Substation detection method based on multi-scale edge sensing network
By using the PEA-YOLO11 neural network model based on a multi-scale edge sensing network, the problems of low efficiency and insufficient accuracy in traditional substation equipment detection are solved, realizing high-precision and automated equipment status monitoring, and improving the robustness and detection accuracy of substation equipment detection.
Patent Information
- Application Number
- CN202510810667.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-10-17
AI Technical Summary
Traditional substation equipment inspection relies on manual inspection, which is inefficient and easily affected by subjective experience and environmental factors. Furthermore, traditional image processing technology is difficult to cope with the diverse equipment failure modes and complex environments, resulting in insufficient detection accuracy and poor robustness. In particular, it is easy to misjudge or miss when identifying subtle defects.
A PEA-YOLO11 neural network model based on a multi-scale edge-aware network is adopted. The PyramidEdgeAggregator module is used for multi-scale feature extraction and edge information enhancement. The EdgeRefinementBlock module is combined to extract edge information. A PyramidEdgeAggregator (PEA) module is designed to replace the C3K2 module in the YOLO11 backbone network, which enhances the model's sensitivity to edges and performs feature fusion.
It improves the accuracy and robustness of substation equipment inspection, enabling better identification of subtle defects in equipment under multi-scale and complex environments, and achieving high-precision automated inspection.
Smart Images
Figure CN120808101A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of image recognition, and in particular to a substation detection method based on a multi-scale edge perception network. BACKGROUND
[0002] With the continuous expansion of the power system, the stable operation of the substation equipment, as the core hub of the power transmission and distribution network, is directly related to the safety and reliability of the power grid. Traditional substation equipment detection mainly relies on manual inspection, and professional personnel periodically visually inspect the equipment. However, manual inspection has the defects of low efficiency, high labor intensity, and being easily affected by subjective experience and environmental factors (such as light and weather), especially in scenes with a large number of equipment types and complex distribution, which is prone to missed detection or misjudgment. In addition, although traditional image processing technology has achieved automatic detection to some extent, it relies on hand-designed features and fixed rules, and is difficult to cope with the challenges of diverse equipment failure modes and complex and variable environments, resulting in insufficient detection accuracy and poor robustness. For example, when identifying subtle defects such as insulator cracks and disconnecting switch oxidation, traditional methods often fail due to insufficient feature extraction. Therefore, there is an urgent need for a high-precision, automated and adaptable detection method to improve the intelligent level of substation equipment state monitoring. SUMMARY
[0003] The present application provides a substation detection method based on a multi-scale edge perception network to solve the problem that existing image recognition technology cannot accurately identify substation equipment due to different equipment scales or blurred equipment edges, which is prone to misjudgment or missed detection.
[0004] To solve the above technical problems, the present application provides a substation detection method based on a multi-scale edge perception network, comprising:
[0005] Collecting substation equipment image data and labeling, and randomly dividing the training set, test set and validation set;
[0006] The obtained sample set is preprocessed, data enhanced, etc.
[0007] A PEA-YOLO11 neural network model is constructed, and then the training set and the validation set are used to train the improved neural network model to obtain the PEA-YOLO11 neural network model after training is completed;
[0008] The test set is input into the obtained PEA-YOLO11 neural network model to obtain the substation equipment image test result and score the model performance.
[0009] Further, the substation equipment image data is collected, including: obtaining substation equipment image data with equipment category real labels, and performing image data preprocessing on the substation equipment image data with equipment category real labels to obtain processed substation equipment image data with equipment category real labels; and the processed substation equipment image data with equipment category real labels is randomly divided to obtain a training set, a test set and a verification set with equipment category real labels.
[0010] The PEA-YOLO11 neural network model is constructed, including: a PyramidEdgeAggregator (PEA) module is designed to replace the C3K2 module in the yolo11 backbone network, which combines multi-scale feature extraction, edge information enhancement and convolution operation.
