A power transformation equipment detection method, system, device and medium
By improving the YOLOv8n model, fusing local and contextual features, and introducing a multi-dimensional attention mechanism and feature modulation module, the problems of occlusion and small targets in substation equipment detection are solved, achieving more efficient detection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU POWER SUPPLY BUREAU GUANGDONG POWER GRID CO LTD
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-12
AI Technical Summary
The existing YOLOv8n target detection algorithm has difficulty effectively detecting obstructed and small targets in the complex environment of substations, resulting in poor detection performance.
By introducing the CGBlock context guidance module and the Triplet Attention mechanism, combined with the SAFM spatial adaptive feature modulation module, local features and surrounding context features are fused to generate channel attention weights. The enhanced feature maps are then evaluated and rearranged in multiple dimensions of importance. The feature maps are segmented and parallel feature transformation operations are performed. The optimal feature representation is selected for reconstruction.
It improves the robustness and accuracy of detection under obstructed and small target conditions, and enhances the accuracy and speed of power equipment detection.
Smart Images
Figure CN122200104A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target detection technology, and in particular to a method, system, equipment and medium for detecting power equipment. Background Technology
[0002] Power equipment inspection falls under the category of object detection. In recent years, with the rapid development of deep learning, deep learning-based object detection has largely replaced traditional image processing-based methods. Currently, deep learning-based object detection can be divided into two-stage and single-stage algorithms. A representative two-stage algorithm is the R-CNN series of region convolutional neural networks, which first generates candidate regions and then performs classification and regression, such as... Figure 1 As shown. The YOLO series is a representative single-stage algorithm, which omits the candidate region generation step and directly performs classification and regression, such as... Figure 2 As shown, the two-stage algorithm has a higher accuracy, but it is difficult to implement, has a huge computational load, a slow detection speed, and high requirements for computer equipment performance; the single-stage algorithm has a slightly lower accuracy, but a faster detection speed, making it more suitable for real-time tasks.
[0003] The existing technology includes the YOLOv8n target detection algorithm, which has the advantages of low computational cost and fast inference speed, and can efficiently detect and identify different targets. However, in real substation environments, various power equipment is usually placed in a complex and intertwined manner, which can easily obscure each other. Furthermore, due to differences in observation positions, some power equipment may appear as small targets in the image, which increases the difficulty of the original target detection algorithm in detecting power equipment in these extreme cases, thus greatly reducing the effectiveness of power equipment detection. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, device, and medium for testing power equipment, thereby solving the problems in the prior art.
[0005] The present invention specifically provides the following technical solution: a method for testing power equipment, comprising the following steps: Collect images of power equipment; Feature extraction is performed on the image of the power equipment to obtain an initial feature map; Local features and their surrounding context features are extracted from the initial feature map. The local features and surrounding context features are fused to generate channel attention weights. The channel attention weights are then used to weight and adjust the initial feature map to obtain an enhanced feature map. The importance of information in the enhanced feature map is evaluated from multiple dimensions, and the information in the enhanced feature map in different dimensions is rearranged and fused according to the evaluation results to obtain an aggregated feature map. The aggregated feature map is divided into multiple feature components. The first feature transformation operation and the second feature transformation operation are performed in parallel on different feature components. The optimal feature representation is adaptively selected from the transformation results. The image is reconstructed based on the selected optimal feature representation. The reconstructed image is then classified and identified, and the detection and classification results of the substation equipment are output.
[0006] Preferably, the substation equipment image is detected using an improved YOLOv8n model, and the detection and classification results of the substation equipment are output: The improved YOLOv8n model is based on the basic YOLOv8n network. Multiple CGBlock context guidance modules are introduced into the backbone and neck parts of the basic YOLOv8n network, and a Triplet Attention mechanism is introduced before the SPPF layer of the backbone network. The Upsample layer of the YOLOv8n baseline network is replaced with the SAFM spatial adaptive feature modulation module.
[0007] Preferably, the multiple CGBlock context bootstrapping modules replace the Conv layers after the first C2f layer and the third C2f layer in the backbone, respectively, and replace the Conv layers after the second C2f layer in the neck.
