Transformer substation defect detection method guided by changing attention

Abstract

The invention discloses a transformer substation defect detection method guided by changing attention. The method comprises the following steps: step 1, acquiring a reference image and an inspection image by transformer substation inspection equipment; 2, the reference image and the inspection image which are input in a double-time-phase mode are sent into the same backbone network for feature extraction, and multi-scale feature maps are obtained through extraction; 3, performing fusion processing on the multi-scale feature map of the reference image and the multi-scale feature map of the inspection image through a change attention guiding module to obtain an enhanced multi-scale feature map; 4, sending the enhanced multi-scale feature map into a feature pyramid network for further feature fusion and mapping to obtain a multi-scale pyramid feature map; and 5, finally, feeding the multi-scale pyramid feature map into a detection head to realize the detection of a defect target. The method is favorable for suppressing background interference and reducing the complexity of model design.

Description

Technical Field

[0001] The invention relates to the technical field of transformer substation defect detection, in particular to a transformer substation defect detection method guided by variable attention. Background Art

[0002] The power system is a critical infrastructure. As the central link of "generation, transmission, transformation, distribution and use", the timely detection and handling of substation defects is of great significance to maintaining the stable operation of the power grid. Affected by the natural environment, high-voltage loads and human factors, substations are prone to equipment defects such as foreign objects, breakage, oil leakage, as well as safety risks such as personnel smoking and not wearing safety helmets. For a long time, substation defects have relied on operation and maintenance personnel to discover through manual inspections. There are problems such as high labor intensity, low work efficiency, and many missed detections and false detections, which have brought hidden dangers to the safe operation of the power system. With the introduction of the concept of power vision and the development of the ubiquitous power Internet of Things, the automatic detection of substation defects based on computer vision technology and artificial intelligence technology has become a hot topic in research and application. That is, the surveillance cameras and inspection robots deployed in the substation are used to collect images of substation equipment, and then artificial intelligence algorithms are used to intelligently analyze and detect defects on the collected images. For example Figure 1 As shown, common substation defects include: broken meters, broken insulators, oil leakage, metal rust, broken respirators, abnormal door closure, oil stains on component surfaces, hanging suspended matter, bird nests, damaged doors, windows, walls and floors, broken covers, unlocked frame ladders, intrusion across the line, missing covers, not wearing safety helmets, not wearing work clothes, not wearing safety ropes, smoking, people falling to the ground, fire smoke, blurred dials, wall leakage, small animals intruding, abnormal meter readings, abnormal oil level of respirator oil seals, discoloration of silicone, water on the ground, roof leakage, etc.

[0003] The samples collected by inspection equipment such as substation inspection robots or surveillance cameras at different times and at the same point contain multi-temporal information, and the substation defects correspond to regional changes in multi-temporal samples. Figure 2 As shown, the first row is normal samples collected by the substation inspection equipment at different points, and the second row is defective samples collected by the inspection equipment at the same points but at different times. The first two columns correspond to abnormal door closure defects, and the third column corresponds to the frame ladder unlocked defect.

[0004] However, most of the existing substation defect detection methods directly use one-stage (such as YOLO) or two-stage (such as Faster-RCNN) target detection methods to perform defect detection on inspection images, without using the temporal information contained in the inspection images. A few methods based on multi-temporal change detection first perform multi-scale fusion on differential features, and then directly send them to the defect detection head to locate defects, that is, defect detection is performed on the differential feature map. This type of method only considers feature differences and fails to fully utilize the defect feature information in the inspection feature map. It is easily affected by environmental changes and produces false detection. The method used in the present invention is significantly different from the above method. The present invention maps the differential features as attention weights, enhances the inspection image feature map, and then sends them to the defect detection head to locate defects, that is, defect detection is performed on the inspection feature map enhanced with differential features. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art.

[0006] To this end, the present invention proposes a substation defect detection method guided by change attention, which utilizes the multi-temporal information contained in the substation inspection samples, embeds a change attention guidance module in the substation defect detection network, and reveals the existence of defects through feature changes, which is conducive to suppressing background interference and reducing the complexity of model design.

