-wise by addition to generate an enhanced convolutional feature map. The enhanced convolutional feature map is processed by 1×1 convolution to perform channel compression mapping, and nonlinear mapping is performed through two MLP layers to generate semantic fusion feature maps. The semantic fusion feature map is upsampled to be consistent with the spatial size of the original image, and a probability map of the hidden danger area is generated by the Sigmoid activation function. Set a hazard threshold, and judge the hazard for each pixel in the hazard area probability map. If the probability value corresponding to each pixel is less than the hazard threshold, set the pixel value to 0; otherwise, set the pixel value to 1. Then, use the hazard area probability map after hazard judgment as the hazard area mask map.

5. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 4, characterized in that, The Mix Transformer encoding module includes four Transformer encoding layers, which perform alternating attention operations on the mixed patch feature vector sequence, specifically including: The first and third Transformer encoders perform multi-head self-attention operations within a local window on the input feature vector sequence, while the second and fourth Transformer encoders perform pooling-enhanced attention operations on the input feature vector sequence. The pooling-enhanced attention operation specifically involves: The input feature vector sequence is processed by two-dimensional average pooling and two-dimensional max pooling to obtain the average pooling feature vector sequence and the max pooling feature vector sequence, respectively. The average pooling feature vector sequence and the max pooling feature vector sequence are concatenated along the channel dimension, and a channel transformation is performed through a one-dimensional convolution layer to generate a pooling fusion vector sequence. The pooling-sensing fusion vector sequence is fused with the input feature vector sequence through element-wise addition to form an enhanced feature vector sequence. The enhanced feature vector sequence is subjected to self-attention operation to generate the output feature vector sequence of the current layer's Transformer encoder; The input to the first layer Transformer encoder is the sequence of mixed patch feature vectors, and the input to the (n+1)th layer Transformer encoder is the sequence of output feature vectors of the nth layer Transformer encoder, where n takes the values ​​1, 2, and 3. The output feature vector sequence of the 4th layer Transformer encoder is used as the modality encoding feature vector sequence, and the modality encoding feature vector sequence is reshaped into a two-dimensional spatial structure to generate a modality semantic feature map.

6. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, Step four specifically includes: The fused feature tensor is upsampled to the same spatial size as the original image through bilinear interpolation to generate a spatially aligned feature tensor. Copy the mask guide map along the channel dimension to form a guide mask tensor with the same channel dimension as the spatially aligned feature tensor. The spatial alignment feature tensor and the guiding mask tensor are multiplied element-wise at the channel level to generate the guiding fusion feature tensor. The guided fusion feature tensor is transformed by a 3×3 convolution and a ReLU activation function to generate an intermediate risk feature tensor; the intermediate risk feature tensor is channel compressed and mapped by a 1×1 convolution and then transformed into a risk level prediction vector for each pixel position by a Softmax classifier. The risk level prediction vector is a three-dimensional vector, which corresponds to the prediction probability values ​​of the current pixel position belonging to the low risk level, the medium risk level, and the high risk level, respectively. The category index corresponding to the maximum predicted probability value in the risk level prediction vector of each pixel location is used as the risk label, and a risk level label map is constructed; the risk labels are 0: low risk, 1: medium risk and 2: high risk.

7. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, Step five specifically includes: Based on the spatial coordinates of each tower site in the tower site spatial vector layer, a spatial window with a set side length is constructed in the risk level label map with the spatial coordinates as the center. Extract the risk labels corresponding to all pixels within the spatial window, and count the number of pixels with risk labels of 0, 1, and 2 respectively, and record them as the number of low-risk pixels, medium-risk pixels, and high-risk pixels respectively. Calculate the proportions of low-risk, medium-risk, and high-risk pixels to the total number of pixels in the spatial window, respectively, to obtain the low-risk proportion, medium-risk proportion, and high-risk proportion, and then form the risk level distribution probability vector of the current tower site in order. Obtain the area that intersects with the spatial window in the hazard area mask image, and take it as the hazard area segment of the current tower site. Count the number of pixels with a mask value of 1 in the hazard area segment as the number of hazard area pixels of the current tower site. The spatial coordinates of the current tower site, the probability vector of risk level distribution, and the number of pixels in the hidden danger area are sequentially used to construct the tower site risk record vector. Organize the risk record vectors of all tower sites into a tower site risk assessment table by row.

8. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, Step six specifically includes: Project the tower site risk assessment table onto geospatial coordinates to generate a tower site risk level distribution layer; The mask image of the potential hazard area is overlaid on the remote sensing image layer to form a spatial mask layer of the potential hazard area. The risk level label map is subjected to pseudo-color mapping to generate a risk level distribution layer, wherein 0 is mapped to green, 1 to yellow, and 2 to red in the pseudo-color mapping. The tower location risk visualization layer, the hidden danger area spatial masking layer, and the risk level distribution layer are aligned and overlaid based on spatial coordinates to generate a spatial risk visualization layer set, which is then output to the user interaction terminal for display.

9. The method for identifying power grid construction risks and hidden dangers based on multi-source spatial data according to claim 1, characterized in that, The user interaction feedback data includes risk level label correction information, hazard area boundary adjustment information, and supplementary labeling of hazard areas; By fusing user interaction feedback data with multi-source spatial data, an interaction feedback training sample set is constructed; Incremental updates are performed on the improved SegFormer network and the risk level prediction model based on the interactive feedback training sample set.

Images