[0008] Preferably, the importance of information in the enhanced feature map is evaluated from multiple dimensions, and the information in the enhanced feature map is rearranged and fused according to the evaluation results to obtain an aggregated feature map, specifically: In the Triplet Attention mechanism, the tensor of shape C×H×W in the enhanced feature map is rotated 90° counterclockwise along the H axis and Z-pooling is performed. The result of Z-pooling is then processed by convolution and batch normalization to obtain a tensor of shape 1×H×C. The sigmoid activation function is applied to the tensor of shape 1×H×C to generate attention weights in channel dimension C and spatial dimension H. The attention weights are applied to the rotated tensor and then rotated 90° clockwise along the H axis to restore it to the original input shape, which is then used as the first output. Set the rotation axis to the W axis, obtain a tensor of shape 1×C×W, and generate attention weights between the channel dimension C and the spatial dimension W. Apply the attention weights to the rotated tensor and rotate it 90° clockwise along the W axis to restore it to the original input shape, which is then used as the second output. The number of channels in the tensor of the enhanced feature map is reduced to two dimensions by Z-pooling, and batch normalization of the convolution kernel is performed to obtain a tensor of shape 1×H×W. Attention weights are generated by the Sigmoid function and applied to the original input tensor as the third output. The first, second, and third outputs are averaged and aggregated as different permutations to obtain the final aggregated feature map.
[0009] Preferably, the aggregated feature map is divided into multiple feature components, and a first feature transformation operation and a second feature transformation operation are performed in parallel on different feature components. The optimal feature representation is adaptively selected from the transformation results, specifically as follows: In the SAFM spatial adaptive feature modulation module, the aggregated feature map is segmented into channels to obtain multiple feature components; Multiple feature components are subjected to a first feature transformation operation and a second feature transformation operation respectively to learn the long-distance dependency relationship of the feature components. Based on the long-distance dependency relationship, features with a correlation greater than a threshold are dynamically selected for high-resolution image reconstruction. The first feature transformation operation is a 3×3 depth convolution process, and the second feature transformation operation is an adaptive max pooling process.
[0010] Preferably, when fusing local features and surrounding context features to generate channel attention weights, the local features and surrounding context features are concatenated in multiple CGBlock context guidance modules. The concatenated data is activated by batch normalization (BN) and parameterized ReLU to obtain joint features that fuse local and surrounding context. Global average pooling is then used to aggregate global context information to generate channel attention weights.
[0011] This invention provides a power equipment testing system, comprising: The data acquisition module is used to collect images of power equipment. An initial feature extraction module is used to extract features from the image of the substation equipment to obtain an initial feature map; The enhanced feature extraction module is used to extract local features and their surrounding context features from the initial feature map, fuse the local features and surrounding context features to generate channel attention weights, and use the channel attention weights to perform weighted adjustments on the initial feature map to obtain the enhanced feature map. The attention output module is used to evaluate the importance of information in the enhanced feature map from multiple dimensions, and rearrange and fuse the information in the enhanced feature map in different dimensions according to the evaluation results to obtain an aggregated feature map. The detection module is used to segment the aggregated feature map into multiple feature components, perform a first feature transformation operation and a second feature transformation operation in parallel on different feature components, adaptively select the optimal feature representation from the transformation results, reconstruct the image based on the selected optimal feature representation, classify and identify the reconstructed image, and output the detection and classification results of the substation equipment.
[0012] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method for detecting power equipment.
[0013] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for detecting power equipment.