[0007] A method for detecting substation defects guided by variable attention according to an embodiment of the present invention comprises the following steps: Step 1: Dual-phase input: Reference images collected by substation inspection equipment and inspection images , where the reference image It is a normal sample without defects collected by the substation inspection equipment at a certain inspection point. The samples collected by the substation inspection equipment at the same point and at different times need to be judged whether they contain defects; inspection images Automatically determine whether there are defects through the substation defect detection algorithm; Step 2: Feature extraction: Reference image of bi-phase input and inspection images They are respectively sent to the same backbone network for feature extraction, and multi-scale feature maps are extracted respectively; Step 3: Fusion Processing: Reference Image Multi-scale feature maps and inspection images The multi-scale feature map is fused by the variable attention guidance module to obtain the enhanced multi-scale feature map; Step 4, feature fusion and mapping: Send the enhanced multi-scale feature map to the feature pyramid network for further feature fusion and mapping to obtain a multi-scale pyramid feature map; Step 5: Obtain defect detection results: Finally, the multi-scale pyramid feature map is sent to the detection head to detect defect targets.

[0008] According to one embodiment of the present invention, in step 1, the substation defect detection algorithm adopts the Faster-RCNN framework.

[0009] According to one embodiment of the present invention, in step 2, the backbone network is a ResNet50 backbone network.

[0010] According to one embodiment of the present invention, in the ResNet50 backbone network, the size of the input image is W×H×3, where W is the width, H is the height, and 3 is the number of channels; the size of the feature map output after the convolution layer Conv1 is W / 2×H / 2×64; the second layer feature map output by the backbone network after the convolution layer Conv2_x is The size is W / 4×H / 4×256; the third layer feature map output by the backbone network after the convolution layer Conv3_x The size is W / 8×H / 8×512; the fourth layer feature map output by the backbone network after the convolution layer Conv4_x The size is W / 16×H / 16×1024; the fifth layer feature map output by the backbone network after the convolution layer Conv5_x The dimensions are W / 32×H / 32×2048.

[0011] According to one embodiment of the present invention, in the feature pyramid network, The calculation process is: the fifth layer feature map After processing by Conv2d_1, we get , and then processed by Conv2d_2, we get ; Among them, Conv2d_1 represents a convolutional layer with a convolution kernel size of 1×1 and 256 channels; Conv2d_2 represents a convolutional layer with a convolution kernel size of 3×3 and 256 channels; The calculation process is: the fourth layer feature map After processing by Conv2d_1, we get , and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain ; The calculation process is: the third layer feature map After processing by Conv2d_1, we get , and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain ; The calculation process is: the second layer feature map After processing by Conv2d_1, we get , and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain ;in, It is the fifth layer feature map Input the FPN feature fusion network to obtain the feature map; It is the fourth layer feature map Input the FPN feature fusion network to obtain the feature map; It is the third layer feature map Input the FPN feature fusion network to obtain the feature map; It is the second layer feature map Input the FPN feature fusion network to obtain the feature map.

[0012] According to one embodiment of the present invention, in step 2, the reference image Send it to the backbone network for feature extraction, and extract the multi-scale feature map , , and ;in, The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, Corresponding feature map; inspection image Send it to the backbone network for feature extraction, and extract the multi-scale feature map , , and ;in, The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map.

[0013] According to one embodiment of the present invention, in the third step, the process of the change attention guidance module fusion processing is: first, input and Perform a difference operation, that is, subtract element by element ; Among them, input and Corresponding to the reference image and inspection images The feature map input of the same scale, that is and Corresponding to and ,or and Corresponding to and ,or and Corresponding to and ,or and Corresponding to and ;Then, and The feature map obtained after element-by-element subtraction is sent to the convolution module, and the features are mapped by the convolution module; then, the Sigmoid activation function is used to map the feature response to the (0,1) interval to obtain the weight ;at last, and Multiply element by element to get enhanced features .

[0014] According to one embodiment of the present invention, the width, height and number of channels of the input feature map are set to be , and ; The convolution module includes a first convolution layer, a second convolution layer, a third convolution layer and a fourth convolution layer; the convolution kernel size of the first convolution layer is 3×3, the feature map filling value p=1, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the second convolution layer is 5×5, the feature map padding value p=2, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the third convolution layer is 7×7, the feature map filling value p=3, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the fourth convolution layer is 1×1, the feature map filling value p=0, the step value s=1, and the number of convolution kernels ; Input feature map The first convolutional layer, the second convolutional layer, and the third convolutional layer extract feature maps in parallel, and then splice the feature maps to output Feature map, concatenated feature map Then pass through the fourth convolutional layer to output the feature map .