[0011] Further, the PyramidEdgeAggregator (PEA) module includes:
[0012] Multi-scale feature extraction: multi-scale pooling is performed through average pooling to extract local information of different sizes, which helps to capture multi-level features of the image.
[0013] Edge enhancement: an EdgeRefinementBlock module is designed to extract edge information, so that the sensitivity of the network to the edge is enhanced, which is important for many visual tasks such as target detection, semantic segmentation, etc.
[0014] Feature fusion: the features extracted at different scales are aligned to the same scale through interpolation operation, then they are spliced together, and finally they are fused into a unified feature representation through convolution layer, which can improve the perception of multi-scale features of the model.
[0015] Further, the EdgeRefinementBlock module includes:
[0016] Average pooling (low-pass filtering): first, the input feature map is processed through a 3x3 average pooling operation to extract low-frequency information. This pooling operation will remove the high-frequency details of the image to obtain a smoothed feature map. The pooling operation uses a stride of 1 and a padding of 1, so that the size of the pooled feature map is the same as that of the input feature map.
[0017] Edge extraction (high-pass filtering): by subtracting the smoothed feature map (low-frequency part) from the original feature map (high-frequency part), the edge information (high-frequency part) is obtained. This process can be regarded as removing the background from the original image, highlighting the edges and details.
[0018] Convolutional processing of edge information: the extracted edge information is processed using a convolutional layer to further process the edge information, and a Sigmoid activation function is used to limit the output value to [0, 1]. This process makes the edge information smoother and more expressive.
[0019] Enhanced output: finally, the processed edge information is added to the original input feature map to obtain the enhanced output feature map. In this way, the edge details of the image are enhanced, so that the model can better focus on the important structures in the image.
[0020] Further, the multi-scale feature extraction includes:
[0021] The input feature map is processed by multiple branches, each branch corresponding to a different scale. Each branch uses average pooling to pool the input feature map, and the goal of pooling is to reduce the feature map to a preset resolution parameter, which is equivalent to extracting features at different scales. Then, two convolutional layers are used to process the pooled feature map.
[0022] Further, the feature fusion includes applying a convolutional layer to the original input feature map to extract local features. This convolutional operation does not perform pooling or scaling, and can capture detailed information in the image.
[0023] Finally, all feature maps of different scales are spliced, including the output of local convolution and the enhanced edge features of each scale. The spliced feature map is processed by a final convolutional layer to restore the channel number to the original input channel number. This feature fusion method allows the network to utilize edge information at different scales, thereby improving the processing capability of different object sizes and spatial context.
[0024] Further, the obtaining of the substation equipment image test result and the scoring of the model performance includes:
[0025] The test set is input into the trained PEA-YOLO11 neural network model to obtain the substation equipment categories output by the model during testing. The evaluation indicators are calculated based on the substation equipment categories output by the model during testing and the real labels of the equipment categories.
[0026] Compared with the prior art, the present application has the following beneficial effects:
[0027] The substation equipment image data acquired in real time is input into the trained PEA-YOLO11 neural network model, so that the trained PEA-YOLO11 neural network model extracts features of the substation equipment image data through a backbone network layer, obtains feature information maps of different dimensions, fuses features of the substation equipment feature maps of different dimensions through a neck network layer, obtains feature fusion maps of different dimensions, performs image recognition on the feature fusion maps of different dimensions through a head network layer, obtains image recognition results of different dimensions, and finally determines the final substation equipment category according to the image recognition results of different dimensions.