[0014] Compared with the prior art, the present invention has the following advantages: This invention generates channel attention weights by fusing local features with surrounding contextual features, thereby providing enhanced feature maps with higher information completeness. This fundamentally alleviates the feature loss problem caused by occlusion. Furthermore, it performs cross-dimensional importance assessment and rearrangement of the enhanced feature maps, selecting the most critical responses from multiple dimensions such as space and channels. This allows for the focusing and reorganization of originally weak and scattered features, ensuring that features of small targets are not obscured in subsequent processing. Additionally, it segments the aggregated feature maps and executes different feature transformation operations in parallel, enabling more efficient capture of deep patterns in different dimensions. Based on the specific situation of the current image, it intelligently selects the optimal feature representation for reconstruction and classification, improving the robustness and accuracy of detection in occlusion and small target conditions. Attached Figure Description
[0015] Figure 1 Here is a flowchart of the two-stage detection process in the background technology; Figure 2 Here is a flowchart of a single-stage detection process in the background technology; Figure 3 This is the original YOLOv8n network diagram in this embodiment of the invention; Figure 4 The Triplet Attention structure diagram provided by this invention; Figure 5 The CGBlock structure diagram provided by this invention; Figure 6 The SAFM structure diagram provided by this invention; Figure 7 This is a structural diagram of the improved YOLOv8n of this invention; Figure 8 This is a thermal comparison diagram of the busbar before and after the improvement of this invention; wherein, Figure 8 (a) is the busbar thermal diagram before the improvement. Figure 8 (b) is the improved busbar thermal diagram; Figure 9 This is a thermal comparison diagram of the switch before and after the improvement of this invention; wherein, Figure 9 (a) is the switch heat map before the improvement. Figure 9 (b) is the improved switch heat map; Figure 10 This is a diagram illustrating the detection effect of small targets and occlusion conditions in this invention; wherein, Figure 10 (a) is a small target image of a sparse scene. Figure 10 (b) is a small target image of a dense scene switch. Figure 10 (c) represents occlusion of a sparse scene switch. Figure 10 (d) is a switch occlusion map of a dense scene; Figure 11 This invention demonstrates the effectiveness of busbar shading detection. Figure 12 The flowchart illustrates a method for testing power equipment provided by this invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0017] like Figure 12 As shown, an embodiment of the present invention provides a method for testing power equipment, including the following steps: Step 1: Collect images of the power equipment.
[0018] Step 2: Extract features from the image of the power equipment to obtain an initial feature map.
[0019] Step 3: Extract local features and their surrounding context features from the initial feature map, fuse the local features and surrounding context features to generate channel attention weights, and use the channel attention weights to weight and adjust the initial feature map to obtain the enhanced feature map.
[0020] An improved YOLOv8n model is used to detect images of substation equipment, and the detection and classification results of the substation equipment are output. The improved YOLOv8n model is as follows: A basic YOLOv8n network is built based on the backbone, neck, and head, such as... Figure 3 As shown.
[0021] Multiple CGBlock context guidance modules are introduced into the backbone and neck parts of the basic YOLOv8n network. A Triplet Attention mechanism is introduced before the SPPF layer of the backbone network. The 13th upsample layer of the YOLOv8n baseline network is replaced with a SAFM spatial adaptive feature modulation module. Specifically, the multiple CGBlock context guidance modules replace the Conv layers after the first and third C2f layers in the backbone, and the Conv layers after the second C2f layer in the neck, respectively. The improved YOLOv8n network is shown below. Figure 7 As shown.
[0022] The YOLOv8n baseline network has a limited receptive field for its Conv convolutions, which easily overlooks contextual information and has a large number of parameters. To expand the receptive field and reduce the number of parameters, this invention introduces a CGBlock (Context Guided Block) module in the backbone and neck regions, with the structure as follows: Figure 5 As shown.
[0023] The input feature map is adjusted for channel count using 1×1 convolutions and then fed in parallel to the local feature extractor and the surrounding context extractor. The local feature extractor uses 3×3 channel-separating convolutions to learn local features from the eight neighboring feature vectors. The surrounding context extractor utilizes 3×3 dilated convolutions to more efficiently capture surrounding contextual information, expanding the receptive field without increasing the number of parameters and avoiding the parameter explosion caused by stacking standard convolutional layers. In multiple CGBlock context-guided modules, the features output from these two branches are concatenated along the channel dimension in the joint feature extractor and activated by batch normalization (BN) and parameterized ReLU (PReLU) to form a joint feature that integrates local and surrounding contextual information. Subsequently, the global context extractor aggregates global contextual information through global average pooling, generates channel attention weights using a multilayer perceptron, and then adaptively reweights the joint features to enhance key features and suppress redundant responses.
[0024] Step 4: Evaluate the importance of information in the enhanced feature map from multiple dimensions, and rearrange and fuse the information in the enhanced feature map in different dimensions according to the evaluation results to obtain the aggregated feature map.