[0015] The beneficial effect of the present invention is that the samples collected by inspection equipment such as substation inspection robots or surveillance cameras at different times and the same points contain multi-temporal information. The substation defects correspond to regional changes in the multi-temporal samples. The regional change information is extracted by using the change attention module, which can effectively enhance the defect features in the backbone network extraction feature map, suppress background interference, and improve the accuracy of defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 This is a schematic diagram of common substation defects; Figure 2 It is a schematic diagram of normal samples and defective samples collected by substation inspection equipment; Figure 3 This is the overall structure diagram of the model.

[0018] Figure 4 It is a structural diagram of the ResNet50 backbone network and the FPN feature fusion network; Figure 5 is a schematic diagram of the structure of the variable attention guidance module; Figure 6 It is the structure diagram of the convolution module; Figure 7 It is the visualization result diagram of the differential feature map; Figure 8 It is a flow chart of the substation defect detection method guided by the change attention of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] A method for detecting substation defects guided by variable attention according to an embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] See Figure 8 A method for detecting substation defects guided by a change attention system of the present invention comprises the following steps: Step 1: Dual-phase input: Reference images collected by substation inspection equipment and inspection images , where the reference image It is a normal sample without defects collected by the substation inspection equipment at a certain inspection point. The samples collected by the substation inspection equipment at the same point and at different times need to be judged whether they contain defects; inspection images The substation defect detection algorithm automatically determines whether there are defects. In other words, the inspection image These are samples that may contain defects and are collected by substation inspection equipment at the same location and at different times.

[0022] See Figure 3 , The reference image is a normal sample collected by the substation inspection equipment at a certain inspection point. The sample is manually judged in advance whether it is a normal sample. The inspection images collected by the substation inspection equipment at the same location and at different times are also samples that need to be automatically determined by the substation defect detection algorithm to determine whether they contain defects.

[0023] The substation defect detection algorithm uses the Faster-RCNN framework, which is mainly composed of the ResNet50 backbone network, the FPN (Feature Pyramid Networks) module and the detection head. The ResNet50 backbone network is responsible for extracting the multi-scale feature map of the input image, the FPN module is responsible for fusing and mapping the multi-scale feature map to obtain the enhanced multi-scale pyramid feature map, and the detection head is responsible for classifying and regressing the defect target on the enhanced feature map.

[0024] Step 2: Feature extraction: Reference image of bi-phase input and inspection images They are respectively sent to the same backbone network for feature extraction, and multi-scale feature maps are extracted respectively; the backbone network is the ResNet50 backbone network.

[0025] The specific structure of the ResNet50 backbone network and the FPN feature fusion network is as follows Figure 4 shown.

[0026] In the ResNet50 backbone network, the size of the input image is W×H×3, where W is the width, H is the height, and 3 is the number of channels; the feature map size output after the convolution layer Conv1 is W / 2×H / 2×64; the second layer feature map output by the backbone network after the convolution layer Conv2_x is The size is W / 4×H / 4×256; the third layer feature map output by the backbone network after the convolution layer Conv3_x The size is W / 8×H / 8×512; the fourth layer feature map output by the backbone network after the convolution layer Conv4_x The size is W / 16×H / 16×1024; the fifth layer feature map output by the backbone network after the convolution layer Conv5_x The dimensions are W / 32×H / 32×2048.

[0027] Preferably, in the ResNet50 backbone network, the size of the input image is (width, height, and number of channels respectively), the output feature map size after convolution layer Conv1 is (width, height, and number of channels, respectively), and then pass through the convolutional layers Conv2_x, Conv3_x, Conv4_x, and Conv5_x in sequence for feature extraction. The width and height of the feature map output by each layer are 1 / 2 of the previous layer, respectively, and the number of channels is twice that of the previous layer. Specifically, after the convolutional layer Conv2_x, the output feature map size is 160×160×256; after the convolutional layer Conv3_x, the output feature map size is 80×80×512; after the convolutional layer Conv4_x, the output feature map size is 40×40×1024; after the convolutional layer Conv5_x, the output feature map size is 20×20×2048. The feature maps output by the convolutional layers Conv2_x, Conv3_x, Conv4_x, and Conv5_x are respectively recorded as , , and ,in, Represents the second layer features output by the backbone network, Represents the third layer features output by the backbone network, Represents the fourth layer features output by the backbone network, Represents the fifth layer features output by the backbone network.