[0028] BRIEF DESCRIPTION OF DRAWINGS (DELETE IF NO DRAWINGS)
[0029] Figure 1 : a step flow chart of a substation detection method based on a multi-scale edge perception network according to an embodiment of the present application;
[0030] Figure 2 : a model structure diagram of a PEA-YOLO11 neural network model provided by an embodiment of the present application;
[0031] Figure 3 : a structure diagram of a PEA module provided by an embodiment of the present application
[0032] Figure 4 : a structure diagram of an EdgeRefinementBlock module provided by an embodiment of the present application; DETAILED DESCRIPTION
[0033] REFERENCE Figure 1 and Figure 2 , Figure 1 : a step flow chart of a substation detection method based on a multi-scale edge perception network according to an embodiment of the present application, Figure 2 : a method provided by an embodiment of the present application; the method comprises at least the following steps:
[0034] Step S1: collect substation equipment image data and label, and randomly divide a training set, a test set and a validation set;
[0035] Step S3: pre-process the obtained sample set, and perform operations such as data enhancement;
[0036] Step S3: construct a PEA-YOLO11 neural network model, and then train the improved neural network model using the training set and the validation set to obtain the PEA-YOLO11 neural network model after training is completed;
[0037] Step S4: input the test set into the obtained PEA-YOLO11 neural network model to obtain a substation equipment image test result and score the performance of the model.
[0038] In the embodiment, the substation equipment image data is collected, including: obtaining substation equipment image data with equipment category real labels, and performing image data preprocessing on the substation equipment image data with equipment category real labels to obtain processed substation equipment image data with equipment category real labels; and the processed substation equipment image data with equipment category real labels is randomly divided to obtain a training set, a test set and a verification set with equipment category real labels.
[0039] Reference Figure 2 In the embodiment, the PEA-YOLO11 neural network model is constructed, including: a PyramidEdgeAggregator (PEA) module is designed to replace the C3K2 module in the yolo11 backbone network, which combines multi-scale feature extraction, edge information enhancement and convolution operation.
[0040] As Figure 3 shown, in the embodiment, the design of the PyramidEdgeAggregator (PEA) module includes:
[0041] Multi-scale feature extraction: multi-scale pooling is performed through average pooling to extract local information of different sizes, which helps to capture multi-level features of the image.
[0042] Edge enhancement: an EdgeRefinementBlock module is designed to extract edge information, so that the sensitivity of the network to the edge is enhanced, which is important for many visual tasks such as target detection, semantic segmentation, etc.
[0043] Feature fusion: the features extracted at different scales are aligned to the same scale through interpolation operation, then they are spliced together, and finally they are fused into a unified feature representation through convolution layer, which can improve the perception of multi-scale features of the model.
[0044] As Figure 4 In the embodiment, the design of the EdgeRefinementBlock module includes:
[0045] Average pooling (low-pass filtering): first, the input feature map is processed through a 3x3 average pooling operation to extract low-frequency information. This pooling operation will remove the high-frequency details of the image to obtain a smoothed feature map. The pooling operation uses a stride of 1 and a padding of 1, so that the size of the pooled feature map is the same as that of the input feature map.
[0046] Edge Extraction (High Pass Filtering): By subtracting the smoothed feature map (low frequency part) from the original feature map, the edge information (high frequency part) is obtained. This process can be seen as removing the background from the original image, highlighting the edges and details.
[0047] Convolution Processing Edge Information: The extracted edge information is processed using a convolution layer to further process the edge information and limit the output value between [0, 1] through a Sigmoid activation function. This process makes the edge information smoother and more expressive.
[0048] Enhanced Output: Finally, the processed edge information is added to the original input feature map to obtain the enhanced output feature map. In this way, the edge details of the image are enhanced, so that the model can better focus on the important structures in the image.
[0049] The formula is as follows:
[0050] edge = Sigmond(Conv(Input - AvgPool(Input))) 1
[0051] output = Input + edge 2
[0052] Where Input represents the input feature map, AvgPool(Input) represents the average pooling result, and output represents the output feature map.
[0053] The core purpose of the EdgeRefinementBlock module is to enhance the edge information in the input feature map. By subtracting the low frequency part to get the high frequency part, the edge and detail information in the image is emphasized. The role of the convolution layer is to further refine these edge information, so that they are smoother and more expressive. Finally, combining the enhanced edge information with the original input can improve the clarity and detail performance of the image, especially in image segmentation, object detection and other tasks, edge information is crucial for accurate recognition.