[0025] When training the YOLOv8 object detection model, the network needs to process a large amount of input data, but only a small portion of the input data is of substantial importance, with the majority being noise. To make the model focus more on the features of the substation equipment rather than background information, this invention introduces a Triplet Attention mechanism before the SPPF layer of the backbone network, with the structure as follows: Figure 4 As shown.
[0026] In the Triplet Attention mechanism, the upper branch is responsible for calculating the attention weights for the channel dimension C and the spatial dimension H. It transforms the input tensor of shape C×H×W in the enhanced feature map into a shape of W×H×C by rotating it 90° counterclockwise along the H-axis, performing Z-pooling, and then passing the result of Z-pooling through standard convolution and batch normalization to obtain a tensor of shape 1×H×C. The sigmoid activation function is then applied to this 1×H×C tensor to generate the attention weights for the channel dimension C and the spatial dimension H. These weights are applied to the rotated tensor, which is then rotated 90° clockwise along the H-axis to restore the original input shape, serving as the first output.
[0027] The middle branch is responsible for capturing the dependency between the channel dimension C and the spatial dimension W. Except that the rotation axis is the W axis, it performs the same operation as the upper branch. It sets the rotation axis to the W axis, obtains a tensor of shape 1×C×W, and generates attention weights between the channel dimension C and the spatial dimension W. The attention weights are applied to the rotated tensor and rotated 90° clockwise along the W axis to restore it to the original input shape, which is the second output.
[0028] The lower branch is used to capture the dependencies between spatial dimensions. This branch does not change the input, reduces the number of channels of the tensor in the enhanced feature map to two dimensions through Z-pooling, performs batch normalization of the convolution kernel to obtain a tensor of shape 1×H×W, and generates attention weights through the Sigmoid function and applies them to the original input tensor as the third output.
[0029] After processing the enhanced feature map using a triple attention mechanism, each branch generates attention weights, and the enhanced features are permuted using these attention weights. The first, second, and third outputs are then averaged and aggregated as different permutation results, ultimately yielding the triple attention output (i.e., the aggregated feature map). This three-branch structure, through different rotation and permutation operations, can integrate information from different dimensions, better capturing the intrinsic features of the data. Furthermore, this method is computationally efficient, enhancing the network's ability to understand and process complex data structures with almost no increase in model complexity or computational burden.
[0030] Step 5: Divide the aggregated feature map into multiple feature components, and perform the first feature transformation operation and the second feature transformation operation in parallel on different feature components. Adaptively select the optimal feature representation from the transformation results, reconstruct the image based on the selected optimal feature representation, and classify and identify the reconstructed image to output the detection and classification results of the substation equipment.
[0031] The YOLOv8n baseline network upsamples only local features, ignoring global features, thus affecting detection accuracy. This invention replaces the 13th upsample layer of the YOLOv8n baseline network with a SAFM (Spatially-Adaptive Feature Modulation) module, the structure of which is as follows: Figure 6 As shown.
[0032] In the SAFM spatial adaptive feature modulation module, the normalized aggregated feature map is input into the SAFM spatial adaptive feature modulation module for channel segmentation to obtain multiple feature components. In one embodiment, four components are obtained. The multiple feature components are subjected to a first feature transformation operation and a second feature transformation operation, respectively. Specifically, one part is processed by 3×3 depth convolution, and the rest are fed into a multi-scale feature generation unit. At the same time, adaptive max pooling is applied to the input features to generate multi-scale features, thereby learning long-distance dependencies from multi-scale feature representations and dynamically selecting representative features for high-resolution image reconstruction to better mine useful features. The first feature transformation operation is 3×3 depth convolution processing, and the second feature transformation operation is adaptive max pooling processing.
[0033] In one experiment, 3216 images of power equipment were collected at a substation to construct a dataset. The average accuracy (mAP) at 0.5:0.95, the number of parameters, the number of floating-point operations per second (FLOPs), and the weights were used as model evaluation metrics. To verify the effectiveness of the three improvements, an ablation experiment was conducted, and the results are shown in Table 1.
[0034] Table 1 Ablation Experiment Results In the table, A - baseline network YOLOv8n; B - YOLOv8n + Triplet Attention; C - YOLOv8n + Triplet Attention + CGBlock; D - YOLOv8n + Triplet Attention + CGBlock + SAFM.