[0028] Reference Images Send it to the backbone network for feature extraction, and extract the multi-scale feature map , , and ;in, The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map.

[0029] Inspection image Send it to the backbone network for feature extraction, and extract the multi-scale feature map , , and ;in, The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map.

[0030] Step 3: Fusion Processing: Reference Image Multi-scale feature maps and inspection images The multi-scale feature map of is fused by the variable attention guidance module to obtain an enhanced multi-scale feature map; specifically, Figure 3 As shown, reference image Multi-scale feature maps are obtained after extraction by the ResNet50 backbone network , , and Inspection image Multi-scale feature maps are obtained after extraction by the ResNet50 backbone network , , and . Reference Image and inspection images The multi-scale feature maps are fused by the variational attention guidance module.

[0031] The structure of the changing attention guidance module is as follows Figure 5 As shown, input and Corresponding to the reference image and inspection images The feature map input of the same scale, that is , , or . and First perform the difference operation, that is, subtract element by element: , and then the features are mapped through the convolution module, and then the Sigmoid activation function is used to map the feature response to the (0,1) interval to obtain the weight ,at last and Multiply element by element to get enhanced features .

[0032] and are two matrices with the same width and height but different numbers of channels, The size is ,and The size is , element-wise multiplication means taking and The elements of the corresponding positions of each channel matrix are multiplied together, and the calculation formula is as follows:

[0033] in, Represents the element of the cth channel, ith row, and jth column. The value of c is .

[0034] Specifically, the process of changing the attention guidance module fusion processing is: first, input and Perform a difference operation, that is, subtract element by element ; Among them, input and Corresponding to the reference image and inspection images The feature map input of the same scale, that is and Corresponding to and ,or and Corresponding to and ,or and Corresponding to and ,or and Corresponding to and ;Then, and The feature map obtained after element-by-element subtraction is sent to the convolution module, and the features are mapped by the convolution module; then, the Sigmoid activation function is used to map the feature response to the (0,1) interval to obtain the weight ;at last, and Multiply element by element to get enhanced features .

[0035] The core idea of ​​​​variation attention is: if the inspection image With reference image Compared with the defect, the color, texture and other features of the defect area will change more, while the feature changes of the background area (referring to the area where the defect is not) are much smaller (there will also be a small change due to the influence of lighting, environment, etc.). After the ResNet50 backbone network extracts features, the value difference of the feature map corresponding to the changed area is large, and the value difference of the feature map corresponding to the background area is small. The result of the calculation is that the defect area has a strong response value and the background area has a weak response value. The convolution module is responsible for further mapping the features, strengthening the defect area and suppressing the background response. The output value of the convolution module is mapped by the activation function to obtain a weight between 0 and 1. , the weight value of the changed area (that is, the defect area) is close to 1, while the weight value of the background area is close to 0. and After element-wise multiplication, The feature values ​​of the changed area are retained, and the feature values ​​of the background area are suppressed, thus achieving feature enhancement. The enhanced features are then sent to the FPN network for further feature fusion and mapping, and finally sent to the detection head to detect defect targets.

[0036] Among them, the detection head directly uses the structure in Faster-RCNN, including RPN (region proposalNetwork), ROI Align, classification sub-network, and regression sub-network. The input of the RPN network is the feature map extracted by the backbone network, and the output is the candidate target box, that is, the box with a relatively high probability of containing the target to be detected. The function of ROI Align is to project the candidate region box output by the RPN network onto the feature map extracted by the backbone network, that is, to extract the local feature map corresponding to the candidate target box. After obtaining the extracted local feature map, it passes through two fully connected layers for feature mapping, and the obtained feature vectors are sent to the classification sub-network and regression sub-network composed of a fully connected layer, respectively, to output the category probability and coordinates of the target box.