[0054] As shown in Figure 3 In this embodiment, the multi-scale feature extraction includes:
[0055] The input feature map is processed through multiple branches, each branch corresponding to a different scale. Each branch uses average pooling to pool the input feature map, and the goal of pooling is to reduce the feature map to a preset resolution parameter, which is equivalent to extracting features at different scales. Then, two convolution layers are used to process the pooled feature map. The first convolution layer reduces the number of channels from the input channel number to the input channel number / / preset resolution parameter group
[0056] The formula is as follows:
[0057]
[0058] where pool represents the average-pooled feature map, inc is the input channel number, bins is the preset resolution array, len(bins) represents the number of branches, and 1 indicates that the convolution kernel size is 1x1.
[0059] The second convolutional layer uses a 3x3 convolution, and a grouped convolution is used to further optimize the feature representation.
[0060] The formula is as follows:
[0061]
[0062] where inc is the feature map output by the first convolutional layer, 3 indicates that the convolution kernel size is 3x3, and g represents the number of grouped convolution groups, each of which is an independent convolution.
[0063] In this way, each branch can extract features at different scales and capture edge information at different spatial resolutions.
[0064] For each scale of feature map, the EdgeRefinementBlock module is used to enhance the edge information. In this way, the edge information at each scale can be enhanced, so that the model can better focus on the structural information in the image.
[0065] Since the feature maps at different scales have different sizes, the interpolate is used to upsample the feature maps to the size of the original input feature map.
[0066] The formula is as follows:
[0067] upsample = F.interpolate(output, x size[2：] ) 5
[0068] where output is the output feature map of the EdgeRefinementBlock module, x size[2：] is the height and width (H, W) of the input feature map, F.interpolate is a bilinear interpolation upsample operation, and upsample is the upsampled feature map,
[0069] As shown in Figure 3 In this example, the feature fusion includes applying a convolutional layer on the original input feature map to extract local features. This convolutional operation does not perform pooling or scaling, and can capture detailed information in the image. The formula is as follows:
[0070] Local feature = local conv(x)6
[0071] where x denotes the input feature map, local conv denotes the local convolution operation, and the size of the convolution kernel is 3x3. feature denotes the feature map output by the local convolution.
[0072] Finally, all feature maps of different scales are spliced, including the output of the local convolution and the enhanced edge features of each scale. The spliced feature map is processed by a final convolution layer to restore the channel number to the original input channel number. This feature fusion method enables the network to comprehensively utilize edge information at different scales, thereby improving the processing capability for different object sizes and spatial context.
[0073] The formula is as follows:
[0074] Concat out =torch.cat([local featyre ,upsample1,...upsample n ],1) 7
[0075] where upsampl n e denotes the feature map after upsampling on the nth branch, torch.cat denotes splicing along the channel dimension (dim=1), and concat out denotes the spliced feature map.
[0076] final out =final conv (concat out ) 8
[0077] where final conv denotes the final convolution operation, and the final output of the final out module.
[0078] In this embodiment, the obtaining of the substation equipment image test result and the scoring of the model performance comprises:
[0079] The test set is input into the trained PEA-YOLO11 neural network model to obtain the substation equipment category output by the model during testing, and evaluation indicators are calculated according to the substation equipment category output by the model during testing and the real label of the equipment category.
[0080] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.
Claims
1. A substation detection method based on a multi-scale edge perception network, characterized in that: include: Step 1: Collect and annotate substation equipment image data, and randomly divide it into training set, test set and validation set; Step 2: Preprocess the obtained sample set, perform data enhancement and other operations; Step 3: Build a PEA-YOLO11 neural network model, and then use the training set and validation set to train the improved neural network model to obtain the trained PEA-YOLO11 neural network model; Step 4: Input the test set into the obtained PEA-YOLO11 neural network model to obtain the substation equipment image test results and score the model performance.