[0035] As can be seen, after adding the Triplet Attention module, the mAP@0.5:0.95 metric increased by 0.5%, with almost no increase in the number of model parameters or computational complexity, only an increase of 0.01M in model weights; after introducing the CGBlock context guidance module, the mAP@0.5:0.95 metric increased by 0.5% compared to model B, and the number of model parameters, floating-point operations per second, and weight size decreased by 0.1M, 0.1GFLOPs, and 0.19M, respectively; after adding the SAFM spatial adaptive feature modulation module, the mAP@0.5:0.95 metric reached 77.7%, an increase of 2.3% compared to model C, and an improvement of 3.3% compared to the original YOLOv8n baseline network.
[0036] To demonstrate the superiority of the proposed algorithm, it was compared with other mainstream target detection algorithms. The experimental results are shown in Table 2.
[0037] Table 2 Comparison of experimental results It can be seen that the algorithm proposed in this invention has significantly better accuracy than other models, and also has certain advantages in terms of parameter quantity and model size, achieving an excellent balance between detection accuracy and speed.
[0038] The improved YOLOv8n model was used to test two types of transformer equipment: busbars and switches. The thermal comparison diagram is shown below. Figure 8 , Figure 9 As shown.
[0039] As can be seen, the improved YOLOv8n heatmap focuses on more accurate feature information, indicating that the improvements effectively remove a large amount of irrelevant interference, allowing the network to focus more on the characteristics of the substation equipment. The results are shown in the actual switch and busbar images. Figure 10 , Figure 11 As shown.
[0040] It is easy to see that the algorithm of this invention can effectively detect switches with occlusion and small targets in both dense and sparse scenes; for buses with overlapping occlusion, the algorithm of this invention can still effectively separate and detect them without mixing them together.
[0041] Based on the above method, the present invention provides a power equipment detection system, including: a data acquisition module, an initial feature extraction module, an enhanced feature extraction module, an attention output module, and a detection module.
[0042] The system comprises the following modules: a data acquisition module for collecting images of substation equipment; an initial feature extraction module for extracting features from the substation equipment images to obtain an initial feature map; an enhanced feature extraction module for extracting local features and their surrounding context features from the initial feature map, fusing the local features and surrounding context features to generate channel attention weights, and using the channel attention weights to weight and adjust the initial feature map to obtain an enhanced feature map; an attention output module for evaluating the importance of information in the enhanced feature map from multiple dimensions, and rearranging and fusing the information in the enhanced feature map in different dimensions based on the evaluation results to obtain an aggregated feature map; and a detection module for segmenting the aggregated feature map into multiple feature components, performing a first feature transformation operation and a second feature transformation operation in parallel on different feature components, adaptively selecting the optimal feature representation from the transformation results, reconstructing the image based on the selected optimal feature representation, classifying and recognizing the reconstructed image, and outputting the detection and classification results of the substation equipment.
[0043] The present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a method for detecting power equipment.
[0044] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).
[0045] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a method for detecting power equipment.
[0046] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0047] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.
Claims
1. A method for testing power equipment, characterized in that, Includes the following steps: Collect images of power equipment; Feature extraction is performed on the image of the power equipment to obtain an initial feature map; Local features and their surrounding context features are extracted from the initial feature map. The local features and surrounding context features are fused to generate channel attention weights. The channel attention weights are then used to weight and adjust the initial feature map to obtain an enhanced feature map. The importance of information in the enhanced feature map is evaluated from multiple dimensions, and the information in the enhanced feature map in different dimensions is rearranged and fused according to the evaluation results to obtain an aggregated feature map. The aggregated feature map is divided into multiple feature components. The first feature transformation operation and the second feature transformation operation are performed in parallel on different feature components. The optimal feature representation is adaptively selected from the transformation results. The image is reconstructed based on the selected optimal feature representation. The reconstructed image is then classified and identified, and the detection and classification results of the substation equipment are output.