[0037] Step 4, feature fusion and mapping: Send the enhanced multi-scale feature map to the feature pyramid network for further feature fusion and mapping to obtain a multi-scale pyramid feature map; In the feature pyramid network (i.e., in the FPN feature fusion network), Conv2d_1 represents a convolutional layer with a convolution kernel size of 1×1 and 256 channels; Conv2d_2 represents a convolutional layer with a convolution kernel size of 3×3 and 256 channels. and The calculation process and similar.

[0038] The calculation process is: the fifth layer feature map After processing by Conv2d_1, we get (size is 20×20×256), and then processed by Conv2d_2, we get (Size is 20×20×256); The calculation process is: the fourth layer feature map After processing by Conv2d_1, we get (size is 40×40×256), and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain (Size is 40×40×256); The calculation process is: the third layer feature map After processing by Conv2d_1, we get (size is 80×80×256), and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain (Size is 80×80×256); The calculation process is: the second layer feature map After processing by Conv2d_1, we get (size is 160×160×256), and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain (Size is 160×160×256); in, It is the fifth layer feature map Input the FPN feature fusion network to obtain the feature map; It is the fourth layer feature map Input the FPN feature fusion network to obtain the feature map; It is the third layer feature map Input the FPN feature fusion network to obtain the feature map; It is the second layer feature map Input the FPN feature fusion network to obtain the feature map.

[0039] In the standard FPN network, Perform the maximum pooling process and get the size of The present invention does not adopt This is partly because defects usually account for a smaller proportion (smaller pixel width and height) in substation inspection images compared to the original input. The size of the image has been significantly downsampled, and the defect information is has been basically lost, so the present invention does not adopt part.

[0040] The design of the convolution module is as follows Figure 6 As shown in the figure, it is assumed that the width, height and number of channels of the input feature map are h, w and C respectively. The input feature map is processed in parallel by three convolutional layers, and the sizes of the convolution kernels are 3×3, 5×5 and 7×7 respectively. p represents the padding value of the feature map, s represents the stride value, and o represents the number of convolution kernels (i.e. the number of output feature maps). Convolution kernels of different sizes have receptive fields of different scales. Feature maps are extracted in parallel by multiple convolution kernels, and then the feature maps are spliced ​​(output ), which can realize the simultaneous extraction and fusion of features of different scales. The concatenated feature map is then subjected to 1×1 convolution and the output .

[0041] Specifically, the width, height and number of channels of the input feature map are set to , and ; The convolution module includes the first convolution layer, the second convolution layer, the third convolution layer and the fourth convolution layer; the convolution kernel size of the first convolution layer is 3×3, the feature map filling value p=1, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the second convolution layer is 5×5, the feature map padding value p=2, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the third convolution layer is 7×7, the feature map padding value p=3, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the fourth convolution layer is 1×1, the feature map padding value p=0, the step value s=1, and the number of convolution kernels ; Input feature map The first convolutional layer, the second convolutional layer, and the third convolutional layer extract feature maps in parallel, and then splice the feature maps to output Feature map, concatenated feature map Then pass through the fourth convolutional layer to output the feature map .

[0042] Step 5: Obtain defect detection results: Finally, the multi-scale pyramid feature map is sent to the detection head to detect defect targets.

[0043] like Figure 7 As shown in the figure, the two figures on the left are the inspection image (upper) and the reference image (lower), and the right is the inspection image and the reference image after the backbone network extracts features. The difference feature map on the layer feature map (i.e. element-by-element subtraction: ) It is easy to see that the differential feature map using dual-phase input can effectively enhance defect features and suppress background interference.

[0044] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting substation defects guided by variable attention, characterized in that: The following steps are involved: Step 1: Dual-phase input: Reference images collected by substation inspection equipment and inspection images , where the reference image It is a normal sample without defects collected by the substation inspection equipment at a certain inspection point. The samples collected by the substation inspection equipment at the same point and at different times need to be judged whether they contain defects; inspection images Automatically determine whether there are defects through the substation defect detection algorithm; Step 2: Feature extraction: Reference image of bi-phase input and inspection images They are respectively sent to the same backbone network for feature extraction, and multi-scale feature maps are extracted respectively; Step 3: Fusion Processing: Reference Image Multi-scale feature maps and inspection images The multi-scale feature map is fused by the variable attention guidance module to obtain the enhanced multi-scale feature map; Step 4, feature fusion and mapping: Send the enhanced multi-scale feature map to the feature pyramid network for further feature fusion and mapping to obtain a multi-scale pyramid feature map; Step 5: Obtain defect detection results: Finally, the multi-scale pyramid feature map is sent to the detection head to detect defect targets.