2. The substation detection method based on a multi-scale edge perception network according to claim 1 is characterized in that: The collecting of substation equipment image data includes: Obtain substation equipment image data with true device category labels, perform image data preprocessing on the substation equipment image data with true device category labels to obtain processed substation equipment image data with true device category labels, and randomly divide the processed substation equipment image data with true device category labels to obtain training set, test set and validation set with true device category labels.
3. The substation detection method based on a multi-scale edge perception network according to claim 1 is characterized in that: The PEA-YOLO11 neural network model is constructed, including: A PyramidEdgeAggregator (PEA) module is designed to replace the C3K2 module in the yolo11 backbone network, which combines multi-scale feature extraction, edge information enhancement and convolution operations.
4. The substation detection method based on a multi-scale edge perception network according to claim 2 is characterized in that: The PyramidEdgeAggregator (PEA) module is designed to include: Multi-scale feature extraction: Multi-scale pooling is performed through average pooling to extract local information of different sizes, which helps to capture the multi-level features of the image; Edge Enhancement: We designed an EdgeRefinementBlock module specifically for extracting edge information, which increases the network's sensitivity to edges. This is important for many visual tasks (such as object detection and semantic segmentation). Feature fusion: Features extracted at different scales are aligned to the same scale through interpolation operations, then spliced together, and finally fused into a unified feature representation through a convolutional layer, which can improve the model's perception of multi-scale features.
5. The substation detection method based on a multi-scale edge perception network according to claim 3 is characterized in that: The EdgeRefinementBlock module is designed to include: Average pooling (low-pass filtering): First, the input feature map is subjected to a 3x3 average pooling operation to extract low-frequency information; this pooling operation removes high-frequency details from the image and obtains a smoothed feature map; the pooling operation uses a stride of 1 and a padding of 1, so that the size of the pooled feature map is the same as the input feature map; Edge extraction (high-pass filtering): By subtracting the smoothed feature map (low-frequency part) from the original feature map, edge information (high-frequency part) is obtained; this process can be regarded as removing the background from the original image and highlighting the edges and details; Convolution processing edge information: The extracted edge information is convolved, further processed using a convolution layer, and the output value is limited to [0, 1] through a Sigmoid activation function; this process makes the edge information smoother and more expressive; Enhanced output: Finally, the processed edge information is added to the original input feature map to obtain the enhanced output feature map; in this way, the edge details of the image are enhanced, allowing the model to better focus on important structures in the image.
6. The substation detection method based on a multi-scale edge perception network according to claim 3 is characterized in that: The multi-scale feature extraction includes: The input feature map is processed by multiple branches, each corresponding to a different scale. Each branch uses average pooling to pool the input feature map. The goal of pooling is to reduce the feature map to a preset resolution parameter, which is equivalent to extracting features at different scales. The pooled feature map is then processed using two convolutional layers.
7. The substation detection method based on a multi-scale edge perception network according to claim 3 is characterized in that: The feature fusion includes: Apply a convolution layer to the original input feature map to extract local features. This convolution operation does not perform pooling or scaling, and can capture the details in the image. Finally, all feature maps at different scales are concatenated, including the output of the local convolution and the enhanced edge features at each scale. The concatenated feature maps are processed through a final convolutional layer to restore the number of channels to the original input channel count. This feature fusion method enables the network to comprehensively utilize edge information at different scales, thereby improving its ability to handle different object sizes and spatial contexts.
8. The substation detection method based on a multi-scale edge perception network according to claim 3 is characterized in that: Obtaining substation equipment image test results and scoring model performance includes: The test set is input into the trained PEA-YOLO11 neural network model to obtain the substation equipment category output during the model test. The evaluation index is calculated based on the substation equipment category output during the model test and the true label of the equipment category.