2. The method for testing power equipment as described in claim 1, characterized in that, The improved YOLOv8n model is used to detect the images of the substation equipment, and the detection and classification results of the substation equipment are output: The improved YOLOv8n model is based on the basic YOLOv8n network. Multiple CGBlock context guidance modules are introduced into the backbone and neck parts of the basic YOLOv8n network, and a Triplet Attention mechanism is introduced before the SPPF layer of the backbone network. The Upsample layer of the YOLOv8n baseline network is replaced with the SAFM spatial adaptive feature modulation module.
3. The method for testing power equipment as described in claim 2, characterized in that, Multiple CGBlock context bootstrapping modules replace the Conv layers after the first C2f layer and the third C2f layer in the backbone, respectively, and replace the Conv layers after the second C2f layer in the neck.
4. The method for testing power equipment as described in claim 2, characterized in that, The importance of information in the enhanced feature map is evaluated from multiple dimensions, and the information in the enhanced feature map is rearranged and fused according to the evaluation results to obtain an aggregated feature map, specifically: In the Triplet Attention mechanism, the tensor of shape C×H×W in the enhanced feature map is rotated 90° counterclockwise along the H axis and Z-pooling is performed. The result of Z-pooling is then processed by convolution and batch normalization to obtain a tensor of shape 1×H×C. The sigmoid activation function is applied to the tensor of shape 1×H×C to generate attention weights in channel dimension C and spatial dimension H. The attention weights are applied to the rotated tensor and then rotated 90° clockwise along the H axis to restore it to the original input shape, which is then used as the first output. Set the rotation axis to the W axis, obtain a tensor of shape 1×C×W, and generate attention weights between the channel dimension C and the spatial dimension W. Apply the attention weights to the rotated tensor and rotate it 90° clockwise along the W axis to restore it to the original input shape, which is then used as the second output. The number of channels in the tensor of the enhanced feature map is reduced to two dimensions by Z-pooling, and batch normalization of the convolution kernel is performed to obtain a tensor of shape 1×H×W. Attention weights are generated by the Sigmoid function and applied to the original input tensor as the third output. The first, second, and third outputs are averaged and aggregated as different permutations to obtain the final aggregated feature map.
5. The method for testing power equipment as described in claim 2, characterized in that, The aggregated feature map is segmented into multiple feature components, and a first feature transformation operation and a second feature transformation operation are performed in parallel on different feature components. The optimal feature representation is adaptively selected from the transformation results. Specifically: In the SAFM spatial adaptive feature modulation module, the aggregated feature map is segmented into channels to obtain multiple feature components; Multiple feature components are subjected to a first feature transformation operation and a second feature transformation operation respectively to learn the long-distance dependency relationship of the feature components. Based on the long-distance dependency relationship, features with a correlation greater than a threshold are dynamically selected for high-resolution image reconstruction. The first feature transformation operation is a 3×3 depth convolution process, and the second feature transformation operation is an adaptive max pooling process.
6. The method for testing power equipment as described in claim 2, characterized in that, When fusing local features and surrounding context features to generate channel attention weights, the local features and surrounding context features are concatenated in multiple CGBlock context guidance modules. The concatenated data is activated by batch normalization (BN) and parameterized ReLU to obtain joint features of the fused local and surrounding context. Global average pooling is then used to aggregate global context information to generate channel attention weights.
7. A power equipment testing system, characterized in that, include: The data acquisition module is used to collect images of power equipment; An initial feature extraction module is used to extract features from the image of the substation equipment to obtain an initial feature map; The enhanced feature extraction module is used to extract local features and their surrounding context features from the initial feature map, fuse the local features and surrounding context features to generate channel attention weights, and use the channel attention weights to perform weighted adjustments on the initial feature map to obtain the enhanced feature map. The attention output module is used to evaluate the importance of information in the enhanced feature map from multiple dimensions, and rearrange and fuse the information in the enhanced feature map in different dimensions according to the evaluation results to obtain an aggregated feature map. The detection module is used to segment the aggregated feature map into multiple feature components, perform a first feature transformation operation and a second feature transformation operation in parallel on different feature components, adaptively select the optimal feature representation from the transformation results, reconstruct the image based on the selected optimal feature representation, classify and identify the reconstructed image, and output the detection and classification results of the substation equipment.
8. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the power equipment detection method as described in any one of claims 1 to 6.
9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the substation equipment detection method according to any one of claims 1 to 6.