2. The method for detecting substation defects guided by variable attention according to claim 1 is characterized in that: In step 1, the substation defect detection algorithm adopts the Faster-RCNN framework.

3. The method for detecting substation defects guided by variable attention according to claim 1 is characterized in that: In step 2, the backbone network is a ResNet50 backbone network.

4. The method for detecting substation defects guided by variable attention according to claim 3 is characterized in that: In the ResNet50 backbone network, the size of the input image is W×H×3, where W is the width, H is the height, and 3 is the number of channels; The feature map size output after the convolution layer Conv1 is W / 2×H / 2×64; The second layer feature map output by the backbone network after the convolution layer Conv2_x The size is W / 4×H / 4×256; The third layer feature map output by the backbone network after the convolution layer Conv3_x The dimensions are W / 8×H / 8×512; The fourth layer feature map output by the backbone network after the convolution layer Conv4_x The size is W / 16×H / 16×1024; The fifth layer feature map output by the backbone network after the convolution layer Conv5_x The dimensions are W / 32×H / 32×2048.

5. A method for detecting substation defects guided by variable attention according to claim 4, characterized in that: In the feature pyramid network, The calculation process is: the fifth layer feature map After processing by Conv2d_1, we get , and then processed by Conv2d_2, we get ; Among them, Conv2d_1 represents a convolutional layer with a convolution kernel size of 1×1 and 256 channels; Conv2d_2 represents a convolutional layer with a convolution kernel size of 3×3 and 256 channels; The calculation process is: the fourth layer feature map After processing by Conv2d_1, we get , and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain ; The calculation process is: the third layer feature map After processing by Conv2d_1, we get , and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain ; The calculation process is: the second layer feature map After processing by Conv2d_1, we get , and then with The upsampled feature maps are added, and finally processed by Conv2d_2 to obtain ; in, It is the fifth layer feature map Input the FPN feature fusion network to obtain the feature map; It is the fourth layer feature map Input the FPN feature fusion network to obtain the feature map; It is the third layer feature map Input the FPN feature fusion network to obtain the feature map; It is the second layer feature map Input the FPN feature fusion network to obtain the feature map.

6. The method for detecting substation defects guided by variable attention according to claim 4 is characterized in that: In step 2, Reference Images Send it to the backbone network for feature extraction, and extract the multi-scale feature map , , and ;in, The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; Inspection image Send it to the backbone network for feature extraction, and extract the multi-scale feature map , , and ;in, The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map; The input image is hour, The corresponding feature map.

7. The method for detecting substation defects guided by variable attention according to claim 6 is characterized in that: In the third step, the process of the change attention guidance module fusion processing is: First, enter and Perform a difference operation, that is, subtract element by element ; Among them, input and Corresponding to the reference image and inspection images The feature map input of the same scale, that is, and Corresponding to and ,or and Corresponding to and ,or and Corresponding to and ,or and Corresponding to and ; Then, and The feature map obtained after element-by-element subtraction is sent to the convolution module, and the features are mapped through the convolution module; Next, the Sigmoid activation function is used to map the feature response to the (0,1) interval to obtain the weight ; at last, and Multiply element by element to get enhanced features .

8. The method for detecting substation defects guided by variable attention according to claim 7 is characterized in that: Set the width, height and number of channels of the input feature map to be , and ; The convolution module includes a first convolution layer, a second convolution layer, a third convolution layer and a fourth convolution layer; The convolution kernel size of the first convolution layer is 3×3, the feature map padding value p=1, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the second convolution layer is 5×5, the feature map padding value p=2, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the third convolution layer is 7×7, the feature map padding value p=3, the step value s=1, and the number of convolution kernels ; The convolution kernel size of the fourth convolution layer is 1×1, the feature map filling value p=0, the step value s=1, and the number of convolution kernels ; Input feature map The first convolutional layer, the second convolutional layer, and the third convolutional layer extract feature maps in parallel, and then splice the feature maps to output Feature map, concatenated feature map Then pass through the fourth convolutional layer to output the feature map .

Images